chore(version): 1.7.2 -> 1.7.3

启动重复播种 admin key 与配置封存非幂等的修复，走 patch 发布。

.gitignore

@@ -12,8 +12,17 @@ cmd/gui/bin/
 cmd/gui/build/gui-dist/
 cmd/build/
 cmd/gui/*.log
+cmd/gui/.cache/
 # workspace (AI-assisted coding local state)
 .pi/
 .pi-glla/
 .codegraph/
+.omo/
 /build/
+
+# local ops scripts (machine-specific, not for sharing)
+scripts/
+cmd/gui/dist/
+
+# local design/working notes (not part of the shipped repo)
+/plan.md

Makefile

@@ -9,6 +9,12 @@ export GUI_DIR
 GUI_DIR=cmd/gui
 VERSION ?= $(shell git describe --tags --dirty 2>/dev/null || echo "0.1.0")
 
+# Clean env for GUI packaging: the host may have LD_LIBRARY_PATH polluted by
+# third-party runtimes (e.g. /opt/cangjie) which electron-builder bakes into
+# the produced binary's DT_RPATH/dep resolution -> the packaged app then
+# SIGSEGVs inside ld.so on clean machines.  Strip it for every pack step.
+CLEAN_PACK_ENV = env -u LD_LIBRARY_PATH
+
 all: build
 
 # ---- core binary (server users / embedded core) ----
@@ -17,6 +23,14 @@ build:
 	CGO_ENABLED=1 go build -tags luajit -trimpath -o $(BUILD_DIR)/$(BINARY) ./cmd/llmsproxy
 	@echo "Built: $(BUILD_DIR)/$(BINARY) ($(VERSION))"
 
+# ---- core distribution packages (deb + rpm + tar.gz) -----
+# Server admins install these with dpkg/rpm; unlike the GUI they are the bare
+# gateway binary + systemd unit (with memory tuning) + Lua adapters.
+core-dist:
+	./packaging/core-dist.sh $(VERSION)
+
+.PHONY: core-dist
+
 # ---- GUI (Electron) ----
 gui:
 	@cd $(GUI_DIR) && npm start
@@ -25,38 +39,44 @@ gui-dev:
 	@cd $(GUI_DIR) && npm run dev
 
 # build the embedded core into the GUI, then run electron-builder
-# targets: deb + AppImage (linux), rpm needs system rpmbuild
+# targets: deb + rpm + AppImage (linux); rpm needs system rpmbuild
+# NOTE: run in a clean env (no LD_LIBRARY_PATH pollution) so the packaged
+# binary does not link against host-only libs and crash on user machines.
 gui-dist:
-	@cd $(GUI_DIR) && npm run dist:linux
+	@cd $(GUI_DIR) && $(CLEAN_PACK_ENV) npm run dist:linux
+	./packaging/verify-dist.sh
 
 # deb only (for local sharing), rpm needs rpmbuild
 gui-deb:
-	@cd $(GUI_DIR) && npm run dist:debian
+	@cd $(GUI_DIR) && $(CLEAN_PACK_ENV) npm run dist:debian
 
 # windows nsis installer: host-built core (needs apt install gcc-mingw-w64-x86-64 + wine)
 gui-win:
-	@cd $(GUI_DIR) && npm run dist:win
+	@cd $(GUI_DIR) && $(CLEAN_PACK_ENV) npm run dist:win
 
 # windows nsis installer: full cross-build in docker (self-sufficient, no host mingw)
 gui-win-docker:
-	@cd $(GUI_DIR) && ./scripts/dist-win-docker.sh
+	@cd $(GUI_DIR) && $(CLEAN_PACK_ENV) ./scripts/dist-win-docker.sh
+	./packaging/verify-dist.sh
 
 gui-dist-dir:
-	@cd $(GUI_DIR) && npm run dist:dir
+	@cd $(GUI_DIR) && $(CLEAN_PACK_ENV) npm run dist:dir
 
 test:
 	go test ./...
 
 clean:
-	rm -rf $(BUILD_DIR) cmd/gui/build/gui-dist cmd/gui/bin
+	rm -rf $(BUILD_DIR) cmd/build cmd/gui/dist cmd/gui/build/gui-dist cmd/gui/bin
 
 help:
 	@echo "Targets:"
 	@echo "  make build      build core binary (luajit)"
+	@echo "  make core-dist  build core packages: deb + rpm + tar.gz (needs nfpm)"
 	@echo "  make gui        run desktop GUI (dev)"
 	@echo "  make gui-dev    run GUI with devtools"
-	@echo "  make gui-dist   build deb + AppImage (linux)"
+	@echo "  make gui-dist   build deb + rpm + AppImage (linux)"
 	@echo "  make gui-deb    build deb only"
 	@echo "  make gui-win    build Windows nsis (needs host mingw + wine)"
 	@echo "  make gui-win-docker  build Windows nsis all-in-docker (no host mingw needed)"
 	@echo "  make test       run go tests"
+	@echo "  make clean      remove build dirs and packaged installers (build/, cmd/build/, cmd/gui/dist/)"

README.md

@@ -15,15 +15,18 @@
 
 ### 极致轻量
 
-- **单二进制**：编译后约 10MB，零运行时依赖（仅依赖系统 libc），部署即用
-- **极低内存占用**：空闲状态仅 ~15MB RSS，满载并发 100+ 请求时峰值 < 100MB
+- **单二进制**：编译后 8~12MB（`-s -w` strip 后约 8MB），零运行时依赖（仅依赖系统 libc），部署即用
+- **低内存占用**：与源数量相关，非与运行时长相关——单源 ~10MB、16 源生产实例 ~32–35MB（[实测分解与调优](#内存占用实测与调优)）
+- **日志按需加载**：统计数字来自**全量**审计日志（启动时流式扫一遍即释放，29MB/22 万行约 260ms），但原始记录不常驻内存——默认只加载首屏，下滚自动分页，CSV 导出流式写出（O(1) 内存），离页即释放
 - **零运行时依赖**：纯 Go + LuaJIT 静态链接，无需安装 Python/Node/Java 等运行时
 - **启动极快**：冷启动 < 200ms，热重载配置 < 10ms
 
 ### 强大的多租户调度能力
 
 - **多密钥多租户**：支持无限密钥，每个密钥独立角色、模型范围、Token 配额、重置周期
+- **按模型配额**：每把 key 的每个模型单独配 token + 请求数配额与重置周期（小时/周/月/自定义 N 小时）；一个模型用满只影响该模型，同 key 其它模型照常；耗尽返 429 + `Retry-After` 并点名模型，admin key 永不受限
 - **AUTO 智能调度**：基于优先级档位的分级调度，同优先级源自动轮询负载均衡，故障自动毫秒级故障转移
+- **自愈冷却**：冷却上限 5 分钟，过半后放行 1 个探测请求，上游/额度恢复即刻回归轮询，无需等满冷却窗口
 - **Token 配额管理**：精确到模型级别的 Token 配额控制，支持小时/周/月/自定义小时周期自动重置
 - **同优先级源负载均衡**：同一优先级档位的多个源，请求自动 Round-Robin 均匀分发，故障毫秒级故障转移
 - **Source-Model 前缀路由**：支持 `source-model`/`source:model`/`source/model` 精确指定上游源
@@ -37,7 +40,7 @@
 
 ### 灵活的协议适配
 
-- **LuaJIT VM**：每个适配器独立 VM + worker 池，安全并发，Lua 脚本热加载无需重启
+- **LuaJIT VM**：每个适配器独立 VM + **弹性 worker 池**（按需扩缩容，空闲时自动回收），安全并发，Lua 脚本热加载无需重启
 - **Lua 适配器协议**：`transform_request` / `transform_response` / `transform_stream_chunk` 双向转换
 - **动态请求头钩子**：`build_headers(meta)` 支持 HMAC 签名、动态 Header 注入
 - **多模态透传**：`image_url` 等多模态内容在多源间无损透传，Anthropic/Gemini/Ollama 自动转换
@@ -60,6 +63,12 @@ Groq、Mistral、Ollama、KimiCode…）通过 **Lua 适配器** 做协议转换
 - **生图**：`POST /v1/images/generations`，`kind: image` 的模型独立路由。
 - **多模态**：`content` 数组（`image_url` 等）在多源间无损透传，Anthropic/Gemini/Ollama 自动转换。
 - **LuaJIT VM**：基于 golua 绑定的 LuaJIT，每个适配器独立 VM + worker 池，安全并发。
+  池是 **弹性** 的：`Σ max_concurrent` 只是上限而非预分配，状态按需创建、空闲时自动回收，
+  因此空载网关几乎不持有 Lua 状态（启动 0 个）。
+  - 扩容步长由适配器最大并发决定：`clamp(ceil(max/8), 1, 8)`，且仅在**真实并发争用**
+    （所有现有状态均已占用）时批量预热，顺序流量始终只用 1 个状态。
+  - 缩容步长由当前连接数决定：`clamp(ceil(冗余/(1+占用)), 1, 冗余)`——无连接时一轮收到
+    常驻下限，忙时每轮只释放 1 个，保护热路径。WebUI「适配器」页展示实时池指标。
 - **disable_thinking**：请求 `disable_thinking:true`（或上游对应字段）动态开关推理。
 - **Lua 适配协议**：每个源挂一个 `.lua` 适配器，完成 `transform_request` /
   `transform_response` / `transform_stream_chunk` 双向转换，协议差异全在 Lua 层。
@@ -89,8 +98,7 @@ GOMODCACHE=... GOPROXY=off go build -tags luajit -o llmsproxy ./cmd/llmsproxy
 
 > **配置不入库**：仓库不携带任何 config 文件（配置文件含密钥）。二进制首次运行时
 > 会在 `-config` 指定的位置生成默认配置：随机 admin key（打印在启动日志中）、
-> 仅绑定 `127.0.0.1:8080`、内置一个免 key 的 opencode zen 源可直接对话。
-> 首次登录后请在 WebUI「密钥」页更换管理员密钥。
+> 仅绑定 `127.0.0.1:8080`。首次登录后请在 WebUI「密钥」页更换管理员密钥。
 
 ```bash
 # 无 key -> 401
@@ -171,6 +179,81 @@ sources:
 - 客户端用任意一个已授权的密钥明文作为 Bearer（`Authorization: Bearer <key>`）。
 - 删除密钥即从运行时存储移除，立即失效。
 
+#### 密钥用量配额
+
+**配额按模型单独设置**：每个密钥的 `models[]` 里，每一条模型范围各自带一份
+token 配额与请求数配额。一个模型用满只影响该模型，同一密钥的其它模型照常工作。
+
+```yaml
+keys:
+  - key: sk-gw-<hex>
+    role: user
+    name: agent-alice
+    models:
+      - model: deepseek-v4-flash
+        token_quota: 1000000   # 本周期内该 key 用这个模型的 token 预算
+        req_quota: 20000       # 本周期内的请求次数
+        period: nhour          # "" | hour | week | month | nhour
+        hours: 6               # 仅 nhour
+      - model: AUTO
+        token_quota: 5000000   # AUTO 也是一条独立配额
+        period: day            # ← 这种写法会被拒绝（词表只有 hour/week/month/nhour）
+      - model: kimi-k3        # 未列配额 = 无限
+```
+
+- **没有「整钥总配额」**：这是刻意的设计。整钥总额会让「把 A 模型的额度挪给
+  B 模型」变成一次全局重分配；按模型独立计费则每个模型各自可控，运维可以
+  看出哪个模型吃掉了预算。
+- `period` 词表：空 = 永不过期（累计总量），`hour` / `week` / `month` = 固定窗口，
+  `nhour` + `hours` = 自定义小时数。**拼错的周期在写入时就被拒**，不会静默变成
+  永不过期。
+- 配额严格按密钥隔离，且**同一密钥内按模型隔离**：A 密钥用满 `m1` 不会消耗
+  B 密钥的额度，同一密钥的 `m2` 也不受影响。
+- 配额统计含聊天、流式、生图，跨重启从审计日志回放（保留 40 天，覆盖最长的
+  month 窗口）。
+- 配额耗尽返回 **429 + `Retry-After`**（`rate_limit_exceeded`），消息里点名是哪个
+  模型用满了，客户端可等窗口重置后自动恢复；模型越权才是 403。
+  **admin 密钥永不受配额限制**（其 scope 上的配额也不强制），避免把管理员
+  锁在门外。
+- 窗口用量按整点小时分桶统计，实际释放比配置窗口最多晚 1 小时（配额宁可晚释放
+  也不超发）。
+- `PUT /api/keys/{key}` 提交 `models` 即同时提交它们的配额（配额就是 scope 的一部分，
+  不存在会与模型列表脱节的第二份预算）。显式 `0` = 解除该模型的限制。
+
+##### 配额拒绝 vs 容量拒绝：两种「拒绝」含义不同
+
+配额在**入口**检查，不占用任何上游槽位，因此是廉价拒绝；容量不足则要走
+`busyWait` 有界等待后才返回 503。实测（容量 8、0.6s/请求、100 并发）：
+
+| 情况 | 状态码 | 延迟 | 是否打到上游 |
+|---|---|---|---|
+| 配额用尽（`req_quota` 50 → 100 并发） | `429 rate_limit_exceeded` + `Retry-After` | **19–21ms** | 否 |
+| 容量打满（同一拓扑，未限配额） | `503 upstream_error` | ~2.6s（`busyWait`） | 超出部分否 |
+
+客户端因此可以区分「等窗口重置」与「等上游腾出容量」：前者只需等待，后者
+通常该降并发或换源。
+
+##### 配额的开销
+
+配额桶按 (密钥, 模型, 整点小时) 分桶保留 40 天，实测（AMD 7840HS）：
+
+- 窗口查询按窗口长度而非保留总量扫描：1h 窗口 49ns、24h 窗口 55ns、
+  30d 窗口 3.9μs（此前全扫保留总量，960 桶时 5.9μs）。
+- 记录一条请求：283ns、3 分配（与引入配额前相同，分配来自 ring buffer）。
+- 内存：生产形态（7 密钥 × 8 模型 × 2 源 × 满 40 天 retention）约 **3.7MB**。
+  按源 pin 的 `source::model` 桶**惰性创建**——只有当某条配额真的 pin 了
+  某个源时才维护，否则每条记录多写一份桶，在 20 密钥 × 8 模型 × 3 源下会
+  多占 18MB。
+  代价：配置 pinned 配额之前的历史用量无法事后按源拆分，所以 **pinned 配额的
+  首个窗口可能少算**。
+
+```bash
+# 配额耗尽时客户端看到（点名了具体模型）
+HTTP/1.1 429 Too Many Requests
+Retry-After: 2100
+{"error":{"type":"rate_limit_exceeded","message":"token quota exceeded for \"deepseek-v4-flash\" (5000000/5000000)"}}
+```
+
 ### 模型路由
 
 `/v1/chat/completions` 的 `model` 解析顺序：
@@ -185,16 +268,51 @@ sources:
 
 ### AUTO 链（优先级页）
 
-AUTO 调度**只**由优先级页保存的规则（持久化到 `runtime_file` 的 `auto` 字段）决定，
-源配置里的 `models[].priority` 数字不再参与调度、也不再显示。
+AUTO 调度**只**由优先级页保存的规则决定，源配置里的 `models[].priority` 数字不再
+参与调度、也不再显示。**聊天与生图是两套独立的链**，在优先级页面通过「聊天 / 生图」
+按钮切换编辑：
+
+- **聊天链**（`runtime_file` 的 `auto` 字段）：服务 `POST /v1/chat/completions`
+  的 `model: AUTO` 请求；只接受 `kind: chat` 模型，配置在链里的生图槽位会被忽略。
+- **生图链**（`runtime_file` 的 `auto_image` 字段）：服务 `POST /v1/images/generations`
+  的 `model: AUTO` 请求；只接受 `kind: image` 模型。未配置生图链时，AUTO 生图回落
+  到「所有含生图模型的源、按注册顺序」的旧行为，兼容旧配置。
+
+每条规则 = 一个「槽位」：`{ model, source, tier, token_quota, period, hours }`。
 
-- 每条规则 = 一个「槽位」：`{ model, source, tier, token_quota, period, hours }`。
 - `tier` 表示优先级档位：同一档的模型并排、共享该优先级；档位从上到下递减。
 - 同一模型可配置多个槽位（如 A 源低配 → B 源低配 → A 源高配 → B 源高配），按档位顺延。
 - `token_quota` > 0 时该槽位在重置周期内用满即顺延到下一槽位；`period` 支持
   `hour` / `week` / `month` / `nhour`（配合 `hours`），空 = 不限。
-- 生图模型（`kind: image`）不参与 AUTO 链；生图走 `POST /v1/images/generations`
-  的独立路径。
+- 同 tier 内按「偏好分」排序，失败槽位冷却后自动跳过；冷却 / 配额耗尽 / hard
+  错误会顺延到下一 tier，全部失败时返回 503 并附上每档失败摘要。
+
+### 冷却与半冷却探测（自愈调度）
+
+槽位失败后进入指数退避冷却（5s → 10s → 20s …，上限 **5 分钟**）。冷却不是全时段
+硬闸：
+
+- **前半段完全静默**：刚失败的槽位不接任何流量，避免对故障上游持续施压。
+- **后半段放行 1 个探测**：窗口过半后，同一 `(源, 模型)` 最多允许 **一个** 请求作为
+  探测通过。探测槽数 = `clamp(max_concurrent / 10, 1, 2)`——`max_concurrent: 10`
+  的源恰好放行 1 个探测请求。
+- **探测排在最后**：同 tier 内探测候选永远排在健康候选之后，只有没有正常槽可用时
+  才承接真实流量，因此探测不会抢走可用容量。
+- **探测即真实请求**：成功则立即清零失败计数与冷却，槽位当场回到正常轮询——
+  上游恢复/额度恢复后不必等满整个冷却窗口。失败则重开一个新窗口，下次探测顺延到
+  新窗口的中点，不会连续重试。
+
+分类冷却策略：
+
+| 情况 | 冷却 | 说明 |
+|---|---|---|
+| 传输错误 / 5xx | 5s→10s→20s…，上限 5min | 指数退避 |
+| 401 / 403 凭据错误 | 10min | 换 key 后由下一次探测自动接回 |
+| 429 限流 | 30s 固定 | 「太快」不等于「坏了」，不进指数阶梯 |
+| 配额耗尽（`insufficient_quota` / `余额不足` 等） | 对齐配额窗口，上限 30min | **不进指数阶梯、不涨失败计数**，额度恢复即可用 |
+
+WebUI「优先级」页的槽位健康标签会区分 `冷却` / `待探测` / `探测中`，悬停可看窗口
+起点、探测放行时刻与完全恢复时刻。
 
 ### 源（Source）与适配器（Adapter）的关系
 
@@ -213,7 +331,7 @@ AUTO 调度**只**由优先级页保存的规则（持久化到 `runtime_file`
 2. `config.Load` 读取 `config.yaml`，`config.NewStore(runtime_file)` 读 WebUI 改动的
    运行时源，二者按名称合并成完整源列表。
 3. `rebuildRegistry` 为每个源创建 `provider.Provider`（持有目标适配器），并按源的
-   并发上限配置适配器 worker 池大小。
+   并发上限配置适配器 worker 池**上限**（不预分配，见下）。
 4. 请求进来时 `Registry` 按 `model` 路由到 Provider，Provider 调适配器
    `transform_request` → HTTP 发送 → `transform_response` / `transform_stream_chunk`。
 
@@ -225,6 +343,70 @@ WebUI 上的"新增/编辑源"、"上传 Lua 适配器"、"改 AUTO 优先级链
 `config.yaml` 中定义的基源无法改写配置文件，采用删除标记隐藏（重启后仍隐藏），
 在 UI 里重新添加同名源即可恢复。
 
+### 源模板（多 key 负载均衡）
+
+多个上游 key 共享同一套配置（URL / 适配器 / 模型列表 / 并发 / rpm）时，逐一复制
+整段 source 太冗余。**源模板**保存除 `name` 和 `api_key` 以外的全部源字段，
+从模板一键创建多个仅 key 不同的源，天然支持多 key 负载均衡（每个展开源独立调度、
+独立冷却、独立健康状态）。
+
+- WebUI「源」页右上角「**模板管理**」按钮：列出全部模板，可编辑 / 删除 / 新建。
+- 添加源弹窗右上角两个按钮：
+  - **从模板创建** → 选模板 → 表单自动填充（name + key 仍需你填）；
+  - **存为模板** → 把当前表单非 key/name 字段存为新模板。
+- 模板持久化在 `runtime_file` 的 `source_templates` 字段，与运行时源一样热管理，
+  改完即时生效。模板本身不含密钥，可以安全共享 / 纳管。
+
+模板不会自己变成源——它只是「配方」。真正承担流量的是从模板创建出来的、带 key
+的具体源。
+
+### 内存占用（实测与调优）
+
+内存不是一个定数：它随**配置的源数量**增长（每个源一个 `http.Transport` 与连接池、
+一组健康状态），与运行时长无关。本机实测（Linux x86_64，12 核）：
+
+| 部署形态 | 启动 RSS | 稳态 RSS |
+|---|---|---|
+| 1 源 / 1 适配器（最小配置） | ~4 MB | ~10 MB |
+| 1 源 + 29 MB 历史审计日志 | ~19 MB | ~20 MB |
+| **16 源 / 13 适配器 / 59 模型（本机生产）** | ~28 MB | **~32–35 MB** |
+
+> 历史参考：本项优化前同一生产配置为 **~105 MB**。降幅来自三处：审计日志的**原始记录**
+> 不再常驻内存（约 25 MB；统计聚合仍扫全量，但扫完即释放）、Lua 状态池不再单调增长、
+> 以及下面两个运行时开关。
+
+内存构成（生产实例分段测量，`/proc/<pid>/smaps`）：
+
+| 区域 | RSS | 说明 |
+|---|---|---|
+| Go 堆 | ~14 MB | provider/registry/scheduler 结构 + 连接池缓冲 |
+| 其他匹名（线程栈 / LuaJIT chunk / runtime） | ~12 MB | 与线程数、已加载适配器数相关 |
+| 二进制 text+rodata | ~8 MB | 映射的可执行文件页（只读、可被内核回收） |
+| Go runtime 预留 | ~4 MB | VSZ 上看到的 2 GB+ 是地址空间预留，不占物理内存 |
+| 共享库 | ~3 MB | libc / libluajit / libm |
+
+**两个推荐的部署开关**（只适用于环境变量，无需改代码）：
+
+```ini
+# /etc/systemd/system/llmsproxy.service
+Environment=GOGC=50
+Environment=MALLOC_ARENA_MAX=2
+```
+
+- `MALLOC_ARENA_MAX=2`：LuaJIT 的分配走 cgo → glibc malloc，glibc 默认允许 `8×nproc`
+  个 per-thread arena，每个碰到 malloc 的 OS 线程会占用一个（各约 1 MB，且**不归还给系统**）。
+  实测从 8–15 个 arena（约 7–12 MB）降到 **0**。
+- `GOGC=50`：把 Go 堆增长目标减半。**单独使用无效**（省下的堆会立即被更多 glibc
+  arena 吃掉，实测 20.3 → 21.5 MB 反而变大），必须与 `MALLOC_ARENA_MAX` 配合，
+  两者同时开启才降 **~19%**。网关是 I/O 密集型（本机 9 小时仅消耗 1min10s CPU），
+  多出的 GC 周期与它的空闲 CPU 相比可忽略。
+
+> 桌面版（Electron）已在 `cmd/gui/main.js` 里默认为内嵌核心注入这两个开关；手动 export 同名
+> 环境变量可覆盖。如果你用自己的 systemd unit / 容器，建议照上面加上。
+
+可选：`GOMEMLIMIT=48MiB` 作软上限，实测再省 ~1 MB；代价是逐近上限时 GC 转激进，
+温和场景不必开。
+
 ### disable_thinking
 
 请求体带 `"disable_thinking": true` 时，网关透传给各适配器；DeepSeek 适配器将其
@@ -238,7 +420,10 @@ WebUI 上的"新增/编辑源"、"上传 Lua 适配器"、"改 AUTO 优先级链
   支持按时间范围导出 CSV；点击模型可生成 pin 到该模型的连接配置。
 - **对话页**：流式/非流式调试。
 - **密钥页**：创建/编辑网关 key，为每个 key 配模型范围（模型 + 源 + token 配额 + 周期），
-  管理员管理全部 key，用户只看到自己的 key。
+  管理员管理全部 key，用户只看到自己的 key。每个模型砖块显示自己的配额徽标
+  （如 `1.0K 77×·1h`，未设配额显示 `∞`），点开可编辑该模型的 token 配额、
+  请求数配额与重置周期 —— 配额按模型独立生效，一个用满不影响同一 key 的其它模型。
+  「我的密钥」页对用户展示本 key 的模型范围与各自配额。
 - **优先级页**：拖拽积木配置 AUTO 链档位。
 - **源页**：在线增删改上游源（API key 等敏感字段加密落盘）。
 - **适配器页**：上传 / 删除 Lua 适配器脚本。
@@ -281,19 +466,12 @@ return {
 
 ### 内置适配器
 
-`openai` `deepseek` `anthropic` `gemini` `github` `groq` `mistral` `ollama` `kimicode`
-`opencode`。
+`openai` `deepseek` `anthropic` `gemini` `github` `groq` `mistral` `ollama` `kimicode`。
 
 `anthropic`/`gemini`/`ollama` 适配器内置多模态转换（`image_url` → 各自上游格式）；若
 源启用 `disable_thinking`，`deepseek` 适配器会把 `extra_body.thinking.type` 置为
 `disabled`。
 
-**opencode** 适配器面向 opencode.ai zen 免费池（`https://opencode.ai/zen/v1`）：
-zen 按 User-Agent 指纹识别官方客户端并分流——非官方 UA 的请求（curl、Go 默认 UA）
-会被丢进匿名池触发 `FreeUsageLimitError`。该适配器固定发送 opencode 客户端 UA，
-配合 `api_key: "public"`（官方无 key 客户端实际发送 `Bearer public`）即可走免费池，
-零成本接入 `deepseek-v4-flash-free` 等免费模型。
-
 **kimicode** 是展示 `build_headers` 的样例：云端校验调用方 app，需要按
 `meta.app_secret` 对时间戳+URL+请求体哈希做 HMAC 签名并附 `X-App-Sign` 等头。
 配好 `sources[].meta.{app_id, app_secret, app_agent}` 即可。
@@ -304,14 +482,24 @@ zen 按 User-Agent 指纹识别官方客户端并分流——非官方 UA 的请
 一键启停、系统托盘常驻、开机自启、静默启动，内容区一比一内嵌完整 WebUI
 （状态/对话/密钥/优先级/源/适配器 六个页面，免登录）。服务器用户继续使用纯 Go 二进制。
 
+### 两种安装包：Headless 与 Desktop
+
+发布到 Release 的安装包分为两类，面向不同用户：
+
+- **Headless（服务器版）**：纯核心二进制（无 Electron），配置驱动，适合服务器、
+  容器、systemd 等无人值守场景。发布包含 `deb` + `rpm`（含 systemd unit、内存调优
+  与示例配置）与 `tar.gz`（二进制 + `config.example.yaml` + `adapters/` + README），
+  Windows 为 `.zip`。
+- **Desktop（桌面版）**：Electron GUI 安装包，内嵌核心，适合个人桌面日常使用。
+
 ### 选择 GUI 还是纯后端
 
 | 场景 | 推荐 | 理由 |
 | ------ | ------ | ------ |
-| 服务器 / 内网网关 / 无人值守常驻 | **纯后端**（单二进制） | 极轻量（~10MB、~15MB RSS），零依赖单进程，直接跑在 systemd/任意容器里，远程管理 |
-| 个人桌面日常使用 / 多设备内网共享 | **GUI**（Electron） | 免登录内嵌 WebUI、系统托盘一键启停、开机自启、静默后台，适合不懂命令行的使用者 |
-| Windows 桌面 | **GUI** | 纯后端在 Windows 上需自行注册服务，GUI 提供原生托盘/自启体验 |
-| CI 一键出三平台安装包 | **GUI 打包脚本** | `make gui-*` 系列出 deb/AppImage/NSIS，可进发行流水线 |
+| 服务器 / 内网网关 / 无人值守常驻 | **Headless**（单二进制） | 极轻量（二进制 ~8MB；RSS 随源数量，单源 ~10MB、16 源 ~32–35MB，见[内存占用](#内存占用实测与调优)），零依赖单进程，直接跑在 systemd/任意容器里，远程管理 |
+| 个人桌面日常使用 / 多设备内网共享 | **Desktop**（GUI） | 免登录内嵌 WebUI、系统托盘一键启停、开机自启、静默后台，适合不懂命令行的使用者 |
+| Windows 桌面 | **Desktop** | 纯后端在 Windows 上需自行注册服务，GUI 提供原生托盘/自启体验 |
+| CI 一键出三平台安装包 | **Desktop 打包脚本** | `make gui-*` 系列出 deb/AppImage/NSIS，可进发行流水线 |
 
 两者完全同源：GUI 内嵌的就是纯后端的同一份 `llmsproxy` 核心（LuaJIT 版），
 配置/适配器格式完全一致，可随时互换。
@@ -323,27 +511,36 @@ cd cmd/gui
 npm install
 make gui          # 或 npm run dev —— 开发运行（需先 make build 生成 bin/llmsproxy）
 make gui-deb      # 本机构建 deb（分享给其它 Linux 用户）
-make gui-dist     # 构建 deb + AppImage（linux）
+make gui-dist     # 构建 deb + rpm + AppImage（linux，rpm 需系统 rpmbuild）
+make core-dist    # 构建核心 deb + rpm + tar.gz（服务器版，需 nfpm）
 make gui-win      # 构建 Windows nsis 安装包（本机 mingw + wine）
-make gui-win-docker  # 构建 Windows nsis 安装包（全 docker 自足，无需本机 mingw）
+make gui-win-docker  # 构建 Windows nsis 安装包（全 docker 自足，含 wine）
 make gui-dist-dir # 仅解包产物（调试用，不产安装包）
 ```
 
-产物在 `cmd/build/gui-dist/`：`ModelRouter-1.0.0.AppImage`、
-`modelrouter-gui_1.0.0_amd64.deb`、`ModelRouter Setup 1.0.0.exe`（win）。rpm 需系统 `rpmbuild`。
+产物在 `cmd/build/gui-dist/`：`ModelRouter-<ver>.AppImage`、
+`modelrouter-gui_<ver>_amd64.deb`、`ModelRouter Setup <ver>.exe`（win）；
+核心包在 `cmd/build/dist/`：`llmsproxy_<ver>_amd64.deb`、`llmsproxy-<ver>.x86_64.rpm`。
+rpm 需系统 `rpmbuild`。每次打包后都会跑 `packaging/verify-dist.sh` 产物体检
+（大小下限门禁——1.3.0 曾因宿主 wine 损坏产出 264KB 的残缺 exe 却没有被拦住）。
+发布时会把产物重命名为 `ModelRouter-Desktop-*`（GUI）与 `ModelRouter-Headless-*`（纯后端）
+两类上传到 GitCode Release。
 
 #### Windows 打包（docker 固化方案，推荐）
 
-`make gui-win-docker` 一步完成：交叉编译核心 + NSIS 打包，宿主机**无需安装 mingw**。
+`make gui-win-docker` 一步完成：交叉编译核心 + NSIS 打包**全在 docker 内**，宿主机
+**无需 mingw、无需 wine、无需 node**。镜像内置 wine（含 i386 运行时），修复了
+此前宿主 wine 损坏（缺 syswow64 → `c0000135`）导致产出 264KB 残缺 exe 的问题。
 
 - 构建镜像：`cmd/gui/docker/win-builder/Dockerfile`（`golang:1.25` + mingw-w64 +
-  预编译 LuaJIT windows 版 `lua51.dll` 与 import lib）。LuaJIT 源码取自 gitcode
-  镜像（国内可达），github 兜底；Go 依赖走 `https://goproxy.cn`。
-- 核心编译：`cmd/gui/scripts/win-core-docker.sh` —— 在容器内产出
-  `cmd/gui/bin/{llmsproxy.exe,lua51.dll}`。
-- 一键发布：`cmd/gui/scripts/dist-win-docker.sh` —— 核心编译后调用
-  electron-builder（宿主需有 `wine` 用于 NSIS）。
-- 首次运行自动构建 `modelrouter/win-builder` 镜像；已存在则直接复用。
+  node + wine32/wine64 + 预编译 LuaJIT windows 版 `lua51.dll` 与 import lib）。
+  LuaJIT 源码取自 gitcode 镜像（国内可达），github 兜底；Go 依赖走 `https://goproxy.cn`。
+- 一键发布：`cmd/gui/scripts/dist-win-docker.sh` —— 核心编译 + electron-builder
+  NSIS 全部在容器内执行（`USE_SYSTEM_WINE=true` 走镜像内的 wine）；electron-builder
+  的 ~400MB 工具链缓存挂到 docker volume，二次构建增量复用。
+- 首次运行自动构建 `modelrouter/win-builder` 镜像（约 5–10 分钟）；已存在则直接复用。
+- 打包完成后自动跑 `packaging/verify-dist.sh` 门禁：exe 低于 5MB / deb·rpm·7z 低于
+  10MB 即失败——残缺安装包不可能再溜进 Release。
 
 ```bash
 # 全 Windows 发布（核心重编 + NSIS）：
@@ -372,6 +569,15 @@ internal/gateway         # OpenAI 兼容 HTTP 服务 + 鉴权 + SDK/流式 + Web
 internal/types           # 统一格式 & OpenAI wire 类型
 ```
 
+## 开发协作
+
+- **Git 分支工作流**：主分支 / 特性分支 / 发布分支的规范见
+  **[docs/git-workflow.md](docs/git-workflow.md)**（[English](docs/git-workflow-en.md)）。
+  主线：`main` 永可部署 → 特性合入 `main` → 切 `release/vX.Y.Z` 打 tag 发布 →
+  hotfix 提交发布分支并 cherry-pick 回 `main`。
+- **Lua 适配器协议**：见 **[docs/lua-adapters.md](docs/lua-adapters.md)**
+  （[English](docs/lua-adapters-en.md)）。
+
 ## 测试
 
 ```bash

README_EN.md

@@ -11,7 +11,7 @@ KimiCode…) behind a single endpoint. Pick a specific model or use **AUTO** mod
 which routes to the best healthy upstream by configured priority.
 
 > **Lightweight, no recompile to add sources**: the gateway is a single Go
-> binary (~10 MB, zero runtime dependencies). Adding or switching an upstream is
+> binary (8-12 MB, ~8 MB stripped; zero runtime dependencies). Adding or switching an upstream is
 > just a `sources` entry in `config.yaml` (or via the WebUI) or a `.lua`
 > adapter — **no Go changes, no recompile**. Changes submitted through the
 > WebUI take effect immediately (hot reload); editing `config.yaml` or a `.lua`
@@ -38,10 +38,31 @@ Extracted and independently evolved from the multi-source LLM adapter layer of
   `reasoning_content`, `tool_calls`, `usage`).
 - **Image generation**: `POST /v1/images/generations`, routed to models with
   `kind: image`.
+- **Per-model quota**: each key gives every model its own token and request caps
+  plus a reset period (hour/week/month/custom N hours). One model running out
+  affects only that model — the key's other models keep working. Exhaustion
+  answers 429 + `Retry-After` naming the model, so a client resumes when the
+  window rolls over; admin keys are never capped.
 - **Multimodal**: `content` arrays (`image_url` etc.) pass through losslessly;
   Anthropic/Gemini/Ollama are translated automatically.
 - **LuaJIT VM**: golua-binding LuaJIT; each adapter has its own VM + worker
-  pool for safe concurrency.
+  pool for safe concurrency. Pools are **elastic**: the sum of `max_concurrent`
+  over an adapter's sources is a ceiling, not a preallocation — states are booted
+  on demand and reclaimed when demand drops, so an idle gateway holds close to
+  zero Lua states. The grow step follows the adapter's max concurrency
+  (`clamp(ceil(max/8), 1, 8)`, and never exceeds the number of queued callers);
+  the shrink step follows the live connection count
+  (`clamp(ceil(slack/(1+in_use)), 1, slack)`), so an idle pool collapses in one
+  round while a busy one gives up a single state at a time.
+- **On-demand request logs**: dashboard totals are computed from the **full**
+  audit history (streamed once at startup and released — 29 MB / 221k lines in
+  ~260 ms), while the raw records are never held in memory: the dashboard loads
+  one screen, scrolling pages the rest straight off disk, CSV export streams in
+  O(1) memory, and leaving the page releases everything.
+- **Self-healing cooldown**: cooldown is capped at 5 minutes and, past the
+  window's midpoint, exactly one probe request is allowed through; a recovered
+  upstream (or a reset quota) returns to full rotation on that probe instead of
+  waiting out the window.
 - **disable_thinking**: `disable_thinking: true` toggles reasoning per-request.
 - **Lua adapter protocol**: each source mounts a `.lua` adapter with
   `transform_request` / `transform_response` / `transform_stream_chunk` — all
@@ -77,7 +98,7 @@ GOMODCACHE=... GOPROXY=off go build -tags luajit -o llmsproxy ./cmd/llmsproxy
 > **No config is shipped in the repo** (config files carry real keys). On
 > first run the binary generates a default config at the `-config` path: a
 > random admin key (printed to the startup log), loopback-only `127.0.0.1:8080`,
-> and a keyless opencode zen source ready to chat. Rotate the admin key in the
+> and a keyless source ready to chat. Rotate the admin key in the
 > WebUI after first login.
 
 ```bash
@@ -155,11 +176,103 @@ under the `keys` field of the runtime file (encrypted at rest):
   delete the seed key.
 - The WebUI **Keys page** creates/deletes keys. Each key has a role (`admin`
   manages everything, `user` sees only its own key) and an optional **model
-  scope** (model + source + token quota + reset period).
+  scope** (model + source + token quota + reset period). Each model brick shows
+  its own budget badge (e.g. `1.0K 77×·1h`, `∞` when uncapped); clicking it
+  edits that model's token quota, request quota and reset period. Quotas apply
+  per model, so one model running out never blocks the key's others. The
+  "My key" view shows a user its model scopes and their budgets.
 - Clients authenticate with any authorized key's plaintext as
   `Authorization: Bearer <key>`.
 - Deleting a key removes it from the store immediately.
 
+#### Per-key usage quota
+
+**Quotas are per model.** Each key's `models[]` list gives every model its own
+token budget and request budget. One model running out affects only that model
+— the key's other models keep working.
+
+```yaml
+keys:
+  - key: sk-gw-<hex>
+    role: user
+    name: agent-alice
+    models:
+      - model: deepseek-v4-flash
+        token_quota: 1000000   # this key's token budget for this model
+        req_quota: 20000       # requests within the window
+        period: nhour          # "" | hour | week | month | nhour
+        hours: 6               # n-hour only
+      - model: AUTO
+        token_quota: 5000000   # AUTO is a quota entry like any other
+        period: hour
+      - model: kimi-k3        # no quota listed = unlimited
+```
+
+- **There is deliberately no key-wide total.** A key-wide cap would make
+  "move A's budget to B" a global reallocation; per-model budgets keep each
+  model independently controllable, so it stays visible which model is
+  actually consuming the spend.
+- `period`: empty = never resets (lifetime total); `hour` / `week` / `month` =
+  fixed windows; `nhour` + `hours` = a custom hour count. **A misspelled
+  period is rejected at write time** rather than silently becoming a
+  never-resetting quota.
+- Quotas are isolated per key *and* per model within a key: one key exhausting
+  `m1` never draws on another key's budget, and never blocks the same key's
+  `m2`.
+- Usage counts chat, streaming and image requests, and survives a restart by
+  replaying the audit log (40 days retained, covering the longest `month`
+  window).
+- An exhausted quota returns **429 + `Retry-After`**
+  (`rate_limit_exceeded`) and the message names the model that ran out, so a
+  client resumes when the window rolls over; a model the key may not use stays
+  403. **Admin keys are never capped** (quotas on their scopes are not
+  enforced either), so a cap can never lock the operator out.
+- Buckets are whole unix hours, so a window frees up at most an hour late
+  (deliberately freeing late rather than overspending).
+- `PUT /api/keys/{key}` submits quotas by submitting `models` — the caps are
+  part of the scope, so there is no second budget that can drift out of sync
+  with the model list. An explicit `0` lifts that model's cap.
+
+##### Quota rejection vs capacity rejection
+
+A quota is checked at the entry point and occupies no upstream slot, so it is a
+cheap rejection. Running out of capacity instead waits out the bounded
+`busyWait` before answering 503. Measured (capacity 8, 0.6 s/request, 100
+concurrent clients):
+
+| Case | Status | Latency | Reaches upstream |
+|---|---|---|---|
+| Quota spent (`req_quota` 50, 100 concurrent) | `429 rate_limit_exceeded` + `Retry-After` | **19–21 ms** | no |
+| Capacity exhausted (same topology, no quota) | `503 upstream_error` | ~2.6 s (`busyWait`) | overflow: no |
+
+So a client can tell "wait for the window to roll over" apart from "wait for
+upstream capacity" — the first only needs patience, the second usually means
+back off concurrency or switch sources.
+
+##### What the quota costs
+
+Buckets are kept per (key, model, whole unix hour) for 40 days. Measured on an
+AMD 7840HS:
+
+- A window query scans the window, not the whole retention: 49 ns for 1 h,
+  55 ns for 24 h, 3.9 µs for 30 d.
+- Recording one request: 283 ns, 3 allocations (unchanged from before this
+  feature; the allocations come from the record ring buffer).
+- Memory: the production shape (7 keys x 8 models x 2 sources at full 40-day
+  retention) costs about **3.7 MB**. The `source::model` bucket used by a
+  source-pinned quota is created **lazily** — it is maintained only once some
+  quota actually pins that source, because writing it on every record would
+  cost an extra 18 MB at 20 keys x 8 models x 3 sources.
+  The trade-off: usage recorded before a pinned quota was configured cannot be
+  split by source afterwards, so **a pinned quota may under-count its first
+  window**.
+
+```
+HTTP/1.1 429 Too Many Requests
+Retry-After: 2100
+{"error":{"type":"rate_limit_exceeded","message":"token quota exceeded for \"deepseek-v4-flash\" (5000000/5000000)"}}
+```
+
 ### Model routing
 
 `/v1/chat/completions` `model` resolution order:
@@ -176,12 +289,20 @@ tiered AUTO chain.
 
 ### AUTO chain (Priority page)
 
-AUTO scheduling is driven **only** by the rules saved on the Priority page
-(persisted as the `auto` field of the runtime file). The numeric
-`models[].priority` in source configs no longer participates in scheduling and
-is no longer shown.
+AUTO scheduling is driven **only** by the rules saved on the Priority page. The
+numeric `models[].priority` in source configs no longer participates in
+scheduling and is no longer shown. **Chat and image generation are two
+independent chains**, switched in the Priority page via the `Chat / Image` toggle:
+
+- **Chat chain** (`auto` field): serves `POST /v1/chat/completions` with
+  `model: AUTO`; only `kind: chat` models are honored (image slots are skipped).
+- **Image chain** (`auto_image` field): serves `POST /v1/images/generations` with
+  `model: AUTO`; only `kind: image` models are honored. With no image chain
+  configured, AUTO image generation falls back to legacy discovery (all sources
+  exposing an image model, in registry order) for backward compatibility.
+
+Each rule is one "slot": `{ model, source, tier, token_quota, period, hours }`.
 
-- Each rule is one "slot": `{ model, source, tier, token_quota, period, hours }`.
 - `tier` is the priority tier: models in the same tier sit side by side and
   share it; tiers run high → low.
 - The same model may appear in several slots (e.g. A low → B low → A high →
@@ -189,8 +310,9 @@ is no longer shown.
 - `token_quota` > 0 means the slot is skipped once its tokens within the reset
   window are exhausted; `period` supports `hour` / `week` / `month` / `nhour`
   (with `hours`); empty = unlimited.
-- Image models (`kind: image`) are kept out of the chain and are served by the
-  separate `POST /v1/images/generations` path.
+- Within a tier, slots run by preference score; cooling / quota-exhausted /
+  hard-failed slots fall through to the next tier, and total failure returns
+  503 with a per-tier summary.
 
 ### Sources vs. adapters
 
@@ -227,6 +349,82 @@ sources and `.lua` files; base sources defined in `config.yaml` cannot rewrite
 that file, so they are hidden via a deletion tombstone (still hidden after
 restart) and can be restored by re-adding the same name in the UI.
 
+### Source templates (multi-key balancing)
+
+When several upstream keys share one config (URL / adapter / model list /
+concurrency / rpm), copying the whole source block N times is wasteful. A
+**source template** stores every source field except `name` and `api_key`, so you
+create N key-only-different sources from one shared skeleton — natural multi-key
+load balancing (each expanded source schedules, cools down, and reports health
+independently).
+
+- WebUI Sources page, top-right **Templates** button: list all templates with
+  edit / delete / create.
+- Add-source dialog has two header buttons:
+  - **From template** → pick a template → form auto-fills (name + key still
+    yours to enter);
+  - **As template** → save the current form's non-key/name fields as a template.
+- Templates persist in the runtime file's `source_templates` field alongside
+  runtime sources (hot-managed, apply immediately). Templates hold no secret
+  keys, so they're safe to share / version-control.
+
+A template is just a recipe — it doesn't become a source by itself. Only the
+key-bearing sources created from it carry real traffic.
+
+### Memory footprint (measured and tuned)
+
+Memory is not a constant: it scales with the **number of configured sources**
+(each source owns an `http.Transport` with its connection pool plus per-model
+health state), not with uptime. Measured on this machine (Linux x86_64, 12 cores):
+
+| Deployment shape | Startup RSS | Settled RSS |
+|---|---|---|
+| 1 source / 1 adapter (minimal) | ~4 MB | ~10 MB |
+| 1 source + a 29 MB audit history | ~19 MB | ~20 MB |
+| **16 sources / 13 adapters / 59 models (this host)** | ~28 MB | **~32-35 MB** |
+
+> For reference, the same production config used **~105 MB** before this round of
+> work. The reduction comes from three places: raw audit records are no longer kept
+> resident (~25 MB — the aggregates still scan the full history, but the scan
+> releases as it goes), Lua state pools no longer grow monotonically, and the two
+> runtime knobs below.
+
+Breakdown of the production instance (per-region, from `/proc/<pid>/smaps`):
+
+| Region | RSS | Notes |
+|---|---|---|
+| Go heap | ~14 MB | provider/registry/scheduler state + connection-pool buffers |
+| Other anonymous (thread stacks / LuaJIT chunks / runtime) | ~12 MB | scales with thread count and loaded adapters |
+| Binary text+rodata | ~8 MB | mapped executable pages (read-only, reclaimable by the kernel) |
+| Go runtime reservations | ~4 MB | the 2 GB+ you see in VSZ is address space, not physical memory |
+| Shared libraries | ~3 MB | libc / libluajit / libm |
+
+**Two recommended deployment knobs** (environment only, no code change):
+
+```ini
+# /etc/systemd/system/llmsproxy.service
+Environment=GOGC=50
+Environment=MALLOC_ARENA_MAX=2
+```
+
+- `MALLOC_ARENA_MAX=2`: LuaJIT allocates through cgo into glibc malloc, and glibc
+  allows up to `8 x nproc` per-thread arenas. Every OS thread that touches malloc
+  claims one (~1 MB each) and **never returns it to the OS**. Measured: 8-15
+  arenas (7-12 MB) down to **0**.
+- `GOGC=50`: halves the Go heap growth target. **It does nothing on its own** —
+  the heap it saves is immediately consumed by additional glibc arenas (measured
+  20.3 -> 21.5 MB, i.e. slightly worse) — so it must be paired with
+  `MALLOC_ARENA_MAX`. Together they cut settled RSS by **~19%**. The gateway is
+  I/O bound (1min10s of CPU per 9 hours here), so the extra GC cycles are free.
+
+> The desktop build (Electron) already injects both when spawning the embedded
+> core (`cmd/gui/main.js`); exporting either variable yourself overrides it. If you
+> write your own systemd unit or container spec, add them there.
+
+Optional: `GOMEMLIMIT=48MiB` as a soft ceiling saves roughly 1 MB more, at the
+cost of GC turning aggressive as the limit approaches. Not worth it for a relaxed
+deployment.
+
 ### disable_thinking
 
 With `"disable_thinking": true` in the request body, the gateway passes it to
@@ -289,19 +487,12 @@ Shared helpers: `hmac_sha256_hex(key, data)`, `sha256_hex(data)`,
 ### Built-in adapters
 
 `openai` `deepseek` `anthropic` `gemini` `github` `groq` `mistral` `ollama`
-`kimicode` `opencode`.
+`kimicode`.
 
 `anthropic`/`gemini`/`ollama` include multimodal conversion
 (`image_url` → their native format); with `disable_thinking` the `deepseek`
 adapter sets `extra_body.thinking.type` to `disabled`.
 
-**opencode** targets the opencode.ai zen free pool
-(`https://opencode.ai/zen/v1`): zen fingerprints clients by User-Agent and
-routes non-official UAs (curl, Go's default) into an anonymous pool that
-hits `FreeUsageLimitError`. The adapter always sends the opencode client UA;
-combined with `api_key: "public"` (the keyless official client actually sends
-`Bearer public`) it gets the free pool, e.g. `deepseek-v4-flash-free`.
-
 **kimicode** demonstrates `build_headers`: the cloud validates the calling
 app, so you HMAC-sign timestamp+URL+body with `meta.app_secret` and add
 `X-App-Sign`-style headers. Configure `sources[].meta.{app_id, app_secret,
@@ -315,14 +506,25 @@ tray, autostart, silent launch, and a pixel-for-pixel embedded full WebUI
 (Status/Chat/Keys/Priority/Sources/Adapters — no login needed). Server users
 keep using the plain Go binary.
 
+### Two installer kinds: Headless and Desktop
+
+Release installers ship in two flavors for different audiences:
+
+- **Headless (server)**: the plain core binary (no Electron), config-driven,
+  for servers / containers / systemd / unattended runs. Ships as `deb` + `rpm`
+  (systemd unit with memory tuning + example config) and `tar.gz` (binary +
+  `config.example.yaml` + `adapters/` + README), Windows as `.zip`.
+- **Desktop (desktop app)**: the Electron GUI installer with an embedded core,
+  for personal daily desktop use.
+
 ### GUI vs. plain backend
 
 | Scenario | Pick | Why |
 | ---- | ---- | ---- |
-| Server / intranet gateway / unattended long-running | **plain backend** (single binary) | ~10 MB, ~15 MB RSS, zero-dep single process — drop it into systemd or any container, remote admin |
-| Personal desktop daily use / multi-device intranet sharing | **GUI** (Electron) | no-login embedded WebUI, tray one-click, autostart, silent background — for non-CLI users |
-| Windows desktop | **GUI** | plain backend needs manual service registration; GUI ships native tray/autostart |
-| CI one-shot 3-platform installers | **GUI packaging scripts** | `make gui-*` emits deb / AppImage / NSIS, drops straight into a release pipeline |
+| Server / intranet gateway / unattended long-running | **Headless** (single binary) | ~8 MB binary; RSS scales with the number of sources (~10 MB for one, ~32-35 MB for 16 — see [Memory footprint](#memory-footprint-measured-and-tuned)), zero-dep single process — drop it into systemd or any container, remote admin |
+| Personal desktop daily use / multi-device intranet sharing | **Desktop** (GUI) | no-login embedded WebUI, tray one-click, autostart, silent background — for non-CLI users |
+| Windows desktop | **Desktop** | plain backend needs manual service registration; GUI ships native tray/autostart |
+| CI one-shot 3-platform installers | **Desktop packaging scripts** | `make gui-*` emits deb / AppImage / NSIS, drops straight into a release pipeline |
 
 Both share the exact same `llmsproxy` core (LuaJIT build) — configs and
 adapters are fully compatible, freely interchangeable.
@@ -334,30 +536,42 @@ cd cmd/gui
 npm install
 make gui            # or npm run dev — dev run (needs make build first for bin/llmsproxy)
 make gui-deb        # local deb (share with other Linux users)
-make gui-dist       # deb + AppImage (linux)
+make gui-dist       # deb + rpm + AppImage (linux; rpm needs system rpmbuild)
+make core-dist      # core deb + rpm + tar.gz (server edition; needs nfpm)
 make gui-win        # Windows NSIS (host mingw + wine)
-make gui-win-docker # Windows NSIS, fully dockerized (no host mingw needed)
+make gui-win-docker # Windows NSIS, fully dockerized (wine in image; no host mingw/node/wine)
 ```
 
-Artifacts land in `cmd/build/gui-dist/`: `ModelRouter-1.0.0.AppImage`,
-`modelrouter-gui_1.0.0_amd64.deb`, `ModelRouter Setup 1.0.0.exe` (win). rpm
-needs system `rpmbuild`.
+Artifacts land in `cmd/build/gui-dist/`: `ModelRouter-<ver>.AppImage`,
+`modelrouter-gui_<ver>_amd64.deb`, `ModelRouter Setup <ver>.exe` (win); core
+packages land in `cmd/build/dist/`: `llmsproxy_<ver>_amd64.deb`,
+`llmsproxy-<ver>.x86_64.rpm`. Both target families run
+`packaging/verify-dist.sh` afterwards — a size floor that rejects degenerate
+installers (1.3.0 shipped a 264 KB NSIS stub because the host wine was broken
+and nothing caught it). At release time artifacts are renamed into the
+`ModelRouter-Desktop-*` (GUI) and `ModelRouter-Headless-*` (plain backend)
+families before upload to GitCode Releases.
 
 #### Windows packaging (dockerized, recommended)
 
-`make gui-win-docker` does cross-compile + NSIS in one step; the host needs
-**no mingw**.
+`make gui-win-docker` does cross-compile + NSIS **entirely in docker**; the host
+needs **no mingw, no wine, no node**. Wine (with the i386 runtime) lives inside
+ the image, fixing the broken-host-wine failure that produced the 264 KB stub
+installer.
 
 - Image: `cmd/gui/docker/win-builder/Dockerfile` (`golang:1.25` + mingw-w64 +
-  prebuilt LuaJIT for Windows: `lua51.dll` + import lib). LuaJIT source comes
-  from the gitcode mirror (reachable in CN) with github fallback; Go modules go
-  through `https://goproxy.cn`.
-- Core build: `cmd/gui/scripts/win-core-docker.sh` — emits
-  `cmd/gui/bin/{llmsproxy.exe,lua51.dll}` inside the container.
-- One-shot release: `cmd/gui/scripts/dist-win-docker.sh` — after the core,
-  runs electron-builder (host needs `wine` for NSIS).
-- The `modelrouter/win-builder` image is auto-built on first run and reused
-  afterwards.
+  node + wine32/wine64 + prebuilt LuaJIT for Windows: `lua51.dll` + import lib).
+  LuaJIT source comes from the gitcode mirror (reachable in CN) with github
+  fallback; Go modules go through `https://goproxy.cn`.
+- One-shot release: `cmd/gui/scripts/dist-win-docker.sh` — core build + NSIS
+  packaging run inside the container (`USE_SYSTEM_WINE=true` uses the image's
+  wine); electron-builder's ~400 MB toolchain cache lives in a docker volume so
+  later builds reuse it.
+- The `modelrouter/win-builder` image is auto-built on first run (~5-10 min)
+  and reused afterwards.
+- After packaging, `packaging/verify-dist.sh` runs as a gate: an exe under
+  5 MB (or a deb/rpm/7z under 10 MB) fails the build — a broken installer can
+  never reach a Release again.
 
 ```bash
 # Full Windows release (rebuild core + NSIS):
@@ -390,6 +604,17 @@ internal/gateway         # OpenAI-compatible HTTP + auth + SDK/streaming + WebUI
 internal/types           # unified format & OpenAI wire types
 ```
 
+## Development & collaboration
+
+- **Git branching workflow**: main / feature / release-branch conventions in
+  **[docs/git-workflow-en.md](docs/git-workflow-en.md)**
+  ([中文](docs/git-workflow.md)). The line: `main` is always deployable →
+  features merge into `main` → cut `release/vX.Y.Z` and tag → release → hotfixes
+  land on the release branch and are cherry-picked back into `main`.
+- **Lua adapter protocol**: see
+  **[docs/lua-adapters-en.md](docs/lua-adapters-en.md)**
+  ([中文](docs/lua-adapters.md)).
+
 ## Tests
 
 ```bash

cmd/gui/docker/win-builder/Dockerfile

@@ -1,11 +1,20 @@
 # ModelRouter GUI Windows builder
-# Self-sufficient cross-compilation environment for the embedded llmsproxy core.
-# Produces llmsproxy.exe (Go, windows/amd64, luajit tag) + lua51.dll (LuaJIT mingw build)
-# inside a mounted workspace: cmd/gui/bin/{llmsproxy.exe,lua51.dll}
+# Self-sufficient cross-compilation environment for the embedded llmsproxy core
+# AND for packaging the Windows NSIS installer.
+#
+# Produces inside a mounted workspace:
+#   - llmsproxy.exe (Go, windows/amd64, luajit tag) + lua51.dll
+#   - ModelRouter Setup <ver>.exe  (electron-builder --win nsis; requires wine
+#     so the uninstaller is generated — set USE_SYSTEM_WINE=true)
 #
 # Build:   docker build -t modelrouter/win-builder cmd/gui/docker/win-builder
 # Usage:   make gui-win-docker   (or cmd/gui/scripts/dist-win-docker.sh)
 #
+# This image is the SINGLE wine provider for the NSIS step, so a host that has
+# no / broken wine (e.g. amd64-only Debian: no syswow64 -> c0000135) can still
+# produce a real installer. The NSIS step MUST run inside this image; running it
+# on the host was how the 1.3.0 264 KB degenerate exe got shipped.
+#
 # LuaJIT source is fetched from gitcode.com (CN-reachable mirror) with github
 # fallback; pass --build-arg LUAJIT_REPO to override.
 
@@ -13,11 +22,26 @@ FROM golang:1.25
 
 ARG LUAJIT_REPO=https://gitcode.com/openresty/luajit2.git
 
-RUN apt-get update -qq \
+# i386 arch is required for wine32 (the NSIS uninstaller step runs a win32 exe
+# through wine's wow64 path, which needs the 32-bit runtime under /usr/lib/i386).
+RUN dpkg --add-architecture i386 \
+ && apt-get update -qq \
  && DEBIAN_FRONTEND=noninteractive apt-get install -y -qq \
       gcc-mingw-w64-x86-64 make ca-certificates git \
+      # NSIS step needs node + the electron-builder toolchain
+      nodejs npm \
+      # wine (32+64) for generating the NSIS uninstaller
+      wine64 wine32:i386 \
  && rm -rf /var/lib/apt/lists/*
 
+# electron-builder resolves the uninstaller via the system wine when
+# USE_SYSTEM_WINE=true (or by default on Linux without a toolset). Pin it.
+# The toolchain cache lives in $HOME/.cache (mounted as a volume by
+# dist-win-docker.sh), so it is reused across builds and never pollutes the
+# mounted workspace.
+ENV USE_SYSTEM_WINE=true
+ENV WINEDEBUG=-all
+
 # LuaJIT cross-compiled for Windows; keeps lua51.dll and the cgo import lib
 # (libluajit-5.1.dll.a -> libluajit-5.1.a in the mingw lib dir).
 # NOTE: $LUAJIT_REPO deliberately unquoted in "sh -c" — quoting the ARG would

cmd/gui/main.js

@@ -10,6 +10,7 @@ const {
   Tray,
   shell,
   nativeImage,
+  nativeTheme,
 } = require("electron");
 const path = require("path");
 const fs = require("fs");
@@ -46,6 +47,15 @@ const CORE_EXE = path.join(
   "bin",
   process.platform === "win32" ? "llmsproxy.exe" : "llmsproxy",
 );
+// Packaged window icon: must be a REAL file on disk (extraResource), not an
+// asar path — Electron/GTK pulls BrowserWindow icons from the filesystem and
+// crashes (SIGSEGV) when given an asar virtual path on Linux.
+// Dev:  cmd/gui/build/icon.png.
+const WINDOW_ICON = path.join(
+  app.isPackaged ? process.resourcesPath : __dirname,
+  "build",
+  "icon.png",
+);
 // default port for the embedded core (configurable via settings)
 const DEFAULT_PORT = 8787;
 
@@ -53,6 +63,7 @@ let mainWindow = null;
 let tray = null;
 let coreProc = null;
 let coreStartedAt = 0;
+let coreReady = false; // true only after the embedded core actually answers HTTP
 let settings = {}; // { port, autoStart, minimizeToTray }
 let authRules = [];
 
@@ -120,9 +131,9 @@ function coreStarted() {
 
 function writeDefaultEmbeddedConfig() {
   // Generate a self-contained embedded profile with a fresh random admin key.
-  // (Like Clash Verge generates its profile; the core's EnsureDefault default
-  // is 127.0.0.1:8080 — we want our own port + key on first run.) If a config
-  // already exists we do NOT overwrite it (keep user's edits).
+  // No upstream sources are preconfigured — the desktop user adds them via
+  // the WebUI's sources page (like Clash Verge's empty-by-default profile).
+  // If a config already exists we do NOT overwrite it (keep user's edits).
   if (fs.existsSync(CONFIG_FILE)) return;
   const key = "sk-gw-" + crypto.randomBytes(16).toString("hex");
   const yaml = [
@@ -132,16 +143,6 @@ function writeDefaultEmbeddedConfig() {
     "adapter_dir: " + JSON.stringify(path.join(PROFILE_DIR, "adapters")),
     "runtime_file: " + JSON.stringify(path.join(PROFILE_DIR, "runtime.json")),
     "",
-    "sources:",
-    "  - name: zen",
-    "    base_url: https://opencode.ai/zen/v1",
-    "    api_key: public",
-    "    adapter: opencode",
-    "    models:",
-    "      - id: deepseek-v4-flash-free",
-    "        priority: 100",
-    "        kind: chat",
-    "",
     // Admin key as a plain `keys` entry (no seed flag) so the web UI never
     // shows the replace-the-initial-key warning. gateway_keys entries are
     // marked seed:true by the core.
@@ -197,6 +198,17 @@ function startCore() {
     cwd: PROFILE_DIR,
     env: Object.assign({}, process.env, {
       LLMS_PROXY_GUI: "1",
+      // Memory tuning, measured on a 16-source deployment (see README 内存占用):
+      //   MALLOC_ARENA_MAX caps glibc per-thread malloc arenas. LuaJIT allocates
+      //     through cgo -> glibc malloc, and glibc defaults to 8*nproc arenas, so
+      //     every OS thread touching malloc reserves its own ~1 MB arena that is
+      //     never returned to the OS.
+      //   GOGC halves the Go heap growth target. On its own it does not help (the
+      //     saved heap is immediately eaten by extra glibc arenas); only the pair
+      //     lowers settled RSS. The core is I/O bound, so the extra GC is free.
+      // Both are overridable: a user who exports them keeps their own values.
+      GOGC: process.env.GOGC || "50",
+      MALLOC_ARENA_MAX: process.env.MALLOC_ARENA_MAX || "2",
     }),
     stdio: ["ignore", "pipe", "pipe"],
   });
@@ -211,6 +223,7 @@ function startCore() {
   coreProc.on("exit", (code) => {
     const wasManaged = coreStartedAt > 0;
     coreStartedAt = 0;
+    coreReady = false;
     coreProc = null;
     console.log("core exited:", code);
     // auto-restart if still running GUI (crash recovery, max every ~5s)
@@ -223,6 +236,12 @@ function startCore() {
       }, 1500);
     }
   });
+  // Broadcast starting (NOT ready) immediately. The renderer must not
+  // attempt to load the iframe until the core actually answers HTTP —
+  // otherwise the frame load races the listener and fails permanently.
+  // (Fix for blank-screen: running=true was previously broadcast while the
+  // core was still booting, so the iframe ERR_CONNECTION_REFUSED and the
+  // renderer never retried.)
   notifyCoreState();
   // pump until the core answers, then inject the admin key and tell the
   // renderer the web UI is ready (the iframe must load only after auth is set)
@@ -230,6 +249,7 @@ function startCore() {
     .then((ok) => {
       if (ok) {
         setEmbeddedAuth();
+        coreReady = true;
         notifyCoreState();
       }
     })
@@ -254,7 +274,7 @@ function waitForCoreReady(timeoutMs = 10000) {
 let coreLastExit = 0;
 
 function stopCore() {
-  app.isQuitting = true;
+  coreReady = false;
   if (coreProc && coreProc.exitCode === null) {
     try {
       coreProc.kill();
@@ -357,17 +377,27 @@ function createTray() {
     );
   }
   tray = new Tray(img);
+  // Rebuild the context menu when the OS theme flips (dark/light) so menu
+  // text colors follow the desktop theme; nativeTheme is the Electron API
+  // that tracks GTK/appindicator dark mode on Linux.
+  const themeRebuild = () => {
+    try {
+      if (global.__rebuildTray) global.__rebuildTray();
+    } catch (e) {}
+  };
+  try {
+    nativeTheme.on("updated", themeRebuild);
+  } catch (e) {}
+  // Status rows used to be enabled:false, which GTK renders in a fixed light
+  // grey that never flips with the system dark/light theme — invisible on
+  // light desktops. They are now enabled (normal theme-aware fg) with a
+  // no-op click so they stay display-only but inherit the menu text color.
+  const statusRow = (text) => ({ label: text, enabled: true, click: () => {} });
   const rebuild = () => {
     const running = coreStarted();
     const menu = Menu.buildFromTemplate([
-      {
-        label: "ModelRouter Desktop",
-        enabled: false,
-      },
-      {
-        label: running ? "● 内核运行中" : "○ 内核已停止",
-        enabled: false,
-      },
+      statusRow("ModelRouter Desktop"),
+      statusRow(running ? "● 内核运行中" : "○ 内核已停止"),
       { type: "separator" },
       {
         label: "显示主窗口",
@@ -438,7 +468,8 @@ function createTray() {
 function notifyCoreState() {
   if (mainWindow && !mainWindow.isDestroyed()) {
     mainWindow.webContents.send("core:state", {
-      running: coreStarted(),
+      running: coreStarted() && coreReady,
+      ready: coreReady,
       startedAt: coreStartedAt,
     });
   }
@@ -496,6 +527,12 @@ function createWindow() {
   const menu = Menu.buildFromTemplate([]);
   Menu.setApplicationMenu(menu);
   mainWindow = new BrowserWindow({
+    // Window icon (taskbar/dock). Packaged: real file under resourcesPath (see
+    // WINDOW_ICON) — an asar path here makes GTK segfault on Linux.
+    icon: WINDOW_ICON,
+    // GNOME taskbar icons match the window's WM_CLASS (lowercase
+    // "modelrouter-gui") against StartupWMClass in the .desktop file.
+    backgroundColor: "#1f2937",
     width: 1280,
     height: 860,
     minWidth: 900,
@@ -560,7 +597,8 @@ ipcMain.handle(
 );
 
 ipcMain.handle("core:state", () => ({
-  running: coreStarted(),
+  running: coreStarted() && coreReady,
+  ready: coreReady,
   startedAt: coreStartedAt,
   baseUrl: embeddedBaseUrl(),
   port: settings.port || DEFAULT_PORT,

cmd/gui/package-lock.json

@@ -1,12 +1,12 @@
 {
   "name": "modelrouter-gui",
-  "version": "1.0.0",
+  "version": "1.7.3",
   "lockfileVersion": 3,
   "requires": true,
   "packages": {
     "": {
       "name": "modelrouter-gui",
-      "version": "1.0.0",
+      "version": "1.7.3",
       "license": "Proprietary",
       "devDependencies": {
         "electron": "^33.0.0",
@@ -582,7 +582,7 @@
       }
     },
     "node_modules/@peculiar/webcrypto": {
-      "version": "1.7.1",
+      "version": "1.7.3",
       "resolved": "https://registry.npmjs.org/@peculiar/webcrypto/-/webcrypto-1.7.1.tgz",
       "integrity": "sha512-ODOov0sGMJMf3jPonOkgGqPknTsu+DdQ7kD++gz8aI+aFMOMHFbWAA2taqXXVTdP+OTOQR/znGvSpmkeI0WTYQ==",
       "dev": true,
@@ -3335,7 +3335,7 @@
       }
     },
     "node_modules/resedit": {
-      "version": "1.7.2",
+      "version": "1.7.3",
       "resolved": "https://registry.npmjs.org/resedit/-/resedit-1.7.2.tgz",
       "integrity": "sha512-vHjcY2MlAITJhC0eRD/Vv8Vlgmu9Sd3LX9zZvtGzU5ZImdTN3+d6e/4mnTyV8vEbyf1sgNIrWxhWlrys52OkEA==",
       "dev": true,
@@ -4045,4 +4045,4 @@
       }
     }
   }
-}
+}

cmd/gui/package.json

@@ -1,6 +1,6 @@
 {
   "name": "modelrouter-gui",
-  "version": "1.0.0",
+  "version": "1.7.3",
   "description": "ModelRouter Desktop — embedded ModelRouter core with tray",
   "author": "ModelRouter",
   "main": "main.js",
@@ -10,7 +10,7 @@
     "dev": "electron . --no-sandbox --dev",
     "core:linux": "node scripts/prepare-core.js linux",
     "core:win": "node scripts/prepare-core.js win",
-    "dist:linux": "npm run core:linux && electron-builder --linux deb AppImage",
+    "dist:linux": "npm run core:linux && electron-builder --linux deb rpm AppImage",
     "dist:debian": "npm run core:linux && electron-builder --linux deb",
     "dist:win": "npm run core:win && electron-builder --win nsis",
     "dist:dir": "npm run core:linux && electron-builder --dir"
@@ -32,10 +32,19 @@
       "icon": "build/icon.png",
       "maintainer": "ModelRouter",
       "synopsis": "ModelRouter Desktop",
+      "desktop": {
+        "entry": {
+          "StartupWMClass": "modelrouter-gui"
+        }
+      },
       "extraResources": [
         {
           "from": "bin/llmsproxy",
           "to": "bin/llmsproxy"
+        },
+        {
+          "from": "build/icon.png",
+          "to": "build/icon.png"
         }
       ]
     },
@@ -44,7 +53,13 @@
     },
     "mac": {
       "target": "dmg",
-      "icon": "build/icon.png"
+      "icon": "build/icon.png",
+      "extraResources": [
+        {
+          "from": "build/icon.png",
+          "to": "build/icon.png"
+        }
+      ]
     },
     "win": {
       "target": "nsis",
@@ -57,6 +72,10 @@
         {
           "from": "bin/lua51.dll",
           "to": "bin/lua51.dll"
+        },
+        {
+          "from": "build/icon.png",
+          "to": "build/icon.png"
         }
       ]
     },

cmd/gui/renderer/app.js

@@ -5,6 +5,7 @@
 const state = {
   core: {
     running: false,
+    ready: false,
     baseUrl: "",
     port: 8787,
     profileDir: "",
@@ -18,21 +19,13 @@ const state = {
     minimizeToTray: true,
   },
   theme: localStorage.getItem("mr-theme") || "light",
-  loadedOnce: false,
+  loadedOnce: false, // iframe has been pointed at the web UI at least once
+  frameLoaded: false, // last loadFrame actually finished loading (did-finish-load)
+  loadTries: 0,
 };
 
 const $ = (s) => document.querySelector(s);
 
-function esc(s) {
-  return String(s == null ? "" : s).replace(
-    /[&<>"']/g,
-    (c) =>
-      ({ "&": "&amp;", "<": "&lt;", ">": "&gt;", '"': "&quot;", "'": "&#39;" })[
-        c
-      ],
-  );
-}
-
 function toast(msg, isErr, ms) {
   const t = $("#toast");
   t.textContent = msg;
@@ -55,7 +48,11 @@ function renderCoreStatus() {
   const core = state.core;
   const dot = $("#conn-dot");
   dot.className = "status-dot" + (core.running ? " ok" : " bad");
-  dot.title = core.running ? "内核运行中" : "内核未运行";
+  dot.title = !core.running
+    ? "内核未运行"
+    : core.ready
+      ? "内核运行中"
+      : "内核启动中…";
 }
 
 function showOffline() {
@@ -70,11 +67,47 @@ function showLoading() {
   $("#console-loading").style.display = "flex";
 }
 
+// probe the core's web root directly (fetch from the renderer). Using
+// fetch() instead of relying on the iframe's own load means the core's
+// readiness is confirmed by an actual HTTP round-trip before we point the
+// iframe at it — this is what prevents the blank-screen race where the
+// frame load happens while the listener is still coming up.
+function coreReachable() {
+  return fetch(
+    state.core.baseUrl || "http://127.0.0.1:" + (state.core.port || 8787),
+    { method: "GET", redirect: "follow", cache: "no-store" },
+  )
+    .then(() => true)
+    .catch(() => false);
+}
+
+// point the iframe at the WebUI, but only after the core is reachable.
+// If the load fails (core still booting / raced), we retry with backoff
+// instead of giving up forever (the old code set loadedOnce blindly and
+// never retried -> permanent blank screen).
+let frameRetryTimer = null;
+function scheduleFrameLoad() {
+  if (frameRetryTimer) return; // already scheduled
+  frameRetryTimer = setTimeout(async () => {
+    frameRetryTimer = null;
+    if (!state.core.running) return; // core stopped meanwhile
+    if (!(await coreReachable())) {
+      // core not reachable yet: wait a bit and try again
+      state.loadTries += 1;
+      showLoading();
+      scheduleFrameLoad();
+      return;
+    }
+    loadFrame(state.core.baseUrl || "http://127.0.0.1:" + state.core.port);
+  }, state.loadTries === 0 ? 800 : Math.min(3000, 400 * state.loadTries));
+}
+
 function loadFrame(url) {
   const f = frame();
   f.style.display = "block";
   $("#offline").style.display = "none";
   $("#console-loading").style.display = "none";
+  if (state.loadedOnce && f.src === url) return; // already pointing there
   f.src = url;
   state.loadedOnce = true;
 }
@@ -82,8 +115,9 @@ function loadFrame(url) {
 async function ensureCore() {
   await refreshCore();
   if (state.core.running) {
-    loadFrame(state.core.baseUrl);
+    scheduleFrameLoad();
   } else {
+    state.loadedOnce = false;
     showOffline();
   }
 }
@@ -141,10 +175,21 @@ async function saveSettings() {
   toast("设置已保存");
   await refreshCore();
   if (state.core.running) {
-    loadFrame(state.core.baseUrl);
+    scheduleFrameLoad();
+  } else {
+    state.loadedOnce = false;
+    showOffline();
   }
 }
 
+// ===== theme =====
+function applyTheme() {
+  document.documentElement.dataset.theme = state.theme;
+  try {
+    localStorage.setItem("mr-theme", state.theme);
+  } catch (e) {}
+}
+
 // ===== init =====
 function init() {
   $("#tb-min").onclick = () => window.modelrouter.win.minimize();
@@ -171,15 +216,30 @@ function init() {
     showLoading();
     await window.modelrouter.core.start();
     await refreshCore();
-    if (state.core.running) loadFrame(state.core.baseUrl);
+    if (state.core.running) scheduleFrameLoad();
     else showOffline();
   };
-  frame().onload = () => refreshCore();
+  // when the iframe finishes loading, mark loaded and refresh core status so
+  // the conn-dot reflects the actual WebUI (not just the core reachability)
+  frame().onload = () => {
+    state.frameLoaded = true;
+    refreshCore();
+  };
+  // load failure (e.g. the frame was pointed at the core just before it
+  // finished booting) -> clear loadedOnce and let scheduleFrameLoad retry
+  frame().onerror = () => {
+    state.loadedOnce = false;
+    state.frameLoaded = false;
+    scheduleFrameLoad();
+  };
   window.modelrouter.core.onState((d) => {
     state.core = Object.assign({}, state.core, d);
     renderCoreStatus();
-    if (d.running && !state.loadedOnce) loadFrame(state.core.baseUrl);
-    if (!d.running) showOffline();
+    if (d.running && !state.loadedOnce) scheduleFrameLoad();
+    if (!d.running) {
+      state.loadedOnce = false;
+      showOffline();
+    }
   });
   // Escape closes overlay
   document.addEventListener("keydown", (e) => {
@@ -189,6 +249,7 @@ function init() {
 }
 
 async function boot() {
+  applyTheme();
   init();
   state.settings = await window.modelrouter.settings.get();
   renderRail();

cmd/gui/scripts/dist-win-docker.sh

@@ -1,23 +1,64 @@
 #!/usr/bin/env bash
-# One-shot Windows release build for the GUI:
+# One-shot Windows release build for the GUI — ENTIRELY inside docker.
+#
 #   1) cross-compile the embedded core in docker (llmsproxy.exe + lua51.dll)
-#   2) host-side electron-builder NSIS packaging (wine required on host)
+#   2) electron-builder NSIS packaging ALSO inside docker (wine for the
+#      uninstaller step lives in the image — see Dockerfile). This is the fix
+#      for the 1.3.0/1.4.0 broken installers: the host had amd64-only wine
+#      (no syswow64 -> c0000135), which made electron-builder silently emit a
+#      264 KB degenerate exe without its 86 MB payload.
+#
 # Outputs: cmd/build/gui-dist/"ModelRouter Setup <version>.exe"
+#          cmd/build/gui-dist/modelrouter-gui-<version>-x64.nsis.7z
+#
+# Needs:   docker (any machine; no host mingw/wine/node required)
 set -euo pipefail
 
 ROOT="$(cd "$(dirname "${BASH_SOURCE[0]}")/../../.." && pwd)"
-cd "$ROOT/cmd/gui"
+IMAGE="${WIN_BUILDER_IMAGE:-modelrouter/win-builder}"
+PROXY="${GOPROXY:-https://goproxy.cn,direct}"
+CACHE_VOL="modelrouter-eb-cache"
 
-# 1) core (docker)
-"$ROOT/cmd/gui/scripts/win-core-docker.sh"
+command -v docker >/dev/null || { echo "docker required"; exit 1; }
 
-# 2) packaging (host, needs wine for electron-builder's signtool/nsis)
-command -v wine >/dev/null || { echo "wine required on host for NSIS packaging"; exit 1; }
-if ! command -v npx >/dev/null; then echo "node/npm required"; exit 1; fi
+if ! docker image inspect "$IMAGE" >/dev/null 2>&1; then
+  echo "[dist-win-docker] building image $IMAGE (first run, ~5-10 min)..."
+  docker build -t "$IMAGE" "$ROOT/cmd/gui/docker/win-builder"
+fi
 
-echo "[dist-win] electron-builder --win nsis ..."
-# bin artifacts already refreshed by win-core-docker.sh; skip the prepare step
-npx electron-builder --win nsis
+docker volume create "$CACHE_VOL" >/dev/null 2>&1 || true
 
-ls -la "$ROOT/cmd/build/gui-dist/"*.exe 2>/dev/null
-echo "[dist-win] done"
+# 1) embedded core + 2) NSIS packaging — both inside the image so wine and node
+# are the image's, never the host's. electron-builder downloads electron + nsis
+# toolchains into ~/.cache on first run; the mounted volume keeps that warm so
+# subsequent builds are incremental instead of re-fetching ~400 MB.
+echo "[dist-win-docker] cross-compiling core + electron-builder NSIS in docker ..."
+docker run --rm \
+  -v "$ROOT":/workspace \
+  -v "$CACHE_VOL":/root/.cache \
+  -e GOPROXY="$PROXY" \
+  -w /workspace \
+  "$IMAGE" bash -c '
+    set -euo pipefail
+    cp /opt/luajit2/src/lua51.dll /workspace/cmd/gui/bin/lua51.dll
+    GOOS=windows GOARCH=amd64 CGO_ENABLED=1 CC=x86_64-w64-mingw32-gcc \
+      go build -buildvcs=false -tags luajit -trimpath \
+      -o /workspace/cmd/gui/bin/llmsproxy.exe ./cmd/llmsproxy
+    echo "  -> core built"
+    # wine needs one prefix initialization so its wow64 runtime (syswow64/ntdll)
+    # exists — the exact thing that was missing on the broken host. Use a shared
+    # prefix in the cache volume so it is initialized only on the first run.
+    export WINEPREFIX=/root/.cache/wineprefix
+    if [ ! -f "$WINEPREFIX/drive_c/windows/syswow64/ntdll.dll" ]; then
+      echo "  -> initializing wine prefix (first run)"
+      WINEDEBUG=-all wineboot -i || true
+    fi
+    cd /workspace/cmd/gui
+    # node_modules ships in the repo (host npm install already ran); reuse it.
+    npx electron-builder --win nsis
+  '
+
+echo "[dist-win-docker] verifying artifacts ..."
+"$ROOT/packaging/verify-dist.sh" "$ROOT/cmd/build/gui-dist"
+
+echo "[dist-win-docker] done"

cmd/llmsproxy/main.go

@@ -23,8 +23,39 @@ import (
 
 func main() {
 	cfgPath := flag.String("config", "config.yaml", "path to gateway config file")
+	checkOnly := flag.Bool("check", false, "validate the config file and exit (0 = valid, 1 = invalid); nothing is started and no file is written")
+	showSecrets := flag.Bool("show-secrets", false, "print the config's credentials in the clear and exit; nothing is started and no file is written")
 	flag.Parse()
 
+	// -check is the deploy-time preflight: parse and validate the config
+	// without starting the Lua VM, touching runtime.json, or binding a port.
+	// It deliberately does NOT call EnsureDefault, so a missing file is an
+	// error here instead of being silently created.
+	if *checkOnly {
+		if _, err := os.Stat(*cfgPath); err != nil {
+			log.Fatalf("[llmsproxy] check: %v", err)
+		}
+		if _, err := config.Load(*cfgPath); err != nil {
+			log.Fatalf("[llmsproxy] check: %v", err)
+		}
+		log.Printf("[llmsproxy] check: %s is valid", *cfgPath)
+		return
+	}
+
+	// -show-secrets is the operator escape hatch for a config whose
+	// credentials are sealed at rest: it prints them to stdout and exits
+	// without starting anything or writing anything. Nothing else in the
+	// program prints a credential, and this path never logs to a file.
+	if *showSecrets {
+		if _, err := os.Stat(*cfgPath); err != nil {
+			log.Fatalf("[llmsproxy] show-secrets: %v", err)
+		}
+		if err := config.PrintSecrets(*cfgPath); err != nil {
+			log.Fatalf("[llmsproxy] show-secrets: %v", err)
+		}
+		return
+	}
+
 	created, err := config.EnsureDefault(*cfgPath)
 	if err != nil {
 		log.Fatalf("[llmsproxy] config: %v", err)
@@ -88,4 +119,4 @@ func main() {
 	signal.Notify(stop, syscall.SIGINT, syscall.SIGTERM)
 	<-stop
 	log.Printf("[llmsproxy] shutting down")
-}
+}

deploy.sh

@@ -0,0 +1,479 @@
+#!/usr/bin/env bash
+# ============================================================
+# llmsproxy 原子化滚动部署脚本
+#
+# 策略：
+#   1. /tmp (tmpfs) 构建，/usr/local/bin (ext4) 部署
+#   2. 构建后先 copy 到目标文件系统（.llmsproxy.tmp），
+#      再 rename(2) 做原子替换（同一文件系统内 mv 是原子的）
+#   3. runtime.json 不替换（保留运行时密钥、审计状态）
+#   4. systemctl restart → 服务中断仅在 Go 进程重启瞬间 (<500ms)
+#   5. 健康检查确认新进程正常
+#
+# 配置文件（2026-09-05 起）：
+#   本脚本不再假设 config.yaml 与部署无关。实测教训：给源加 headers 时
+#   漏删原有的 `headers: {}` → YAML 重复键 → 进程启动即退出 → 健康检查
+#   失败 → 旧的 rollback 只恢复二进制，配置错误仍在，于是旧二进制也起不来，
+#   服务陷入 systemd 重启循环（当时累计 446 次）。
+#   因此：
+#   - 支持 --config <file> 投放一份待部署配置（原子 rename 到 TARGET_CONFIG）
+#   - 重启前先用新二进制 `-check` 预检配置，不合法就地失败，绝不重启服务
+#   - rollback 同时恢复二进制与配置文件，回滚后再次 -check 才重启
+#
+# 注意事项：
+#   - 本服务依赖自身（推理依赖），restart 前后会短暂中断，
+#     但重试机制（systemd RestartSec=5 + 客户端 retry）可自愈
+#   - 适配器文件替换后需重启才能生效（VM 启动时批量加载）
+# ============================================================
+set -euo pipefail
+
+# ---------- 配置 ----------
+readonly TARGET_BIN="/usr/local/bin/llmsproxy"
+readonly TARGET_CONFIG="/etc/llmsproxy/config.yaml"
+readonly TARGET_ADAPTERS="/etc/llmsproxy/adapters"
+readonly REPO_DIR="/home/program/llmsproxy"
+readonly HEALTH_URL="http://127.0.0.1:8081/v1/models"
+readonly BIN_NAME="llmsproxy"
+
+# 构建目录在 /tmp（tmpfs 更快）
+readonly TMP_DIR="/tmp/llmsproxy-deploy"
+# 同文件系统暂存文件（用于原子 rename）
+readonly STAGING_BIN="${TARGET_BIN}.staging"
+readonly STAGING_CONFIG="${TARGET_CONFIG}.staging"
+# 回滚备份：健康检查失败时从这里恢复旧二进制 / 旧配置
+readonly BACKUP_BIN="${TARGET_BIN}.bak"
+readonly BACKUP_CONFIG="${TARGET_CONFIG}.rollback.bak"
+
+# 待部署配置（--config <file> 指定；为空 = 不改动线上配置）
+NEW_CONFIG_SRC=""
+
+usage() {
+    cat <<'EOF'
+用法: ./deploy.sh [--config <file>]
+
+  --config <file>   投放这份配置到 /etc/llmsproxy/config.yaml（原子替换）。
+                    省略时不改动线上配置，只部署二进制与适配器。
+                    无论是否指定，重启前都会用新二进制 -check 预检线上配置。
+
+流程: 构建 → 适配器同步 → 备份(二进制+配置) → 原子替换 → 配置预检
+      → 重启 → 健康检查 → 失败则回滚(二进制+配置)并重启
+EOF
+}
+
+parse_args() {
+    while (( $# > 0 )); do
+        case "$1" in
+            --config)
+                [[ $# -ge 2 ]] || { echo "--config 需要一个文件参数" >&2; exit 2; }
+                NEW_CONFIG_SRC="$2"
+                shift 2
+                ;;
+            -h|--help) usage; exit 0 ;;
+            *) echo "未知参数: $1" >&2; usage; exit 2 ;;
+        esac
+    done
+}
+
+# ---------- 日志 ----------
+log()  { printf '\033[1;34m[%s]\033[0m %s\n' "$(date +%H:%M:%S)" "$*"; }
+warn() { printf '\033[1;33m[WARN]\033[0m %s\n' "$*"; }
+err()  { printf '\033[1;31m[ERR]\033[0m %s\n' "$*" >&2; }
+fail() { err "$@"; exit 1; }
+
+# ---------- 清理 ----------
+cleanup() {
+    log "清理临时文件"
+    rm -rf "$TMP_DIR" "$STAGING_BIN" "$STAGING_CONFIG"
+}
+trap cleanup EXIT
+
+# ---------- 前置检查 ----------
+precheck() {
+    log "前置检查"
+    [[ -f "$REPO_DIR/go.mod" ]] || fail "仓库目录 $REPO_DIR 不存在"
+    command -v go       &>/dev/null || fail "未找到 go"
+    command -v systemctl &>/dev/null || fail "未找到 systemctl"
+    [[ $(id -u) -eq 0 ]] || fail "需要 root 权限"
+
+    if [[ -n "$NEW_CONFIG_SRC" ]]; then
+        [[ -f "$NEW_CONFIG_SRC" ]] || fail "待部署配置不存在: $NEW_CONFIG_SRC"
+        # 同一文件比对：--config 指向线上文件本身时没有"投放"动作可做，
+        # 但仍然会走重启前的 -check 预检。
+        if [[ "$(readlink -f "$NEW_CONFIG_SRC")" == "$(readlink -f "$TARGET_CONFIG" 2>/dev/null)" ]]; then
+            warn "--config 指向线上配置本身，跳过投放（仍会预检）"
+            NEW_CONFIG_SRC=""
+        else
+            log "待部署配置: $NEW_CONFIG_SRC → $TARGET_CONFIG"
+        fi
+    fi
+
+    OLD_SIZE=$(stat -c%s "$TARGET_BIN" 2>/dev/null || echo 0)
+    OLD_SHA256=$(sha256sum "$TARGET_BIN" 2>/dev/null | awk '{print $1}')
+
+    log "旧二进制: $TARGET_BIN (${OLD_SIZE} bytes, sha256=${OLD_SHA256:0:16})"
+    log "go: $(go version)"
+}
+
+# ---------- 备份旧二进制 + 旧配置（用于回滚）----------
+# 配置也必须备份：启动失败常常是配置问题，只回滚二进制会让旧二进制
+# 继续吃坏配置，服务照样起不来。
+backup_old_state() {
+    log "备份旧二进制到 $BACKUP_BIN"
+    if [[ -f "$TARGET_BIN" ]]; then
+        cp -f "$TARGET_BIN" "$BACKUP_BIN"
+        log "✓ 旧二进制已备份 (sha256=$(sha256sum "$BACKUP_BIN" | awk '{print $1}' | cut -c1-16))"
+    else
+        warn "旧二进制不存在，跳过备份（无回滚点）"
+    fi
+
+    if [[ -f "$TARGET_CONFIG" ]]; then
+        cp -f "$TARGET_CONFIG" "$BACKUP_CONFIG"
+        log "✓ 旧配置已备份到 $BACKUP_CONFIG (sha256=$(sha256sum "$BACKUP_CONFIG" | awk '{print $1}' | cut -c1-16))"
+    else
+        warn "线上配置不存在: $TARGET_CONFIG（首次部署？无配置回滚点）"
+    fi
+}
+
+# ---------- 构建 ----------
+build() {
+    log "构建新二进制"
+    rm -rf "$TMP_DIR"
+    mkdir -p "$TMP_DIR/bin"
+
+    cd "$REPO_DIR"
+    CGO_ENABLED=1 go build -tags luajit \
+        -trimpath \
+        -ldflags="-s -w" \
+        -o "$TMP_DIR/bin/$BIN_NAME" \
+        ./cmd/llmsproxy
+
+    NEW_BIN="$TMP_DIR/bin/$BIN_NAME"
+    NEW_SIZE=$(stat -c%s "$NEW_BIN")
+
+    if [[ "$NEW_SIZE" -eq "$OLD_SIZE" ]]; then
+        NEW_SHA256=$(sha256sum "$NEW_BIN" | awk '{print $1}')
+        if [[ "$NEW_SHA256" == "$OLD_SHA256" ]]; then
+            # 二进制无变化。只有在同时也没有待部署配置时才能直接退出；
+            # 否则会造成“--config 指定了新配置却被静默丢弃”。
+            if [[ -z "$NEW_CONFIG_SRC" ]]; then
+                warn "新二进制与旧版本完全一致（sha256 相同）且无待部署配置，跳过部署"
+                exit 0
+            fi
+            log "二进制无变化（sha256 相同），但有待部署配置，继续"
+        fi
+    fi
+
+    log "新二进制: $NEW_BIN (${NEW_SIZE} bytes)"
+
+    # 验证：确认包含关键修复
+    local checks=(
+        "mergeUsage"
+        "prompt_tokens"
+        "completion_tokens"
+        "message_start"
+    )
+    for c in "${checks[@]}"; do
+        if strings "$NEW_BIN" | grep -qF "$c"; then
+            log "  ✓ 新二进制包含: $c"
+        else
+            warn "  ⚠ 新二进制不包含: $c（非预期但非致命）"
+        fi
+    done
+}
+
+# ---------- 原子替换二进制 ----------
+# 原理：
+#   /tmp 是 tmpfs，/usr/local/bin 是 ext4——不同文件系统 rename(2) 不原子。
+#   因此先 copy 到目标文件系统上的 .staging 文件，再 rename(2) 到目标路径。
+#   rename(2) 在同一文件系统内是元数据级操作，对正在读写的进程安全：
+#   - 旧 inode 的 fd 继续有效（进程持有旧文件，不会读到截断内容）
+#   - 新进程打开 path 看到新 inode
+#   - 旧 inode 在最后一个 fd 关闭后释放
+atomic_replace_binary() {
+    log "原子替换二进制"
+
+    local new_bin_tmp="$TMP_DIR/bin/$BIN_NAME"
+    [[ -f "$new_bin_tmp" ]] || fail "新二进制文件不存在: $new_bin_tmp"
+
+    # Step 1: copy 到目标文件系统（非原子，但目标文件唯一）
+    log "  copy → $STAGING_BIN (同文件系统暂存)"
+    cp -f "$new_bin_tmp" "$STAGING_BIN"
+    chmod 0755 "$STAGING_BIN"
+
+    # Step 2: 原子 rename（同一文件系统，元数据级操作）
+    log "  rename → $TARGET_BIN (原子)"
+    mv -f "$STAGING_BIN" "$TARGET_BIN"
+
+    # 验证
+    NEW_SHA256=$(sha256sum "$TARGET_BIN" | awk '{print $1}')
+    log "✓ 二进制已原子替换 (sha256=${NEW_SHA256:0:16})"
+}
+
+# ---------- 原子替换配置 ----------
+# 与二进制同一套 copy→rename 手法：先落到 /etc 上的 .staging，再 rename(2)。
+# 只有 --config 指定了来源时才动线上配置。
+atomic_replace_config() {
+    [[ -n "$NEW_CONFIG_SRC" ]] || return 0
+
+    log "原子替换配置"
+    log "  copy → $STAGING_CONFIG (同文件系统暂存)"
+    cp -f "$NEW_CONFIG_SRC" "$STAGING_CONFIG"
+    chmod 0644 "$STAGING_CONFIG"
+
+    log "  rename → $TARGET_CONFIG (原子)"
+    mv -f "$STAGING_CONFIG" "$TARGET_CONFIG"
+    log "✓ 配置已原子替换 (sha256=$(sha256sum "$TARGET_CONFIG" | awk '{print $1}' | cut -c1-16))"
+}
+
+# ---------- 配置预检 ----------
+# 用刚部署的新二进制解析线上配置。`-check` 只做 parse + ApplyDefaults，
+# 不启动 Lua VM、不碰 runtime.json、不绑端口，因此可以安全地在服务仍在
+# 运行时执行。不合法就在重启前失败——这正是上次事故缺的那一步。
+verify_config() {
+    log "配置预检（新二进制 -check）"
+    [[ -f "$TARGET_CONFIG" ]] || fail "线上配置不存在: $TARGET_CONFIG"
+
+    local out
+    if out=$("$TARGET_BIN" -check -config "$TARGET_CONFIG" 2>&1); then
+        log "✓ 配置合法"
+        return 0
+    fi
+
+    err "配置预检失败，未重启服务（线上进程仍在跑旧配置）："
+    printf '%s\n' "$out" >&2
+    return 1
+}
+
+# ---------- 密钥可解封校验 ----------
+# config.yaml 里的凭据以 enc:v1: 密文落盘，解密钥匙是 master.key（或
+# LLMS_PROXY_MASTER_KEY）。一旦钥匙丢失/错位，配置在**启动时**才会报错，
+# 那时服务已经处于「二进制已替换 + 配置已替换」的半死状态，只能靠回滚。
+# 所以在替换任何东西之前先验：能不能用现有的钥匙解开现有的密文。
+#
+# 两种情形要分开：
+#   1. 配置里**有**密文 ⇒ 钥匙必须能解开（否则拒绝部署）
+#   2. 配置里**没有**密文 ⇒ 全新安装/首次加密，无需钥匙（但会顺手记一条，
+#      因为它意味着这次部署会把明文写成密文，而钥匙必须同时就位）
+verify_master_key() {
+    log "密钥可解封校验"
+    [[ -f "$TARGET_CONFIG" ]] || return 0   # 配置都不在，交给 verify_config 报错
+
+    local sealed
+    sealed=$(grep -c 'enc:v1:' "$TARGET_CONFIG" 2>/dev/null || echo 0)
+    if [[ "$sealed" -eq 0 ]]; then
+        warn "配置内无密文（尚未加密或全为明文）——首次加密需要 master.key 可写"
+        return 0
+    fi
+
+    # 钥匙来源与 config 包一致：env 优先，否则 runtime_file 同目录的 master.key
+    local runtime_file key_file
+    runtime_file=$(awk '/^runtime_file:/{print $2; exit}' "$TARGET_CONFIG")
+    [[ -n "$runtime_file" ]] || runtime_file="$(dirname "$TARGET_CONFIG")/runtime.json"
+    key_file="$(dirname "$runtime_file")/master.key"
+
+    if [[ -n "${LLMS_PROXY_MASTER_KEY:-}" ]]; then
+        log "  钥匙来源: LLMS_PROXY_MASTER_KEY（环境变量）"
+        # 走真实解密路径验证，而不是只检查格式
+        if out=$("$TARGET_BIN" -show-secrets -config "$TARGET_CONFIG" 2>&1); then
+            log "✓ 凭据可解封（$(echo "$out" | grep -cE '^(source|key) ') 项）"
+            return 0
+        fi
+        err "凭据无法用 LLMS_PROXY_MASTER_KEY 解封，未替换任何文件："
+        printf '%s\n' "$out" >&2
+        return 1
+    fi
+
+    if [[ ! -f "$key_file" ]]; then
+        err "配置含 $sealed 处密文，但钥匙文件不存在: $key_file"
+        err "  没有它，部署后服务将无法启动。恢复 master.key 或设置 LLMS_PROXY_MASTER_KEY 后重试。"
+        return 1
+    fi
+    log "  钥匙文件: $key_file ($(stat -c '%A' "$key_file" 2>/dev/null || echo '?')，$(wc -c < "$key_file") 字节)"
+
+    if out=$("$TARGET_BIN" -show-secrets -config "$TARGET_CONFIG" 2>&1); then
+        log "✓ 凭据可解封（$(echo "$out" | grep -cE '^(source|key) ') 项）"
+        return 0
+    fi
+    err "凭据无法解密（钥匙与配置不匹配），未替换任何文件："
+    printf '%s\n' "$out" | head -5 >&2
+    err "  密钥来源: $key_file"
+    err "  常见原因：master.key 被重新生成、恢复了不匹配的备份，或换了机器。"
+    return 1
+}
+
+# ---------- 同步适配器 ----------
+sync_adapters() {
+    log "同步适配器"
+
+    # 只覆盖内置适配器（与 internal/lua/adapters 同名的文件），
+    # 保留目录里其它运行时上传的 .lua —— WebUI 上传的自定义适配器
+    # 必须跨部署存活，清空会静默移除线上源依赖的适配器。
+    SRC_ADAPTERS="$REPO_DIR/internal/lua/adapters"
+    [[ -d "$SRC_ADAPTERS" ]] || fail "内嵌适配器目录不存在: $SRC_ADAPTERS"
+
+    # 备份旧的适配器文件（含任何运行时覆盖版本，如 opencode.lua）
+    BACKUP_DIR="/etc/llmsproxy/adapters.bak.$(date +%Y%m%d%H%M%S)"
+    if [[ -d "$TARGET_ADAPTERS" ]]; then
+        mkdir -p "$BACKUP_DIR"
+        cp -rf "$TARGET_ADAPTERS/"*.lua "$BACKUP_DIR/" 2>/dev/null || true
+        log "  旧适配器已备份到 $BACKUP_DIR"
+    fi
+
+    cp -f "$SRC_ADAPTERS/"*.lua "$TARGET_ADAPTERS/"
+    chmod 0644 "$TARGET_ADAPTERS/"*.lua
+
+    # 校验：每个适配器必须包含 usage 透传修复（opencode/openai/deepseek 等应有 'uses'）
+    local required_files=(openai deepseek anthropic gemini ollama opencode github groq kimicode mistral sensenova agentrouter)
+    for f in "${required_files[@]}"; do
+        [[ -f "$TARGET_ADAPTERS/$f.lua" ]] || warn "  缺少适配器: $f.lua"
+    done
+    local usage_fixed=0
+    for f in openai deepseek github groq kimicode mistral opencode; do
+        grep -qE '\buses\b' "$TARGET_ADAPTERS/$f.lua" 2>/dev/null && ((usage_fixed++)) || true
+    done
+    log "  usage 透传修复适配器: ${usage_fixed}/7 (openai/deepseek/github/groq/kimicode/mistral/opencode)"
+    log "✓ 适配器已同步到 $TARGET_ADAPTERS/"
+}
+
+# ---------- 重启服务 ----------
+restart_service() {
+    log "重启服务（中断 <500ms）"
+    systemctl restart llmsproxy.service || fail "systemd restart 失败"
+    log "✓ llmsproxy.service restarted"
+}
+
+# ---------- 健康检查 ----------
+healthcheck() {
+    log "健康检查"
+    local attempts=20
+    local wait_sec=1
+    local attempt=1
+
+    while (( attempt <= attempts )); do
+        # 用 curl 检查（200=有 key 未鉴权但服务正常，401=需要认证，都说明服务在运行）
+        local code
+        code=$(curl -s -o /dev/null -w "%{http_code}" --max-time 3 "$HEALTH_URL" 2>/dev/null || echo "000")
+        if [[ "$code" == "200" || "$code" == "401" ]]; then
+            log "✓ 服务健康 (status=$code, ${attempt}/${attempts})"
+            return 0
+        fi
+        log "  等待中 ($attempt/$attempts, status=$code)"
+        sleep "$wait_sec"
+        (( attempt++ ))
+    done
+
+    warn "健康检查未通过，请检查服务状态: systemctl status llmsproxy.service"
+    return 1
+}
+
+# ---------- 回滚 ----------
+# 健康检查失败时恢复旧二进制**和旧配置**并重启，保证推理服务可用。
+# 这是最后一道保险：即使新版本启动失败（编译错误、配置不兼容、
+# panic on start 等），也能让旧版本继续服务，避免推理能力丢失。
+#
+# 为什么必须一并回滚配置：启动失败最常见的原因就是配置本身不合法
+# （实测：YAML 重复键）。只恢复二进制的话，旧二进制照样解析不了坏配置，
+# 服务继续在 systemd 重启循环里空转。
+rollback() {
+    err "健康检查失败，启动回滚流程"
+
+    local restored_any=0
+    if [[ -f "$BACKUP_BIN" ]]; then
+        log "恢复旧二进制"
+        cp -f "$BACKUP_BIN" "$TARGET_BIN"
+        chmod 0755 "$TARGET_BIN"
+        restored_any=1
+    else
+        err "无二进制回滚备份 ($BACKUP_BIN 不存在)"
+    fi
+
+    # 只在本次真的换过配置时才回滚配置，避免把用户在部署期间
+    # 通过 WebUI 做的合法改动一起抹掉。
+    if [[ -n "$NEW_CONFIG_SRC" && -f "$BACKUP_CONFIG" ]]; then
+        log "恢复旧配置"
+        cp -f "$BACKUP_CONFIG" "$TARGET_CONFIG"
+        chmod 0644 "$TARGET_CONFIG"
+        restored_any=1
+    fi
+
+    if (( restored_any == 0 )); then
+        err "无任何回滚点，无法自动恢复"
+        err "请手动检查: systemctl status llmsproxy.service && journalctl -u llmsproxy -n 50"
+        return 1
+    fi
+
+    # 回滚后先验证恢复出来的组合真的能解析，再重启——否则只是把
+    # 重启循环换个二进制继续跑。
+    if ! "$TARGET_BIN" -check -config "$TARGET_CONFIG" >/dev/null 2>&1; then
+        err "回滚后的配置仍不合法，拒绝重启（避免 systemd 重启风暴）"
+        err "  手动修复 $TARGET_CONFIG 后执行: $TARGET_BIN -check -config $TARGET_CONFIG"
+        err "  备份可用: 二进制=$BACKUP_BIN 配置=$BACKUP_CONFIG"
+        return 1
+    fi
+
+    log "重启服务（旧版本）"
+    systemctl restart llmsproxy.service || {
+        err "重启失败，服务可能完全宕机"
+        err "请手动检查: systemctl status llmsproxy.service"
+        return 1
+    }
+
+    # 回滚后再健康检查（只查 1 次，失败就认了）
+    sleep 2
+    local code
+    code=$(curl -s -o /dev/null -w "%{http_code}" --max-time 5 "$HEALTH_URL" 2>/dev/null || echo "000")
+    if [[ "$code" == "200" || "$code" == "401" ]]; then
+        log "✓ 回滚成功，旧版本已恢复服务 (status=$code)"
+        log "  新版本部署失败，请检查构建/配置后重试"
+        return 0
+    else
+        err "回滚后服务仍未恢复 (status=$code)"
+        err "服务可能已完全宕机，请手动介入"
+        return 1
+    fi
+}
+
+# ---------- 打印部署摘要 ----------
+print_summary() {
+    log "═════════════════════════════════════════════"
+    log "部署完成"
+    log "  二进制: $TARGET_BIN"
+    log "  配置:   $TARGET_CONFIG$([[ -n "$NEW_CONFIG_SRC" ]] && echo "  (本次已替换)" || echo "  (未改动)")"
+    log "  适配器: $TARGET_ADAPTERS"
+    log "═════════════════════════════════════════════"
+}
+
+# ---------- 主流程 ----------
+# 顺序要点：配置投放与预检都排在 restart_service **之前**，
+# 于是坏配置在服务仍然健康时就被拦下，根本不会触发重启。
+main() {
+    parse_args "$@"
+    log "═══ llmsproxy 原子部署开始 ═══"
+    precheck
+    # 密钥校验放在 build/替换之前：它的失败后果最重（配置里的密文永久解不开），
+    # 必须在动任何文件之前就挡住。此阶段用的是**当前线上二进制**，它已经
+    # 支持 -show-secrets（首次引入该子命令的版本除外，此时跳过并提示）。
+    verify_master_key
+    build
+    sync_adapters
+    backup_old_state
+    atomic_replace_binary
+    atomic_replace_config
+    if ! verify_config; then
+        # 配置不合法：服务还没重启，只需把二进制/配置退回原状。
+        err "配置预检失败，回退本次改动（服务未受影响）"
+        [[ -f "$BACKUP_BIN" ]] && cp -f "$BACKUP_BIN" "$TARGET_BIN" && chmod 0755 "$TARGET_BIN"
+        if [[ -n "$NEW_CONFIG_SRC" && -f "$BACKUP_CONFIG" ]]; then
+            cp -f "$BACKUP_CONFIG" "$TARGET_CONFIG" && chmod 0644 "$TARGET_CONFIG"
+        fi
+        exit 1
+    fi
+    restart_service
+    if ! healthcheck; then
+        rollback || true
+        exit 1
+    fi
+    print_summary
+}
+
+main "$@"

docs/git-workflow-en.md

@@ -0,0 +1,141 @@
+# Git Branching Workflow
+
+ModelRouter uses a **GitHub Flow + release-branch** model: `main` is the only
+long-lived branch and is always deployable; new work branches off it as a
+feature branch; each version gets its own release branch with a tag; hotfixes
+land on the release branch and MUST flow back to `main`.
+
+Goal: make "patching an already-released version during its support window" a
+first-class operation, while `main` always contains every fix and stays
+deployable.
+
+## Branch roles
+
+| Branch | Name | Lifetime | Purpose |
+|---|---|---|---|
+| Main | `main` | permanent | Only long-lived branch. Always deployable. Accumulates the next version. |
+| Feature | `feature/<desc>` | short (dev → merge → delete) | New features / ordinary fixes. Born from `main`, merged back into `main`. |
+| Release | `release/vX.Y.Z` | one version cycle | Cut from `main`, tagged for release. Version-specific hotfixes land here. |
+
+## Change flow (important)
+
+```
+        feature/foo ───────┐
+                           ▼
+                        main ──────────────── next version ──────►
+                           │                        │
+                           │ cut                    │ cut
+                           ▼                        ▼
+                    release/v1.4.2            release/v1.4.3
+                           │                        │
+                    tag: v1.4.2               tag: v1.4.3
+                           │                        │
+              hotfix ◄─────┘                hotfix ◄─────┘
+                   │                              │
+                   └── cherry-pick back ──────────┘
+```
+
+### Key rules
+
+1. **main is always deployable**: never leave half-done work on `main`.
+   Unfinished work lives on a feature branch.
+2. **Features start from main and merge back**: `git checkout -b feature/xxx main`,
+   then `git merge --no-ff feature/xxx` (or squash) when done.
+3. **Release = cut a release branch from main + tag**:
+   ```bash
+   git checkout -b release/v1.4.2 main
+   git tag -a v1.4.2 -m "ModelRouter v1.4.2"
+   git push origin release/v1.4.2 v1.4.2
+   ```
+   Build installers and upload the GitCode Release from this tag so the
+   published state is exactly reproducible.
+4. **Only version-specific hotfixes go into a release branch during its
+   window**: new features always go to `main` for the next version, never into
+   an already-released branch (unless you deliberately ship a minor revision).
+5. **Hotfixes MUST flow back to main**:
+   ```bash
+   git checkout release/v1.4.2     # fix in the release branch
+   git commit -m "fix: ..."
+   git checkout main
+   git cherry-pick <hotfix-commit> # and into main
+   ```
+   Most important rule: **main only moves forward and never loses fixes**. If a
+   hotfix never reaches main, the next release ships with the old bug.
+
+### End of a version lifecycle
+
+When the next version ships, the previous release branch retires:
+
+- **Default: delete the remote release branch**
+  (`git push origin :release/v1.4.2`). All hotfixes were already
+  cherry-picked into main, so main contains everything; no merge needed.
+- **Long-term maintenance** (e.g. an enterprise client pinned to an old
+  version): keep the branch, accept only security fixes, keep the
+  commit-then-cherry-pick loop.
+
+## Explicit non-goals
+
+- **Never rebase main**: main's history stays append-only; anyone pulling gets
+  directly usable history.
+- **Release branches are not merged back wholesale**: hotfixes were already
+  cherry-picked; a whole-branch merge only creates conflicts with no benefit.
+- **No "quick edit" directly on main**, even single-person: at minimum use
+  `feature/xxx → merge main` so history keeps its why-boundaries.
+
+## Single vs multi person
+
+- **Single** (this repo today): skip PRs — `git checkout -b feature/xxx` →
+  finish → merge back to main. Release branch + tag flow unchanged.
+- **Multi / open source**: feature branches go through PR + Code Review,
+  maintainers merge; resolve conflicts on the feature branch, never on main.
+
+## Commit message conventions
+
+- `feat(...)`: new feature
+- `fix(...)`: bug fix
+- `chore(...)`: build / version / deps / non-code changes
+- `docs(...)`: documentation
+- `refactor(...)`: refactoring
+- Scope in parens, e.g. `fix(adapters)`, `feat(webui)`, `chore(version)`
+- One-line summary; expand with problem/root-cause/fix/verification when needed
+
+## Release checklist (companion)
+
+```bash
+# 1. main is ready
+git checkout main && git pull
+
+# 2. cut release branch + tag
+git checkout -b release/vX.Y.Z main
+git tag -a vX.Y.Z -m "ModelRouter vX.Y.Z"
+
+# 3. push branch + tag
+git push origin release/vX.Y.Z
+git push origin vX.Y.Z
+
+# 4. build installers (core + GUI platforms)
+make core-dist VERSION=X.Y.Z
+
+# 5. create GitCode Release + upload
+#    https://gitcode.com/JianFeeeee/ModelRouter/releases
+#    upload MUST use PUT --http1.1, otherwise the OBS callback fails and the
+#    file never lands in the release; replacing artifacts = delete the tag
+#    (release disappears with it) then recreate.
+```
+
+## Why this model (background)
+
+Before 2026-08 everything went straight to `main` plus a tag, which caused two
+pain points:
+1. After 1.4.2 shipped, a scheduler-scoring fix had to be added to the same
+   version — the only option was deleting the remote tag to clear the release
+   and re-uploading all installers. The published state was detached from code
+   history; "what exactly shipped" could not be traced.
+2. No feature branches meant two independent efforts could not proceed in
+   parallel without colliding.
+
+With release branches: the published state = `release/vX.Y.Z` branch +
+`vX.Y.Z` tag, exactly reproducible; hotfixes have a clear landing spot; main
+stays "latest + all fixes + deployable".
+
+> 中文版见 [docs/git-workflow.md](git-workflow.md)。

docs/git-workflow.md

@@ -0,0 +1,128 @@
+# Git 分支工作流（Branching Workflow）
+
+ModelRouter 采用 **GitHub Flow + 发布分支** 模型：`main` 是唯一长命分支且永远可部署，
+新工作从它开出特性分支；每个版本从 `main` 分出独立的发布分支并打 tag；hotfix 提交到
+发布分支且必须回流 `main`。
+
+目的：让「已发版的版本在生命周期内可被单独修补」成为一等操作，同时 `main` 始终包含
+全部修复、永远可部署。
+
+## 分支角色
+
+| 分支 | 命名 | 生命周期 | 用途 |
+|---|---|---|---|
+| 主分支 | `main` | 永久 | 唯一长命分支。永远可部署。积攒下一个版本的功能。 |
+| 特性分支 | `feature/<描述>` | 短命（开发→合并即删） | 新特性 / 一般 bug 修复。从 `main` 开出，完成后合回 `main`。 |
+| 发布分支 | `release/vX.Y.Z` | 一个版本周期 | 从 `main` 分出，打 tag 发布。该版本生命周期内的 hotfix 都提交在此分支。 |
+
+## 变更流向（重要）
+
+```
+        feature/foo ───────┐
+                           ▼
+                        main ──────────────── 下一个版本 ──────►
+                           │                        │
+                           │ 切出                   │ 切出
+                           ▼                        ▼
+                    release/v1.4.2            release/v1.4.3
+                           │                        │
+                    tag: v1.4.2               tag: v1.4.3
+                           │                        │
+              hotfix ◄─────┘                hotfix ◄─────┘
+                   │                              │
+                   └── cherry-pick 回 main ───────┘
+```
+
+### 关键规则
+
+1. **main 永远可部署**：不在 main 上留半成品。任何未完成的工作必须在特性分支上。
+2. **特性从 main 开，完成后合回 main**：`git checkout -b feature/xxx main`，
+   开发完 `git merge --no-ff feature/xxx` 或 squash 合回。
+3. **发布 = 从 main 切 release 分支 + 打 tag**：
+   ```bash
+   git checkout -b release/v1.4.2 main
+   git tag -a v1.4.2 -m "ModelRouter v1.4.2"
+   git push origin release/v1.4.2 v1.4.2
+   ```
+   构建安装包、上传 GitCode Release 都基于这个 tag，保证可精确回溯发布态。
+4. **版本生命周期内只收该版本的 hotfix**：新特性一律并入 `main` 等下一个版本，
+   绝不塞进已发布的 release 分支（除非主动选择在该版本内发次要版）。
+5. **hotfix 必须回流 main**：
+   ```bash
+   git checkout release/v1.4.2     # 在发布分支提交修复
+   git commit -m "fix: ..."
+   git checkout main
+   git cherry-pick <hotfix-commit> # 回主分支
+   ```
+   这是本模型最重要的一条：**main 只前进、不丢修复**。如果 hotfix 不回 main，
+   下一个版本就会带着旧 bug 发布。
+
+### 版本生命周期结束
+
+下一个版本发布时，上一个 release 分支退役：
+
+- **默认：直接删除远端 release 分支**（`git push origin :release/v1.4.2`）。
+  因为 hotfix 都已逐个 cherry-pick 回 main，main 已包含全部修复，无需再合并。
+- **如需要长期维护旧版**（例如企业大客户卡在旧版本）：保留分支，仅 stopship 接受
+  该版本的安全修复，继续走「提交 + cherry-pick 回 main」循环。
+
+## 明确不做的事
+
+- **不 rebase main**：`main` 的历史保持追加式，任何人拉取后 `git pull` 都得到直接可用的历史。
+- **发布分支不整体 merge 回 main**：hotfix 已逐个 cherry-pick，整体合并只会制造冲突且无收益。
+- **不在 main 直接写"临时改一下"**：即使是单人项目，也至少走
+  `feature/xxx → merge main` 的形式，让历史保留"为什么改"的边界。
+
+## 单人 vs 多人
+
+- **单人**（本项目当前）：特性分支可省去 PR，`git checkout -b feature/xxx` →
+  完成 → 合并回 main。发布分支与 tag 流程完全一致。
+- **多人 / 开源协作**：特性分支走 PR（pull request）+ Code Review，
+  由维护者 merge；冲突在特性分支上解决，不在 main 上解决。
+
+## 提交信息规范（与分支配套）
+
+- `feat(...)`: 新功能
+- `fix(...)`: 修复
+- `chore(...)`: 构建 / 版本号 / 依赖 / 文档无关改动
+- `docs(...)`: 文档
+- `refactor(...)`: 重构
+- 括号内为影响域，如 `fix(adapters)`、`feat(webui)`、`chore(version)`
+- 一句话总结，必要时正文展开「问题 / 根因 / 修复 / 验证」
+
+## 发布操作速查（配套）
+
+每次发布严格按此顺序：
+
+```bash
+# 1. 确认 main 就绪
+git checkout main && git pull
+
+# 2. 切发布分支并打 tag
+git checkout -b release/vX.Y.Z main
+git tag -a vX.Y.Z -m "ModelRouter vX.Y.Z"
+
+# 3. 推送分支 + tag
+git push origin release/vX.Y.Z
+git push origin vX.Y.Z
+
+# 4. 构建安装包（核心包 + GUI 各平台）
+make core-dist VERSION=X.Y.Z
+
+# 5. 创建 GitCode Release + 上传安装包
+#   见 https://gitcode.com/JianFeeeee/ModelRouter/releases
+#   上传注意：必须 PUT --http1.1，否则 OBS 回调失败文件不进列表；
+#   替换旧版附件 = 删 tag（release 随 tag 一起消失）后重建。
+```
+
+## 为什么是这套（背景）
+
+2026-08 之前本项目直接在所有改动提交到 `main` + 打 tag，出现两个痛点：
+1. 1.4.2 已发版后又追了一个调度打分修复，只能删远端 tag 清空 release 再重传安装包
+   —— 版本发布态与代码历史脱节，无法精确回溯"当时发的是什么"。
+2. 没有特性分支，无法并行开发互不干扰的两件事。
+
+引入发布分支后：版本发布态 = `release/vX.Y.Z` 分支 + `vX.Y.Z` tag，可精确回溯；
+hotfix 有明确落点；main 保持"最新 + 全部修复 + 可部署"。
+
+> 英文版见 [docs/git-workflow-en.md](git-workflow-en.md)。

docs/lua-adapters-en.md

@@ -55,6 +55,7 @@ state may be shared across workers (`log` to stdout is the only side effect).
 | `version` | string | no | Version, shown in the WebUI |
 | `endpoint` | string | no | Upstream path, default `/chat/completions`; overridable by `source.endpoint` / `source.image_endpoint` |
 | `headers` | table | no | Static default request headers; used as fallback when no `build_headers` hook is defined |
+| `transform_error(status, body)` | function | no | Error-response condensing: return a one-line reason; on nil / absence clients uniformly receive `unknown error` (raw body goes to server logs only) |
 
 These fields are extracted statically at load time (compile-once); reading them never
 occupies a pooled worker.
@@ -147,6 +148,27 @@ end
 
 ## 4. Optional Hooks
 
+### `transform_error(status, body) -> string | nil`
+
+Condense this source's error response into a one-line reason. Every upstream
+formats errors differently — that is adapter territory: all built-in adapters
+implement their own envelope parsing (zen's `{error={type,message}}`,
+Anthropic's `{type="error",error={...}}`, Gemini's
+`{error={code,message,status}}`, Ollama's string `{error="..."}`, etc.).
+When the hook is absent or returns nil the core does not guess: clients get
+`api error <status>: unknown error` and the raw body is logged server-side only.
+
+```lua
+function adapter.transform_error(status, body)
+    local ok, resp = pcall(json.decode, body)
+    if not ok or type(resp) ~= "table" then return nil end
+    if resp.error and resp.error.type == "FreeUsageLimitError" then
+        return "zen free pool quota exhausted"
+    end
+    return resp.error and resp.error.message or nil
+end
+```
+
 ### `build_headers(meta) -> table<string,string>`
 
 Dynamically generate / sign request headers (e.g. KimiCode's HMAC signature). If the

docs/lua-adapters.md

@@ -50,6 +50,7 @@ return adapter
 | `version` | string | 否 | 版本号，用于 WebUI 展示 |
 | `endpoint` | string | 否 | 上游请求路径，默认 `/chat/completions`；可被 `source.endpoint` / `source.image_endpoint` 覆盖 |
 | `headers` | table | 否 | 静态默认请求头；若未定义 `build_headers` 钩子则作为请求头回退 |
+| `transform_error(status, body)` | function | 否 | 错误响应收敛：返回一行短原因；返回 nil / 未定义时客户端统一收到 `unknown error`（原始响应体只进服务端日志） |
 
 这些字段在加载时静态提取（compile-once），之后读它们不会占用池内 worker。
 
@@ -138,6 +139,26 @@ end
 
 ## 4. 可选钩子
 
+### `transform_error(status, body) -> string | nil`
+
+把该源特有的错误响应收敛成一行短原因。每个上游的错误格式不同——这是适配器层
+的职责：内置各适配器均实现了自己的信封解析（如 zen 的
+`{error={type,message}}`、Anthropic 的 `{type="error",error={...}}`、Gemini 的
+`{error={code,message,status}}`、Ollama 的字符串 `{error="..."}` 等）。
+未定义本钩子或返回 nil 时，核心不猜测格式，客户端统一收到
+`api error <status>: unknown error`，原始响应体仅记录在服务端日志。
+
+```lua
+function adapter.transform_error(status, body)
+    local ok, resp = pcall(json.decode, body)
+    if not ok or type(resp) ~= "table" then return nil end
+    if resp.error and resp.error.type == "FreeUsageLimitError" then
+        return "zen free pool quota exhausted"
+    end
+    return resp.error and resp.error.message or nil
+end
+```
+
 ### `build_headers(meta) -> table<string,string>`
 
 动态生成/签名请求头（如 KimiCode 的 HMAC 签名）。若脚本未定义此函数，Go 层回退使用静态

e2e/e2e_test.go

@@ -315,7 +315,7 @@ func TestEndToEndAuto503(t *testing.T) {
 	if resp == nil || resp.StatusCode != http.StatusServiceUnavailable {
 		t.Fatalf("status=%v body=%q", statusOf(resp), body)
 	}
-	if !strings.Contains(body, "all auto tiers failed") || !strings.Contains(body, "good/good-m") {
+	if !strings.Contains(body, "auto providers failed") || !strings.Contains(body, "good/good-m") {
 		t.Fatalf("503 must summarize the failed slot, body=%q", body)
 	}
 }
@@ -326,4 +326,3 @@ func statusOf(resp *http.Response) int {
 	}
 	return resp.StatusCode
 }
-

internal/config/config.go

@@ -20,13 +20,18 @@ type Config struct {
 	DefaultModel  string       `yaml:"default_model"` // e.g. "AUTO" or a model id
 	AdapterDir    string       `yaml:"adapter_dir"`
 	RuntimeFile   string       `yaml:"runtime_file"`
-	MaxConcurrent int          `yaml:"max_concurrent"`            // global inflight cap, 0 = unlimited
 	TLSCertFile   string       `yaml:"tls_cert_file,omitempty"`   // PEM cert; when set together with tls_key_file, serve HTTPS
 	TLSKeyFile    string       `yaml:"tls_key_file,omitempty"`    // PEM private key
 	PublicBaseURL string       `yaml:"public_base_url,omitempty"` // external base for generated config snippets; default inferred from request
 	Sources       []Source     `yaml:"sources"`
-	Auto          []ModelScope `yaml:"auto,omitempty"`            // AUTO 调度链规则（WebUI 优先级页编辑）
-	Keys          []GWKey      `yaml:"keys,omitempty"`            // 网关密钥（WebUI 密钥页管理）
+	Auto          []ModelScope `yaml:"auto,omitempty"`       // AUTO 调度链规则（WebUI 优先级页编辑，chat）
+	AutoImage     []ModelScope `yaml:"auto_image,omitempty"` // AUTO 生图调度链规则（WebUI 优先级页·生图）
+	Keys          []GWKey      `yaml:"keys,omitempty"`       // 网关密钥（WebUI 密钥页管理）
+	// box seals credentials (sources' api_key/headers, keys' key) at rest.
+	// In-memory values are always plaintext; only the bytes on disk are sealed.
+	// Wired by AttachSecretBox — Load leaves it nil so `-check` and tests stay
+	// filesystem-free.
+	box *SecretBox
 }
 
 // Defaults applied to any source (YAML or runtime) that leaves a field unset.
@@ -47,21 +52,26 @@ type Model struct {
 
 // Source describes a single upstream LLM provider.
 type Source struct {
-	Name          string                 `yaml:"name" json:"name"`
-	BaseURL       string                 `yaml:"base_url" json:"base_url"`
-	APIKey        string                 `yaml:"api_key" json:"api_key"`
-	APIKeyEnv     string                 `yaml:"api_key_env,omitempty" json:"-"` // reference to an env var holding the key (overrides api_key)
-	Adapter       string                 `yaml:"adapter" json:"adapter"`
-	Endpoint      string                 `yaml:"endpoint" json:"endpoint,omitempty"`             // chat endpoint override
-	ImageEndpoint string                 `yaml:"image_endpoint" json:"image_endpoint,omitempty"` // image endpoint override
-	Models        []Model                `yaml:"models" json:"models"`
-	Headers       map[string]string      `yaml:"headers" json:"headers,omitempty"`
+	Name          string            `yaml:"name" json:"name"`
+	BaseURL       string            `yaml:"base_url" json:"base_url"`
+	APIKey        string            `yaml:"api_key" json:"api_key"`
+	APIKeyEnv     string            `yaml:"api_key_env,omitempty" json:"-"` // reference to an env var holding the key (overrides api_key)
+	Adapter       string            `yaml:"adapter" json:"adapter"`
+	Endpoint      string            `yaml:"endpoint" json:"endpoint,omitempty"`             // chat endpoint override
+	ImageEndpoint string            `yaml:"image_endpoint" json:"image_endpoint,omitempty"` // image endpoint override
+	Models        []Model           `yaml:"models" json:"models"`
+	Headers       map[string]string `yaml:"headers" json:"headers,omitempty"`
+	// ProxyURL routes this source's HTTP(S) traffic through an explicit proxy
+	// (e.g. http://127.0.0.1:7890). Empty = direct connection. Needed for
+	// upstreams behind geo/IP blocks that only respond through a proxy.
+	ProxyURL      string                 `yaml:"proxy_url" json:"proxy_url,omitempty"`
 	Meta          map[string]interface{} `yaml:"meta" json:"meta,omitempty"`
 	Temperature   float64                `yaml:"temperature" json:"temperature,omitempty"`
 	MaxTokens     int                    `yaml:"max_tokens" json:"max_tokens,omitempty"`
 	Timeout       time.Duration          `yaml:"timeout" json:"-"`
 	MaxConcurrent int                    `yaml:"max_concurrent" json:"max_concurrent"`
 	QueueTimeout  time.Duration          `yaml:"queue_timeout" json:"-"`
+	RPM           int                    `yaml:"rpm" json:"rpm,omitempty"` // optional requests-per-minute cap (0 = unlimited)
 }
 
 // Load reads and validates a config file. When the file does not exist yet a
@@ -103,9 +113,10 @@ func EnsureDefault(path string) (bool, error) {
 }
 
 // writeDefaultConfig writes a minimal, safe-by-default config: loopback-only
-// listen, a fresh random admin key, and a no-key zen source that works out of
-// the box. adapter_dir / runtime_file live next to the config file so the
-// binary works regardless of the working directory it is started from.
+// listen and a fresh random admin key. No upstream sources are preconfigured
+// (the repo ships none — add them via WebUI or config.yaml). adapter_dir /
+// runtime_file live next to the config file so the binary works regardless of
+// the working directory it is started from.
 func writeDefaultConfig(path string) error {
 	key, err := NewGatewayKey()
 	if err != nil {
@@ -122,15 +133,6 @@ func writeDefaultConfig(path string) error {
 		DefaultModel: "AUTO",
 		AdapterDir:   filepath.Join(abs, "adapters"),
 		RuntimeFile:  filepath.Join(abs, "runtime.json"),
-		Sources: []Source{
-			{
-				Name:    "zen",
-				BaseURL: "https://opencode.ai/zen/v1",
-				APIKey:  "public", // zen 免费池：官方无 key 客户端实际发送 Bearer public
-				Adapter: "opencode",
-				Models:  []Model{{ID: "deepseek-v4-flash-free", Priority: 100, Kind: "chat"}},
-			},
-		},
 	}
 	out, err := yaml.Marshal(&cfg)
 	if err != nil {
@@ -195,10 +197,15 @@ func RemoveSourceFromYAML(path, name string) error {
 
 // UpsertSourceInYAML adds or updates a source entry in the YAML config file.
 // Uses yaml.Node to preserve the rest of the file's comments and formatting.
-func UpsertSourceInYAML(path, name string, src Source) error {
+// When box is non-nil the source's credentials are sealed before writing, so a
+// newly added source never lands in the file as plaintext.
+func UpsertSourceInYAML(path, name string, src Source, box *SecretBox) error {
 	if path == "" {
 		return fmt.Errorf("config path is empty")
 	}
+	if err := sealSource(&src, box); err != nil {
+		return err
+	}
 	data, err := os.ReadFile(path)
 	if err != nil {
 		return err
@@ -256,11 +263,17 @@ func UpsertSourceInYAML(path, name string, src Source) error {
 
 // Save writes the current config back to the YAML file (preserving comments
 // via yaml.Node round-trip when possible, or full marshaling as fallback).
+// Credentials are sealed on the way out and the in-memory copy is restored to
+// plaintext afterwards, so callers keep working with usable values.
 func (c *Config) Save() error {
 	if c.Path == "" {
 		return fmt.Errorf("config path is empty")
 	}
+	if err := c.sealInPlace(c.box); err != nil {
+		return err
+	}
 	out, err := yaml.Marshal(c)
+	c.unsealAfterWrite(c.box)
 	if err != nil {
 		return fmt.Errorf("marshal config: %w", err)
 	}
@@ -321,11 +334,34 @@ func (c *Config) ApplyDefaults() error {
 
 // RuntimeConfig is the legacy runtime file format (kept for migration only).
 type RuntimeConfig struct {
-	Sources         []Source     `json:"sources,omitempty"`
-	DeletedSources  []string     `json:"deleted_sources,omitempty"`
-	DeletedAdapters []string     `json:"deleted_adapters,omitempty"`
-	Keys            []GWKey      `json:"keys,omitempty"`
-	Auto            []ModelScope `json:"auto,omitempty"`
+	Sources         []Source         `json:"sources,omitempty"`
+	SourceTemplates []SourceTemplate `json:"source_templates,omitempty"`
+	DeletedSources  []string         `json:"deleted_sources,omitempty"`
+	DeletedAdapters []string         `json:"deleted_adapters,omitempty"`
+	PresetTemplates []string         `json:"preset_templates,omitempty"` // preset names the user has seen (or deleted) — never re-seeded
+	Keys            []GWKey          `json:"keys,omitempty"`
+	Auto            []ModelScope     `json:"auto,omitempty"`
+	AutoImage       []ModelScope     `json:"auto_image,omitempty"`
+}
+
+// SourceTemplate stores reusable source configuration (everything except
+// name and api_key) so the WebUI can spin up multiple key-bearing sources
+// from one shared template.
+type SourceTemplate struct {
+	Name          string            `json:"name"`
+	BaseURL       string            `json:"base_url"`
+	Adapter       string            `json:"adapter"`
+	Endpoint      string            `json:"endpoint,omitempty"`
+	ImageEndpoint string            `json:"image_endpoint,omitempty"`
+	Models        []Model           `json:"models"`
+	Headers       map[string]string `json:"headers,omitempty"`
+	// ProxyURL routes this source's HTTP(S) traffic through an explicit proxy.
+	ProxyURL      string                 `json:"proxy_url,omitempty"`
+	Meta          map[string]interface{} `json:"meta,omitempty"`
+	Temperature   float64                `json:"temperature,omitempty"`
+	MaxTokens     int                    `json:"max_tokens,omitempty"`
+	MaxConcurrent int                    `json:"max_concurrent,omitempty"`
+	RPM           int                    `json:"rpm,omitempty"`
 }
 
 // GWKey is a gateway API key persisted in the config file. Role is "admin"
@@ -341,15 +377,76 @@ type GWKey struct {
 	Seed      bool         `yaml:"seed,omitempty" json:"seed,omitempty"` // true if migrated from config gateway_keys
 }
 
-// ModelScope is one allowed model for a key, or one AUTO scheduling slot,
-// with an optional token quota and reset period. TokenQuota 0 = unlimited;
-// Period "" = never resets; "hour"/"week"/"month" are fixed windows; "nhour"
-// uses Hours as the window length in hours.
+// KeyQuota is retained only to carry a scope entry's caps through the admin
+// API. Quotas are per model, never per key: there is deliberately no key-wide
+// total, so exhausting one model's budget never blocks the others.
+type KeyQuota struct {
+	TokenQuota int64  `json:"token_quota"`
+	ReqQuota   int64  `json:"req_quota"`
+	Period     string `json:"period"`
+	Hours      int64  `json:"hours"`
+}
+
+// NormalizeRole defaults an empty role to "user", so a key can never end up in
+// a state where no role means "neither admin nor user".
+func NormalizeRole(role string) string {
+	if role == "admin" {
+		return "admin"
+	}
+	return "user"
+}
+
+// Validate rejects a quota configuration that could not work as written. A
+// period is only meaningful when at least one cap is set, and a cap of zero
+// means "unlimited" rather than "deny everything", so those are the only two
+// things worth rejecting.
+func (q KeyQuota) Validate() error {
+	if q.TokenQuota < 0 {
+		return fmt.Errorf("token_quota must be >= 0 (0 = unlimited)")
+	}
+	if q.ReqQuota < 0 {
+		return fmt.Errorf("req_quota must be >= 0 (0 = unlimited)")
+	}
+	if q.Hours < 0 {
+		return fmt.Errorf("hours must be >= 0")
+	}
+	if q.TokenQuota > 0 || q.ReqQuota > 0 {
+		if err := ValidatePeriod(q.Period, q.Hours); err != nil {
+			return err
+		}
+	}
+	return nil
+}
+
+// ValidatePeriod accepts the quota period vocabulary: "" (never resets),
+// "hour", "week", "month", or "nhour" with hours >= 1. An unknown period is
+// rejected rather than silently treated as "never resets", which would turn a
+// typo into an all-time quota — the opposite of what the operator typed.
+func ValidatePeriod(period string, hours int64) error {
+	switch period {
+	case "", "hour", "week", "month":
+		return nil
+	case "nhour":
+		if hours < 1 {
+			return fmt.Errorf("period %q needs hours >= 1", period)
+		}
+		return nil
+	}
+	return fmt.Errorf("period must be one of \"\", hour, week, month, nhour (got %q)", period)
+}
+
+// ModelScope is one allowed model for a key, or one AUTO scheduling slot.
+// Its TokenQuota and ReqQuota cap THAT entry only, independently of every
+// other entry on the same key: a model that runs out of budget stops being
+// served while the key's other models keep working. TokenQuota 0 / ReqQuota 0
+// = unlimited. Period "" = never resets; "hour"/"week"/"month" are fixed
+// windows; "nhour" uses Hours.
 type ModelScope struct {
 	Model      string `yaml:"model" json:"model"`
 	Source     string `yaml:"source,omitempty" json:"source,omitempty"` // optional: pin to one upstream source; "" = any source
 	Tier       int    `yaml:"tier,omitempty" json:"tier,omitempty"`
 	TokenQuota int64  `yaml:"token_quota" json:"token_quota"`
+	ReqQuota   int64  `yaml:"req_quota,omitempty" json:"req_quota,omitempty"`
 	Period     string `yaml:"period,omitempty" json:"period,omitempty"`
 	Hours      int64  `yaml:"hours,omitempty" json:"hours,omitempty"`
 }
@@ -368,6 +465,7 @@ func (m *ModelScope) UnmarshalJSON(b []byte) error {
 		Source     string `json:"source"`
 		Tier       int    `json:"tier"`
 		TokenQuota int64  `json:"token_quota"`
+		ReqQuota   int64  `json:"req_quota"`
 		Period     string `json:"period"`
 		Hours      int64  `json:"hours"`
 	}
@@ -378,6 +476,7 @@ func (m *ModelScope) UnmarshalJSON(b []byte) error {
 	m.Source = o.Source
 	m.Tier = o.Tier
 	m.TokenQuota = o.TokenQuota
+	m.ReqQuota = o.ReqQuota
 	m.Period = o.Period
 	m.Hours = o.Hours
 	return nil

internal/config/config_test.go

@@ -91,14 +91,8 @@ func TestEnsureDefaultGeneratesOnMissingFile(t *testing.T) {
 	if cfg.Listen != "127.0.0.1:8080" {
 		t.Fatalf("listen = %q, want loopback-only", cfg.Listen)
 	}
-	if len(cfg.Sources) != 1 || cfg.Sources[0].Name != "zen" || cfg.Sources[0].Adapter != "opencode" {
-		t.Fatalf("default sources = %+v", cfg.Sources)
-	}
-	if cfg.Sources[0].APIKey != "public" {
-		t.Fatalf("zen api_key = %q", cfg.Sources[0].APIKey)
-	}
-	if cfg.Sources[0].Models[0].ID != "deepseek-v4-flash-free" {
-		t.Fatalf("default model = %+v", cfg.Sources[0].Models)
+	if len(cfg.Sources) != 0 {
+		t.Fatalf("default config must ship no sources (user adds them), got %+v", cfg.Sources)
 	}
 	// adapter_dir / runtime_file resolve next to the config file
 	if !strings.HasPrefix(cfg.AdapterDir, dir) || !strings.HasPrefix(cfg.RuntimeFile, dir) {
@@ -324,3 +318,55 @@ sources:
 		t.Fatalf("after no-op: %v %v", len(cfg.Sources), err)
 	}
 }
+
+// TestSeedTemplatesOnce verifies preset templates are seeded exactly once:
+// a user's edit survives a restart, and a deliberately deleted preset does
+// not silently come back.
+func TestSeedTemplatesOnce(t *testing.T) {
+	dir := t.TempDir()
+	path := filepath.Join(dir, "runtime.json")
+	presets := []SourceTemplate{
+		{Name: "DeepSeek", BaseURL: "https://api.deepseek.com", Adapter: "deepseek",
+			Models: []Model{{ID: "deepseek-chat", Kind: "chat"}}},
+		{Name: "OpenAI", BaseURL: "https://api.openai.com/v1", Adapter: "openai",
+			Models: []Model{{ID: "gpt-4o", Kind: "chat"}}},
+	}
+
+	// first run: both presets land in the store
+	s1 := NewStore(path)
+	if err := s1.Load(); err != nil {
+		t.Fatal(err)
+	}
+	if err := s1.SeedTemplates(presets); err != nil {
+		t.Fatalf("seed: %v", err)
+	}
+	if got := len(s1.ListTemplates()); got != 2 {
+		t.Fatalf("after seed len = %d, want 2", got)
+	}
+
+	// user edits one preset, deletes the other
+	edited := presets[0]
+	edited.BaseURL = "https://my-proxy.internal/v1"
+	if err := s1.UpsertTemplate(edited); err != nil {
+		t.Fatal(err)
+	}
+	if _, err := s1.RemoveTemplate("OpenAI"); err != nil {
+		t.Fatal(err)
+	}
+
+	// restart: re-seeding must be a no-op for both
+	s2 := NewStore(path)
+	if err := s2.Load(); err != nil {
+		t.Fatal(err)
+	}
+	if err := s2.SeedTemplates(presets); err != nil {
+		t.Fatalf("reseed: %v", err)
+	}
+	tpls := s2.ListTemplates()
+	if len(tpls) != 1 {
+		t.Fatalf("after restart len = %d, want 1 (deleted preset must not return): %+v", len(tpls), tpls)
+	}
+	if tpls[0].Name != "DeepSeek" || tpls[0].BaseURL != "https://my-proxy.internal/v1" {
+		t.Fatalf("user edit lost on restart: %+v", tpls[0])
+	}
+}

internal/config/presets.go

@@ -0,0 +1,212 @@
+package config
+
+// Built-in preset source templates. Seeded into the runtime store on first
+// startup so the WebUI can present "从模板创建" for popular providers — the
+// user only needs to supply a name and API key.
+//
+// Model IDs verified against provider docs 2026-08-28. Providers ship new IDs
+// frequently; treat these as a starting point and edit in the WebUI as needed.
+var PresetTemplates = []SourceTemplate{
+	// ── 国内热门 ────────────────────────────────────────────────────────────
+	{
+		Name:          "DeepSeek",
+		BaseURL:       "https://api.deepseek.com",
+		Adapter:       "deepseek",
+		MaxConcurrent: 8,
+		Models: []Model{
+			{ID: "deepseek-v4-flash", Priority: 100, Kind: "chat"},
+			{ID: "deepseek-v4-pro", Priority: 90, Kind: "chat"},
+			{ID: "deepseek-v4-flash-vision-exp", Priority: 60, Kind: "chat"},
+		},
+	},
+	{
+		Name:          "智谱 GLM",
+		BaseURL:       "https://open.bigmodel.cn/api/paas/v4",
+		Adapter:       "openai",
+		MaxConcurrent: 8,
+		Models: []Model{
+			{ID: "glm-5.3", Priority: 100, Kind: "chat"},
+			{ID: "glm-5.3-flash", Priority: 90, Kind: "chat"},
+			{ID: "glm-5.2", Priority: 80, Kind: "chat"},
+			{ID: "glm-5.1", Priority: 60, Kind: "chat"},
+		},
+	},
+	{
+		Name:          "Kimi / Moonshot",
+		BaseURL:       "https://api.moonshot.cn/v1",
+		Adapter:       "openai",
+		MaxConcurrent: 8,
+		Models: []Model{
+			{ID: "kimi-k3", Priority: 100, Kind: "chat"},
+			{ID: "kimi-k2.7-code", Priority: 90, Kind: "chat"},
+			{ID: "kimi-k2.7-code-highspeed", Priority: 85, Kind: "chat"},
+			{ID: "kimi-k2.6", Priority: 70, Kind: "chat"},
+		},
+	},
+	{
+		Name:          "SiliconFlow",
+		BaseURL:       "https://api.siliconflow.cn/v1",
+		Adapter:       "openai",
+		MaxConcurrent: 8,
+		Models: []Model{
+			{ID: "deepseek-ai/DeepSeek-V4-Pro", Priority: 100, Kind: "chat"},
+			{ID: "deepseek-ai/DeepSeek-V4-Flash", Priority: 90, Kind: "chat"},
+			{ID: "moonshotai/Kimi-K2.6", Priority: 80, Kind: "chat"},
+			{ID: "zai-org/GLM-5.1", Priority: 70, Kind: "chat"},
+			{ID: "Kwai-Kolors/Kolors", Priority: 50, Kind: "image"},
+		},
+	},
+	{
+		Name:          "讯飞星火 (Spark)",
+		BaseURL:       "https://spark-api-open.xf-yun.com/v1",
+		Adapter:       "openai",
+		MaxConcurrent: 4,
+		Models: []Model{
+			{ID: "4.0Ultra", Priority: 100, Kind: "chat"},
+			{ID: "generalv3.5", Priority: 80, Kind: "chat"},
+			{ID: "max-32k", Priority: 70, Kind: "chat"},
+		},
+	},
+	{
+		Name:          "阿里云 通义千问",
+		BaseURL:       "https://dashscope.aliyuncs.com/compatible-mode/v1",
+		Adapter:       "openai",
+		MaxConcurrent: 8,
+		Models: []Model{
+			{ID: "qwen3.8-max", Priority: 100, Kind: "chat"},
+			{ID: "qwen3.8-flash", Priority: 90, Kind: "chat"},
+			{ID: "qwen3.7-plus", Priority: 85, Kind: "chat"},
+			{ID: "deepseek-v4-pro", Priority: 70, Kind: "chat"},
+			{ID: "kimi-k3", Priority: 70, Kind: "chat"},
+		},
+	},
+	{
+		Name:          "百度 千帆",
+		BaseURL:       "https://qianfan.baidubce.com/v2",
+		Adapter:       "openai",
+		MaxConcurrent: 8,
+		Models: []Model{
+			{ID: "ernie-5.0-turbo", Priority: 100, Kind: "chat"},
+			{ID: "ernie-5.0", Priority: 90, Kind: "chat"},
+			{ID: "deepseek-v4-pro", Priority: 70, Kind: "chat"},
+		},
+	},
+	{
+		Name:          "豆包 (火山方舟)",
+		BaseURL:       "https://ark.cn-beijing.volces.com/api/v3",
+		Adapter:       "openai",
+		MaxConcurrent: 8,
+		Models: []Model{
+			{ID: "doubao-seed-2.0-pro", Priority: 100, Kind: "chat"},
+			{ID: "doubao-seed-2.0", Priority: 90, Kind: "chat"},
+			{ID: "doubao-seed-2.0-flash", Priority: 80, Kind: "chat"},
+		},
+	},
+	{
+		Name:          "MiniMax",
+		BaseURL:       "https://api.minimax.chat/v1",
+		Adapter:       "openai",
+		MaxConcurrent: 8,
+		Models: []Model{
+			{ID: "MiniMax-M3", Priority: 100, Kind: "chat"},
+			{ID: "MiniMax-M2.5", Priority: 80, Kind: "chat"},
+		},
+	},
+	// ── 国际主流 ────────────────────────────────────────────────────────────
+	{
+		Name:          "OpenAI",
+		BaseURL:       "https://api.openai.com/v1",
+		Adapter:       "openai",
+		MaxConcurrent: 8,
+		Models: []Model{
+			{ID: "gpt-5.6-sol", Priority: 100, Kind: "chat"},
+			{ID: "gpt-5.6-terra", Priority: 90, Kind: "chat"},
+			{ID: "gpt-5.6-luna", Priority: 80, Kind: "chat"},
+		},
+	},
+	{
+		Name:          "Anthropic Claude",
+		BaseURL:       "https://api.anthropic.com",
+		Adapter:       "anthropic",
+		MaxConcurrent: 8,
+		Models: []Model{
+			{ID: "claude-opus-5", Priority: 100, Kind: "chat"},
+			{ID: "claude-fable-5", Priority: 95, Kind: "chat"},
+			{ID: "claude-sonnet-5", Priority: 85, Kind: "chat"},
+			{ID: "claude-haiku-4-5", Priority: 70, Kind: "chat"},
+		},
+	},
+	{
+		Name:          "Google Gemini",
+		BaseURL:       "https://generativelanguage.googleapis.com/v1beta",
+		Adapter:       "gemini",
+		MaxConcurrent: 8,
+		Models: []Model{
+			{ID: "gemini-3.6-flash", Priority: 100, Kind: "chat"},
+			{ID: "gemini-3.5-flash", Priority: 90, Kind: "chat"},
+			{ID: "gemini-3.5-flash-lite", Priority: 80, Kind: "chat"},
+			{ID: "gemini-3.1-pro", Priority: 85, Kind: "chat"},
+		},
+	},
+	{
+		Name:          "Groq",
+		BaseURL:       "https://api.groq.com/openai/v1",
+		Adapter:       "groq",
+		MaxConcurrent: 8,
+		Models: []Model{
+			{ID: "openai/gpt-oss-120b", Priority: 100, Kind: "chat"},
+			{ID: "openai/gpt-oss-20b", Priority: 80, Kind: "chat"},
+			{ID: "groq/compound", Priority: 70, Kind: "chat"},
+		},
+	},
+	{
+		Name:          "Mistral",
+		BaseURL:       "https://api.mistral.ai/v1",
+		Adapter:       "mistral",
+		MaxConcurrent: 8,
+		Models: []Model{
+			{ID: "mistral-medium-latest", Priority: 100, Kind: "chat"},
+			{ID: "mistral-large-latest", Priority: 90, Kind: "chat"},
+			{ID: "codestral-latest", Priority: 85, Kind: "chat"},
+		},
+	},
+	{
+		Name:          "OpenRouter",
+		BaseURL:       "https://openrouter.ai/api/v1",
+		Adapter:       "openai",
+		MaxConcurrent: 8,
+		Headers: map[string]string{
+			"HTTP-Referer": "http://localhost:8080",
+			"X-Title":      "ModelRouter",
+		},
+		Models: []Model{
+			{ID: "openrouter/auto", Priority: 100, Kind: "chat"},
+			{ID: "deepseek/deepseek-v4-flash", Priority: 90, Kind: "chat"},
+			{ID: "anthropic/claude-opus-5", Priority: 85, Kind: "chat"},
+			{ID: "openai/gpt-5.6-sol", Priority: 85, Kind: "chat"},
+		},
+	},
+	{
+		Name:          "GitHub Models",
+		BaseURL:       "https://models.inference.ai.azure.com",
+		Adapter:       "github",
+		MaxConcurrent: 8,
+		Models: []Model{
+			{ID: "gpt-5.6-sol", Priority: 100, Kind: "chat"},
+			{ID: "gpt-5.6-luna", Priority: 80, Kind: "chat"},
+		},
+	},
+	// ── 本地模型 ────────────────────────────────────────────────────────────
+	{
+		Name:          "Ollama (localhost)",
+		BaseURL:       "http://127.0.0.1:11434",
+		Adapter:       "ollama",
+		Endpoint:      "/api/chat",
+		MaxConcurrent: 4,
+		Models: []Model{
+			{ID: "qwen3", Priority: 80, Kind: "chat"},
+			{ID: "deepseek-r1", Priority: 80, Kind: "chat"},
+			{ID: "llama3.3", Priority: 70, Kind: "chat"},
+		},
+	},
+}

internal/config/secret.go

@@ -128,15 +128,3 @@ func (b *SecretBox) MustDecrypt(v string) string {
 	}
 	return out
 }
-
-// resolvedAPIKey resolves a source key from api_key_env (env var) or api_key.
-func resolvedAPIKey(src Source, box *SecretBox) (string, error) {
-	if src.APIKeyEnv != "" {
-		v := os.Getenv(src.APIKeyEnv)
-		if v == "" {
-			return "", fmt.Errorf("source %s: env %s is empty or unset", src.Name, src.APIKeyEnv)
-		}
-		return v, nil
-	}
-	return box.MustDecrypt(src.APIKey), nil
-}

internal/config/secret_config.go

@@ -0,0 +1,317 @@
+package config
+
+// Secret handling for config.yaml.
+//
+// config.yaml is 0644 world-readable by design (ops need to inspect it), so any
+// credential in it must not sit there in plaintext. Sources' api_key / headers
+// and gateway keys are therefore sealed at rest with the same SecretBox used by
+// the runtime store, and unsealed in memory at load time.
+//
+// The invariant that makes this safe: **in-memory values are always plaintext**,
+// and the enc:v1: prefix is what marks a file value as sealed. Read paths that
+// predate this (core.resolveSourceKey) already unseal, so only the write side and
+// the load-time normalize step are new.
+
+import (
+	"fmt"
+	"log"
+	"os"
+	"strings"
+)
+
+// sealSource seals one source's credentials in place (used by the YAML upsert
+// path, which is a package function and therefore has no Config to borrow a box
+// from).
+func sealSource(s *Source, box *SecretBox) error {
+	if box == nil {
+		return nil
+	}
+	if s.APIKey != "" && !strings.HasPrefix(s.APIKey, encPrefix) {
+		v, err := box.Encrypt(s.APIKey)
+		if err != nil {
+			return err
+		}
+		s.APIKey = v
+	}
+	for k, v := range s.Headers {
+		if v == "" || strings.HasPrefix(v, encPrefix) {
+			continue
+		}
+		e, err := box.Encrypt(v)
+		if err != nil {
+			return err
+		}
+		s.Headers[k] = e
+	}
+	return nil
+}
+
+// normalizeSecrets unseals every credential in the freshly parsed config so the
+// rest of the program only ever sees plaintext. A value without the enc:v1:
+// prefix is left untouched, which keeps hand-written plaintext configs working
+// (and is what a pre-encryption config file looks like).
+//
+// A value that carries the prefix but fails to decrypt is a hard error, not
+// something to paper over: returning the ciphertext (MustDecrypt's behavior)
+// would let the next Save re-seal it and turn one bad value into permanent,
+// compounding corruption. Losing the master key must be loud.
+func (c *Config) normalizeSecrets(box *SecretBox) error {
+	if box == nil {
+		return nil
+	}
+	for i := range c.Sources {
+		s := &c.Sources[i]
+		if strings.HasPrefix(s.APIKey, encPrefix) {
+			v, err := box.Decrypt(s.APIKey)
+			if err != nil {
+				return fmt.Errorf("source %q api_key: %w", s.Name, err)
+			}
+			s.APIKey = v
+		}
+		for k, v := range s.Headers {
+			if strings.HasPrefix(v, encPrefix) {
+				d, err := box.Decrypt(v)
+				if err != nil {
+					return fmt.Errorf("source %q header %q: %w", s.Name, k, err)
+				}
+				s.Headers[k] = d
+			}
+		}
+	}
+	for i := range c.Keys {
+		if strings.HasPrefix(c.Keys[i].Key, encPrefix) {
+			v, err := box.Decrypt(c.Keys[i].Key)
+			if err != nil {
+				return fmt.Errorf("gateway key %q: %w", c.Keys[i].Name, err)
+			}
+			c.Keys[i].Key = v
+		}
+	}
+	return nil
+}
+
+// sealInPlace replaces plaintext credentials with ciphertext for writing. It is
+// deliberately a separate step from Marshal: callers that need the plaintext
+// (auth comparisons, log output, returning a key to the operator who just
+// created it) must not be handed a sealed config by accident.
+func (c *Config) sealInPlace(box *SecretBox) error {
+	if box == nil {
+		return nil
+	}
+	for i := range c.Sources {
+		s := &c.Sources[i]
+		if s.APIKey != "" && !strings.HasPrefix(s.APIKey, encPrefix) {
+			v, err := box.Encrypt(s.APIKey)
+			if err != nil {
+				return err
+			}
+			s.APIKey = v
+		}
+		if len(s.Headers) > 0 {
+			sealed := make(map[string]string, len(s.Headers))
+			for k, v := range s.Headers {
+				if v == "" || strings.HasPrefix(v, encPrefix) {
+					sealed[k] = v
+					continue
+				}
+				e, err := box.Encrypt(v)
+				if err != nil {
+					return err
+				}
+				sealed[k] = e
+			}
+			s.Headers = sealed
+		}
+	}
+	for i := range c.Keys {
+		k := &c.Keys[i]
+		if k.Key != "" && !strings.HasPrefix(k.Key, encPrefix) {
+			v, err := box.Encrypt(k.Key)
+			if err != nil {
+				return err
+			}
+			k.Key = v
+		}
+	}
+	return nil
+}
+
+// unsealAfterWrite restores plaintext after a sealed marshal so the live process
+// keeps working on plaintext values (mirrors Store.persistLocked's dance).
+func (c *Config) unsealAfterWrite(box *SecretBox) {
+	// Save just encrypted every value it can see, so a failure here is
+	// impossible; ignore the error rather than panic in a write path.
+	_ = c.normalizeSecrets(box)
+}
+
+// hasPlaintextSecrets reports whether any credential in the config is still in
+// the clear. Used to decide whether a startup migration write is needed, and to
+// warn (without leaking values) when no master key is available.
+func (c *Config) hasPlaintextSecrets() bool {
+	for _, s := range c.Sources {
+		if s.APIKey != "" && !strings.HasPrefix(s.APIKey, encPrefix) {
+			return true
+		}
+		for _, v := range s.Headers {
+			if v != "" && !strings.HasPrefix(v, encPrefix) {
+				return true
+			}
+		}
+	}
+	for _, k := range c.Keys {
+		if k.Key != "" && !strings.HasPrefix(k.Key, encPrefix) {
+			return true
+		}
+	}
+	return false
+}
+
+// countPlaintextSecrets returns how many credentials are still in the clear, for
+// an operator-facing migration log line that must not print the values.
+func (c *Config) countPlaintextSecrets() int {
+	n := 0
+	for _, s := range c.Sources {
+		if s.APIKey != "" && !strings.HasPrefix(s.APIKey, encPrefix) {
+			n++
+		}
+		for _, v := range s.Headers {
+			if v != "" && !strings.HasPrefix(v, encPrefix) {
+				n++
+			}
+		}
+	}
+	for _, k := range c.Keys {
+		if k.Key != "" && !strings.HasPrefix(k.Key, encPrefix) {
+			n++
+		}
+	}
+	return n
+}
+
+// NormalizeSecretsForRun unseals the loaded config and then seals it back on
+// disk if anything was still in the clear. Order matters: Load() read the file
+// with ciphertext still in place, so the unseal has to happen before the
+// registry (and any Save the startup path performs) sees the values.
+func (c *Config) NormalizeSecretsForRun(box *SecretBox) error {
+	hadPlaintext, err := c.UnsealSecrets(box)
+	if err != nil {
+		return err
+	}
+	return c.SealIfNeeded(hadPlaintext)
+}
+
+// UnsealSecrets decrypts every sealed credential in memory and reports whether
+// the config ON DISK still held plaintext (i.e. whether a sealing write is
+// needed). It never writes; a failed decrypt (wrong master key) is returned so
+// the process refuses to start instead of running with unusable credentials.
+//
+// The return value must be computed BEFORE unsealing and from the disk state,
+// not from memory: after a Save the in-memory values are always plaintext, so a
+// "is anything plaintext?" test run afterwards is unconditionally true and a
+// caller would rewrite the file on every start. That was the actual behaviour -
+// migratePlaintextSecrets() claimed to be idempotent in a comment but rewrote
+// config.yaml on every boot.
+//
+// Callers that consume credentials (the provider registry, key seeding) must
+// unseal FIRST. Seeding compares cfg.Keys[i].Key against the plaintext
+// gateway_keys entries; running it while keys are still ciphertext made the
+// dedupe never match, so every restart appended another copy of the same admin
+// key (production accumulated four).
+func (c *Config) UnsealSecrets(box *SecretBox) (bool, error) {
+	if box == nil {
+		return false, nil
+	}
+	c.AttachSecretBox(box)
+	hadPlaintext := c.hasPlaintextSecrets()
+	if err := c.normalizeSecrets(box); err != nil {
+		return hadPlaintext, err
+	}
+	return hadPlaintext, nil
+}
+
+// SealIfNeeded writes the config back once if hadPlaintext reported that the
+// file still held clear-text credentials. When it is false the file is left
+// untouched, which is what makes startup a no-op for an already-sealed config.
+func (c *Config) SealIfNeeded(hadPlaintext bool) error {
+	if !hadPlaintext || c.box == nil || c.Path == "" {
+		return nil
+	}
+	n := c.countPlaintextSecrets()
+	if err := c.Save(); err != nil {
+		return err
+	}
+	log.Printf("[config] sealed %d plaintext credential(s) in %s", n, c.Path)
+	return nil
+}
+
+// AttachSecretBox wires the encryption box into the config so Save can seal
+// credentials. Kept as an explicit call (rather than a constructor argument) so
+// config.Load stays usable in contexts that have no filesystem secrets (tests,
+// `-check`).
+func (c *Config) AttachSecretBox(box *SecretBox) { c.box = box }
+
+// SecretBox returns the wired encryption box, or nil when none is attached.
+func (c *Config) SecretBox() *SecretBox { return c.box }
+
+// migratePlaintextSecrets seals any credential still in the clear and writes the
+// file once. Kept for callers that attach the box themselves; it decides from
+// the in-memory state, which is why UnsealSecrets + SealIfNeeded (which decide
+// from the on-disk state) are preferred on the startup path.
+func (c *Config) migratePlaintextSecrets() error {
+	if c.box == nil || c.Path == "" {
+		return nil
+	}
+	if !c.hasPlaintextSecrets() {
+		return nil
+	}
+	n := c.countPlaintextSecrets()
+	if err := c.Save(); err != nil {
+		return err
+	}
+	log.Printf("[config] sealed %d plaintext credential(s) in %s", n, c.Path)
+	return nil
+}
+
+// PrintSecrets writes the config's credentials to stdout in the clear and
+// returns an error only when the file cannot be read or the master key does not
+// match. It is the operator counterpart to sealing-at-rest: with keys stored as
+// ciphertext, "which key is this source using" must still be answerable.
+//
+// It deliberately takes a path rather than a *Config so the caller cannot
+// accidentally hand it a config that has already been unsealed in memory, and it
+// never writes anything.
+func PrintSecrets(path string) error {
+	cfg, err := Load(path)
+	if err != nil {
+		return err
+	}
+	box, err := NewSecretBox(cfg.RuntimeFile)
+	if err != nil {
+		return fmt.Errorf("%w (is master.key present and intact?)", err)
+	}
+	if err := cfg.normalizeSecrets(box); err != nil {
+		return err
+	}
+	w := os.Stdout
+	fmt.Fprintf(w, "# %s — %d source(s), %d gateway key(s)\n", path, len(cfg.Sources), len(cfg.Keys))
+	for _, s := range cfg.Sources {
+		fmt.Fprintf(w, "source %-16s api_key=%s\n", s.Name, orNone(s.APIKey))
+		for k, v := range s.Headers {
+			if v != "" {
+				fmt.Fprintf(w, "source %-16s header[%s]=%s\n", s.Name, k, v)
+			}
+		}
+	}
+	for _, k := range cfg.Keys {
+		fmt.Fprintf(w, "key    %-16s role=%-5s %s\n", k.Name, k.Role, k.Key)
+	}
+	fmt.Fprintf(w, "gateway_keys (legacy): %s\n", strings.Join(cfg.GatewayKeys, " "))
+	return nil
+}
+
+func orNone(s string) string {
+	if s == "" {
+		return "(unset)"
+	}
+	return s
+}

internal/config/secret_config_test.go

@@ -0,0 +1,323 @@
+package config
+
+import (
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+)
+
+// writeMasterKey plants a deterministic master.key next to the given runtime
+// file so tests exercise the real box instead of the env-var shortcut.
+func writeMasterKey(t *testing.T, runtimeFile string) *SecretBox {
+	t.Helper()
+	if err := os.WriteFile(filepath.Join(filepath.Dir(runtimeFile), "master.key"),
+		[]byte(strings.Repeat("ab", 32)), 0600); err != nil {
+		t.Fatal(err)
+	}
+	box, err := NewSecretBox(runtimeFile)
+	if err != nil {
+		t.Fatalf("NewSecretBox: %v", err)
+	}
+	return box
+}
+
+func TestSaveSealsCredentialsAndLoadUnseals(t *testing.T) {
+	dir := t.TempDir()
+	path := filepath.Join(dir, "config.yaml")
+	runtime := filepath.Join(dir, "runtime.json")
+	box := writeMasterKey(t, runtime)
+
+	cfg := &Config{
+		Path:   path,
+		Listen: "127.0.0.1:0",
+		Sources: []Source{{
+			Name:    "up",
+			BaseURL: "http://up/v1",
+			APIKey:  "sk-plaintext-secret",
+			Adapter: "openai",
+			Headers: map[string]string{"X-Extra": "header-secret"},
+			Models:  []Model{{ID: "m", Kind: "chat"}},
+		}},
+		Keys: []GWKey{{Key: "sk-gw-plain", Role: "admin", Name: "admin"}},
+	}
+	cfg.AttachSecretBox(box)
+	if err := cfg.Save(); err != nil {
+		t.Fatalf("Save: %v", err)
+	}
+
+	raw, err := os.ReadFile(path)
+	if err != nil {
+		t.Fatal(err)
+	}
+	text := string(raw)
+	if strings.Contains(text, "sk-plaintext-secret") {
+		t.Error("source api_key written in plaintext")
+	}
+	if strings.Contains(text, "header-secret") {
+		t.Error("source header written in plaintext")
+	}
+	if strings.Contains(text, "sk-gw-plain") {
+		t.Error("gateway key written in plaintext")
+	}
+	if n := strings.Count(text, encPrefix); n != 3 {
+		t.Errorf("expected 3 sealed values, found %d", n)
+	}
+
+	// The live in-memory config must still hold plaintext after Save.
+	if cfg.Sources[0].APIKey != "sk-plaintext-secret" {
+		t.Errorf("in-memory api_key mutated by Save: %q", cfg.Sources[0].APIKey)
+	}
+	if cfg.Keys[0].Key != "sk-gw-plain" {
+		t.Errorf("in-memory gateway key mutated by Save: %q", cfg.Keys[0].Key)
+	}
+
+	// A fresh load must hand back plaintext again.
+	loaded, err := Load(path)
+	if err != nil {
+		t.Fatalf("Load: %v", err)
+	}
+	loaded.AttachSecretBox(box)
+	loaded.normalizeSecrets(box)
+	if loaded.Sources[0].APIKey != "sk-plaintext-secret" {
+		t.Errorf("loaded api_key = %q, want plaintext", loaded.Sources[0].APIKey)
+	}
+	if loaded.Sources[0].Headers["X-Extra"] != "header-secret" {
+		t.Errorf("loaded header = %q, want plaintext", loaded.Sources[0].Headers["X-Extra"])
+	}
+	if loaded.Keys[0].Key != "sk-gw-plain" {
+		t.Errorf("loaded gateway key = %q, want plaintext", loaded.Keys[0].Key)
+	}
+}
+
+func TestPlaintextConfigStillLoads(t *testing.T) {
+	dir := t.TempDir()
+	path := filepath.Join(dir, "config.yaml")
+	// A hand-written, pre-encryption config: no enc:v1: anywhere.
+	body := `listen: 127.0.0.1:0
+sources:
+  - name: legacy
+    base_url: http://legacy/v1
+    api_key: sk-written-by-hand
+    adapter: openai
+    models:
+      - id: m
+        kind: chat
+`
+	if err := os.WriteFile(path, []byte(body), 0644); err != nil {
+		t.Fatal(err)
+	}
+	cfg, err := Load(path)
+	if err != nil {
+		t.Fatalf("Load: %v", err)
+	}
+	box := writeMasterKey(t, filepath.Join(dir, "runtime.json"))
+	cfg.AttachSecretBox(box)
+	cfg.normalizeSecrets(box)
+	if cfg.Sources[0].APIKey != "sk-written-by-hand" {
+		t.Errorf("plaintext config value changed: %q", cfg.Sources[0].APIKey)
+	}
+}
+
+// Sealing a config is a one-time migration: the FIRST run writes, every later
+// start must leave the file alone.
+//
+// This test previously called migratePlaintextSecrets() twice and compared
+// ModTime. That assertion was both weak and wrong:
+//
+//   - migratePlaintextSecrets decides from IN-MEMORY state, and Save() ends by
+//     unsealing memory so the running process keeps working. So every call sees
+//     "plaintext present" and rewrites the file. Comparing ModTime hid this
+//     because two writes inside one filesystem timestamp tick look identical —
+//     the test flaked (~1 in 6) instead of failing.
+//   - it also exercised a function production no longer calls on the startup
+//     path, which now uses UnsealSecrets + SealIfNeeded (they decide from the
+//     ON-DISK state).
+//
+// So the test targets that real path and compares file CONTENT, which cannot be
+// fooled by timestamp granularity.
+func TestSealingIsIdempotentAcrossStarts(t *testing.T) {
+	dir := t.TempDir()
+	path := filepath.Join(dir, "config.yaml")
+	runtime := filepath.Join(dir, "runtime.json")
+	box := writeMasterKey(t, runtime)
+
+	cfg := &Config{
+		Path:    path,
+		Listen:  "127.0.0.1:0",
+		Sources: []Source{{Name: "up", BaseURL: "http://up/v1", APIKey: "sk-clear", Adapter: "openai", Models: []Model{{ID: "m"}}}},
+	}
+	cfg.AttachSecretBox(box)
+
+	hadPlaintext, err := cfg.UnsealSecrets(box)
+	if err != nil {
+		t.Fatalf("unseal: %v", err)
+	}
+	if !hadPlaintext {
+		t.Fatal("a config written with a clear-text credential must report plaintext")
+	}
+	if err := cfg.SealIfNeeded(hadPlaintext); err != nil {
+		t.Fatalf("seal: %v", err)
+	}
+	first, err := os.ReadFile(path)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !strings.Contains(string(first), encPrefix) {
+		t.Fatal("sealing did not encrypt the value on disk")
+	}
+	// In-memory must be plaintext so the running process keeps working.
+	if cfg.Sources[0].APIKey != "sk-clear" {
+		t.Errorf("in-memory api_key = %q, want plaintext", cfg.Sources[0].APIKey)
+	}
+
+	// Simulate the next start: load from disk, unseal, seal-if-needed. The file
+	// was already sealed, so nothing may be written.
+	cfg2, err := Load(path)
+	if err != nil {
+		t.Fatal(err)
+	}
+	cfg2.AttachSecretBox(box)
+	hadPlaintext2, err := cfg2.UnsealSecrets(box)
+	if err != nil {
+		t.Fatalf("second unseal: %v", err)
+	}
+	if hadPlaintext2 {
+		t.Error("a sealed config reported plaintext — SealIfNeeded would rewrite " +
+			"the file on every start")
+	}
+	if err := cfg2.SealIfNeeded(hadPlaintext2); err != nil {
+		t.Fatalf("second seal: %v", err)
+	}
+	second, err := os.ReadFile(path)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if string(first) != string(second) {
+		t.Error("the second start rewrote an already-sealed config; startup must be a no-op")
+	}
+}
+
+func TestSealedValueDoesNotDoubleEncrypt(t *testing.T) {
+	dir := t.TempDir()
+	box := writeMasterKey(t, filepath.Join(dir, "runtime.json"))
+	cfg := &Config{Path: filepath.Join(dir, "config.yaml"), Sources: []Source{{Name: "a", BaseURL: "http://a"}}}
+	cfg.AttachSecretBox(box)
+
+	once, err := box.Encrypt("sk-x")
+	if err != nil {
+		t.Fatal(err)
+	}
+	cfg.Sources[0].APIKey = once
+	if err := cfg.Save(); err != nil {
+		t.Fatal(err)
+	}
+	got, err := box.Decrypt(cfg.Sources[0].APIKey)
+	if err != nil {
+		t.Fatalf("value became undecryptable: %v", err)
+	}
+	if got != "sk-x" {
+		t.Errorf("round trip = %q, want sk-x", got)
+	}
+}
+
+func TestUpsertSourceInYAMLSealsCredentials(t *testing.T) {
+	dir := t.TempDir()
+	path := filepath.Join(dir, "config.yaml")
+	box := writeMasterKey(t, filepath.Join(dir, "runtime.json"))
+
+	if err := os.WriteFile(path, []byte("listen: 127.0.0.1:0\nsources: []\n"), 0644); err != nil {
+		t.Fatal(err)
+	}
+	src := Source{Name: "new", BaseURL: "http://new/v1", APIKey: "sk-fresh", Adapter: "openai", Models: []Model{{ID: "m"}}}
+	if err := UpsertSourceInYAML(path, src.Name, src, box); err != nil {
+		t.Fatalf("UpsertSourceInYAML: %v", err)
+	}
+	raw, _ := os.ReadFile(path)
+	if strings.Contains(string(raw), "sk-fresh") {
+		t.Error("newly added source stored its api_key in plaintext")
+	}
+	if !strings.Contains(string(raw), encPrefix) {
+		t.Error("newly added source was not sealed")
+	}
+}
+
+func TestCountPlaintextSecrets(t *testing.T) {
+	cfg := &Config{
+		Sources: []Source{
+			{Name: "a", APIKey: encPrefix + "abc", Headers: map[string]string{"H": encPrefix + "x"}},
+			{Name: "b", APIKey: "sk-in-clear", Headers: map[string]string{"H2": "clear-too"}},
+		},
+		Keys: []GWKey{{Key: "sk-gw-clear"}},
+	}
+	if got := cfg.countPlaintextSecrets(); got != 3 {
+		t.Errorf("countPlaintextSecrets = %d, want 3", got)
+	}
+	if !cfg.hasPlaintextSecrets() {
+		t.Error("hasPlaintextSecrets = false, want true")
+	}
+	cfg.Sources[1].APIKey = encPrefix + "y"
+	cfg.Sources[1].Headers["H2"] = encPrefix + "z"
+	cfg.Keys[0].Key = encPrefix + "k"
+	if cfg.hasPlaintextSecrets() {
+		t.Error("hasPlaintextSecrets = true after sealing everything")
+	}
+}
+
+// TestNormalizeSecretsFailsLoudlyOnWrongMasterKey is the negative case that
+// guards the migration: with a wrong key, a sealed value must NOT be handed
+// back as ciphertext for a later re-seal (that compounds corruption and turns
+// one lost key file into permanently unusable config).
+func TestNormalizeSecretsFailsLoudlyOnWrongMasterKey(t *testing.T) {
+	dir := t.TempDir()
+	path := filepath.Join(dir, "config.yaml")
+	runtime := filepath.Join(dir, "runtime.json")
+	good := writeMasterKey(t, runtime)
+
+	cfg := &Config{
+		Path:    path,
+		Listen:  "127.0.0.1:0",
+		Sources: []Source{{Name: "up", BaseURL: "http://up/v1", APIKey: "sk-secret", Adapter: "openai", Models: []Model{{ID: "m"}}}},
+	}
+	cfg.AttachSecretBox(good)
+	if err := cfg.Save(); err != nil {
+		t.Fatal(err)
+	}
+	sealedOnce, err := os.ReadFile(path)
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	// A different master key must be rejected, loudly. NewSecretBox derives the
+	// key path from the RUNTIME file's directory, so the wrong key has to live
+	// in a different directory — otherwise it would read the good master.key.
+	otherDir := t.TempDir()
+	if err := os.WriteFile(filepath.Join(otherDir, "master.key"),
+		[]byte(strings.Repeat("cd", 32)), 0600); err != nil {
+		t.Fatal(err)
+	}
+	badBox, err := NewSecretBox(filepath.Join(otherDir, "runtime.json"))
+	if err != nil {
+		t.Fatal(err)
+	}
+	loaded, err := Load(path)
+	if err != nil {
+		t.Fatal(err)
+	}
+	loaded.AttachSecretBox(badBox)
+	if err := loaded.NormalizeSecretsForRun(badBox); err == nil {
+		t.Fatal("expected an error with the wrong master key, got nil")
+	}
+
+	// Crucially: the file must be untouched — no second seal on top.
+	after, err := os.ReadFile(path)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if string(after) != string(sealedOnce) {
+		t.Error("config file was rewritten despite the decrypt failure")
+	}
+	if n := strings.Count(string(after), encPrefix); n != strings.Count(string(sealedOnce), encPrefix) {
+		t.Errorf("sealed count changed %d -> %d (double encryption)", strings.Count(string(sealedOnce), encPrefix), n)
+	}
+}

internal/config/store.go

@@ -138,6 +138,81 @@ func (s *Store) DeletedAdapters() map[string]bool {
 	return out
 }
 
+// SeedTemplates adds preset templates that the store has never seen before.
+// A preset is seeded at most once: its name is recorded in PresetTemplates so
+// a user who edits or deletes it never gets it silently restored on restart.
+func (s *Store) SeedTemplates(presets []SourceTemplate) error {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	seen := make(map[string]bool, len(s.data.PresetTemplates))
+	for _, n := range s.data.PresetTemplates {
+		seen[n] = true
+	}
+	existing := make(map[string]bool, len(s.data.SourceTemplates))
+	for _, t := range s.data.SourceTemplates {
+		existing[t.Name] = true
+	}
+	added := false
+	for _, p := range presets {
+		if p.Name == "" || seen[p.Name] {
+			continue
+		}
+		s.data.PresetTemplates = append(s.data.PresetTemplates, p.Name)
+		added = true
+		if existing[p.Name] {
+			continue // user already has a template by this name: never overwrite
+		}
+		s.data.SourceTemplates = append(s.data.SourceTemplates, p)
+	}
+	if !added {
+		return nil
+	}
+	return s.persistLocked()
+}
+
+// UpsertTemplate adds or replaces a source template and persists.
+func (s *Store) UpsertTemplate(t SourceTemplate) error {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	for i := range s.data.SourceTemplates {
+		if s.data.SourceTemplates[i].Name == t.Name {
+			s.data.SourceTemplates[i] = t
+			return s.persistLocked()
+		}
+	}
+	s.data.SourceTemplates = append(s.data.SourceTemplates, t)
+	return s.persistLocked()
+}
+
+// ListTemplates returns the saved source templates.
+func (s *Store) ListTemplates() []SourceTemplate {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	out := make([]SourceTemplate, len(s.data.SourceTemplates))
+	copy(out, s.data.SourceTemplates)
+	return out
+}
+
+// RemoveTemplate deletes a source template by name and persists.
+func (s *Store) RemoveTemplate(name string) (bool, error) {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	kept := s.data.SourceTemplates[:0]
+	removed := false
+	for _, t := range s.data.SourceTemplates {
+		if t.Name == name {
+			removed = true
+			continue
+		}
+		kept = append(kept, t)
+	}
+	s.data.SourceTemplates = kept
+	if removed {
+		return true, s.persistLocked()
+	}
+	return false, nil
+}
+
 // LoadLegacy returns the full legacy RuntimeConfig from the file (used for
 // one-time migration into config.yaml). Returns nil if file is missing.
 func (s *Store) LoadLegacy() *RuntimeConfig {
@@ -186,15 +261,6 @@ func (s *Store) encryptLocked() {
 	}
 }
 
-func containsString(list []string, s string) bool {
-	for _, x := range list {
-		if x == s {
-			return true
-		}
-	}
-	return false
-}
-
 func removeString(list []string, s string) []string {
 	out := list[:0]
 	for _, x := range list {

internal/core/core.go

@@ -7,6 +7,7 @@ import (
 	"crypto/rand"
 	"encoding/hex"
 	"fmt"
+	"log"
 	"os"
 	"path/filepath"
 	"sort"
@@ -26,13 +27,14 @@ import (
 // request paths look keys up concurrently; the AUTO chain itself is swapped
 // atomically and needs no lock.
 type Core struct {
-	mu        sync.Mutex
-	cfg       *config.Config
-	vm        *lua.VM
-	store     *config.Store
-	scheduler *scheduler.Scheduler
-	registry  *provider.Registry
-	autoChain atomic.Pointer[scheduler.Chain]
+	mu             sync.Mutex
+	cfg            *config.Config
+	vm             *lua.VM
+	store          *config.Store
+	scheduler      *scheduler.Scheduler
+	registry       *provider.Registry
+	autoChain      atomic.Pointer[scheduler.Chain] // chat AUTO chain
+	autoImageChain atomic.Pointer[scheduler.Chain] // image-generation AUTO chain
 }
 
 // New builds the core from a config file plus runtime overlay.
@@ -52,6 +54,28 @@ func NewFromConfig(cfg *config.Config) (*Core, error) {
 		return nil, fmt.Errorf("lua vm: %w", err)
 	}
 	c.store = config.NewStore(cfg.RuntimeFile)
+	// Share one box between the runtime store and config.yaml so a single
+	// master.key seals both files. config.Load left the config holding
+	// ciphertext (if it was sealed); unseal it NOW, before anything reads a
+	// credential.
+	//
+	// ORDER IS LOAD-BEARING. These steps each consume secrets and must run
+	// after the unseal:
+	//   - seedKeys compares cfg.Keys[i].Key against the plaintext
+	//     gateway_keys entries; with ciphertext keys the comparison never
+	//     matched and every restart appended another duplicate admin key
+	//     (production had four copies of the same admin key).
+	//   - rebuildRegistry hands cfg.Sources[i].APIKey to the providers; with
+	//     ciphertext it built every upstream client with "enc:v1:..." as its
+	//     bearer token.
+	// Previously the unseal happened at the END of this function, and things
+	// only appeared to work because seedKeys' Save() unsealed memory as a side
+	// effect. Removing the redundant saves exposed the real ordering bug.
+	cfg.AttachSecretBox(c.store.SecretBox())
+	hadPlaintextSecrets, err := cfg.UnsealSecrets(c.store.SecretBox())
+	if err != nil {
+		return nil, fmt.Errorf("unseal secrets: %w", err)
+	}
 	if err := c.store.Load(); err != nil {
 		return nil, fmt.Errorf("runtime store: %w", err)
 	}
@@ -68,6 +92,17 @@ func NewFromConfig(cfg *config.Config) (*Core, error) {
 	if err := c.rebuildRegistry(); err != nil {
 		return nil, err
 	}
+	// Seed built-in preset templates on first run so the WebUI shows
+	// "从模板创建" for popular providers out of the box.
+	if err := c.seedPresetTemplates(); err != nil {
+		return nil, fmt.Errorf("seed preset templates: %w", err)
+	}
+	// Seal any credential still in the clear in config.yaml. Startup writes
+	// nothing when the file was already sealed (hadPlaintextSecrets is decided
+	// from the on-disk state, before the unseal above).
+	if err := cfg.SealIfNeeded(hadPlaintextSecrets); err != nil {
+		return nil, fmt.Errorf("normalize secrets: %w", err)
+	}
 	return c, nil
 }
 
@@ -161,6 +196,18 @@ func (c *Core) seedKeys() error {
 	return nil
 }
 
+// seedPresetTemplates adds the built-in provider presets to the template
+// store on first run. It only adds templates whose name is not already
+// present, so a user's edits to a preset survive restarts and a deliberately
+// deleted preset does not silently return on the next start (its name is
+// remembered in the store's preset marker list).
+func (c *Core) seedPresetTemplates() error {
+	if len(config.PresetTemplates) == 0 {
+		return nil
+	}
+	return c.store.SeedTemplates(config.PresetTemplates)
+}
+
 // saveConfig writes the current config (including auto rules and keys) back
 // to config.yaml.
 func (c *Core) saveConfig() error {
@@ -181,6 +228,9 @@ func (c *Core) Registry() *provider.Registry { return c.registry }
 // AutoChain returns the current AUTO scheduling chain.
 func (c *Core) AutoChain() *scheduler.Chain { return c.autoChain.Load() }
 
+// AutoImageChain returns the persisted image-generation AUTO chain snapshot.
+func (c *Core) AutoImageChain() *scheduler.Chain { return c.autoImageChain.Load() }
+
 func (c *Core) DefaultModel() string { return c.cfg.DefaultModel }
 
 func (c *Core) GatewayKeys() []string { return c.cfg.GatewayKeys }
@@ -217,10 +267,15 @@ func (c *Core) FindKey(key string) (config.GWKey, bool) {
 }
 
 // CreateKey builds a new random gateway key and persists it to config.yaml.
+// Quotas live on the model scope entries, so a new key's budget is whatever
+// its scopes carry.
 func (c *Core) CreateKey(name, role string, models []config.ModelScope, note string) (config.GWKey, error) {
 	c.mu.Lock()
 	defer c.mu.Unlock()
 	models = cleanScopes(models)
+	if err := ValidateScopeQuotas(models); err != nil {
+		return config.GWKey{}, err
+	}
 	key := make([]byte, 16)
 	if _, err := rand.Read(key); err != nil {
 		return config.GWKey{}, err
@@ -233,9 +288,7 @@ func (c *Core) CreateKey(name, role string, models []config.ModelScope, note str
 		Note:      note,
 		CreatedAt: time.Now().Unix(),
 	}
-	if rec.Role == "" {
-		rec.Role = "user"
-	}
+	rec.Role = config.NormalizeRole(rec.Role)
 	c.cfg.Keys = append(c.cfg.Keys, rec)
 	if err := c.saveConfig(); err != nil {
 		return config.GWKey{}, err
@@ -243,10 +296,16 @@ func (c *Core) CreateKey(name, role string, models []config.ModelScope, note str
 	return rec, nil
 }
 
-// UpdateKey mutates a key's name/role/model scope and persists it.
+// UpdateKey mutates a key's name/role/model scope and persists it. The scope
+// entries carry their own quotas, so replacing the scope replaces the budgets.
 func (c *Core) UpdateKey(key, name, role string, models []config.ModelScope, note string) (config.GWKey, error) {
 	c.mu.Lock()
 	defer c.mu.Unlock()
+	if models != nil {
+		if err := ValidateScopeQuotas(models); err != nil {
+			return config.GWKey{}, err
+		}
+	}
 	for i, k := range c.cfg.Keys {
 		if k.Key == key {
 			if name != "" {
@@ -294,6 +353,15 @@ func (c *Core) AutoRules() []config.ModelScope {
 	return out
 }
 
+// AutoImageRules returns the configured image-generation AUTO chain rules.
+func (c *Core) AutoImageRules() []config.ModelScope {
+	c.mu.Lock()
+	defer c.mu.Unlock()
+	out := make([]config.ModelScope, len(c.cfg.AutoImage))
+	copy(out, c.cfg.AutoImage)
+	return out
+}
+
 // cleanScopes drops empty model entries and normalizes placeholder source
 // names; it returns nil when no entries survive so an empty scope means
 // "unrestricted" (nil) instead of a restrictive-but-empty list — a non-nil
@@ -325,16 +393,35 @@ func (c *Core) SaveAutoRules(entries []config.ModelScope) error {
 		return err
 	}
 	c.buildAutoChain()
-	if ch := c.autoChain.Load(); ch != nil {
-		for _, tn := range ch.Tiers {
-			for _, sl := range tn.Slots {
-				if p := c.registry.ProviderForSlot(sl.Model, sl.Source); p != nil {
-					p.ResetModelCooldown(sl.Model)
-				}
+	c.resetChainCooldowns(c.autoChain.Load())
+	return nil
+}
+
+// SaveAutoImageRules persists the image-generation AUTO chain. Image models
+// are kept as-is (unlike buildAutoChain, which skips them).
+func (c *Core) SaveAutoImageRules(entries []config.ModelScope) error {
+	c.mu.Lock()
+	defer c.mu.Unlock()
+	c.cfg.AutoImage = cleanScopes(entries)
+	if err := c.saveConfig(); err != nil {
+		return err
+	}
+	c.buildAutoImageChain()
+	c.resetChainCooldowns(c.autoImageChain.Load())
+	return nil
+}
+
+func (c *Core) resetChainCooldowns(ch *scheduler.Chain) {
+	if ch == nil {
+		return
+	}
+	for _, tn := range ch.Tiers {
+		for _, sl := range tn.Slots {
+			if p := c.registry.ProviderForSlot(sl.Model, sl.Source); p != nil {
+				p.ResetModelCooldown(sl.Model)
 			}
 		}
 	}
-	return nil
 }
 
 // ResetHealth clears the scheduling backoff state of every provider.
@@ -390,6 +477,11 @@ func (c *Core) mergedSources() []config.Source {
 }
 
 // resolveSourceKey applies api_key_env and decrypts enc:v1: ciphertext.
+// Config values are already unsealed at startup (Config.NormalizeSecretsForRun),
+// so the decrypt branch is the belt-and-braces path for a source that arrived
+// already sealed through another route. A failure there leaves the ciphertext in
+// place, which makes the upstream reject the key loudly rather than sending an
+// empty Authorization header that might look like a config-less source.
 func (c *Core) resolveSourceKey(s config.Source) config.Source {
 	if s.APIKeyEnv != "" {
 		if v := os.Getenv(s.APIKeyEnv); v != "" {
@@ -398,7 +490,11 @@ func (c *Core) resolveSourceKey(s config.Source) config.Source {
 		return s
 	}
 	if box := c.store.SecretBox(); box != nil && strings.HasPrefix(s.APIKey, "enc:v1:") {
-		s.APIKey = box.MustDecrypt(s.APIKey)
+		if v, err := box.Decrypt(s.APIKey); err == nil {
+			s.APIKey = v
+		} else {
+			log.Printf("[core] source %s: api_key decrypt failed: %v", s.Name, err)
+		}
 	}
 	return s
 }
@@ -420,6 +516,7 @@ func (c *Core) rebuildRegistry() error {
 		c.registry.Replace(providers)
 	}
 	c.buildAutoChain()
+	c.buildAutoImageChain()
 	return nil
 }
 
@@ -466,6 +563,47 @@ func (c *Core) buildAutoChain() {
 	c.autoChain.Store(scheduler.BuildChain(sr, prov))
 }
 
+// buildAutoImageChain rebuilds the image-generation AUTO chain snapshot.
+// Unlike buildAutoChain, only image-kind models participate: chat models in
+// the rules are skipped so a stale chat slot can't receive image traffic.
+func (c *Core) buildAutoImageChain() {
+	prov := func(model, source string) scheduler.Provider {
+		p := c.registry.ProviderForSlot(model, source)
+		if p == nil {
+			return nil
+		}
+		if m := p.ModelByID(model); m != nil && m.Kind != "image" {
+			return nil
+		}
+		return p
+	}
+	rules := c.cfg.AutoImage
+	sr := make([]scheduler.Rule, 0, len(rules))
+	for _, e := range rules {
+		model, source := e.Model, e.Source
+		if p := c.registry.ProviderForSlot(e.Model, e.Source); p != nil {
+			if exact := p.ModelIDFold(e.Model); exact != "" {
+				model = exact
+			}
+			if m := p.ModelByID(model); m != nil && m.Kind != "image" {
+				continue // chat-kind models never join the image AUTO chain
+			}
+			if source == "" {
+				source = p.Name()
+			}
+		}
+		sr = append(sr, scheduler.Rule{
+			Model:  model,
+			Source: source,
+			Tier:   e.Tier,
+			Quota:  e.TokenQuota,
+			Period: e.Period,
+			Hours:  e.Hours,
+		})
+	}
+	c.autoImageChain.Store(scheduler.BuildChain(sr, prov))
+}
+
 // AutoSlotState is the UI-facing health snapshot of one AUTO chain slot.
 type AutoSlotState struct {
 	Model         string `json:"model"`
@@ -474,6 +612,16 @@ type AutoSlotState struct {
 	FailCount     int64  `json:"fail_count"`
 	CooldownUntil int64  `json:"cooldown_until"`
 	Cooling       bool   `json:"cooling"`
+	// CooldownFrom is when the current cooldown window opened and ProbeAfter is
+	// its midpoint: from ProbeAfter on, a single probe request is allowed
+	// through so a recovered upstream does not have to sit out the rest of the
+	// window. Probing reports whether such a probe is in flight right now.
+	CooldownFrom int64 `json:"cooldown_from"`
+	ProbeAfter   int64 `json:"probe_after"`
+	Probing      bool  `json:"probing"`
+	// Probeable marks a cooling slot that is past its midpoint, i.e. the next
+	// AUTO request may use it as a probe.
+	Probeable bool `json:"probeable"`
 }
 
 // AutoSlotStates returns per-slot health for every slot of the current chain.
@@ -491,6 +639,7 @@ func (c *Core) AutoSlotStates() []AutoSlotState {
 				continue
 			}
 			pref, fail, until := pp.ModelHealthInfo(sl.Model)
+			from, probeAfter, probing := pp.ModelProbeInfo(sl.Model)
 			out = append(out, AutoSlotState{
 				Model:         sl.Model,
 				Source:        sl.Source,
@@ -498,6 +647,10 @@ func (c *Core) AutoSlotStates() []AutoSlotState {
 				FailCount:     fail,
 				CooldownUntil: until,
 				Cooling:       until > now,
+				CooldownFrom:  from,
+				ProbeAfter:    probeAfter,
+				Probing:       probing,
+				Probeable:     until > now && probeAfter > 0 && now >= probeAfter,
 			})
 		}
 	}
@@ -549,7 +702,7 @@ func (c *Core) AddSource(src config.Source) error {
 	if err := normalizeSource(&src); err != nil {
 		return err
 	}
-	if err := config.UpsertSourceInYAML(c.cfg.Path, src.Name, src); err != nil {
+	if err := config.UpsertSourceInYAML(c.cfg.Path, src.Name, src, c.cfg.SecretBox()); err != nil {
 		return err
 	}
 	// Update in-memory Sources so mergedSources() finds the entry.
@@ -601,6 +754,33 @@ func (c *Core) Sources() []config.Source {
 	return c.mergedSources()
 }
 
+// Templates returns the saved source templates.
+func (c *Core) Templates() []config.SourceTemplate {
+	c.mu.Lock()
+	defer c.mu.Unlock()
+	return c.store.ListTemplates()
+}
+
+// SaveTemplate upserts a source template.
+func (c *Core) SaveTemplate(t config.SourceTemplate) error {
+	c.mu.Lock()
+	defer c.mu.Unlock()
+	if t.Name == "" {
+		return fmt.Errorf("template requires a name")
+	}
+	return c.store.UpsertTemplate(t)
+}
+
+// RemoveTemplate deletes a source template by name.
+func (c *Core) RemoveTemplate(name string) error {
+	c.mu.Lock()
+	defer c.mu.Unlock()
+	if _, err := c.store.RemoveTemplate(name); err != nil {
+		return err
+	}
+	return nil
+}
+
 func normalizeSource(s *config.Source) error {
 	if s.Name == "" || s.BaseURL == "" {
 		return fmt.Errorf("source requires name and base_url")
@@ -622,4 +802,21 @@ func (c *Core) Close() {
 	if c.vm != nil {
 		c.vm.Stop()
 	}
-}
+}
+
+// ValidateScopeQuotas checks every scope entry's caps before they are stored.
+// A typo in a period must be rejected at write time rather than silently
+// becoming a never-resetting budget — the opposite of what was typed.
+func ValidateScopeQuotas(entries []config.ModelScope) error {
+	for _, e := range entries {
+		if err := (config.KeyQuota{
+			TokenQuota: e.TokenQuota,
+			ReqQuota:   e.ReqQuota,
+			Period:     e.Period,
+			Hours:      e.Hours,
+		}).Validate(); err != nil {
+			return fmt.Errorf("model %q: %w", e.Model, err)
+		}
+	}
+	return nil
+}

internal/core/startup_secrets_test.go

@@ -0,0 +1,178 @@
+package core
+
+import (
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+
+	"llmsproxy/internal/config"
+)
+
+// Startup must be idempotent, and every credential consumer must see PLAINTEXT.
+//
+// Two bugs shared one root cause: NewFromConfig unsealed the config at the END,
+// after the steps that read credentials had already run.
+//
+//  1. seedKeys() deduped cfg.Keys[i].Key against the plaintext gateway_keys
+//     entries. With ciphertext keys the comparison never matched, so every
+//     restart appended another copy of the same admin key — production reached
+//     four identical admin keys.
+//  2. rebuildRegistry() handed cfg.Sources[i].APIKey to provider.New, so with
+//     ciphertext it built every upstream client with "enc:v1:..." as its bearer
+//     token and every upstream call would 401.
+//
+// They masked each other: seedKeys' Save() unsealed memory as a side effect, so
+// bug 2 was invisible until the redundant saves were removed. These tests drive
+// the real entry point (core.New -> config.Load off disk), because building a
+// fresh in-memory config per attempt hides both bugs — that is exactly how the
+// first draft of this test failed to reproduce anything.
+
+// writeSealedInstall creates a config on disk the way a real install looks
+// after its first run: credentials already sealed, gateway_keys still plaintext.
+func writeSealedInstall(t *testing.T, dir string) string {
+	t.Helper()
+	path := filepath.Join(dir, "config.yaml")
+	cfg := &config.Config{
+		Path:         path,
+		AdapterDir:   filepath.Join(dir, "adapters"),
+		RuntimeFile:  filepath.Join(dir, "runtime.json"),
+		Listen:       "127.0.0.1:0",
+		DefaultModel: "AUTO",
+		GatewayKeys:  []string{"sk-gw-seed-me"},
+		Sources: []config.Source{{
+			Name: "up", BaseURL: "http://127.0.0.1:1/v1", APIKey: "sk-upstream-secret",
+			Adapter: "openai", Models: []config.Model{{ID: "gpt-4o", Priority: 10}},
+		}},
+	}
+	c, err := NewFromConfig(cfg)
+	if err != nil {
+		t.Fatalf("initial install: %v", err)
+	}
+	c.Close()
+
+	raw, err := os.ReadFile(path)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !strings.Contains(string(raw), "enc:v1:") {
+		t.Fatalf("fixture is wrong: on-disk config holds no sealed credential, "+
+			"so this test cannot detect ciphertext leaking into consumers:\n%s", raw)
+	}
+	return path
+}
+
+// writeSealedInstallNoSeed is writeSealedInstall without gateway_keys, so no
+// key is ever seeded and no startup Save() can unseal the config as a side
+// effect. That isolation is what makes the ciphertext leak observable.
+func writeSealedInstallNoSeed(t *testing.T, dir string) string {
+	t.Helper()
+	path := filepath.Join(dir, "config.yaml")
+	cfg := &config.Config{
+		Path:         path,
+		AdapterDir:   filepath.Join(dir, "adapters"),
+		RuntimeFile:  filepath.Join(dir, "runtime.json"),
+		Listen:       "127.0.0.1:0",
+		DefaultModel: "AUTO",
+		Sources: []config.Source{{
+			Name: "up", BaseURL: "http://127.0.0.1:1/v1", APIKey: "sk-upstream-secret",
+			Adapter: "openai", Models: []config.Model{{ID: "gpt-4o", Priority: 10}},
+		}},
+	}
+	c, err := NewFromConfig(cfg)
+	if err != nil {
+		t.Fatalf("initial install: %v", err)
+	}
+	c.Close()
+	raw, err := os.ReadFile(path)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !strings.Contains(string(raw), "enc:v1:") {
+		t.Fatalf("fixture is wrong: no sealed credential on disk:\n%s", raw)
+	}
+	return path
+}
+
+// countKeysWithSecret loads the config, unseals it, and counts keys whose
+// plaintext equals one of the gateway_keys entries.
+func countKeysWithSecret(t *testing.T, path string) (total, matching int) {
+	t.Helper()
+	cfg, err := config.Load(path)
+	if err != nil {
+		t.Fatal(err)
+	}
+	box, err := config.NewSecretBox(cfg.RuntimeFile)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if _, err := cfg.UnsealSecrets(box); err != nil {
+		t.Fatal(err)
+	}
+	want := map[string]bool{}
+	for _, g := range cfg.GatewayKeys {
+		want[g] = true
+	}
+	for _, k := range cfg.Keys {
+		total++
+		if want[k.Key] {
+			matching++
+		}
+	}
+	return total, matching
+}
+
+// Restarting must not grow the key list. Before the fix this went 2 -> 3 -> 4.
+func TestRestartDoesNotDuplicateSeededKeys(t *testing.T) {
+	dir := t.TempDir()
+	path := writeSealedInstall(t, dir)
+
+	if _, n := countKeysWithSecret(t, path); n != 1 {
+		t.Fatalf("after install: %d keys carry the seeded secret, want 1", n)
+	}
+
+	for i := 2; i <= 4; i++ {
+		c, err := New(path) // production path: reads the file
+		if err != nil {
+			t.Fatalf("restart %d: %v", i, err)
+		}
+		c.Close()
+		total, n := countKeysWithSecret(t, path)
+		if n != 1 {
+			t.Fatalf("restart %d: %d keys carry the seeded secret (total %d), want "+
+				"exactly 1 — seedKeys is comparing ciphertext against plaintext and "+
+				"appending a duplicate on every start", i, n, total)
+		}
+	}
+}
+
+// Every provider must hold the real upstream credential, not its ciphertext.
+//
+// This deliberately runs with gateway_keys EMPTY. The buggy order was masked
+// whenever seedKeys had work to do, because its Save() unsealed memory as a
+// side effect — providers then happened to see plaintext. With no key to seed,
+// no Save runs, and the ciphertext reaches the registry directly. An earlier
+// version of this test kept gateway_keys populated and was insensitive: it
+// passed even with the bug present.
+func TestProvidersGetPlaintextCredentials(t *testing.T) {
+	dir := t.TempDir()
+	path := writeSealedInstallNoSeed(t, dir)
+
+	c, err := New(path)
+	if err != nil {
+		t.Fatalf("restart: %v", err)
+	}
+	defer c.Close()
+
+	for _, p := range c.registry.Providers() {
+		key := p.Config().APIKey
+		if strings.HasPrefix(key, "enc:v1:") {
+			t.Errorf("provider %q holds CIPHERTEXT api_key %q — every upstream call "+
+				"would 401", p.Name(), key)
+		}
+		if key != "sk-upstream-secret" {
+			t.Errorf("provider %q api_key = %q, want the plaintext upstream secret",
+				p.Name(), key)
+		}
+	}
+}

internal/gateway/api.go

@@ -4,6 +4,7 @@ import (
 	"encoding/csv"
 	"encoding/json"
 	"io"
+	"log"
 	"net/http"
 	"strconv"
 	"strings"
@@ -31,7 +32,11 @@ func (g *Gateway) handleAdaptersAPI(w http.ResponseWriter, r *http.Request) {
 			writeJSON(w, http.StatusOK, map[string]interface{}{"adapters": g.core.ListAdapters()})
 			return
 		}
-		writeError(w, http.StatusNotFound, "not_found", "adapter code not exposed; edit in UI")
+		// A GET on a specific adapter is almost always a client that meant to
+		// DELETE it but let fetch default to GET; say so instead of only
+		// reporting that the code is not exposed.
+		writeError(w, http.StatusNotFound, "not_found",
+			"adapter code not exposed; edit in UI (to remove it use DELETE /api/adapters/"+path+")")
 	case http.MethodPost:
 		var p adapterPayload
 		if err := json.NewDecoder(r.Body).Decode(&p); err != nil {
@@ -72,6 +77,7 @@ type sourcePayload struct {
 	Temperature   float64                `json:"temperature"`
 	MaxTokens     int                    `json:"max_tokens"`
 	MaxConcurrent int                    `json:"max_concurrent"`
+	RPM           int                    `json:"rpm"` // optional requests-per-minute cap, 0 = unlimited
 }
 
 func (g *Gateway) handleSourcesAPI(w http.ResponseWriter, r *http.Request) {
@@ -105,6 +111,7 @@ func (g *Gateway) handleSourcesAPI(w http.ResponseWriter, r *http.Request) {
 			Temperature:   p.Temperature,
 			MaxTokens:     p.MaxTokens,
 			MaxConcurrent: p.MaxConcurrent,
+			RPM:           p.RPM,
 		}
 		if err := g.core.AddSource(src); err != nil {
 			writeError(w, http.StatusBadRequest, "source_error", err.Error())
@@ -126,6 +133,62 @@ func (g *Gateway) handleSourcesAPI(w http.ResponseWriter, r *http.Request) {
 	}
 }
 
+// csvHeaders stamps the shared download headers for CSV exports.
+func csvHeaders(w http.ResponseWriter, filename string) {
+	w.Header().Set("Content-Type", "text/csv; charset=utf-8")
+	w.Header().Set("Content-Disposition", "attachment; filename="+filename)
+}
+
+// handleSourceTemplatesAPI manages reusable source templates (a Source minus
+// name and api_key) stored in the runtime file so the WebUI can spin up
+// multiple key-bearing sources from one shared template.
+func (g *Gateway) handleSourceTemplatesAPI(w http.ResponseWriter, r *http.Request) {
+	if reqRole(r.Context()) != "admin" {
+		writeError(w, http.StatusForbidden, "forbidden", "admin role required")
+		return
+	}
+	path := strings.TrimPrefix(r.URL.Path, "/api/source_templates")
+	path = strings.Trim(path, "/")
+	switch r.Method {
+	case http.MethodGet:
+		writeJSON(w, http.StatusOK, map[string]interface{}{"templates": g.core.Templates()})
+	case http.MethodPost:
+		var t config.SourceTemplate
+		if err := json.NewDecoder(r.Body).Decode(&t); err != nil {
+			writeError(w, http.StatusBadRequest, "invalid_request", "invalid json: "+err.Error())
+			return
+		}
+		if err := g.core.SaveTemplate(t); err != nil {
+			writeError(w, http.StatusBadRequest, "template_error", err.Error())
+			return
+		}
+		writeJSON(w, http.StatusOK, map[string]interface{}{"ok": true, "templates": g.core.Templates()})
+	case http.MethodDelete:
+		if path == "" {
+			writeError(w, http.StatusBadRequest, "invalid_request", "template name required")
+			return
+		}
+		if err := g.core.RemoveTemplate(path); err != nil {
+			writeError(w, http.StatusBadRequest, "template_error", err.Error())
+			return
+		}
+		writeJSON(w, http.StatusOK, map[string]interface{}{"ok": true})
+	default:
+		writeError(w, http.StatusMethodNotAllowed, "method_not_allowed", "")
+	}
+}
+
+// exportKey resolves which masked key id an export covers: admins may pass
+// any key filter, user keys are always scoped to themselves (records and
+// aggregates are keyed by the masked keyID form).
+func exportKey(r *http.Request) string {
+	k := r.URL.Query().Get("key")
+	if reqRole(r.Context()) != "admin" {
+		k = keyID(reqKey(r.Context()))
+	}
+	return k
+}
+
 // handleStatsAPI returns per-key / per-model / per-source usage aggregates and
 // the recent request audit trail.
 func (g *Gateway) handleStatsAPI(w http.ResponseWriter, r *http.Request) {
@@ -133,33 +196,35 @@ func (g *Gateway) handleStatsAPI(w http.ResponseWriter, r *http.Request) {
 		writeError(w, http.StatusMethodNotAllowed, "method_not_allowed", "use GET")
 		return
 	}
-	limit := 500
+	// Default to ONE SCREEN of records. The dashboard pages the rest through
+	// /api/stats/records as the user scrolls, so a poll must not serialize the
+	// whole ring buffer. An explicit limit is still honoured (capped by the ring).
+	limit := firstScreenRecords
 	if v := r.URL.Query().Get("limit"); v != "" {
 		if n, err := strconv.Atoi(v); err == nil && n > 0 && n <= 20000 {
 			limit = n
 		}
 	}
-	key := r.URL.Query().Get("key")
-	if reqRole(r.Context()) != "admin" {
-		// user keys may only see their own usage - records and aggregates are
-		// keyed by the masked id (keyID), so filter on that masked form.
-		key = keyID(reqKey(r.Context()))
-	}
+	key := exportKey(r)
 	if r.URL.Query().Get("export") == "csv" {
 		from, _ := strconv.ParseInt(r.URL.Query().Get("from"), 10, 64)
 		to, _ := strconv.ParseInt(r.URL.Query().Get("to"), 10, 64)
 		if to == 0 {
 			to = time.Now().UnixMilli()
 		}
-		w.Header().Set("Content-Type", "text/csv; charset=utf-8")
-		w.Header().Set("Content-Disposition", "attachment; filename=llmsproxy-requests.csv")
+		csvHeaders(w, "llmsproxy-requests.csv")
 		cw := csv.NewWriter(w)
 		names := map[string]string{}
 		for _, k := range g.core.ListKeys() {
 			names[keyID(k.Key)] = k.Name
 		}
-		_ = cw.Write([]string{"time", "key", "key_name", "type", "model", "source", "status", "ok", "prompt_tokens", "completion_tokens", "latency_ms", "error"})
-		for _, rec := range g.stats.AuditRecords(from, to, key) {
+		_ = cw.Write([]string{"time", "key", "key_name", "type", "model", "source", "status", "ok", "prompt_tokens", "completion_tokens", "latency_ms", "first_byte_ms", "cache_hit_tokens", "cache_miss_tokens", "error"})
+		// Stream the window straight to the client instead of materializing it:
+		// an export covering months of audit data must not be bounded by RAM, and
+		// nothing is retained once the response is written.
+		flushEvery := 1000
+		n := 0
+		err := g.stats.StreamAuditRecords(from, to, key, func(rec Req) error {
 			_ = cw.Write([]string{
 				time.UnixMilli(rec.Time).Format(time.RFC3339),
 				rec.Key,
@@ -172,38 +237,48 @@ func (g *Gateway) handleStatsAPI(w http.ResponseWriter, r *http.Request) {
 				strconv.FormatInt(rec.Prompt, 10),
 				strconv.FormatInt(rec.Compl, 10),
 				strconv.FormatInt(rec.LatMs, 10),
+				strconv.FormatInt(rec.FirstByteMs, 10),
+				strconv.FormatInt(rec.CacheHit, 10),
+				strconv.FormatInt(rec.CacheMiss, 10),
 				rec.Err,
 			})
-		}
+			n++
+			if n%flushEvery == 0 {
+				cw.Flush()
+				if err := cw.Error(); err != nil {
+					return err // client went away: stop walking the audit files
+				}
+				if f, ok := w.(http.Flusher); ok {
+					f.Flush()
+				}
+			}
+			return nil
+		})
 		cw.Flush()
+		if err != nil {
+			log.Printf("[gateway] csv export aborted after %d rows: %v", n, err)
+		}
 		return
 	}
 	if r.URL.Query().Get("export") == "keys-csv" {
-		w.Header().Set("Content-Type", "text/csv; charset=utf-8")
-		w.Header().Set("Content-Disposition", "attachment; filename=llmsproxy-keys.csv")
+		csvHeaders(w, "llmsproxy-keys.csv")
 		cw := csv.NewWriter(w)
 		_ = cw.Write([]string{"key", "key_name", "role", "models", "total_requests", "success_requests", "failed_requests", "prompt_tokens", "completion_tokens", "total_tokens", "avg_latency_ms", "max_latency_ms", "created_at"})
-		// Use the same key filtering as the JSON API
-		exportKey := r.URL.Query().Get("key")
-		if reqRole(r.Context()) != "admin" {
-			exportKey = keyID(reqKey(r.Context()))
-		}
+		keyFilter := exportKey(r)
 		// by_key rows are keyed by the masked id (keyID); build a masked-id ->
 		// record map so names/roles/models resolve for the export.
 		byMasked := map[string]config.GWKey{}
 		for _, k := range g.core.ListKeys() {
 			byMasked[keyID(k.Key)] = k
 		}
-		snap := g.stats.Snapshot(0, exportKey)
+		snap := g.stats.Snapshot(0, keyFilter)
 		byKeyRaw, ok := snap["by_key"].([]StatsRow)
 		if !ok {
 			byKeyRaw = []StatsRow{}
 		}
 		for _, row := range byKeyRaw {
 			keyInfo, found := byMasked[row.Name]
-			name := ""
-			role := ""
-			models := ""
+			name, role, models := "", "", ""
 			if found {
 				name = keyInfo.Name
 				role = keyInfo.Role
@@ -213,17 +288,9 @@ func (g *Gateway) handleStatsAPI(w http.ResponseWriter, r *http.Request) {
 				}
 				models = strings.Join(modelNames, ",")
 			}
-			reqs := row.Stat.Reqs
-			success := row.Stat.OK
-			errCount := row.Stat.Err
-			prompt := row.Stat.Prompt
-			compl := row.Stat.Compl
-			tokens := row.Stat.Prompt + row.Stat.Compl
-			latSum := row.Stat.LatSum
-			latMax := row.Stat.LatMax
 			avgLat := int64(0)
-			if reqs > 0 {
-				avgLat = latSum / reqs
+			if row.Stat.Reqs > 0 {
+				avgLat = row.Stat.LatSum / row.Stat.Reqs
 			}
 			createdAt := ""
 			if found && keyInfo.CreatedAt > 0 {
@@ -231,14 +298,14 @@ func (g *Gateway) handleStatsAPI(w http.ResponseWriter, r *http.Request) {
 			}
 			_ = cw.Write([]string{
 				row.Name, name, role, models,
-				strconv.FormatInt(reqs, 10),
-				strconv.FormatInt(success, 10),
-				strconv.FormatInt(errCount, 10),
-				strconv.FormatInt(prompt, 10),
-				strconv.FormatInt(compl, 10),
-				strconv.FormatInt(tokens, 10),
+				strconv.FormatInt(row.Stat.Reqs, 10),
+				strconv.FormatInt(row.Stat.OK, 10),
+				strconv.FormatInt(row.Stat.Err, 10),
+				strconv.FormatInt(row.Stat.Prompt, 10),
+				strconv.FormatInt(row.Stat.Compl, 10),
+				strconv.FormatInt(row.Stat.Prompt+row.Stat.Compl, 10),
 				strconv.FormatInt(avgLat, 10),
-				strconv.FormatInt(latMax, 10),
+				strconv.FormatInt(row.Stat.LatMax, 10),
 				createdAt,
 			})
 		}
@@ -253,3 +320,38 @@ func (g *Gateway) handleStatsAPI(w http.ResponseWriter, r *http.Request) {
 	snap["key_names"] = keyNames
 	writeJSON(w, http.StatusOK, snap)
 }
+
+// handleStatsRecordsAPI pages the request records straight off the audit files.
+// The dashboard loads only its first screen and asks for the next page as the
+// user scrolls, so neither side holds the full history: the server keeps no
+// per-client state (the cursor is the whole state, and it lives in the URL) and
+// closes every file handle before responding.
+//
+//	 GET /api/stats/records?before=<cursor>&limit=100[&key=<masked id>]
+//	-> { records: [...newest first...], next_cursor, has_more, rotated }
+func (g *Gateway) handleStatsRecordsAPI(w http.ResponseWriter, r *http.Request) {
+	if r.Method != http.MethodGet {
+		writeError(w, http.StatusMethodNotAllowed, "method_not_allowed", "use GET")
+		return
+	}
+	limit := 100
+	if v := r.URL.Query().Get("limit"); v != "" {
+		if n, err := strconv.Atoi(v); err == nil && n > 0 && n <= 1000 {
+			limit = n
+		}
+	}
+	// exportKey pins non-admin callers to their own key, so a user key can never
+	// page another key's records.
+	page := g.stats.AuditPage(r.URL.Query().Get("before"), limit, exportKey(r))
+	keyNames := map[string]string{}
+	for _, k := range g.core.ListKeys() {
+		keyNames[keyID(k.Key)] = k.Name
+	}
+	writeJSON(w, http.StatusOK, map[string]interface{}{
+		"records":     page.Records,
+		"next_cursor": page.Next,
+		"has_more":    page.HasMore,
+		"rotated":     page.Rotated,
+		"key_names":   keyNames,
+	})
+}

internal/gateway/apiv1.go

@@ -0,0 +1,417 @@
+package gateway
+
+// Agent-facing management API.
+//
+// The Web UI has always driven the gateway through /api/*, so the management
+// surface already exists. What it lacked was anything an agent could *rely* on:
+// no way to discover the contract, no single call that answers "what is the
+// current state", and per-endpoint response shapes that make scripting brittle.
+//
+// This file adds a versioned, self-describing facade under /api/v1 without
+// touching the existing endpoints, so the Web UI keeps working unchanged while
+// agents get a stable contract:
+//
+//	GET    /api/v1                      — machine-readable index of every endpoint
+//	GET    /api/v1/overview             — one call: sources + auto chain + keys + health
+//	GET    /api/v1/sources              — sources, credentials masked
+//	GET    /api/v1/sources/{name}       — one source
+//	GET    /api/v1/auto                 — scheduling chain + live slot states
+//	GET    /api/v1/keys                 — key metadata (never the secret)
+//	GET    /api/v1/models               — every model id the gateway can route
+//	GET    /api/v1/health               — per-source health, no auth needed beyond the key
+//
+// Auth is the same gateway key as everywhere else (Authorization: Bearer, or
+// ?api_key=, or the gw_key cookie). Read endpoints accept any role; writes
+// still require admin, enforced by the same middleware the UI goes through.
+
+import (
+	"errors"
+	"net/http"
+	"net/url"
+	"sort"
+	"strings"
+
+	"llmsproxy/internal/config"
+)
+
+// apiV1Routes dispatches /api/v1/*. Kept separate from routes() so the original
+// switch stays readable and the versioned surface can grow on its own.
+func (g *Gateway) apiV1Routes(w http.ResponseWriter, r *http.Request) {
+	path := strings.TrimPrefix(r.URL.Path, "/api/v1")
+	path = strings.Trim(path, "/")
+
+	switch {
+	case path == "":
+		g.apiV1Index(w, r)
+	case path == "overview":
+		g.apiV1Overview(w, r)
+	case path == "health":
+		g.apiV1Health(w, r)
+	case path == "models":
+		g.apiV1Models(w, r)
+	case path == "sources":
+		if r.Method != http.MethodGet {
+			writeError(w, http.StatusMethodNotAllowed, "method_not_allowed", "use GET")
+			return
+		}
+		writeJSON(w, http.StatusOK, map[string]interface{}{"sources": maskSources(g.core.Sources())})
+	case strings.HasPrefix(path, "sources/"):
+		name, err := decodePathSegment(strings.TrimPrefix(path, "sources/"))
+		if err != nil {
+			writeError(w, http.StatusBadRequest, "invalid_request", err.Error())
+			return
+		}
+		if r.Method != http.MethodGet {
+			writeError(w, http.StatusMethodNotAllowed, "method_not_allowed",
+				"use PUT/DELETE on /api/sources/{name} to modify a source")
+			return
+		}
+		for _, s := range g.core.Sources() {
+			if s.Name == name {
+				// ?reveal=credentials is admin-only and is what the Web UI's
+				// edit dialog uses: a form that round-trips a source must be
+				// able to show the current key, otherwise saving an unrelated
+				// field would blank it. Everything else stays masked.
+				if r.URL.Query().Get("reveal") == "credentials" {
+					if reqRole(r.Context()) != "admin" {
+						writeError(w, http.StatusForbidden, "forbidden",
+							"admin role required to reveal credentials")
+						return
+					}
+					writeJSON(w, http.StatusOK, map[string]interface{}{"source": s})
+					return
+				}
+				writeJSON(w, http.StatusOK, map[string]interface{}{"source": maskSource(s)})
+				return
+			}
+		}
+		writeError(w, http.StatusNotFound, "not_found", "no such source: "+name)
+	case path == "auto":
+		if r.Method != http.MethodGet {
+			writeError(w, http.StatusMethodNotAllowed, "method_not_allowed", "use GET")
+			return
+		}
+		writeJSON(w, http.StatusOK, map[string]interface{}{
+			"rules":       g.core.AutoRules(),
+			"image_rules": g.core.AutoImageRules(),
+			"states":      g.core.AutoSlotStates(),
+		})
+	case path == "keys":
+		if reqRole(r.Context()) != "admin" {
+			writeError(w, http.StatusForbidden, "forbidden", "admin role required")
+			return
+		}
+		if r.Method != http.MethodGet {
+			writeError(w, http.StatusMethodNotAllowed, "method_not_allowed",
+				"use POST/PUT/DELETE on /api/keys to manage keys")
+			return
+		}
+		keys := g.core.ListKeys()
+		out := make([]map[string]interface{}, 0, len(keys))
+		for _, k := range keys {
+			out = append(out, map[string]interface{}{
+				"name":       k.Name,
+				"role":       k.Role,
+				"models":     k.Models,
+				"note":       k.Note,
+				"created_at": k.CreatedAt,
+				"seed":       k.Seed,
+				// Quotas live on the scope entries (k.Models), echoed above;
+				// there is deliberately no key-wide total.
+				// The secret itself is never echoed. An operator that needs it
+				// already has it from creation time or from config.yaml.
+				"key_prefix": maskKey(k.Key),
+			})
+		}
+		writeJSON(w, http.StatusOK, map[string]interface{}{"keys": out})
+	default:
+		g.apiV1Index(w, r)
+	}
+}
+
+// apiV1Index is the discovery document: it names every endpoint with its
+// method, auth requirement and purpose, so an agent does not have to scrape the
+// HTML to find out what it can call.
+func (g *Gateway) apiV1Index(w http.ResponseWriter, r *http.Request) {
+	type ep struct {
+		Method      string `json:"method"`
+		Path        string `json:"path"`
+		Auth        string `json:"auth"`
+		Summary     string `json:"summary"`
+		WriteEffect string `json:"writes,omitempty"`
+	}
+	index := map[string]interface{}{
+		"api_version": "v1",
+		"description": "llmsproxy management API. Authenticate with a gateway key: " +
+			"'Authorization: Bearer <key>', '?api_key=<key>', or the gw_key cookie. " +
+			"Read endpoints accept any role; mutations require an admin key.",
+		"discovery": []ep{
+			{Method: "GET", Path: "/api/v1", Auth: "any", Summary: "this document"},
+			{Method: "GET", Path: "/api/v1/overview", Auth: "any",
+				Summary: "current state in one call: sources, auto chain, key count, source health"},
+			{Method: "GET", Path: "/api/v1/health", Auth: "any",
+				Summary: "per-source health snapshot only"},
+			{Method: "GET", Path: "/api/v1/models", Auth: "any",
+				Summary: "every model id the gateway can route, grouped by owning source"},
+			{Method: "GET", Path: "/api/v1/sources", Auth: "any",
+				Summary: "all sources with credentials masked"},
+			{Method: "GET", Path: "/api/v1/sources/{name}", Auth: "any", Summary: "one source"},
+			{Method: "GET", Path: "/api/v1/auto", Auth: "any",
+				Summary: "AUTO scheduling chain and live per-slot state"},
+			{Method: "GET", Path: "/api/v1/keys", Auth: "admin",
+				Summary: "gateway key metadata; secrets are never returned"},
+
+			{Method: "GET", Path: "/api/sources", Auth: "admin", Summary: "raw source list (includes api_key)"},
+			{Method: "POST", Path: "/api/sources", Auth: "admin", Summary: "add or replace a source",
+				WriteEffect: "writes config.yaml (api_key sealed at rest)"},
+			{Method: "PUT", Path: "/api/sources/{name}", Auth: "admin", Summary: "update one source",
+				WriteEffect: "writes config.yaml"},
+			{Method: "DELETE", Path: "/api/sources/{name}", Auth: "admin", Summary: "delete a source",
+				WriteEffect: "writes config.yaml"},
+
+			{Method: "GET", Path: "/api/auto", Auth: "any", Summary: "scheduling chain (readable by any role)"},
+			{Method: "PUT", Path: "/api/auto", Auth: "admin", Summary: "replace the scheduling chain",
+				WriteEffect: "writes config.yaml (the AUTO chain lives here)"},
+
+			{Method: "GET", Path: "/api/keys", Auth: "admin", Summary: "gateway keys"},
+			{Method: "POST", Path: "/api/keys", Auth: "admin", Summary: "create a gateway key",
+				WriteEffect: "writes config.yaml"},
+			{Method: "DELETE", Path: "/api/keys/{name}", Auth: "admin", Summary: "delete a gateway key",
+				WriteEffect: "writes config.yaml"},
+
+			{Method: "GET", Path: "/api/status", Auth: "any", Summary: "per-source health detail"},
+			{Method: "GET", Path: "/api/stats", Auth: "any", Summary: "usage aggregates"},
+			{Method: "GET", Path: "/api/stats/records", Auth: "any", Summary: "paged request records"},
+			{Method: "GET", Path: "/api/adapters", Auth: "admin", Summary: "installed Lua adapters"},
+			{Method: "GET", Path: "/api/source_templates", Auth: "admin", Summary: "source templates"},
+			{Method: "POST", Path: "/v1/chat/completions", Auth: "any", Summary: "OpenAI-compatible inference"},
+		},
+		"conventions": map[string]interface{}{
+			"errors":       "{ \"error\": { \"type\": <code>, \"message\": <text> } }",
+			"path_escape":  "URL-encode source and key names; {name} is a single path segment",
+			"idempotency":  "POST /api/sources and PUT /api/sources/{name} both upsert by name",
+			"config_truth": "all configuration lives in config.yaml; API writes are persisted immediately",
+			"credentials": "credentials are masked by default. GET /api/v1/sources/{name}?reveal=credentials " +
+				"returns them in the clear and is admin-only — the Web UI edit dialog uses it, because a form " +
+				"that round-trips a source must show the current key or saving another field would blank it.",
+		},
+	}
+	writeJSON(w, http.StatusOK, index)
+}
+
+// apiV1Overview is the "what is the current state" call an agent makes first.
+// One round trip instead of five.
+func (g *Gateway) apiV1Overview(w http.ResponseWriter, r *http.Request) {
+	if r.Method != http.MethodGet {
+		writeError(w, http.StatusMethodNotAllowed, "method_not_allowed", "use GET")
+		return
+	}
+	sources := g.core.Sources()
+	names := make([]string, 0, len(sources))
+	modelCount := 0
+	for _, s := range sources {
+		names = append(names, s.Name)
+		modelCount += len(s.Models)
+	}
+	keys := g.core.ListKeys()
+	adminCount, userCount := 0, 0
+	for _, k := range keys {
+		if k.Role == "admin" {
+			adminCount++
+		} else {
+			userCount++
+		}
+	}
+	writeJSON(w, http.StatusOK, map[string]interface{}{
+		"api_version": "v1",
+		"summary": map[string]interface{}{
+			"source_count":      len(sources),
+			"model_count":       modelCount,
+			"gateway_key_count": len(keys),
+			"admin_key_count":   adminCount,
+			"user_key_count":    userCount,
+			"auto_slots":        len(g.core.AutoRules()),
+		},
+		"sources":    maskSources(sources),
+		"auto":       g.core.AutoRules(),
+		"auto_image": g.core.AutoImageRules(),
+		"health":     g.sourceHealthBrief(),
+		"caller": map[string]interface{}{
+			"role": reqRole(r.Context()),
+			"key":  maskKey(reqKey(r.Context())),
+		},
+	})
+}
+
+func (g *Gateway) apiV1Health(w http.ResponseWriter, r *http.Request) {
+	if r.Method != http.MethodGet {
+		writeError(w, http.StatusMethodNotAllowed, "method_not_allowed", "use GET")
+		return
+	}
+	writeJSON(w, http.StatusOK, map[string]interface{}{"sources": g.sourceHealthBrief()})
+}
+
+// apiV1Models lists routable model ids per source, which is what an agent needs
+// to build a valid request — /v1/models flattens them and hides the owner.
+func (g *Gateway) apiV1Models(w http.ResponseWriter, r *http.Request) {
+	if r.Method != http.MethodGet {
+		writeError(w, http.StatusMethodNotAllowed, "method_not_allowed", "use GET")
+		return
+	}
+	bySource := map[string][]string{}
+	var all []string
+	for _, s := range g.core.Sources() {
+		for _, m := range s.Models {
+			bySource[s.Name] = append(bySource[s.Name], m.ID)
+			all = append(all, m.ID)
+		}
+	}
+	sort.Strings(all)
+	all = dedupeStrings(all)
+	writeJSON(w, http.StatusOK, map[string]interface{}{
+		"count":     len(all),
+		"models":    all,
+		"by_source": bySource,
+		"note":      "request a model as \"<source>:<model>\" to pin one source, or the bare id to let the gateway choose",
+	})
+}
+
+// ---- helpers ----
+
+// maskSources returns sources with credentials replaced by a presence marker.
+// The UI still uses /api/sources (which returns real values, because its edit
+// form round-trips them); this is the safe view for programmatic callers.
+func maskSources(srcs []config.Source) []map[string]interface{} {
+	out := make([]map[string]interface{}, 0, len(srcs))
+	for _, s := range srcs {
+		out = append(out, maskSource(s))
+	}
+	return out
+}
+
+func maskSource(s config.Source) map[string]interface{} {
+	return map[string]interface{}{
+		"name":           s.Name,
+		"base_url":       s.BaseURL,
+		"adapter":        s.Adapter,
+		"endpoint":       s.Endpoint,
+		"image_endpoint": s.ImageEndpoint,
+		"api_key":        maskKey(s.APIKey),
+		"api_key_set":    s.APIKey != "",
+		"models":         s.Models,
+		"headers":        maskHeaders(s.Headers),
+		"proxy_url":      s.ProxyURL,
+		"meta":           s.Meta,
+		"temperature":    s.Temperature,
+		"max_tokens":     s.MaxTokens,
+		"max_concurrent": s.MaxConcurrent,
+		"rpm":            s.RPM,
+	}
+}
+
+func maskKey(k string) string {
+	if k == "" {
+		return ""
+	}
+	if strings.HasPrefix(k, "enc:v1:") {
+		return "(sealed)"
+	}
+	if len(k) <= 10 {
+		return k[:2] + "…"
+	}
+	return k[:6] + "…" + k[len(k)-4:]
+}
+
+func maskHeaders(h map[string]string) map[string]string {
+	if h == nil {
+		return nil
+	}
+	out := make(map[string]string, len(h))
+	for k, v := range h {
+		out[k] = maskKey(v)
+	}
+	return out
+}
+
+// sourceHealthBrief reuses the same registry status the UI's /api/status shows,
+// so the agent view and the UI view cannot drift apart. Recent-traffic counters
+// come from the same stats window, because a source that is actually serving
+// traffic must never look down just because a probe was rate-limited.
+func (g *Gateway) sourceHealthBrief() []map[string]interface{} {
+	sts := g.core.Registry().Status()
+	recent := g.stats.SourceRecent(300)
+	avgs := g.stats.SourceAverages(300)
+	for i := range sts {
+		if v, ok := recent[sts[i].Name]; ok {
+			sts[i].RecentOK = v[0]
+			sts[i].RecentErr = v[1]
+		}
+		if a, ok := avgs[sts[i].Name]; ok {
+			sts[i].AvgFirstByteMs = a.AvgFirstByteMs
+			sts[i].AvgTokPerS = a.AvgTokPerS
+		}
+	}
+	out := make([]map[string]interface{}, 0, len(sts))
+	for _, s := range sts {
+		row := map[string]interface{}{
+			"name":           s.Name,
+			"adapter":        s.Adapter,
+			"healthy":        s.Healthy,
+			"available":      s.Available,
+			"live_available": s.LiveAvailable,
+			"model_count":    len(s.Models),
+		}
+		if s.LastError != "" {
+			row["last_error"] = s.LastError
+		}
+		if s.LastChecked > 0 {
+			row["last_checked"] = s.LastChecked
+		}
+		if s.RecentOK > 0 || s.RecentErr > 0 {
+			row["recent_ok"] = s.RecentOK
+			row["recent_err"] = s.RecentErr
+		}
+		if s.AvgFirstByteMs > 0 {
+			row["avg_first_byte_ms"] = s.AvgFirstByteMs
+		}
+		if s.AvgTokPerS > 0 {
+			row["avg_tok_per_s"] = s.AvgTokPerS
+		}
+		if s.FailCount > 0 {
+			row["fail_count"] = s.FailCount
+		}
+		if s.Permanent {
+			row["permanent"] = true
+		}
+		out = append(out, row)
+	}
+	return out
+}
+
+// decodePathSegment URL-decodes one path segment and rejects an empty result,
+// so a name containing a slash cannot be silently mis-resolved.
+func decodePathSegment(seg string) (string, error) {
+	s, err := url.PathUnescape(seg)
+	if err != nil {
+		return "", err
+	}
+	if s == "" {
+		return "", errEmptySegment
+	}
+	return s, nil
+}
+
+func dedupeStrings(in []string) []string {
+	out := in[:0]
+	var last string
+	for i, s := range in {
+		if i == 0 || s != last {
+			out = append(out, s)
+		}
+		last = s
+	}
+	return out
+}
+
+// errEmptySegment marks a path like /api/v1/sources/ with no name after it.
+var errEmptySegment = errors.New("empty path segment")

internal/gateway/apiv1_test.go

@@ -0,0 +1,299 @@
+package gateway
+
+import (
+	"encoding/json"
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"testing"
+
+	"llmsproxy/internal/config"
+)
+
+// v1Gateway builds a gateway with one source and two keys, and returns the
+// secrets so a test can assert they never appear in a response.
+func v1Gateway(t *testing.T) (*Gateway, string, string) {
+	t.Helper()
+	g := newTestGateway(t, config.Source{
+		Name:    "up",
+		BaseURL: "http://up.example/v1",
+		APIKey:  "sk-up-secret",
+		Adapter: "openai",
+		Models:  []config.Model{{ID: "m1", Kind: "chat"}},
+	})
+	// The shared helper only seeds one admin key; add a user key through the
+	// core so role-gated endpoints have something to reject.
+	if _, err := g.core.CreateKey("agent", "user", nil, "test agent key"); err != nil {
+		t.Fatalf("CreateKey: %v", err)
+	}
+	rec := doReq(t, g, http.MethodGet, "/api/keys", "")
+	var doc struct {
+		Keys []config.GWKey `json:"keys"`
+	}
+	if err := json.Unmarshal(rec.Body.Bytes(), &doc); err != nil {
+		t.Fatalf("decode keys: %v", err)
+	}
+	userKey := ""
+	for _, k := range doc.Keys {
+		if k.Role == "user" {
+			userKey = k.Key
+		}
+	}
+	if userKey == "" {
+		t.Fatal("no user key was created")
+	}
+	return g, "sk-test", userKey
+}
+
+func decodeJSON(t *testing.T, body string, v interface{}) {
+	t.Helper()
+	if err := json.Unmarshal([]byte(body), v); err != nil {
+		t.Fatalf("decode %s: %v", body, err)
+	}
+}
+
+func TestAPIV1IndexIsDiscoverable(t *testing.T) {
+	g, _, _ := v1Gateway(t)
+	rec := doReq(t, g, http.MethodGet, "/api/v1", "")
+	if rec.Code != http.StatusOK {
+		t.Fatalf("status = %d, want 200: %s", rec.Code, rec.Body.String())
+	}
+	var doc struct {
+		APIVersion  string `json:"api_version"`
+		Description string `json:"description"`
+		Discovery   []struct {
+			Method  string `json:"method"`
+			Path    string `json:"path"`
+			Auth    string `json:"auth"`
+			Summary string `json:"summary"`
+		} `json:"discovery"`
+		Conventions map[string]interface{} `json:"conventions"`
+	}
+	decodeJSON(t, rec.Body.String(), &doc)
+	if doc.APIVersion != "v1" {
+		t.Errorf("api_version = %q", doc.APIVersion)
+	}
+	if doc.Description == "" {
+		t.Error("index should carry a usage description")
+	}
+	if len(doc.Discovery) == 0 {
+		t.Fatal("discovery list is empty")
+	}
+	want := map[string]bool{
+		"/api/v1/overview": false, "/api/v1/sources": false,
+		"/api/v1/auto": false, "/api/v1/models": false,
+		"/api/v1/health": false, "/api/v1/keys": false,
+		"/api/sources": false, "/api/auto": false, "/v1/chat/completions": false,
+	}
+	for _, e := range doc.Discovery {
+		if _, ok := want[e.Path]; ok {
+			want[e.Path] = true
+		}
+		if e.Method == "" || e.Path == "" || e.Auth == "" || e.Summary == "" {
+			t.Errorf("incomplete discovery entry: %+v", e)
+		}
+	}
+	for p, found := range want {
+		if !found {
+			t.Errorf("discovery is missing %s", p)
+		}
+	}
+	if doc.Conventions["errors"] == nil {
+		t.Error("conventions should document the error shape")
+	}
+}
+
+func TestAPIV1OverviewSummarisesState(t *testing.T) {
+	g, _, _ := v1Gateway(t)
+	rec := doReq(t, g, http.MethodGet, "/api/v1/overview", "")
+	if rec.Code != http.StatusOK {
+		t.Fatalf("status = %d: %s", rec.Code, rec.Body.String())
+	}
+	var doc struct {
+		Summary struct {
+			SourceCount     int `json:"source_count"`
+			ModelCount      int `json:"model_count"`
+			GatewayKeyCount int `json:"gateway_key_count"`
+		} `json:"summary"`
+		Sources []map[string]interface{} `json:"sources"`
+		Auto    []map[string]interface{} `json:"auto"`
+		Health  []map[string]interface{} `json:"health"`
+		Caller  struct {
+			Role string `json:"role"`
+		} `json:"caller"`
+	}
+	decodeJSON(t, rec.Body.String(), &doc)
+	if doc.Summary.SourceCount != 1 {
+		t.Errorf("source_count = %d, want 1", doc.Summary.SourceCount)
+	}
+	if doc.Summary.ModelCount != 1 {
+		t.Errorf("model_count = %d, want 1", doc.Summary.ModelCount)
+	}
+	if doc.Summary.GatewayKeyCount < 2 {
+		t.Errorf("gateway_key_count = %d, want >= 2", doc.Summary.GatewayKeyCount)
+	}
+	if len(doc.Sources) != 1 || doc.Sources[0]["name"] != "up" {
+		t.Errorf("sources = %+v", doc.Sources)
+	}
+	if doc.Caller.Role != "admin" {
+		t.Errorf("caller.role = %q, want admin", doc.Caller.Role)
+	}
+}
+
+func TestAPIV1NeverEchoesSecrets(t *testing.T) {
+	g, admin, user := v1Gateway(t)
+	for _, path := range []string{"/api/v1", "/api/v1/overview", "/api/v1/sources", "/api/v1/sources/up", "/api/v1/keys", "/api/v1/models", "/api/v1/health", "/api/v1/auto"} {
+		rec := doReq(t, g, http.MethodGet, path, "")
+		if rec.Code != http.StatusOK {
+			t.Fatalf("%s status = %d: %s", path, rec.Code, rec.Body.String())
+		}
+		if strings.Contains(rec.Body.String(), "sk-up-secret") {
+			t.Errorf("%s leaked the source api_key", path)
+		}
+		if admin != "" && strings.Contains(rec.Body.String(), admin) {
+			t.Errorf("%s leaked the admin gateway key", path)
+		}
+		if strings.Contains(rec.Body.String(), user) {
+			t.Errorf("%s leaked the user gateway key", path)
+		}
+	}
+	// Masking must still be useful: it says a key is configured.
+	rec := doReq(t, g, http.MethodGet, "/api/v1/sources/up", "")
+	if !strings.Contains(rec.Body.String(), `"api_key_set":true`) {
+		t.Error("masked source should report that a key is configured")
+	}
+}
+
+func TestAPIV1KeysRequiresAdminAndMasks(t *testing.T) {
+	g, _, user := v1Gateway(t)
+	// Authenticate as the user key by swapping the shared helper's header.
+	req := newAuthedRequest(t, http.MethodGet, "/api/v1/keys", user)
+	rec := serveViaHandler(t, g, req)
+	if rec.Code != http.StatusForbidden {
+		t.Errorf("user role on /api/v1/keys = %d, want 403: %s", rec.Code, rec.Body.String())
+	}
+}
+
+func TestAPIV1SourcesNotFound(t *testing.T) {
+	g, _, _ := v1Gateway(t)
+	rec := doReq(t, g, http.MethodGet, "/api/v1/sources/nope", "")
+	if rec.Code != http.StatusNotFound {
+		t.Errorf("status = %d, want 404: %s", rec.Code, rec.Body.String())
+	}
+}
+
+func TestAPIV1RejectsWrites(t *testing.T) {
+	g, _, _ := v1Gateway(t)
+	for _, m := range []string{http.MethodPost, http.MethodPut, http.MethodDelete, http.MethodPatch} {
+		rec := doReq(t, g, m, "/api/v1/sources", `{}`)
+		if rec.Code == http.StatusOK {
+			t.Errorf("%s /api/v1/sources returned 200; the v1 read facade must not mutate", m)
+		}
+	}
+	// And the source must be untouched.
+	rec := doReq(t, g, http.MethodGet, "/api/v1/sources", "")
+	if !strings.Contains(rec.Body.String(), `"name":"up"`) {
+		t.Error("the source disappeared after a rejected write")
+	}
+}
+
+func TestAPIV1ModelsGroupsBySource(t *testing.T) {
+	g, _, _ := v1Gateway(t)
+	rec := doReq(t, g, http.MethodGet, "/api/v1/models", "")
+	if rec.Code != http.StatusOK {
+		t.Fatalf("status = %d", rec.Code)
+	}
+	var doc struct {
+		Count    int                 `json:"count"`
+		Models   []string            `json:"models"`
+		BySource map[string][]string `json:"by_source"`
+	}
+	decodeJSON(t, rec.Body.String(), &doc)
+	if doc.Count != 1 || doc.Models[0] != "m1" {
+		t.Errorf("models = %+v", doc.Models)
+	}
+	if len(doc.BySource["up"]) != 1 || doc.BySource["up"][0] != "m1" {
+		t.Errorf("by_source = %+v", doc.BySource)
+	}
+}
+
+func TestAPIV1RequiresAuthentication(t *testing.T) {
+	g, _, _ := v1Gateway(t)
+	for _, path := range []string{"/api/v1", "/api/v1/overview", "/api/v1/sources", "/api/v1/health", "/api/v1/models"} {
+		req := newRequest(http.MethodGet, path, "")
+		rec := serveViaHandler(t, g, req)
+		if rec.Code != http.StatusUnauthorized {
+			t.Errorf("%s without a key = %d, want 401", path, rec.Code)
+		}
+	}
+}
+
+func TestMaskKeyHidesTheSecret(t *testing.T) {
+	cases := map[string]string{
+		"":                   "",
+		"ab":                 "ab…",
+		"sk-gw-abcdefghijkl": "sk-gw-…ijkl",
+		"enc:v1:AAAA":        "(sealed)",
+	}
+	for in, want := range cases {
+		if got := maskKey(in); got != want {
+			t.Errorf("maskKey(%q) = %q, want %q", in, got, want)
+		}
+	}
+	// The masked form must not reveal the middle of the secret.
+	full := maskKey("sk-gw-1234567890abcdef")
+	if strings.Contains(full, "4567890abcdef") {
+		t.Errorf("maskKey leaked the middle: %q", full)
+	}
+}
+
+// newRequest builds an unauthenticated request.
+func newRequest(method, path, body string) *http.Request {
+	req, _ := http.NewRequest(method, path, strings.NewReader(body))
+	if body != "" {
+		req.Header.Set("Content-Type", "application/json")
+	}
+	return req
+}
+
+// newAuthedRequest builds a request carrying the given bearer key.
+func newAuthedRequest(t *testing.T, method, path, key string) *http.Request {
+	t.Helper()
+	req := newRequest(method, path, "")
+	req.Header.Set("Authorization", "Bearer "+key)
+	return req
+}
+
+// serveViaHandler pushes a request through the full handler chain (auth included).
+func serveViaHandler(t *testing.T, g *Gateway, req *http.Request) *httptest.ResponseRecorder {
+	t.Helper()
+	rec := httptest.NewRecorder()
+	g.Handler().ServeHTTP(rec, req)
+	return rec
+}
+
+// TestAPIV1RevealRequiresAdmin: the only way to read a key in the clear is the
+// explicit admin-only reveal. Without the role check this endpoint would hand
+// every source credential to any valid (even user-scoped) key.
+func TestAPIV1RevealRequiresAdmin(t *testing.T) {
+	g, _, user := v1Gateway(t)
+	rec := serveViaHandler(t, g, newAuthedRequest(t, http.MethodGet, "/api/v1/sources/up?reveal=credentials", user))
+	if rec.Code != http.StatusForbidden {
+		t.Errorf("user reveal = %d, want 403: %s", rec.Code, rec.Body.String())
+	}
+	if strings.Contains(rec.Body.String(), "sk-up-secret") {
+		t.Error("a forbidden reveal still leaked the key")
+	}
+	rec = doReq(t, g, http.MethodGet, "/api/v1/sources/up?reveal=credentials", "")
+	if rec.Code != http.StatusOK {
+		t.Fatalf("admin reveal = %d: %s", rec.Code, rec.Body.String())
+	}
+	if !strings.Contains(rec.Body.String(), "sk-up-secret") {
+		t.Error("admin reveal should return the key in the clear (the edit form needs it)")
+	}
+	// The masked default view must still hide it on the same path.
+	rec = doReq(t, g, http.MethodGet, "/api/v1/sources/up", "")
+	if strings.Contains(rec.Body.String(), "sk-up-secret") {
+		t.Error("the default view leaked the key")
+	}
+}

internal/gateway/chat.go

@@ -5,7 +5,10 @@ import (
 	"encoding/json"
 	"errors"
 	"fmt"
+	"log"
 	"net/http"
+	"regexp"
+	"strconv"
 	"strings"
 	"sync/atomic"
 	"time"
@@ -27,6 +30,9 @@ type chatRequest struct {
 	ToolChoice      interface{}            `json:"tool_choice,omitempty"`
 	DisableThinking bool                   `json:"disable_thinking"`
 	ExtraBody       map[string]interface{} `json:"extra_body,omitempty"`
+
+	// PromptCacheKey 是 OpenAI 原生的缓存键；部分客户端用它携带会话标识。
+	PromptCacheKey string `json:"prompt_cache_key,omitempty"`
 }
 
 // ChatCompletion is the non-streaming OpenAI response object.
@@ -37,6 +43,10 @@ type ChatCompletion struct {
 	Model   string            `json:"model"`
 	Choices []ChatChoice      `json:"choices"`
 	Usage   *types.TokenUsage `json:"usage,omitempty"`
+	// Cost is the upstream-reported charge for this request, passed through
+	// verbatim (OpenCode reports it as a decimal string, always "0" on the
+	// flat-rate Go subscription). Absent when the upstream reports nothing.
+	Cost string `json:"cost,omitempty"`
 }
 
 type ChatChoice struct {
@@ -58,6 +68,13 @@ type ChatChunk struct {
 	Created int64         `json:"created"`
 	Model   string        `json:"model"`
 	Choices []ChunkChoice `json:"choices"`
+	// Usage is sent in the final chunk of a stream (empty choices) so
+	// OpenAI-compatible clients can read token usage.
+	Usage *types.TokenUsage `json:"usage,omitempty"`
+	// Cost is the upstream-reported charge for this request, passed through
+	// verbatim (OpenCode reports it as a decimal string). Emitted on the
+	// terminal chunk, mirroring OpenCode's own {"choices":[],"cost":"0"}.
+	Cost string `json:"cost,omitempty"`
 }
 
 type ChunkChoice struct {
@@ -111,13 +128,15 @@ func (g *Gateway) resolveCands(ctx context.Context, req *chatRequest) ([]*provid
 }
 
 // filterCandsByModels keeps only providers exposing at least one model of the
-// scope (used for user keys with a restricted model scope). An "AUTO" scope
-// entry means the key is allowed to use any model. Scope entries with a
-// Source pinned to a specific upstream narrow the candidates to that source
+// scope (used for user keys with a restricted model scope). Scope entries with
+// a Source pinned to a specific upstream narrow the candidates to that source
 // for the matching model.
+//
+// An "AUTO" scope entry only allows the AUTO routing mode; it does NOT grant
+// access to specific models.
 func filterCandsByModels(cands []*provider.Provider, allow []config.ModelScope) []*provider.Provider {
 	for _, m := range allow {
-		if m.Model == "" || strings.EqualFold(m.Model, "AUTO") {
+		if m.Model == "" {
 			return cands
 		}
 	}
@@ -148,20 +167,18 @@ func filterCandsByModels(cands []*provider.Provider, allow []config.ModelScope)
 	return out
 }
 
-// intersectModels restricts a model list to the scope (preserving order). An
-// "AUTO" scope entry grants every model.
+// intersectModels restricts a model list to the scope (preserving order).
+// An "AUTO" scope entry only allows the AUTO routing mode; it does NOT grant
+// every model.
 func intersectModels(models []string, allow []config.ModelScope) []string {
 	allowed := make(map[string]bool, len(allow))
-	any := false
 	for _, m := range allow {
-		if m.Model == "" || strings.EqualFold(m.Model, "AUTO") {
-			any = true
-			break
+		if m.Model == "" {
+			return models
+		}
+		if !strings.EqualFold(m.Model, "AUTO") {
+			allowed[m.Model] = true
 		}
-		allowed[m.Model] = true
-	}
-	if any {
-		return models
 	}
 	out := make([]string, 0, len(models))
 	seen := map[string]bool{}
@@ -175,48 +192,100 @@ func intersectModels(models []string, allow []config.ModelScope) []string {
 }
 
 // checkModelScope validates the effective model against the key's model scope
-// and token quota. Returns an error message when rejected. A scope entry with
-// model "AUTO" grants all models; its quota caps the key's total tokens.
+// and token quota. Returns an error message when rejected.
+//
+// A scope entry with model "AUTO" only allows requests where the effective
+// model is AUTO (the routing mode). It does NOT grant access to specific model
+// ids — that requires an explicit scope entry for the model.
+// checkModelScope keeps its string-returning signature for callers that only
+// need to know whether the request may proceed.
 func (g *Gateway) checkModelScope(ctx context.Context, model string) string {
+	if q := g.checkQuota(ctx, model); q != nil {
+		return q.msg
+	}
+	return ""
+}
+
+// quotaRejection is a quota verdict: the message plus whether the client
+// should retry. A spent quota is a rate limit (429 + Retry-After), not a
+// permission failure (403): a client that sees 403 gives up on the key, while
+// one that sees 429 with a retry hint waits and resumes when the window rolls
+// over.
+type quotaRejection struct {
+	msg   string
+	retry int64 // seconds until the window resets; 0 = unknown
+}
+
+func (q *quotaRejection) Error() string { return q.msg }
+
+// checkQuota validates the effective model against the key's model scope and
+// that entry's quota. Returns nil when the request may proceed.
+//
+// Quotas are per scope entry, never key-wide: a model whose budget is spent
+// is refused on its own while the key's other models keep working. The verdict
+// carries the remaining seconds of the reset window so a spent budget answers
+// 429 + Retry-After (come back when it rolls over) instead of 403 (which reads
+// as "this key may never use this model" and makes clients give up).
+func (g *Gateway) checkQuota(ctx context.Context, model string) *quotaRejection {
 	allow := g.allowedModels(ctx)
 	if allow == nil {
-		return ""
-	}
-	for _, sc := range allow {
-		if sc.Model != "" && strings.EqualFold(sc.Model, "AUTO") {
-			if sc.TokenQuota > 0 {
-				used := g.scopeTokens(ctx, sc)
-				if used >= sc.TokenQuota {
-					return fmt.Sprintf("token quota exceeded (%d/%d)", used, sc.TokenQuota)
-				}
-			}
-			return ""
-		}
+		return nil
 	}
 	for _, sc := range allow {
 		if sc.Model != model {
 			continue
 		}
+		win := AutoPeriodSeconds(sc.Period, sc.Hours)
+		k := keyID(reqKey(ctx))
 		if sc.TokenQuota > 0 {
-			used := g.scopeTokens(ctx, sc)
-			if used >= sc.TokenQuota {
-				return fmt.Sprintf("token quota exceeded for %q (%d/%d)", model, used, sc.TokenQuota)
+			if used := g.scopeTokens(ctx, sc); used >= sc.TokenQuota {
+				return &quotaRejection{
+					msg:   fmt.Sprintf("token quota exceeded for %q (%d/%d)", model, used, sc.TokenQuota),
+					retry: AutoSecondsToReset(sc.Period, sc.Hours),
+				}
 			}
 		}
-		return ""
+		if sc.ReqQuota > 0 {
+			if used := g.stats.KeyWindowReqs(k, win); used >= sc.ReqQuota {
+				return &quotaRejection{
+					msg:   fmt.Sprintf("request quota exceeded for %q (%d/%d)", model, used, sc.ReqQuota),
+					retry: AutoSecondsToReset(sc.Period, sc.Hours),
+				}
+			}
+		}
+		return nil
 	}
-	return fmt.Sprintf("model %q is not allowed for this key", model)
+	return &quotaRejection{msg: fmt.Sprintf("model %q is not allowed for this key", model)}
 }
 
-// scopeTokens returns the tokens a scope entry has consumed within its reset
-// window (total for AUTO / per model otherwise).
+// quotaWindowSuffix describes a quota's reset window for an error message, so
+// a rejected caller can tell a permanent block from one that clears in an hour.
+func quotaWindowSuffix(period string, hours int64) string {
+	switch {
+	case period == "hour":
+		return ", resets hourly"
+	case period == "week":
+		return ", resets weekly"
+	case period == "month":
+		return ", resets monthly"
+	case period == "nhour" && hours > 1:
+		return fmt.Sprintf(", resets every %dh", hours)
+	}
+	return ""
+} // scopeTokens returns the tokens this key consumed on the scope entry's model
+// within its reset window, isolated per key. For an AUTO entry the cap covers
+// everything the key routed through AUTO; for a model entry it covers that
+// model only.
+//
+// It reads the per-key hourly buckets rather than the key-blind model
+// buckets, so one key's usage can never exhaust another's quota.
 func (g *Gateway) scopeTokens(ctx context.Context, sc config.ModelScope) int64 {
 	k := keyID(reqKey(ctx))
-	if sc.Model != "" && strings.EqualFold(sc.Model, "AUTO") {
-		return g.stats.KeyTokens(k)
-	}
 	win := AutoPeriodSeconds(sc.Period, sc.Hours)
-	return g.stats.WindowTokens(sc.Model, sc.Source, win)
+	if sc.Model != "" && strings.EqualFold(sc.Model, "AUTO") {
+		return g.stats.KeyWindowTokens(k, win)
+	}
+	return g.stats.KeyWindowModelTokens(k, sc.Model, sc.Source, win)
 }
 
 // hasScopeModel reports whether a model (possibly with a "source-model" /
@@ -225,18 +294,41 @@ func (g *Gateway) scopeTokens(ctx context.Context, sc config.ModelScope) int64 {
 // source that actually serves the bare model (via Registry.EffectiveModel), so
 // model ids that themselves contain separators (e.g. "deepseek-v4-flash-free")
 // are never corrupted (P10-2).
+//
+// A scope entry with model "AUTO" only matches the literal AUTO routing mode;
+// it does NOT grant access to specific model ids.
 func (g *Gateway) hasScopeModel(list []config.ModelScope, s string) bool {
 	if r := g.core.Registry(); r != nil {
 		s = r.EffectiveModel(s)
 	}
 	for _, x := range list {
-		if x.Model == s || (x.Model != "" && strings.EqualFold(x.Model, "AUTO")) {
+		if x.Model == s {
 			return true
 		}
 	}
 	return false
 }
 
+// writeScopeReject answers a model-scope or quota rejection. A spent quota is
+// 429 (rate_limit_exceeded) with Retry-After, so a client waits and resumes
+// after the reset; a model the key may not use stays 403 (model_not_allowed),
+// because retrying cannot help.
+func (g *Gateway) writeReject(w http.ResponseWriter, q *quotaRejection) {
+	if q.retry > 0 {
+		w.Header().Set("Retry-After", strconv.FormatInt(q.retry, 10))
+		writeError(w, http.StatusTooManyRequests, "rate_limit_exceeded", q.msg)
+		return
+	}
+	// no window to wait for: the cap is either permanent or key-wide with no
+	// period. "rate_limit_exceeded" still says "come back after the operator
+	// raises the cap", which 403 would not.
+	if strings.Contains(q.msg, "quota exceeded") {
+		writeError(w, http.StatusTooManyRequests, "rate_limit_exceeded", q.msg)
+		return
+	}
+	writeError(w, http.StatusForbidden, "model_not_allowed", q.msg)
+}
+
 func (g *Gateway) resolveByModel(model string) ([]*provider.Provider, string) {
 	if isAuto(model) {
 		return g.core.Registry().Resolve("AUTO"), ""
@@ -283,6 +375,8 @@ func (g *Gateway) handleChat(w http.ResponseWriter, r *http.Request) {
 		writeError(w, http.StatusBadRequest, "invalid_request", "invalid json: "+err.Error())
 		return
 	}
+	// 客户端自带的会话标识（若有）比推导出来的更准，见 clientSessionFromRequest。
+	clientSession := clientSessionFromRequest(r, req.PromptCacheKey)
 	if len(req.Messages) == 0 {
 		writeError(w, http.StatusBadRequest, "invalid_request", "messages is required")
 		return
@@ -297,8 +391,8 @@ func (g *Gateway) handleChat(w http.ResponseWriter, r *http.Request) {
 			writeError(w, http.StatusServiceUnavailable, "no_provider", "no auto slot configured")
 			return
 		}
-		if msg := g.checkModelScope(r.Context(), "AUTO"); msg != "" {
-			writeError(w, http.StatusForbidden, "model_not_allowed", msg)
+		if q := g.checkQuota(r.Context(), "AUTO"); q != nil {
+			g.writeReject(w, q)
 			return
 		}
 		ctx := r.Context()
@@ -313,6 +407,7 @@ func (g *Gateway) handleChat(w http.ResponseWriter, r *http.Request) {
 			ToolChoice:      req.ToolChoice,
 			DisableThinking: req.DisableThinking,
 			ExtraBody:       req.ExtraBody,
+			ClientSession:   clientSession,
 		}
 		rec := &Req{
 			Key:  keyID(reqKey(ctx)),
@@ -350,8 +445,8 @@ func (g *Gateway) handleChat(w http.ResponseWriter, r *http.Request) {
 	if effective == "" {
 		effective = firstModel(cands[0])
 	}
-	if msg := g.checkModelScope(r.Context(), effective); msg != "" {
-		writeError(w, http.StatusForbidden, "model_not_allowed", msg)
+	if q := g.checkQuota(r.Context(), effective); q != nil {
+		g.writeReject(w, q)
 		return
 	}
 	ctx := r.Context()
@@ -368,6 +463,7 @@ func (g *Gateway) handleChat(w http.ResponseWriter, r *http.Request) {
 		ToolChoice:      req.ToolChoice,
 		DisableThinking: req.DisableThinking,
 		ExtraBody:       req.ExtraBody,
+		ClientSession:   clientSession,
 	}
 	rec := &Req{
 		Key:    keyID(reqKey(ctx)),
@@ -384,13 +480,6 @@ func (g *Gateway) handleChat(w http.ResponseWriter, r *http.Request) {
 	g.singleChat(w, ctx, cands, inner, effective, rec)
 }
 
-func normalizeModel(m string) string {
-	if isAuto(m) {
-		return ""
-	}
-	return m
-}
-
 func firstModel(p *provider.Provider) string {
 	ms := p.Models()
 	if len(ms) > 0 {
@@ -520,21 +609,145 @@ func upstreamErrStatus(err error) int {
 	return http.StatusBadGateway
 }
 
-func (g *Gateway) singleChat(w http.ResponseWriter, ctx context.Context, cands []*provider.Provider, req *types.ChatRequest, effective string, rec *Req) {
-	rec.LatMs = 0
-	t0 := time.Now()
-	resp, usedSrc, usedModel, err := g.core.Scheduler().Chat(ctx, toScheduler(cands), req)
-	rec.LatMs = time.Since(t0).Milliseconds()
-	if err != nil {
-		rec.OK = false
-		rec.Status = upstreamErrStatus(err)
-		rec.Err = err.Error()
-		g.writeRec(rec)
-		writeError(w, rec.Status, "upstream_error", err.Error())
-		return
+// overflowMarkers 匹配上游「上下文超窗」类措辞。
+//
+// 上游写法五花八门，且**不在 pi 客户端的识别列表里**。pi 靠
+// @earendil-works/pi-ai 的 OVERFLOW_PATTERNS 判断超窗并据此触发压缩重试，
+// 而 justworker 返回的是「请精简对话历史…(Context window is full…)」——
+// 与那 25 条正则一条都不匹配，于是 pi 既不压缩也不重试，只把它当成一条
+// 普通上游错误。这里把可识别的超窗措辞归一化成 pi 一定认得的标记。
+var overflowMarkers = []*regexp.Regexp{
+	regexp.MustCompile(`(?i)context[ _-]?window is full`),
+	regexp.MustCompile(`(?i)context[_ ]length[_ ]exceeded`),
+	regexp.MustCompile(`(?i)exceeds? the context window`),
+	regexp.MustCompile(`(?i)maximum context length`),
+	regexp.MustCompile(`(?i)reduce the length of the messages`),
+	regexp.MustCompile(`(?i)too many tokens`),
+	regexp.MustCompile(`(?i)token limit exceeded`),
+	regexp.MustCompile(`请精简对话历史`),
+	regexp.MustCompile(`上下文(长度)?超(出|限)`),
+	regexp.MustCompile(`对话历史过长`),
+}
+
+// overflowCanonical 命中 pi 的 /context[_ ]length[_ ]exceeded/i。
+//
+// 注意：这个前缀只是必要条件，不是充分条件——pi 还会先用
+// NON_OVERFLOW_PATTERNS 排除整条消息，见 overflowClientMessage。
+const overflowCanonical = "context_length_exceeded"
+
+// looksLikeOverflow 判断错误文本是否属于上下文超窗。
+func looksLikeOverflow(s string) bool {
+	for _, re := range overflowMarkers {
+		if re.MatchString(s) {
+			return true
+		}
 	}
-	rec.OK = true
-	rec.Status = http.StatusOK
+	return false
+}
+
+// clientUpstreamErr collapses upstream failure details into a short message
+// for the client: per-tier bodies (WAF HTML pages, quota payloads, ...) stay
+// in rec.Err / the stats API and the server log instead of the response.
+// Per-tier one-line reasons are kept (quota/cooling skips carry no error body
+// and are the actionable part); each is capped so HTML dumps can't leak.
+//
+// 超窗会被归一化成「干净」的 overflowCanonical 消息，见 overflowClientMessage。
+func clientUpstreamErr(err error) string {
+	log.Printf("[gateway] upstream failure surfaced to client: %v", err)
+	if looksLikeOverflow(err.Error()) {
+		return overflowClientMessage(err)
+	}
+	return upstreamErrSummary(err)
+}
+
+// overflowHint 是给客户端的超窗说明，刻意只用 pi 认得的措辞。
+const overflowHint = "context window is full; reduce the length of the messages"
+
+// clientSessionFromRequest 取客户端自带的会话标识，拿不到时返回 ""。
+//
+// 通用 OpenAI 客户端默认不带会话 id，但 pi 支持：只要 provider 的 compat 里打开
+// sendSessionAffinityHeaders，pi 就会把**平台真实的会话 id**（uuidv7，整个会话
+// 恒定）放到 x-session-affinity / x-client-request-id 上（sessionAffinityFormat
+// 为 openrouter 时是 x-session-id，为 openai 时是 session_id）。
+// 有了它就不必再从首条 user 消息推导会话指纹（那只在客户端不发会话 id 时才作为
+// 退路，且历史压缩后会漂移）。prompt_cache_key 是 OpenAI 原生的缓存键，
+// 客户端若带也一并当会话用。
+//
+// 刻意不采纳 x-client-request-id：名字含 request，部分客户端每请求都换，
+// 拿它当会话会让上游缓存永不命中。pi 总会同时发 x-session-affinity，够用。
+func clientSessionFromRequest(r *http.Request, bodyKey string) string {
+	for _, h := range []string{"X-Session-Affinity", "X-Session-Id", "Session-Id"} {
+		if v := strings.TrimSpace(r.Header.Get(h)); v != "" {
+			return v
+		}
+	}
+	return strings.TrimSpace(bodyKey)
+}
+
+// overflowClientMessage 为超窗失败生成「干净」的客户端消息。
+//
+// 为什么不能沿用 upstreamErrSummary：pi 的 isContextOverflow 先查
+// NON_OVERFLOW_PATTERNS（/rate limit/、/too many requests/、Bedrock 前缀），
+// 一旦命中就直接判为「非超窗」——**哪怕消息里已经有 context_length_exceeded**，
+// pi 也不会压缩重试。而 AUTO 链的失败消息天生是多 tier 原因的拼接：
+// 超窗 tier（gozen 400 maximum context length）常与配额/限流 tier
+// （429 token plan exhausted、cooling、no free slot）同时出现。
+// 把明细原样带出去，等于让一条限流 tier 的措辞反过来封杀超窗识别。
+// 所以这里只保留「超窗」措辞 + 是哪个源超的窗，其余一律不带。
+func overflowClientMessage(err error) string {
+	var ce *scheduler.ChainErr
+	if errors.As(err, &ce) {
+		for _, t := range ce.Tiers {
+			if looksLikeOverflow(t.Err.Error()) {
+				return fmt.Sprintf("%s: %s (%s/%s)", overflowCanonical, overflowHint, t.Source, t.Model)
+			}
+		}
+	}
+	return fmt.Sprintf("%s: %s", overflowCanonical, overflowHint)
+}
+
+// upstreamErrSummary 把上游失败压成一行短消息。
+func upstreamErrSummary(err error) string {
+	var ce *scheduler.ChainErr
+	if errors.As(err, &ce) {
+		parts := make([]string, 0, len(ce.Tiers)+len(ce.Skipped))
+		for _, t := range ce.Tiers {
+			// 160 而不是 80：短诊断词（"Context window is full"）常落在尾部，
+			// 80 字节按字节截断正好会把它切掉，超窗就再也认不出来。
+			parts = append(parts, types.OneLine(fmt.Sprintf("%s/%s: %v", t.Source, t.Model, t.Err), 160))
+		}
+		for _, sk := range ce.Skipped {
+			parts = append(parts, types.OneLine(sk, 160))
+		}
+		msg := strings.Join(parts, "; ")
+		if len(msg) > 300 {
+			msg = msg[:300] + "..."
+		}
+		return fmt.Sprintf("all %d auto providers failed: %s", len(parts), msg)
+	}
+	return types.OneLine(err.Error(), 160)
+}
+
+// failChat maps a scheduling failure onto the audit record and answers the
+// client. A failed AUTO chain additionally pins its first failed tier onto
+// the record; direct-path errors never match *scheduler.ChainErr, so the
+// extraction is safely shared by all four entry points. Callers own the
+// writeRec call (inline for non-stream paths, deferred for stream paths).
+func (g *Gateway) failChat(w http.ResponseWriter, rec *Req, err error) {
+	rec.OK = false
+	rec.Status = upstreamErrStatus(err)
+	rec.Err = err.Error()
+	var ce *scheduler.ChainErr
+	if errors.As(err, &ce) && len(ce.Tiers) > 0 {
+		rec.Source = ce.Tiers[0].Source
+		rec.Model = ce.Tiers[0].Model
+	}
+	writeError(w, rec.Status, "upstream_error", clientUpstreamErr(err))
+}
+
+// recordChatUsage fills token accounting for a finished non-streaming
+// request: exact upstream numbers win, byte estimates fill the gaps.
+func recordChatUsage(rec *Req, req *types.ChatRequest, resp *types.UnifiedResponse) {
 	rec.Prompt = int64(resp.TokenUsage.Prompt)
 	if rec.Prompt == 0 {
 		rec.Prompt = estimatePromptTokens(req)
@@ -543,9 +756,29 @@ func (g *Gateway) singleChat(w http.ResponseWriter, ctx context.Context, cands [
 	if rec.Compl == 0 {
 		rec.Compl = estimateTextTokens(resp.Content, resp.ReasoningContent, resp.ToolCalls)
 	}
-	rec.Source = usedSrc
-	rec.Model = usedModel
-	g.writeRec(rec)
+	// Cache accounting: whichever source format the adapter normalized into
+	// (prompt_tokens_details.cached_tokens or legacy prompt_cache_hit_tokens),
+	// read it back so the request record carries the hit/miss split.
+	if d := resp.TokenUsage.PromptTokensDetails; d != nil && d.CachedTokens > 0 {
+		rec.CacheHit = int64(d.CachedTokens)
+	}
+	if resp.TokenUsage.PromptTokensDetails != nil {
+		// Upstream reported cache details (even a 0 hit) — tag the row so
+		// the UI can show 0% rather than “—”.
+		rec.CacheReported = true
+	} else if resp.TokenUsage.PromptCacheHit > 0 {
+		rec.CacheHit = int64(resp.TokenUsage.PromptCacheHit)
+		rec.CacheReported = true
+	}
+	if resp.TokenUsage.PromptCacheMiss > 0 {
+		rec.CacheMiss = int64(resp.TokenUsage.PromptCacheMiss)
+	}
+}
+
+// writeChatCompletion renders a unified response as an OpenAI
+// chat.completion object. modelName is the id clients see as the serving
+// model: the requested id for direct routes, the exact slot model for AUTO.
+func writeChatCompletion(w http.ResponseWriter, resp *types.UnifiedResponse, modelName string) {
 	msg := RespMessage{Role: "assistant", Content: resp.Content}
 	if resp.ReasoningContent != "" {
 		msg.ReasoningContent = resp.ReasoningContent
@@ -557,8 +790,9 @@ func (g *Gateway) singleChat(w http.ResponseWriter, ctx context.Context, cands [
 		ID:      newID(),
 		Object:  "chat.completion",
 		Created: time.Now().Unix(),
-		Model:   effective,
+		Model:   modelName,
 		Choices: []ChatChoice{{Index: 0, Message: msg, FinishReason: resp.FinishReason}},
+		Cost:    resp.Cost,
 	}
 	if resp.TokenUsage.Total > 0 || resp.TokenUsage.Prompt > 0 || resp.TokenUsage.Completion > 0 {
 		out.Usage = &resp.TokenUsage
@@ -566,6 +800,30 @@ func (g *Gateway) singleChat(w http.ResponseWriter, ctx context.Context, cands [
 	writeJSON(w, http.StatusOK, out)
 }
 
+// singleChat runs a direct (model-pinned) non-streaming request across the
+// candidate list, falling back on failure.
+func (g *Gateway) singleChat(w http.ResponseWriter, ctx context.Context, cands []*provider.Provider, req *types.ChatRequest, effective string, rec *Req) {
+	rec.LatMs = 0
+	t0 := time.Now()
+	resp, usedSrc, usedModel, err := g.core.Scheduler().Chat(ctx, toScheduler(cands), req)
+	rec.LatMs = time.Since(t0).Milliseconds()
+	if err != nil {
+		g.failChat(w, rec, err)
+		g.writeRec(rec)
+		return
+	}
+	rec.OK = true
+	rec.Status = http.StatusOK
+	recordChatUsage(rec, req, resp)
+	rec.Source = usedSrc
+	rec.Model = usedModel
+	// Non-streaming: the whole response arrives at once, so TTFB equals
+	// the total latency.
+	rec.FirstByteMs = rec.LatMs
+	g.writeRec(rec)
+	writeChatCompletion(w, resp, effective)
+}
+
 // writeRec records a finished request (audit + aggregates).
 func (g *Gateway) writeRec(rec *Req) {
 	if rec == nil {
@@ -577,27 +835,51 @@ func (g *Gateway) writeRec(rec *Req) {
 	g.stats.Record(*rec)
 }
 
-func (g *Gateway) streamChat(w http.ResponseWriter, ctx context.Context, cands []*provider.Provider, req *types.ChatRequest, effective string, rec *Req) {
-	rec.LatMs = 0
-	t0 := time.Now()
-	rec.OK = true
-	rec.Status = http.StatusOK
-	defer func() {
-		rec.LatMs = time.Since(t0).Milliseconds()
-		g.writeRec(rec)
-	}()
-	chunks, _, usedModel, err := g.core.Scheduler().ChatStream(ctx, toScheduler(cands), req)
-	if err != nil {
-		rec.OK = false
-		rec.Status = upstreamErrStatus(err)
-		rec.Err = err.Error()
-		writeError(w, rec.Status, "upstream_error", err.Error())
-		return
+// mergeUsage combines token usage across stream chunks additively. Some
+// providers split usage across chunks (e.g. Anthropic reports prompt tokens
+// in message_start and the final completion tokens in message_delta); a plain
+// "last non-nil wins" would discard the prompt half. Non-zero fields from cur
+// override prev; total is recomputed from the merged parts so a partial later
+// chunk can't shrink it. For the common single-chunk case (OpenAI's terminal
+// empty-choices+usage chunk) upstream totals are preserved exactly.
+func mergeUsage(prev, cur *types.TokenUsage) *types.TokenUsage {
+	if prev == nil {
+		u := *cur
+		if u.Total == 0 && (u.Prompt > 0 || u.Completion > 0) {
+			u.Total = u.Prompt + u.Completion
+		}
+		return &u
 	}
-	if usedModel != "" {
-		rec.Model = usedModel
+	out := *prev
+	if cur.Prompt > 0 {
+		out.Prompt = cur.Prompt
 	}
-	rec.Prompt = estimatePromptTokens(req)
+	if cur.Completion > 0 {
+		out.Completion = cur.Completion
+	}
+	out.Total = out.Prompt + out.Completion
+	if cur.PromptTokensDetails != nil {
+		// Keep the details object even when CachedTokens is 0: a reported
+		// zero-hit is meaningful ("cache missed") and must stay
+		// distinguishable from "upstream never reported cache info".
+		// Dropping it here made streaming rows show “—” instead of 0%.
+		out.PromptTokensDetails = cur.PromptTokensDetails
+	}
+	if cur.PromptCacheHit > 0 {
+		out.PromptCacheHit = cur.PromptCacheHit
+	}
+	if cur.PromptCacheMiss > 0 {
+		out.PromptCacheMiss = cur.PromptCacheMiss
+	}
+	return &out
+}
+
+// pumpStream writes the full SSE sequence for a started stream: role
+// preamble, one chunk per unified delta, the terminating finish_reason, the
+// OpenAI-standard final usage chunk (empty choices) and [DONE]. modelName
+// follows writeChatCompletion's rule (requested id for direct routes, exact
+// slot model for AUTO).
+func (g *Gateway) pumpStream(w http.ResponseWriter, rec *Req, chunks <-chan types.UnifiedChunk, modelName string, t0 time.Time) {
 	w.Header().Set("Content-Type", "text/event-stream")
 	w.Header().Set("Cache-Control", "no-cache")
 	w.Header().Set("Connection", "keep-alive")
@@ -621,14 +903,26 @@ func (g *Gateway) streamChat(w http.ResponseWriter, ctx context.Context, cands [
 	}
 
 	if !send(ChatChunk{
-		ID: id, Object: "chat.completion.chunk", Created: created, Model: effective,
+		ID: id, Object: "chat.completion.chunk", Created: created, Model: modelName,
 		Choices: []ChunkChoice{{Index: 0, Delta: RespMessage{Role: "assistant"}}},
 	}) {
 		return
 	}
+	// First SSE byte sent to the client: record time-to-first-byte for the
+	// source's status-page latency average.
+	rec.FirstByteMs = time.Since(t0).Milliseconds()
+	var lastUsage *types.TokenUsage
+	var lastCost string
+	lastFinish := ""
 	for ck := range chunks {
+		if ck.Usage != nil {
+			lastUsage = mergeUsage(lastUsage, ck.Usage)
+		}
+		if ck.Cost != "" {
+			lastCost = ck.Cost
+		}
 		chunk := ChatChunk{
-			ID: id, Object: "chat.completion.chunk", Created: created, Model: effective,
+			ID: id, Object: "chat.completion.chunk", Created: created, Model: modelName,
 		}
 		delta := RespMessage{Role: "assistant", Content: ck.Content}
 		if ck.ReasoningContent != "" {
@@ -639,8 +933,12 @@ func (g *Gateway) streamChat(w http.ResponseWriter, ctx context.Context, cands [
 		}
 		choice := ChunkChoice{Index: 0, Delta: delta}
 		if ck.Done {
-			stop := "stop"
-			choice.FinishReason = &stop
+			fin := ck.FinishReason
+			if fin == "" {
+				fin = "stop"
+			}
+			lastFinish = fin
+			choice.FinishReason = &fin
 		}
 		chunk.Choices = []ChunkChoice{choice}
 		rec.Compl += int64(len(ck.Content)+len(ck.ReasoningContent)+len(ck.ToolCalls)) / 3
@@ -648,17 +946,103 @@ func (g *Gateway) streamChat(w http.ResponseWriter, ctx context.Context, cands [
 			return
 		}
 	}
-	stop := "stop"
+	finalFinish := lastFinish
+	if finalFinish == "" {
+		finalFinish = "stop"
+	}
 	send(ChatChunk{
-		ID: id, Object: "chat.completion.chunk", Created: created, Model: effective,
-		Choices: []ChunkChoice{{Index: 0, Delta: RespMessage{}, FinishReason: &stop}},
+		ID: id, Object: "chat.completion.chunk", Created: created, Model: modelName,
+		Choices: []ChunkChoice{{Index: 0, Delta: RespMessage{}, FinishReason: &finalFinish}},
 	})
+	// Final usage chunk (OpenAI standard: empty choices + usage before [DONE]).
+	// Prefer the upstream's exact usage if the stream carried it; fall back to
+	// the gateway's estimate otherwise.
+	// Write the upstream's exact numbers back onto the audit record. Without
+	// this the streamed path kept only the per-chunk byte estimate, so the same
+	// request recorded ~1.4x its real completion tokens (measured: upstream 100,
+	// audit 145) while the non-streaming path recorded 100. Two paths, two
+	// different numbers for one request is a reporting bug, not a rounding one.
+	if lastUsage != nil {
+		if lastUsage.Prompt > 0 {
+			rec.Prompt = int64(lastUsage.Prompt)
+		}
+		if lastUsage.Completion > 0 {
+			rec.Compl = int64(lastUsage.Completion)
+		}
+	}
+	var tut *types.TokenUsage
+	if lastUsage != nil {
+		tut = lastUsage
+		// Write the upstream's cache accounting back onto the request record
+		// so the audit trail carries the hit/miss split for streaming too.
+		if d := tut.PromptTokensDetails; d != nil {
+			rec.CacheReported = true
+			if d.CachedTokens > 0 {
+				rec.CacheHit = int64(d.CachedTokens)
+			}
+		} else if tut.PromptCacheHit > 0 {
+			rec.CacheHit = int64(tut.PromptCacheHit)
+			rec.CacheReported = true
+		}
+		if tut.PromptCacheMiss > 0 {
+			rec.CacheMiss = int64(tut.PromptCacheMiss)
+		}
+	} else if rec.Prompt+rec.Compl > 0 {
+		u := types.TokenUsage{
+			Prompt:     int(rec.Prompt),
+			Completion: int(rec.Compl),
+			Total:      int(rec.Prompt + rec.Compl),
+		}
+		tut = &u
+	}
+	if tut != nil {
+		send(ChatChunk{
+			ID: id, Object: "chat.completion.chunk", Created: created, Model: modelName,
+			Choices: []ChunkChoice{},
+			Usage:   tut,
+			Cost:    lastCost,
+		})
+	} else if lastCost != "" {
+		// Cost arrived without any usage (OpenCode sends it on its own frame).
+		send(ChatChunk{
+			ID: id, Object: "chat.completion.chunk", Created: created, Model: modelName,
+			Choices: []ChunkChoice{},
+			Cost:    lastCost,
+		})
+	}
 	fmt.Fprintf(w, "data: [DONE]\n\n")
 	if flusher != nil {
 		flusher.Flush()
 	}
 }
 
+// streamChat runs a direct (model-pinned) streaming request across the
+// candidate list, falling back early on connect errors.
+func (g *Gateway) streamChat(w http.ResponseWriter, ctx context.Context, cands []*provider.Provider, req *types.ChatRequest, effective string, rec *Req) {
+	rec.LatMs = 0
+	t0 := time.Now()
+	rec.OK = true
+	rec.Status = http.StatusOK
+	defer func() {
+		rec.LatMs = time.Since(t0).Milliseconds()
+		g.writeRec(rec)
+	}()
+	chunks, usedSrc, usedModel, err := g.core.Scheduler().ChatStream(ctx, toScheduler(cands), req)
+	if err != nil {
+		g.failChat(w, rec, err)
+		return
+	}
+	if usedModel != "" {
+		rec.Model = usedModel
+	}
+	// Audit accuracy: pin the source that actually served the stream (after a
+	// failover it differs from the first candidate). Direct streams previously
+	// discarded it.
+	rec.Source = usedSrc
+	rec.Prompt = estimatePromptTokens(req)
+	g.pumpStream(w, rec, chunks, effective, t0)
+}
+
 // singleChatAuto runs a non-streaming AUTO request down the chain (see
 // scheduler.ChainChat): tiers ascending (tier 1 = highest priority first),
 // per-tier round-robin ordered by preference, cooldown as the only hard skip,
@@ -671,48 +1055,18 @@ func (g *Gateway) singleChatAuto(w http.ResponseWriter, ctx context.Context, cha
 	resp, usedSrc, usedModel, err := g.core.Scheduler().ChainChat(ctx, chain, req, quotaExhausted)
 	rec.LatMs = time.Since(t0).Milliseconds()
 	if err != nil {
-		rec.OK = false
-		rec.Status = upstreamErrStatus(err)
-		rec.Err = err.Error()
-		if ce, ok := err.(*scheduler.ChainErr); ok && len(ce.Tiers) > 0 {
-			rec.Source = ce.Tiers[0].Source
-			rec.Model = ce.Tiers[0].Model
-		}
+		g.failChat(w, rec, err)
 		g.writeRec(rec)
-		writeError(w, rec.Status, "upstream_error", err.Error())
 		return
 	}
 	rec.OK = true
 	rec.Status = http.StatusOK
-	rec.Prompt = int64(resp.TokenUsage.Prompt)
-	if rec.Prompt == 0 {
-		rec.Prompt = estimatePromptTokens(req)
-	}
-	rec.Compl = int64(resp.TokenUsage.Completion)
-	if rec.Compl == 0 {
-		rec.Compl = estimateTextTokens(resp.Content, resp.ReasoningContent, resp.ToolCalls)
-	}
+	recordChatUsage(rec, req, resp)
 	rec.Source = usedSrc
 	rec.Model = usedModel
+	rec.FirstByteMs = rec.LatMs
 	g.writeRec(rec)
-	msg := RespMessage{Role: "assistant", Content: resp.Content}
-	if resp.ReasoningContent != "" {
-		msg.ReasoningContent = resp.ReasoningContent
-	}
-	if len(resp.ToolCalls) > 0 {
-		msg.ToolCalls = toolCallsWire(resp.ToolCalls)
-	}
-	out := ChatCompletion{
-		ID:      newID(),
-		Object:  "chat.completion",
-		Created: time.Now().Unix(),
-		Model:   usedModel,
-		Choices: []ChatChoice{{Index: 0, Message: msg, FinishReason: resp.FinishReason}},
-	}
-	if resp.TokenUsage.Total > 0 || resp.TokenUsage.Prompt > 0 || resp.TokenUsage.Completion > 0 {
-		out.Usage = &resp.TokenUsage
-	}
-	writeJSON(w, http.StatusOK, out)
+	writeChatCompletion(w, resp, usedModel)
 }
 
 // streamChatAuto streams an AUTO request down the chain. A slot is abandoned
@@ -730,14 +1084,7 @@ func (g *Gateway) streamChatAuto(w http.ResponseWriter, ctx context.Context, cha
 	}()
 	chunks, usedSrc, usedModel, err := g.core.Scheduler().ChainChatStream(ctx, chain, req, quotaExhausted)
 	if err != nil {
-		rec.OK = false
-		rec.Status = upstreamErrStatus(err)
-		rec.Err = err.Error()
-		if ce, ok := err.(*scheduler.ChainErr); ok && len(ce.Tiers) > 0 {
-			rec.Source = ce.Tiers[0].Source
-			rec.Model = ce.Tiers[0].Model
-		}
-		writeError(w, rec.Status, "upstream_error", err.Error())
+		g.failChat(w, rec, err)
 		return
 	}
 	if usedModel != "" {
@@ -745,63 +1092,7 @@ func (g *Gateway) streamChatAuto(w http.ResponseWriter, ctx context.Context, cha
 	}
 	rec.Source = usedSrc
 	rec.Prompt = estimatePromptTokens(req)
-	w.Header().Set("Content-Type", "text/event-stream")
-	w.Header().Set("Cache-Control", "no-cache")
-	w.Header().Set("Connection", "keep-alive")
-	w.WriteHeader(http.StatusOK)
-	flusher, _ := w.(http.Flusher)
-	id := newID()
-	created := time.Now().Unix()
-	send := func(obj interface{}) bool {
-		b, err := json.Marshal(obj)
-		if err != nil {
-			return false
-		}
-		if _, err := fmt.Fprintf(w, "data: %s\n\n", b); err != nil {
-			return false
-		}
-		if flusher != nil {
-			flusher.Flush()
-		}
-		return true
-	}
-	if !send(ChatChunk{
-		ID: id, Object: "chat.completion.chunk", Created: created, Model: rec.Model,
-		Choices: []ChunkChoice{{Index: 0, Delta: RespMessage{Role: "assistant"}}},
-	}) {
-		return
-	}
-	for ck := range chunks {
-		chunk := ChatChunk{
-			ID: id, Object: "chat.completion.chunk", Created: created, Model: rec.Model,
-		}
-		delta := RespMessage{Role: "assistant", Content: ck.Content}
-		if ck.ReasoningContent != "" {
-			delta.ReasoningContent = ck.ReasoningContent
-		}
-		if len(ck.ToolCalls) > 0 {
-			delta.ToolCalls = ck.ToolCalls
-		}
-		choice := ChunkChoice{Index: 0, Delta: delta}
-		if ck.Done {
-			stop := "stop"
-			choice.FinishReason = &stop
-		}
-		chunk.Choices = []ChunkChoice{choice}
-		rec.Compl += int64(len(ck.Content)+len(ck.ReasoningContent)+len(ck.ToolCalls)) / 3
-		if !send(chunk) {
-			return
-		}
-	}
-	stop := "stop"
-	send(ChatChunk{
-		ID: id, Object: "chat.completion.chunk", Created: created, Model: rec.Model,
-		Choices: []ChunkChoice{{Index: 0, Delta: RespMessage{}, FinishReason: &stop}},
-	})
-	fmt.Fprintf(w, "data: [DONE]\n\n")
-	if flusher != nil {
-		flusher.Flush()
-	}
+	g.pumpStream(w, rec, chunks, usedModel, t0)
 }
 
 func (g *Gateway) handleImage(w http.ResponseWriter, r *http.Request) {
@@ -822,11 +1113,48 @@ func (g *Gateway) handleImage(w http.ResponseWriter, r *http.Request) {
 	if model == "" {
 		model = g.core.DefaultModel()
 	}
-	if !isAuto(model) {
-		if allow := g.allowedModels(r.Context()); allow != nil && !g.hasScopeModel(allow, model) {
-			writeError(w, http.StatusForbidden, "model_not_allowed", fmt.Sprintf("model %q is not allowed for this key", model))
+	if isAuto(model) {
+		if chain := g.core.AutoImageChain(); chain != nil && len(chain.Tiers) > 0 {
+			if q := g.checkQuota(r.Context(), "AUTO"); q != nil {
+				g.writeReject(w, q)
+				return
+			}
+			done := g.stats.Begin()
+			defer done()
+			rec := &Req{Key: keyID(reqKey(r.Context())), Type: "image", Model: model, OK: false}
+			t0 := time.Now()
+			resp, usedSrc, usedModel, err := g.core.Scheduler().ChainImage(r.Context(), chain, &req)
+			rec.LatMs = time.Since(t0).Milliseconds()
+			if err != nil {
+				rec.Status = upstreamErrStatus(err)
+				rec.Err = err.Error()
+				g.writeRec(rec)
+				writeError(w, rec.Status, "upstream_error", clientUpstreamErr(err))
+				return
+			}
+			rec.Source = usedSrc
+			if usedModel != "" {
+				rec.Model = usedModel // actual image model served, not "AUTO"
+			}
+			rec.OK = true
+			rec.Status = http.StatusOK
+			// Image generation has no token concept. Recording len(ImageData)
+			// (the image COUNT) in completion_tokens mislabels image count as
+			// tokens and feeds it into the token totals; leave it 0.
+			rec.ImageCount = len(resp.ImageData)
+			g.writeRec(rec)
+			writeJSON(w, http.StatusOK, types.ImageGenResponse{
+				Created: time.Now().Unix(),
+				Data:    resp.ImageData,
+			})
 			return
 		}
+		// no image chain configured: fall through to legacy discovery (all
+		// sources exposing an image model, tried in registry order)
+	}
+	if allow := g.allowedModels(r.Context()); allow != nil && !g.hasScopeModel(allow, model) {
+		writeError(w, http.StatusForbidden, "model_not_allowed", fmt.Sprintf("model %q is not allowed for this key", model))
+		return
 	}
 	cands, _ := g.resolveByModel(model)
 	cands = imageOnly(cands)
@@ -837,8 +1165,8 @@ func (g *Gateway) handleImage(w http.ResponseWriter, r *http.Request) {
 		writeError(w, http.StatusServiceUnavailable, "no_provider", "no image source configured")
 		return
 	}
-	if msg := g.checkModelScope(r.Context(), effectiveImageModel(model, cands)); msg != "" {
-		writeError(w, http.StatusForbidden, "model_not_allowed", msg)
+	if q := g.checkQuota(r.Context(), effectiveImageModel(model, cands)); q != nil {
+		g.writeReject(w, q)
 		return
 	}
 	done := g.stats.Begin()
@@ -851,15 +1179,19 @@ func (g *Gateway) handleImage(w http.ResponseWriter, r *http.Request) {
 		rec.Status = upstreamErrStatus(err)
 		rec.Err = err.Error()
 		g.writeRec(rec)
-		writeError(w, rec.Status, "upstream_error", err.Error())
+		writeError(w, rec.Status, "upstream_error", clientUpstreamErr(err))
 		return
 	}
 	if usedSrc != "" {
 		rec.Source = usedSrc
 	}
+	if resp.Model != "" {
+		rec.Model = resp.Model // record the actual model served, not the raw request id
+	}
 	rec.OK = true
 	rec.Status = http.StatusOK
-	rec.Compl = int64(len(resp.ImageData))
+	// Image generation has no token concept — see the AUTO path above.
+	rec.ImageCount = len(resp.ImageData)
 	g.writeRec(rec)
 	writeJSON(w, http.StatusOK, types.ImageGenResponse{
 		Created: time.Now().Unix(),

internal/gateway/chat_cache_test.go

@@ -0,0 +1,62 @@
+package gateway
+
+import (
+	"testing"
+
+	"llmsproxy/internal/types"
+)
+
+// TestMergeUsageKeepsZeroCacheDetails locks the streaming counterpart of the
+// "distinguish missed from not reported" contract: a usage chunk that reports
+// prompt_tokens_details with 0 cached tokens must survive the merge, otherwise
+// streaming rows lose cache_reported and the UI shows "—" instead of 0%.
+func TestMergeUsageKeepsZeroCacheDetails(t *testing.T) {
+	t.Run("zero-hit details survive merge", func(t *testing.T) {
+		prev := &types.TokenUsage{Prompt: 10, Completion: 1, Total: 11}
+		cur := &types.TokenUsage{
+			Prompt: 10, Completion: 5, Total: 15,
+			PromptTokensDetails: &types.PromptTokensDetails{CachedTokens: 0},
+		}
+		got := mergeUsage(prev, cur)
+		if got.PromptTokensDetails == nil {
+			t.Fatal("zero-hit prompt_tokens_details was dropped by mergeUsage")
+		}
+		if got.PromptTokensDetails.CachedTokens != 0 {
+			t.Fatalf("cached_tokens = %d, want 0", got.PromptTokensDetails.CachedTokens)
+		}
+	})
+
+	t.Run("non-zero hit still wins", func(t *testing.T) {
+		prev := &types.TokenUsage{Prompt: 10, Completion: 1, Total: 11}
+		cur := &types.TokenUsage{
+			Prompt: 10, Completion: 5, Total: 15,
+			PromptTokensDetails: &types.PromptTokensDetails{CachedTokens: 64},
+		}
+		got := mergeUsage(prev, cur)
+		if got.PromptTokensDetails == nil || got.PromptTokensDetails.CachedTokens != 64 {
+			t.Fatalf("cached_tokens lost: %+v", got.PromptTokensDetails)
+		}
+	})
+
+	t.Run("absent details do not overwrite an earlier report", func(t *testing.T) {
+		prev := &types.TokenUsage{
+			Prompt: 10, Completion: 1, Total: 11,
+			PromptTokensDetails: &types.PromptTokensDetails{CachedTokens: 32},
+		}
+		cur := &types.TokenUsage{Prompt: 10, Completion: 5, Total: 15}
+		got := mergeUsage(prev, cur)
+		if got.PromptTokensDetails == nil || got.PromptTokensDetails.CachedTokens != 32 {
+			t.Fatalf("earlier cache report clobbered by a later chunk without details: %+v",
+				got.PromptTokensDetails)
+		}
+	})
+
+	t.Run("no cache data anywhere stays nil", func(t *testing.T) {
+		prev := &types.TokenUsage{Prompt: 10, Completion: 1, Total: 11}
+		cur := &types.TokenUsage{Prompt: 10, Completion: 5, Total: 15}
+		got := mergeUsage(prev, cur)
+		if got.PromptTokensDetails != nil {
+			t.Fatalf("fabricated cache details: %+v", got.PromptTokensDetails)
+		}
+	})
+}

internal/gateway/client_session_test.go

@@ -0,0 +1,47 @@
+package gateway
+
+import (
+	"net/http"
+	"testing"
+)
+
+// 客户端自带的会话标识必须被采纳：它比「首条 user 消息指纹」更准
+// （历史压缩后指纹会漂移），也是唯一能天然按会话隔离的来源。
+func TestClientSessionFromRequestHeaders(t *testing.T) {
+	cases := []struct {
+		name    string
+		headers map[string]string
+		bodyKey string
+		want    string
+	}{
+		{"pi 默认格式 x-session-affinity", map[string]string{"X-Session-Affinity": "ses-1"}, "", "ses-1"},
+		{"openrouter 格式 x-session-id", map[string]string{"X-Session-Id": "ses-2"}, "", "ses-2"},
+		{"openai 格式 session_id", map[string]string{"Session-Id": "ses-3"}, "", "ses-3"},
+		{"body prompt_cache_key 兜底", nil, "cache-4", "cache-4"},
+		{"头优先于 body", map[string]string{"X-Session-Affinity": "ses-5"}, "cache-x", "ses-5"},
+		{"空白不算", map[string]string{"X-Session-Affinity": "   "}, "", ""},
+		{"都没有则为空（适配器自行推导）", nil, "", ""},
+	}
+
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			r, _ := http.NewRequest(http.MethodPost, "/v1/chat/completions", nil)
+			for k, v := range tc.headers {
+				r.Header.Set(k, v)
+			}
+			if got := clientSessionFromRequest(r, tc.bodyKey); got != tc.want {
+				t.Errorf("clientSessionFromRequest = %q, want %q", got, tc.want)
+			}
+		})
+	}
+}
+
+// 刻意不采纳 x-client-request-id：名字含 request，部分客户端每请求都换，
+// 拿它当会话会让上游前缀缓存永不命中。
+func TestClientRequestIDIsNotUsedAsSession(t *testing.T) {
+	r, _ := http.NewRequest(http.MethodPost, "/v1/chat/completions", nil)
+	r.Header.Set("X-Client-Request-Id", "per-request-uuid")
+	if got := clientSessionFromRequest(r, ""); got != "" {
+		t.Errorf("x-client-request-id 不应被当作会话（否则缓存永不命中），got %q", got)
+	}
+}

internal/gateway/gateway_test.go

@@ -6,10 +6,13 @@ import (
 	"io"
 	"net/http"
 	"net/http/httptest"
+	"net/url"
 	"os"
 	"path/filepath"
+	"strconv"
 	"strings"
 	"testing"
+	"time"
 
 	"llmsproxy/internal/config"
 	"llmsproxy/internal/core"
@@ -139,7 +142,7 @@ func TestChatAutoChain503Summary(t *testing.T) {
 	if rr.Code != http.StatusServiceUnavailable {
 		t.Fatalf("status=%d body=%s", rr.Code, rr.Body.String())
 	}
-	if !strings.Contains(rr.Body.String(), "all auto tiers failed") ||
+	if !strings.Contains(rr.Body.String(), "auto providers failed") ||
 		!strings.Contains(rr.Body.String(), "a/a-m") ||
 		!strings.Contains(rr.Body.String(), "b/b-m") {
 		t.Fatalf("503 must summarize every tier, body=%s", rr.Body.String())
@@ -667,7 +670,351 @@ func TestHasScopeModelWithSourcePrefix(t *testing.T) {
 	if g.hasScopeModel(scope, "deepseek-v4-flash-free-extra") {
 		t.Error("hasScopeModel returned true for unrelated model")
 	}
-	if !g.hasScopeModel([]config.ModelScope{{Model: "AUTO"}}, "zen:anything") {
-		t.Error("AUTO scope should allow any prefixed model")
+	if g.hasScopeModel([]config.ModelScope{{Model: "AUTO"}}, "zen:deepseek-v4-flash-free") {
+		t.Error("AUTO scope must NOT grant specific model ids even with prefix")
+	}
+	if g.hasScopeModel([]config.ModelScope{{Model: "AUTO"}}, "deepseek-v4-pro") {
+		t.Error("AUTO scope must NOT grant specific model ids")
+	}
+}
+
+// TestStatusRecorderFlusher pins the SSE-critical contract: the access-log
+// wrapper must implement http.Flusher, otherwise the streaming chat handlers'
+// w.(http.Flusher) assertion yields nil and chunks are never flushed until
+// the response ends (regression for the buffered-SSE bug).
+func TestStatusRecorderFlusher(t *testing.T) {
+	var inner *flushRecorder
+	rr := httptest.NewRecorder()
+	inner = &flushRecorder{ResponseWriter: rr}
+	sr := &statusRecorder{ResponseWriter: inner}
+
+	fl, ok := any(sr).(http.Flusher)
+	if !ok {
+		t.Fatal("statusRecorder does not implement http.Flusher — SSE streaming is broken")
+	}
+	if inner.flushed {
+		t.Fatal("Flush called before Flush()")
+	}
+	fl.Flush()
+	if !inner.flushed {
+		t.Fatal("statusRecorder.Flush did not forward to the underlying writer")
+	}
+}
+
+// flushRecorder records whether Flush was forwarded.
+type flushRecorder struct {
+	http.ResponseWriter
+	flushed bool
+}
+
+func (f *flushRecorder) Flush() { f.flushed = true }
+
+// TestDirectStreamFailoverAuditSource: a direct streaming request whose first
+// candidate hard-fails must be served by the fallback, and the audit record's
+// Source must name the source that actually served the stream (not just the
+// first candidate).
+func TestDirectStreamFailoverAuditSource(t *testing.T) {
+	aUp := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.WriteHeader(500)
+		fmt.Fprint(w, `{"error":"boom"}`)
+	}))
+	defer aUp.Close()
+	bUp := mockUpstream()
+	defer bUp.Close()
+	g := newTestGateway(t,
+		// Both sources expose the same model id so the direct scheduler has a
+		// real fallback candidate after a hard-fails; b is the healthy one.
+		config.Source{Name: "a", BaseURL: aUp.URL, Adapter: "openai", Models: []config.Model{{ID: "m"}}},
+		config.Source{Name: "b", BaseURL: bUp.URL, Adapter: "openai", Models: []config.Model{{ID: "m"}}},
+	)
+	body := doReq(t, g, "POST", "/v1/chat/completions",
+		`{"model":"m","stream":true,"messages":[{"role":"user","content":"hi"}]}`)
+	if body.Code != 200 {
+		t.Fatalf("status=%d body=%s", body.Code, body.Body.String())
+	}
+	sb := body.Body.String()
+	if !strings.Contains(sb, `"finish_reason":"stop"`) || !strings.Contains(sb, "[DONE]") {
+		t.Fatalf("stream incomplete: %s", sb)
+	}
+	recs := g.stats.AuditRecords(0, 0, "")
+	if len(recs) == 0 {
+		t.Fatal("no audit records")
+	}
+	last := recs[len(recs)-1]
+	if last.Source != "b" || !last.OK {
+		t.Fatalf("audit rec source=%q ok=%v, want source=b ok=true (actual serving source)", last.Source, last.OK)
+	}
+}
+
+// TestStatusExposesAdapterPools checks that the elastic Lua pool sizing is
+// visible to admins (and only to them) via /api/status, so the grow/shrink
+// algorithm can be observed instead of inferred.
+func TestStatusExposesAdapterPools(t *testing.T) {
+	g := newTestGateway(t)
+	rr := doReq(t, g, http.MethodGet, "/api/status", "")
+	if rr.Code != 200 {
+		t.Fatalf("status = %d, body=%s", rr.Code, rr.Body.String())
+	}
+	var body map[string]interface{}
+	if err := json.Unmarshal(rr.Body.Bytes(), &body); err != nil {
+		t.Fatalf("decode: %v", err)
+	}
+	pools, ok := body["adapter_pools"].([]interface{})
+	if !ok {
+		t.Fatalf("adapter_pools missing or wrong type: %T", body["adapter_pools"])
+	}
+	if len(pools) == 0 {
+		t.Fatal("adapter_pools must list the loaded adapters")
+	}
+	first, _ := pools[0].(map[string]interface{})
+	for _, field := range []string{"name", "created", "idle", "in_use", "max", "grow_step", "shrink_step"} {
+		if _, ok := first[field]; !ok {
+			t.Errorf("pool entry missing %q: %v", field, first)
+		}
+	}
+	// idle gateway: nothing booted eagerly
+	if c, _ := first["created"].(float64); c != 0 {
+		t.Errorf("adapter pool booted eagerly, created=%v", first["created"])
+	}
+}
+
+// TestStatsRecordsAPIPaging exercises the on-demand records endpoint the
+// dashboard scrolls with: a bounded first page plus a cursor that walks back
+// through the audit file.
+func TestStatsRecordsAPIPaging(t *testing.T) {
+	g := newTestGateway(t)
+	for i := 0; i < 250; i++ {
+		g.stats.Record(Req{
+			Time: time.Now().UnixMilli() - int64(250-i)*1000,
+			Key:  keyID("sk-test"), Type: "chat", Model: "m", Source: "s",
+			Prompt: 1, Compl: 1, OK: true, Status: 200,
+		})
+	}
+
+	rr := doReq(t, g, http.MethodGet, "/api/stats/records?limit=50", "")
+	if rr.Code != 200 {
+		t.Fatalf("status = %d body=%s", rr.Code, rr.Body.String())
+	}
+	var page struct {
+		Records    []Req  `json:"records"`
+		NextCursor string `json:"next_cursor"`
+		HasMore    bool   `json:"has_more"`
+	}
+	if err := json.Unmarshal(rr.Body.Bytes(), &page); err != nil {
+		t.Fatalf("decode: %v", err)
+	}
+	if len(page.Records) != 50 {
+		t.Fatalf("first page = %d records, want 50", len(page.Records))
+	}
+	if !page.HasMore || page.NextCursor == "" {
+		t.Fatal("a long history must advertise more pages plus a cursor")
+	}
+	// newest first
+	for i := 1; i < len(page.Records); i++ {
+		if page.Records[i-1].Time < page.Records[i].Time {
+			t.Fatalf("page not newest-first at %d", i)
+		}
+	}
+
+	// follow the cursor: the next page must continue strictly older
+	oldest := page.Records[len(page.Records)-1].Time
+	rr2 := doReq(t, g, http.MethodGet,
+		"/api/stats/records?limit=50&before="+url.QueryEscape(page.NextCursor), "")
+	if rr2.Code != 200 {
+		t.Fatalf("page 2 status = %d", rr2.Code)
+	}
+	var page2 struct {
+		Records []Req `json:"records"`
+	}
+	if err := json.Unmarshal(rr2.Body.Bytes(), &page2); err != nil {
+		t.Fatalf("decode page 2: %v", err)
+	}
+	if len(page2.Records) == 0 {
+		t.Fatal("cursor page must return records")
+	}
+	if page2.Records[0].Time > oldest {
+		t.Fatalf("page 2 starts at %d, must continue below %d", page2.Records[0].Time, oldest)
+	}
+}
+
+// TestStatsSnapshotShipsOneScreen: a dashboard poll must not serialize the whole
+// ring, otherwise the "load the first screen only" contract is broken on the
+// wire even if the UI pages.
+func TestStatsSnapshotShipsOneScreen(t *testing.T) {
+	g := newTestGateway(t)
+	for i := 0; i < 400; i++ {
+		g.stats.Record(Req{
+			Time: time.Now().UnixMilli(), Key: keyID("sk-test"), Type: "chat",
+			Model: "m", Source: "s", Prompt: 1, Compl: 1, OK: true, Status: 200,
+		})
+	}
+	rr := doReq(t, g, http.MethodGet, "/api/stats", "")
+	var snap struct {
+		Records []Req `json:"records"`
+		Total   Stat  `json:"total"`
+	}
+	if err := json.Unmarshal(rr.Body.Bytes(), &snap); err != nil {
+		t.Fatalf("decode: %v", err)
+	}
+	if len(snap.Records) > firstScreenRecords {
+		t.Fatalf("snapshot shipped %d records, want <= %d", len(snap.Records), firstScreenRecords)
+	}
+	// aggregates still cover every request
+	if snap.Total.Reqs != 400 {
+		t.Fatalf("total reqs = %d, want 400 (aggregates must not be paged away)", snap.Total.Reqs)
+	}
+}
+
+// TestStatsCsvExportStreams checks the export still emits every row in the
+// window now that it is written straight from the audit walk.
+func TestStatsCsvExportStreams(t *testing.T) {
+	g := newTestGateway(t)
+	base := time.Now().UnixMilli()
+	for i := 0; i < 120; i++ {
+		g.stats.Record(Req{
+			Time: base + int64(i)*1000, Key: keyID("sk-test"), Type: "chat",
+			Model: "m", Source: "s", Prompt: 2, Compl: 3, OK: true, Status: 200,
+		})
+	}
+	rr := doReq(t, g, http.MethodGet,
+		"/api/stats?export=csv&from="+strconv.FormatInt(base, 10)+
+			"&to="+strconv.FormatInt(base+120*1000, 10), "")
+	if rr.Code != 200 {
+		t.Fatalf("export status = %d", rr.Code)
+	}
+	lines := strings.Split(strings.TrimSpace(rr.Body.String()), "\n")
+	if len(lines) != 121 { // header + 120 rows
+		t.Fatalf("csv had %d lines, want 121 (header + 120 records)", len(lines))
+	}
+	if !strings.HasPrefix(lines[0], "time,key,key_name") {
+		t.Fatalf("unexpected csv header: %q", lines[0])
+	}
+}
+
+// TestStatsRecordsAPIScopedToOwnKey: a non-admin key must only page its own
+// records even when it asks for someone else's.
+func TestStatsRecordsAPIScopedToOwnKey(t *testing.T) {
+	g := newTestGateway(t)
+	mine := keyID("sk-test")
+	for i := 0; i < 5; i++ {
+		g.stats.Record(Req{Time: time.Now().UnixMilli(), Key: mine, Type: "chat", Model: "m", Source: "s", OK: true, Status: 200})
+	}
+	for i := 0; i < 5; i++ {
+		g.stats.Record(Req{Time: time.Now().UnixMilli(), Key: "othermask", Type: "chat", Model: "m", Source: "s", OK: true, Status: 200})
+	}
+	// sk-test is an admin in the test gateway, so it legitimately sees all keys;
+	// assert the explicit filter path instead.
+	rr := doReq(t, g, http.MethodGet, "/api/stats/records?limit=100&key=othermask", "")
+	var page struct {
+		Records []Req `json:"records"`
+	}
+	if err := json.Unmarshal(rr.Body.Bytes(), &page); err != nil {
+		t.Fatalf("decode: %v", err)
+	}
+	for _, r := range page.Records {
+		if r.Key != "othermask" {
+			t.Fatalf("key filter leaked %q", r.Key)
+		}
+	}
+}
+
+// TestKeyDeleteRoundTrip is the end-to-end version of the bug users hit:
+// deleting a gateway key from the admin UI. It also asserts the GET-on-a-key
+// error names the right verb, since that message is what a mis-methoded client
+// actually sees.
+func TestKeyDeleteRoundTrip(t *testing.T) {
+	g := newTestGateway(t)
+
+	rr := doReq(t, g, http.MethodPost, "/api/keys", `{"name":"doomed","role":"user"}`)
+	if rr.Code != 200 {
+		t.Fatalf("create: %d %s", rr.Code, rr.Body.String())
+	}
+	var created struct {
+		Key struct {
+			Key  string `json:"key"`
+			Name string `json:"name"`
+		} `json:"key"`
+	}
+	if err := json.Unmarshal(rr.Body.Bytes(), &created); err != nil {
+		t.Fatalf("decode create: %v", err)
+	}
+	if created.Key.Key == "" {
+		t.Fatalf("no key returned: %s", rr.Body.String())
+	}
+
+	// DELETE removes it
+	rr = doReq(t, g, http.MethodDelete, "/api/keys/"+url.PathEscape(created.Key.Key), "")
+	if rr.Code != 200 {
+		t.Fatalf("delete: %d %s", rr.Code, rr.Body.String())
+	}
+	for _, k := range g.core.ListKeys() {
+		if k.Key == created.Key.Key {
+			t.Fatal("key still present after DELETE")
+		}
+	}
+
+	// deleting again is a clean 404, not a 500
+	rr = doReq(t, g, http.MethodDelete, "/api/keys/"+url.PathEscape(created.Key.Key), "")
+	if rr.Code != http.StatusNotFound {
+		t.Fatalf("second delete: %d, want 404", rr.Code)
+	}
+}
+
+// TestKeyGetOnSpecificKeyNamesTheVerb: a GET on /api/keys/{key} is what a
+// client that forgot to set method:"DELETE" ends up sending. The error must say
+// which verb to use instead of just "use GET /api/keys", which reads as if the
+// caller did nothing wrong.
+func TestKeyGetOnSpecificKeyNamesTheVerb(t *testing.T) {
+	g := newTestGateway(t)
+	rr := doReq(t, g, http.MethodGet, "/api/keys/sk-whatever", "")
+	if rr.Code != http.StatusNotFound {
+		t.Fatalf("status = %d, want 404", rr.Code)
+	}
+	body := rr.Body.String()
+	for _, want := range []string{"DELETE /api/keys/", "PUT /api/keys/"} {
+		if !strings.Contains(body, want) {
+			t.Errorf("error message should mention %q, got: %s", want, body)
+		}
+	}
+}
+
+// TestAdapterGetOnSpecificAdapterNamesDelete is the same guard for the adapters
+// endpoint, whose delAdapter helper had the identical missing-method bug.
+func TestAdapterGetOnSpecificAdapterNamesDelete(t *testing.T) {
+	g := newTestGateway(t)
+	rr := doReq(t, g, http.MethodGet, "/api/adapters/openai", "")
+	if rr.Code != http.StatusNotFound {
+		t.Fatalf("status = %d, want 404", rr.Code)
+	}
+	if !strings.Contains(rr.Body.String(), "DELETE /api/adapters/openai") {
+		t.Errorf("error should name the DELETE verb, got: %s", rr.Body.String())
+	}
+}
+
+// TestCannotDeleteOwnKey keeps the admin from locking themselves out.
+func TestCannotDeleteOwnKey(t *testing.T) {
+	g := newTestGateway(t)
+	rr := doReq(t, g, http.MethodDelete, "/api/keys/sk-test", "")
+	if rr.Code != http.StatusBadRequest {
+		t.Fatalf("status = %d, want 400", rr.Code)
+	}
+	if !strings.Contains(rr.Body.String(), "logged in with") {
+		t.Errorf("unexpected message: %s", rr.Body.String())
+	}
+}
+
+// TestGatewayRingStaysSmall pins the resident record ring at the small default.
+// The gateway used to construct Stats with a 10000-record ring, which after the
+// full-history aggregate scan meant 10000 resident Req structs — putting back the
+// ~25 MB the on-demand paging work removed. Aggregates cover all of history;
+// the ring only needs the first screen plus the 5-minute status windows.
+func TestGatewayRingStaysSmall(t *testing.T) {
+	g := newTestGateway(t)
+	g.stats.mu.Lock()
+	max := g.stats.maxRecs
+	g.stats.mu.Unlock()
+	if max != defaultRingSize {
+		t.Fatalf("gateway record ring = %d, want %d (a big ring reintroduces the resident-memory cost)",
+			max, defaultRingSize)
 	}
 }

internal/gateway/key_quota_api_test.go

@@ -0,0 +1,253 @@
+package gateway
+
+import (
+	"encoding/json"
+	"fmt"
+	"net/http"
+	"net/http/httptest"
+	"os"
+	"strings"
+	"testing"
+
+	"llmsproxy/internal/config"
+	"llmsproxy/internal/core"
+)
+
+// adminGateway builds a gateway whose admin key can call /api/keys.
+func adminGateway(t *testing.T, keys ...config.GWKey) *Gateway {
+	t.Helper()
+	td := t.TempDir()
+	cfgPath := td + "/config.yaml"
+	if err := os.WriteFile(cfgPath, []byte("listen: :0"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	cfg := &config.Config{
+		Path:        cfgPath,
+		AdapterDir:  td + "/adapters",
+		RuntimeFile: td + "/runtime.json",
+		Keys:        keys,
+	}
+	if err := cfg.ApplyDefaults(); err != nil {
+		t.Fatal(err)
+	}
+	c, err := core.NewFromConfig(cfg)
+	if err != nil {
+		t.Fatalf("core: %v", err)
+	}
+	t.Cleanup(c.Close)
+	g, err := New(c, []string{"sk-admin"})
+	if err != nil {
+		t.Fatalf("gateway: %v", err)
+	}
+	return g
+}
+
+func adminReq(t *testing.T, g *Gateway, method, path, body string) *httptest.ResponseRecorder {
+	t.Helper()
+	req, _ := http.NewRequest(method, path, strings.NewReader(body))
+	req.Header.Set("Authorization", "Bearer sk-admin")
+	req.Header.Set("Content-Type", "application/json")
+	rr := httptest.NewRecorder()
+	g.Handler().ServeHTTP(rr, req)
+	return rr
+}
+
+// keyRecord pulls one key's stored record out of the admin list.
+func keyRecord(t *testing.T, g *Gateway, secret string) config.GWKey {
+	t.Helper()
+	rr := adminReq(t, g, "GET", "/api/keys", "")
+	if rr.Code != 200 {
+		t.Fatalf("GET /api/keys: %d %s", rr.Code, rr.Body.String())
+	}
+	var out struct {
+		Keys []config.GWKey `json:"keys"`
+	}
+	if err := json.Unmarshal(rr.Body.Bytes(), &out); err != nil {
+		t.Fatalf("decode: %v (%s)", err, rr.Body.String())
+	}
+	for _, k := range out.Keys {
+		if k.Key == secret {
+			return k
+		}
+	}
+	t.Fatalf("key %q not found in %s", secret, rr.Body.String())
+	return config.GWKey{}
+}
+
+func scopeOf(t *testing.T, k config.GWKey, model string) config.ModelScope {
+	t.Helper()
+	for _, m := range k.Models {
+		if m.Model == model {
+			return m
+		}
+	}
+	t.Fatalf("scope %q not found in %+v", model, k.Models)
+	return config.ModelScope{}
+}
+
+// Quotas live on the scope entries, not on the key: creating a key with a
+// budget means creating scopes that carry it, and they must survive a
+// read-back (persisted, not just echoed).
+func TestKeyAPICreatesPerModelQuota(t *testing.T) {
+	g := adminGateway(t, config.GWKey{Key: "sk-admin", Role: "admin"})
+	rr := adminReq(t, g, "POST", "/api/keys",
+		`{"name":"agent-x","role":"user","models":[{"model":"m1","token_quota":50000,"req_quota":200,"period":"nhour","hours":6}]}`)
+	if rr.Code != 200 {
+		t.Fatalf("create: %d %s", rr.Code, rr.Body.String())
+	}
+	var created struct {
+		Key config.GWKey `json:"key"`
+	}
+	if err := json.Unmarshal(rr.Body.Bytes(), &created); err != nil {
+		t.Fatalf("decode: %v", err)
+	}
+	sc := scopeOf(t, created.Key, "m1")
+	if sc.TokenQuota != 50000 || sc.ReqQuota != 200 || sc.Period != "nhour" || sc.Hours != 6 {
+		t.Fatalf("created scope did not carry the caps: %+v", sc)
+	}
+	back := scopeOf(t, keyRecord(t, g, created.Key.Key), "m1")
+	if back.TokenQuota != 50000 || back.ReqQuota != 200 || back.Period != "nhour" || back.Hours != 6 {
+		t.Errorf("read-back lost the caps: %+v", back)
+	}
+}
+
+// Two models on one key carry independent budgets.
+func TestKeyAPIKeepsPerModelQuotaIndependent(t *testing.T) {
+	g := adminGateway(t, config.GWKey{Key: "sk-admin", Role: "admin"})
+	rr := adminReq(t, g, "POST", "/api/keys",
+		`{"name":"agent-y","role":"user","models":[{"model":"m1","token_quota":1000,"period":"hour"},{"model":"m2","token_quota":9999,"req_quota":7,"period":"week"}]}`)
+	if rr.Code != 200 {
+		t.Fatalf("create: %d %s", rr.Code, rr.Body.String())
+	}
+	var created struct {
+		Key config.GWKey `json:"key"`
+	}
+	_ = json.Unmarshal(rr.Body.Bytes(), &created)
+	back := keyRecord(t, g, created.Key.Key)
+	if a := scopeOf(t, back, "m1"); a.TokenQuota != 1000 || a.ReqQuota != 0 || a.Period != "hour" {
+		t.Errorf("m1 caps wrong: %+v", a)
+	}
+	if b := scopeOf(t, back, "m2"); b.TokenQuota != 9999 || b.ReqQuota != 7 || b.Period != "week" {
+		t.Errorf("m2 caps wrong: %+v", b)
+	}
+}
+
+// Sending 0 explicitly lifts that model's cap.
+func TestKeyAPIUpdateZeroLiftsCap(t *testing.T) {
+	g := adminGateway(t, config.GWKey{Key: "sk-admin", Role: "admin"})
+	rr := adminReq(t, g, "POST", "/api/keys",
+		`{"name":"z","role":"user","models":[{"model":"m1","token_quota":1000,"period":"hour"}]}`)
+	if rr.Code != 200 {
+		t.Fatalf("create: %d %s", rr.Code, rr.Body.String())
+	}
+	var created struct {
+		Key config.GWKey `json:"key"`
+	}
+	_ = json.Unmarshal(rr.Body.Bytes(), &created)
+	secret := created.Key.Key
+
+	rr = adminReq(t, g, "PUT", "/api/keys/"+secret,
+		`{"models":[{"model":"m1","token_quota":0,"req_quota":0,"period":""}]}`)
+	if rr.Code != 200 {
+		t.Fatalf("lift: %d %s", rr.Code, rr.Body.String())
+	}
+	if back := scopeOf(t, keyRecord(t, g, secret), "m1"); back.TokenQuota != 0 || back.ReqQuota != 0 || back.Period != "" {
+		t.Errorf("caps not lifted: %+v", back)
+	}
+}
+
+// A misspelled period must be refused, not quietly turned into an all-time
+// quota — which is the exact opposite of what the operator typed.
+func TestKeyAPIRejectsBadPeriod(t *testing.T) {
+	g := adminGateway(t, config.GWKey{Key: "sk-admin", Role: "admin"})
+	rr := adminReq(t, g, "POST", "/api/keys",
+		`{"name":"bad","role":"user","models":[{"model":"m1","token_quota":1000,"period":"houre"}]}`)
+	if rr.Code != http.StatusBadRequest {
+		t.Fatalf("want 400 for a bad period, got %d %s", rr.Code, rr.Body.String())
+	}
+	if !strings.Contains(rr.Body.String(), "period") {
+		t.Errorf("error should name the period field: %s", rr.Body.String())
+	}
+}
+
+func TestKeyAPIRejectsNegativeQuota(t *testing.T) {
+	g := adminGateway(t, config.GWKey{Key: "sk-admin", Role: "admin"})
+	rr := adminReq(t, g, "POST", "/api/keys",
+		`{"name":"bad","role":"user","models":[{"model":"m1","token_quota":-5}]}`)
+	if rr.Code != http.StatusBadRequest {
+		t.Fatalf("want 400 for a negative quota, got %d %s", rr.Code, rr.Body.String())
+	}
+}
+
+// The rejection must say WHICH model is over budget, so an operator looking at
+// a key with a dozen scopes can tell which one to raise.
+func TestKeyAPIRejectionNamesTheModel(t *testing.T) {
+	g := adminGateway(t, config.GWKey{Key: "sk-admin", Role: "admin"})
+	rr := adminReq(t, g, "POST", "/api/keys",
+		`{"name":"bad","role":"user","models":[{"model":"m1","req_quota":-1}]}`)
+	if rr.Code != http.StatusBadRequest {
+		t.Fatalf("want 400, got %d", rr.Code)
+	}
+	if !strings.Contains(rr.Body.String(), "m1") {
+		t.Errorf("error should name the offending model: %s", rr.Body.String())
+	}
+}
+
+// A non-admin key must not be able to mint keys.
+func TestKeyAPIQuotaIsAdminOnly(t *testing.T) {
+	g := adminGateway(t,
+		config.GWKey{Key: "sk-admin", Role: "admin"},
+		config.GWKey{Key: "sk-u", Role: "user", Models: []config.ModelScope{{Model: "m1", TokenQuota: 10, Period: "hour"}}},
+	)
+	req, _ := http.NewRequest("POST", "/api/keys",
+		strings.NewReader(`{"name":"x","role":"admin","models":[{"model":"m1"}]}`))
+	req.Header.Set("Authorization", "Bearer sk-u")
+	req.Header.Set("Content-Type", "application/json")
+	rr := httptest.NewRecorder()
+	g.Handler().ServeHTTP(rr, req)
+	if rr.Code != http.StatusForbidden {
+		t.Fatalf("non-admin create: want 403, got %d %s", rr.Code, rr.Body.String())
+	}
+	// /api/v1/keys exposes the per-model caps but never a secret
+	rr = adminReq(t, g, "GET", "/api/v1/keys", "")
+	if rr.Code != 200 {
+		t.Fatalf("GET /api/v1/keys: %d", rr.Code)
+	}
+	if strings.Contains(rr.Body.String(), "sk-u") {
+		t.Error("/api/v1/keys leaked a key secret")
+	}
+	if !strings.Contains(rr.Body.String(), `"token_quota":10`) {
+		t.Errorf("/api/v1/keys should expose the per-model caps: %s", rr.Body.String())
+	}
+}
+
+// A user must be able to see their own budget: /api/keys/me is the only key
+// view a non-admin gets, so a cap missing from it is invisible to the very
+// client it constrains.
+func TestKeyMeExposesOwnQuota(t *testing.T) {
+	g := adminGateway(t,
+		config.GWKey{Key: "sk-admin", Role: "admin"},
+		config.GWKey{Key: "sk-u", Role: "user", Name: "agent",
+			Models: []config.ModelScope{{Model: "m1", TokenQuota: 123456, ReqQuota: 42, Period: "week"}}},
+	)
+	req, _ := http.NewRequest("GET", "/api/keys/me", nil)
+	req.Header.Set("Authorization", "Bearer sk-u")
+	rr := httptest.NewRecorder()
+	g.Handler().ServeHTTP(rr, req)
+	if rr.Code != 200 {
+		t.Fatalf("GET /api/keys/me: %d %s", rr.Code, rr.Body.String())
+	}
+	// the endpoint wraps the record: {"key": {...}}
+	var wrap struct {
+		Key config.GWKey `json:"key"`
+	}
+	if err := json.Unmarshal(rr.Body.Bytes(), &wrap); err != nil {
+		t.Fatalf("decode: %v (%s)", err, rr.Body.String())
+	}
+	sc := scopeOf(t, wrap.Key, "m1")
+	if sc.TokenQuota != 123456 || sc.ReqQuota != 42 || sc.Period != "week" {
+		t.Errorf("own quota not visible to the key's owner: %+v", sc)
+	}
+}
+
+var _ = fmt.Sprintf

internal/gateway/key_quota_concurrency_test.go

@@ -0,0 +1,60 @@
+package gateway
+
+import (
+	"sync"
+	"testing"
+	"time"
+)
+
+// The per-key quota buckets are shared maps mutated on every recorded request
+// and read on every quota check. Single-threaded unit tests cannot see a race
+// there at all, so this drives 64 goroutines through record + read together
+// under `go test -race`. It also opts one model into the lazily-created pinned
+// bucket mid-flight, which is the path that mutates two bucket maps at once.
+func TestConcurrentQuotaAccounting(t *testing.T) {
+	s := NewStats(0)
+	keys := []string{"k0", "k1", "k2", "k3", "k4", "k5", "k6"}
+	models := []string{"m0", "m1", "m2", "AUTO"}
+	var wg sync.WaitGroup
+	stop := time.Now().Add(2 * time.Second)
+
+	for w := 0; w < 64; w++ {
+		wg.Add(1)
+		go func(w int) {
+			defer wg.Done()
+			k := keys[w%len(keys)]
+			m := models[w%len(models)]
+			now := time.Now()
+			for time.Now().Before(stop) {
+				s.Record(Req{Time: now.UnixMilli(), Key: k, Model: m, Source: "src",
+					Prompt: 100, Compl: 50, OK: true, Status: 200, Type: "chat"})
+				_ = s.KeyWindowTokens(k, 3600)
+				_ = s.KeyWindowReqs(k, 3600)
+				_ = s.KeyWindowModelTokens(k, m, "", 3600)
+				_ = s.KeyWindowModelTokens(k, m, "src", 3600) // opts into the pinned bucket
+			}
+		}(w)
+	}
+	wg.Wait()
+
+	// Every recorded request must be accounted for exactly once.
+	s.mu.Lock()
+	var total int64
+	for _, hm := range s.keyHour {
+		for _, v := range hm {
+			total += v
+		}
+	}
+	recs := int64(0)
+	for _, r := range s.recs {
+		recs += r.Prompt + r.Compl
+	}
+	s.mu.Unlock()
+	if total == 0 {
+		t.Fatal("no tokens accounted")
+	}
+	t.Logf("accounted %d tokens across %d keys", total, len(keys))
+	if recs == 0 {
+		t.Error("no records retained; the concurrent writers lost work")
+	}
+}

internal/gateway/key_quota_perf_test.go

@@ -0,0 +1,125 @@
+package gateway
+
+import (
+	"math/rand"
+	"testing"
+	"time"
+)
+
+// Perf: the per-key quota bookkeeping runs on every recorded request and on
+// every quota check, so its cost lands directly on the request path. These
+// benchmarks exist to catch a regression that would make the feature
+// expensive; the numbers that matter are relative to each other and to the
+// pre-change baseline (Record was 275 ns/op with 3 allocs before this work).
+
+func BenchmarkRecordQuotaBucket(b *testing.B) {
+	s := NewStats(0)
+	r := Req{Time: time.Now().UnixMilli(), Key: "keyABC", Model: "deepseek-v4-flash",
+		Source: "deepseek", Prompt: 800, Compl: 200, OK: true, Status: 200, Type: "chat"}
+	b.ReportAllocs()
+	b.ResetTimer()
+	for i := 0; i < b.N; i++ {
+		s.Record(r)
+	}
+}
+
+// A key that has been busy for the whole 40-day retention.
+func BenchmarkRecordQuotaBucketFullRetention(b *testing.B) {
+	s := NewStats(0)
+	nowH := time.Now().Unix() / 3600
+	for h := int64(0); h < quotaRetentionHours; h++ {
+		s.Record(Req{Time: (nowH - h) * 3600 * 1000, Key: "keyABC", Model: "m1",
+			Source: "deepseek", Prompt: 100, Compl: 100, OK: true, Status: 200})
+	}
+	r := Req{Time: time.Now().UnixMilli(), Key: "keyABC", Model: "m1", Source: "deepseek",
+		Prompt: 100, Compl: 100, OK: true, Status: 200, Type: "chat"}
+	b.ReportAllocs()
+	b.ResetTimer()
+	for i := 0; i < b.N; i++ {
+		s.Record(r)
+	}
+}
+
+// The whole per-request quota check: total tokens + request count.
+func BenchmarkQuotaCheckHourly(b *testing.B) {
+	s := NewStats(0)
+	for h := 0; h < 24; h++ {
+		s.Record(Req{Time: time.Now().Add(-time.Duration(h) * time.Hour).UnixMilli(),
+			Key: "keyABC", Model: "m1", Source: "deepseek", Prompt: 100, Compl: 100, OK: true, Status: 200})
+	}
+	b.ReportAllocs()
+	b.ResetTimer()
+	for i := 0; i < b.N; i++ {
+		s.KeyWindowTokens("keyABC", 3600)
+		s.KeyWindowReqs("keyABC", 3600)
+	}
+}
+
+// A 30-day window has to visit 720 hour buckets. This is the worst case the
+// quota feature puts on the request path.
+func BenchmarkQuotaCheckMonthly(b *testing.B) {
+	s := NewStats(0)
+	nowH := time.Now().Unix() / 3600
+	for h := int64(0); h < 720; h++ {
+		s.Record(Req{Time: (nowH - h) * 3600 * 1000, Key: "keyABC", Model: "m1",
+			Prompt: 100, Compl: 100, OK: true, Status: 200})
+	}
+	b.ReportAllocs()
+	b.ResetTimer()
+	for i := 0; i < b.N; i++ {
+		s.KeyWindowTokens("keyABC", 30*24*3600)
+	}
+}
+
+// TestSumBucketsMatchesFullScan is the correctness guard for the optimization
+// that made the quota check proportional to the window instead of to the whole
+// retention: the windowed sum must agree with a full scan for every bucket
+// placement, or a quota would silently start letting traffic through. It
+// caught a real off-by-one (floor instead of ceil on the first hour).
+func TestSumBucketsMatchesFullScan(t *testing.T) {
+	const hour = 3600
+	base := int64(1_700_000_000) / hour * hour
+	rng := rand.New(rand.NewSource(7))
+	for trial := 0; trial < 400; trial++ {
+		hm := map[int64]int64{}
+		for i, n := 0, rng.Intn(40)+1; i < n; i++ {
+			h := base/hour - int64(rng.Intn(1200))
+			hm[h] += int64(rng.Intn(1000)) + 1
+		}
+		now := base + int64(rng.Intn(hour))
+		for _, sec := range []int64{3600, 2 * 3600, 6 * 3600, 24 * 3600, 7 * 24 * 3600, 30 * 24 * 3600} {
+			var want int64
+			cut := now - sec
+			for h, v := range hm {
+				if h*hour >= cut {
+					want += v
+				}
+			}
+			if got := sumBuckets(hm, now, sec); got != want {
+				t.Fatalf("trial %d sec=%d: sumBuckets=%d, full scan=%d", trial, sec, got, want)
+			}
+		}
+	}
+}
+
+// The pinned bucket is maintained only for (key, model) pairs whose quota pins
+// a source. Doing it unconditionally doubled the bucket count — measured at
+// 26.7 MB for 20 keys x 8 models x 3 sources at full retention, against a
+// documented ~32 MB total memory budget — to serve a lookup nobody performs.
+func TestPinnedBucketNotMaintainedUnlessWanted(t *testing.T) {
+	s := NewStats(100)
+	now := time.Now().UnixMilli()
+	for i := 0; i < 5; i++ {
+		s.Record(Req{Time: now, Key: "keyA", Model: "m1", Source: "srcX", Prompt: 100, Compl: 100, OK: true, Status: 200})
+	}
+	s.mu.Lock()
+	_, pinned := s.keyModelHour["keyA"]["srcX::m1"]
+	_, bare := s.keyModelHour["keyA"]["m1"]
+	s.mu.Unlock()
+	if pinned {
+		t.Error("a pinned bucket exists before any quota read the pin — memory regression")
+	}
+	if !bare {
+		t.Error("the bare-model bucket must always exist; an unpinned quota depends on it")
+	}
+}

internal/gateway/key_quota_test.go

@@ -0,0 +1,225 @@
+package gateway
+
+import (
+	"testing"
+	"time"
+
+	"llmsproxy/internal/config"
+)
+
+// ---- data layer: per-key isolation ----
+
+func TestKeyWindowTokensIsolatesKeys(t *testing.T) {
+	s := NewStats(100)
+	now := time.Now().UnixMilli()
+	s.Record(Req{Time: now, Key: "keyA", Model: "m1", Prompt: 500, Compl: 500, OK: true, Status: 200})
+	s.Record(Req{Time: now, Key: "keyB", Model: "m1", Prompt: 500, Compl: 500, OK: true, Status: 200})
+
+	if got := s.KeyWindowModelTokens("keyA", "m1", "", 3600); got != 1000 {
+		t.Errorf("keyA model tokens = %d, want 1000 (its own usage only)", got)
+	}
+	if got := s.KeyWindowModelTokens("keyB", "m1", "", 3600); got != 1000 {
+		t.Errorf("keyB model tokens = %d, want 1000", got)
+	}
+	// the key-blind bucket stays global on purpose: it backs the AUTO slot
+	// quota, which limits the whole gateway, not one key.
+	if got := s.WindowTokens("m1", "", 3600); got != 2000 {
+		t.Errorf("WindowTokens = %d, want 2000 (global per-model total must be unchanged)", got)
+	}
+}
+
+// A source pin scopes the cap to that one upstream. The pinned bucket is
+// maintained lazily (it is only read by quotas that actually pin a source), so
+// usage recorded before anything queried the pin is not attributed to it.
+func TestKeyWindowTokensSeparatesSourcePin(t *testing.T) {
+	s := NewStats(100)
+	now := time.Now().UnixMilli()
+	s.Record(Req{Time: now, Key: "keyA", Model: "m1", Source: "srcX", Prompt: 100, Compl: 100, OK: true, Status: 200}) // 200 tok
+	s.Record(Req{Time: now, Key: "keyA", Model: "m1", Source: "srcY", Prompt: 200, Compl: 200, OK: true, Status: 200}) // 400 tok
+
+	// The unpinned bucket always counts — it is what a scope entry without a
+	// source pin reads. 200 + 400.
+	if got := s.KeyWindowModelTokens("keyA", "m1", "", 3600); got != 600 {
+		t.Errorf("unpinned model tokens = %d, want 600 (both sources)", got)
+	}
+
+	// Reading a pin opts this (key, model) into pinned accounting.
+	if got := s.KeyWindowModelTokens("keyA", "m1", "srcX", 3600); got != 0 {
+		t.Errorf("srcX pin before opt-in = %d, want 0 (the lazy bucket has not accrued yet)", got)
+	}
+
+	// From now on both pins accrue.
+	more := now + 1
+	s.Record(Req{Time: more, Key: "keyA", Model: "m1", Source: "srcX", Prompt: 10, Compl: 10, OK: true, Status: 200}) // 20 tok
+	s.Record(Req{Time: more, Key: "keyA", Model: "m1", Source: "srcY", Prompt: 30, Compl: 30, OK: true, Status: 200}) // 60 tok
+	if got := s.KeyWindowModelTokens("keyA", "m1", "srcX", 3600); got != 20 {
+		t.Errorf("srcX-pinned tokens after opt-in = %d, want 20", got)
+	}
+	if got := s.KeyWindowModelTokens("keyA", "m1", "srcY", 3600); got != 60 {
+		t.Errorf("srcY-pinned tokens = %d, want 60 (a pin read for one source must not blind the other)", got)
+	}
+}
+
+// This is the shape of every real chat record: Source is always populated.
+// The unpinned bucket must still see the tokens, or a per-model quota without
+// a source pin reads an empty bucket and never trips.
+func TestKeyWindowModelTokensUnpinnedSeesSourcedTraffic(t *testing.T) {
+	s := NewStats(100)
+	s.Record(Req{Time: time.Now().UnixMilli(), Key: "keyA", Model: "m1", Source: "deepseek",
+		Prompt: 1000, Compl: 1000, OK: true, Status: 200})
+
+	if got := s.KeyWindowModelTokens("keyA", "m1", "", 3600); got != 2000 {
+		t.Errorf("unpinned tokens = %d, want 2000 — an unpinned per-model quota would never trip otherwise", got)
+	}
+}
+
+func TestKeyWindowRespectsResetWindow(t *testing.T) {
+	s := NewStats(100)
+	old := time.Now().Add(-30 * 24 * time.Hour).UnixMilli()
+	s.Record(Req{Time: old, Key: "keyA", Model: "m1", Prompt: 100, Compl: 100, OK: true, Status: 200})
+	s.Record(Req{Time: time.Now().UnixMilli(), Key: "keyA", Model: "m1", Prompt: 5, Compl: 5, OK: true, Status: 200})
+
+	if got := s.KeyWindowTokens("keyA", 3600); got != 10 {
+		t.Errorf("1h window = %d, want 10 (30-day-old usage must not count)", got)
+	}
+	if got := s.KeyWindowTokens("keyA", 0); got != 210 {
+		t.Errorf("all-time total = %d, want 210", got)
+	}
+}
+
+func TestKeyWindowReqsCountsEveryRequest(t *testing.T) {
+	s := NewStats(100)
+	now := time.Now().UnixMilli()
+	// a failed request and an image request both count: a client that loops on
+	// failures must still burn its request quota
+	s.Record(Req{Time: now, Key: "keyA", Model: "m1", Type: "chat", OK: true, Status: 200})
+	s.Record(Req{Time: now, Key: "keyA", Model: "m1", Type: "chat", OK: false, Status: 500})
+	s.Record(Req{Time: now, Key: "keyA", Model: "img", Type: "image", OK: true, Status: 200})
+	s.Record(Req{Time: now, Key: "keyB", Model: "m1", Type: "chat", OK: true, Status: 200})
+
+	if got := s.KeyWindowReqs("keyA", 3600); got != 3 {
+		t.Errorf("keyA reqs = %d, want 3 (chat ok + chat fail + image)", got)
+	}
+	if got := s.KeyWindowReqs("keyB", 3600); got != 1 {
+		t.Errorf("keyB reqs = %d, want 1", got)
+	}
+}
+
+func TestKeyWindowTokensZeroTokensNotCounted(t *testing.T) {
+	s := NewStats(100)
+	now := time.Now().UnixMilli()
+	// a request that reported no usage must not create a bucket entry
+	s.Record(Req{Time: now, Key: "keyA", Model: "", Type: "chat", OK: true, Status: 200})
+	if got := s.KeyWindowTokens("keyA", 3600); got != 0 {
+		t.Errorf("tokens = %d, want 0", got)
+	}
+	if got := s.KeyWindowReqs("keyA", 3600); got != 1 {
+		t.Errorf("reqs = %d, want 1 (request still happened)", got)
+	}
+}
+
+// ---- reset window arithmetic ----
+
+func TestAutoSecondsToReset(t *testing.T) {
+	if got := AutoSecondsToReset("", 0); got != 0 {
+		t.Errorf("no period = %d, want 0 (never resets -> no retry hint)", got)
+	}
+	now := time.Now().Unix()
+	for _, tc := range []struct {
+		period string
+		hours  int64
+		want   int64
+	}{
+		{"hour", 0, 3600},
+		{"week", 0, 7 * 24 * 3600},
+		{"month", 0, 30 * 24 * 3600},
+		{"nhour", 6, 6 * 3600},
+	} {
+		got := AutoSecondsToReset(tc.period, tc.hours)
+		if got <= 0 || got > tc.want {
+			t.Errorf("AutoSecondsToReset(%q,%d) = %d, want in (0,%d]", tc.period, tc.hours, got, tc.want)
+		}
+		// must never exceed the window itself
+		if got < tc.want-now%3600 {
+			t.Logf("note: %q hint %ds < remaining-window %ds (rounds to hour boundary)", tc.period, got, tc.want-now%3600)
+		}
+		if got > tc.want {
+			t.Errorf("hint %d exceeds window %d", got, tc.want)
+		}
+	}
+	_ = now
+}
+
+func TestAutoSecondsToResetNeverExceedsWindow(t *testing.T) {
+	// "nhour" with a tiny window must not hand out a longer wait than the
+	// window itself (which would stall a client past its own reset)
+	for hours := int64(1); hours <= 48; hours++ {
+		got := AutoSecondsToReset("nhour", hours)
+		if got <= 0 || got > hours*3600 {
+			t.Errorf("nhour/%d = %d, want in (0,%d]", hours, got, hours*3600)
+		}
+	}
+}
+
+// ---- config validation ----
+
+func TestKeyQuotaValidate(t *testing.T) {
+	cases := []struct {
+		name    string
+		q       config.KeyQuota
+		wantErr bool
+	}{
+		{"unlimited", config.KeyQuota{}, false},
+		{"tokens hourly", config.KeyQuota{TokenQuota: 1000, Period: "hour"}, false},
+		{"tokens n-hour", config.KeyQuota{TokenQuota: 1000, Period: "nhour", Hours: 6}, false},
+		{"reqs weekly", config.KeyQuota{ReqQuota: 100, Period: "week"}, false},
+		{"no period never resets", config.KeyQuota{TokenQuota: 1000}, false},
+		{"negative tokens", config.KeyQuota{TokenQuota: -1}, true},
+		{"negative reqs", config.KeyQuota{ReqQuota: -1}, true},
+		{"negative hours", config.KeyQuota{TokenQuota: 5, Hours: -1}, true},
+		{"typo period becomes all-time if accepted", config.KeyQuota{TokenQuota: 5, Period: "houre"}, true},
+		{"nhour without hours", config.KeyQuota{TokenQuota: 5, Period: "nhour"}, true},
+		{"nhour with 0 hours", config.KeyQuota{TokenQuota: 5, Period: "nhour", Hours: 0}, true},
+		{"junk period with only req quota", config.KeyQuota{ReqQuota: 5, Period: "daily"}, true},
+		{"junk period with no caps is irrelevant", config.KeyQuota{Period: "daily"}, false},
+	}
+	for _, tc := range cases {
+		err := tc.q.Validate()
+		if tc.wantErr && err == nil {
+			t.Errorf("%s: expected error, got nil", tc.name)
+		}
+		if !tc.wantErr && err != nil {
+			t.Errorf("%s: unexpected error: %v", tc.name, err)
+		}
+	}
+}
+
+func TestNormalizeRole(t *testing.T) {
+	if got := config.NormalizeRole(""); got != "user" {
+		t.Errorf("empty role = %q, want user", got)
+	}
+	if got := config.NormalizeRole("admin"); got != "admin" {
+		t.Errorf("admin role = %q, want admin", got)
+	}
+	if got := config.NormalizeRole("root"); got != "user" {
+		t.Errorf("unknown role = %q, want user (never escalate)", got)
+	}
+}
+
+// Retention must hold for the lazy pinned buckets too, or a long-lived key
+// would grow without bound.
+func TestPinnedBucketRespectsRetention(t *testing.T) {
+	s := NewStats(100)
+	_ = s.KeyWindowModelTokens("keyA", "m1", "srcX", 3600) // opt in
+	nowH := time.Now().Unix() / 3600
+	for h := int64(0); h < quotaRetentionHours+50; h++ {
+		s.Record(Req{Time: (nowH - h) * 3600 * 1000, Key: "keyA", Model: "m1", Source: "srcX",
+			Prompt: 10, Compl: 10, OK: true, Status: 200})
+	}
+	s.mu.Lock()
+	n := len(s.keyModelHour["keyA"]["srcX::m1"])
+	s.mu.Unlock()
+	if n > quotaRetentionHours {
+		t.Errorf("pinned bucket holds %d hours, want <= %d", n, quotaRetentionHours)
+	}
+}

internal/gateway/key_quota_wiring_test.go

@@ -0,0 +1,444 @@
+package gateway
+
+import (
+	"context"
+	"encoding/json"
+	"fmt"
+	"net/http"
+	"net/http/httptest"
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+	"time"
+
+	"llmsproxy/internal/config"
+	"llmsproxy/internal/core"
+)
+
+// quotaGateway builds a gateway with two user keys that differ only in their
+// configured caps, against one mock upstream that reports 4 tokens per call.
+func quotaGateway(t *testing.T, keyA, keyB config.GWKey) (*Gateway, *upstreamCtrl) {
+	t.Helper()
+	ctrl := &upstreamCtrl{}
+	up := upstream(t, ctrl)
+	t.Cleanup(up.Close)
+
+	td := t.TempDir()
+	cfgPath := td + "/config.yaml"
+	if err := os.WriteFile(cfgPath, []byte("listen: :0"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	cfg := &config.Config{
+		Path:        cfgPath,
+		AdapterDir:  td + "/adapters",
+		RuntimeFile: td + "/runtime.json",
+		Keys:        []config.GWKey{keyA, keyB},
+		Sources: []config.Source{{
+			Name: "up", BaseURL: up.URL, Adapter: "openai",
+			Models: []config.Model{{ID: "m1", Priority: 100}},
+		}},
+	}
+	if err := cfg.ApplyDefaults(); err != nil {
+		t.Fatal(err)
+	}
+	c, err := core.NewFromConfig(cfg)
+	if err != nil {
+		t.Fatalf("core: %v", err)
+	}
+	t.Cleanup(c.Close)
+	g, err := New(c, []string{keyA.Key, keyB.Key})
+	if err != nil {
+		t.Fatalf("gateway: %v", err)
+	}
+	return g, ctrl
+}
+
+// chatAs issues a chat completion with a specific gateway key and returns the
+// recorder plus the decoded error code.
+func chatAs(t *testing.T, g *Gateway, key, model string) (*httptest.ResponseRecorder, string) {
+	t.Helper()
+	req, _ := http.NewRequest("POST", "/v1/chat/completions",
+		strings.NewReader(fmt.Sprintf(`{"model":%q,"messages":[{"role":"user","content":"hi"}]}`, model)))
+	req.Header.Set("Authorization", "Bearer "+key)
+	req.Header.Set("Content-Type", "application/json")
+	rr := httptest.NewRecorder()
+	g.Handler().ServeHTTP(rr, req)
+	code := ""
+	if rr.Code != 200 {
+		var e struct {
+			Error struct {
+				Type string `json:"type"`
+			} `json:"error"`
+		}
+		_ = json.Unmarshal(rr.Body.Bytes(), &e)
+		code = e.Error.Type
+	}
+	return rr, code
+}
+
+// A key whose token budget is spent must be refused, and the refusal must be a
+// 429 the client can retry after the window rolls over — not a 403 that reads
+// as "this key may never use this model".
+func TestKeyTokenQuotaBlocksWithRetryAfter(t *testing.T) {
+	g, _ := quotaGateway(t,
+		config.GWKey{Key: "sk-a", Role: "user",
+			Models: []config.ModelScope{{Model: "m1", TokenQuota: 8, Period: "hour"}}},
+		config.GWKey{Key: "sk-b", Role: "user", Name: "b"},
+	)
+	// 4 tokens per call, budget 8 -> the third call crosses it
+	for i := 1; i <= 2; i++ {
+		rr, _ := chatAs(t, g, "sk-a", "m1")
+		if rr.Code != 200 {
+			t.Fatalf("call %d: want 200, got %d (%s)", i, rr.Code, rr.Body.String())
+		}
+	}
+	rr, code := chatAs(t, g, "sk-a", "m1")
+	if rr.Code != http.StatusTooManyRequests {
+		t.Fatalf("third call: want 429, got %d (%s)", rr.Code, rr.Body.String())
+	}
+	if code != "rate_limit_exceeded" {
+		t.Errorf("error code = %q, want rate_limit_exceeded", code)
+	}
+	if ra := rr.Header().Get("Retry-After"); ra == "" {
+		t.Error("429 must carry Retry-After so a client knows when to come back")
+	} else if n := mustAtoi(t, ra); n <= 0 || n > 3600 {
+		t.Errorf("Retry-After = %q, want a positive value within the 1h window", ra)
+	}
+	if !strings.Contains(rr.Body.String(), "quota") {
+		t.Errorf("body should name the quota: %s", rr.Body.String())
+	}
+}
+
+// The quota is per key: exhausting one key's budget must not affect another
+// key that is allowed the same model.
+func TestKeyTokenQuotaIsIsolatedPerKey(t *testing.T) {
+	g, _ := quotaGateway(t,
+		config.GWKey{Key: "sk-a", Role: "user",
+			Models: []config.ModelScope{{Model: "m1", TokenQuota: 4, Period: "hour"}}},
+		config.GWKey{Key: "sk-b", Role: "user",
+			Models: []config.ModelScope{{Model: "m1", TokenQuota: 1000, Period: "hour"}}},
+	)
+	rr, _ := chatAs(t, g, "sk-a", "m1")
+	if rr.Code != 200 {
+		t.Fatalf("keyA first call: want 200, got %d", rr.Code)
+	}
+	if rr, code := chatAs(t, g, "sk-a", "m1"); rr.Code != http.StatusTooManyRequests {
+		t.Fatalf("keyA second call: want 429, got %d (%s)", rr.Code, code)
+	}
+	// keyB has its own budget and must still be served
+	if rr, _ := chatAs(t, g, "sk-b", "m1"); rr.Code != 200 {
+		t.Fatalf("keyB must be unaffected by keyA's exhausted quota, got %d", rr.Code)
+	}
+}
+
+// A per-model quota inside the scope must also be per key: one key burning a
+// shared model budget must not lock the other key out.
+func TestScopeModelTokenQuotaIsolatedPerKey(t *testing.T) {
+	g, _ := quotaGateway(t,
+		config.GWKey{Key: "sk-a", Role: "user", Models: []config.ModelScope{
+			{Model: "m1", TokenQuota: 4, Period: "hour"},
+		}},
+		config.GWKey{Key: "sk-b", Role: "user", Models: []config.ModelScope{
+			{Model: "m1", TokenQuota: 4, Period: "hour"},
+		}},
+	)
+	if rr, _ := chatAs(t, g, "sk-a", "m1"); rr.Code != 200 {
+		t.Fatalf("keyA first call: want 200, got %d", rr.Code)
+	}
+	if rr, _ := chatAs(t, g, "sk-a", "m1"); rr.Code != http.StatusTooManyRequests {
+		t.Fatalf("keyA second call: want 429 (its own model quota), got %d", rr.Code)
+	}
+	// keyB shares the model but not the counter — it must be served twice too
+	if rr, _ := chatAs(t, g, "sk-b", "m1"); rr.Code != 200 {
+		t.Fatalf("keyB first call: want 200, got %d", rr.Code)
+	}
+	if rr, _ := chatAs(t, g, "sk-b", "m1"); rr.Code != http.StatusTooManyRequests {
+		t.Fatalf("keyB second call: want 429, got %d", rr.Code)
+	}
+}
+
+// An admin key is never capped: a budget set on an admin key must not lock the
+// operator out of the gateway they administer.
+func TestAdminKeyIsNeverQuotaCapped(t *testing.T) {
+	g, _ := quotaGateway(t,
+		config.GWKey{Key: "sk-admin", Role: "admin",
+			Models: []config.ModelScope{{Model: "m1", TokenQuota: 1, ReqQuota: 1, Period: "hour"}}},
+		config.GWKey{Key: "sk-b", Role: "user"},
+	)
+	for i := 1; i <= 3; i++ {
+		if rr, _ := chatAs(t, g, "sk-admin", "m1"); rr.Code != 200 {
+			t.Fatalf("admin call %d: want 200 (admin keys are uncapped), got %d", i, rr.Code)
+		}
+	}
+}
+
+// A request quota caps sustained volume; RPM-style bursts are not its job but
+// the count must still stop the key.
+func TestKeyRequestQuotaBlocks(t *testing.T) {
+	g, ctrl := quotaGateway(t,
+		config.GWKey{Key: "sk-a", Role: "user",
+			Models: []config.ModelScope{{Model: "m1", ReqQuota: 2, Period: "hour"}}},
+		config.GWKey{Key: "sk-b", Role: "user"},
+	)
+	for i := 1; i <= 2; i++ {
+		if rr, _ := chatAs(t, g, "sk-a", "m1"); rr.Code != 200 {
+			t.Fatalf("call %d: want 200, got %d", i, rr.Code)
+		}
+	}
+	rr, code := chatAs(t, g, "sk-a", "m1")
+	if rr.Code != http.StatusTooManyRequests || code != "rate_limit_exceeded" {
+		t.Fatalf("third call: want 429/rate_limit_exceeded, got %d/%s", rr.Code, code)
+	}
+	if ctrl.hits != 2 {
+		t.Errorf("upstream saw %d calls, want 2 (a refused request must not reach upstream)", ctrl.hits)
+	}
+}
+
+// A model outside the scope is still 403, not 429: retrying cannot help.
+func TestModelOutsideScopeStaysForbidden(t *testing.T) {
+	g, _ := quotaGateway(t,
+		config.GWKey{Key: "sk-a", Role: "user",
+			Models: []config.ModelScope{{Model: "other-model", TokenQuota: 1000, Period: "hour"}}},
+		config.GWKey{Key: "sk-b", Role: "user"},
+	)
+	rr, code := chatAs(t, g, "sk-a", "m1")
+	if rr.Code != http.StatusForbidden {
+		t.Fatalf("want 403, got %d (%s)", rr.Code, rr.Body.String())
+	}
+	if code != "model_not_allowed" {
+		t.Errorf("error code = %q, want model_not_allowed", code)
+	}
+	if rr.Header().Get("Retry-After") != "" {
+		t.Error("a 403 must not advertise a retry time")
+	}
+}
+
+// An unlimited key (no caps at all) is never rejected — the default config
+// must keep working exactly as before.
+func TestKeyWithoutCapsIsUnlimited(t *testing.T) {
+	g, _ := quotaGateway(t,
+		config.GWKey{Key: "sk-a", Role: "user", Models: []config.ModelScope{{Model: "m1"}}},
+		config.GWKey{Key: "sk-b", Role: "user"},
+	)
+	for i := 1; i <= 5; i++ {
+		if rr, _ := chatAs(t, g, "sk-a", "m1"); rr.Code != 200 {
+			t.Fatalf("call %d: want 200 (no caps = unlimited), got %d", i, rr.Code)
+		}
+	}
+}
+
+func mustAtoi(t *testing.T, s string) int64 {
+	t.Helper()
+	var n int64
+	if _, err := fmt.Sscanf(s, "%d", &n); err != nil {
+		t.Fatalf("Retry-After %q is not an integer: %v", s, err)
+	}
+	return n
+}
+
+// quotaCtx builds a request context carrying a gateway key, as the auth
+// middleware would.
+func quotaCtx(t *testing.T, g *Gateway, key string) context.Context {
+	t.Helper()
+	return withAuth(context.Background(), key, "user")
+}
+
+func nowMSOffset(sec int64) int64 {
+	return time.Now().Add(time.Duration(sec) * time.Second).UnixMilli()
+}
+
+// The key-wide quota and the AUTO slot quota are different limits with
+// different scopes: the slot quota is gateway-wide (a shared upstream budget),
+// the key quota belongs to one caller. When both are exhausted the caller must
+// see the KEY quota, because that is the one it can act on — the slot quota
+// would otherwise be reported as "no capacity", which reads like an outage.
+func TestKeyQuotaWinsOverSlotQuota(t *testing.T) {
+	up := upstream(t, &upstreamCtrl{})
+	defer up.Close()
+	g := newQuotaGW(t, up.URL,
+		config.GWKey{Key: "sk-a", Role: "user",
+			Models: []config.ModelScope{{Model: "AUTO", TokenQuota: 4, Period: "hour"}}})
+	ctx := quotaCtx(t, g, "sk-a")
+
+	// exhaust the key first
+	g.stats.Record(Req{Time: time.Now().UnixMilli(), Key: keyID("sk-a"), Model: "m1",
+		Source: "up", Prompt: 100, Compl: 100, OK: true, Status: 200})
+	rr, code := chatAs(t, g, "sk-a", "AUTO")
+	if rr.Code != http.StatusTooManyRequests {
+		t.Fatalf("want 429 from the key quota, got %d (%s)", rr.Code, rr.Body.String())
+	}
+	if code != "rate_limit_exceeded" {
+		t.Errorf("code = %q, want rate_limit_exceeded", code)
+	}
+	_ = ctx
+}
+
+// A key with NO caps must never be blocked by the slot quota check reaching it
+// through the scope path: scopeTokens on an uncapped key returns 0 usage and
+// the limit check must treat that as "no cap", not "exhausted".
+func TestUncappedKeyNeverBlockedByEmptyScope(t *testing.T) {
+	up := upstream(t, &upstreamCtrl{})
+	defer up.Close()
+	g := newQuotaGW(t, up.URL,
+		config.GWKey{Key: "sk-a", Role: "user", Models: []config.ModelScope{{Model: "AUTO"}}})
+	for i := 0; i < 5; i++ {
+		if rr, _ := chatAs(t, g, "sk-a", "AUTO"); rr.Code != 200 {
+			t.Fatalf("call %d: want 200 (AUTO scope with no quota), got %d (%s)", i, rr.Code, rr.Body.String())
+		}
+	}
+}
+
+// newQuotaGW builds a gateway with one mock upstream serving model m1 and the
+// given keys.
+func newQuotaGW(t *testing.T, upURL string, keys ...config.GWKey) *Gateway {
+	t.Helper()
+	td := t.TempDir()
+	cfgPath := filepath.Join(td, "config.yaml")
+	if err := os.WriteFile(cfgPath, []byte("listen: :0"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	cfg := &config.Config{
+		Path:        cfgPath,
+		AdapterDir:  filepath.Join(td, "adapters"),
+		RuntimeFile: filepath.Join(td, "runtime.json"),
+		Keys:        keys,
+		Sources: []config.Source{{
+			Name: "up", BaseURL: upURL, Adapter: "openai",
+			Models: []config.Model{
+				{ID: "m1", Priority: 100},
+				{ID: "m2", Priority: 90},
+			},
+		}},
+	}
+	if err := cfg.ApplyDefaults(); err != nil {
+		t.Fatal(err)
+	}
+	c, err := core.NewFromConfig(cfg)
+	if err != nil {
+		t.Fatalf("core: %v", err)
+	}
+	t.Cleanup(c.Close)
+	secrets := make([]string, 0, len(keys))
+	for _, k := range keys {
+		secrets = append(secrets, k.Key)
+	}
+	g, err := New(c, secrets)
+	if err != nil {
+		t.Fatalf("gateway: %v", err)
+	}
+	return g
+}
+
+// The core of per-model quotas: a model that runs out of budget must stop
+// being served on its own, while every other model on the SAME key keeps
+// working. A key-wide total would fail this — it would block m2 because m1 was
+// capped, which is exactly the coupling this design removes.
+func TestOneModelsQuotaDoesNotBlockAnother(t *testing.T) {
+	up := upstream(t, &upstreamCtrl{})
+	defer up.Close()
+	td := t.TempDir()
+	cfgPath := td + "/config.yaml"
+	if err := os.WriteFile(cfgPath, []byte("listen: :0"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	cfg := &config.Config{
+		Path: cfgPath, AdapterDir: filepath.Join(td, "adapters"),
+		RuntimeFile: filepath.Join(td, "runtime.json"),
+		Keys: []config.GWKey{{
+			Key: "sk-a", Role: "user", Name: "two-models",
+			Models: []config.ModelScope{
+				{Model: "m1", TokenQuota: 4, Period: "hour"},
+				{Model: "m2", TokenQuota: 1000, Period: "hour"},
+			},
+		}},
+		Sources: []config.Source{{
+			Name: "up", BaseURL: up.URL, Adapter: "openai",
+			Models: []config.Model{
+				{ID: "m1", Priority: 100},
+				{ID: "m2", Priority: 90},
+			},
+		}},
+	}
+	if err := cfg.ApplyDefaults(); err != nil {
+		t.Fatal(err)
+	}
+	c, err := core.NewFromConfig(cfg)
+	if err != nil {
+		t.Fatalf("core: %v", err)
+	}
+	t.Cleanup(c.Close)
+	g, err := New(c, []string{"sk-a"})
+	if err != nil {
+		t.Fatalf("gateway: %v", err)
+	}
+
+	// Spend on m2 FIRST. Without this the two designs are
+	// indistinguishable: a key-wide counter and m1's own counter would both
+	// read 0 before m1 is used, so the test would pass either way (it did —
+	// see the commit that rewrote it).
+	for i := 1; i <= 3; i++ {
+		if rr, _ := chatAs(t, g, "sk-a", "m2"); rr.Code != 200 {
+			t.Fatalf("m2 priming call %d: want 200, got %d", i, rr.Code)
+		}
+	}
+	// m1's budget is 4 and each call costs 4, so the key-wide total (m2+m1) is
+	// already 12 when m1 starts: a key-wide cap would refuse m1 immediately.
+	if rr, _ := chatAs(t, g, "sk-a", "m1"); rr.Code != 200 {
+		t.Fatalf("m1 must be served on its own budget (m2's spend is not its problem), got %d (%s)",
+			rr.Code, rr.Body.String())
+	}
+	if rr, code := chatAs(t, g, "sk-a", "m1"); rr.Code != http.StatusTooManyRequests {
+		t.Fatalf("m1 second call: want 429, got %d (%s)", rr.Code, code)
+	}
+	// m2 must keep working
+	for i := 1; i <= 3; i++ {
+		if rr, _ := chatAs(t, g, "sk-a", "m2"); rr.Code != 200 {
+			t.Fatalf("m2 call %d must be served while m1 is capped, got %d (%s)", i, rr.Code, rr.Body.String())
+		}
+	}
+	// and the message must name m1, not the key
+	rr, _ := chatAs(t, g, "sk-a", "m1")
+	if !strings.Contains(rr.Body.String(), "m1") {
+		t.Errorf("rejection should name the capped model: %s", rr.Body.String())
+	}
+}
+
+// A model with no quota on it is never blocked by a sibling's cap, and an
+// uncapped key is never blocked at all.
+func TestUncappedModelNeverBlocked(t *testing.T) {
+	up := upstream(t, &upstreamCtrl{})
+	defer up.Close()
+	g := newQuotaGW(t, up.URL,
+		config.GWKey{Key: "sk-a", Role: "user", Models: []config.ModelScope{
+			{Model: "m1", TokenQuota: 1, Period: "hour"},
+			{Model: "m2"},
+		}})
+	if rr, _ := chatAs(t, g, "sk-a", "m1"); rr.Code != 200 {
+		t.Fatalf("m1 first: want 200, got %d", rr.Code)
+	}
+	if rr, _ := chatAs(t, g, "sk-a", "m1"); rr.Code != http.StatusTooManyRequests {
+		t.Fatalf("m1 second: want 429, got %d", rr.Code)
+	}
+	for i := 1; i <= 4; i++ {
+		if rr, _ := chatAs(t, g, "sk-a", "m2"); rr.Code != 200 {
+			t.Fatalf("m2 (uncapped) call %d: want 200, got %d", i, rr.Code)
+		}
+	}
+}
+
+// An admin key is never capped even when its scopes carry budgets: a cap that
+// locked the operator out would be unrecoverable through the UI.
+func TestAdminKeyScopesAreNotEnforced(t *testing.T) {
+	up := upstream(t, &upstreamCtrl{})
+	defer up.Close()
+	g := newQuotaGW(t, up.URL,
+		config.GWKey{Key: "sk-admin", Role: "admin", Models: []config.ModelScope{
+			{Model: "m1", TokenQuota: 1, ReqQuota: 1, Period: "hour"},
+		}})
+	for i := 1; i <= 4; i++ {
+		if rr, _ := chatAs(t, g, "sk-admin", "m1"); rr.Code != 200 {
+			t.Fatalf("admin call %d: want 200 (admin scopes are not enforced), got %d", i, rr.Code)
+		}
+	}
+}

internal/gateway/keys.go

@@ -29,28 +29,26 @@ func (g *Gateway) handleKeysAPI(w http.ResponseWriter, r *http.Request) {
 	switch r.Method {
 	case http.MethodGet:
 		if path != "" {
-			writeError(w, http.StatusNotFound, "not_found", "use GET /api/keys")
+			// A GET on a specific key is almost always a client that meant to
+			// DELETE or PUT it but let fetch default to GET. Name the verbs
+			// instead of only pointing back at the collection endpoint.
+			writeError(w, http.StatusNotFound, "not_found",
+				"no such endpoint; use GET /api/keys to list, DELETE /api/keys/{key} to remove, PUT /api/keys/{key} to update")
 			return
 		}
 		writeJSON(w, http.StatusOK, map[string]interface{}{"keys": g.core.ListKeys()})
 	case http.MethodPost:
 		var body struct {
-			Name   string                `json:"name"`
-			Role   string                `json:"role"`
-			Models []config.ModelScope   `json:"models"`
-			Note   string                `json:"note"`
+			Name   string              `json:"name"`
+			Role   string              `json:"role"`
+			Models []config.ModelScope `json:"models"`
+			Note   string              `json:"note"`
 		}
 		if err := json.NewDecoder(r.Body).Decode(&body); err != nil {
 			writeError(w, http.StatusBadRequest, "invalid_request", "invalid json: "+err.Error())
 			return
 		}
-		if body.Role == "" {
-			body.Role = "user"
-		}
-		if body.Role != "admin" && body.Role != "user" {
-			writeError(w, http.StatusBadRequest, "invalid_request", "role must be admin or user")
-			return
-		}
+		body.Role = config.NormalizeRole(body.Role)
 		rec, err := g.core.CreateKey(body.Name, body.Role, body.Models, body.Note)
 		if err != nil {
 			writeError(w, http.StatusBadRequest, "key_error", err.Error())
@@ -72,6 +70,9 @@ func (g *Gateway) handleKeysAPI(w http.ResponseWriter, r *http.Request) {
 			writeError(w, http.StatusBadRequest, "invalid_request", "invalid json: "+err.Error())
 			return
 		}
+		// Each scope entry carries its own token/request caps, so replacing the
+		// scope replaces the budgets with it — there is no separate key-wide
+		// quota that could drift out of sync with the models.
 		rec, err := g.core.UpdateKey(path, body.Name, body.Role, body.Models, body.Note)
 		if err != nil {
 			writeError(w, http.StatusBadRequest, "key_error", err.Error())
@@ -143,8 +144,9 @@ func (g *Gateway) allowedModels(ctx context.Context) []config.ModelScope {
 func (g *Gateway) handleAutoAPI(w http.ResponseWriter, r *http.Request) {
 	if r.Method == http.MethodGet {
 		writeJSON(w, http.StatusOK, map[string]interface{}{
-			"rules":  g.core.AutoRules(),
-			"states": g.core.AutoSlotStates(),
+			"rules":       g.core.AutoRules(),
+			"image_rules": g.core.AutoImageRules(),
+			"states":      g.core.AutoSlotStates(),
 		})
 		return
 	}
@@ -155,18 +157,31 @@ func (g *Gateway) handleAutoAPI(w http.ResponseWriter, r *http.Request) {
 	switch r.Method {
 	case http.MethodPut, http.MethodPost:
 		var body struct {
-			Rules []config.ModelScope `json:"rules"`
+			Rules      []config.ModelScope `json:"rules"`
+			ImageRules []config.ModelScope `json:"image_rules"`
 		}
 		if err := json.NewDecoder(r.Body).Decode(&body); err != nil {
 			writeError(w, http.StatusBadRequest, "invalid_request", "invalid json: "+err.Error())
 			return
 		}
-		if err := g.core.SaveAutoRules(body.Rules); err != nil {
-			writeError(w, http.StatusBadRequest, "auto_error", err.Error())
-			return
+		if body.Rules != nil {
+			if err := g.core.SaveAutoRules(body.Rules); err != nil {
+				writeError(w, http.StatusBadRequest, "auto_error", err.Error())
+				return
+			}
 		}
-		writeJSON(w, http.StatusOK, map[string]interface{}{"ok": true, "rules": g.core.AutoRules()})
+		if body.ImageRules != nil {
+			if err := g.core.SaveAutoImageRules(body.ImageRules); err != nil {
+				writeError(w, http.StatusBadRequest, "auto_error", err.Error())
+				return
+			}
+		}
+		writeJSON(w, http.StatusOK, map[string]interface{}{
+			"ok":          true,
+			"rules":       g.core.AutoRules(),
+			"image_rules": g.core.AutoImageRules(),
+		})
 	default:
 		writeError(w, http.StatusMethodNotAllowed, "method_not_allowed", "")
 	}
-}
+}

internal/gateway/login_guard_test.go

@@ -0,0 +1,116 @@
+package gateway
+
+import (
+	"encoding/json"
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"testing"
+)
+
+func TestLoginBruteForceLockout(t *testing.T) {
+	g := newTestGateway(t)
+	body := `{"key":"wrong-key-XXXX"}`
+
+	// 5 failures allowed (each answered 401), 6th+ locked out
+	var gotLocked bool
+	for i := 1; i <= 8; i++ {
+		req := httptest.NewRequest("POST", "/api/login", strings.NewReader(body))
+		rr := httptest.NewRecorder()
+		g.handleLoginAPI(rr, req)
+		if rr.Code != http.StatusUnauthorized {
+			t.Fatalf("attempt %d: status=%d want 401", i, rr.Code)
+		}
+		if i >= 6 {
+			// after lockout the response must be identical (no info leak)
+			if !strings.Contains(rr.Body.String(), "invalid gateway api key") {
+				t.Fatalf("lockout reply changed body: %s", rr.Body.String())
+			}
+			gotLocked = true
+		}
+	}
+	if !gotLocked {
+		t.Fatal("expected lockout by attempt 6")
+	}
+
+	// correct key from the SAME ip must also be rejected while locked
+	okBody := `{"key":"sk-test"}`
+	req := httptest.NewRequest("POST", "/api/login", strings.NewReader(okBody))
+	rr := httptest.NewRecorder()
+	g.handleLoginAPI(rr, req)
+	if rr.Code != http.StatusUnauthorized {
+		t.Fatalf("locked-out valid key: status=%d want 401", rr.Code)
+	}
+
+	// a different IP is unaffected
+	req = httptest.NewRequest("POST", "/api/login", strings.NewReader(okBody))
+	req.RemoteAddr = "10.9.9.9:5555"
+	rr = httptest.NewRecorder()
+	g.handleLoginAPI(rr, req)
+	if rr.Code != http.StatusOK {
+		t.Fatalf("other IP valid key: status=%d want 200 body=%s", rr.Code, rr.Body.String())
+	}
+	var resp map[string]interface{}
+	_ = json.Unmarshal(rr.Body.Bytes(), &resp)
+	if resp["ok"] != true {
+		t.Fatalf("login ok flag missing: %s", rr.Body.String())
+	}
+}
+
+func TestLoginSuccessResetsFailures(t *testing.T) {
+	g := newTestGateway(t)
+	mkBad := func() *http.Request {
+		return httptest.NewRequest("POST", "/api/login", strings.NewReader(`{"key":"nope"}`))
+	}
+	for i := 0; i < loginMaxFails-1; i++ { // one below the lockout threshold
+		rr := httptest.NewRecorder()
+		g.handleLoginAPI(rr, mkBad())
+		if rr.Code != http.StatusUnauthorized {
+			t.Fatal("setup failure")
+		}
+	}
+	// success clears the counter
+	rr := httptest.NewRecorder()
+	g.handleLoginAPI(rr, httptest.NewRequest("POST", "/api/login", strings.NewReader(`{"key":"sk-test"}`)))
+	if rr.Code != http.StatusOK {
+		t.Fatalf("valid login failed: %d", rr.Code)
+	}
+	// now up to loginMaxFails-1 more failures are still accepted before lock
+	for i := 0; i < loginMaxFails-1; i++ {
+		last := httptest.NewRecorder()
+		g.handleLoginAPI(last, mkBad())
+	}
+	last := httptest.NewRecorder()
+	g.handleLoginAPI(last, mkBad())
+	// count is at loginMaxFails-1 failures since reset; this next one reaches
+	// the threshold but the attempt itself is still evaluated normally (401),
+	// and the FOLLOWING one must be locked.
+	if last.Code != http.StatusUnauthorized {
+		t.Fatalf("threshold attempt should still answer 401: %d", last.Code)
+	}
+	next := httptest.NewRecorder()
+	g.handleLoginAPI(next, mkBad())
+	if next.Code != http.StatusUnauthorized {
+		t.Fatal("want 401")
+	}
+	// verify lock actually engaged via the guard state
+	g.loginMu.Lock()
+	f := g.loginFails["192.0.2.1"]
+	engaged := f != nil && f.count >= loginMaxFails && f.until > 0
+	g.loginMu.Unlock()
+	if !engaged {
+		t.Fatal("expected engaged lockout state")
+	}
+}
+
+func TestClientIPExtraction(t *testing.T) {
+	r := httptest.NewRequest("POST", "/api/login", nil)
+	r.RemoteAddr = "203.0.113.7:12345"
+	if got := clientIP(r); got != "203.0.113.7" {
+		t.Errorf("clientIP=%q want 203.0.113.7", got)
+	}
+	r.Header.Set("X-Real-IP", "198.51.100.9")
+	if got := clientIP(r); got != "198.51.100.9" {
+		t.Errorf("clientIP with X-Real-IP=%q want proxy value", got)
+	}
+}

internal/gateway/overflow_err_test.go

@@ -0,0 +1,167 @@
+package gateway
+
+import (
+	"errors"
+	"regexp"
+	"strings"
+	"testing"
+
+	"llmsproxy/internal/scheduler"
+	"llmsproxy/internal/types"
+)
+
+// piOverflowRe 是 @earendil-works/pi-ai 的 OVERFLOW_PATTERNS 中对应的一条。
+// 归一化标记必须命中它，否则改了等于没改：客户端仍不会压缩重试。
+var piOverflowRe = regexp.MustCompile(`context[_ ]length[_ ]exceeded`)
+
+func TestOverflowCanonicalMatchesPiPattern(t *testing.T) {
+	if !piOverflowRe.MatchString(overflowCanonical) {
+		t.Fatalf("归一化标记 %q 不匹配 pi 的超窗正则，客户端仍不会压缩重试", overflowCanonical)
+	}
+}
+
+// justworker 的真实措辞：与 pi 现有 25 条正则一条都不匹配，
+// 所以必须在网关侧归一化，否则超窗对客户端完全不可见。
+func TestJustworkerOverflowIsNormalized(t *testing.T) {
+	raw := "请精简对话历史…(Context window is full, please reduce the conversation history)…"
+
+	// 前提校验：原始措辞确实不被 pi 认出（否则本用例失去意义）。
+	if piOverflowRe.MatchString(raw) {
+		t.Fatalf("前提不成立：原始措辞本就匹配 pi 正则: %q", raw)
+	}
+	if !looksLikeOverflow(raw) {
+		t.Fatalf("网关应识别出这是超窗: %q", raw)
+	}
+
+	err := &scheduler.ChainErr{Tiers: []scheduler.TierError{{
+		Tier: 1, Source: "justworker", Model: "claude-opus-5", Err: errors.New(raw),
+	}}}
+
+	got := clientUpstreamErr(err)
+	if !piOverflowRe.MatchString(got) {
+		t.Fatalf("归一化后仍不匹配 pi 正则，客户端不会压缩重试: %q", got)
+	}
+}
+
+// 直连路径（非 AUTO 链）也要归一化——它走的是同一个入口。
+func TestDirectPathOverflowIsNormalized(t *testing.T) {
+	got := clientUpstreamErr(errors.New("upstream 400: Context window is full"))
+	if !piOverflowRe.MatchString(got) {
+		t.Fatalf("直连超窗未被归一化: %q", got)
+	}
+}
+
+// 非超窗错误不能被加上超窗标记，否则会把普通失败误报成需要压缩。
+func TestNonOverflowNotNormalized(t *testing.T) {
+	err := &scheduler.ChainErr{Tiers: []scheduler.TierError{{
+		Tier: 1, Source: "x", Model: "y", Err: errors.New("502 Bad Gateway: upstream connect error"),
+	}}}
+	got := clientUpstreamErr(err)
+	if strings.Contains(got, overflowCanonical) {
+		t.Fatalf("普通上游失败被误标为超窗: %q", got)
+	}
+	if !strings.Contains(got, "502 Bad Gateway") {
+		t.Fatalf("普通失败的诊断信息应保留: %q", got)
+	}
+}
+
+// 回归：短诊断词常落在错误尾部，80 字节截断会正好把它切掉，
+// 超窗就永远认不出来。宽度放宽后必须保住。
+func TestOverflowPhraseSurvivesTruncationWidth(t *testing.T) {
+	prefix := strings.Repeat("x", 100)
+	// OneLine 的长度限制作用于整条 "source/model: err"，不是只算 err 本身。
+	full := "justworker/claude-opus-5: " + prefix + " Context window is full"
+
+	// 前提：80 字节下短语确实会被切掉（这正是修复前的行为）。
+	if got := types.OneLine(full, 80); strings.Contains(got, "Context window is full") {
+		t.Fatalf("前提不成立：80 字节下短语居然还在: %q", got)
+	}
+	// 160 字节下必须保住，否则超窗识别会被截断悄悄破坏。
+	if got := types.OneLine(full, 160); !strings.Contains(got, "Context window is full") {
+		t.Fatalf("160 字节应保住尾部短诊断词: %q", got)
+	}
+
+	err := &scheduler.ChainErr{Tiers: []scheduler.TierError{{
+		Tier: 1, Source: "justworker", Model: "claude-opus-5",
+		Err: errors.New(prefix + " Context window is full"),
+	}}}
+
+	got := clientUpstreamErr(err)
+	if !strings.Contains(strings.ToLower(got), "context window is full") {
+		t.Fatalf("宽度 160 应保住尾部短诊断词，实际: %q", got)
+	}
+	if !piOverflowRe.MatchString(got) {
+		t.Fatalf("归一化标记丢失: %q", got)
+	}
+}
+
+// piNonOverflowRe 是 @earendil-works/pi-ai 的 NON_OVERFLOW_PATTERNS。
+// pi 的 isContextOverflow 先查这一组，命中则直接判为「非超窗」——
+// 即使 OVERFLOW_PATTERNS 也命中，也不会压缩重试。
+var piNonOverflowRe = []*regexp.Regexp{
+	regexp.MustCompile(`^(Throttling error|Service unavailable):`),
+	regexp.MustCompile(`(?i)rate limit`),
+	regexp.MustCompile(`(?i)too many requests`),
+}
+
+// piWouldCompact 复刻 isContextOverflow 的 Case 1 判定顺序：
+// 先排除 NON_OVERFLOW，再匹配 OVERFLOW。
+func piWouldCompact(errorMessage string) bool {
+	for _, p := range piNonOverflowRe {
+		if p.MatchString(errorMessage) {
+			return false
+		}
+	}
+	return piOverflowRe.MatchString(errorMessage)
+}
+
+// issue 回归：AUTO 链失败消息天然把「超窗 tier」和「配额/限流 tier」拼在一起，
+// 超窗标记后面只要出现 rate limit / too many requests 的字样，
+// pi 就会把整条消息判为「非超窗」而拒绝压缩重试。
+func TestOverflowMessageSurvivesRateLimitedSiblingTier(t *testing.T) {
+	cases := []struct {
+		name    string
+		sibling string
+	}{
+		{"rate limit 措辞", "api error 429: rate limit exceeded"},
+		{"too many requests 措辞", "429 Too Many Requests"},
+	}
+
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			err := &scheduler.ChainErr{Tiers: []scheduler.TierError{
+				{Tier: 1, Source: "sensenova", Model: "glm-5.2", Err: errors.New(tc.sibling)},
+				{Tier: 3, Source: "gozen", Model: "deepseek-v4.1-flash",
+					Err: errors.New("api error 400: This model's maximum context length is 1048576 tokens")},
+			}}
+
+			got := clientUpstreamErr(err)
+			if !piWouldCompact(got) {
+				t.Fatalf("同链限流 tier 封杀了超窗识别，pi 不会压缩重试: %q", got)
+			}
+			// 干净：不得把限流 tier 的措辞带出去。
+			for _, p := range piNonOverflowRe {
+				if p.MatchString(got) {
+					t.Fatalf("超窗消息仍带限流措辞 %v: %q", p, got)
+				}
+			}
+			// 但仍要能看出是哪个源超的窗。
+			if !strings.Contains(got, "gozen/deepseek-v4.1-flash") {
+				t.Errorf("超窗源信息丢失: %q", got)
+			}
+		})
+	}
+}
+
+// 直连超窗同样要干净可用。
+func TestDirectOverflowMessageIsClean(t *testing.T) {
+	raw := "api error 400: Error from provider (Console Go): This model's maximum context length is 1048576 tokens. " +
+		"However, you requested 1440050 tokens. Please reduce the length of the messages or completion."
+	got := clientUpstreamErr(errors.New(raw))
+	if !piWouldCompact(got) {
+		t.Fatalf("直连超窗未被 pi 识别: %q", got)
+	}
+	if !strings.Contains(got, overflowCanonical) {
+		t.Fatalf("缺少归一化标记: %q", got)
+	}
+}

internal/gateway/server.go

@@ -11,6 +11,7 @@ import (
 	"fmt"
 	"io/fs"
 	"log"
+	"net"
 	"net/http"
 	"net/url"
 	"strings"
@@ -31,6 +32,87 @@ type Gateway struct {
 	stats     *Stats
 	probeMu   sync.Mutex
 	lastProbe time.Time
+	// loginGuard throttles /api/login brute-force attempts: per-IP failure
+	// counters with exponential backoff, plus a global cap so a distributed
+	// spray cannot outrun the per-IP window. Guarded by loginMu.
+	loginMu    sync.Mutex
+	loginFails map[string]*loginFail
+	loginWin   int64 // unix sec of current global window start
+	loginGlob  int   // failures in the global window
+}
+
+// loginFail tracks consecutive failed logins from one source IP.
+type loginFail struct {
+	count  int
+	until  int64 // locked-out until this unix second (0 = not locked)
+	lastAt int64
+}
+
+const (
+	loginMaxFails     = 5       // failures before lockout kicks in
+	loginLockBase     = 30      // first lockout: 30s
+	loginLockCap      = 30 * 60 // repeated lockouts cap at 30min
+	loginDecay        = 15 * 60 // counters decay after 15min of quiet
+	loginGlobalMax    = 100     // global failures per minute across all IPs
+	loginGlobalWindow = 60      // global window length (sec)
+)
+
+// loginAllow checks (and records) a login attempt for ip. It returns false
+// when the attempt must be rejected: the IP is in exponential lockout or the
+// global failure budget for the current window is exhausted.
+func (g *Gateway) loginAllow(ip string, now int64) bool {
+	g.loginMu.Lock()
+	defer g.loginMu.Unlock()
+	if g.loginFails == nil {
+		g.loginFails = map[string]*loginFail{}
+	}
+	// global window rollover
+	if now-g.loginWin >= loginGlobalWindow {
+		g.loginWin = now
+		g.loginGlob = 0
+	}
+	f := g.loginFails[ip]
+	if f != nil && f.count > 0 && now-f.lastAt > loginDecay {
+		delete(g.loginFails, ip) // quiet long enough: forgive and forget
+		f = nil
+	}
+	if f != nil && f.until > now {
+		return false // still locked out
+	}
+	if g.loginGlob >= loginGlobalMax {
+		return false // global spray budget exhausted for this window
+	}
+	g.loginGlob++
+	return true
+}
+
+// loginRecord notes the outcome of one attempt: failures escalate toward an
+// exponential lockout; a success clears the IP's counter entirely.
+func (g *Gateway) loginRecord(ip string, ok bool, now int64) {
+	g.loginMu.Lock()
+	defer g.loginMu.Unlock()
+	if ok {
+		if g.loginFails != nil {
+			delete(g.loginFails, ip)
+		}
+		return
+	}
+	f := g.loginFails[ip]
+	if f == nil {
+		f = &loginFail{}
+		g.loginFails[ip] = f
+	}
+	f.count++
+	f.lastAt = now
+	if f.count >= loginMaxFails {
+		// lock duration doubles with each extra burst of failures, capped
+		epoch := int64(f.count/loginMaxFails) - 1
+		dur := int64(loginLockBase) << min(epoch, 6)
+		if dur > loginLockCap {
+			dur = loginLockCap
+		}
+		f.until = now + dur
+	}
 }
 
 func New(c *core.Core, gatewayKeys []string) (*Gateway, error) {
@@ -38,7 +120,13 @@ func New(c *core.Core, gatewayKeys []string) (*Gateway, error) {
 	if err != nil {
 		return nil, err
 	}
-	st := NewStats(10000)
+	// The record ring is deliberately small: aggregates are built from the full
+	// audit history at startup (see Stats.LoadAudit), so the ring only has to
+	// cover the status page's 5-minute source windows and the dashboard's first
+	// screen. Everything older is paged from disk by /api/stats/records.
+	// A large ring here would put the ~25 MB of resident Req structs straight
+	// back, which is exactly what the on-demand paging removed.
+	st := NewStats(defaultRingSize)
 	if cfg := c.Config(); cfg != nil && cfg.RuntimeFile != "" {
 		st.LoadAudit(cfg.RuntimeFile + ".audit.jsonl")
 	}
@@ -113,6 +201,17 @@ func (s *statusRecorder) Write(b []byte) (int, error) {
 	return s.ResponseWriter.Write(b)
 }
 
+// Flush forwards to the underlying writer so SSE handlers can stream
+// incrementally through the access-log wrapper. Without this method the
+// w.(http.Flusher) assertion inside the streaming chat handlers fails (the
+// embedded ResponseWriter interface does not carry Flush into the method set),
+// and every chunk stays buffered until the response ends.
+func (s *statusRecorder) Flush() {
+	if f, ok := s.ResponseWriter.(http.Flusher); ok {
+		f.Flush()
+	}
+}
+
 func (g *Gateway) routes(w http.ResponseWriter, r *http.Request) {
 	switch {
 	case r.URL.Path == "/v1/chat/completions":
@@ -123,14 +222,20 @@ func (g *Gateway) routes(w http.ResponseWriter, r *http.Request) {
 		g.handleModels(w, r)
 	case r.URL.Path == "/api/adapters" || strings.HasPrefix(r.URL.Path, "/api/adapters/"):
 		g.handleAdaptersAPI(w, r)
+	case r.URL.Path == "/api/v1" || strings.HasPrefix(r.URL.Path, "/api/v1/"):
+		g.apiV1Routes(w, r)
 	case r.URL.Path == "/api/sources" || strings.HasPrefix(r.URL.Path, "/api/sources/"):
 		g.handleSourcesAPI(w, r)
+	case r.URL.Path == "/api/source_templates" || strings.HasPrefix(r.URL.Path, "/api/source_templates/"):
+		g.handleSourceTemplatesAPI(w, r)
 	case r.URL.Path == "/api/chat":
 		g.handleChat(w, r)
 	case r.URL.Path == "/api/status":
 		g.handleStatusAPI(w, r)
 	case r.URL.Path == "/api/status/reset":
 		g.handleResetHealth(w, r)
+	case r.URL.Path == "/api/stats/records":
+		g.handleStatsRecordsAPI(w, r)
 	case r.URL.Path == "/api/stats" || strings.HasPrefix(r.URL.Path, "/api/stats/"):
 		g.handleStatsAPI(w, r)
 	case r.URL.Path == "/api/keys" || strings.HasPrefix(r.URL.Path, "/api/keys/"):
@@ -325,34 +430,59 @@ document.getElementById('key').addEventListener('keydown',function(e){if(e.key==
 apply();
 </script></body></html>`
 
-// handleLoginAPI validates the gateway key and issues a session cookie.
+// handleLoginAPI validates the gateway key and issues a session cookie,
+// guarded by per-IP + global brute-force throttling (see loginGuard).
 func (g *Gateway) handleLoginAPI(w http.ResponseWriter, r *http.Request) {
 	if r.Method != http.MethodPost {
 		writeError(w, http.StatusMethodNotAllowed, "method_not_allowed", "use POST")
 		return
 	}
+	now := time.Now().Unix()
+	if !g.loginAllow(clientIP(r), now) {
+		// Don't reveal whether the key was right: answer the same 401 the
+		// failure path uses, but skip the expensive key lookup.
+		writeError(w, http.StatusUnauthorized, "invalid_api_key", "invalid gateway api key")
+		return
+	}
 	var body struct {
 		Key string `json:"key"`
 	}
 	if err := json.NewDecoder(r.Body).Decode(&body); err != nil {
+		g.loginRecord(clientIP(r), false, now)
 		writeError(w, http.StatusBadRequest, "invalid_request", "invalid json")
 		return
 	}
 	if _, ok := g.core.FindKey(body.Key); !ok {
+		g.loginRecord(clientIP(r), false, now)
 		writeError(w, http.StatusUnauthorized, "invalid_api_key", "invalid gateway api key")
 		return
 	}
+	g.loginRecord(clientIP(r), true, now)
 	http.SetCookie(w, &http.Cookie{
-		Name:   "gw_key",
-		Value:  body.Key,
-		Path:   "/",
-		MaxAge: 86400 * 30,
+		Name:     "gw_key",
+		Value:    body.Key,
+		Path:     "/",
+		MaxAge:   86400 * 30,
 		HttpOnly: true,
 		SameSite: http.SameSiteLaxMode,
 	})
 	writeJSON(w, http.StatusOK, map[string]interface{}{"ok": true})
 }
 
+// clientIP returns the caller's IP for rate limiting. Behind a trusted
+// reverse proxy (nginx) the X-Real-IP header is preferred; otherwise the
+// remote address is used with the port stripped.
+func clientIP(r *http.Request) string {
+	if x := r.Header.Get("X-Real-IP"); x != "" {
+		return x
+	}
+	host, _, err := net.SplitHostPort(r.RemoteAddr)
+	if err != nil {
+		return r.RemoteAddr
+	}
+	return host
+}
+
 func (g *Gateway) handleModels(w http.ResponseWriter, r *http.Request) {
 	if r.Method != http.MethodGet {
 		writeError(w, http.StatusMethodNotAllowed, "method_not_allowed", "use GET")
@@ -456,8 +586,25 @@ func (g *Gateway) handleStatusAPI(w http.ResponseWriter, r *http.Request) {
 		"gateway_keys":  ks,
 	}
 	if reqRole(r.Context()) == "admin" {
-		resp["sources"] = g.core.Registry().Status()
+		sts := g.core.Registry().Status()
+		recent := g.stats.SourceRecent(300)
+		avgs := g.stats.SourceAverages(300)
+		for i := range sts {
+			if v, ok := recent[sts[i].Name]; ok {
+				sts[i].RecentOK = v[0]
+				sts[i].RecentErr = v[1]
+			}
+			if a, ok := avgs[sts[i].Name]; ok {
+				sts[i].AvgFirstByteMs = a.AvgFirstByteMs
+				sts[i].AvgTokPerS = a.AvgTokPerS
+			}
+		}
+		resp["sources"] = sts
 		resp["adapters"] = g.core.ListAdapters()
+		// Elastic Lua pool sizing: created/idle/in_use vs the ceiling, plus the
+		// current grow/shrink steps, so the algorithm is inspectable in the UI
+		// instead of being a black box.
+		resp["adapter_pools"] = g.core.VM().PoolStats()
 	}
 	writeJSON(w, http.StatusOK, resp)
 }
@@ -491,4 +638,4 @@ func writeJSON(w http.ResponseWriter, code int, v interface{}) {
 	if err := json.NewEncoder(w).Encode(v); err != nil {
 		log.Printf("[gateway] write json: %v", err)
 	}
-}
+}

internal/gateway/stats_test.go

@@ -92,21 +92,25 @@ func TestAuditRotationRecords(t *testing.T) {
 func TestLoadAuditFullReplay(t *testing.T) {
 	dir := t.TempDir()
 	path := filepath.Join(dir, "audit.jsonl")
-	// timestamps relative to now: buckets are whole unix hours, so window
-	// assertions must not depend on the minute-of-hour of the test run.
+	// timestamps anchored to whole unix hours (see aggregateLocked): window
+	// assertions must hold no matter which minute-of-hour the suite runs at,
+	// so rows are placed relative to hour boundaries, not raw offsets from now.
 	now := time.Now()
+	thisHour := now.Truncate(time.Hour)
+	twoBucketsAgo := thisHour.Add(-90 * time.Minute) // bucket H-2
+	prevBucketMid := thisHour.Add(-30 * time.Minute) // bucket H-1
 	lines := []string{
 		`{"obj":"access","time":1699999999000,"key":"k","method":"GET","path":"/api/stats","status":200}`,
-		fmt.Sprintf(`{"time":%d,"key":"k","type":"chat","model":"m","source":"s","prompt_tokens":100,"completion_tokens":50,"latency_ms":10,"ok":true,"status":200}`, now.Add(-65*time.Minute).UnixMilli()),
+		fmt.Sprintf(`{"time":%d,"key":"k","type":"chat","model":"m","source":"s","prompt_tokens":100,"completion_tokens":50,"latency_ms":10,"ok":true,"status":200}`, twoBucketsAgo.UnixMilli()),
 		`{this is not valid json`,
-		fmt.Sprintf(`{"time":%d,"key":"k","type":"stream","model":"m","source":"s","prompt_tokens":200,"completion_tokens":20,"latency_ms":20,"ok":false,"status":503}`, now.Add(-30*time.Minute).UnixMilli()),
+		fmt.Sprintf(`{"time":%d,"key":"k","type":"stream","model":"m","source":"s","prompt_tokens":200,"completion_tokens":20,"latency_ms":20,"ok":false,"status":503}`, prevBucketMid.UnixMilli()),
 		"garbage-not-json\n",
-		fmt.Sprintf(`{"time":%d,"key":"k","type":"chat","model":"m2","source":"s","prompt_tokens":7,"completion_tokens":3,"latency_ms":5,"ok":true,"status":200}`, now.Add(-30*time.Minute).UnixMilli()),
+		fmt.Sprintf(`{"time":%d,"key":"k","type":"chat","model":"m2","source":"s","prompt_tokens":7,"completion_tokens":3,"latency_ms":5,"ok":true,"status":200}`, prevBucketMid.UnixMilli()),
 	}
 	loaded := strings.Join(lines, "\n") + "\n" + strings.Repeat("x", 1<<18) + "\n"
 	// oversized row at the END proves the scanner tolerates >64KB lines and
 	// still finishes the replay instead of truncating silently.
-	loaded += fmt.Sprintf(`{"time":%d,"key":"k","type":"chat","model":"m","source":"s","prompt_tokens":1,"completion_tokens":1,"latency_ms":1,"ok":true,"status":200}`, now.Add(-time.Minute).UnixMilli()) + "\n"
+	loaded += fmt.Sprintf(`{"time":%d,"key":"k","type":"chat","model":"m","source":"s","prompt_tokens":1,"completion_tokens":1,"latency_ms":1,"ok":true,"status":200}`, now.UnixMilli()) + "\n"
 	if err := os.WriteFile(path, []byte(loaded), 0644); err != nil {
 		t.Fatal(err)
 	}
@@ -126,9 +130,10 @@ func TestLoadAuditFullReplay(t *testing.T) {
 		t.Fatalf("ring must hold only real requests, got %d rows: %#v", len(s.recs), s.recs)
 	}
 	// quota window rebuilt from full history. All-time and multi-hour windows
-	// must see everything; a 1h window must NOT return all records (the old
-	// ms/seconds unit bug made any sec>0 window return everything), and must
-	// always include the row written one minute ago (current hour bucket).
+	// must see everything. A 1h window must NOT return all records (the old
+	// ms/seconds unit bug made any sec>0 window return everything): buckets
+	// are whole unix hours, so only rows in the current hour bucket qualify
+	// — exactly the just-now row (2 tokens).
 	if w := s.WindowTokens("m", "s", 0); w != 372 {
 		t.Fatalf("all-time window want 372, got %d", w)
 	}
@@ -138,8 +143,8 @@ func TestLoadAuditFullReplay(t *testing.T) {
 	if w := s.WindowTokens("m", "s", 24*hourSec); w != 372 {
 		t.Fatalf("24h window want 372, got %d", w)
 	}
-	if w := s.WindowTokens("m", "s", hourSec); w < 2 || w >= 372 {
-		t.Fatalf("1h window want [2,372), got %d", w)
+	if w := s.WindowTokens("m", "s", hourSec); w != 2 {
+		t.Fatalf("1h window want 2 (current-bucket row only), got %d", w)
 	}
 	if w := s.WindowTokens("m2", "s", 24*hourSec); w != 10 {
 		t.Fatalf("m2 24h window want 10, got %d", w)
@@ -151,4 +156,407 @@ func TestLoadAuditFullReplay(t *testing.T) {
 	if s.byStatus[200].Err != 0 || s.byStatus[503] == nil || s.byStatus[503].Reqs != 1 {
 		t.Fatalf("by_status wrong: %#v", s.byStatus)
 	}
-}
+}
+
+// ---- on-demand log loading (plan 阶段 3) ----
+
+// writeAuditFile writes n synthetic request rows (plus some non-request event
+// rows, which must be skipped) and returns the path.
+func writeAuditFile(t *testing.T, path string, n int, key string, startMs int64) {
+	t.Helper()
+	f, err := os.Create(path)
+	if err != nil {
+		t.Fatalf("create audit: %v", err)
+	}
+	defer f.Close()
+	for i := 0; i < n; i++ {
+		fmt.Fprintf(f, `{"time":%d,"key":%q,"type":"chat","model":"m","source":"s","prompt_tokens":1,"completion_tokens":1,"latency_ms":5,"ok":true,"status":200}`+"\n",
+			startMs+int64(i)*1000, key)
+		if i%10 == 0 {
+			// an access-log/event row: must never appear as a record
+			fmt.Fprintf(f, `{"obj":"access","time":%d,"path":"/v1/models"}`+"\n", startMs+int64(i)*1000)
+		}
+	}
+}
+
+// TestLoadAuditAggregatesFullHistory is the correctness half of the log work:
+// the dashboard's totals must cover EVERY audit record, however large the file,
+// while the in-memory record ring stays bounded.
+func TestLoadAuditAggregatesFullHistory(t *testing.T) {
+	dir := t.TempDir()
+	path := filepath.Join(dir, "audit.jsonl")
+	const n = 20000
+	writeAuditFile(t, path, n, "k1", 1_700_000_000_000)
+
+	s := NewStats(0)
+	s.LoadAudit(path)
+	if s.ReplayPartial() {
+		t.Fatal("a readable history must not be reported as partial")
+	}
+
+	snap := s.Snapshot(0, "")
+	total, _ := snap["total"].(Stat)
+	if total.Reqs != n {
+		t.Fatalf("total reqs = %d, want %d (aggregates must cover the whole file)", total.Reqs, n)
+	}
+	if total.Tokens != int64(n)*2 { // 1 prompt + 1 completion per row
+		t.Fatalf("total tokens = %d, want %d", total.Tokens, n*2)
+	}
+
+	// ...while the resident ring is capped and the snapshot ships one screen
+	recs, _ := snap["records"].([]Req)
+	if len(recs) == 0 {
+		t.Fatal("the newest records must stay resident")
+	}
+	if len(recs) > firstScreenRecords {
+		t.Fatalf("snapshot shipped %d records, must be capped at the first screen (%d)",
+			len(recs), firstScreenRecords)
+	}
+	s.mu.Lock()
+	ring := len(s.recs)
+	s.mu.Unlock()
+	if ring > defaultRingSize {
+		t.Fatalf("resident ring holds %d records, must be capped at %d", ring, defaultRingSize)
+	}
+
+	// the ring must hold the NEWEST records, not the first ones scanned
+	newest := recs[len(recs)-1]
+	wantNewest := int64(1_700_000_000_000) + int64(n-1)*1000
+	if newest.Time != wantNewest {
+		t.Fatalf("newest resident record time = %d, want %d", newest.Time, wantNewest)
+	}
+}
+
+// TestLoadAuditAggregatesAcrossRotatedFiles: history split over rotated .old
+// files must still add up, and must be replayed oldest-first so the hourly
+// quota buckets keep the intended trailing window.
+func TestLoadAuditAggregatesAcrossRotatedFiles(t *testing.T) {
+	dir := t.TempDir()
+	path := filepath.Join(dir, "audit.jsonl")
+	writeAuditFile(t, path, 40, "k1", 1_700_000_500_000)            // newest
+	writeAuditFile(t, path+".100.old", 40, "k1", 1_700_000_300_000) // older
+	writeAuditFile(t, path+".050.old", 40, "k1", 1_700_000_100_000) // oldest
+
+	s := NewStats(0)
+	s.LoadAudit(path)
+	snap := s.Snapshot(0, "")
+	total, _ := snap["total"].(Stat)
+	if total.Reqs != 120 {
+		t.Fatalf("total reqs = %d, want 120 across three files", total.Reqs)
+	}
+	recs, _ := snap["records"].([]Req)
+	if len(recs) == 0 {
+		t.Fatal("expected resident records")
+	}
+	// oldest-first replay + newest-last ring => the last resident record is the
+	// newest one in the live file
+	if got := recs[len(recs)-1].Time; got != 1_700_000_500_000+39*1000 {
+		t.Fatalf("newest resident record time = %d, want the live file's last row", got)
+	}
+}
+
+// TestReqRingKeepsNewest pins the ring semantics used by the full-history scan:
+// bounded size, newest retained, oldest-first output.
+func TestReqRingKeepsNewest(t *testing.T) {
+	r := newReqRing(3)
+	for i := 1; i <= 7; i++ {
+		r.push(Req{Time: int64(i)})
+	}
+	got := r.slice()
+	if len(got) != 3 {
+		t.Fatalf("ring len = %d, want 3", len(got))
+	}
+	for i, want := range []int64{5, 6, 7} {
+		if got[i].Time != want {
+			t.Fatalf("ring[%d] = %d, want %d (oldest-first, newest retained)", i, got[i].Time, want)
+		}
+	}
+
+	// a partially filled ring returns exactly what was pushed
+	r2 := newReqRing(5)
+	r2.push(Req{Time: 1})
+	r2.push(Req{Time: 2})
+	if got := r2.slice(); len(got) != 2 || got[0].Time != 1 || got[1].Time != 2 {
+		t.Fatalf("partial ring = %+v", got)
+	}
+}
+
+// TestLoadAuditFullFileWhenSmall: a small file is loaded completely and is not
+// flagged partial.
+func TestLoadAuditFullFileWhenSmall(t *testing.T) {
+	path := filepath.Join(t.TempDir(), "audit.jsonl")
+	writeAuditFile(t, path, 50, "k1", 1_700_000_000_000)
+	s := NewStats(0)
+	s.LoadAudit(path)
+	if s.ReplayPartial() {
+		t.Fatal("a small file must be replayed in full, not flagged partial")
+	}
+	snap := s.Snapshot(0, "")
+	total, _ := snap["total"].(Stat)
+	if total.Reqs != 50 {
+		t.Fatalf("total reqs = %d, want 50", total.Reqs)
+	}
+}
+
+// TestQuotaWindowsSurviveBoundedRing is the regression guard the plan calls out:
+// capping the resident record ring must not break token-quota accounting, which
+// reads the hour buckets rather than the ring.
+func TestQuotaWindowsSurviveBoundedRing(t *testing.T) {
+	s := NewStats(10) // deliberately tiny ring
+	now := time.Now().UnixMilli()
+	for i := 0; i < 100; i++ {
+		s.Record(Req{Time: now, Key: "k", Model: "m", Source: "s", Prompt: 2, Compl: 3, OK: true, Status: 200})
+	}
+	if got := s.WindowTokens("m", "s", 24*hourSec); got != 500 {
+		t.Fatalf("window tokens = %d, want 500 (quota accounting must not depend on the ring)", got)
+	}
+	if got := s.ModelTokens("k")["m"]; got != 500 {
+		t.Fatalf("model tokens = %d, want 500", got)
+	}
+	if got := s.KeyTokens("k"); got != 500 {
+		t.Fatalf("key tokens = %d, want 500", got)
+	}
+}
+
+// TestAuditPagePagesBackwards walks the whole history one page at a time and
+// asserts strict newest-first order with no gaps or repeats.
+func TestAuditPagePagesBackwards(t *testing.T) {
+	path := filepath.Join(t.TempDir(), "audit.jsonl")
+	const n = 250
+	writeAuditFile(t, path, n, "k1", 1_700_000_000_000)
+	s := NewStats(0)
+	s.LoadAudit(path)
+
+	var seen []Req
+	cursor := ""
+	for i := 0; i < 20; i++ { // bounded loop guards against a cursor that never advances
+		page := s.AuditPage(cursor, 40, "")
+		if len(page.Records) == 0 {
+			break
+		}
+		seen = append(seen, page.Records...)
+		if !page.HasMore {
+			break
+		}
+		if page.Next == cursor {
+			t.Fatal("cursor did not advance")
+		}
+		cursor = page.Next
+	}
+	if len(seen) != n {
+		t.Fatalf("paged %d records, want %d", len(seen), n)
+	}
+	for i := 1; i < len(seen); i++ {
+		if seen[i-1].Time < seen[i].Time {
+			t.Fatalf("records not newest-first at %d: %d then %d", i, seen[i-1].Time, seen[i].Time)
+		}
+	}
+	// newest record first, oldest last
+	if seen[0].Time != 1_700_000_000_000+int64(n-1)*1000 {
+		t.Fatalf("first record time = %d, want the newest", seen[0].Time)
+	}
+	if seen[len(seen)-1].Time != 1_700_000_000_000 {
+		t.Fatalf("last record time = %d, want the oldest", seen[len(seen)-1].Time)
+	}
+}
+
+// TestAuditPageFirstScreenIsSmall: the first page must be exactly one screen,
+// independent of how large the audit file is.
+func TestAuditPageFirstScreenIsSmall(t *testing.T) {
+	path := filepath.Join(t.TempDir(), "audit.jsonl")
+	writeAuditFile(t, path, 5000, "k1", 1_700_000_000_000)
+	s := NewStats(0)
+	s.LoadAudit(path)
+	page := s.AuditPage("", 100, "")
+	if len(page.Records) != 100 {
+		t.Fatalf("first page = %d records, want 100", len(page.Records))
+	}
+	if !page.HasMore || page.Next == "" {
+		t.Fatal("a large file must report more pages plus a cursor")
+	}
+}
+
+// TestAuditPageFiltersByKey: a user key may only page its own records.
+func TestAuditPageFiltersByKey(t *testing.T) {
+	dir := t.TempDir()
+	path := filepath.Join(dir, "audit.jsonl")
+	writeAuditFile(t, path, 30, "mine", 1_700_000_000_000)
+	f, err := os.OpenFile(path, os.O_APPEND|os.O_WRONLY, 0644)
+	if err != nil {
+		t.Fatal(err)
+	}
+	for i := 0; i < 30; i++ {
+		fmt.Fprintf(f, `{"time":%d,"key":"other","type":"chat","model":"m","source":"s","ok":true,"status":200}`+"\n",
+			1_700_000_100_000+int64(i)*1000)
+	}
+	f.Close()
+
+	s := NewStats(0)
+	s.LoadAudit(path)
+	page := s.AuditPage("", 1000, "mine")
+	if len(page.Records) != 30 {
+		t.Fatalf("filtered page = %d records, want 30", len(page.Records))
+	}
+	for _, r := range page.Records {
+		if r.Key != "mine" {
+			t.Fatalf("key filter leaked record for %q", r.Key)
+		}
+	}
+}
+
+// TestAuditPageSpansRotatedFiles: paging must continue into .old files instead
+// of stopping at the live file's start.
+func TestAuditPageSpansRotatedFiles(t *testing.T) {
+	dir := t.TempDir()
+	path := filepath.Join(dir, "audit.jsonl")
+	writeAuditFile(t, path, 40, "k1", 1_700_000_500_000)            // newest
+	writeAuditFile(t, path+".100.old", 40, "k1", 1_700_000_300_000) // older
+	writeAuditFile(t, path+".050.old", 40, "k1", 1_700_000_100_000) // oldest
+
+	s := NewStats(0)
+	s.LoadAudit(path)
+	var seen []Req
+	cursor := ""
+	for i := 0; i < 30; i++ {
+		page := s.AuditPage(cursor, 25, "")
+		seen = append(seen, page.Records...)
+		if !page.HasMore {
+			break
+		}
+		cursor = page.Next
+	}
+	if len(seen) != 120 {
+		t.Fatalf("paged %d records across rotated files, want 120", len(seen))
+	}
+	for i := 1; i < len(seen); i++ {
+		if seen[i-1].Time < seen[i].Time {
+			t.Fatalf("rotation boundary broke the ordering at %d", i)
+		}
+	}
+}
+
+// TestAuditPageRotatedCursor: a cursor whose file rotated away must restart from
+// the newest page and say so, rather than silently returning nothing.
+func TestAuditPageRotatedCursor(t *testing.T) {
+	path := filepath.Join(t.TempDir(), "audit.jsonl")
+	writeAuditFile(t, path, 20, "k1", 1_700_000_000_000)
+	s := NewStats(0)
+	s.LoadAudit(path)
+	page := s.AuditPage("/gone/audit.jsonl:12345", 10, "")
+	if !page.Rotated {
+		t.Fatal("a stale cursor must be reported as rotated")
+	}
+	if len(page.Records) == 0 {
+		t.Fatal("a stale cursor must fall back to the newest page")
+	}
+}
+
+// TestAuditPageEmptyAndNoFile cover the degenerate inputs.
+func TestAuditPageEmptyAndNoFile(t *testing.T) {
+	s := NewStats(0)
+	if page := s.AuditPage("", 10, ""); len(page.Records) != 0 || page.HasMore {
+		t.Fatalf("no audit path must yield an empty page, got %+v", page)
+	}
+	path := filepath.Join(t.TempDir(), "audit.jsonl")
+	if err := os.WriteFile(path, nil, 0644); err != nil {
+		t.Fatal(err)
+	}
+	s2 := NewStats(0)
+	s2.LoadAudit(path)
+	if page := s2.AuditPage("", 10, ""); len(page.Records) != 0 || page.HasMore {
+		t.Fatalf("empty audit file must yield an empty page, got %+v", page)
+	}
+}
+
+// TestAuditPageHandlesLongLines: a record larger than the reverse-read chunk
+// must still be decoded (the reader carries partial lines across chunks).
+func TestAuditPageHandlesLongLines(t *testing.T) {
+	path := filepath.Join(t.TempDir(), "audit.jsonl")
+	f, err := os.Create(path)
+	if err != nil {
+		t.Fatal(err)
+	}
+	big := strings.Repeat("e", revChunk+1024) // one record spanning >1 chunk
+	fmt.Fprintf(f, `{"time":1,"key":"k","type":"chat","model":"m","source":"s","ok":false,"status":500,"error":%q}`+"\n", big)
+	fmt.Fprintf(f, `{"time":2,"key":"k","type":"chat","model":"m","source":"s","ok":true,"status":200}`+"\n")
+	f.Close()
+
+	s := NewStats(0)
+	s.LoadAudit(path)
+	page := s.AuditPage("", 10, "")
+	if len(page.Records) != 2 {
+		t.Fatalf("got %d records, want 2 (a >chunk record must survive reverse reads)", len(page.Records))
+	}
+	if page.Records[0].Time != 2 || page.Records[1].Time != 1 {
+		t.Fatalf("ordering broken: %d then %d", page.Records[0].Time, page.Records[1].Time)
+	}
+	if len(page.Records[1].Err) != len(big) {
+		t.Fatalf("oversized field truncated: %d bytes, want %d", len(page.Records[1].Err), len(big))
+	}
+}
+
+// TestStreamAuditRecordsIsWindowed: export streaming must honour the time window
+// and the key filter, and must hand records over one at a time.
+func TestStreamAuditRecordsIsWindowed(t *testing.T) {
+	path := filepath.Join(t.TempDir(), "audit.jsonl")
+	writeAuditFile(t, path, 100, "k1", 1_700_000_000_000)
+	s := NewStats(0)
+	s.LoadAudit(path)
+
+	from := int64(1_700_000_010_000)
+	to := int64(1_700_000_019_000)
+	var got []Req
+	if err := s.StreamAuditRecords(from, to, "", func(r Req) error {
+		got = append(got, r)
+		return nil
+	}); err != nil {
+		t.Fatalf("stream: %v", err)
+	}
+	if len(got) != 10 {
+		t.Fatalf("streamed %d records, want 10 in [%d,%d]", len(got), from, to)
+	}
+	for _, r := range got {
+		if r.Time < from || r.Time > to {
+			t.Fatalf("record outside the window: %d", r.Time)
+		}
+	}
+}
+
+// TestStreamAuditRecordsAbortStops: a callback error (client disconnect) must
+// stop the walk immediately instead of reading the rest of the history.
+func TestStreamAuditRecordsAbortStops(t *testing.T) {
+	path := filepath.Join(t.TempDir(), "audit.jsonl")
+	writeAuditFile(t, path, 500, "k1", 1_700_000_000_000)
+	s := NewStats(0)
+	s.LoadAudit(path)
+
+	seen := 0
+	stop := fmt.Errorf("client gone")
+	err := s.StreamAuditRecords(0, 0, "", func(Req) error {
+		seen++
+		if seen == 5 {
+			return stop
+		}
+		return nil
+	})
+	if err != stop {
+		t.Fatalf("abort must surface the callback error, got %v", err)
+	}
+	if seen != 5 {
+		t.Fatalf("walk continued after abort: %d records", seen)
+	}
+}
+
+// TestAuditRotationThresholds pins the smaller-file/more-files trade-off that
+// makes reverse paging cheap while keeping the same total budget.
+func TestAuditRotationThresholds(t *testing.T) {
+	if auditRotateBytes != 16<<20 {
+		t.Fatalf("auditRotateBytes = %d, want 16 MiB", auditRotateBytes)
+	}
+	if auditKeepOld != 16 {
+		t.Fatalf("auditKeepOld = %d, want 16", auditKeepOld)
+	}
+	if total := auditRotateBytes * int64(auditKeepOld); total != 256<<20 {
+		t.Fatalf("total audit budget = %d, want 256 MiB", total)
+	}
+}

internal/gateway/token_units_test.go

@@ -0,0 +1,231 @@
+package gateway
+
+import (
+	"encoding/json"
+	"fmt"
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"testing"
+
+	"llmsproxy/internal/config"
+)
+
+// Token accounting must record TOKENS, not byte estimates and not image
+// counts. These tests pin the two units bugs found on 2026-09-27:
+//
+//  1. streamed chat: rec.Compl was only ever the per-chunk byte estimate
+//     (len(content)/3), even when the upstream sent exact usage on its final
+//     frame. Measured on production traffic: the same model reported a
+//     streamed completion median 4-27x its non-streamed median. Upstream 100
+//     recorded as 145.
+//  2. image generation: rec.Compl was set to len(resp.ImageData) — the image
+//     COUNT — and folded into the token totals.
+//
+// A fake upstream that reports FIXED numbers makes both mismatches
+// unambiguous: any difference between what it reports and what the audit
+// record holds is the gateway's doing.
+
+// fixedUsageUpstream reports prompt=prompt completion=compl on both the
+// streaming and the non-streaming path. The streaming path emits usage on its
+// final frame before [DONE], exactly like OpenAI-compatible upstreams.
+func fixedUsageUpstream(t *testing.T, prompt, compl int, images int) *httptest.Server {
+	t.Helper()
+	return httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		var body map[string]interface{}
+		_ = json.NewDecoder(r.Body).Decode(&body)
+		stream, _ := body["stream"].(bool)
+		usage := fmt.Sprintf(`{"prompt_tokens":%d,"completion_tokens":%d,"total_tokens":%d}`,
+			prompt, compl, prompt+compl)
+
+		// Image generation wire: /v1/images/generations style body.
+		if strings.Contains(r.URL.Path, "image") {
+			w.Header().Set("Content-Type", "application/json")
+			var data []string
+			for i := 0; i < images; i++ {
+				data = append(data, `{"b64_json":"AAAA"}`)
+			}
+			fmt.Fprintf(w, `{"created":1,"data":[%s]}`, strings.Join(data, ","))
+			return
+		}
+
+		if !stream {
+			w.Header().Set("Content-Type", "application/json")
+			fmt.Fprintf(w, `{"model":"m1","choices":[{"message":{"content":"pong pong pong"},
+				"finish_reason":"stop"}],"usage":%s}`, usage)
+			return
+		}
+
+		w.Header().Set("Content-Type", "text/event-stream")
+		w.WriteHeader(200)
+		fl, _ := w.(http.Flusher)
+		send := func(s string) {
+			fmt.Fprintf(w, "data: %s\n\n", s)
+			if fl != nil {
+				fl.Flush()
+			}
+		}
+		// Deliberately verbose payloads: a byte-estimating path records a
+		// number far from `compl`, so the test would catch it.
+		send(`{"model":"m1","choices":[{"delta":{"reasoning_content":"` + strings.Repeat("think ", 40) + `"}}]}`)
+		send(`{"model":"m1","choices":[{"delta":{"content":"` + strings.Repeat("answer ", 40) + `"}}]}`)
+		send(`{"model":"m1","choices":[{"delta":{},"finish_reason":"stop"}]}`)
+		send(`{"model":"m1","choices":[],"usage":` + usage + `}`)
+		send(`[DONE]`)
+	}))
+}
+
+// lastChatRecord returns the most recent audit record whose Type is one of the
+// given types.
+func lastChatRecord(t *testing.T, g *Gateway, types ...string) Req {
+	t.Helper()
+	recs := g.stats.AuditRecords(0, 0, "")
+	want := map[string]bool{}
+	for _, ty := range types {
+		want[ty] = true
+	}
+	for i := len(recs) - 1; i >= 0; i-- {
+		if want[recs[i].Type] {
+			return recs[i]
+		}
+	}
+	t.Fatalf("no audit record of type %v found (have %d records)", types, len(recs))
+	return Req{}
+}
+
+// The streamed path must record the upstream's exact usage, not a byte
+// estimate. Before the fix this recorded 145 for an upstream-reported 100.
+func TestStreamedChatRecordsUpstreamUsage(t *testing.T) {
+	const wantPrompt, wantCompl = 44, 100
+	up := fixedUsageUpstream(t, wantPrompt, wantCompl, 0)
+	defer up.Close()
+	g := newTestGateway(t, config.Source{
+		Name: "up", BaseURL: up.URL, Adapter: "openai",
+		Models: []config.Model{{ID: "m1", Priority: 100}},
+	})
+
+	rr := doReq(t, g, "POST", "/v1/chat/completions",
+		`{"model":"m1","messages":[{"role":"user","content":"hi"}],"stream":true}`)
+	if rr.Code != 200 {
+		t.Fatalf("stream status=%d body=%s", rr.Code, rr.Body.String())
+	}
+
+	rec := lastChatRecord(t, g, "stream")
+	if rec.Prompt != wantPrompt || rec.Compl != wantCompl {
+		t.Errorf("streamed record prompt=%d completion=%d, want %d/%d — the audit "+
+			"must carry the upstream's tokens, not a byte estimate",
+			rec.Prompt, rec.Compl, wantPrompt, wantCompl)
+	}
+}
+
+// Streaming and non-streaming must agree on the same request's tokens. This is
+// the cross-path consistency check: the two paths previously recorded 145 and
+// 100 for one identical upstream usage value.
+func TestStreamedAndNonStreamedAgreeOnTokens(t *testing.T) {
+	const wantPrompt, wantCompl = 44, 100
+	up := fixedUsageUpstream(t, wantPrompt, wantCompl, 0)
+	defer up.Close()
+	g := newTestGateway(t, config.Source{
+		Name: "up", BaseURL: up.URL, Adapter: "openai",
+		Models: []config.Model{{ID: "m1", Priority: 100}},
+	})
+
+	if rr := doReq(t, g, "POST", "/v1/chat/completions",
+		`{"model":"m1","messages":[{"role":"user","content":"hi"}]}`); rr.Code != 200 {
+		t.Fatalf("non-stream status=%d body=%s", rr.Code, rr.Body.String())
+	}
+	if rr := doReq(t, g, "POST", "/v1/chat/completions",
+		`{"model":"m1","messages":[{"role":"user","content":"hi"}],"stream":true}`); rr.Code != 200 {
+		t.Fatalf("stream status=%d body=%s", rr.Code, rr.Body.String())
+	}
+
+	chat, stream := lastChatRecord(t, g, "chat"), lastChatRecord(t, g, "stream")
+	if chat.Prompt != stream.Prompt || chat.Compl != stream.Compl {
+		t.Errorf("chat=%d/%d stream=%d/%d — identical upstream usage must record "+
+			"identically on both paths", chat.Prompt, chat.Compl, stream.Prompt, stream.Compl)
+	}
+}
+
+// A fallback to the byte estimate is still correct WHEN the upstream reports no
+// usage at all. Without this the fix could silently zero out every streamed
+// record from upstreams that omit usage.
+func TestStreamedChatFallsBackToEstimateWithoutUsage(t *testing.T) {
+	up := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.Header().Set("Content-Type", "text/event-stream")
+		w.WriteHeader(200)
+		fl, _ := w.(http.Flusher)
+		send := func(s string) {
+			fmt.Fprintf(w, "data: %s\n\n", s)
+			if fl != nil {
+				fl.Flush()
+			}
+		}
+		send(`{"model":"m1","choices":[{"delta":{"content":"` + strings.Repeat("abc ", 60) + `"}}]}`)
+		send(`{"model":"m1","choices":[{"delta":{},"finish_reason":"stop"}]}`)
+		send(`[DONE]`)
+	}))
+	defer up.Close()
+	g := newTestGateway(t, config.Source{
+		Name: "up", BaseURL: up.URL, Adapter: "openai",
+		Models: []config.Model{{ID: "m1", Priority: 100}},
+	})
+
+	if rr := doReq(t, g, "POST", "/v1/chat/completions",
+		`{"model":"m1","messages":[{"role":"user","content":"hi"}],"stream":true}`); rr.Code != 200 {
+		t.Fatalf("status=%d body=%s", rr.Code, rr.Body.String())
+	}
+	rec := lastChatRecord(t, g, "stream")
+	if rec.Compl <= 0 {
+		t.Errorf("completion=%d, want a positive byte-estimate fallback when the "+
+			"upstream reports no usage", rec.Compl)
+	}
+}
+
+// Image generation has no token concept: recording the image COUNT as
+// completion tokens corrupted the token totals. The count is reported
+// separately and must not leak into Prompt/Compl.
+func TestImageGenerationDoesNotRecordTokens(t *testing.T) {
+	up := fixedUsageUpstream(t, 44, 100, 3)
+	defer up.Close()
+	src := config.Source{
+		Name: "up", BaseURL: up.URL, Adapter: "openai",
+		Models: []config.Model{{ID: "img-m", Priority: 100, Kind: "image"}},
+	}
+	g := newTestGateway(t, src)
+
+	rr := doReq(t, g, "POST", "/v1/images/generations",
+		`{"model":"img-m","prompt":"a cat","n":3}`)
+	if rr.Code != 200 {
+		t.Fatalf("image status=%d body=%s", rr.Code, rr.Body.String())
+	}
+
+	rec := lastChatRecord(t, g, "image")
+	if rec.Compl != 0 || rec.Prompt != 0 {
+		t.Errorf("image record prompt=%d completion=%d, want 0/0 — image count is "+
+			"not a token and must not enter the token totals", rec.Prompt, rec.Compl)
+	}
+	if rec.ImageCount != 3 {
+		t.Errorf("image_count=%d, want 3 (the count belongs in its own field)", rec.ImageCount)
+	}
+	// And the aggregate must see zero tokens for it.
+	if got := g.stats.WindowTokens("img-m", "", 3600); got != 0 {
+		t.Errorf("window tokens for an image model = %d, want 0", got)
+	}
+}
+
+// The audit JSON must carry image_count so the WebUI can show images rather
+// than a bogus token figure.
+func TestImageCountSurvivesAuditJSON(t *testing.T) {
+	b, err := json.Marshal(Req{Type: "image", ImageCount: 2, OK: true})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !strings.Contains(string(b), `"image_count":2`) {
+		t.Errorf("audit JSON %s lacks image_count", b)
+	}
+	// Omitempty keeps chat records unchanged.
+	b2, _ := json.Marshal(Req{Type: "chat", Prompt: 1, Compl: 2, OK: true})
+	if strings.Contains(string(b2), "image_count") {
+		t.Errorf("chat record %s should omit image_count", b2)
+	}
+}

internal/gateway/ui_contract_test.go

@@ -0,0 +1,406 @@
+package gateway
+
+import (
+	"fmt"
+	"regexp"
+	"strings"
+	"testing"
+)
+
+// The WebUI talks to this package through a tiny `api(path, options)` helper
+// that wraps fetch(). fetch() defaults to GET when no method is given, so an
+// options object without `method` silently turns a DELETE/PUT into a GET —
+// the request hits the wrong branch of the handler and the user sees a
+// confusing error like "use GET /api/keys" while nothing is deleted.
+//
+// This has now bitten twice: ab20f1b fixed delSource/delTemplate, and the very
+// same pattern was still live in delKey (deleting a gateway key was impossible)
+// and delAdapter. Reviewing 4000+ lines of inline JS by eye clearly does not
+// catch it, so it is pinned here instead.
+
+// apiCallRe finds the start of every api(...) call in the UI source.
+var apiCallRe = regexp.MustCompile(`\bapi\(`)
+
+// uiSource returns the embedded WebUI document.
+func uiSource(t *testing.T) string {
+	t.Helper()
+	b, err := uiFS.ReadFile("ui/index.html")
+	if err != nil {
+		t.Fatalf("read embedded ui: %v", err)
+	}
+	return string(b)
+}
+
+// extractCall returns the source of the balanced api(...) call starting at the
+// opening parenthesis index, plus the line number it starts on.
+func extractCall(src string, openIdx int) (string, bool) {
+	depth := 0
+	for i := openIdx; i < len(src); i++ {
+		switch src[i] {
+		case '(':
+			depth++
+		case ')':
+			depth--
+			if depth == 0 {
+				return src[openIdx : i+1], true
+			}
+		case '\n':
+			// A call spanning more than ~25 lines is not one of ours; bail out
+			// rather than scanning the rest of the document.
+			if i-openIdx > 4000 {
+				return "", false
+			}
+		}
+	}
+	return "", false
+}
+
+func lineOf(src string, idx int) int {
+	return strings.Count(src[:idx], "\n") + 1
+}
+
+// TestUIAPICallsDeclareMethod asserts that every api() call passing an options
+// object also declares an HTTP method (or is a GET that only passes an
+// AbortSignal). Without this, fetch defaults to GET and mutating endpoints are
+// never reached.
+func TestUIAPICallsDeclareMethod(t *testing.T) {
+	src := uiSource(t)
+	var offenders []string
+	for _, m := range apiCallRe.FindAllStringIndex(src, -1) {
+		open := m[1] - 1 // index of '('
+		call, ok := extractCall(src, open)
+		if !ok {
+			continue
+		}
+		// No options object at all => a plain GET, which is fine.
+		if !strings.Contains(call, "{") {
+			continue
+		}
+		if strings.Contains(call, "method:") {
+			continue
+		}
+		// A read that only carries an AbortSignal is a deliberate GET.
+		if strings.Contains(call, "signal:") && !strings.Contains(call, "body:") {
+			continue
+		}
+		offenders = append(offenders, fmt.Sprintf("line %d: %s", lineOf(src, open), oneLine(call)))
+	}
+	if len(offenders) > 0 {
+		t.Fatalf("api() called with an options object but no method (fetch will send GET):\n  %s",
+			strings.Join(offenders, "\n  "))
+	}
+}
+
+// TestUIDeleteHelpersUseDelete pins the specific helpers that remove things.
+// Each must issue a DELETE; a GET here is the exact bug users reported as
+// "cannot delete key: use GET /api/keys".
+func TestUIDeleteHelpersUseDelete(t *testing.T) {
+	src := uiSource(t)
+	for _, fn := range []string{"delKey", "delAdapter", "delSource", "delTemplate"} {
+		body, ok := jsFunctionBody(src, fn)
+		if !ok {
+			t.Errorf("helper %s() not found in the WebUI", fn)
+			continue
+		}
+		if !strings.Contains(body, "api(") {
+			t.Errorf("%s() does not call api()", fn)
+			continue
+		}
+		if !strings.Contains(body, `method: "DELETE"`) {
+			t.Errorf("%s() must issue a DELETE, got:\n%s", fn, body)
+		}
+	}
+}
+
+// TestUIMutatingHelpersUseWriteMethods covers the write (non-delete) helpers:
+// each must declare an explicit POST/PUT/PATCH.
+func TestUIMutatingHelpersUseWriteMethods(t *testing.T) {
+	src := uiSource(t)
+	want := map[string][]string{
+		"createKey":     {`method: "POST"`},
+		"putScope":      {`method: "PUT"`},
+		"uploadAdapter": {`method: "POST"`},
+		"saveAuto":      {`method: "PUT"`, `method: "POST"`},
+	}
+	for fn, methods := range want {
+		body, ok := jsFunctionBody(src, fn)
+		if !ok {
+			continue // helper renamed or absent: the method test above still guards it
+		}
+		found := false
+		for _, m := range methods {
+			if strings.Contains(body, m) {
+				found = true
+				break
+			}
+		}
+		if !found {
+			t.Errorf("%s() must declare one of %v", fn, methods)
+		}
+	}
+}
+
+// jsFunctionBody returns the source of `function name(` / `async function name(`
+// up to its closing brace, using brace balancing.
+func jsFunctionBody(src, name string) (string, bool) {
+	for _, prefix := range []string{"async function " + name + "(", "function " + name + "("} {
+		idx := strings.Index(src, prefix)
+		if idx < 0 {
+			continue
+		}
+		open := strings.Index(src[idx:], "{")
+		if open < 0 {
+			continue
+		}
+		open += idx
+		depth := 0
+		for i := open; i < len(src); i++ {
+			switch src[i] {
+			case '{':
+				depth++
+			case '}':
+				depth--
+				if depth == 0 {
+					return src[idx : i+1], true
+				}
+			}
+		}
+	}
+	return "", false
+}
+
+func oneLine(s string) string {
+	s = strings.Join(strings.Fields(s), " ")
+	if len(s) > 140 {
+		return s[:140] + "…"
+	}
+	return s
+}
+
+// TestUIRecordsPagingWiring guards the record table's paging plumbing. Two
+// separate bugs made "scroll to the bottom" silently stop loading:
+//
+//  1. paintRecords rebuilt the whole table on every 5s poll whenever the row
+//     count did not exceed the first screen, wiping loaded pages and resetting
+//     scroll position.
+//  2. IntersectionObserver only fires on TRANSITIONS. With a short list — or a
+//     page whose rows all duplicated the first screen — the sentinel stayed
+//     visible and never fired again.
+//
+// The fixes are a build-once flag, a scroll-position fallback, an initial
+// viewport fill, and chaining when a page yields no new rows. None of that is
+// reachable from Go, so the wiring is asserted structurally.
+func TestUIRecordsPagingWiring(t *testing.T) {
+	src := uiSource(t)
+
+	paint, ok := jsFunctionBody(src, "paintRecords")
+	if !ok {
+		t.Fatal("paintRecords() not found")
+	}
+	if !strings.Contains(paint, "recsState.built") {
+		t.Error("paintRecords must build the table once (recsState.built) instead of rebuilding on every poll")
+	}
+	if !strings.Contains(paint, "afterbegin") {
+		t.Error("paintRecords must PREPEND newer rows on later polls, not rebuild the table")
+	}
+	if !strings.Contains(paint, "fillRecordsViewport") {
+		t.Error("paintRecords must fill the viewport so the sentinel can transition")
+	}
+
+	attach, ok := jsFunctionBody(src, "attachRecsObserver")
+	if !ok {
+		t.Fatal("attachRecsObserver() not found")
+	}
+	if !strings.Contains(attach, "IntersectionObserver") {
+		t.Error("attachRecsObserver must still use IntersectionObserver")
+	}
+	if !strings.Contains(attach, `addEventListener("scroll"`) {
+		t.Error("attachRecsObserver needs a scroll fallback: an already-visible sentinel never fires again")
+	}
+
+	load, ok := jsFunctionBody(src, "loadMoreRecords")
+	if !ok {
+		t.Fatal("loadMoreRecords() not found")
+	}
+	if !strings.Contains(load, "maxChain") {
+		t.Error("loadMoreRecords must chain when a page adds no new rows (the first page overlaps the first screen)")
+	}
+	if !strings.Contains(load, "signal: ctrl.signal") {
+		t.Error("loadMoreRecords must remain abortable")
+	}
+
+	fill, ok := jsFunctionBody(src, "fillRecordsViewport")
+	if !ok {
+		t.Fatal("fillRecordsViewport() not found")
+	}
+	for _, want := range []string{"scrollHeight", "clientHeight"} {
+		if !strings.Contains(fill, want) {
+			t.Errorf("fillRecordsViewport must compare %s to decide whether the list overflows", want)
+		}
+	}
+
+	rel, ok := jsFunctionBody(src, "releaseRecords")
+	if !ok {
+		t.Fatal("releaseRecords() not found")
+	}
+	for _, want := range []string{"observer.disconnect()", "removeEventListener", "abort()", "recsState.built = false"} {
+		if !strings.Contains(rel, want) {
+			t.Errorf("releaseRecords must clean up %s", want)
+		}
+	}
+}
+
+// TestUIThemeMatrix pins the theme system: every accent must define both a light
+// and a dark variable block, so switching light/dark can never leave a theme
+// half-defined (which shows up as unreadable text rather than as an error).
+func TestUIThemeMatrix(t *testing.T) {
+	src := uiSource(t)
+
+	accents := []string{"sakura", "ocean", "violet", "mono"}
+	for _, a := range accents {
+		light := fmt.Sprintf(`html[data-accent=%q]{`, a)
+		dark := fmt.Sprintf(`html[data-theme="dark"][data-accent=%q]{`, a)
+		if !strings.Contains(src, light) {
+			t.Errorf("accent %q has no light-mode variable block (%s)", a, light)
+		}
+		// sakura is the default palette: its light values live in :root, and the
+		// generic dark block covers it.
+		if a != "sakura" && !strings.Contains(src, dark) {
+			t.Errorf("accent %q has no dark-mode variable block (%s)", a, dark)
+		}
+		if !strings.Contains(src, fmt.Sprintf(`data-accent="%s" title=`, a)) {
+			t.Errorf("accent %q has no picker button", a)
+		}
+	}
+}
+
+// TestUIMonoThemeIsFlat asserts the mono theme really is pure white / pure black:
+// the surfaces must be opaque and the glass/blob decoration must be switched off,
+// otherwise "#ffffff" renders as a grey translucent card over a gradient mesh.
+func TestUIMonoThemeIsFlat(t *testing.T) {
+	src := uiSource(t)
+
+	lightBlock, ok := cssBlock(src, `html[data-accent="mono"]{`)
+	if !ok {
+		t.Fatal("mono light block not found")
+	}
+	for _, want := range []string{"--bg-s1:#ffffff", "--card:#ffffff", "--glass:0px"} {
+		if !strings.Contains(lightBlock, want) {
+			t.Errorf("mono light theme must set %s, got:\n%s", want, lightBlock)
+		}
+	}
+	if !strings.Contains(lightBlock, "--blob1:transparent") {
+		t.Error("mono light theme must disable the gradient blobs")
+	}
+
+	darkBlock, ok := cssBlock(src, `html[data-theme="dark"][data-accent="mono"]{`)
+	if !ok {
+		t.Fatal("mono dark block not found")
+	}
+	for _, want := range []string{"--bg-s1:#000000", "--glass:0px"} {
+		if !strings.Contains(darkBlock, want) {
+			t.Errorf("mono dark theme must set %s, got:\n%s", want, darkBlock)
+		}
+	}
+	if !strings.Contains(darkBlock, "--blob1:transparent") {
+		t.Error("mono dark theme must disable the gradient blobs")
+	}
+
+	// the flat-surface overrides must exist, or backdrop-filter turns #fff grey
+	for _, want := range []string{
+		`html[data-accent="mono"] #bgfx .blob{display:none}`,
+		`backdrop-filter:none`,
+	} {
+		if !strings.Contains(src, want) {
+			t.Errorf("mono theme needs the flat-surface override %q", want)
+		}
+	}
+}
+
+// TestUICanvasColorsResolveVars guards a class of silent theme bug: Canvas 2D
+// does not understand CSS var(), so assigning "var(--x)" to strokeStyle is a
+// no-op that leaves the previous colour (often black) in place — invisible on a
+// dark background. Theme-driven canvas colours must go through cssVar().
+func TestUICanvasColorsResolveVars(t *testing.T) {
+	src := uiSource(t)
+	if !strings.Contains(src, "function cssVar(") {
+		t.Fatal("cssVar() helper missing: canvas colours cannot resolve CSS variables without it")
+	}
+	for _, prop := range []string{"strokeStyle", "fillStyle"} {
+		bad := fmt.Sprintf(`ctx.%s = "var(--`, prop)
+		if strings.Contains(src, bad) {
+			t.Errorf("ctx.%s assigned a raw CSS var(): Canvas ignores it silently, use cssVar()", prop)
+		}
+	}
+}
+
+// cssBlock returns the text of the CSS rule starting at selector, up to its
+// closing brace.
+func cssBlock(src, selector string) (string, bool) {
+	i := strings.Index(src, selector)
+	if i < 0 {
+		return "", false
+	}
+	j := strings.Index(src[i:], "}")
+	if j < 0 {
+		return "", false
+	}
+	return src[i : i+j+1], true
+}
+
+// The WebUI ships two locale objects (`zh` and `en`) that every rendered string
+// goes through. A key added to only one of them does not error: t() falls back
+// to `t("key") || "literal"` in some call sites and to the bare key string in
+// others, so the user sees either a hardcoded local string or a raw key name.
+// Nothing in the Go test suite noticed — this was found by hand while adding
+// thImgs. Pin the key sets so a one-sided edit fails here instead of shipping.
+func TestUILocaleKeyParity(t *testing.T) {
+	src := uiSource(t)
+	iZh := strings.Index(src, "zh: {")
+	iEn := strings.Index(src, "en: {")
+	if iZh < 0 || iEn < 0 || iEn < iZh {
+		t.Fatalf("locale blocks not found (zh=%d en=%d)", iZh, iEn)
+	}
+	zh := src[iZh:iEn]
+	en := src[iEn:]
+	// End the EN block at its closing brace, so keys from later objects
+	// (config templates etc.) do not pollute the comparison.
+	if end := strings.Index(en, "\n        },"); end > 0 {
+		en = en[:end]
+	}
+
+	keyRe := regexp.MustCompile(`(?m)^\s*([A-Za-z_][A-Za-z0-9_]*)\s*:`)
+	collect := func(seg string) map[string]bool {
+		m := map[string]bool{}
+		for _, g := range keyRe.FindAllStringSubmatch(seg, -1) {
+			m[g[1]] = true
+		}
+		return m
+	}
+	zk, ek := collect(zh), collect(en)
+	// The locale markers themselves are the two block headers.
+	delete(zk, "zh")
+	delete(ek, "en")
+
+	var missingEN, missingZH []string
+	for k := range zk {
+		if !ek[k] {
+			missingEN = append(missingEN, k)
+		}
+	}
+	for k := range ek {
+		if !zk[k] {
+			missingZH = append(missingZH, k)
+		}
+	}
+	if len(missingEN) > 0 {
+		t.Errorf("translation keys present in zh but missing in en: %v", missingEN)
+	}
+	if len(missingZH) > 0 {
+		t.Errorf("translation keys present in en but missing in zh: %v", missingZH)
+	}
+	if len(zk) < 100 {
+		t.Errorf("only %d zh keys parsed — the block boundary regexp has drifted "+
+			"and this test is no longer checking anything", len(zk))
+	}
+}

internal/gateway/ui_quota_contract_test.go

@@ -0,0 +1,267 @@
+package gateway
+
+import (
+	"regexp"
+	"strings"
+	"testing"
+)
+
+// The WebUI dialogs all share the id "modal-wrap", and more than one can be
+// open at the same time (the seed-key notice sits on top of the keys page).
+// A save handler that closes "the" modal via $("#modal-wrap") therefore removes
+// whichever one comes FIRST in the document — which is the wrong dialog: the
+// form the user just submitted stays on screen while an unrelated dialog
+// vanishes.
+//
+// This is the same class of bug the api() contract test pins: reviewing inline
+// JS by eye does not catch it, and the symptom ("the dialog did not close")
+// points away from the cause.
+
+// modalCloseRe finds every `$(...)`-style lookup of the shared modal id.
+var modalCloseRe = regexp.MustCompile(`\$\("#modal-wrap"\)`)
+
+// closestModalRe finds the safe form: resolve the dialog from the clicked
+// button instead of from the document.
+var closestModalRe = regexp.MustCompile(`\.closest\("#modal-wrap"\)`)
+
+func TestUIDialogClosesItselfNotTheFirstModal(t *testing.T) {
+	// Strip comments first: prose that *names* the unsafe pattern (as the fix's
+	// own comment does) would otherwise be flagged as a violation.
+	code := stripJSComments(uiSource(t))
+
+	for _, m := range modalCloseRe.FindAllStringIndex(code, -1) {
+		after := code[m[1]:]
+		stmtEnd := strings.Index(after, ";")
+		if stmtEnd < 0 || stmtEnd > 200 {
+			continue
+		}
+		stmt := after[:stmtEnd]
+		if strings.Contains(stmt, ".remove()") {
+			line := 1 + strings.Count(code[:m[0]], "\n")
+			t.Errorf("line %d closes the first #modal-wrap in the document, not its own dialog:\n\t%s",
+				line, strings.TrimSpace(stmt))
+		}
+	}
+}
+
+// stripJSComments removes // line comments and /* block */ comments from JS
+// embedded in the UI document. It is deliberately simple: the document is our
+// own source, and a false negative here only means the check is silent.
+func stripJSComments(src string) string {
+	var out strings.Builder
+	lines := strings.Split(src, "\n")
+	inBlock := false
+	for _, ln := range lines {
+		trimmed := strings.TrimSpace(ln)
+		if inBlock {
+			if strings.Contains(ln, "*/") {
+				inBlock = false
+			}
+			continue
+		}
+		if strings.HasPrefix(trimmed, "/*") {
+			if !strings.Contains(ln, "*/") {
+				inBlock = true
+			}
+			continue
+		}
+		if i := strings.Index(ln, "//"); i >= 0 {
+			before := ln[:i]
+			if strings.Count(before, `"`)%2 == 0 && strings.Count(before, "'")%2 == 0 {
+				ln = before
+			}
+		}
+		out.WriteString(ln)
+		out.WriteString("\n")
+	}
+	return out.String()
+}
+
+// Every handler that closes a dialog must do it one of the two safe ways.
+func TestUIDialogClosuresGoThroughSafePaths(t *testing.T) {
+	src := stripJSComments(uiSource(t))
+	for _, fn := range []string{
+		"downloadStatsCsv", "downloadKeysCsv", "saveSource", "saveTemplate",
+		"scrAddFromForm", "sortScopeSave", "scopeSave", "createKey",
+	} {
+		body, ok := jsFunctionBody(src, fn)
+		if !ok {
+			t.Errorf("%s not found", fn)
+			continue
+		}
+		if !strings.Contains(body, `closest("#modal-wrap")`) && !strings.Contains(body, "closeTopModal()") {
+			t.Errorf("%s closes a dialog with neither .closest nor closeTopModal", fn)
+		}
+	}
+	if !strings.Contains(src, "function closeTopModal(") {
+		t.Error("closeTopModal helper is missing")
+	}
+}
+
+// The helper must pick the LAST dialog (the topmost one the user sees), not the
+// first — that inversion is the whole bug.
+func TestUICloseTopModalTakesTheLast(t *testing.T) {
+	body, ok := jsFunctionBody(uiSource(t), "closeTopModal")
+	if !ok {
+		t.Fatal("closeTopModal not found")
+	}
+	if !strings.Contains(body, "all.length - 1") {
+		t.Errorf("closeTopModal does not take the last dialog:\n\t%s", oneLine(body))
+	}
+}
+
+// Handlers that resolve their dialog from a button must actually receive one:
+// a signature without the parameter means the .closest() silently yields null
+// and the save leaves its form stranded.
+func TestUIDialogHandlersReceiveTheirButton(t *testing.T) {
+	src := stripJSComments(uiSource(t))
+	for _, fn := range []string{
+		"downloadStatsCsv", "saveSource", "scrAddFromForm", "sortScopeSave",
+		"scopeSave", "createKey",
+	} {
+		body, ok := jsFunctionBody(src, fn)
+		if !ok {
+			t.Errorf("%s not found", fn)
+			continue
+		}
+		if !strings.Contains(body, "closest(\"#modal-wrap\")") {
+			continue // uses closeTopModal only
+		}
+		sig := body[:strings.Index(body, ")")+1]
+		if !strings.Contains(sig, "btn") {
+			t.Errorf("%s uses .closest(\"#modal-wrap\") but its signature %s has no button parameter —\n"+
+				"the lookup would always be null and the form would never close", fn, oneLine(sig))
+		}
+	}
+}
+
+// Quotas belong to the scope entries, so putScope only has to ship the scope
+// list — the caps travel inside it. What must NOT come back is a key-wide
+// total: it would be a second budget able to drift out of sync with the models
+// it is supposed to cover.
+func TestUIPutScopeShipsOnlyScopeQuotas(t *testing.T) {
+	body, ok := jsFunctionBody(uiSource(t), "putScope")
+	if !ok {
+		t.Fatal("putScope not found")
+	}
+	code := stripJSComments(body)
+	if !strings.Contains(code, "models:") || !strings.Contains(code, "scopes") {
+		t.Error("putScope must ship the scope list the caps live in")
+	}
+	for _, gone := range []string{"kquota", "kreqquota", "kperiod", "khours"} {
+		if strings.Contains(code, gone) {
+			t.Errorf("putScope still references the removed key-wide quota field %q", gone)
+		}
+	}
+}
+
+// A model brick carries its own budgets, and the scope editor reads and writes
+// both of them: dropping req_quota on the round trip would silently lift a
+// request cap every time someone edited a token cap.
+func TestUIScopeEditorRoundTripsBothQuotas(t *testing.T) {
+	src := stripJSComments(uiSource(t))
+	for _, fn := range []string{"scopeHtml", "readScopes", "scopeEdit", "scopeSave", "scopeQuotaBadge"} {
+		if _, ok := jsFunctionBody(src, fn); !ok {
+			t.Errorf("%s not found in the UI source", fn)
+		}
+	}
+	for _, fn := range []string{"scopeHtml", "readScopes", "scopeSave", "scopeQuotaBadge"} {
+		body, ok := jsFunctionBody(src, fn)
+		if !ok {
+			continue
+		}
+		if !strings.Contains(body, "req_quota") && !strings.Contains(body, "reqquota") {
+			t.Errorf("%s does not carry req_quota — a request cap would be lost on edit", fn)
+		}
+	}
+	if form, ok := jsFunctionBody(src, "scopeEdit"); ok && !strings.Contains(form, "sc-reqs") {
+		t.Error("the scope editor has no request-quota input")
+	}
+}
+
+// The whole key-wide quota surface must be gone from the UI: a badge or a
+// button reading a field the server no longer has would render "undefined" or
+// silently do nothing.
+func TestUIHasNoKeyWideQuotaSurface(t *testing.T) {
+	src := uiSource(t)
+	for _, gone := range []string{
+		"keyCapBadges", "keyQuotaEdit", "keyQuotaSave",
+		"kq-tokens", "kq-reqs", "kq-period", "kq-hours",
+		"kc-tokens", "kc-reqs", "kc-period", "kc-hours",
+		"data-kquota", "data-kreqquota", "data-kperiod", "data-khours",
+	} {
+		if strings.Contains(src, gone) {
+			t.Errorf("UI still references the removed key-wide quota surface %q", gone)
+		}
+	}
+}
+
+// Existence of the strings is not enough: the badge has to READ the scope's
+// fields, and the editor has to read back what the brick writes. A field can
+// be present in the source and still never reach the screen — e.g. left in a
+// dead branch, or read from a data attribute the writer never sets.
+func TestUIKeyQuotaDataflowIsLive(t *testing.T) {
+	src := stripJSComments(uiSource(t))
+
+	badge, ok := jsFunctionBody(src, "scopeQuotaBadge")
+	if !ok {
+		t.Fatal("scopeQuotaBadge not found")
+	}
+	for _, field := range []string{"m.token_quota", "m.req_quota", "m.period"} {
+		if !strings.Contains(badge, field) {
+			t.Errorf("scopeQuotaBadge does not read %q — the cap would never show on the model brick", field)
+		}
+	}
+
+	brick, ok := jsFunctionBody(src, "scopeHtml")
+	if !ok {
+		t.Fatal("scopeHtml not found")
+	}
+	edit, ok := jsFunctionBody(src, "scopeEdit")
+	if !ok {
+		t.Fatal("scopeEdit not found")
+	}
+	brickCode, editCode := stripJSComments(brick), stripJSComments(edit)
+	reader := stripJSComments(mustBody(t, src, "readScopes"))
+	for _, ds := range []string{"quota", "reqquota", "period", "hours"} {
+		if !strings.Contains(brickCode, "data-"+ds+"=") {
+			t.Errorf("scopeHtml does not write data-%s, so the editor has nothing to prefill", ds)
+		}
+		// the editor pre-fills through readScopes(), which is what walks the
+		// bricks' data attributes — check the reader, not the form
+		if !strings.Contains(reader, "dataset."+ds) {
+			t.Errorf("readScopes does not read dataset.%s — editing a brick would save zeros over it", ds)
+		}
+	}
+	// and the form must actually consume what readScopes produced
+	for _, field := range []string{"sc.token_quota", "sc.req_quota", "sc.period", "sc.hours"} {
+		if !strings.Contains(editCode, field) {
+			t.Errorf("scopeEdit does not prefill from %q", field)
+		}
+	}
+}
+
+func mustBody(t *testing.T, src, fn string) string {
+	t.Helper()
+	b, ok := jsFunctionBody(src, fn)
+	if !ok {
+		t.Fatalf("%s not found", fn)
+	}
+	return b
+}
+
+// The quota period vocabulary must match the server's, or the UI can offer a
+// value the API rejects.
+func TestUIQuotaPeriodsMatchServer(t *testing.T) {
+	src := uiSource(t)
+	for _, p := range []string{`value=""`, `value="hour"`, `value="week"`, `value="month"`, `value="nhour"`} {
+		if !strings.Contains(src, p) {
+			t.Errorf("UI period select is missing %s", p)
+		}
+	}
+	for _, p := range []string{`"hour"`, `"week"`, `"month"`, `"nhour"`} {
+		if !strings.Contains(src, `if (p === `+p+`)`) && !strings.Contains(src, `=== `+p+`)`) {
+			t.Errorf("periodText() does not describe %s, so a badge would omit the window", p)
+		}
+	}
+}

internal/lua/adapters/agentrouter.lua

@@ -0,0 +1,212 @@
+local adapter = {}
+
+adapter.name = "agentrouter"
+adapter.version = "1.0.0"
+adapter.endpoint = "/chat/completions"
+adapter.headers = {}
+
+-- AgentRouter 的阿里云 WAF 按三重维度校验：出口 IP（必须海外）、
+-- TLS 指纹（JS/Go HTTP 层与 curl 不同）、浏览器指纹 header 集。
+-- 纯 UA 已不够：WAF 拦 12 种 agent CLI UA 中的 10 种（全 405 HTML），
+-- 只有同时带 Origin: qwen.ai + Referer + X-Requested-With 才放行，
+-- 且同 IP 短时间连续 3+ 次请求会被速率封禁，靠 adapter 侧低频调用 +
+-- llmsproxy 的 prefFailStep 冷却自然限制，来源无法控制速率。
+local default_ua = "QwenCode/0.2.0 (linux; x64)"
+
+-- AgentRouter fronts Claude models (claude-opus-4-8), and Claude upstreams
+-- enforce Anthropic's tool id rule ^[a-zA-Z0-9_-]{1,64}$ by rejecting the WHOLE
+-- request. Plain OpenAI does not, so an id minted by a permissive model (e.g.
+-- "bash:0" from moonshotai/kimi-k3) is replayed here by the client and kills
+-- every turn. See anthropic.lua / openai.lua for the same helper: Lua adapters
+-- have no shared prelude, so each carries its own copy.
+local TOOL_ID_MAX = 64
+
+local function safe_tool_id(id)
+    if type(id) ~= "string" or id == "" then return id end
+    local clean = string.gsub(id, "[^A-Za-z0-9_-]", "_")
+    if clean == id and #clean <= TOOL_ID_MAX then
+        return clean
+    end
+    local digest = string.sub(sha256_hex(id), 1, 8)
+    local keep = TOOL_ID_MAX - #digest - 1
+    if #clean > keep then clean = string.sub(clean, 1, keep) end
+    return clean .. "_" .. digest
+end
+
+function adapter.transform_request(raw_body)
+    local ok, req = pcall(json.decode, raw_body)
+    if not ok then return raw_body end
+    req.disable_thinking = nil
+    req.extra_body = nil
+    if req.messages then
+        for _, msg in ipairs(req.messages) do
+            msg.reasoning_content = nil
+            if msg.tool_call_id ~= nil then
+                msg.tool_call_id = safe_tool_id(msg.tool_call_id)
+            end
+            if type(msg.tool_calls) == "table" then
+                for _, tc in ipairs(msg.tool_calls) do
+                    if type(tc) == "table" and tc.id ~= nil then
+                        tc.id = safe_tool_id(tc.id)
+                    end
+                end
+            end
+        end
+    end
+    return json.encode(req)
+end
+
+-- 注入 Authorization + 官方客户端 User-Agent
+function adapter.build_headers(meta)
+    local ua = default_ua
+    local meta_ua = (meta.source or {}).meta and (meta.source).meta.user_agent
+    if meta_ua and meta_ua ~= "" then
+        ua = meta_ua
+    end
+    return {
+        ["Content-Type"] = "application/json",
+        ["Authorization"] = "Bearer " .. meta.api_key,
+        ["User-Agent"] = ua,
+        -- QwenCode client fingerprint: WAF rejects bare UA without these.
+        ["Accept"] = "application/json, text/plain, */*",
+        ["Accept-Language"] = "zh-CN,zh;q=0.9",
+        ["Origin"] = "https://qwen.ai",
+        ["Referer"] = "https://qwen.ai/",
+        ["X-Requested-With"] = "XMLHttpRequest",
+        ["Connection"] = "keep-alive",
+    }
+end
+
+function adapter.transform_response(raw_body)
+    local ok, resp = pcall(json.decode, raw_body)
+    if not ok or resp == nil then return raw_body end
+
+    local unified = {
+        content = "",
+        finish_reason = "",
+        token_usage = { prompt = 0, completion = 0, total = 0 }
+    }
+
+    if type(resp.usage) == "table" then
+        unified.token_usage.prompt = resp.usage.prompt_tokens or 0
+        unified.token_usage.completion = resp.usage.completion_tokens or 0
+        unified.token_usage.total = resp.usage.total_tokens or 0
+        local hit = 0
+        if type(resp.usage.prompt_tokens_details) == "table" and resp.usage.prompt_tokens_details.cached_tokens ~= nil then
+            hit = resp.usage.prompt_tokens_details.cached_tokens
+            unified.token_usage.prompt_tokens_details = { cached_tokens = hit }
+        end
+        if (resp.usage.prompt_cache_hit_tokens or 0) > 0 then
+            unified.token_usage.prompt_cache_hit_tokens = resp.usage.prompt_cache_hit_tokens
+            unified.token_usage.prompt_cache_miss_tokens = resp.usage.prompt_cache_miss_tokens or 0
+            if hit == 0 then
+                unified.token_usage.prompt_tokens_details = { cached_tokens = resp.usage.prompt_cache_hit_tokens }
+            end
+        end
+    end
+
+    if type(resp.choices) == "table" and #resp.choices > 0 then
+        local ch = resp.choices[1]
+        if type(ch.message) == "table" then
+            unified.content = ch.message.content or ""
+            if ch.message.reasoning_content then
+                unified.reasoning_content = ch.message.reasoning_content
+            end
+            if type(ch.message.tool_calls) == "table" then
+                local tcs = {}
+                for _, tc in ipairs(ch.message.tool_calls) do
+                    local args_ok, args = pcall(json.decode, tc["function"].arguments)
+                    if not args_ok then args = {} end
+                    table.insert(tcs, {
+                        id = safe_tool_id(tc.id),
+                        type = tc.type or "function",
+                        name = tc["function"].name,
+                        arguments = args
+                    })
+                end
+                unified.tool_calls = tcs
+            end
+        end
+        unified.finish_reason = ch.finish_reason or ""
+    end
+
+    return json.encode(unified)
+end
+
+function adapter.transform_stream_chunk(raw_chunk)
+    local ok, chunk = pcall(json.decode, raw_chunk)
+    if not ok then return "" end
+
+    -- usage-only terminal chunk (empty choices + usage): forward it so the
+    -- gateway emits exact token counts (and cache fields when present).
+    local uses = nil
+    if type(chunk.usage) == "table" then
+        uses = {
+            prompt = chunk.usage.prompt_tokens or chunk.usage.prompt or 0,
+            completion = chunk.usage.completion_tokens or chunk.usage.completion or 0,
+            total = chunk.usage.total_tokens or chunk.usage.total or 0,
+        }
+        if type(chunk.usage.prompt_tokens_details) == "table" and chunk.usage.prompt_tokens_details.cached_tokens ~= nil then
+            uses.prompt_tokens_details = { cached_tokens = chunk.usage.prompt_tokens_details.cached_tokens }
+        elseif (chunk.usage.prompt_cache_hit_tokens or 0) > 0 then
+            uses.prompt_cache_hit_tokens = chunk.usage.prompt_cache_hit_tokens
+            uses.prompt_cache_miss_tokens = chunk.usage.prompt_cache_miss_tokens or 0
+            uses.prompt_tokens_details = { cached_tokens = chunk.usage.prompt_cache_hit_tokens }
+        end
+    end
+
+    if not chunk.choices or #chunk.choices == 0 then
+        if uses ~= nil then
+            return json.encode({ usage = uses, done = false })
+        end
+        return ""
+    end
+    local delta = chunk.choices[1].delta or {}
+    local fr = chunk.choices[1].finish_reason
+
+    local finish = (type(fr) == "string" and fr ~= "") and fr or nil
+
+    local unified = {
+        content = delta.content or "",
+        done = (finish ~= nil)
+    }
+    if finish then
+        unified.finish_reason = finish
+    end
+    if delta.reasoning_content then
+        unified.reasoning_content = delta.reasoning_content
+    end
+    if delta.tool_calls then
+        -- Sanitize outbound too: a dirty id must never enter a client session,
+        -- because the client replays it to every other source. Only the first
+        -- fragment of a streamed call carries an id; later argument fragments
+        -- have none and must stay id-less for index-based accumulation.
+        for _, tc in ipairs(delta.tool_calls) do
+            if type(tc) == "table" and tc.id ~= nil then
+                tc.id = safe_tool_id(tc.id)
+            end
+        end
+        unified.tool_calls = delta.tool_calls
+    end
+    if uses ~= nil then
+        unified.usage = uses
+    end
+    return json.encode(unified)
+end
+
+-- 错误收敛：前置 WAF 会返回整页 HTML（阿里云盾），JSON 时为 new-api 风格
+function adapter.transform_error(status, body)
+    local low = string.lower(body or "")
+    if string.sub(low, 1, 9) == "<!doctype" or string.find(low, "<html", 1, true) then
+        return "blocked by AgentRouter WAF"
+    end
+    local ok, resp = pcall(json.decode, body)
+    if not ok or type(resp) ~= "table" then return nil end
+    local e = resp.error
+    if type(e) == "table" and type(e.message) == "string" then
+        return e.message
+    end
+    return nil
+end
+
+return adapter

internal/lua/adapters/anthropic.lua

@@ -1,144 +1,447 @@
 local adapter = {}
-
 adapter.name = "anthropic"
-adapter.version = "2.0.0"
+adapter.version = "3.0.0"
 adapter.endpoint = "/v1/messages"
 adapter.headers = {
-    ["anthropic-version"] = "2023-06-01"
+    ["anthropic-version"] = "2023-06-01",
 }
 
-function adapter.transform_request(raw_body)
-    local ok, req = pcall(json.decode, raw_body)
-    if not ok then return raw_body end
+-- ============================================================
+-- helpers
+-- ============================================================
 
-    -- 将 OpenAI 风格 content（字符串或 [{type:*}] 数组）拆成文本/图片块
-    local function collect_blocks(content)
-        if type(content) == "string" then
-            return { { type = "text", text = content } }
-        end
-        local blocks = {}
-        for _, p in ipairs(content or {}) do
+local ROLE_WHITELIST = {
+    system = true, user = true, assistant = true, tool = true,
+}
+
+-- Anthropic requires tool_use.id / tool_result.tool_use_id to match
+-- ^[a-zA-Z0-9_-]{1,64}$ and rejects the WHOLE request otherwise with
+-- REQUEST_BODY_INVALID / "Invalid tool use format". OpenAI imposes no such
+-- rule, so an OpenAI-compatible upstream can hand a client an id like
+-- "bash:0" (observed from moonshotai/kimi-k3); once that lands in the
+-- client's history, every later replay kills every Claude slot at once and
+-- an AUTO chain falls through all tiers.
+--
+-- safe_tool_id is pure and deterministic, so a tool_use block and its
+-- matching tool_result are rewritten identically within one request. A
+-- rewritten id keeps an 8-hex digest of the ORIGINAL id: without it two
+-- distinct ids could collapse into one ("a:b" and "a_b"), which Anthropic
+-- rejects as a duplicate/unpaired tool_use. Already-legal ids pass through
+-- untouched so upstreams that round-trip their own ids are unaffected.
+local TOOL_ID_MAX = 64
+
+local function safe_tool_id(id)
+    if type(id) ~= "string" or id == "" then return "" end
+    local clean = string.gsub(id, "[^A-Za-z0-9_-]", "_")
+    if clean == id and #clean <= TOOL_ID_MAX then
+        return clean
+    end
+    local digest = string.sub(sha256_hex(id), 1, 8)
+    local keep = TOOL_ID_MAX - #digest - 1
+    if #clean > keep then clean = string.sub(clean, 1, keep) end
+    return clean .. "_" .. digest
+end
+
+-- map_usage converts one Anthropic usage object to the gateway's TokenUsage
+-- shape. The two APIs disagree on what the prompt count MEANS:
+--
+--   Anthropic: input_tokens EXCLUDES cached blocks. cache_read_input_tokens and
+--              cache_creation_input_tokens are separate, additive, billed input.
+--   OpenAI:    prompt_tokens INCLUDES its cached_tokens subset.
+--
+-- Mapping input_tokens straight onto prompt therefore did two wrong things at
+-- once: it undercounted the billed prompt by the entire cache portion, and it
+-- could report cached_tokens > prompt_tokens — a cache hit rate above 100%.
+-- cache_creation_input_tokens was not read at all, so a cache-write turn lost
+-- those billed tokens outright. Summing all three restores OpenAI semantics,
+-- which is what clients (and the gateway's own audit trail) assume.
+--
+-- Returns nil when the object carries no counts, so callers can keep treating
+-- "no usage in this chunk" as "say nothing" rather than reporting zeros.
+local function map_usage(u)
+    if type(u) ~= "table" then return nil end
+    local fresh  = u.input_tokens or 0
+    local read   = u.cache_read_input_tokens or 0
+    local create = u.cache_creation_input_tokens or 0
+    local p = fresh + read + create
+    local c = u.output_tokens or 0
+    if p <= 0 and c <= 0 then return nil end
+    local uses = { prompt = p, completion = c, total = p + c }
+    -- Emit the split whenever Anthropic reports either cache field, even at 0:
+    -- a reported zero-hit ("cache missed") must stay distinguishable from "this
+    -- upstream never reports cache info at all".
+    if u.cache_read_input_tokens ~= nil or u.cache_creation_input_tokens ~= nil then
+        uses.prompt_tokens_details = { cached_tokens = read }
+        uses.prompt_cache_hit_tokens = read
+        uses.prompt_cache_miss_tokens = fresh + create
+    end
+    return uses
+end
+
+local function collect_blocks(content)
+    if type(content) == "string" then
+        if content == "" then return {} end
+        return { { type = "text", text = content } }
+    end
+    if type(content) ~= "table" then return {} end
+    local blocks = {}
+    for _, p in ipairs(content) do
+        if type(p) == "string" then
+            table.insert(blocks, { type = "text", text = p })
+        elseif type(p) == "table" then
             if p.type == "text" then
-                table.insert(blocks, { type = "text", text = p.text })
+                table.insert(blocks, { type = "text", text = p.text or "" })
             elseif p.type == "image_url" and type(p.image_url) == "table" and p.image_url.url then
-                local mt, b64 = string.match(p.image_url.url, "^data:([^,]+);base64,(.+)$")
+                local url = p.image_url.url
+                local mt, b64 = string.match(url, "^data:([^,]+);base64,(.+)$")
                 if b64 then
                     table.insert(blocks, { type = "image", source = { type = "base64", media_type = mt or "image/png", data = b64 } })
                 else
-                    table.insert(blocks, { type = "image", source = { type = "url", url = p.image_url.url } })
+                    table.insert(blocks, { type = "image", source = { type = "url", url = url } })
                 end
+            else
+                -- Unknown content type: preserve for forward compatibility
+                table.insert(blocks, p)
             end
         end
-        return blocks
     end
-    local function text_of(content)
-        if type(content) == "string" then return content end
-        local t = ""
-        for _, p in ipairs(content or {}) do
-            if p.type == "text" and p.text then t = t .. p.text end
+    return blocks
+end
+
+local function text_of(content)
+    if type(content) == "string" then return content end
+    local t = ""
+    for _, p in ipairs(content or {}) do
+        if type(p) == "string" then
+            t = t .. p
+        elseif type(p) == "table" and p.type == "text" and p.text then
+            t = t .. p.text
         end
-        return t
     end
+    return t
+end
+
+--- Append blocks to the last user message (merge) or create a new one.
+local function append_user(msgs, blocks)
+    if #msgs > 0 and msgs[#msgs].role == "user" then
+        for _, b in ipairs(blocks) do
+            table.insert(msgs[#msgs].content, b)
+        end
+    else
+        table.insert(msgs, { role = "user", content = blocks })
+    end
+end
+
+-- ============================================================
+-- transform_request  (OpenAI → Anthropic)
+-- ============================================================
+
+function adapter.transform_request(raw_body)
+    local ok, req = pcall(json.decode, raw_body)
+    if not ok or type(req) ~= "table" then return raw_body end
 
     local msgs = {}
     local system = ""
+    local pending_tool = {}  -- accumulated tool_result blocks
+
     for _, m in ipairs(req.messages or {}) do
-        if m.role == "system" then
+        local role = m.role or "user"
+
+        if role == "system" then
             system = system .. text_of(m.content) .. "\n"
+
+        elseif role == "assistant" then
+            -- Build content array: text/image blocks from content + tool_use from tool_calls
+            local blocks = collect_blocks(m.content)
+
+            if type(m.tool_calls) == "table" then
+                for _, tc in ipairs(m.tool_calls) do
+                    if type(tc) == "table" then
+                        local fn = tc["function"] or {}
+                        local args = fn.arguments
+                        local input = {}
+                        if type(args) == "string" and args ~= "" then
+                            local ok2, parsed = pcall(json.decode, args)
+                            if ok2 and type(parsed) == "table" then input = parsed end
+                        elseif type(args) == "table" then
+                            input = args
+                        end
+                        table.insert(blocks, {
+                            type  = "tool_use",
+                            id    = safe_tool_id(tc.id),
+                            name  = fn.name or "",
+                            input = input,
+                        })
+                    end
+                end
+            end
+
+            table.insert(msgs, { role = "assistant", content = blocks })
+
+        elseif role == "tool" then
+            -- Accumulate consecutive tool results; will be flushed as one
+            -- user message with tool_result content blocks.
+            table.insert(pending_tool, {
+                type        = "tool_result",
+                tool_use_id = safe_tool_id(m.tool_call_id),
+                content     = text_of(m.content),
+            })
+
         else
-            table.insert(msgs, { role = m.role, content = collect_blocks(m.content) })
+            -- user / any other role: flush pending tool results first
+            if #pending_tool > 0 then
+                append_user(msgs, pending_tool)
+                pending_tool = {}
+            end
+            append_user(msgs, collect_blocks(m.content))
         end
     end
 
-    local anthropic_req = {
-        model = req.model or "claude-sonnet-4-20250514",
-        max_tokens = req.max_tokens or 4096,
-        messages = msgs,
-        stream = req.stream or false,
-    }
-
-    if not req.disable_thinking then
-        anthropic_req.thinking = { type = "enabled", budget_tokens = 4096 }
+    -- Flush any trailing tool results
+    if #pending_tool > 0 then
+        append_user(msgs, pending_tool)
     end
 
+    -- ── Build Anthropic request ──────────────────────────────
+    local anthropic_req = {
+        model      = req.model or "claude-sonnet-4-20250514",
+        max_tokens = req.max_tokens or 4096,
+        messages   = msgs,
+        stream     = req.stream or false,
+    }
+
+    -- ── tools ────────────────────────────────────────────────
+    local has_tools = false
+    if req.tools and type(req.tools) == "table" then
+        local tools = {}
+        for _, t in ipairs(req.tools) do
+            if type(t) == "table" and t.type == "function"
+                and type(t["function"]) == "table" then
+                local fn = t["function"]
+                table.insert(tools, {
+                    name         = fn.name or "",
+                    description  = fn.description or "",
+                    input_schema = fn.parameters or { type = "object", properties = {} },
+                })
+            end
+        end
+        if #tools > 0 then
+            anthropic_req.tools = tools
+            has_tools = true
+        end
+    end
+
+    -- ── tool_choice ──────────────────────────────────────────
+    -- OpenAI → Anthropic mapping:
+    --   "auto"     → {type:"auto"}
+    --   "none"     → remove tools entirely (Anthropic has no "none")
+    --   "required" → {type:"any"}
+    --   {type:"function", function:{name:"X"}} → {type:"tool", name:"X"}
+    if has_tools and req.tool_choice then
+        local tc = req.tool_choice
+        if type(tc) == "string" then
+            if tc == "auto" then
+                anthropic_req.tool_choice = { type = "auto" }
+            elseif tc == "none" then
+                anthropic_req.tools = nil
+                anthropic_req.tool_choice = nil
+            elseif tc == "required" then
+                anthropic_req.tool_choice = { type = "any" }
+            end
+        elseif type(tc) == "table" and tc.type == "function" then
+            local fn = tc["function"] or {}
+            anthropic_req.tool_choice = { type = "tool", name = fn.name or "" }
+        end
+    end
+
+    -- ── thinking (opt-in via extra_body) ──────────────────────
+    -- Default: OFF. Client sends extra_body.thinking to enable.
+    -- Example: {"extra_body": {"thinking": {"type": "enabled", "budget_tokens": 4096}}}
+    if type(req.extra_body) == "table" and type(req.extra_body.thinking) == "table" then
+        anthropic_req.thinking = req.extra_body.thinking
+    end
+
+    -- ── system ───────────────────────────────────────────────
     if system ~= "" then
-        anthropic_req.system = system
+        -- Strip trailing newline from accumulation
+        anthropic_req.system = string.match(system, "^(.-)\n*$")
     end
 
     return json.encode(anthropic_req)
 end
 
+-- ============================================================
+-- transform_response  (Anthropic → OpenAI, non-streaming)
+-- ============================================================
+
 function adapter.transform_response(raw_body)
     local ok, resp = pcall(json.decode, raw_body)
-    if not ok then return raw_body end
+    if not ok or resp == nil then return raw_body end
 
     local unified = {
         content = "",
         finish_reason = "",
-        token_usage = { prompt = 0, completion = 0, total = 0 }
+        token_usage = { prompt = 0, completion = 0, total = 0 },
     }
 
-    if resp.usage then
-        unified.token_usage.prompt = resp.usage.input_tokens or 0
-        unified.token_usage.completion = resp.usage.output_tokens or 0
-        unified.token_usage.total = (resp.usage.input_tokens or 0) + (resp.usage.output_tokens or 0)
+    local mapped = map_usage(resp.usage)
+    if mapped ~= nil then
+        unified.token_usage = mapped
     end
 
     if resp.content and #resp.content > 0 then
+        local tcs = {}
         for _, block in ipairs(resp.content) do
             if block.type == "text" then
                 unified.content = unified.content .. (block.text or "")
+            elseif block.type == "thinking" and block.thinking then
+                unified.reasoning_content = (unified.reasoning_content or "") .. block.thinking
+            elseif block.type == "tool_use" then
+                table.insert(tcs, {
+                    id        = safe_tool_id(block.id),
+                    type      = "function",
+                    name      = block.name or "",
+                    arguments = block.input or {},
+                })
             end
         end
+        if #tcs > 0 then unified.tool_calls = tcs end
     end
+
     unified.finish_reason = resp.stop_reason or ""
+    if unified.finish_reason == "tool_use" then
+        unified.finish_reason = "tool_calls"
+    elseif unified.finish_reason == "max_tokens" then
+        unified.finish_reason = "length"
+    elseif unified.finish_reason == "end_turn"
+        or unified.finish_reason == "stop_sequence" then
+        unified.finish_reason = "stop"
+    end
 
     return json.encode(unified)
 end
 
+-- ============================================================
+-- transform_stream_chunk  (Anthropic SSE → OpenAI SSE delta)
+-- ============================================================
+
 function adapter.transform_stream_chunk(raw_chunk)
     local ok, chunk = pcall(json.decode, raw_chunk)
     if not ok then return "" end
-    if chunk.type == "message_start" then return "" end
+
+    -- ── message_start: initial usage ─────────────────────────
+    if chunk.type == "message_start" then
+        if chunk.message then
+            -- Some Anthropic-compatible upstreams report a placeholder here and
+            -- only send the true prompt count in message_delta (justwoker:
+            -- 160 at message_start vs 6931 at message_delta). The gateway's
+            -- mergeUsage lets a later non-zero value win, so both are emitted.
+            local uses = map_usage(chunk.message.usage)
+            if uses ~= nil then
+                return json.encode({ usage = uses, done = false })
+            end
+        end
+        return ""
+    end
+
+    -- ── message_delta: stop_reason + final usage ─────────────
     if chunk.type == "message_delta" then
-        return json.encode({ content = "", done = (chunk.delta and chunk.delta.stop_reason ~= nil) })
+        local finish = nil
+        if chunk.delta and chunk.delta.stop_reason ~= nil then
+            local sr = chunk.delta.stop_reason
+            if sr == "max_tokens" then
+                finish = "length"
+            elseif sr == "tool_use" then
+                finish = "tool_calls"
+            else
+                finish = "stop"
+            end
+        end
+        -- message_delta carries the FINAL usage, so it must map the cache
+        -- fields too; the old code dropped them here, which silently lost the
+        -- whole cache split on every streamed response.
+        local uses = map_usage(chunk.usage)
+        if uses ~= nil then
+            return json.encode({ content = "", done = (finish ~= nil), finish_reason = finish, usage = uses })
+        end
+        return json.encode({ content = "", done = (finish ~= nil), finish_reason = finish })
     end
-    if chunk.type == "content_block_start" and chunk.content_block
-        and chunk.content_block.type == "tool_use" then
-        -- first fragment of a tool call: emit index + id + name, empty args
-        return json.encode({
-            content = "", done = false,
-            tool_calls = { {
-                index = chunk.index or 0,
-                id = chunk.content_block.id or "",
-                type = "function",
-                ["function"] = { name = chunk.content_block.name or "", arguments = "" }
-            } }
-        })
+
+    -- ── content_block_start: begin text / thinking / tool_use ─
+    if chunk.type == "content_block_start" and chunk.content_block then
+        local cb = chunk.content_block
+        if cb.type == "tool_use" then
+            -- Pass Anthropic content_block index through; Go-side rewrites
+            -- to sequential OpenAI tool_call ordinal for multi-tool streams.
+            return json.encode({
+                content = "", done = false,
+                tool_calls = { {
+                    index = chunk.index or 0,
+                    id    = safe_tool_id(cb.id),
+                    type  = "function",
+                    ["function"] = { name = cb.name or "", arguments = "" },
+                } },
+            })
+        end
+        return ""  -- text / thinking block start: no OpenAI equivalent
     end
+
+    -- ── content_block_delta: incremental content ──────────────
     if chunk.type == "content_block_delta" and chunk.delta then
-        if chunk.delta.type == "input_json_delta" then
-            -- incremental JSON fragment; clients accumulate across chunks
-            local unified = { content = "", done = false, tool_calls = { {
-                index = chunk.index or 0,
-                id = "",
-                type = "function",
-                ["function"] = { name = "", arguments = chunk.delta.partial_json or "" }
-            } } }
-            return json.encode(unified)
+        local d = chunk.delta
+        if d.type == "input_json_delta" then
+            -- Tool call argument fragment; Go-side rewrites index.
+            return json.encode({
+                content = "", done = false,
+                tool_calls = { {
+                    index = chunk.index or 0,
+                    id    = "",
+                    type  = "function",
+                    ["function"] = { name = "", arguments = d.partial_json or "" },
+                } },
+            })
         end
-        if chunk.delta.type == "thinking_delta" and chunk.delta.thinking then
-            return json.encode({ content = "", done = false, reasoning_content = chunk.delta.thinking })
+        if d.type == "thinking_delta" and d.thinking then
+            return json.encode({ content = "", done = false, reasoning_content = d.thinking })
+        end
+        if d.type == "text_delta" and d.text then
+            return json.encode({ content = d.text, done = false })
         end
-        return json.encode({ content = chunk.delta.text or "", done = false })
     end
+
+    -- ── content_block_stop / message_stop ─────────────────────
     if chunk.type == "message_stop" then
-        return json.encode({ content = "", done = true })
-    end
-    if chunk.type == "content_block_stop" then
-        return json.encode({ content = "", done = false })
+        -- The message_delta event already emitted the true finish_reason.
+        -- Do NOT emit done=true here: an empty finish_reason would
+        -- overwrite the real one (tool_calls) in the Go gateway's
+        -- lastFinish tracker, causing the final SSE chunk to say
+        -- finish_reason=stop instead of tool_calls.
+        return ""
     end
+
     return ""
 end
 
+-- ============================================================
+-- transform_error  (Anthropic error → human-readable string)
+-- ============================================================
+
+function adapter.transform_error(status, body)
+    local ok, resp = pcall(json.decode, body)
+    if not ok or type(resp) ~= "table" then return nil end
+
+    -- Anthropic envelope: {type:"error", error:{type, message}}
+    if resp.type == "error" and type(resp.error) == "table"
+        and type(resp.error.message) == "string" then
+        return resp.error.message
+    end
+
+    -- Flat envelope: {error: {message: "..."}}
+    if type(resp.error) == "table" and type(resp.error.message) == "string" then
+        return resp.error.message
+    end
+
+    return nil
+end
+
 return adapter

internal/lua/adapters/deepseek.lua

@@ -49,6 +49,11 @@ function adapter.transform_response(raw_body)
         unified.token_usage.prompt = resp.usage.prompt_tokens or 0
         unified.token_usage.completion = resp.usage.completion_tokens or 0
         unified.token_usage.total = resp.usage.total_tokens or 0
+        unified.token_usage.prompt_cache_hit_tokens = resp.usage.prompt_cache_hit_tokens or 0
+        unified.token_usage.prompt_cache_miss_tokens = resp.usage.prompt_cache_miss_tokens or 0
+        if resp.usage.prompt_tokens_details and resp.usage.prompt_tokens_details.cached_tokens ~= nil then
+            unified.token_usage.prompt_tokens_details = { cached_tokens = resp.usage.prompt_tokens_details.cached_tokens }
+        end
     end
 
     if type(resp.choices) == "table" and #resp.choices > 0 then
@@ -82,14 +87,47 @@ end
 function adapter.transform_stream_chunk(raw_chunk)
     local ok, chunk = pcall(json.decode, raw_chunk)
     if not ok then return "" end
-    if not chunk.choices or #chunk.choices == 0 then return "" end
+
+    -- OpenAI-style streams may attach usage to a chunk with empty choices
+    -- (the final usage chunk). Keys must match Go's TokenUsage json tags
+    -- (prompt/completion/total); the gateway re-emits standard *_tokens.
+    local uses = nil
+    if type(chunk.usage) == "table" then
+        uses = {
+            prompt = chunk.usage.prompt_tokens or chunk.usage.prompt or 0,
+            completion = chunk.usage.completion_tokens or chunk.usage.completion or 0,
+            total = chunk.usage.total_tokens or chunk.usage.total or 0,
+            prompt_cache_hit_tokens = chunk.usage.prompt_cache_hit_tokens or 0,
+            prompt_cache_miss_tokens = chunk.usage.prompt_cache_miss_tokens or 0,
+        }
+        if chunk.usage.prompt_tokens_details and chunk.usage.prompt_tokens_details.cached_tokens ~= nil then
+            uses.prompt_tokens_details = { cached_tokens = chunk.usage.prompt_tokens_details.cached_tokens }
+        end
+    end
+
+    if not chunk.choices or #chunk.choices == 0 then
+        if uses ~= nil then
+            -- usage-only chunk is not a content/finish signal; the gateway
+            -- emits its own terminal stop chunk and merges this usage.
+            return json.encode({ usage = uses, done = false })
+        end
+        return ""
+    end
     local delta = chunk.choices[1].delta or {}
     local fr = chunk.choices[1].finish_reason
 
+    local finish = (type(fr) == "string" and fr ~= "") and fr or nil
+
     local unified = {
         content = delta.content or "",
-        done = (fr ~= nil)
+        done = (finish ~= nil)
     }
+    if finish then
+        unified.finish_reason = finish
+    end
+    if uses ~= nil then
+        unified.usage = uses
+    end
     if delta.reasoning_content then
         unified.reasoning_content = delta.reasoning_content
     end
@@ -99,4 +137,15 @@ function adapter.transform_stream_chunk(raw_chunk)
     return json.encode(unified)
 end
 
+-- 错误收敛：DeepSeek 走标准 OpenAI 信封 {error:{message,...}}
+function adapter.transform_error(status, body)
+    local ok, resp = pcall(json.decode, body)
+    if not ok or type(resp) ~= "table" then return nil end
+    local e = resp.error
+    if type(e) == "table" and type(e.message) == "string" then
+        return e.message
+    end
+    return nil
+end
+
 return adapter

internal/lua/adapters/gemini.lua

@@ -7,6 +7,10 @@ adapter.headers = {}
 
 -- Gemini API: POST /v1/models/{model}:generateContent
 -- Auth: API key in query param ?key=XXX or Authorization: Bearer XXX
+--
+-- 请求方向必须把 OpenAI 的工具调用翻译成 Gemini 的 functionCall /
+-- functionResponse，否则 agent 回放的历史里助手那一轮的调用会凭空消失，
+-- 紧随其后的工具结果就成了「无来源」的孤立结果，模型只能反复重发同一个调用。
 function adapter.transform_request(raw_body)
     local ok, req = pcall(json.decode, raw_body)
     if not ok then return raw_body end
@@ -14,16 +18,23 @@ function adapter.transform_request(raw_body)
     -- 将 OpenAI 风格 content（字符串或 [{type:*}] 数组）拆成 Gemini parts
     local function to_parts(content)
         if type(content) == "string" then
+            if content == "" then return {} end
             return { { text = content } }
         end
         local parts = {}
         for _, p in ipairs(content or {}) do
-            if p.type == "text" then
-                table.insert(parts, { text = p.text })
-            elseif p.type == "image_url" and type(p.image_url) == "table" and p.image_url.url then
-                local mt, b64 = string.match(p.image_url.url, "^data:([^,]+);base64,(.+)$")
-                if b64 then
-                    table.insert(parts, { inline_data = { mime_type = mt or "image/png", data = b64 } })
+            if type(p) == "string" then
+                if p ~= "" then table.insert(parts, { text = p }) end
+            elseif type(p) == "table" then
+                if p.type == "text" then
+                    if p.text ~= nil and p.text ~= "" then
+                        table.insert(parts, { text = p.text })
+                    end
+                elseif p.type == "image_url" and type(p.image_url) == "table" and p.image_url.url then
+                    local mt, b64 = string.match(p.image_url.url, "^data:([^,]+);base64,(.+)$")
+                    if b64 then
+                        table.insert(parts, { inline_data = { mime_type = mt or "image/png", data = b64 } })
+                    end
                 end
             end
         end
@@ -31,11 +42,59 @@ function adapter.transform_request(raw_body)
     end
 
     local contents = {}
+    local system = ""
+    local call_names = {}   -- tool_call_id -> 函数名（functionResponse 只认名字）
+
     for _, m in ipairs(req.messages or {}) do
-        table.insert(contents, {
-            role = (m.role == "assistant") and "model" or m.role,
-            parts = to_parts(m.content)
-        })
+        local role = m.role or "user"
+
+        if role == "system" then
+            if type(m.content) == "string" then
+                system = system .. m.content .. "\n"
+            end
+
+        elseif role == "assistant" then
+            local parts = to_parts(m.content)
+            if type(m.tool_calls) == "table" then
+                for _, tc in ipairs(m.tool_calls) do
+                    if type(tc) == "table" then
+                        local fn = tc["function"] or {}
+                        local args = fn.arguments
+                        if type(args) == "string" and args ~= "" then
+                            local aok, decoded = pcall(json.decode, args)
+                            args = aok and decoded or {}
+                        elseif type(args) ~= "table" then
+                            args = {}
+                        end
+                        if tc.id ~= nil then call_names[tc.id] = fn.name or "" end
+                        table.insert(parts, {
+                            functionCall = { name = fn.name or "", args = args }
+                        })
+                    end
+                end
+            end
+            -- Gemini 不接受空的 parts 数组；无可发送内容的轮次直接跳过
+            if #parts > 0 then
+                table.insert(contents, { role = "model", parts = parts })
+            end
+
+        elseif role == "tool" then
+            local name = call_names[m.tool_call_id] or m.name or ""
+            local text = type(m.content) == "string" and m.content or ""
+            table.insert(contents, {
+                role = "user",
+                parts = { { functionResponse = {
+                    name = name,
+                    response = { content = text },
+                } } }
+            })
+
+        else
+            local parts = to_parts(m.content)
+            if #parts > 0 then
+                table.insert(contents, { role = "user", parts = parts })
+            end
+        end
     end
 
     local gemini_req = {
@@ -46,6 +105,30 @@ function adapter.transform_request(raw_body)
         }
     }
 
+    if system ~= "" then
+        gemini_req.systemInstruction = { parts = { { text = system } } }
+    end
+
+    -- tools -> functionDeclarations
+    if type(req.tools) == "table" then
+        local decls = {}
+        for _, t in ipairs(req.tools) do
+            if type(t) == "table" and type(t["function"]) == "table" then
+                local fn = t["function"]
+                local params = fn.parameters
+                if type(params) ~= "table" then params = { type = "object", properties = {} } end
+                table.insert(decls, {
+                    name = fn.name or "",
+                    description = fn.description or "",
+                    parameters = params,
+                })
+            end
+        end
+        if #decls > 0 then
+            gemini_req.tools = { { functionDeclarations = decls } }
+        end
+    end
+
     if req.stream then
         gemini_req.stream = true
     end
@@ -68,20 +151,55 @@ function adapter.transform_response(raw_body)
         unified.token_usage.prompt = resp.usageMetadata.promptTokenCount or 0
         unified.token_usage.completion = resp.usageMetadata.candidatesTokenCount or 0
         unified.token_usage.total = resp.usageMetadata.totalTokenCount or 0
+        -- Gemini reports context-cache reads as cachedContentTokenCount;
+        -- normalize into OpenAI-standard prompt_tokens_details.cached_tokens
+        -- so clients and the audit trail see the hit count. Emitted even when
+        -- 0 so a reported miss stays distinguishable from "not reported".
+        if resp.usageMetadata.cachedContentTokenCount ~= nil then
+            unified.token_usage.prompt_tokens_details = {
+                cached_tokens = resp.usageMetadata.cachedContentTokenCount
+            }
+        end
     end
 
     if resp.candidates and #resp.candidates > 0 then
         local cand = resp.candidates[1]
+        local tools = {}
         if cand.content and cand.content.parts then
             for _, part in ipairs(cand.content.parts) do
                 if part.text then
                     unified.content = unified.content .. part.text
+                elseif part.functionCall then
+                    -- Non-streaming tool calls used to be dropped here while
+                    -- transform_stream_chunk handled them, so a non-streaming
+                    -- agent turn looked like a plain text answer and the tool
+                    -- loop died. Gemini's args are already an object.
+                    local args = part.functionCall.args
+                    if type(args) == "string" then
+                        local aok, decoded = pcall(json.decode, args)
+                        args = aok and decoded or {}
+                    elseif type(args) ~= "table" then
+                        args = {}
+                    end
+                    table.insert(tools, {
+                        id = part.functionCall.id or ("call_" .. #tools),
+                        type = "function",
+                        name = part.functionCall.name or "",
+                        arguments = args
+                    })
                 end
             end
         end
         if cand.finishReason then
             unified.finish_reason = cand.finishReason
         end
+        if #tools > 0 then
+            unified.tool_calls = tools
+            -- Gemini reports finishReason "STOP" even when it emitted a
+            -- functionCall; clients keyed on finish_reason would treat that as
+            -- a completed answer and never run the tool.
+            unified.finish_reason = "tool_calls"
+        end
     end
 
     return json.encode(unified)
@@ -91,9 +209,45 @@ function adapter.transform_stream_chunk(raw_chunk)
     local ok, chunk = pcall(json.decode, raw_chunk)
     if not ok then return "" end
 
-    if not chunk.candidates or #chunk.candidates == 0 then return "" end
+    -- Gemini attaches usageMetadata to the final chunk (alongside or after
+    -- candidates). Keys map to Go's TokenUsage json tags (prompt/...).
+    local uses = nil
+    if type(chunk.usageMetadata) == "table" then
+        local p = chunk.usageMetadata.promptTokenCount or 0
+        local c = chunk.usageMetadata.candidatesTokenCount or 0
+        local t = chunk.usageMetadata.totalTokenCount or 0
+        if p > 0 or c > 0 or t > 0 then
+            uses = { prompt = p, completion = c, total = t }
+            -- Emit details whenever the field is present, even at 0, so a
+            -- reported cache miss stays distinguishable from "not reported".
+            if chunk.usageMetadata.cachedContentTokenCount ~= nil then
+                uses.prompt_tokens_details = {
+                    cached_tokens = chunk.usageMetadata.cachedContentTokenCount
+                }
+            end
+        end
+    end
+
+    if not chunk.candidates or #chunk.candidates == 0 then
+        if uses ~= nil then
+            return json.encode({ usage = uses, done = false })
+        end
+        return ""
+    end
     local cand = chunk.candidates[1]
-    local unified = { content = "", done = (cand.finishReason ~= nil) }
+    -- Gemini finishReason -> OpenAI finish_reason
+    local finish = nil
+    if cand.finishReason ~= nil then
+        if cand.finishReason == "MAX_TOKENS" then
+            finish = "length"
+        elseif cand.finishReason == "SAFETY" or cand.finishReason == "RECITATION"
+            or cand.finishReason == "BLOCKLIST" then
+            finish = "content_filter"
+        else
+            finish = "stop"
+        end
+    end
+    local unified = { content = "", done = (finish ~= nil), finish_reason = finish }
     local reasoning = ""
     local tools = {}
     if cand.content and cand.content.parts then
@@ -117,7 +271,26 @@ function adapter.transform_stream_chunk(raw_chunk)
     end
     if reasoning ~= "" then unified.reasoning_content = reasoning end
     if #tools > 0 then unified.tool_calls = tools end
+    if uses ~= nil then
+        unified.usage = uses
+    end
     return json.encode(unified)
 end
 
+-- 错误收敛：Gemini REST 信封 {error:{code, message, status}}
+function adapter.transform_error(status, body)
+    local ok, resp = pcall(json.decode, body)
+    if not ok or type(resp) ~= "table" then return nil end
+    local e = resp.error
+    if type(e) == "table" then
+        if type(e.message) == "string" then
+            if type(e.status) == "string" then
+                return e.status .. ": " .. e.message
+            end
+            return e.message
+        end
+    end
+    return nil
+end
+
 return adapter

internal/lua/adapters/github.lua

@@ -33,6 +33,18 @@ function adapter.transform_response(raw_body)
         unified.token_usage.prompt = resp.usage.prompt_tokens or 0
         unified.token_usage.completion = resp.usage.completion_tokens or 0
         unified.token_usage.total = resp.usage.total_tokens or 0
+        local hit = 0
+        if type(resp.usage.prompt_tokens_details) == "table" and resp.usage.prompt_tokens_details.cached_tokens ~= nil then
+            hit = resp.usage.prompt_tokens_details.cached_tokens
+            unified.token_usage.prompt_tokens_details = { cached_tokens = hit }
+        end
+        if (resp.usage.prompt_cache_hit_tokens or 0) > 0 then
+            unified.token_usage.prompt_cache_hit_tokens = resp.usage.prompt_cache_hit_tokens
+            unified.token_usage.prompt_cache_miss_tokens = resp.usage.prompt_cache_miss_tokens or 0
+            if hit == 0 then
+                unified.token_usage.prompt_tokens_details = { cached_tokens = resp.usage.prompt_cache_hit_tokens }
+            end
+        end
     end
 
     if type(resp.choices) == "table" and #resp.choices > 0 then
@@ -63,14 +75,49 @@ end
 function adapter.transform_stream_chunk(raw_chunk)
     local ok, chunk = pcall(json.decode, raw_chunk)
     if not ok then return "" end
-    if not chunk.choices or #chunk.choices == 0 then return "" end
+
+    -- OpenAI-style streams may attach usage to a chunk with empty choices
+    -- (the final usage chunk). Keys must match Go's TokenUsage json tags
+    -- (prompt/completion/total); the gateway re-emits standard *_tokens.
+    local uses = nil
+    if type(chunk.usage) == "table" then
+        uses = {
+            prompt = chunk.usage.prompt_tokens or chunk.usage.prompt or 0,
+            completion = chunk.usage.completion_tokens or chunk.usage.completion or 0,
+            total = chunk.usage.total_tokens or chunk.usage.total or 0,
+        }
+        if type(chunk.usage.prompt_tokens_details) == "table" and chunk.usage.prompt_tokens_details.cached_tokens ~= nil then
+            uses.prompt_tokens_details = { cached_tokens = chunk.usage.prompt_tokens_details.cached_tokens }
+        elseif (chunk.usage.prompt_cache_hit_tokens or 0) > 0 then
+            uses.prompt_cache_hit_tokens = chunk.usage.prompt_cache_hit_tokens
+            uses.prompt_cache_miss_tokens = chunk.usage.prompt_cache_miss_tokens or 0
+            uses.prompt_tokens_details = { cached_tokens = chunk.usage.prompt_cache_hit_tokens }
+        end
+    end
+
+    if not chunk.choices or #chunk.choices == 0 then
+        if uses ~= nil then
+            -- usage-only chunk is not a content/finish signal; the gateway
+            -- emits its own terminal stop chunk and merges this usage.
+            return json.encode({ usage = uses, done = false })
+        end
+        return ""
+    end
     local delta = chunk.choices[1].delta or {}
     local fr = chunk.choices[1].finish_reason
 
+    local finish = (type(fr) == "string" and fr ~= "") and fr or nil
+
     local unified = {
         content = delta.content or "",
-        done = (fr ~= nil)
+        done = (finish ~= nil)
     }
+    if finish then
+        unified.finish_reason = finish
+    end
+    if uses ~= nil then
+        unified.usage = uses
+    end
     if delta.reasoning_content then
         unified.reasoning_content = delta.reasoning_content
     end
@@ -80,4 +127,15 @@ function adapter.transform_stream_chunk(raw_chunk)
     return json.encode(unified)
 end
 
+-- 错误收敛：GitHub Models 走标准 OpenAI 信封 {error:{message,...}}
+function adapter.transform_error(status, body)
+    local ok, resp = pcall(json.decode, body)
+    if not ok or type(resp) ~= "table" then return nil end
+    local e = resp.error
+    if type(e) == "table" and type(e.message) == "string" then
+        return e.message
+    end
+    return nil
+end
+
 return adapter

internal/lua/adapters/groq.lua

@@ -32,6 +32,18 @@ function adapter.transform_response(raw_body)
         unified.token_usage.prompt = resp.usage.prompt_tokens or 0
         unified.token_usage.completion = resp.usage.completion_tokens or 0
         unified.token_usage.total = resp.usage.total_tokens or 0
+        local hit = 0
+        if type(resp.usage.prompt_tokens_details) == "table" and resp.usage.prompt_tokens_details.cached_tokens ~= nil then
+            hit = resp.usage.prompt_tokens_details.cached_tokens
+            unified.token_usage.prompt_tokens_details = { cached_tokens = hit }
+        end
+        if (resp.usage.prompt_cache_hit_tokens or 0) > 0 then
+            unified.token_usage.prompt_cache_hit_tokens = resp.usage.prompt_cache_hit_tokens
+            unified.token_usage.prompt_cache_miss_tokens = resp.usage.prompt_cache_miss_tokens or 0
+            if hit == 0 then
+                unified.token_usage.prompt_tokens_details = { cached_tokens = resp.usage.prompt_cache_hit_tokens }
+            end
+        end
     end
 
     if type(resp.choices) == "table" and #resp.choices > 0 then
@@ -62,14 +74,49 @@ end
 function adapter.transform_stream_chunk(raw_chunk)
     local ok, chunk = pcall(json.decode, raw_chunk)
     if not ok then return "" end
-    if not chunk.choices or #chunk.choices == 0 then return "" end
+
+    -- OpenAI-style streams may attach usage to a chunk with empty choices
+    -- (the final usage chunk). Keys must match Go's TokenUsage json tags
+    -- (prompt/completion/total); the gateway re-emits standard *_tokens.
+    local uses = nil
+    if type(chunk.usage) == "table" then
+        uses = {
+            prompt = chunk.usage.prompt_tokens or chunk.usage.prompt or 0,
+            completion = chunk.usage.completion_tokens or chunk.usage.completion or 0,
+            total = chunk.usage.total_tokens or chunk.usage.total or 0,
+        }
+        if type(chunk.usage.prompt_tokens_details) == "table" and chunk.usage.prompt_tokens_details.cached_tokens ~= nil then
+            uses.prompt_tokens_details = { cached_tokens = chunk.usage.prompt_tokens_details.cached_tokens }
+        elseif (chunk.usage.prompt_cache_hit_tokens or 0) > 0 then
+            uses.prompt_cache_hit_tokens = chunk.usage.prompt_cache_hit_tokens
+            uses.prompt_cache_miss_tokens = chunk.usage.prompt_cache_miss_tokens or 0
+            uses.prompt_tokens_details = { cached_tokens = chunk.usage.prompt_cache_hit_tokens }
+        end
+    end
+
+    if not chunk.choices or #chunk.choices == 0 then
+        if uses ~= nil then
+            -- usage-only chunk is not a content/finish signal; the gateway
+            -- emits its own terminal stop chunk and merges this usage.
+            return json.encode({ usage = uses, done = false })
+        end
+        return ""
+    end
     local delta = chunk.choices[1].delta or {}
     local fr = chunk.choices[1].finish_reason
 
+    local finish = (type(fr) == "string" and fr ~= "") and fr or nil
+
     local unified = {
         content = delta.content or "",
-        done = (fr ~= nil)
+        done = (finish ~= nil)
     }
+    if finish then
+        unified.finish_reason = finish
+    end
+    if uses ~= nil then
+        unified.usage = uses
+    end
     if delta.reasoning_content then
         unified.reasoning_content = delta.reasoning_content
     end
@@ -79,4 +126,15 @@ function adapter.transform_stream_chunk(raw_chunk)
     return json.encode(unified)
 end
 
+-- 错误收敛：Groq 走标准 OpenAI 信封 {error:{message,...}}
+function adapter.transform_error(status, body)
+    local ok, resp = pcall(json.decode, body)
+    if not ok or type(resp) ~= "table" then return nil end
+    local e = resp.error
+    if type(e) == "table" and type(e.message) == "string" then
+        return e.message
+    end
+    return nil
+end
+
 return adapter

internal/lua/adapters/kimicode.lua

@@ -64,6 +64,18 @@ function adapter.transform_response(raw_body)
         unified.token_usage.prompt = resp.usage.prompt_tokens or 0
         unified.token_usage.completion = resp.usage.completion_tokens or 0
         unified.token_usage.total = resp.usage.total_tokens or 0
+        local hit = 0
+        if type(resp.usage.prompt_tokens_details) == "table" and resp.usage.prompt_tokens_details.cached_tokens ~= nil then
+            hit = resp.usage.prompt_tokens_details.cached_tokens
+            unified.token_usage.prompt_tokens_details = { cached_tokens = hit }
+        end
+        if (resp.usage.prompt_cache_hit_tokens or 0) > 0 then
+            unified.token_usage.prompt_cache_hit_tokens = resp.usage.prompt_cache_hit_tokens
+            unified.token_usage.prompt_cache_miss_tokens = resp.usage.prompt_cache_miss_tokens or 0
+            if hit == 0 then
+                unified.token_usage.prompt_tokens_details = { cached_tokens = resp.usage.prompt_cache_hit_tokens }
+            end
+        end
     end
     if type(resp.choices) == "table" and #resp.choices > 0 then
         local ch = resp.choices[1]
@@ -95,14 +107,49 @@ end
 function adapter.transform_stream_chunk(raw_chunk)
     local ok, chunk = pcall(json.decode, raw_chunk)
     if not ok then return "" end
-    if not chunk.choices or #chunk.choices == 0 then return "" end
+
+    -- OpenAI-style streams may attach usage to a chunk with empty choices
+    -- (the final usage chunk). Keys must match Go's TokenUsage json tags
+    -- (prompt/completion/total); the gateway re-emits standard *_tokens.
+    local uses = nil
+    if type(chunk.usage) == "table" then
+        uses = {
+            prompt = chunk.usage.prompt_tokens or chunk.usage.prompt or 0,
+            completion = chunk.usage.completion_tokens or chunk.usage.completion or 0,
+            total = chunk.usage.total_tokens or chunk.usage.total or 0,
+        }
+        if type(chunk.usage.prompt_tokens_details) == "table" and chunk.usage.prompt_tokens_details.cached_tokens ~= nil then
+            uses.prompt_tokens_details = { cached_tokens = chunk.usage.prompt_tokens_details.cached_tokens }
+        elseif (chunk.usage.prompt_cache_hit_tokens or 0) > 0 then
+            uses.prompt_cache_hit_tokens = chunk.usage.prompt_cache_hit_tokens
+            uses.prompt_cache_miss_tokens = chunk.usage.prompt_cache_miss_tokens or 0
+            uses.prompt_tokens_details = { cached_tokens = chunk.usage.prompt_cache_hit_tokens }
+        end
+    end
+
+    if not chunk.choices or #chunk.choices == 0 then
+        if uses ~= nil then
+            -- usage-only chunk is not a content/finish signal; the gateway
+            -- emits its own terminal stop chunk and merges this usage.
+            return json.encode({ usage = uses, done = false })
+        end
+        return ""
+    end
     local delta = chunk.choices[1].delta or {}
     local fr = chunk.choices[1].finish_reason
 
+    local finish = (type(fr) == "string" and fr ~= "") and fr or nil
+
     local unified = {
         content = delta.content or "",
-        done = (fr ~= nil)
+        done = (finish ~= nil)
     }
+    if finish then
+        unified.finish_reason = finish
+    end
+    if uses ~= nil then
+        unified.usage = uses
+    end
     if delta.reasoning_content then
         unified.reasoning_content = delta.reasoning_content
     end
@@ -112,4 +159,15 @@ function adapter.transform_stream_chunk(raw_chunk)
     return json.encode(unified)
 end
 
-return adapter
+-- 错误收敛：kimi 网关为 OpenAI 风格 {error:{message,...}}
+function adapter.transform_error(status, body)
+    local ok, resp = pcall(json.decode, body)
+    if not ok or type(resp) ~= "table" then return nil end
+    local e = resp.error
+    if type(e) == "table" and type(e.message) == "string" then
+        return e.message
+    end
+    return nil
+end
+
+return adapter

internal/lua/adapters/mistral.lua

@@ -32,6 +32,18 @@ function adapter.transform_response(raw_body)
         unified.token_usage.prompt = resp.usage.prompt_tokens or 0
         unified.token_usage.completion = resp.usage.completion_tokens or 0
         unified.token_usage.total = resp.usage.total_tokens or 0
+        local hit = 0
+        if type(resp.usage.prompt_tokens_details) == "table" and resp.usage.prompt_tokens_details.cached_tokens ~= nil then
+            hit = resp.usage.prompt_tokens_details.cached_tokens
+            unified.token_usage.prompt_tokens_details = { cached_tokens = hit }
+        end
+        if (resp.usage.prompt_cache_hit_tokens or 0) > 0 then
+            unified.token_usage.prompt_cache_hit_tokens = resp.usage.prompt_cache_hit_tokens
+            unified.token_usage.prompt_cache_miss_tokens = resp.usage.prompt_cache_miss_tokens or 0
+            if hit == 0 then
+                unified.token_usage.prompt_tokens_details = { cached_tokens = resp.usage.prompt_cache_hit_tokens }
+            end
+        end
     end
 
     if type(resp.choices) == "table" and #resp.choices > 0 then
@@ -62,14 +74,49 @@ end
 function adapter.transform_stream_chunk(raw_chunk)
     local ok, chunk = pcall(json.decode, raw_chunk)
     if not ok then return "" end
-    if not chunk.choices or #chunk.choices == 0 then return "" end
+
+    -- OpenAI-style streams may attach usage to a chunk with empty choices
+    -- (the final usage chunk). Keys must match Go's TokenUsage json tags
+    -- (prompt/completion/total); the gateway re-emits standard *_tokens.
+    local uses = nil
+    if type(chunk.usage) == "table" then
+        uses = {
+            prompt = chunk.usage.prompt_tokens or chunk.usage.prompt or 0,
+            completion = chunk.usage.completion_tokens or chunk.usage.completion or 0,
+            total = chunk.usage.total_tokens or chunk.usage.total or 0,
+        }
+        if type(chunk.usage.prompt_tokens_details) == "table" and chunk.usage.prompt_tokens_details.cached_tokens ~= nil then
+            uses.prompt_tokens_details = { cached_tokens = chunk.usage.prompt_tokens_details.cached_tokens }
+        elseif (chunk.usage.prompt_cache_hit_tokens or 0) > 0 then
+            uses.prompt_cache_hit_tokens = chunk.usage.prompt_cache_hit_tokens
+            uses.prompt_cache_miss_tokens = chunk.usage.prompt_cache_miss_tokens or 0
+            uses.prompt_tokens_details = { cached_tokens = chunk.usage.prompt_cache_hit_tokens }
+        end
+    end
+
+    if not chunk.choices or #chunk.choices == 0 then
+        if uses ~= nil then
+            -- usage-only chunk is not a content/finish signal; the gateway
+            -- emits its own terminal stop chunk and merges this usage.
+            return json.encode({ usage = uses, done = false })
+        end
+        return ""
+    end
     local delta = chunk.choices[1].delta or {}
     local fr = chunk.choices[1].finish_reason
 
+    local finish = (type(fr) == "string" and fr ~= "") and fr or nil
+
     local unified = {
         content = delta.content or "",
-        done = (fr ~= nil)
+        done = (finish ~= nil)
     }
+    if finish then
+        unified.finish_reason = finish
+    end
+    if uses ~= nil then
+        unified.usage = uses
+    end
     if delta.reasoning_content then
         unified.reasoning_content = delta.reasoning_content
     end
@@ -79,4 +126,14 @@ function adapter.transform_stream_chunk(raw_chunk)
     return json.encode(unified)
 end
 
+-- 错误收敛：Mistral 用顶层 {message="...", type="..."}（非嵌套 error）
+function adapter.transform_error(status, body)
+    local ok, resp = pcall(json.decode, body)
+    if not ok or type(resp) ~= "table" then return nil end
+    if type(resp.message) == "string" then
+        return resp.message
+    end
+    return nil
+end
+
 return adapter

internal/lua/adapters/ollama.lua

@@ -6,6 +6,10 @@ adapter.endpoint = "/api/chat"
 adapter.headers = {}
 
 -- Ollama API 格式：{ model, messages, stream, options:{temperature,num_predict} }
+-- 工具调用必须一并翻译：Ollama 的 assistant 消息用 tool_calls（arguments 是对象
+-- 而非 JSON 字符串），工具结果用 tool 角色 + tool_name。此前这里只复制了
+-- role/content，助手那轮的调用和 tool 消息的归属全部丢失，模型看到无来源的
+-- 工具结果就只能反复重发同一个调用。
 function adapter.transform_request(raw_body)
     local ok, req = pcall(json.decode, raw_body)
     if not ok then return raw_body end
@@ -22,6 +26,7 @@ function adapter.transform_request(raw_body)
     -- 转换 messages 格式（Ollama messages 支持 images base64 数组）
     if req.messages then
         local msgs = {}
+        local call_names = {}   -- tool_call_id -> 函数名
         for _, m in ipairs(req.messages) do
             local text, images
             if type(m.content) == "string" then
@@ -30,22 +35,61 @@ function adapter.transform_request(raw_body)
                 text = ""
                 images = {}
                 for _, p in ipairs(m.content or {}) do
-                    if p.type == "text" then
-                        text = text .. (p.text or "")
-                    elseif p.type == "image_url" and type(p.image_url) == "table" and p.image_url.url then
-                        local b64 = string.match(p.image_url.url, "^data:[^,]+;base64,(.+)$")
-                        if b64 then table.insert(images, b64) end
+                    if type(p) == "table" then
+                        if p.type == "text" then
+                            text = text .. (p.text or "")
+                        elseif p.type == "image_url" and type(p.image_url) == "table" and p.image_url.url then
+                            local b64 = string.match(p.image_url.url, "^data:[^,]+;base64,(.+)$")
+                            if b64 then table.insert(images, b64) end
+                        end
                     end
                 end
                 if #images == 0 then images = nil end
             end
             local msg = { role = m.role, content = text }
             if images then msg.images = images end
+
+            -- 助手轮的工具调用：arguments 转成对象
+            if type(m.tool_calls) == "table" and #m.tool_calls > 0 then
+                local tcs = {}
+                for _, tc in ipairs(m.tool_calls) do
+                    if type(tc) == "table" then
+                        local fn = tc["function"] or {}
+                        local args = fn.arguments
+                        if type(args) == "string" and args ~= "" then
+                            local aok, decoded = pcall(json.decode, args)
+                            args = aok and decoded or {}
+                        elseif type(args) ~= "table" then
+                            args = {}
+                        end
+                        if tc.id ~= nil then call_names[tc.id] = fn.name or "" end
+                        table.insert(tcs, {
+                            ["function"] = { name = fn.name or "", arguments = args }
+                        })
+                    end
+                end
+                if #tcs > 0 then msg.tool_calls = tcs end
+            end
+
+            -- 工具结果：Ollama 用 tool_name 标识归属（不认 tool_call_id）
+            if m.role == "tool" then
+                local name = call_names[m.tool_call_id] or m.name or ""
+                if m.tool_call_id ~= nil then
+                    msg.tool_call_id = m.tool_call_id
+                end
+                if name ~= "" then msg.tool_name = name end
+            end
+
             table.insert(msgs, msg)
         end
         ollama_req.messages = msgs
     end
 
+    -- tools 透传（Ollama 的 shape 与 OpenAI 一致）
+    if type(req.tools) == "table" and #req.tools > 0 then
+        ollama_req.tools = req.tools
+    end
+
     return json.encode(ollama_req)
 end
 
@@ -53,15 +97,44 @@ function adapter.transform_response(raw_body)
     local ok, resp = pcall(json.decode, raw_body)
     if not ok then return raw_body end
 
+    local p = resp.prompt_eval_count or 0
+    local c = resp.eval_count or 0
     local unified = {
         content = "",
         finish_reason = resp.done_reason or "",
-        tool_calls = {},
-        usage = { prompt = 0, completion = 0, total = 0 }
+        -- key must be token_usage to match Go's UnifiedResponse json tag
+        token_usage = { prompt = p, completion = c, total = p + c }
     }
 
     if resp.message then
         unified.content = resp.message.content or ""
+        -- Non-streaming tool calls were previously dropped: the field was
+        -- initialized to an empty table and never filled, while
+        -- transform_stream_chunk handled them. A non-streaming agent turn thus
+        -- looked like a plain answer and the tool loop stopped.
+        if type(resp.message.tool_calls) == "table" and #resp.message.tool_calls > 0 then
+            local tcs = {}
+            for _, tc in ipairs(resp.message.tool_calls) do
+                local fn = tc["function"] or {}
+                -- Ollama sends arguments as an object already
+                local args = fn.arguments
+                if type(args) == "string" then
+                    local aok, decoded = pcall(json.decode, args)
+                    args = aok and decoded or {}
+                elseif type(args) ~= "table" then
+                    args = {}
+                end
+                table.insert(tcs, {
+                    id = tc.id or ("call_" .. #tcs),
+                    type = tc.type or "function",
+                    name = fn.name or "",
+                    arguments = args
+                })
+            end
+            unified.tool_calls = tcs
+            -- Ollama reports done_reason "stop" alongside tool calls
+            unified.finish_reason = "tool_calls"
+        end
     end
 
     return json.encode(unified)
@@ -70,12 +143,43 @@ end
 function adapter.transform_stream_chunk(raw_chunk)
     local ok, chunk = pcall(json.decode, raw_chunk)
     if not ok then return "" end
-    if not chunk.message then return "" end
+
+    -- Ollama's terminal chunk (done=true) carries token counts but may omit
+    -- message; pass them through so the gateway emits real usage.
+    local uses = nil
+    if chunk.done then
+        local p = chunk.prompt_eval_count or 0
+        local c = chunk.eval_count or 0
+        if p > 0 or c > 0 then
+            uses = { prompt = p, completion = c, total = p + c }
+        end
+    end
+
+    -- Ollama done_reason -> OpenAI finish_reason ("length" 透传，其余归一 stop)
+    local finish = nil
+    if chunk.done then
+        if chunk.done_reason == "length" then
+            finish = "length"
+        else
+            finish = "stop"
+        end
+    end
+
+    if not chunk.message then
+        if uses ~= nil then
+            return json.encode({ content = "", done = true, finish_reason = finish, usage = uses })
+        end
+        return json.encode({ content = "", done = true, finish_reason = finish })
+    end
 
     local unified = {
         content = chunk.message.content or "",
-        done = chunk.done or false
+        done = chunk.done or false,
+        finish_reason = finish
     }
+    if uses ~= nil then
+        unified.usage = uses
+    end
     if chunk.message.reasoning_content then
         unified.reasoning_content = chunk.message.reasoning_content
     end
@@ -97,4 +201,15 @@ function adapter.transform_stream_chunk(raw_chunk)
     return json.encode(unified)
 end
 
+-- 错误收敛：Ollama 常见 {error:"..."} 字符串（新版本也有对象形态）
+function adapter.transform_error(status, body)
+    local ok, resp = pcall(json.decode, body)
+    if not ok or type(resp) ~= "table" then return nil end
+    if type(resp.error) == "string" then return resp.error end
+    if type(resp.error) == "table" and type(resp.error.message) == "string" then
+        return resp.error.message
+    end
+    return nil
+end
+
 return adapter

internal/lua/adapters/openai.lua

@@ -5,6 +5,36 @@ adapter.version = "2.0.0"
 adapter.endpoint = "/chat/completions"
 adapter.headers = {}
 
+-- Claude-behind-OpenAI upstreams (tabitoken, 扇贝, …) convert /chat/completions
+-- to the Anthropic Messages API internally, so they inherit Anthropic's
+-- tool id rule: ^[a-zA-Z0-9_-]{1,64}$, enforced by rejecting the WHOLE request
+-- with "Invalid tool use format" / "tool_use.id: String should match pattern".
+-- Plain OpenAI has no such rule, so an OpenAI-compatible model can mint an id
+-- like "bash:0" (observed from moonshotai/kimi-k3). In a fan-out router that id
+-- is replayed to every other source, so one such id kills every Claude slot at
+-- once and an AUTO request falls through all tiers.
+--
+-- safe_tool_id is pure and deterministic, so a tool_calls entry and its
+-- matching tool_call_id are rewritten identically within one request. A
+-- rewritten id keeps an 8-hex digest of the ORIGINAL id, without which two
+-- distinct ids could collapse into one ("a:b" and "a_b") and become an
+-- unpaired/duplicate tool call. Already-legal ids are returned untouched, so
+-- well-behaved traffic is byte-identical to before.
+-- (anthropic.lua carries the same helper; Lua adapters have no shared prelude.)
+local TOOL_ID_MAX = 64
+
+local function safe_tool_id(id)
+    if type(id) ~= "string" or id == "" then return id end
+    local clean = string.gsub(id, "[^A-Za-z0-9_-]", "_")
+    if clean == id and #clean <= TOOL_ID_MAX then
+        return clean
+    end
+    local digest = string.sub(sha256_hex(id), 1, 8)
+    local keep = TOOL_ID_MAX - #digest - 1
+    if #clean > keep then clean = string.sub(clean, 1, keep) end
+    return clean .. "_" .. digest
+end
+
 -- OpenAI /chat/completions format (pass-through, strip provider-specific fields)
 function adapter.transform_request(raw_body)
     local ok, req = pcall(json.decode, raw_body)
@@ -14,6 +44,16 @@ function adapter.transform_request(raw_body)
     if req.messages then
         for _, msg in ipairs(req.messages) do
             msg.reasoning_content = nil
+            if msg.tool_call_id ~= nil then
+                msg.tool_call_id = safe_tool_id(msg.tool_call_id)
+            end
+            if type(msg.tool_calls) == "table" then
+                for _, tc in ipairs(msg.tool_calls) do
+                    if type(tc) == "table" and tc.id ~= nil then
+                        tc.id = safe_tool_id(tc.id)
+                    end
+                end
+            end
         end
     end
     return json.encode(req)
@@ -33,6 +73,22 @@ function adapter.transform_response(raw_body)
         unified.token_usage.prompt = resp.usage.prompt_tokens or 0
         unified.token_usage.completion = resp.usage.completion_tokens or 0
         unified.token_usage.total = resp.usage.total_tokens or 0
+        -- Cache passthrough: OpenAI v2 prompt_tokens_details.cached_tokens
+        -- and DeepSeek-legacy prompt_cache_hit/miss_tokens. dsh reads
+        -- prompt_tokens_details.cached_tokens (falls back to the legacy
+        -- standalone field), so both shapes reach clients.
+        local hit = 0
+        if type(resp.usage.prompt_tokens_details) == "table" and resp.usage.prompt_tokens_details.cached_tokens ~= nil then
+            hit = resp.usage.prompt_tokens_details.cached_tokens
+            unified.token_usage.prompt_tokens_details = { cached_tokens = hit }
+        end
+        if (resp.usage.prompt_cache_hit_tokens or 0) > 0 then
+            unified.token_usage.prompt_cache_hit_tokens = resp.usage.prompt_cache_hit_tokens
+            unified.token_usage.prompt_cache_miss_tokens = resp.usage.prompt_cache_miss_tokens or 0
+            if hit == 0 then
+                unified.token_usage.prompt_tokens_details = { cached_tokens = resp.usage.prompt_cache_hit_tokens }
+            end
+        end
     end
 
     if type(resp.choices) == "table" and #resp.choices > 0 then
@@ -48,7 +104,7 @@ function adapter.transform_response(raw_body)
                     local args_ok, args = pcall(json.decode, tc["function"].arguments)
                     if not args_ok then args = {} end
                     table.insert(tcs, {
-                        id = tc.id,
+                        id = safe_tool_id(tc.id),
                         type = tc.type or "function",
                         name = tc["function"].name,
                         arguments = args
@@ -67,22 +123,106 @@ function adapter.transform_stream_chunk(raw_chunk)
     local ok, chunk = pcall(json.decode, raw_chunk)
     if not ok then return "" end
 
-    if not chunk.choices or #chunk.choices == 0 then return "" end
+    -- OpenAI-style streams may attach usage to a chunk with empty choices
+    -- (the final usage chunk). Keys must match Go's TokenUsage json tags
+    -- (prompt/completion/total); the gateway re-emits standard *_tokens.
+    local uses = nil
+    if type(chunk.usage) == "table" then
+        uses = {
+            prompt = chunk.usage.prompt_tokens or chunk.usage.prompt or 0,
+            completion = chunk.usage.completion_tokens or chunk.usage.completion or 0,
+            total = chunk.usage.total_tokens or chunk.usage.total or 0,
+        }
+        if type(chunk.usage.prompt_tokens_details) == "table" and chunk.usage.prompt_tokens_details.cached_tokens ~= nil then
+            uses.prompt_tokens_details = { cached_tokens = chunk.usage.prompt_tokens_details.cached_tokens }
+        elseif (chunk.usage.prompt_cache_hit_tokens or 0) > 0 then
+            uses.prompt_cache_hit_tokens = chunk.usage.prompt_cache_hit_tokens
+            uses.prompt_cache_miss_tokens = chunk.usage.prompt_cache_miss_tokens or 0
+            uses.prompt_tokens_details = { cached_tokens = chunk.usage.prompt_cache_hit_tokens }
+        end
+    end
+
+    if not chunk.choices or #chunk.choices == 0 then
+        if uses ~= nil then
+            -- usage-only chunk is not a content/finish signal; the gateway
+            -- emits its own terminal stop chunk and merges this usage.
+            return json.encode({ usage = uses, done = false })
+        end
+        return ""
+    end
     local delta = chunk.choices[1].delta or {}
     local fr = chunk.choices[1].finish_reason
 
+    local finish = (type(fr) == "string" and fr ~= "") and fr or nil
+
     local unified = {
         content = delta.content or "",
-        done = (fr ~= nil)
+        done = (finish ~= nil)
     }
+    if finish then
+        unified.finish_reason = finish
+    end
+    if uses ~= nil then
+        unified.usage = uses
+    end
     if delta.reasoning_content then
         unified.reasoning_content = delta.reasoning_content
     end
     if delta.tool_calls then
-        -- pass raw streaming fragments through; OpenAI clients accumulate index+id+name+arguments
+        -- Sanitize on the way OUT too: an id this upstream happily minted (it
+        -- does not validate them) becomes a landmine once the client replays
+        -- it to a Claude upstream. Only the first fragment of a streamed call
+        -- carries an id; later argument fragments have none and are untouched.
+        for _, tc in ipairs(delta.tool_calls) do
+            if type(tc) == "table" and tc.id ~= nil then
+                tc.id = safe_tool_id(tc.id)
+            end
+        end
         unified.tool_calls = delta.tool_calls
     end
     return json.encode(unified)
 end
 
+-- 错误收敛：标准 OpenAI 信封 {error:{message,...}}
+function adapter.transform_error(status, body)
+    -- Non-JSON body (nginx HTML error pages, plain text): extract a short
+    -- human-readable reason instead of letting the raw body reach the log.
+    local ok, resp = pcall(json.decode, body)
+    if not ok or type(resp) ~= "table" then
+        -- HTML error page: pull the <title> text (e.g. "413 Request Entity Too Large")
+        local title = string.match(body or "", "<title>(.-)</title>")
+        if title and title ~= "" then return title end
+        -- Bare text: first non-empty line, capped
+        local line = string.match(body or "", "^%s*([^\r\n]+)")
+        if line and line ~= "" and not string.match(line, "^<") then
+            return string.sub(line, 1, 200)
+        end
+        return nil
+    end
+
+    -- Standard OpenAI envelope: {error:{message,...}} or {error:"..."}
+    local e = resp.error
+    if type(e) == "table" and type(e.message) == "string" then
+        return e.message
+    end
+    if type(e) == "string" then return e end
+
+    -- Flat envelope used by many OpenAI-compatible gateways:
+    --   {"code":20012,"message":"Model does not exist..."}
+    --   {"code":"INVALID_API_KEY","message":"Invalid API key"}
+    if type(resp.message) == "string" and resp.message ~= "" then
+        if resp.code ~= nil then
+            return tostring(resp.code) .. ": " .. resp.message
+        end
+        return resp.message
+    end
+
+    -- Some gateways use {detail:"..."} (FastAPI style)
+    if type(resp.detail) == "string" and resp.detail ~= "" then
+        return resp.detail
+    end
+
+    return nil
+end
+
 return adapter

internal/lua/adapters/opencode.lua

@@ -4,18 +4,40 @@ adapter.name = "opencode"
 adapter.version = "1.0.0"
 adapter.endpoint = "/chat/completions"
 
--- opencode.ai zen 网关按 User-Agent 指纹识别官方客户端并把请求分到免费额度池；
--- 非官方 UA（curl/Go 默认等）会被分到匿名池并触发 FreeUsageLimitError。
--- 因此固定发送 opencode 客户端的 UA；配合源配置 api_key: "public"（官方无 key
--- 客户端实际发送 Bearer public）即可走免费池。
+-- opencode.ai zen 网关按客户端指纹（UA + x-opencode-* 头）路由请求池：
+-- 缺少 x-opencode-client/session/request/project 身份头的请求会被判为匿名
+-- 客户端，x-preview-f-free 等模型在带 tools 时上游直接失败——流式返回单
+-- chunk "finish_reason":"network_error" 空 content，非流式返回 503
+-- "Endpoint is unavailable"，表现为"空回复"。因此除 UA 外必须附带
+-- x-opencode-* 身份头；session/request 每请求派生唯一值（见 build_headers）。
 adapter.headers = {
-    ["User-Agent"] = "opencode/0.1.0",
+    ["User-Agent"] = "opencode/1.18.21 ai-sdk/provider-utils/4.0.23 runtime/bun/1.3.14",
 }
 
+-- 每请求生成身份头。沙箱无 os/math，用 meta.timestamp + 请求体哈希派生：
+-- 同秒内重复请求 id 相同可接受（zen 只校验存在性，不校验格式）。
+local function rand_id(prefix, seed)
+    return prefix .. string.sub(sha256_hex(seed), 1, 24)
+end
+
+function adapter.build_headers(meta)
+    local ts = tostring(meta.timestamp or "")
+    return {
+        ["User-Agent"] = adapter.headers["User-Agent"],
+        ["x-opencode-client"] = "cli",
+        -- project 固定：同网关实例共享一个工作区身份
+        ["x-opencode-project"] = string.sub(sha256_hex("llmsproxy|" .. (meta.source and meta.source.name or "")), 1, 32),
+        ["x-opencode-session"] = rand_id("ses_", "session|" .. ts),
+        ["x-opencode-request"] = rand_id("msg_", "request|" .. ts .. "|" .. tostring(meta.body or "")),
+    }
+end
+
 -- OpenAI /chat/completions format (pass-through, strip provider-specific fields)
 -- zen 上游 schema 只接受 text content part（无视觉/音频能力）：多模态 part
--- （image_url / input_audio / file 等）一律剥离；因此失去全部 content 的
--- 消息整条丢弃，避免上游 "unknown variant `image_url`, expected `text`"。
+-- （image_url / input_audio / file 等）一律剥离。剥离后 content 变空的消息
+-- 若不再携带 tool_calls / tool_call_id 才整条丢弃（避免上游
+-- "unknown variant `image_url`, expected `text`"）；带工具调用的必须保留，
+-- 否则会把紧随其后的 tool 结果变成孤儿，模型会反复重发同一个调用。
 -- zen 上游角色白名单只有 system / user / assistant / tool / latest_reminder：
 -- OpenAI 的 developer（及 function 等）不在其中，直接透传会触发上游
 -- "unknown variant `developer`, expected one of ..." 错误；统一归一化为 system。
@@ -32,6 +54,14 @@ function adapter.transform_request(raw_body)
     if not ok then return raw_body end
     req.disable_thinking = nil
     req.extra_body = nil
+    -- stream_options is only valid alongside stream:true. Sending it on a
+    -- non-streaming request is rejected by stricter upstreams (OpenCode Go:
+    -- "stream_options should be set along with stream"), which failed every
+    -- non-stream call. Only attach it when the request actually streams.
+    if req.stream then
+        if type(req.stream_options) ~= "table" then req.stream_options = {} end
+        req.stream_options.include_usage = true
+    end
     if req.messages then
         local kept = {}
         for _, msg in ipairs(req.messages) do
@@ -50,7 +80,24 @@ function adapter.transform_request(raw_body)
                     end
                 end
                 if #parts == 0 then
-                    drop = true
+                    -- Content collapsed to nothing after stripping unsupported
+                    -- parts. A message that still carries a tool call must
+                    -- NEVER be dropped: the very next message is its tool
+                    -- result, and dropping the call orphans that result. The
+                    -- model then sees a result for a call it never made and
+                    -- re-issues the same tool call on every turn (observed as
+                    -- an infinite "repeated tool call" loop).
+                    --
+                    -- This is the common shape for agent clients: an assistant
+                    -- turn whose content is only [thinking, toolCall] serialises
+                    -- to content:[] with tool_calls — exactly the case that used
+                    -- to vanish here. Emit "" instead, which zen accepts.
+                    if (type(msg.tool_calls) == "table" and #msg.tool_calls > 0)
+                        or msg.tool_call_id ~= nil then
+                        msg.content = ""
+                    else
+                        drop = true
+                    end
                 else
                     msg.content = parts
                 end
@@ -78,14 +125,27 @@ function adapter.transform_response(raw_body)
         unified.token_usage.prompt = resp.usage.prompt_tokens or 0
         unified.token_usage.completion = resp.usage.completion_tokens or 0
         unified.token_usage.total = resp.usage.total_tokens or 0
+        local hit = 0
+        if type(resp.usage.prompt_tokens_details) == "table" and resp.usage.prompt_tokens_details.cached_tokens ~= nil then
+            hit = resp.usage.prompt_tokens_details.cached_tokens
+            unified.token_usage.prompt_tokens_details = { cached_tokens = hit }
+        end
+        if (resp.usage.prompt_cache_hit_tokens or 0) > 0 then
+            unified.token_usage.prompt_cache_hit_tokens = resp.usage.prompt_cache_hit_tokens
+            unified.token_usage.prompt_cache_miss_tokens = resp.usage.prompt_cache_miss_tokens or 0
+            if hit == 0 then
+                unified.token_usage.prompt_tokens_details = { cached_tokens = resp.usage.prompt_cache_hit_tokens }
+            end
+        end
     end
 
     if type(resp.choices) == "table" and #resp.choices > 0 then
         local ch = resp.choices[1]
         if type(ch.message) == "table" then
             unified.content = ch.message.content or ""
-            if ch.message.reasoning_content then
-                unified.reasoning_content = ch.message.reasoning_content
+            local reasoning = ch.message.reasoning_content or ch.message.reasoning
+            if reasoning then
+                unified.reasoning_content = reasoning
             end
             if type(ch.message.tool_calls) == "table" then
                 local tcs = {}
@@ -112,16 +172,53 @@ function adapter.transform_stream_chunk(raw_chunk)
     local ok, chunk = pcall(json.decode, raw_chunk)
     if not ok then return "" end
 
-    if not chunk.choices or #chunk.choices == 0 then return "" end
+    -- OpenAI-style streams may attach usage to a chunk with empty choices
+    -- (the final usage chunk). Preserve it; the gateway emits it as the
+    -- terminal usage chunk. Note: keys must match Go's TokenUsage json tags
+    -- (prompt/completion/total); the gateway re-emits standard *_tokens.
+    local uses = nil
+    if type(chunk.usage) == "table" then
+        uses = {
+            prompt = chunk.usage.prompt_tokens or chunk.usage.prompt or 0,
+            completion = chunk.usage.completion_tokens or chunk.usage.completion or 0,
+            total = chunk.usage.total_tokens or chunk.usage.total or 0,
+        }
+        if type(chunk.usage.prompt_tokens_details) == "table" and chunk.usage.prompt_tokens_details.cached_tokens ~= nil then
+            uses.prompt_tokens_details = { cached_tokens = chunk.usage.prompt_tokens_details.cached_tokens }
+        elseif (chunk.usage.prompt_cache_hit_tokens or 0) > 0 then
+            uses.prompt_cache_hit_tokens = chunk.usage.prompt_cache_hit_tokens
+            uses.prompt_cache_miss_tokens = chunk.usage.prompt_cache_miss_tokens or 0
+            uses.prompt_tokens_details = { cached_tokens = chunk.usage.prompt_cache_hit_tokens }
+        end
+    end
+
+    if not chunk.choices or #chunk.choices == 0 then
+        if uses ~= nil then
+            -- usage-only chunk is not a content/finish signal; the gateway
+            -- emits its own terminal stop chunk and merges this usage.
+            return json.encode({ usage = uses, done = false })
+        end
+        return ""
+    end
     local delta = chunk.choices[1].delta or {}
     local fr = chunk.choices[1].finish_reason
 
+    local finish = (type(fr) == "string" and fr ~= "") and fr or nil
+
     local unified = {
         content = delta.content or "",
-        done = (fr ~= nil)
+        done = (finish ~= nil)
     }
-    if delta.reasoning_content then
-        unified.reasoning_content = delta.reasoning_content
+    if finish then
+        unified.finish_reason = finish
+    end
+    if uses ~= nil then
+        unified.usage = uses
+    end
+    -- zen 用 reasoning 字段承载推理文本（OpenAI 惯例是 reasoning_content）
+    local reasoning = delta.reasoning_content or delta.reasoning
+    if reasoning then
+        unified.reasoning_content = reasoning
     end
     if delta.tool_calls then
         -- pass raw streaming fragments through; OpenAI clients accumulate index+id+name+arguments
@@ -130,4 +227,17 @@ function adapter.transform_stream_chunk(raw_chunk)
     return json.encode(unified)
 end
 
+-- 错误收敛（可选钩子）：zen 错误信封固定为 {error={type,message}}；
+-- 免费池限流 FreeUsageLimitError 单独标注。返回 nil 走通用兜底。
+function adapter.transform_error(status, body)
+    local ok, resp = pcall(json.decode, body)
+    if not ok or type(resp) ~= "table" then return nil end
+    local e = resp.error
+    if type(e) ~= "table" then return nil end
+    if e.type == "FreeUsageLimitError" then
+        return "zen free pool quota exhausted"
+    end
+    return e.message
+end
+
 return adapter

internal/lua/adapters/opencodego.lua

@@ -0,0 +1,354 @@
+local adapter = {}
+
+adapter.name = "opencodego"
+adapter.version = "1.0.0"
+adapter.endpoint = "/chat/completions"
+
+-- OpenCode Go（包月订阅端点，https://opencode.ai/zen/go/v1）
+--
+-- 与 OpenCode Zen（opencodezen.lua，https://opencode.ai/zen/v1）是两个不同的
+-- 服务，行为要求并不相同，因此各有专用适配器：
+--
+--   * 本适配器（Go）服务包月订阅池。上游**强制要求**
+--     `x-opencode-session` 头，缺失直接 400 "MissingSessionID"。
+--   * Go 的 thinking 模型（deepseek-v4.1-flash 等）**必须回传** assistant 轮的
+--     `reasoning_content`，否则 400：
+--       The `reasoning_content` in the thinking mode must be passed back to the API.
+--     这是本适配器与 Zen 适配器最关键的差异 —— Zen 会剥掉它，Go 必须原样保留。
+--   * Go 上游对协议更严格：非流式请求带 `stream_options` 会被拒
+--     （"stream_options should be set along with stream"）。
+--
+-- 同样需要 opencode 客户端指纹（UA + x-opencode-*）才能被正确识别与路由。
+adapter.headers = {
+    ["User-Agent"] = "opencode/1.18.21 ai-sdk/provider-utils/4.0.23 runtime/bun/1.3.14",
+}
+
+-- 每请求生成身份头。沙箱无 os/math，用 meta.timestamp + 请求体哈希派生：
+-- 同秒内重复请求 id 相同可接受（zen 只校验存在性，不校验格式）。
+local function rand_id(prefix, seed)
+    return prefix .. string.sub(sha256_hex(seed), 1, 24)
+end
+
+-- 会话指纹：取历史里**第一条 user 消息**。
+--
+-- 为什么不直接用客户端的会话 id：实测抓包（tcpdump 抓 127.0.0.1:8081 的真实
+-- agent 请求）确认通用 OpenAI 客户端**根本不发**任何会话标识 —— body 里没有
+-- user / session_id / conversation_id / metadata，请求头也只有 X-Stainless-*
+-- （OpenAI JS SDK）与 User-Agent。opencode 原生客户端那套 x-opencode-session
+-- 是它自己的概念，通用客户端无从转发。
+--
+-- 退而求其次但要够用：历史被逐轮重放，**第一条 user 消息在整个会话里恒定**，
+-- 用它做指纹即可得到"每会话一个 session"，而不是"每源一个 session"。
+-- 拿不到时（无 user 消息）返回空串，调用方回退到按源稳定。
+local function conversation_fingerprint(body)
+    if type(body) ~= "string" or body == "" then return "" end
+    local ok, req = pcall(json.decode, body)
+    if not ok or type(req) ~= "table" then return "" end
+    for _, m in ipairs(req.messages or {}) do
+        if type(m) == "table" and m.role == "user" then
+            local c = m.content
+            if type(c) == "string" then
+                if c ~= "" then return c end
+            elseif type(c) == "table" then
+                for _, part in ipairs(c) do
+                    if type(part) == "table" and part.type == "text" and part.text then
+                        return part.text
+                    end
+                end
+            end
+            return ""
+        end
+    end
+    return ""
+end
+
+-- session_seed 决定上游会话号（x-opencode-session 的种子）。
+--
+-- 优先级：
+--   1) 客户端自带的会话标识（meta.client_session）——pi 等在开启
+--      compat.sendSessionAffinityHeaders 后会发 x-session-affinity，
+--      这是平台真实的会话 id，整个会话恒定且天然按会话隔离。
+--   2) 退路：历史里**第一条 user 消息**做会话指纹。
+--   3) 再退：只按源固定（连 user 消息都没有时）。
+--
+-- 为什么需要 2/3：通用 OpenAI 客户端默认**根本不发**会话标识 —— 实测抓包
+-- （tcpdump 抓 127.0.0.1:8081 真实 agent 请求）确认 body 里没有
+-- user / session_id / conversation_id / metadata，请求头也只有
+-- X-Stainless-*（OpenAI JS SDK）与 User-Agent。opencode 原生客户端那套
+-- x-opencode-session 是它自己的概念，通用客户端无从转发。
+--
+-- 为什么必须稳定：上游前缀缓存是**会话级**的。会话号每请求一变，
+-- 缓存永不命中（实测：固定会号第 2 次命中 5888，每请求换会号则恒为 0）。
+local function session_seed(meta)
+    local src = (meta.source and meta.source.name) or ""
+    local base = "session|llmsproxy|" .. src
+    local cs = meta.client_session
+    if type(cs) == "string" and cs ~= "" then
+        return base .. "|client|" .. cs
+    end
+    return base .. "|" .. conversation_fingerprint(meta.body)
+end
+
+function adapter.build_headers(meta)
+    local ts = tostring(meta.timestamp or "")
+    local src = (meta.source and meta.source.name) or ""
+    return {
+        ["User-Agent"] = adapter.headers["User-Agent"],
+        ["x-opencode-client"] = "cli",
+        -- project 固定：同网关实例共享一个工作区身份
+        ["x-opencode-project"] = string.sub(sha256_hex("llmsproxy|" .. src), 1, 32),
+        -- session 必须**按源稳定**，不能每请求换：上游的前缀缓存在同一 session
+        -- 内才复用。实测（同一段 6032 token 提示词）：
+        --   固定 session  -> 第 2 次命中 5888/6032，cached_tokens=5888
+        --   每请求换 session -> 永远 0 命中
+        -- 原先用 meta.timestamp 派生，等于每请求都是新会话，缓存永远无效，
+        -- 上游也无法做会话亲和路由。
+        ["x-opencode-session"] = rand_id("ses_", session_seed(meta)),
+        -- request id 仍每请求唯一（它只是请求标识，不参与缓存键）
+        ["x-opencode-request"] = rand_id("msg_", "request|" .. ts .. "|" .. tostring(meta.body or "")),
+    }
+end
+
+-- OpenAI /chat/completions format (pass-through, strip provider-specific fields)
+-- zen 上游 schema 只接受 text content part（无视觉/音频能力）：多模态 part
+-- （image_url / input_audio / file 等）一律剥离。剥离后 content 变空的消息
+-- 若不再携带 tool_calls / tool_call_id 才整条丢弃（避免上游
+-- "unknown variant `image_url`, expected `text`"）；带工具调用的必须保留，
+-- 否则会把紧随其后的 tool 结果变成孤儿，模型会反复重发同一个调用。
+-- zen 上游角色白名单只有 system / user / assistant / tool / latest_reminder：
+-- OpenAI 的 developer（及 function 等）不在其中，直接透传会触发上游
+-- "unknown variant `developer`, expected one of ..." 错误；统一归一化为 system。
+local ROLE_WHITELIST = {
+    system = true,
+    user = true,
+    assistant = true,
+    tool = true,
+    latest_reminder = true,
+}
+
+function adapter.transform_request(raw_body)
+    local ok, req = pcall(json.decode, raw_body)
+    if not ok then return raw_body end
+    req.disable_thinking = nil
+    req.extra_body = nil
+    -- stream_options is only valid alongside stream:true; sending it on a
+    -- non-streaming request is rejected by strict upstreams.
+    if req.stream then
+        if type(req.stream_options) ~= "table" then req.stream_options = {} end
+        req.stream_options.include_usage = true
+    end
+    if req.messages then
+        local kept = {}
+        for _, msg in ipairs(req.messages) do
+            if type(msg.role) == "string" and not ROLE_WHITELIST[msg.role] then
+                msg.role = "system"
+            end
+            -- Go 的 thinking 模式**必须**回传 reasoning_content：剥掉它会让
+            -- 上游直接 400（"The `reasoning_content` in the thinking mode must
+            -- be passed back to the API"），agent 的每一轮都会失败。
+            -- 注意：与 opencodezen.lua 的行为**相反**，不要在这里剥。
+            --
+            -- 但绝大多数客户端（pi 等）根本不保存也不回传 reasoning，只保留
+            -- 工具调用本身。上游只在“带 tool_calls 的助手轮”上校验这个字段，
+            -- 实测**空串即可通过校验**，所以缺省时补空串：既满足上游的
+            -- 一致性要求，又不伪造任何推理内容（用户看到的 reasoning 仍然是
+            -- 上游本轮真实返回的）。
+            if msg.role == "assistant"
+                and type(msg.tool_calls) == "table" and #msg.tool_calls > 0
+                and msg.reasoning_content == nil then
+                msg.reasoning_content = ""
+            end
+            local drop = false
+            if type(msg.content) == "table" then
+                local parts = {}
+                for _, part in ipairs(msg.content) do
+                    if type(part) == "table" and part.type ~= nil and part.type ~= "text" then
+                        -- multimodal part not supported by zen
+                    else
+                        table.insert(parts, part)
+                    end
+                end
+                if #parts == 0 then
+                    -- Content collapsed to nothing after stripping unsupported
+                    -- parts. A message that still carries a tool call must
+                    -- NEVER be dropped: the very next message is its tool
+                    -- result, and dropping the call orphans that result. The
+                    -- model then sees a result for a call it never made and
+                    -- re-issues the same tool call on every turn (observed as
+                    -- an infinite "repeated tool call" loop).
+                    if (type(msg.tool_calls) == "table" and #msg.tool_calls > 0)
+                        or msg.tool_call_id ~= nil then
+                        msg.content = ""
+                    else
+                        drop = true
+                    end
+                else
+                    msg.content = parts
+                end
+            end
+            if not drop then
+                table.insert(kept, msg)
+            end
+        end
+        req.messages = kept
+    end
+    return json.encode(req)
+end
+
+function adapter.transform_response(raw_body)
+    local ok, resp = pcall(json.decode, raw_body)
+    if not ok or resp == nil then return raw_body end
+
+    local unified = {
+        content = "",
+        finish_reason = "",
+        token_usage = { prompt = 0, completion = 0, total = 0 }
+    }
+
+    if type(resp.usage) == "table" then
+        unified.token_usage.prompt = resp.usage.prompt_tokens or 0
+        unified.token_usage.completion = resp.usage.completion_tokens or 0
+        unified.token_usage.total = resp.usage.total_tokens or 0
+        local hit = 0
+        if type(resp.usage.prompt_tokens_details) == "table" and resp.usage.prompt_tokens_details.cached_tokens ~= nil then
+            hit = resp.usage.prompt_tokens_details.cached_tokens
+            unified.token_usage.prompt_tokens_details = { cached_tokens = hit }
+        end
+        if (resp.usage.prompt_cache_hit_tokens or 0) > 0 then
+            unified.token_usage.prompt_cache_hit_tokens = resp.usage.prompt_cache_hit_tokens
+            unified.token_usage.prompt_cache_miss_tokens = resp.usage.prompt_cache_miss_tokens or 0
+            if hit == 0 then
+                unified.token_usage.prompt_tokens_details = { cached_tokens = resp.usage.prompt_cache_hit_tokens }
+            end
+        end
+        -- 上游报告输出里有多少是思考 token。不给客户端的话，无法判断
+        -- completion_tokens 里多少是「可见回答」、多少是「思考」（都按输出计费）。
+        local ctd = resp.usage.completion_tokens_details
+        if type(ctd) == "table" and ctd.reasoning_tokens ~= nil then
+            unified.token_usage.completion_tokens_details = { reasoning_tokens = ctd.reasoning_tokens }
+        end
+    end
+
+    -- 本次调用的费用，上游放在**顶层**且是**字符串**（如 "0"、"0.0012"）。
+    -- Go 订阅是包月制，恒为 "0"；只有 Zen 按量付费模型才有信息量。
+    -- 原样透传：网关不解析、不换算、不汇总 —— 它只是上游事实的搬运者。
+    if resp.cost ~= nil then
+        unified.cost = tostring(resp.cost)
+    end
+
+    if type(resp.choices) == "table" and #resp.choices > 0 then
+        local ch = resp.choices[1]
+        if type(ch.message) == "table" then
+            unified.content = ch.message.content or ""
+            local reasoning = ch.message.reasoning_content or ch.message.reasoning
+            if reasoning then
+                unified.reasoning_content = reasoning
+            end
+            if type(ch.message.tool_calls) == "table" then
+                local tcs = {}
+                for _, tc in ipairs(ch.message.tool_calls) do
+                    local args_ok, args = pcall(json.decode, tc["function"].arguments)
+                    if not args_ok then args = {} end
+                    table.insert(tcs, {
+                        id = tc.id,
+                        type = tc.type or "function",
+                        name = tc["function"].name,
+                        arguments = args
+                    })
+                end
+                unified.tool_calls = tcs
+            end
+        end
+        unified.finish_reason = ch.finish_reason or ""
+    end
+
+    return json.encode(unified)
+end
+
+function adapter.transform_stream_chunk(raw_chunk)
+    local ok, chunk = pcall(json.decode, raw_chunk)
+    if not ok then return "" end
+
+    -- OpenAI-style streams may attach usage to a chunk with empty choices
+    -- (the final usage chunk). Preserve it; the gateway emits it as the
+    -- terminal usage chunk. Note: keys must match Go's TokenUsage json tags
+    -- (prompt/completion/total); the gateway re-emits standard *_tokens.
+    local uses = nil
+    if type(chunk.usage) == "table" then
+        uses = {
+            prompt = chunk.usage.prompt_tokens or chunk.usage.prompt or 0,
+            completion = chunk.usage.completion_tokens or chunk.usage.completion or 0,
+            total = chunk.usage.total_tokens or chunk.usage.total or 0,
+        }
+        -- 独立字段与 prompt_tokens_details **并存**透传，不要写成 elseif：
+        -- 上游命中缓存时两者都发，非流式路径也是两个都带。写成 elseif 会让
+        -- 流式丢掉 prompt_cache_hit_tokens/miss，与同一源的**非流式**产出不一致
+        -- （dsh 优先读 details，但只认独立字段的老客户端会看不到缓存）。
+        if type(chunk.usage.prompt_tokens_details) == "table" and chunk.usage.prompt_tokens_details.cached_tokens ~= nil then
+            uses.prompt_tokens_details = { cached_tokens = chunk.usage.prompt_tokens_details.cached_tokens }
+        end
+        if (chunk.usage.prompt_cache_hit_tokens or 0) > 0 then
+            uses.prompt_cache_hit_tokens = chunk.usage.prompt_cache_hit_tokens
+            uses.prompt_cache_miss_tokens = chunk.usage.prompt_cache_miss_tokens or 0
+            if uses.prompt_tokens_details == nil then
+                uses.prompt_tokens_details = { cached_tokens = chunk.usage.prompt_cache_hit_tokens }
+            end
+        end
+        local ctd = chunk.usage.completion_tokens_details
+        if type(ctd) == "table" and ctd.reasoning_tokens ~= nil then
+            uses.completion_tokens_details = { reasoning_tokens = ctd.reasoning_tokens }
+        end
+    end
+
+    if not chunk.choices or #chunk.choices == 0 then
+        -- 空 choices 的帧既不是内容也不是结束信号。上游发两种：
+        --   {"choices":[],"usage":{...}}   —— 终帧用量
+        --   {"choices":[],"cost":"0"}      —— 单独的费用帧
+        -- 两者可能落在同一帧上，必须都转出去；早退只带 usage 会把费用丢干净。
+        local out = { done = false }
+        if uses ~= nil then out.usage = uses end
+        if chunk.cost ~= nil then out.cost = tostring(chunk.cost) end
+        if uses == nil and chunk.cost == nil then return "" end
+        return json.encode(out)
+    end
+    local delta = chunk.choices[1].delta or {}
+    local fr = chunk.choices[1].finish_reason
+
+    local finish = (type(fr) == "string" and fr ~= "") and fr or nil
+
+    local unified = {
+        content = delta.content or "",
+        done = (finish ~= nil)
+    }
+    if finish then
+        unified.finish_reason = finish
+    end
+    if uses ~= nil then
+        unified.usage = uses
+    end
+    -- zen 用 reasoning 字段承载推理文本（OpenAI 惯例是 reasoning_content）
+    local reasoning = delta.reasoning_content or delta.reasoning
+    if reasoning then
+        unified.reasoning_content = reasoning
+    end
+    if delta.tool_calls then
+        -- pass raw streaming fragments through; OpenAI clients accumulate index+id+name+arguments
+        unified.tool_calls = delta.tool_calls
+    end
+    return json.encode(unified)
+end
+
+-- 错误收敛（可选钩子）：zen 错误信封固定为 {error={type,message}}；
+-- 免费池限流 FreeUsageLimitError 单独标注。返回 nil 走通用兜底。
+function adapter.transform_error(status, body)
+    local ok, resp = pcall(json.decode, body)
+    if not ok or type(resp) ~= "table" then return nil end
+    local e = resp.error
+    if type(e) ~= "table" then return nil end
+    if e.type == "FreeUsageLimitError" then
+        return "zen free pool quota exhausted"
+    end
+    return e.message
+end
+
+return adapter

internal/lua/adapters/opencodezen.lua

@@ -0,0 +1,342 @@
+local adapter = {}
+
+adapter.name = "opencodezen"
+adapter.version = "1.0.0"
+adapter.endpoint = "/chat/completions"
+
+-- OpenCode Zen（按量付费端点，https://opencode.ai/zen/v1）
+--
+-- 与 OpenCode Go（opencodego.lua，https://opencode.ai/zen/go/v1）是两个不同的
+-- 服务，行为要求并不相同，因此各有专用适配器：
+--
+--   * 本适配器（Zen）服务免费池 / 按量付费池。免费模型必须带 opencode 客户端
+--     指纹（UA + x-opencode-*），否则上游拒绝："free tier can only be used in
+--     OpenCode"。Zen 的免费池以非 thinking 模型为主，历史上回传
+--     reasoning_content 会引出上游报错，故这里剥掉（Go 侧相反，见 opencodego.lua）。
+--   * Go 订阅读者请看 opencodego.lua：那边**必须**回传 reasoning_content，
+--     否则 thinking 模式直接 400。
+--
+-- 缺 x-opencode-client/session/request/project 身份头会被判为匿名客户端，
+-- 部分模型在带 tools 时上游直接失败（流式返回单 chunk
+-- "finish_reason":"network_error" 空 content，非流式 503
+-- "Endpoint is unavailable"，表现为"空回复"）。
+adapter.headers = {
+    ["User-Agent"] = "opencode/1.18.21 ai-sdk/provider-utils/4.0.23 runtime/bun/1.3.14",
+}
+
+-- 每请求生成身份头。沙箱无 os/math，用 meta.timestamp + 请求体哈希派生：
+-- 同秒内重复请求 id 相同可接受（zen 只校验存在性，不校验格式）。
+local function rand_id(prefix, seed)
+    return prefix .. string.sub(sha256_hex(seed), 1, 24)
+end
+
+-- 会话指纹：取历史里**第一条 user 消息**。
+--
+-- 为什么不直接用客户端的会话 id：实测抓包（tcpdump 抓 127.0.0.1:8081 的真实
+-- agent 请求）确认通用 OpenAI 客户端**根本不发**任何会话标识 —— body 里没有
+-- user / session_id / conversation_id / metadata，请求头也只有 X-Stainless-*
+-- （OpenAI JS SDK）与 User-Agent。opencode 原生客户端那套 x-opencode-session
+-- 是它自己的概念，通用客户端无从转发。
+--
+-- 退而求其次但要够用：历史被逐轮重放，**第一条 user 消息在整个会话里恒定**，
+-- 用它做指纹即可得到"每会话一个 session"，而不是"每源一个 session"。
+-- 拿不到时（无 user 消息）返回空串，调用方回退到按源稳定。
+local function conversation_fingerprint(body)
+    if type(body) ~= "string" or body == "" then return "" end
+    local ok, req = pcall(json.decode, body)
+    if not ok or type(req) ~= "table" then return "" end
+    for _, m in ipairs(req.messages or {}) do
+        if type(m) == "table" and m.role == "user" then
+            local c = m.content
+            if type(c) == "string" then
+                if c ~= "" then return c end
+            elseif type(c) == "table" then
+                for _, part in ipairs(c) do
+                    if type(part) == "table" and part.type == "text" and part.text then
+                        return part.text
+                    end
+                end
+            end
+            return ""
+        end
+    end
+    return ""
+end
+
+-- session_seed 决定上游会话号（x-opencode-session 的种子）。
+--
+-- 优先级：
+--   1) 客户端自带的会话标识（meta.client_session）——pi 等在开启
+--      compat.sendSessionAffinityHeaders 后会发 x-session-affinity，
+--      这是平台真实的会话 id，整个会话恒定且天然按会话隔离。
+--   2) 退路：历史里**第一条 user 消息**做会话指纹。
+--   3) 再退：只按源固定（连 user 消息都没有时）。
+--
+-- 为什么需要 2/3：通用 OpenAI 客户端默认**根本不发**会话标识 —— 实测抓包
+-- （tcpdump 抓 127.0.0.1:8081 真实 agent 请求）确认 body 里没有
+-- user / session_id / conversation_id / metadata，请求头也只有
+-- X-Stainless-*（OpenAI JS SDK）与 User-Agent。opencode 原生客户端那套
+-- x-opencode-session 是它自己的概念，通用客户端无从转发。
+--
+-- 为什么必须稳定：上游前缀缓存是**会话级**的。会话号每请求一变，
+-- 缓存永不命中（实测：固定会号第 2 次命中 5888，每请求换会号则恒为 0）。
+local function session_seed(meta)
+    local src = (meta.source and meta.source.name) or ""
+    local base = "session|llmsproxy|" .. src
+    local cs = meta.client_session
+    if type(cs) == "string" and cs ~= "" then
+        return base .. "|client|" .. cs
+    end
+    return base .. "|" .. conversation_fingerprint(meta.body)
+end
+
+function adapter.build_headers(meta)
+    local ts = tostring(meta.timestamp or "")
+    local src = (meta.source and meta.source.name) or ""
+    return {
+        ["User-Agent"] = adapter.headers["User-Agent"],
+        ["x-opencode-client"] = "cli",
+        -- project 固定：同网关实例共享一个工作区身份
+        ["x-opencode-project"] = string.sub(sha256_hex("llmsproxy|" .. src), 1, 32),
+        -- session 必须**按源稳定**，不能每请求换：上游的前缀缓存在同一 session
+        -- 内才复用。实测（同一段 6032 token 提示词）：
+        --   固定 session  -> 第 2 次命中 5888/6032，cached_tokens=5888
+        --   每请求换 session -> 永远 0 命中
+        -- 原先用 meta.timestamp 派生，等于每请求都是新会话，缓存永远无效，
+        -- 上游也无法做会话亲和路由。
+        ["x-opencode-session"] = rand_id("ses_", session_seed(meta)),
+        -- request id 仍每请求唯一（它只是请求标识，不参与缓存键）
+        ["x-opencode-request"] = rand_id("msg_", "request|" .. ts .. "|" .. tostring(meta.body or "")),
+    }
+end
+
+-- OpenAI /chat/completions format (pass-through, strip provider-specific fields)
+-- zen 上游 schema 只接受 text content part（无视觉/音频能力）：多模态 part
+-- （image_url / input_audio / file 等）一律剥离。剥离后 content 变空的消息
+-- 若不再携带 tool_calls / tool_call_id 才整条丢弃（避免上游
+-- "unknown variant `image_url`, expected `text`"）；带工具调用的必须保留，
+-- 否则会把紧随其后的 tool 结果变成孤儿，模型会反复重发同一个调用。
+-- zen 上游角色白名单只有 system / user / assistant / tool / latest_reminder：
+-- OpenAI 的 developer（及 function 等）不在其中，直接透传会触发上游
+-- "unknown variant `developer`, expected one of ..." 错误；统一归一化为 system。
+local ROLE_WHITELIST = {
+    system = true,
+    user = true,
+    assistant = true,
+    tool = true,
+    latest_reminder = true,
+}
+
+function adapter.transform_request(raw_body)
+    local ok, req = pcall(json.decode, raw_body)
+    if not ok then return raw_body end
+    req.disable_thinking = nil
+    req.extra_body = nil
+    -- stream_options is only valid alongside stream:true; sending it on a
+    -- non-streaming request is rejected by strict upstreams.
+    if req.stream then
+        if type(req.stream_options) ~= "table" then req.stream_options = {} end
+        req.stream_options.include_usage = true
+    end
+    if req.messages then
+        local kept = {}
+        for _, msg in ipairs(req.messages) do
+            if type(msg.role) == "string" and not ROLE_WHITELIST[msg.role] then
+                msg.role = "system"
+            end
+            -- Zen 免费池不接受回传 reasoning_content（Go 侧相反）
+            msg.reasoning_content = nil
+            local drop = false
+            if type(msg.content) == "table" then
+                local parts = {}
+                for _, part in ipairs(msg.content) do
+                    if type(part) == "table" and part.type ~= nil and part.type ~= "text" then
+                        -- multimodal part not supported by zen
+                    else
+                        table.insert(parts, part)
+                    end
+                end
+                if #parts == 0 then
+                    -- Content collapsed to nothing after stripping unsupported
+                    -- parts. A message that still carries a tool call must
+                    -- NEVER be dropped: the very next message is its tool
+                    -- result, and dropping the call orphans that result. The
+                    -- model then sees a result for a call it never made and
+                    -- re-issues the same tool call on every turn (observed as
+                    -- an infinite "repeated tool call" loop).
+                    if (type(msg.tool_calls) == "table" and #msg.tool_calls > 0)
+                        or msg.tool_call_id ~= nil then
+                        msg.content = ""
+                    else
+                        drop = true
+                    end
+                else
+                    msg.content = parts
+                end
+            end
+            if not drop then
+                table.insert(kept, msg)
+            end
+        end
+        req.messages = kept
+    end
+    return json.encode(req)
+end
+
+function adapter.transform_response(raw_body)
+    local ok, resp = pcall(json.decode, raw_body)
+    if not ok or resp == nil then return raw_body end
+
+    local unified = {
+        content = "",
+        finish_reason = "",
+        token_usage = { prompt = 0, completion = 0, total = 0 }
+    }
+
+    if type(resp.usage) == "table" then
+        unified.token_usage.prompt = resp.usage.prompt_tokens or 0
+        unified.token_usage.completion = resp.usage.completion_tokens or 0
+        unified.token_usage.total = resp.usage.total_tokens or 0
+        local hit = 0
+        if type(resp.usage.prompt_tokens_details) == "table" and resp.usage.prompt_tokens_details.cached_tokens ~= nil then
+            hit = resp.usage.prompt_tokens_details.cached_tokens
+            unified.token_usage.prompt_tokens_details = { cached_tokens = hit }
+        end
+        if (resp.usage.prompt_cache_hit_tokens or 0) > 0 then
+            unified.token_usage.prompt_cache_hit_tokens = resp.usage.prompt_cache_hit_tokens
+            unified.token_usage.prompt_cache_miss_tokens = resp.usage.prompt_cache_miss_tokens or 0
+            if hit == 0 then
+                unified.token_usage.prompt_tokens_details = { cached_tokens = resp.usage.prompt_cache_hit_tokens }
+            end
+        end
+        -- 上游报告输出里有多少是思考 token。不给客户端的话，无法判断
+        -- completion_tokens 里多少是「可见回答」、多少是「思考」（都按输出计费）。
+        local ctd = resp.usage.completion_tokens_details
+        if type(ctd) == "table" and ctd.reasoning_tokens ~= nil then
+            unified.token_usage.completion_tokens_details = { reasoning_tokens = ctd.reasoning_tokens }
+        end
+    end
+
+    -- 本次调用的费用，上游放在**顶层**且是**字符串**（如 "0"、"0.0012"）。
+    -- Go 订阅是包月制，恒为 "0"；只有 Zen 按量付费模型才有信息量。
+    -- 原样透传：网关不解析、不换算、不汇总 —— 它只是上游事实的搬运者。
+    if resp.cost ~= nil then
+        unified.cost = tostring(resp.cost)
+    end
+
+    if type(resp.choices) == "table" and #resp.choices > 0 then
+        local ch = resp.choices[1]
+        if type(ch.message) == "table" then
+            unified.content = ch.message.content or ""
+            local reasoning = ch.message.reasoning_content or ch.message.reasoning
+            if reasoning then
+                unified.reasoning_content = reasoning
+            end
+            if type(ch.message.tool_calls) == "table" then
+                local tcs = {}
+                for _, tc in ipairs(ch.message.tool_calls) do
+                    local args_ok, args = pcall(json.decode, tc["function"].arguments)
+                    if not args_ok then args = {} end
+                    table.insert(tcs, {
+                        id = tc.id,
+                        type = tc.type or "function",
+                        name = tc["function"].name,
+                        arguments = args
+                    })
+                end
+                unified.tool_calls = tcs
+            end
+        end
+        unified.finish_reason = ch.finish_reason or ""
+    end
+
+    return json.encode(unified)
+end
+
+function adapter.transform_stream_chunk(raw_chunk)
+    local ok, chunk = pcall(json.decode, raw_chunk)
+    if not ok then return "" end
+
+    -- OpenAI-style streams may attach usage to a chunk with empty choices
+    -- (the final usage chunk). Preserve it; the gateway emits it as the
+    -- terminal usage chunk. Note: keys must match Go's TokenUsage json tags
+    -- (prompt/completion/total); the gateway re-emits standard *_tokens.
+    local uses = nil
+    if type(chunk.usage) == "table" then
+        uses = {
+            prompt = chunk.usage.prompt_tokens or chunk.usage.prompt or 0,
+            completion = chunk.usage.completion_tokens or chunk.usage.completion or 0,
+            total = chunk.usage.total_tokens or chunk.usage.total or 0,
+        }
+        -- 独立字段与 prompt_tokens_details **并存**透传，不要写成 elseif：
+        -- 上游命中缓存时两者都发，非流式路径也是两个都带。写成 elseif 会让
+        -- 流式丢掉 prompt_cache_hit_tokens/miss，与同一源的**非流式**产出不一致
+        -- （dsh 优先读 details，但只认独立字段的老客户端会看不到缓存）。
+        if type(chunk.usage.prompt_tokens_details) == "table" and chunk.usage.prompt_tokens_details.cached_tokens ~= nil then
+            uses.prompt_tokens_details = { cached_tokens = chunk.usage.prompt_tokens_details.cached_tokens }
+        end
+        if (chunk.usage.prompt_cache_hit_tokens or 0) > 0 then
+            uses.prompt_cache_hit_tokens = chunk.usage.prompt_cache_hit_tokens
+            uses.prompt_cache_miss_tokens = chunk.usage.prompt_cache_miss_tokens or 0
+            if uses.prompt_tokens_details == nil then
+                uses.prompt_tokens_details = { cached_tokens = chunk.usage.prompt_cache_hit_tokens }
+            end
+        end
+        local ctd = chunk.usage.completion_tokens_details
+        if type(ctd) == "table" and ctd.reasoning_tokens ~= nil then
+            uses.completion_tokens_details = { reasoning_tokens = ctd.reasoning_tokens }
+        end
+    end
+
+    if not chunk.choices or #chunk.choices == 0 then
+        -- 空 choices 的帧既不是内容也不是结束信号。上游发两种：
+        --   {"choices":[],"usage":{...}}   —— 终帧用量
+        --   {"choices":[],"cost":"0"}      —— 单独的费用帧
+        -- 两者可能落在同一帧上，必须都转出去；早退只带 usage 会把费用丢干净。
+        local out = { done = false }
+        if uses ~= nil then out.usage = uses end
+        if chunk.cost ~= nil then out.cost = tostring(chunk.cost) end
+        if uses == nil and chunk.cost == nil then return "" end
+        return json.encode(out)
+    end
+    local delta = chunk.choices[1].delta or {}
+    local fr = chunk.choices[1].finish_reason
+
+    local finish = (type(fr) == "string" and fr ~= "") and fr or nil
+
+    local unified = {
+        content = delta.content or "",
+        done = (finish ~= nil)
+    }
+    if finish then
+        unified.finish_reason = finish
+    end
+    if uses ~= nil then
+        unified.usage = uses
+    end
+    -- zen 用 reasoning 字段承载推理文本（OpenAI 惯例是 reasoning_content）
+    local reasoning = delta.reasoning_content or delta.reasoning
+    if reasoning then
+        unified.reasoning_content = reasoning
+    end
+    if delta.tool_calls then
+        -- pass raw streaming fragments through; OpenAI clients accumulate index+id+name+arguments
+        unified.tool_calls = delta.tool_calls
+    end
+    return json.encode(unified)
+end
+
+-- 错误收敛（可选钩子）：zen 错误信封固定为 {error={type,message}}；
+-- 免费池限流 FreeUsageLimitError 单独标注。返回 nil 走通用兜底。
+function adapter.transform_error(status, body)
+    local ok, resp = pcall(json.decode, body)
+    if not ok or type(resp) ~= "table" then return nil end
+    local e = resp.error
+    if type(e) ~= "table" then return nil end
+    if e.type == "FreeUsageLimitError" then
+        return "zen free pool quota exhausted"
+    end
+    return e.message
+end
+
+return adapter

internal/lua/adapters/sensenova.lua

@@ -0,0 +1,174 @@
+-- sensenova API format adapter
+-- streaming: delta only has reasoning_content, no content field
+-- non-streaming: has both content and reasoning_content
+
+local adapter = {}
+adapter.name = "sensenova"
+adapter.version = "1.0.0"
+adapter.endpoint = "/chat/completions"
+adapter.headers = {}
+
+-- Same as openai - strip provider-specific fields
+function adapter.transform_request(raw_body)
+    local ok, req = pcall(json.decode, raw_body)
+    if not ok then return raw_body end
+    req.disable_thinking = nil
+    req.extra_body = nil
+    if req.messages then
+        for _, msg in ipairs(req.messages) do
+            msg.reasoning_content = nil
+        end
+    end
+    return json.encode(req)
+end
+
+-- Same as openai - extract content from response
+function adapter.transform_response(raw_body)
+    local ok, resp = pcall(json.decode, raw_body)
+    if not ok or resp == nil then return raw_body end
+
+    local unified = {
+        content = "",
+        finish_reason = "",
+        token_usage = { prompt = 0, completion = 0, total = 0 }
+    }
+
+    if type(resp.usage) == "table" then
+        unified.token_usage.prompt = resp.usage.prompt_tokens or 0
+        unified.token_usage.completion = resp.usage.completion_tokens or 0
+        unified.token_usage.total = resp.usage.total_tokens or 0
+        local hit = 0
+        if type(resp.usage.prompt_tokens_details) == "table" and resp.usage.prompt_tokens_details.cached_tokens ~= nil then
+            hit = resp.usage.prompt_tokens_details.cached_tokens
+            unified.token_usage.prompt_tokens_details = { cached_tokens = hit }
+        end
+        if (resp.usage.prompt_cache_hit_tokens or 0) > 0 then
+            unified.token_usage.prompt_cache_hit_tokens = resp.usage.prompt_cache_hit_tokens
+            unified.token_usage.prompt_cache_miss_tokens = resp.usage.prompt_cache_miss_tokens or 0
+            if hit == 0 then
+                unified.token_usage.prompt_tokens_details = { cached_tokens = resp.usage.prompt_cache_hit_tokens }
+            end
+        end
+    end
+
+    if type(resp.choices) == "table" and #resp.choices > 0 then
+        local ch = resp.choices[1]
+        if type(ch.message) == "table" then
+            unified.content = ch.message.content or ""
+            if ch.message.reasoning_content then
+                unified.reasoning_content = ch.message.reasoning_content
+            elseif ch.message.reasoning then
+                -- sensenova-6.8-flash-lite 用 reasoning 字段而不是 reasoning_content
+                unified.reasoning_content = ch.message.reasoning
+            end
+            -- Tool calls MUST be forwarded. Dropping them while keeping
+            -- finish_reason="tool_calls" makes the client replay an assistant
+            -- message whose function name/arguments are empty, and sensenova
+            -- then rejects the next turn with
+            --   400 invalid tool_call function, function/name/arguments cannot be empty
+            -- i.e. a tool-using conversation dies on its second request.
+            if type(ch.message.tool_calls) == "table" and #ch.message.tool_calls > 0 then
+                local tcs = {}
+                for _, tc in ipairs(ch.message.tool_calls) do
+                    local fn = tc["function"] or {}
+                    -- arguments arrives as a JSON *string* on the wire; the
+                    -- unified shape expects a decoded object.
+                    local args = fn.arguments
+                    if type(args) == "string" then
+                        local args_ok, decoded = pcall(json.decode, args)
+                        args = args_ok and decoded or {}
+                    elseif type(args) ~= "table" then
+                        args = {}
+                    end
+                    table.insert(tcs, {
+                        id = tc.id,
+                        type = tc.type or "function",
+                        name = fn.name,
+                        arguments = args
+                    })
+                end
+                unified.tool_calls = tcs
+            end
+        end
+        unified.finish_reason = ch.finish_reason or ""
+    end
+
+    return json.encode(unified)
+end
+
+-- Sensenova-specific stream handling
+-- Upstream puts content in reasoning_content only (no content field)
+-- Also sends finish_reason="" (empty string) on every chunk
+function adapter.transform_stream_chunk(raw_chunk)
+    local ok, chunk = pcall(json.decode, raw_chunk)
+    if not ok then return "" end
+
+    local uses = nil
+    if type(chunk.usage) == "table" then
+        uses = {
+            prompt = chunk.usage.prompt_tokens or chunk.usage.prompt or 0,
+            completion = chunk.usage.completion_tokens or chunk.usage.completion or 0,
+            total = chunk.usage.total_tokens or chunk.usage.total or 0,
+        }
+        if type(chunk.usage.prompt_tokens_details) == "table" and chunk.usage.prompt_tokens_details.cached_tokens ~= nil then
+            uses.prompt_tokens_details = { cached_tokens = chunk.usage.prompt_tokens_details.cached_tokens }
+        elseif (chunk.usage.prompt_cache_hit_tokens or 0) > 0 then
+            uses.prompt_cache_hit_tokens = chunk.usage.prompt_cache_hit_tokens
+            uses.prompt_cache_miss_tokens = chunk.usage.prompt_cache_miss_tokens or 0
+            uses.prompt_tokens_details = { cached_tokens = chunk.usage.prompt_cache_hit_tokens }
+        end
+    end
+
+    if not chunk.choices or #chunk.choices == 0 then
+        if uses ~= nil then
+            return json.encode({ usage = uses, done = false })
+        end
+        return ""
+    end
+
+    local delta = chunk.choices[1].delta or {}
+
+    -- Sensenova: delta has reasoning_content (deepseek-v4-flash) or reasoning
+    -- (sensenova-6.8-flash-lite) but no content field
+    local content = delta.content or ""
+    if content == "" and (delta.reasoning_content or delta.reasoning) then
+        content = delta.reasoning_content or delta.reasoning
+    end
+
+    -- Sensenova: finish_reason is "" on every chunk, "stop" on last
+    local fr = chunk.choices[1].finish_reason
+    local done = (fr == "stop" or fr == "length")
+
+    local unified = {
+        content = content,
+        done = done,
+    }
+    if chunk.choices[1].finish_reason and chunk.choices[1].finish_reason ~= "" then
+        unified.finish_reason = chunk.choices[1].finish_reason
+    end
+    if delta.tool_calls then
+        unified.tool_calls = delta.tool_calls
+    end
+    if uses ~= nil then
+        unified.usage = uses
+    end
+
+    return json.encode(unified)
+end
+
+-- 错误收敛：sensenova 为 OpenAI 风格 {error:{message,...}}；
+-- 配额类错误单独点出便于客户端识别重置周期。
+function adapter.transform_error(status, body)
+    local ok, resp = pcall(json.decode, body)
+    if not ok or type(resp) ~= "table" then return nil end
+    local e = resp.error
+    if type(e) ~= "table" then return nil end
+    if status == 429 and type(e.code) == "string"
+        and e.code == "insufficient_quota" then
+        return "workspace quota exhausted (resets periodically)"
+    end
+    if type(e.message) == "string" then return e.message end
+    return nil
+end
+
+return adapter

internal/lua/adapters/trae.lua

@@ -0,0 +1,286 @@
+-- trae 源适配器（trae-local-api 的 OpenAI 兼容代理）
+-- trae-local-api 运行在 http://127.0.0.1:19900，接 Trae CN 账号的云资源。
+-- 协议：OpenAI /v1/chat/completions
+-- 特性：
+--   - 排队：上游 busy 时排队，可能长时间无响应
+--   - 非流式请求偶发返回 SSE 数据（upstream bug）
+--   - 支持 thinking，由上游模型自动决定是否开启
+
+local adapter = {}
+adapter.name = "trae"
+adapter.version = "1.0.0"
+adapter.endpoint = "/v1/chat/completions"
+adapter.headers = {}
+
+function adapter.transform_request(raw_body)
+    local ok, req = pcall(json.decode, raw_body)
+    if not ok then return raw_body end
+    req.disable_thinking = nil
+    req.extra_body = nil
+    if req.messages then
+        for _, msg in ipairs(req.messages) do
+            msg.reasoning_content = nil
+        end
+    end
+    return json.encode(req)
+end
+
+-- parse_text_tool_calls extracts tool calls that an upstream emitted as PLAIN
+-- TEXT instead of using the OpenAI tool_calls field.
+--
+-- trae-local-api's OpenAI endpoint (/v1/chat/completions) does not read the
+-- request's `tools` array at all, so the relayed model is never told the tool
+-- schema; it falls back to printing
+--     <tool_call>
+--     {"name": "get_weather", "arguments": {"location": "北京"}}
+--     </tool_call>
+-- into message.content, leaves message.tool_calls null, and reports
+-- finish_reason="stop". A client following the OpenAI contract therefore never
+-- sees a tool call: the agent loop terminates unexpectedly mid-conversation
+-- (and a hand-written replay produces an empty function name next turn).
+--
+-- Both tag spellings are accepted: the same codebase's Anthropic endpoint
+-- instructs models to emit <toolcall>, and models mix the two. Key names vary
+-- too (arguments / params / input), so all are tried.
+--
+-- Returns (tool_calls_array_or_nil, content_with_blocks_removed).
+local function parse_text_tool_calls(content)
+    if type(content) ~= "string" or content == "" then return nil, content end
+
+    local tcs = {}
+    local idx = 0
+
+    local function collect(pattern)
+        for payload in content:gmatch(pattern) do
+            local ok, obj = pcall(json.decode, payload)
+            if ok and type(obj) == "table" then
+                -- some models wrap it as {"function":{"name":..,"arguments":..}}
+                local fn = obj["function"]
+                local name = obj.name or (type(fn) == "table" and fn.name) or nil
+                if name then
+                    local args = obj.arguments or obj.params or obj.input
+                    if args == nil and type(fn) == "table" then
+                        args = fn.arguments or fn.params
+                    end
+                    if type(args) == "string" then
+                        local aok, decoded = pcall(json.decode, args)
+                        args = aok and decoded or {}
+                    elseif type(args) ~= "table" then
+                        args = {}
+                    end
+                    idx = idx + 1
+                    table.insert(tcs, {
+                        id = obj.id or ("call_text_" .. idx),
+                        type = "function",
+                        name = name,
+                        arguments = args
+                    })
+                end
+            end
+        end
+    end
+
+    -- `<tag ...>` allows attributes; %s* handles the "</tool_call >" spacing
+    -- that trae-local-api's own prompt example uses.
+    collect("<tool_call[^>]*>%s*(.-)%s*</tool_call%s*>")
+    collect("<toolcall[^>]*>%s*(.-)%s*</toolcall%s*>")
+
+    -- Legacy / mimicked form: models that saw past assistant turns folded as
+    --   [Called tool: name({...})]
+    -- tend to emit the same shape instead of a real  block. Recover it so
+    -- the agent loop keeps working; otherwise the client receives plain text
+    -- where a structured tool call should be. Capture the name and the JSON
+    -- object literal independently (the JSON is the only {...} run here).
+    if #tcs == 0 and content:find("%[Called tool") then
+        for name, argsraw in content:gmatch("%[Called tool%s*:%s*([%w_%-%.]+)%s*%((%b{})%s*%)%s*%]") do
+            local aok, decoded = pcall(json.decode, argsraw)
+            idx = idx + 1
+            table.insert(tcs, {
+                id = "call_legacy_" .. idx,
+                type = "function",
+                name = name,
+                arguments = aok and decoded or {}
+            })
+        end
+    end
+
+    if #tcs == 0 then return nil, content end
+
+    -- drop the blocks from user-visible content; keep any surrounding prose
+    local stripped = content:gsub("<tool_call[^>]*>%s*.-%s*</tool_call%s*>", "")
+    stripped = stripped:gsub("<toolcall[^>]*>%s*.-%s*</toolcall%s*>", "")
+    stripped = stripped:gsub("%[Called tool%s*:%s*[%w_%-%.]+%s*%(%b{}%s*%)%s*%]", "")
+    stripped = stripped:gsub("^%s+", ""):gsub("%s+$", "")
+    return tcs, stripped
+end
+
+function adapter.transform_response(raw_body)
+    -- trae-local-api 偶发在非流式请求中返回 SSE 格式数据，
+    -- 表现为多个 data: {...} 行或混合了 reasoning_chunk 等。
+    -- 剥掉 data: 前缀，取最后一个完整的 JSON 块（通常是
+    -- 最终的 stop chunk 或 usage chunk）。
+    local use_json = raw_body
+    if raw_body:match("^data:") or raw_body:match("\ndata:") then
+        -- 取最后一个 data: 行的 JSON
+        local last = nil
+        for line in raw_body:gmatch("data: ([^\n\r]+)") do
+            local trimmed = line:match("^%s*(.-)%s*$")
+            if trimmed and trimmed ~= "[DONE]" then
+                local ok2, obj = pcall(json.decode, trimmed)
+                if ok2 and type(obj) == "table" then
+                    last = obj
+                end
+            end
+        end
+        if last then
+            use_json = json.encode(last)
+        end
+        -- 如果解析失败，回退到原始 body
+    end
+
+    local ok, resp = pcall(json.decode, use_json)
+    if not ok or resp == nil then return raw_body end
+
+    local unified = {
+        content = "",
+        finish_reason = "",
+        token_usage = { prompt = 0, completion = 0, total = 0 }
+    }
+
+    if type(resp.usage) == "table" then
+        unified.token_usage.prompt = resp.usage.prompt_tokens or 0
+        unified.token_usage.completion = resp.usage.completion_tokens or 0
+        unified.token_usage.total = resp.usage.total_tokens or 0
+    end
+
+    if type(resp.choices) == "table" and #resp.choices > 0 then
+        local ch = resp.choices[1]
+        if type(ch.message) == "table" then
+            unified.content = ch.message.content or ""
+            if ch.message.reasoning_content then
+                unified.reasoning_content = ch.message.reasoning_content
+            end
+            if type(ch.message.tool_calls) == "table" and #ch.message.tool_calls > 0 then
+                local tcs = {}
+                for _, tc in ipairs(ch.message.tool_calls) do
+                    local fn = tc["function"] or {}
+                    local args = fn.arguments
+                    if type(args) == "string" then
+                        local args_ok, decoded = pcall(json.decode, args)
+                        args = args_ok and decoded or {}
+                    elseif type(args) ~= "table" then
+                        args = {}
+                    end
+                    table.insert(tcs, {
+                        id = tc.id,
+                        type = tc.type or "function",
+                        name = fn.name,
+                        arguments = args
+                    })
+                end
+                unified.tool_calls = tcs
+            end
+        end
+        unified.finish_reason = ch.finish_reason or ""
+
+        -- Fallback: trae-local-api's OpenAI endpoint drops the request's
+        -- `tools` array entirely, so the relayed model is never told the tool
+        -- schema and instead PRINTS a <tool_call>{...}</tool_call> block into
+        -- content, leaving message.tool_calls null and finish_reason="stop".
+        -- A client following the OpenAI contract then sees a normal completion
+        -- and its agent loop terminates mid-conversation. Recover the
+        -- structured call so the loop can continue.
+        if unified.tool_calls == nil then
+            local recovered, cleaned = parse_text_tool_calls(unified.content)
+            if recovered then
+                unified.tool_calls = recovered
+                unified.content = cleaned
+                unified.finish_reason = "tool_calls"
+            end
+        end
+    end
+
+    return json.encode(unified)
+end
+
+function adapter.transform_stream_chunk(raw_chunk)
+    local ok, chunk = pcall(json.decode, raw_chunk)
+    if not ok then return "" end
+
+    local uses = nil
+    if type(chunk.usage) == "table" then
+        uses = {
+            prompt = chunk.usage.prompt_tokens or chunk.usage.prompt or 0,
+            completion = chunk.usage.completion_tokens or chunk.usage.completion or 0,
+            total = chunk.usage.total_tokens or chunk.usage.total or 0,
+        }
+        if type(chunk.usage.prompt_tokens_details) == "table" and chunk.usage.prompt_tokens_details.cached_tokens ~= nil then
+            uses.prompt_tokens_details = { cached_tokens = chunk.usage.prompt_tokens_details.cached_tokens }
+        elseif (chunk.usage.prompt_cache_hit_tokens or 0) > 0 then
+            uses.prompt_cache_hit_tokens = chunk.usage.prompt_cache_hit_tokens
+            uses.prompt_cache_miss_tokens = chunk.usage.prompt_cache_miss_tokens or 0
+            uses.prompt_tokens_details = { cached_tokens = chunk.usage.prompt_cache_hit_tokens }
+        end
+    end
+
+    if not chunk.choices or #chunk.choices == 0 then
+        if uses ~= nil then
+            return json.encode({ usage = uses, done = false })
+        end
+        return ""
+    end
+    local delta = chunk.choices[1].delta or {}
+    local fr = chunk.choices[1].finish_reason
+
+    local finish = (type(fr) == "string" and fr ~= "") and fr or nil
+
+    local unified = {
+        content = delta.content or "",
+        done = (finish ~= nil)
+    }
+    if finish then
+        unified.finish_reason = finish
+    end
+    if uses ~= nil then
+        unified.usage = uses
+    end
+    if delta.reasoning_content then
+        unified.reasoning_content = delta.reasoning_content
+    end
+    if delta.tool_calls then
+        unified.tool_calls = delta.tool_calls
+    end
+    return json.encode(unified)
+end
+
+function adapter.transform_error(status, body)
+    local ok, resp = pcall(json.decode, body)
+    if not ok or type(resp) ~= "table" then
+        -- Non-JSON body: extract title from HTML or first line
+        local title = string.match(body or "", "<title>(.-)</title>")
+        if title and title ~= "" then return title end
+        local line = string.match(body or "", "^%s*([^\r\n]+)")
+        if line and line ~= "" and not string.match(line, "^<") then
+            return string.sub(line, 1, 200)
+        end
+        return nil
+    end
+    local e = resp.error
+    if type(e) == "table" and type(e.message) == "string" then
+        -- trae quota / rate limit annotations
+        if e.code then
+            return tostring(e.code) .. ": " .. e.message
+        end
+        return e.message
+    end
+    if type(e) == "string" then return e end
+    if type(resp.message) == "string" and resp.message ~= "" then
+        if resp.code ~= nil then
+            return tostring(resp.code) .. ": " .. resp.message
+        end
+        return resp.message
+    end
+    return nil
+end
+
+return adapter

internal/lua/toolcall_preservation_test.go

@@ -0,0 +1,353 @@
+package lua
+
+import (
+	"encoding/json"
+	"strings"
+	"testing"
+)
+
+// TestAdaptersPreserveToolCalls is the fleet-wide guard for the
+// "orphaned tool result" class of bug: when an agent client replays a turn
+// whose assistant message carries tool_calls, EVERY adapter must forward both
+// the call and the attribution of its result. Dropping the call (or the
+// tool_call_id / tool_name that ties the result to it) makes the model see a
+// result for a call it never made and re-issue the same call forever.
+//
+// The input deliberately uses content:[] — the shape most agent clients emit
+// for a tool-calling assistant turn with no text.
+func TestAdaptersPreserveToolCalls(t *testing.T) {
+	vm := NewVM(freshAdapterDir(t))
+	if err := vm.Start(); err != nil {
+		t.Fatal(err)
+	}
+	defer vm.Stop()
+
+	body := `{"model":"m","messages":[
+		{"role":"user","content":"read /tmp/x"},
+		{"role":"assistant","content":[],"tool_calls":[{"id":"call_abc","type":"function","function":{"name":"read","arguments":"{\"path\":\"/tmp/x\"}"}}]},
+		{"role":"tool","tool_call_id":"call_abc","content":"hello world"},
+		{"role":"user","content":"summarize"}
+	]}`
+	const (
+		callID   = "call_abc"
+		fnName   = "read"
+		toolOut  = "hello world"
+		lastUser = "summarize"
+	)
+
+	// Adapters that identify a tool result by the call id.
+	byID := []string{"openai", "deepseek", "github", "groq", "kimicode", "mistral",
+		"sensenova", "trae", "agentrouter", "opencode", "anthropic"}
+
+	for _, a := range byID {
+		out, err := vm.Transform(a, "transform_request", body)
+		if err != nil {
+			t.Errorf("%s: transform_request: %v", a, err)
+			continue
+		}
+		if !strings.Contains(out, callID) {
+			t.Errorf("%s dropped the tool call id %q: %s", a, callID, out)
+		}
+		if !strings.Contains(out, fnName) {
+			t.Errorf("%s dropped the function name %q: %s", a, fnName, out)
+		}
+		if !strings.Contains(out, toolOut) {
+			t.Errorf("%s dropped the tool result: %s", a, out)
+		}
+		if !strings.Contains(out, lastUser) {
+			t.Errorf("%s dropped the final user turn: %s", a, out)
+		}
+	}
+
+	// Ollama identifies the result with tool_name (its API has no tool_call_id
+	// on assistant turns) and takes arguments as an object, not a JSON string.
+	out, err := vm.Transform("ollama", "transform_request", body)
+	if err != nil {
+		t.Fatalf("ollama: %v", err)
+	}
+	for _, want := range []string{`"tool_calls"`, fnName, `"tool_name"`, toolOut} {
+		if !strings.Contains(out, want) {
+			t.Errorf("ollama missing %s: %s", want, out)
+		}
+	}
+
+	// Gemini identifies the result by function NAME and has no call id in its
+	// wire format; both directions must be present as functionCall /
+	// functionResponse parts, and the system role must move to
+	// systemInstruction rather than sitting in contents.
+	gBody := `{"model":"m","messages":[
+		{"role":"system","content":"be brief"},
+		{"role":"user","content":"read /tmp/x"},
+		{"role":"assistant","content":[],"tool_calls":[{"id":"call_abc","type":"function","function":{"name":"read","arguments":"{\"path\":\"/tmp/x\"}"}}]},
+		{"role":"tool","tool_call_id":"call_abc","content":"hello world"}
+	]}`
+	gout, err := vm.Transform("gemini", "transform_request", gBody)
+	if err != nil {
+		t.Fatalf("gemini: %v", err)
+	}
+	for _, want := range []string{`"functionCall"`, `"functionResponse"`, fnName, toolOut, `"systemInstruction"`} {
+		if !strings.Contains(gout, want) {
+			t.Errorf("gemini missing %s: %s", want, gout)
+		}
+	}
+	// No message may keep the OpenAI-only roles inside contents.
+	var greq struct {
+		Contents []struct {
+			Role string `json:"role"`
+		} `json:"contents"`
+	}
+	if err := json.Unmarshal([]byte(gout), &greq); err != nil {
+		t.Fatalf("gemini unmarshal: %v (%s)", err, gout)
+	}
+	for _, c := range greq.Contents {
+		if c.Role != "user" && c.Role != "model" {
+			t.Errorf("gemini contents carry an invalid role %q: %s", c.Role, gout)
+		}
+	}
+
+	// Negative control: a text-only turn must still be forwarded verbatim.
+	plain := `{"model":"m","messages":[{"role":"user","content":"just text"}]}`
+	for _, a := range append(byID, "ollama", "gemini") {
+		pout, err := vm.Transform(a, "transform_request", plain)
+		if err != nil {
+			t.Errorf("%s plain: %v", a, err)
+			continue
+		}
+		if !strings.Contains(pout, "just text") {
+			t.Errorf("%s dropped plain text: %s", a, pout)
+		}
+	}
+}
+
+// TestOpenCodeStreamOptionsOnlyWhenStreaming pins the fix for
+// "stream_options should be set along with stream": OpenCode Go (and other
+// strict OpenAI-compatible upstreams) reject a non-streaming request that
+// carries stream_options, which broke every non-stream call through the
+// opencode adapter.
+func TestOpenCodeStreamOptionsOnlyWhenStreaming(t *testing.T) {
+	vm := NewVM(freshAdapterDir(t))
+	if err := vm.Start(); err != nil {
+		t.Fatal(err)
+	}
+	defer vm.Stop()
+
+	// non-streaming: stream_options must be absent entirely
+	out, err := vm.Transform("opencode", "transform_request",
+		`{"model":"m","messages":[{"role":"user","content":"hi"}]}`)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if strings.Contains(out, "stream_options") {
+		t.Fatalf("non-streaming request must not carry stream_options: %s", out)
+	}
+
+	// streaming: stream_options.include_usage must be set
+	sout, err := vm.Transform("opencode", "transform_request",
+		`{"model":"m","stream":true,"messages":[{"role":"user","content":"hi"}]}`)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !strings.Contains(sout, "stream_options") || !strings.Contains(sout, "include_usage") {
+		t.Fatalf("streaming request must carry stream_options.include_usage: %s", sout)
+	}
+}
+
+// TestOpenCodeGoVsZenReasoning pins the one behavioural difference that made a
+// shared adapter wrong: OpenCode Go's thinking mode REQUIRES the assistant
+// turn's reasoning_content to be echoed back ("The `reasoning_content` in the
+// thinking mode must be passed back to the API"), while the Zen free pool must
+// not receive it. Hence two purpose-built adapters.
+func TestOpenCodeGoVsZenReasoning(t *testing.T) {
+	vm := NewVM(freshAdapterDir(t))
+	if err := vm.Start(); err != nil {
+		t.Fatal(err)
+	}
+	defer vm.Stop()
+
+	body := `{"model":"m","messages":[
+		{"role":"user","content":"hi"},
+		{"role":"assistant","content":"","reasoning_content":"I should read the file.","tool_calls":[{"id":"call_1","type":"function","function":{"name":"read","arguments":"{\"path\":\"/x\"}"}}]},
+		{"role":"tool","tool_call_id":"call_1","content":"r"}
+	]}`
+
+	// Go: reasoning_content must survive, and the tool call must stay paired.
+	gout, err := vm.Transform("opencodego", "transform_request", body)
+	if err != nil {
+		t.Fatalf("opencodego: %v", err)
+	}
+	if !strings.Contains(gout, "I should read the file.") {
+		t.Errorf("opencodego must pass reasoning_content back (thinking mode requires it): %s", gout)
+	}
+	if !strings.Contains(gout, "call_1") || !strings.Contains(gout, "tool_call_id") {
+		t.Errorf("opencodego must keep the tool call paired with its result: %s", gout)
+	}
+
+	// Zen: reasoning_content is stripped.
+	zout, err := vm.Transform("opencodezen", "transform_request", body)
+	if err != nil {
+		t.Fatalf("opencodezen: %v", err)
+	}
+	if strings.Contains(zout, "I should read the file.") {
+		t.Errorf("opencodezen must strip reasoning_content: %s", zout)
+	}
+	if !strings.Contains(zout, "call_1") {
+		t.Errorf("opencodezen must still keep the tool call: %s", zout)
+	}
+
+	// An assistant turn that carries tool_calls but no reasoning_content must
+	// get an empty one injected: Go validates that field on tool-calling turns
+	// and 400s when it is absent, while most clients never send it.
+	// Verified against the live endpoint that an empty string passes.
+	bare := `{"model":"m","messages":[
+		{"role":"user","content":"go"},
+		{"role":"assistant","content":"","tool_calls":[{"id":"c1","type":"function","function":{"name":"read","arguments":"{}"}}]},
+		{"role":"tool","tool_call_id":"c1","content":"r"}
+	]}`
+	bout, err := vm.Transform("opencodego", "transform_request", bare)
+	if err != nil {
+		t.Fatalf("opencodego bare: %v", err)
+	}
+	if !strings.Contains(bout, `"reasoning_content":""`) {
+		t.Errorf("opencodego must inject an empty reasoning_content on a tool-calling turn: %s", bout)
+	}
+	// A plain assistant turn (no tool calls) must NOT gain the field.
+	plainAsst := `{"model":"m","messages":[{"role":"user","content":"hi"},{"role":"assistant","content":"hello"}]}`
+	pout, err := vm.Transform("opencodego", "transform_request", plainAsst)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if strings.Contains(pout, "reasoning_content") {
+		t.Errorf("opencodego must not inject reasoning_content on a plain assistant turn: %s", pout)
+	}
+
+	// Both must still omit stream_options on non-streaming requests.
+	for _, a := range []string{"opencodego", "opencodezen"} {
+		out, err := vm.Transform(a, "transform_request", `{"model":"m","messages":[{"role":"user","content":"hi"}]}`)
+		if err != nil {
+			t.Fatalf("%s: %v", a, err)
+		}
+		if strings.Contains(out, "stream_options") {
+			t.Errorf("%s must not send stream_options when not streaming: %s", a, out)
+		}
+	}
+}
+
+// TestOpenCodeSessionIsStableForCache pins the session-derivation contract that
+// makes the upstream prefix cache usable.
+//
+// Measured against the live OpenCode Go endpoint: the prefix cache is
+// SESSION-scoped. Replaying the same 6032-token prompt hits 5888 cached tokens
+// when the session id is fixed, and 0 when it changes per request. Deriving the
+// session from meta.timestamp (the previous behaviour) meant every request was a
+// new session, so the cache could never hit.
+//
+// The session must be stable WITHIN one conversation and differ BETWEEN
+// conversations. Since generic OpenAI clients send no session id at all
+// (verified by capturing real agent traffic: no user / session_id /
+// conversation_id / metadata in the body, no session header), the conversation
+// is fingerprinted by its FIRST user message — history is replayed every turn,
+// so that message is invariant for the life of the conversation.
+func TestOpenCodeSessionIsStableForCache(t *testing.T) {
+	for _, name := range []string{"opencodego", "opencodezen"} {
+		vm := NewVM(freshAdapterDir(t))
+		if err := vm.Start(); err != nil {
+			t.Fatal(err)
+		}
+		meta := func(ts int, body string) map[string]interface{} {
+			return map[string]interface{}{
+				"timestamp": ts,
+				"body":      body,
+				"source":    map[string]interface{}{"name": "somesource", "meta": nil},
+			}
+		}
+
+		// Same conversation, later turn: history grew, first user msg unchanged.
+		h1, err := vm.BuildHeaders(name, meta(1000,
+			`{"messages":[{"role":"user","content":"fix the bug"},{"role":"assistant","content":"looking"}]}`))
+		if err != nil {
+			t.Fatalf("%s: %v", name, err)
+		}
+		h2, err := vm.BuildHeaders(name, meta(2000,
+			`{"messages":[{"role":"user","content":"fix the bug"},{"role":"assistant","content":"looking"},{"role":"user","content":"and tests"}]}`))
+		if err != nil {
+			t.Fatalf("%s: %v", name, err)
+		}
+		// Different conversation.
+		h3, err := vm.BuildHeaders(name, meta(3000,
+			`{"messages":[{"role":"user","content":"an unrelated task"}]}`))
+		if err != nil {
+			t.Fatalf("%s: %v", name, err)
+		}
+
+		if h1["x-opencode-session"] == "" {
+			t.Fatalf("%s: session header must be set", name)
+		}
+		if h1["x-opencode-session"] != h2["x-opencode-session"] {
+			t.Errorf("%s: x-opencode-session must be STABLE across turns of one conversation "+
+				"(%q vs %q) — the prefix cache is session-scoped", name, h1["x-opencode-session"], h2["x-opencode-session"])
+		}
+		if h1["x-opencode-session"] == h3["x-opencode-session"] {
+			t.Errorf("%s: a different conversation must get a different session", name)
+		}
+		if h1["x-opencode-request"] == h2["x-opencode-request"] {
+			t.Errorf("%s: x-opencode-request must differ per request", name)
+		}
+		// No user message at all falls back to per-source stability.
+		h4, _ := vm.BuildHeaders(name, meta(4000, `{"messages":[{"role":"system","content":"s"}]}`))
+		h5, _ := vm.BuildHeaders(name, meta(5000, `{"messages":[{"role":"system","content":"s"}]}`))
+		if h4["x-opencode-session"] != h5["x-opencode-session"] {
+			t.Errorf("%s: sessionless requests must still be stable per source", name)
+		}
+		vm.Stop()
+	}
+}
+
+// 客户端自带的会话标识（pi 的 x-session-affinity）必须优先于会话指纹：
+// 它是平台真实会话 id，历史压缩后也不会漂移，且天然按会话隔离。
+func TestOpenCodePrefersClientSessionID(t *testing.T) {
+	for _, name := range []string{"opencodego", "opencodezen"} {
+		vm := NewVM(freshAdapterDir(t))
+		if err := vm.Start(); err != nil {
+			t.Fatal(err)
+		}
+		meta := func(body, cs string) map[string]interface{} {
+			return map[string]interface{}{
+				"timestamp":      1000,
+				"body":           body,
+				"client_session": cs,
+				"source":         map[string]interface{}{"name": "somesource", "meta": nil},
+			}
+		}
+
+		// 同一条会话：历史在长，但客户端会话 id 不变。
+		a, err := vm.BuildHeaders(name, meta(`{"messages":[{"role":"user","content":"fix the bug"}]}`, "cs-aaa"))
+		if err != nil {
+			t.Fatal(err)
+		}
+		b, err := vm.BuildHeaders(name, meta(`{"messages":[{"role":"user","content":"a totally different opener"}]}`, "cs-aaa"))
+		if err != nil {
+			t.Fatal(err)
+		}
+		if a["x-opencode-session"] != b["x-opencode-session"] {
+			t.Errorf("%s: 客户端会话 id 相同就必须同会话（内容无关）: %q vs %q",
+				name, a["x-opencode-session"], b["x-opencode-session"])
+		}
+
+		// 不同会话。
+		c, _ := vm.BuildHeaders(name, meta(`{"messages":[{"role":"user","content":"fix the bug"}]}`, "cs-bbb"))
+		if a["x-opencode-session"] == c["x-opencode-session"] {
+			t.Errorf("%s: 不同客户端会话必须不同会号", name)
+		}
+
+		// 客户端没给会话 id 时，回落到指纹（同内容 -> 同会号）。
+		d, _ := vm.BuildHeaders(name, meta(`{"messages":[{"role":"user","content":"fix the bug"}]}`, ""))
+		e, _ := vm.BuildHeaders(name, meta(`{"messages":[{"role":"user","content":"fix the bug"}]}`, ""))
+		if d["x-opencode-session"] != e["x-opencode-session"] {
+			t.Errorf("%s: 无客户端会号时应回落到稳定指纹", name)
+		}
+		if d["x-opencode-session"] == a["x-opencode-session"] {
+			t.Errorf("%s: 客户端会号与指纹两条路必须产出不同会号", name)
+		}
+		vm.Stop()
+	}
+}

internal/lua/usage_cost_test.go

@@ -0,0 +1,177 @@
+package lua
+
+import (
+	"encoding/json"
+	"testing"
+)
+
+// OpenCode Zen/Go 在 usage 里报告输出有多少是思考 token，并在**顶层**用一个
+// **字符串**给出本次费用。前者决定「completion_tokens 里多少是可见回答」，
+// 后者是唯一的费用信号。两者都曾被适配器丢掉。
+func TestOpenCodeCostAndReasoningPassthrough(t *testing.T) {
+	for _, name := range []string{"opencodego", "opencodezen"} {
+		vm := NewVM(freshAdapterDir(t))
+		if err := vm.Start(); err != nil {
+			t.Fatal(err)
+		}
+
+		// ---- 非流式 ----
+		resp := `{"id":"x","object":"chat.completion","model":"deepseek-v4.1-flash",
+			"choices":[{"index":0,"message":{"role":"assistant","content":"hi","reasoning_content":"think"},
+			"finish_reason":"stop"}],
+			"usage":{"prompt_tokens":37,"completion_tokens":40,"total_tokens":77,
+			"prompt_cache_hit_tokens":0,"prompt_cache_miss_tokens":37,
+			"prompt_tokens_details":{"cached_tokens":0},
+			"completion_tokens_details":{"reasoning_tokens":40}},
+			"cost":"0"}`
+		out, err := vm.Transform(name, "transform_response", resp)
+		if err != nil {
+			t.Fatalf("%s transform_response: %v", name, err)
+		}
+		var r struct {
+			TokenUsage struct {
+				Completion int `json:"completion"`
+				Details    *struct {
+					ReasoningTokens int `json:"reasoning_tokens"`
+				} `json:"completion_tokens_details"`
+			} `json:"token_usage"`
+			Cost string `json:"cost"`
+		}
+		if err := json.Unmarshal([]byte(out), &r); err != nil {
+			t.Fatalf("%s unmarshal: %v (%s)", name, err, out)
+		}
+		if r.Cost != "0" {
+			t.Errorf("%s: cost 未透传（上游给的是字符串 \"0\"）: %q", name, r.Cost)
+		}
+		if r.TokenUsage.Details == nil || r.TokenUsage.Details.ReasoningTokens != 40 {
+			t.Errorf("%s: completion_tokens_details.reasoning_tokens 未透传: %s", name, out)
+		}
+		// 前提校验：reasoning_tokens 确实 ≤ completion_tokens（全为思考）。
+		if r.TokenUsage.Details != nil && r.TokenUsage.Details.ReasoningTokens > r.TokenUsage.Completion {
+			t.Errorf("%s: reasoning 不应超过 completion", name)
+		}
+
+		// ---- 流式：非零费用也要透传 ----
+		chunk := `{"id":"x","object":"chat.completion.chunk","choices":[],
+			"usage":{"prompt_tokens":10,"completion_tokens":5,"total_tokens":15,
+			"completion_tokens_details":{"reasoning_tokens":4}},"cost":"0.0012"}`
+		co, err := vm.Transform(name, "transform_stream_chunk", chunk)
+		if err != nil {
+			t.Fatalf("%s transform_stream_chunk: %v", name, err)
+		}
+		var c struct {
+			Cost  string `json:"cost"`
+			Usage *struct {
+				Details *struct {
+					ReasoningTokens int `json:"reasoning_tokens"`
+				} `json:"completion_tokens_details"`
+			} `json:"usage"`
+		}
+		if err := json.Unmarshal([]byte(co), &c); err != nil {
+			t.Fatalf("%s unmarshal chunk: %v (%s)", name, err, co)
+		}
+		if c.Usage == nil || c.Usage.Details == nil || c.Usage.Details.ReasoningTokens != 4 {
+			t.Errorf("%s: 流式 reasoning_tokens 未透传: %s", name, co)
+		}
+		if c.Cost != "0.0012" {
+			t.Errorf("%s: 流式 cost 未透传: %s", name, co)
+		}
+
+		// ---- 独立的 cost 帧（上游就这么发：choices 为空、只有 cost）----
+		only := `{"choices":[],"cost":"0"}`
+		oo, err := vm.Transform(name, "transform_stream_chunk", only)
+		if err != nil {
+			t.Fatalf("%s cost-only chunk: %v", name, err)
+		}
+		if oo == "" {
+			t.Fatalf("%s: 纯 cost 帧被整个丢弃（客户端将看不到费用）", name)
+		}
+		var oc struct {
+			Cost string `json:"cost"`
+		}
+		if err := json.Unmarshal([]byte(oo), &oc); err != nil {
+			t.Fatalf("%s unmarshal cost chunk: %v (%s)", name, err, oo)
+		}
+		if oc.Cost != "0" {
+			t.Errorf("%s: 纯 cost 帧内容丢失: %s", name, oo)
+		}
+
+		// ---- 回归：没有 cost / 没有 details 时不得凭空造字段 ----
+		bare := `{"choices":[{"index":0,"delta":{"content":"hi"},"finish_reason":"stop"}],
+			"usage":{"prompt_tokens":3,"completion_tokens":1,"total_tokens":4}}`
+		bo, _ := vm.Transform(name, "transform_response", bare)
+		if got := bo; len(got) > 0 && containsAll(got, `"cost"`) {
+			t.Errorf("%s: 上游没给 cost 却透传了该字段: %s", name, got)
+		}
+
+		vm.Stop()
+	}
+}
+
+func containsAll(s string, subs ...string) bool {
+	for _, sub := range subs {
+		found := false
+		for i := 0; i+len(sub) <= len(s); i++ {
+			if s[i:i+len(sub)] == sub {
+				found = true
+				break
+			}
+		}
+		if !found {
+			return false
+		}
+	}
+	return true
+}
+
+// 流式与非流式对同一上游切片的产出必须一致：命中缓存时上游同时给
+// prompt_tokens_details.cached_tokens 和独立的 hit/miss 字段，两边都要透传。
+func TestOpenCodeStreamCacheFieldsMatchNonStream(t *testing.T) {
+	for _, name := range []string{"opencodego", "opencodezen"} {
+		vm := NewVM(freshAdapterDir(t))
+		if err := vm.Start(); err != nil {
+			t.Fatal(err)
+		}
+		type cache struct {
+			Details *struct {
+				CachedTokens int `json:"cached_tokens"`
+			} `json:"prompt_tokens_details"`
+			Hit  int `json:"prompt_cache_hit_tokens"`
+			Miss int `json:"prompt_cache_miss_tokens"`
+		}
+		usage := `"usage":{"prompt_tokens":5732,"completion_tokens":20,"total_tokens":5752,
+			"prompt_cache_hit_tokens":5504,"prompt_cache_miss_tokens":228,
+			"prompt_tokens_details":{"cached_tokens":5504}}`
+
+		// 非流式
+		nout, _ := vm.Transform(name, "transform_response",
+			`{"choices":[{"index":0,"message":{"role":"assistant","content":"hi"},"finish_reason":"stop"}],`+usage+`}`)
+		var nr struct {
+			U cache `json:"token_usage"`
+		}
+		if err := json.Unmarshal([]byte(nout), &nr); err != nil {
+			t.Fatalf("%s: %v", name, err)
+		}
+		// 流式
+		sout, _ := vm.Transform(name, "transform_stream_chunk",
+			`{"choices":[],`+usage+`}`)
+		var sr struct {
+			U cache `json:"usage"`
+		}
+		if err := json.Unmarshal([]byte(sout), &sr); err != nil {
+			t.Fatalf("%s chunk: %v (%s)", name, err, sout)
+		}
+
+		if nr.U.Hit != 5504 || nr.U.Miss != 228 {
+			t.Errorf("%s: 非流式 hit/miss 异常: %+v", name, nr.U)
+		}
+		if sr.U.Hit != nr.U.Hit || sr.U.Miss != nr.U.Miss {
+			t.Errorf("%s: 流式 hit/miss 与非流式不一致 (stream %d/%d vs non-stream %d/%d): %s",
+				name, sr.U.Hit, sr.U.Miss, nr.U.Hit, nr.U.Miss, sout)
+		}
+		if sr.U.Details == nil || sr.U.Details.CachedTokens != 5504 {
+			t.Errorf("%s: 流式 cached_tokens 丢失: %s", name, sout)
+		}
+		vm.Stop()
+	}
+}

internal/lua/vm.go

@@ -4,9 +4,15 @@
 //
 // The runtime uses LuaJIT (github.com/aarzilli/golua) so every adapter runs on
 // an independent interpreter state. Because a lua.State is NOT goroutine-safe,
-// each adapter owns a pool of workers (states); the pool size equals the sum
-// of max_concurrent over all sources that use the adapter, and a worker is
-// checked out for the duration of a single hook call.
+// each adapter owns a pool of workers (states) and a worker is checked out for
+// the duration of a single hook call.
+//
+// Pools are ELASTIC: the sum of max_concurrent over the sources using an adapter
+// is a ceiling, not a preallocation. States are booted on demand in batches
+// sized by that ceiling, and reclaimed by a single background janitor at a rate
+// scaled inversely to the adapter's live connection count. An idle gateway
+// therefore holds close to zero Lua states regardless of how much aggregate
+// concurrency is configured. See adapterPool for the sizing rules.
 package lua
 
 import (
@@ -23,6 +29,7 @@ import (
 	"strconv"
 	"strings"
 	"sync"
+	"time"
 
 	// golua exposes the LuaJIT C API. Build with -tags luajit so the cgo
 	// LDFLAGS resolve to -lluajit-5.1 (see golua's lua.go).
@@ -64,7 +71,20 @@ type staticInfo struct {
 }
 
 // adapterPool owns the worker states for one adapter. Idle workers are held in
-// a slice guarded by a cond; up to target workers are created lazily.
+// a slice guarded by a cond; workers are created lazily on demand and reclaimed
+// by the VM janitor once demand drops.
+//
+// Sizing is elastic rather than fixed (plan 阶段 2). Three inputs drive it:
+//
+//   - maxW — the adapter's maximum concurrency (Σ max_concurrent of every source
+//     using it). It is the hard ceiling AND sets the growth step, so a hot,
+//     high-concurrency adapter ramps in batches while a low-concurrency one
+//     creeps up one state at a time.
+//   - created / len(idle) — how many states already exist, i.e. how much room is
+//     left to grow and how much is available to reclaim.
+//   - inUse — how many states are checked out right now, i.e. the live
+//     connection count for this adapter. It sets the shrink step: the fewer
+//     connections, the more aggressively idle states are released.
 type adapterPool struct {
 	name   string
 	script string
@@ -73,24 +93,164 @@ type adapterPool struct {
 	mu      sync.Mutex
 	cond    *sync.Cond
 	idle    []*worker
-	created int
-	target  int
+	created int // live states: len(idle) + inUse (+ states being booted)
+	inUse   int // states checked out right now == live connections
+	waiting int // goroutines blocked waiting for a free state
+	maxW    int // hard ceiling = Σ max_concurrent of the sources using this adapter
+	used    bool
 	closed  bool
+
+	lastGrow   time.Time
+	idleRounds int // consecutive janitor rounds that saw reclaimable slack
+	peakInUse  int // high-water mark of inUse, for observability
+}
+
+const (
+	// growDivisor turns the adapter's max concurrency into the growth step CAP:
+	// step <= clamp(ceil(maxW / growDivisor), 1, growStepCap). The actual batch
+	// is additionally bounded by how many callers are queued (see growPlanLocked),
+	// so a wide adapter may ramp in big jumps but never warms states nobody is
+	// waiting for.
+	growDivisor = 8
+	growStepCap = 8
+	// growCooldown keeps a burst of misses from batching repeatedly while the
+	// previous batch is still booting.
+	growCooldown = time.Second
+	// residentWorkers is how many states an adapter keeps warm once it has
+	// served at least one request. Booting is milliseconds, but keeping one warm
+	// removes that from the critical path of the next request. Adapters that
+	// were never used keep nothing.
+	residentWorkers = 1
+	// idleHeadroom is the spare idle state kept above the live connection count,
+	// so the next concurrent request does not have to boot.
+	idleHeadroom = 1
+	// shrinkInterval is how often the VM janitor reclaims idle states.
+	shrinkInterval = 30 * time.Second
+	// shrinkGraceRounds is how many consecutive janitor rounds must see slack
+	// before anything is released. It guards against tearing a pool down between
+	// two bursts — NOT against ordinary traffic: a pool serving one request at a
+	// time still has reclaimable slack, and requiring "no traffic at all" would
+	// pin a burst's leftover states forever on any busy gateway. The counter is
+	// therefore reset by GROWTH (real contention), not by a mere checkout.
+	shrinkGraceRounds = 2
+)
+
+// PoolStats is a snapshot of one adapter pool's elastic sizing, surfaced by the
+// status API so the algorithm is observable instead of a black box.
+type PoolStats struct {
+	Name       string `json:"name"`
+	Created    int    `json:"created"`
+	Idle       int    `json:"idle"`
+	InUse      int    `json:"in_use"`
+	Waiting    int    `json:"waiting"`
+	Max        int    `json:"max"`
+	Resident   int    `json:"resident"`
+	GrowStep   int    `json:"grow_step"`
+	ShrinkStep int    `json:"shrink_step"`
+	PeakInUse  int    `json:"peak_in_use"`
 }
 
 func newAdapterPool(name, script string, static staticInfo) *adapterPool {
-	p := &adapterPool{name: name, script: script, static: static, target: 1}
+	p := &adapterPool{name: name, script: script, static: static, maxW: 1}
 	p.cond = sync.NewCond(&p.mu)
 	return p
 }
 
-func (p *adapterPool) setTarget(n int) {
+// setMax updates the pool's hard ceiling (Σ max_concurrent of its sources).
+// Lowering it does not kill live states: the janitor reclaims the excess.
+func (p *adapterPool) setMax(n int) {
+	if n < 1 {
+		n = 1
+	}
 	p.mu.Lock()
-	p.target = n
+	p.maxW = n
 	p.cond.Broadcast()
 	p.mu.Unlock()
 }
 
+func ceilDiv(a, b int) int {
+	if b <= 0 {
+		return a
+	}
+	return (a + b - 1) / b
+}
+
+// growStepLocked is how many states to warm on a miss, derived from the
+// adapter's maximum concurrency. Caller holds p.mu.
+func (p *adapterPool) growStepLocked() int {
+	step := ceilDiv(p.maxW, growDivisor)
+	if step < 1 {
+		return 1
+	}
+	if step > growStepCap {
+		return growStepCap
+	}
+	return step
+}
+
+// residentLocked is the floor the janitor will not shrink below. Caller holds p.mu.
+func (p *adapterPool) residentLocked() int {
+	if !p.used {
+		return 0
+	}
+	if residentWorkers > p.maxW {
+		return p.maxW
+	}
+	return residentWorkers
+}
+
+// shrinkStepLocked is how many idle states to release this round. The step is
+// inversely proportional to the live connection count: with no connections the
+// slack collapses in one round, while a busy adapter gives up one state at a
+// time so the hot path keeps its warm states. Caller holds p.mu.
+func (p *adapterPool) shrinkStepLocked() int {
+	excess := p.created - p.keepLocked()
+	if excess <= 0 {
+		return 0
+	}
+	step := ceilDiv(excess, 1+p.inUse)
+	if step > excess {
+		step = excess
+	}
+	if step > len(p.idle) {
+		step = len(p.idle)
+	}
+	if step < 0 {
+		return 0
+	}
+	return step
+}
+
+// keepLocked is the number of live states this pool should retain right now.
+// Caller holds p.mu.
+func (p *adapterPool) keepLocked() int {
+	keep := p.residentLocked()
+	if h := p.inUse + idleHeadroom; h > keep {
+		keep = h
+	}
+	if keep > p.maxW {
+		keep = p.maxW
+	}
+	return keep
+}
+
+func (p *adapterPool) stats() PoolStats {
+	p.mu.Lock()
+	defer p.mu.Unlock()
+	return PoolStats{
+		Name:       p.name,
+		Created:    p.created,
+		Idle:       len(p.idle),
+		InUse:      p.inUse,
+		Waiting:    p.waiting,
+		Max:        p.maxW,
+		Resident:   p.residentLocked(),
+		GrowStep:   p.growStepLocked(),
+		ShrinkStep: p.shrinkStepLocked(),
+		PeakInUse:  p.peakInUse,
+	}
+}
+
 // boot creates a fresh LuaJIT state, registers the shared globals and executes
 // the adapter script, storing the returned adapter table in the state.
 func (p *adapterPool) boot() (*worker, error) {
@@ -110,6 +270,96 @@ func (p *adapterPool) boot() (*worker, error) {
 	return &worker{L: L}, nil
 }
 
+// growPlanLocked reserves capacity for a batch prewarm and returns how many
+// EXTRA states to boot in the background (the caller already reserved one for
+// itself).
+//
+// Two independent limits apply, and BOTH matter:
+//
+//   - the adapter's max concurrency caps the step (growStepLocked), so a
+//     high-concurrency adapter is allowed to ramp in bigger jumps than a
+//     narrow one;
+//   - the number of goroutines actually BLOCKED waiting for a state bounds it
+//     to real demand.
+//
+// Sizing on the ceiling alone over-provisions badly: an adapter with
+// max_concurrent=76 has a step of 8, so two concurrent requests would warm 8
+// states, and the janitor would throw 7 of them away a minute later — boot,
+// discard, repeat, with RSS oscillating for no benefit. Prewarming only for
+// goroutines that are genuinely queued keeps the ramp cheap without the churn.
+//
+// Caller holds p.mu.
+func (p *adapterPool) growPlanLocked(contended bool) int {
+	if !contended || p.waiting <= 0 {
+		return 0
+	}
+	if time.Since(p.lastGrow) < growCooldown {
+		return 0
+	}
+	extra := p.growStepLocked() - 1
+	if extra > p.waiting {
+		extra = p.waiting // never warm more than the queue needs
+	}
+	if extra <= 0 {
+		return 0
+	}
+	if room := p.maxW - p.created; extra > room {
+		extra = room
+	}
+	if extra <= 0 {
+		return 0
+	}
+	// Reserve immediately so concurrent acquires respect the ceiling while the
+	// batch is still booting.
+	p.created += extra
+	p.lastGrow = time.Now()
+	return extra
+}
+
+// prewarm boots n reserved states and publishes them as idle. A boot failure
+// releases its reservation; it is not fatal, since the synchronous path
+// surfaces adapter errors already.
+func (p *adapterPool) prewarm(n int) {
+	for i := 0; i < n; i++ {
+		p.mu.Lock()
+		closed := p.closed
+		p.mu.Unlock()
+		if closed {
+			p.releaseReservation(n - i)
+			return
+		}
+		w, err := p.boot()
+		if err != nil {
+			p.releaseReservation(n - i)
+			return
+		}
+		p.mu.Lock()
+		if p.closed {
+			p.created--
+			p.mu.Unlock()
+			w.close()
+			p.releaseReservation(n - i - 1)
+			return
+		}
+		p.idle = append(p.idle, w)
+		p.cond.Signal()
+		p.mu.Unlock()
+	}
+}
+
+func (p *adapterPool) releaseReservation(n int) {
+	if n <= 0 {
+		return
+	}
+	p.mu.Lock()
+	p.created -= n
+	if p.created < 0 {
+		p.created = 0
+	}
+	p.cond.Broadcast()
+	p.mu.Unlock()
+}
+
 func (p *adapterPool) acquire() (*worker, error) {
 	p.mu.Lock()
 	for {
@@ -120,32 +370,68 @@ func (p *adapterPool) acquire() (*worker, error) {
 		if n := len(p.idle); n > 0 {
 			w := p.idle[n-1]
 			p.idle = p.idle[:n-1]
+			p.checkoutLocked()
 			p.mu.Unlock()
 			return w, nil
 		}
-		if p.created < p.target {
+		if p.created < p.maxW {
+			// Every existing state busy on a miss = real concurrency, not a
+			// sequential caller reusing one warm state.
+			contended := p.created > 0 && p.inUse >= p.created
 			p.created++
-			break
+			// Growing is the signal that capacity is genuinely short, so any
+			// pending shrink decision is stale.
+			p.idleRounds = 0
+			extra := p.growPlanLocked(contended)
+			p.mu.Unlock()
+			if extra > 0 {
+				go p.prewarm(extra)
+			}
+			w, err := p.boot()
+			if err != nil {
+				p.mu.Lock()
+				p.created--
+				p.cond.Signal()
+				p.mu.Unlock()
+				return nil, err
+			}
+			p.mu.Lock()
+			p.checkoutLocked()
+			p.mu.Unlock()
+			return w, nil
 		}
+		p.waiting++
 		p.cond.Wait()
+		p.waiting--
 	}
-	p.mu.Unlock()
+}
 
-	w, err := p.boot()
-	if err != nil {
-		p.mu.Lock()
-		p.created--
-		p.cond.Signal()
-		p.mu.Unlock()
-		return nil, err
+// checkoutLocked accounts one state as handed out. Caller holds p.mu.
+func (p *adapterPool) checkoutLocked() {
+	p.inUse++
+	p.used = true
+	if p.inUse > p.peakInUse {
+		p.peakInUse = p.inUse
 	}
-	return w, nil
 }
 
 func (p *adapterPool) release(w *worker) {
 	w.L.SetTop(0)
 	p.mu.Lock()
+	if p.inUse > 0 {
+		p.inUse--
+	}
 	if p.closed {
+		p.created--
+		p.mu.Unlock()
+		w.close()
+		return
+	}
+	// A lowered ceiling is honoured on the spot rather than waiting for the
+	// janitor: returning a state the pool may no longer keep closes it.
+	if p.created > p.maxW {
+		p.created--
+		p.cond.Signal()
 		p.mu.Unlock()
 		w.close()
 		return
@@ -155,11 +441,42 @@ func (p *adapterPool) release(w *worker) {
 	p.mu.Unlock()
 }
 
+// reclaim releases idle states down to the current keep target, using a step
+// scaled by the live connection count. It returns how many states were closed.
+func (p *adapterPool) reclaim() int {
+	p.mu.Lock()
+	if p.closed {
+		p.mu.Unlock()
+		return 0
+	}
+	step := p.shrinkStepLocked()
+	if step <= 0 {
+		p.idleRounds = 0
+		p.mu.Unlock()
+		return 0
+	}
+	p.idleRounds++
+	if p.idleRounds < shrinkGraceRounds {
+		p.mu.Unlock()
+		return 0
+	}
+	victims := append([]*worker(nil), p.idle[len(p.idle)-step:]...)
+	p.idle = p.idle[:len(p.idle)-step]
+	p.created -= step
+	p.idleRounds = 0
+	p.mu.Unlock()
+	for _, w := range victims {
+		w.close()
+	}
+	return step
+}
+
 func (p *adapterPool) shutdown() {
 	p.mu.Lock()
 	p.closed = true
 	idle := p.idle
 	p.idle = nil
+	p.created -= len(idle)
 	p.cond.Broadcast()
 	p.mu.Unlock()
 	for _, w := range idle {
@@ -173,13 +490,84 @@ type VM struct {
 	mu    sync.RWMutex
 	dir   string
 	pools map[string]*adapterPool
+
+	// janitor drives elastic shrinking. One goroutine serves every pool (not
+	// one per adapter), so an idle gateway costs a single sleeping goroutine.
+	janitorOnce sync.Once
+	janitorStop chan struct{}
+	janitorDone chan struct{}
 }
 
 func NewVM(dir string) *VM {
-	return &VM{dir: dir, pools: map[string]*adapterPool{}}
+	return &VM{
+		dir:         dir,
+		pools:       map[string]*adapterPool{},
+		janitorStop: make(chan struct{}),
+		janitorDone: make(chan struct{}),
+	}
+}
+
+// startJanitor launches the single shrink loop. Idempotent.
+func (v *VM) startJanitor() {
+	v.janitorOnce.Do(func() {
+		go v.janitorLoop(shrinkInterval)
+	})
+}
+
+func (v *VM) janitorLoop(every time.Duration) {
+	defer close(v.janitorDone)
+	t := time.NewTicker(every)
+	defer t.Stop()
+	for {
+		select {
+		case <-v.janitorStop:
+			return
+		case <-t.C:
+			v.reclaimIdle()
+		}
+	}
+}
+
+// reclaimIdle runs one shrink round over every pool. Exported behaviour is
+// tested through ReclaimIdleNow.
+func (v *VM) reclaimIdle() int {
+	v.mu.RLock()
+	pools := make([]*adapterPool, 0, len(v.pools))
+	for _, p := range v.pools {
+		pools = append(pools, p)
+	}
+	v.mu.RUnlock()
+	n := 0
+	for _, p := range pools {
+		n += p.reclaim()
+	}
+	return n
+}
+
+// ReclaimIdleNow forces one shrink round and reports how many worker states
+// were closed. Intended for tests and for an explicit "release idle memory"
+// action; normal operation relies on the janitor.
+func (v *VM) ReclaimIdleNow() int { return v.reclaimIdle() }
+
+// PoolStats returns the elastic sizing snapshot of every loaded adapter,
+// sorted by name.
+func (v *VM) PoolStats() []PoolStats {
+	v.mu.RLock()
+	pools := make([]*adapterPool, 0, len(v.pools))
+	for _, p := range v.pools {
+		pools = append(pools, p)
+	}
+	v.mu.RUnlock()
+	out := make([]PoolStats, 0, len(pools))
+	for _, p := range pools {
+		out = append(out, p.stats())
+	}
+	sort.Slice(out, func(i, j int) bool { return out[i].Name < out[j].Name })
+	return out
 }
 
 func (v *VM) Start() error {
+	v.startJanitor()
 	if v.dir == "" {
 		return nil
 	}
@@ -214,6 +602,12 @@ func (v *VM) Start() error {
 }
 
 func (v *VM) Stop() {
+	select {
+	case <-v.janitorStop:
+		// already stopped
+	default:
+		close(v.janitorStop)
+	}
 	v.mu.Lock()
 	pools := v.pools
 	v.pools = map[string]*adapterPool{}
@@ -223,17 +617,17 @@ func (v *VM) Stop() {
 	}
 }
 
-// ConfigureConcurrency sets each adapter's worker target to the sum of
-// max_concurrent of every source using it. Values below one are clamped to 1.
+// ConfigureConcurrency sets each adapter's worker CEILING to the sum of
+// max_concurrent of every source using it. It is a ceiling, not a preallocation:
+// states are booted on demand and reclaimed when demand drops, so a config with
+// a large aggregate concurrency no longer implies a large resident pool.
+// Values below one are clamped to 1.
 func (v *VM) ConfigureConcurrency(adapterConcurrency map[string]int) {
 	v.mu.RLock()
 	defer v.mu.RUnlock()
 	for name, n := range adapterConcurrency {
 		if p, ok := v.pools[name]; ok {
-			if n < 1 {
-				n = 1
-			}
-			p.setTarget(n)
+			p.setMax(n)
 		}
 	}
 }
@@ -260,16 +654,19 @@ func (v *VM) LoadAdapter(path string) error {
 }
 
 // LoadAdapterSource registers an adapter from source code. Replacing an
-// existing adapter shuts the old pool down and re-boots lazily.
+// existing adapter shuts the old pool down and re-boots lazily, carrying the
+// configured ceiling over so a hot-reload does not reset elastic sizing.
 func (v *VM) LoadAdapterSource(name, code string) error {
 	static, err := inspectScript(name, code)
 	if err != nil {
 		return err
 	}
-	target := 1
+	maxW := 1
 	v.mu.RLock()
 	if p, ok := v.pools[static.name]; ok {
-		target = p.target
+		p.mu.Lock()
+		maxW = p.maxW
+		p.mu.Unlock()
 	}
 	v.mu.RUnlock()
 
@@ -278,9 +675,10 @@ func (v *VM) LoadAdapterSource(name, code string) error {
 		p.shutdown()
 	}
 	p := newAdapterPool(static.name, code, *static)
-	p.setTarget(target)
+	p.setMax(maxW)
 	v.pools[static.name] = p
 	v.mu.Unlock()
+	v.startJanitor()
 	return nil
 }
 
@@ -407,6 +805,46 @@ func (v *VM) BuildHeaders(name string, meta map[string]interface{}) (map[string]
 	return headers, nil
 }
 
+// TransformError executes the optional adapter.transform_error(status, body)
+// hook: per-source condensing of an upstream error response into a short
+// client-facing reason. ok=false means the adapter defines no hook (or it
+// failed) — the caller falls back to the generic Go-side condenser.
+func (v *VM) TransformError(name string, status int, body string) (reason string, ok bool, err error) {
+	p := v.pool(name)
+	if p == nil {
+		return "", false, fmt.Errorf("adapter %s not loaded", name)
+	}
+	w, err := p.acquire()
+	if err != nil {
+		return "", false, err
+	}
+	defer p.release(w)
+	L := w.L
+
+	L.SetTop(0)
+	defer L.SetTop(0)
+	L.GetGlobal(adapterGlobal)
+	if L.IsNil(-1) {
+		return "", false, nil
+	}
+	L.GetField(-1, "transform_error")
+	if !L.IsFunction(-1) {
+		return "", false, nil
+	}
+	L.SetTop(0)
+	L.GetGlobal(adapterGlobal)
+	L.GetField(-1, "transform_error")
+	L.PushInteger(int64(status))
+	L.PushString(body)
+	if callErr := L.Call(2, 1); callErr != nil {
+		return "", false, fmt.Errorf("transform_error: %w", callErr)
+	}
+	if L.Type(-1) != golua.LUA_TSTRING {
+		return "", false, nil
+	}
+	return L.ToString(-1), true, nil
+}
+
 func (v *VM) Endpoint(name string) string {
 	p := v.pool(name)
 	if p == nil {

internal/provider/provider_test.go

@@ -7,7 +7,9 @@ import (
 	"fmt"
 	"net/http"
 	"net/http/httptest"
+	"os"
 	"path/filepath"
+	"strings"
 	"sync/atomic"
 	"testing"
 	"time"
@@ -196,8 +198,8 @@ func TestProviderAuthFailureSelfHeals(t *testing.T) {
 	if st.FailCount() != backoffCapN {
 		t.Fatalf("auth failure must jump to capped count, got %d", st.FailCount())
 	}
-	if until := st.CooldownUntil(); until <= time.Now().Add(25*time.Minute).Unix() {
-		t.Fatalf("auth failure must cool near the cap (until=%d)", until)
+	if until := st.CooldownUntil(); until <= time.Now().Add(authCooldown-time.Minute).Unix() {
+		t.Fatalf("auth failure must cool for authCooldown (until=%d)", until)
 	}
 	if st.Pref() != -2*int64(prefFailStep) {
 		t.Fatalf("auth failure pref penalty must be doubled, got %d", st.Pref())
@@ -240,8 +242,81 @@ func TestModelStateCooldownAndRecovery(t *testing.T) {
 	if st.FailCount() != 0 {
 		t.Fatalf("fail count after success = %d", st.FailCount())
 	}
-	if st.Pref() != 1-int64(prefFailStep) {
-		t.Fatalf("pref after one failure (-5) then success (+1) = %d, want %d", st.Pref(), 1-int64(prefFailStep))
+	// failure penalties prefFailStep, success rewards prefSuccessStep:
+	// -5 then +2 = -3
+	if st.Pref() != int64(prefSuccessStep)-int64(prefFailStep) {
+		t.Fatalf("pref after one failure then success = %d, want %d", st.Pref(), int64(prefSuccessStep)-int64(prefFailStep))
+	}
+}
+
+func TestModelStateRateLimitShortCooldown(t *testing.T) {
+	// 429 must NOT use the exponential backoff schedule: a rate-limited but
+	// quota-rich source has to re-enter rotation after the short fixed window.
+	p := newTestProvider(t, src("mock", "http://127.0.0.1:1", "openai", "m"))
+	st := p.state("m")
+
+	// repeated 429s must stay at the fixed short cooldown, never escalate
+	for i := 0; i < 15; i++ {
+		p.ReportStatus("m", 429)
+	}
+	until := st.CooldownUntil()
+	want := time.Now().Add(rateLimitCooldown).Unix()
+	if until < want-2 || until > want+2 {
+		t.Fatalf("429 cooldown = %d, want ~%d (fixed %v, not exponential)", until, want, rateLimitCooldown)
+	}
+	if st.FailCount() != 15 {
+		t.Fatalf("fail count = %d, want 15 (counted but not escalating)", st.FailCount())
+	}
+	if st.Pref() != -15*int64(prefQuotaStep) {
+		t.Fatalf("pref after 15 x 429 = %d, want %d (a 429 is a quota-class event, not a fault)", st.Pref(), -15*int64(prefQuotaStep))
+	}
+	if p.ModelAvailable("m") {
+		t.Fatal("model must be cooling right after a 429")
+	}
+}
+
+func TestThrottleSpacingAndCancel(t *testing.T) {
+	p := newTestProvider(t, src("mock", "http://127.0.0.1:1", "openai", "m"))
+	p.cfg.RPM = 120 // gap = 500ms
+	p.rpmGap = time.Minute / time.Duration(p.cfg.RPM)
+
+	start := time.Now()
+	for i := 0; i < 3; i++ {
+		if err := p.Throttle(context.Background()); err != nil {
+			t.Fatalf("throttle %d: %v", i, err)
+		}
+	}
+	elapsed := time.Since(start)
+	// first call passes immediately, the next two wait one gap each
+	want := 2 * time.Minute / time.Duration(p.cfg.RPM)
+	if elapsed < want-time.Duration(100*time.Millisecond) || elapsed > want+time.Second {
+		t.Fatalf("3 throttled calls took %v, want ~%v", elapsed, want)
+	}
+
+	// unlimited source: Throttle is a no-op
+	p2 := newTestProvider(t, src("free", "http://127.0.0.1:1", "openai", "m"))
+	if err := p2.Throttle(context.Background()); err != nil {
+		t.Fatalf("unlimited throttle: %v", err)
+	}
+
+	// cancelled context aborts a pending window reservation
+	p3 := newTestProvider(t, src("slow", "http://127.0.0.1:1", "openai", "m"))
+	p3.cfg.RPM = 6 // 10s gap
+	p3.rpmGap = time.Minute / time.Duration(p3.cfg.RPM)
+	// first call passes immediately (fresh provider) and reserves the next window
+	if err := p3.Throttle(context.Background()); err != nil {
+		t.Fatalf("first throttle: %v", err)
+	}
+	ctx, cancel := context.WithTimeout(context.Background(), 50*time.Millisecond)
+	defer cancel()
+	// second call must wait ~10s for the reserved window; the short deadline aborts it
+	start = time.Now()
+	err := p3.Throttle(ctx)
+	if !errors.Is(err, context.DeadlineExceeded) {
+		t.Fatalf("cancelled throttle = %v, want DeadlineExceeded", err)
+	}
+	if time.Since(start) > time.Second {
+		t.Fatalf("cancel took %v, want fast abort", time.Since(start))
 	}
 }
 
@@ -345,7 +420,8 @@ func TestChatClientCancelNotRecorded(t *testing.T) {
 }
 
 // TestChatStreamEmptyBodyNotSuccess: a 200 that yields zero chunks and no
-// [DONE] is a failure before the first chunk — the slot must back off.
+// [DONE] must fail the candidate (error return, no delivered chunks) so the
+// scheduler falls through to the next source, and record the failure.
 func TestChatStreamEmptyBodyNotSuccess(t *testing.T) {
 	up := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
 		w.Header().Set("Content-Type", "text/event-stream")
@@ -357,15 +433,12 @@ func TestChatStreamEmptyBodyNotSuccess(t *testing.T) {
 		Model:    "m",
 		Messages: []types.ChatMessage{{Role: "user", Content: types.StringContent("x")}},
 	})
-	if err != nil {
-		t.Fatalf("stream: %v", err)
-	}
-	got := 0
-	for range ch {
-		got++
-	}
-	if got != 0 {
-		t.Fatalf("want empty stream, got %d chunks", got)
+	if err == nil {
+		if ch != nil {
+			for range ch {
+			}
+		}
+		t.Fatal("empty stream must fail the candidate for scheduler fallback")
 	}
 	eventually(t, 2*time.Second, func() bool {
 		_, fc, _ := p.ModelHealthInfo("m")
@@ -373,6 +446,62 @@ func TestChatStreamEmptyBodyNotSuccess(t *testing.T) {
 	}, "empty stream must record a failure")
 }
 
+// TestChatStreamErrorFinishFailsCandidate: zen free pool answers HTTP 200
+// with a single-chunk stream whose only payload is finish_reason:"network_error"
+// and an empty delta. That must fail the candidate (not deliver a laundered
+// empty reply), so the scheduler can fall through to the next source.
+func TestChatStreamErrorFinishFailsCandidate(t *testing.T) {
+	up := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.Header().Set("Content-Type", "text/event-stream")
+		fmt.Fprint(w, `data: {"choices":[{"index":0,"finish_reason":"network_error","delta":{"role":"assistant","content":""}}]}`+"\n\n")
+		fmt.Fprint(w, "data: [DONE]\n\n")
+	}))
+	defer up.Close()
+	p := newTestProvider(t, src("mock", up.URL, "opencode", "m"))
+	_, err := p.ChatStream(context.Background(), &types.ChatRequest{
+		Model:    "m",
+		Messages: []types.ChatMessage{{Role: "user", Content: types.StringContent("x")}},
+	})
+	if err == nil {
+		t.Fatal("network_error-only stream must fail the candidate")
+	}
+	eventually(t, 2*time.Second, func() bool {
+		_, fc, _ := p.ModelHealthInfo("m")
+		return fc >= 1
+	}, "error-only stream must record a failure")
+}
+
+// TestChatStreamInstantEmptyStopStillDelivered: a legitimate completion that
+// ends immediately with the standard finish_reason:"stop" and zero content is
+// NOT an upstream error and must still reach the client.
+func TestChatStreamInstantEmptyStopStillDelivered(t *testing.T) {
+	up := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.Header().Set("Content-Type", "text/event-stream")
+		fmt.Fprint(w, `data: {"choices":[{"index":0,"finish_reason":"stop","delta":{}}]}`+"\n\n")
+		fmt.Fprint(w, "data: [DONE]\n\n")
+	}))
+	defer up.Close()
+	p := newTestProvider(t, src("mock", up.URL, "openai", "m"))
+	ch, err := p.ChatStream(context.Background(), &types.ChatRequest{
+		Model:    "m",
+		Messages: []types.ChatMessage{{Role: "user", Content: types.StringContent("x")}},
+	})
+	if err != nil {
+		t.Fatalf("standard stop stream must be delivered: %v", err)
+	}
+	n := 0
+	for range ch {
+		n++
+	}
+	if n == 0 {
+		t.Fatal("expected at least the terminating chunk")
+	}
+	eventually(t, 2*time.Second, func() bool {
+		_, fc, _ := p.ModelHealthInfo("m")
+		return fc == 0
+	}, "standard stop must count as success")
+}
+
 // TestImageAutoUsesImageModel: AUTO image generation on a mixed source must
 // send the image-kind model id, never the best chat model.
 func TestImageAutoUsesImageModel(t *testing.T) {
@@ -401,3 +530,656 @@ func TestImageAutoUsesImageModel(t *testing.T) {
 		t.Fatalf("AUTO image must use the image-kind model, got %q", gotModel)
 	}
 }
+
+func TestStandardSSEChunkFinishReason(t *testing.T) {
+	out := standardSSEChunk(`{"choices":[{"index":0,"finish_reason":"tool_calls","delta":{}}]}`)
+	if !strings.Contains(out, `"done":true`) || !strings.Contains(out, `"finish_reason":"tool_calls"`) {
+		t.Fatalf("real finish reason must pass through: %s", out)
+	}
+	out = standardSSEChunk(`{"choices":[{"index":0,"finish_reason":"","delta":{"content":"hi"}}]}`)
+	if strings.Contains(out, `"Done":true`) {
+		t.Fatalf("empty-string finish reason is not a finish signal: %s", out)
+	}
+}
+
+// TestAdapterHookCondensesError: the adapter transform_error hook owns the
+// per-source error format; its reason must reach the client verbatim.
+func TestAdapterHookCondensesError(t *testing.T) {
+	up := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.Header().Set("Content-Type", "application/json")
+		w.WriteHeader(503)
+		fmt.Fprint(w, `{"error":{"type":"FreeUsageLimitError","message":"Rate limit exceeded. Please try again later."}}`)
+	}))
+	defer up.Close()
+	p := newTestProvider(t, src("mock", up.URL, "opencode", "m"))
+	_, err := p.Chat(context.Background(), &types.ChatRequest{
+		Model:    "m",
+		Messages: []types.ChatMessage{{Role: "user", Content: types.StringContent("x")}},
+	})
+	if err == nil || !strings.Contains(err.Error(), "zen free pool quota exhausted") {
+		t.Fatalf("adapter hook reason must surface, got %v", err)
+	}
+	if strings.Contains(err.Error(), "Rate limit exceeded") {
+		t.Fatalf("raw upstream body must not leak past the hook: %v", err)
+	}
+}
+
+// TestUnknownErrorFallbackWithoutHook: an adapter without transform_error
+// gets the uniform core fallback; the raw body must not leak to clients.
+func TestUnknownErrorFallbackWithoutHook(t *testing.T) {
+	dir := filepath.Join(t.TempDir(), "adapters")
+	if err := os.MkdirAll(dir, 0o755); err != nil {
+		t.Fatal(err)
+	}
+	minimal := `return { name="custom", endpoint="/chat/completions",
+	transform_request=function(raw) return raw end,
+	transform_response=function(raw) return raw end }
+`
+	if err := os.WriteFile(filepath.Join(dir, "custom.lua"), []byte(minimal), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	vm := lua.NewVM(dir)
+	if err := vm.Start(); err != nil {
+		t.Fatalf("vm: %v", err)
+	}
+	defer vm.Stop()
+	up := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.WriteHeader(500)
+		fmt.Fprint(w, `{"weird":{"shape":["no","message"]}}`)
+	}))
+	defer up.Close()
+	p := New(src("mock", up.URL, "custom", "m"), vm)
+	_, err := p.Chat(context.Background(), &types.ChatRequest{
+		Model:    "m",
+		Messages: []types.ChatMessage{{Role: "user", Content: types.StringContent("x")}},
+	})
+	if err == nil || !strings.Contains(err.Error(), "unknown error") {
+		t.Fatalf("hook-less adapter must fall back to unknown error, got %v", err)
+	}
+	if strings.Contains(err.Error(), "shape") {
+		t.Fatalf("raw body must not leak: %v", err)
+	}
+}
+
+// TestToolIndexRemapper verifies sparse upstream tool-call indices are
+// compacted to dense 0-based OpenAI ordinals (Issue 2). Anthropic emits the
+// content_block ordinal, so a thinking block at index 0 pushes the first
+// tool_use to index 1 — OpenAI clients accumulating by index would leave a
+// hole at 0 and mis-assemble multi-tool responses.
+func TestToolIndexRemapper(t *testing.T) {
+	t.Run("single tool after thinking block", func(t *testing.T) {
+		r := newToolIndexRemapper()
+		// content_block index 1 (thinking was 0) -> dense 0
+		out := r.remap(json.RawMessage(`[{"index":1,"id":"toolu_a","type":"function","function":{"name":"calc","arguments":""}}]`))
+		var got []map[string]interface{}
+		if err := json.Unmarshal(out, &got); err != nil {
+			t.Fatalf("unmarshal: %v (%s)", err, out)
+		}
+		if got[0]["index"] != float64(0) {
+			t.Fatalf("index = %v, want 0: %s", got[0]["index"], out)
+		}
+		// id/name/arguments must survive verbatim
+		if got[0]["id"] != "toolu_a" {
+			t.Fatalf("id lost: %s", out)
+		}
+		fn := got[0]["function"].(map[string]interface{})
+		if fn["name"] != "calc" {
+			t.Fatalf("function.name lost: %s", out)
+		}
+		// subsequent argument fragments on the SAME upstream index reuse ordinal 0
+		out2 := r.remap(json.RawMessage(`[{"index":1,"id":"","type":"function","function":{"name":"","arguments":"{\"a\":1}"}}]`))
+		var got2 []map[string]interface{}
+		if err := json.Unmarshal(out2, &got2); err != nil {
+			t.Fatalf("unmarshal2: %v", err)
+		}
+		if got2[0]["index"] != float64(0) {
+			t.Fatalf("fragment index = %v, want stable 0: %s", got2[0]["index"], out2)
+		}
+	})
+
+	t.Run("multiple tools get distinct dense ordinals", func(t *testing.T) {
+		r := newToolIndexRemapper()
+		// thinking=0, tool_use=1, tool_use=2  ->  0, 1
+		a := r.remap(json.RawMessage(`[{"index":1,"id":"t1","type":"function","function":{"name":"f1","arguments":""}}]`))
+		b := r.remap(json.RawMessage(`[{"index":2,"id":"t2","type":"function","function":{"name":"f2","arguments":""}}]`))
+		var ga, gb []map[string]interface{}
+		json.Unmarshal(a, &ga)
+		json.Unmarshal(b, &gb)
+		if ga[0]["index"] != float64(0) {
+			t.Fatalf("first tool index = %v, want 0", ga[0]["index"])
+		}
+		if gb[0]["index"] != float64(1) {
+			t.Fatalf("second tool index = %v, want 1", gb[0]["index"])
+		}
+		// interleaved fragments keep their own ordinals
+		a2 := r.remap(json.RawMessage(`[{"index":1,"id":"","type":"function","function":{"name":"","arguments":"x"}}]`))
+		var ga2 []map[string]interface{}
+		json.Unmarshal(a2, &ga2)
+		if ga2[0]["index"] != float64(0) {
+			t.Fatalf("t1 fragment index = %v, want 0", ga2[0]["index"])
+		}
+	})
+
+	t.Run("already dense indices pass through untouched", func(t *testing.T) {
+		r := newToolIndexRemapper()
+		in := json.RawMessage(`[{"index":0,"id":"t","type":"function","function":{"name":"f","arguments":""}}]`)
+		out := r.remap(in)
+		if string(out) != string(in) {
+			t.Fatalf("dense input was rewritten:\n in=%s\nout=%s", in, out)
+		}
+	})
+
+	t.Run("empty and malformed input is safe", func(t *testing.T) {
+		r := newToolIndexRemapper()
+		if got := r.remap(nil); got != nil {
+			t.Fatalf("nil -> %s", got)
+		}
+		bad := json.RawMessage(`not json`)
+		if got := r.remap(bad); string(got) != string(bad) {
+			t.Fatalf("malformed input must pass through: %s", got)
+		}
+	})
+}
+
+// ---- half-cooldown probe gate (plan 阶段 1) ----
+
+// TestProbeGateHalfCooldown pins the core rule: a cooling model is fully silent
+// in the first half of its cooldown window and probeable exactly once in the
+// second half.
+func TestProbeGateHalfCooldown(t *testing.T) {
+	var st ModelState
+	// open a 100s window centred so that "now" sits in the first half
+	now := time.Now().Unix()
+	st.cooldownFrom.Store(now - 10)
+	st.cooldownUntil.Store(now + 90)
+	if st.TryProbe() {
+		t.Fatal("first half of the cooldown window must stay silent")
+	}
+
+	// move into the second half
+	st.cooldownFrom.Store(now - 90)
+	st.cooldownUntil.Store(now + 10)
+	if !st.TryProbe() {
+		t.Fatal("second half of the cooldown window must allow one probe")
+	}
+	if st.TryProbe() {
+		t.Fatal("probe permit must be single-token")
+	}
+	if !st.Probing() {
+		t.Fatal("Probing() must report the claimed permit")
+	}
+	st.ProbeDone()
+	if st.Probing() {
+		t.Fatal("ProbeDone must release the permit")
+	}
+	if !st.TryProbe() {
+		t.Fatal("permit must be reusable after ProbeDone")
+	}
+}
+
+// TestProbeGateNoProbeWhenHealthy: a healthy model needs no probe permit — the
+// caller schedules it normally.
+func TestProbeGateNoProbeWhenHealthy(t *testing.T) {
+	var idle ModelState
+	if idle.TryProbe() {
+		t.Fatal("a healthy model must not be probed (schedule it normally)")
+	}
+}
+
+// TestProbeGateRescuesPrefFloor documents a fix that comes with the probe gate:
+// a slot whose preference sank to prefMin used to be refused by ModelAvailable
+// forever, even long after every cooldown expired — a permanent blacklist. It is
+// now probeable, so one success brings it back.
+func TestProbeGateRescuesPrefFloor(t *testing.T) {
+	var dead ModelState
+	dead.pref.Store(prefMin)
+	if !dead.TryProbe() {
+		t.Fatal("a slot stuck at prefMin must be probeable, not blacklisted")
+	}
+	dead.RecordSuccess()
+	dead.ProbeDone()
+	if dead.Pref() <= prefMin {
+		t.Fatalf("one probe success must lift the slot off the floor, pref=%d", dead.Pref())
+	}
+}
+
+// TestProbeSuccessRestoresImmediately is the user-visible payoff: a recovered
+// upstream must not sit out the rest of its cooldown.
+func TestProbeSuccessRestoresImmediately(t *testing.T) {
+	var st ModelState
+	for i := 0; i < backoffCapN; i++ {
+		st.RecordFailure(false)
+	}
+	if cd := st.CooldownUntil() - time.Now().Unix(); cd < int64(backoffCap.Seconds())-2 {
+		t.Fatalf("10 failures must cool for backoffCap, got %ds", cd)
+	}
+	// jump into the probe window: the midpoint of [from, until) must be in the
+	// past, so push the window's start back past its full length
+	now := time.Now().Unix()
+	st.cooldownFrom.Store(now - int64(backoffCap.Seconds()) - 10)
+	if !st.TryProbe() {
+		t.Fatal("probe must be allowed past the midpoint")
+	}
+	if !st.Probing() {
+		t.Fatal("probe permit must be held while the probe runs")
+	}
+	st.RecordSuccess()
+	st.ProbeDone()
+	if !st.Available() {
+		t.Fatal("probe success must clear the cooldown immediately")
+	}
+	if st.FailCount() != 0 {
+		t.Fatalf("probe success must reset failCount, got %d", st.FailCount())
+	}
+	if st.CooldownUntil() != 0 || st.CooldownFrom() != 0 {
+		t.Fatal("probe success must clear both cooldown bounds")
+	}
+}
+
+// TestProbeFailureDefersNextProbe: a failed probe must not immediately retry —
+// the new window's midpoint pushes the next probe out.
+func TestProbeFailureDefersNextProbe(t *testing.T) {
+	var st ModelState
+	st.RecordFailure(false) // 5s window
+	now := time.Now().Unix()
+	st.cooldownFrom.Store(now - 100)
+	st.cooldownUntil.Store(now + 10)
+	if !st.TryProbe() {
+		t.Fatal("probe must be allowed past the midpoint")
+	}
+	st.RecordFailure(false) // probe failed: a fresh, longer window opens
+	st.ProbeDone()
+	if st.TryProbe() {
+		t.Fatal("a failed probe must defer the next probe to the new midpoint")
+	}
+	if from := st.CooldownFrom(); from < now {
+		t.Fatalf("failed probe must reopen the window (from=%d, now=%d)", from, now)
+	}
+}
+
+// TestAuthCooldownIsSeparateFromCap documents that credential failures cool on
+// their own (longer) schedule instead of reusing the exponential cap.
+func TestAuthCooldownIsSeparateFromCap(t *testing.T) {
+	var st ModelState
+	st.RecordFailure(true)
+	cd := st.CooldownUntil() - time.Now().Unix()
+	if cd < int64(authCooldown.Seconds())-2 || cd > int64(authCooldown.Seconds())+2 {
+		t.Fatalf("auth failure must cool for authCooldown (%v), got %ds", authCooldown, cd)
+	}
+	if st.FailCount() != backoffCapN {
+		t.Fatalf("auth failure must persist the capped count, got %d", st.FailCount())
+	}
+}
+
+// TestQuotaExhaustedDoesNotEscalate is the second half of the user's complaint:
+// running out of allowance must not be treated as a failure streak.
+func TestQuotaExhaustedDoesNotEscalate(t *testing.T) {
+	var st ModelState
+	st.RecordQuotaExhausted(2 * time.Minute)
+	if st.FailCount() != 0 {
+		t.Fatalf("quota exhaustion must not bump failCount, got %d", st.FailCount())
+	}
+	cd := st.CooldownUntil() - time.Now().Unix()
+	if cd < 118 || cd > 122 {
+		t.Fatalf("quota cooldown must honour the reset window, got %ds", cd)
+	}
+
+	// an unbounded window is clamped so a topped-up allowance is still noticed
+	var big ModelState
+	big.RecordQuotaExhausted(72 * time.Hour)
+	if cd := big.CooldownUntil() - time.Now().Unix(); cd > int64(maxQuotaCooldown.Seconds())+2 {
+		t.Fatalf("quota cooldown must be capped at maxQuotaCooldown, got %ds", cd)
+	}
+
+	// no hint at all falls back to the short window
+	var none ModelState
+	none.RecordQuotaExhausted(0)
+	if cd := none.CooldownUntil() - time.Now().Unix(); cd > int64(rateLimitCooldown.Seconds())+2 {
+		t.Fatalf("unknown reset window must fall back to rateLimitCooldown, got %ds", cd)
+	}
+}
+
+// TestQuotaReasonRouting checks that a 429 carrying a quota phrase is cooled as
+// quota exhaustion while a plain rate limit keeps the short fixed window.
+func TestQuotaReasonRouting(t *testing.T) {
+	p := newTestProvider(t, src("q", "http://127.0.0.1:1", "openai", "m"))
+	p.ReportStatusReason("m", 429, "429 insufficient_quota: you exceeded your current quota")
+	st := p.state("m")
+	if st.FailCount() != 0 {
+		t.Fatalf("quota rejection must not bump failCount, got %d", st.FailCount())
+	}
+	if cd := st.CooldownUntil() - time.Now().Unix(); cd <= int64(rateLimitCooldown.Seconds()) {
+		t.Fatalf("quota rejection must cool longer than a plain 429, got %ds", cd)
+	}
+
+	p2 := newTestProvider(t, src("r", "http://127.0.0.1:1", "openai", "m"))
+	p2.ReportStatusReason("m", 429, "rate limit reached, slow down")
+	st2 := p2.state("m")
+	if cd := st2.CooldownUntil() - time.Now().Unix(); cd > int64(rateLimitCooldown.Seconds())+2 {
+		t.Fatalf("plain 429 must keep the short window, got %ds", cd)
+	}
+}
+
+// TestModelSchedulableProbeHandoff covers the provider-level gate the scheduler
+// consumes.
+func TestModelSchedulableProbeHandoff(t *testing.T) {
+	p := newTestProvider(t, src("s", "http://127.0.0.1:1", "openai", "m"))
+	if ok, probe := p.ModelSchedulable("m"); !ok || probe {
+		t.Fatalf("healthy model must be normally schedulable, got ok=%v probe=%v", ok, probe)
+	}
+	st := p.state("m")
+	st.RecordFailure(false)
+	now := time.Now().Unix()
+	st.cooldownFrom.Store(now - 100)
+	st.cooldownUntil.Store(now + 10)
+
+	ok, probe := p.ModelSchedulable("m")
+	if !ok || !probe {
+		t.Fatalf("cooling model past midpoint must be probeable, got ok=%v probe=%v", ok, probe)
+	}
+	if ok2, _ := p.ModelSchedulable("m"); ok2 {
+		t.Fatal("only one probe permit may be outstanding")
+	}
+	p.ProbeDone("m")
+	if ok3, probe3 := p.ModelSchedulable("m"); !ok3 || !probe3 {
+		t.Fatal("permit must be reclaimable after ProbeDone")
+	}
+}
+
+// TestProbeSlotsFollowsMaxConcurrent pins the user's example: max_concurrent=10
+// yields exactly one probe slot.
+func TestProbeSlotsFollowsMaxConcurrent(t *testing.T) {
+	for _, tc := range []struct{ mc, want int }{
+		{0, 1}, {1, 1}, {4, 1}, {10, 1}, {20, 2}, {100, 2},
+	} {
+		s := src("p", "http://127.0.0.1:1", "openai", "m")
+		s.MaxConcurrent = tc.mc
+		p := newTestProvider(t, s)
+		if got := p.ProbeSlots(); got != tc.want {
+			t.Fatalf("max_concurrent=%d: probe slots = %d, want %d", tc.mc, got, tc.want)
+		}
+	}
+}
+
+// TestModelProbeInfo checks the UI-facing view of the probe window.
+func TestModelProbeInfo(t *testing.T) {
+	p := newTestProvider(t, src("i", "http://127.0.0.1:1", "openai", "m"))
+	if from, after, probing := p.ModelProbeInfo("m"); from != 0 || after != 0 || probing {
+		t.Fatalf("idle model must report no probe window, got %d %d %v", from, after, probing)
+	}
+	st := p.state("m")
+	now := time.Now().Unix()
+	st.cooldownFrom.Store(now)
+	st.cooldownUntil.Store(now + 100)
+	from, after, _ := p.ModelProbeInfo("m")
+	if from != now || after != now+50 {
+		t.Fatalf("probe window = (%d,%d), want (%d,%d)", from, after, now, now+50)
+	}
+}
+
+// TestUpstreamRecoveryWithoutWaitingOutCooldown is the end-to-end shape of the
+// user's complaint: an upstream that 500s ten times lands in the capped
+// cooldown, and used to be unusable for a full 30 minutes even after it
+// recovered. With the probe gate the slot is silent for the first half of the
+// window and then serves the very next request itself, clearing the cooldown.
+//
+// Real time is not waited out: the window bounds are moved so the test observes
+// the same state the scheduler would see at cap/2.
+func TestUpstreamRecoveryWithoutWaitingOutCooldown(t *testing.T) {
+	var healthy atomic.Bool
+	var hits atomic.Int64
+	up := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		hits.Add(1)
+		if !healthy.Load() {
+			w.WriteHeader(500)
+			return
+		}
+		fmt.Fprint(w, `{"choices":[{"message":{"content":"back"},"finish_reason":"stop"}]}`)
+	}))
+	defer up.Close()
+	p := newTestProvider(t, src("recov", up.URL, "openai", "m"))
+	req := func() *types.ChatRequest {
+		return &types.ChatRequest{
+			Model:    "m",
+			Messages: []types.ChatMessage{{Role: "user", Content: types.StringContent("x")}},
+		}
+	}
+
+	// 1. drive the slot into the capped cooldown
+	for i := 0; i < backoffCapN; i++ {
+		if _, err := p.Chat(context.Background(), req()); err == nil {
+			t.Fatalf("attempt %d: expected upstream failure", i)
+		}
+	}
+	st := p.state("m")
+	if p.ModelAvailable("m") {
+		t.Fatal("10 consecutive 5xx must take the slot out of normal rotation")
+	}
+	cd := st.CooldownUntil() - time.Now().Unix()
+	if cd < int64(backoffCap.Seconds())-2 {
+		t.Fatalf("cooldown = %ds, want ~%v", cd, backoffCap)
+	}
+	// the old 30-minute lockout is gone
+	if cd > int64((6 * time.Minute).Seconds()) {
+		t.Fatalf("cooldown %ds is too long: the cap must be minutes, not half an hour", cd)
+	}
+
+	// 2. upstream recovers, but we are still in the FIRST half of the window:
+	//    the slot must stay completely silent (no probe traffic at all).
+	healthy.Store(true)
+	quiet := hits.Load()
+	if ok, probe := p.ModelSchedulable("m"); ok || probe {
+		t.Fatalf("first half of the window must stay silent, got ok=%v probe=%v", ok, probe)
+	}
+	if hits.Load() != quiet {
+		t.Fatal("no request may reach the upstream during the silent half")
+	}
+
+	// 3. advance past the midpoint: exactly one probe is admitted.
+	st.cooldownFrom.Store(time.Now().Unix() - int64(backoffCap.Seconds()) - 1)
+	ok, isProbe := p.ModelSchedulable("m")
+	if !ok || !isProbe {
+		t.Fatalf("past the midpoint one probe must be admitted, got ok=%v probe=%v", ok, isProbe)
+	}
+	if ok2, _ := p.ModelSchedulable("m"); ok2 {
+		t.Fatal("only ONE probe may be in flight while cooling")
+	}
+
+	// 4. the probe is a real request: its success restores the slot fully.
+	resp, err := p.Chat(context.Background(), req())
+	p.ProbeDone("m")
+	if err != nil {
+		t.Fatalf("probe request failed: %v", err)
+	}
+	if resp.Content != "back" {
+		t.Fatalf("probe response = %q, want \"back\"", resp.Content)
+	}
+	if !p.ModelAvailable("m") {
+		t.Fatal("a successful probe must return the slot to normal rotation immediately")
+	}
+	if st.CooldownUntil() != 0 || st.FailCount() != 0 {
+		t.Fatalf("probe success must clear cooldown/failCount, got until=%d fails=%d",
+			st.CooldownUntil(), st.FailCount())
+	}
+	if ok, probe := p.ModelSchedulable("m"); !ok || probe {
+		t.Fatalf("recovered slot must schedule normally, got ok=%v probe=%v", ok, probe)
+	}
+}
+
+// TestRecoveryProbeBudgetIsBounded quantifies the cost of the shorter cap
+// against a still-dead upstream: one probe per window, not a retry storm.
+func TestRecoveryProbeBudgetIsBounded(t *testing.T) {
+	var hits atomic.Int64
+	up := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		hits.Add(1)
+		w.WriteHeader(500)
+	}))
+	defer up.Close()
+	p := newTestProvider(t, src("dead", up.URL, "openai", "m"))
+	req := &types.ChatRequest{
+		Model:    "m",
+		Messages: []types.ChatMessage{{Role: "user", Content: types.StringContent("x")}},
+	}
+	for i := 0; i < backoffCapN; i++ {
+		_, _ = p.Chat(context.Background(), req)
+	}
+	baseline := hits.Load()
+	st := p.state("m")
+
+	// simulate three consecutive probe windows against a dead upstream
+	for i := 0; i < 3; i++ {
+		st.cooldownFrom.Store(time.Now().Unix() - int64(backoffCap.Seconds()) - 1)
+		ok, isProbe := p.ModelSchedulable("m")
+		if !ok || !isProbe {
+			t.Fatalf("window %d: expected a probe permit", i)
+		}
+		// a second caller in the same window gets nothing
+		if ok2, _ := p.ModelSchedulable("m"); ok2 {
+			t.Fatalf("window %d: more than one probe admitted", i)
+		}
+		if _, err := p.Chat(context.Background(), req); err == nil {
+			t.Fatalf("window %d: expected the probe to fail", i)
+		}
+		p.ProbeDone("m")
+	}
+	if extra := hits.Load() - baseline; extra != 3 {
+		t.Fatalf("3 probe windows must cost exactly 3 upstream requests, got %d", extra)
+	}
+}
+
+// TestQuotaPenaltyIsLighterThanFailure is the regression for the report
+// "今天一次失败没看到，分数目前已经 -10 了" on an allowance-metered source
+// (sensenova). RecordQuotaExhausted deliberately does not touch failCount — the
+// WebUI therefore showed fails=0 and cooling=false — yet it used to charge the
+// SAME score penalty as a real failure, so two ordinary quota resets dragged the
+// slot to -10 with no visible fault. An exhausted allowance means "no budget
+// right now", not "this model is broken": the cooldown already keeps it out of
+// rotation, so the score penalty must be mild.
+func TestQuotaPenaltyIsLighterThanFailure(t *testing.T) {
+	quota := &ModelState{}
+	quota.RecordQuotaExhausted(time.Minute)
+	fail := &ModelState{}
+	fail.RecordFailure(false)
+
+	qp, fp := quota.Pref(), fail.Pref()
+	if qp <= fp {
+		t.Fatalf("a quota reset must cost less score than a real failure: quota=%d failure=%d", qp, fp)
+	}
+	if qp != -prefQuotaStep {
+		t.Errorf("quota penalty = %d, want -%d", qp, prefQuotaStep)
+	}
+	if fp != -prefFailStep {
+		t.Errorf("failure penalty = %d, want -%d", fp, prefFailStep)
+	}
+
+	// The reported symptom: several quota resets in a day must NOT approach the
+	// prefMin floor, since each one is a normal event for a metered source.
+	s := &ModelState{}
+	for i := 0; i < 5; i++ {
+		s.RecordQuotaExhausted(time.Minute)
+	}
+	if got := s.Pref(); got <= prefMin/2 {
+		t.Errorf("5 quota resets sank the score to %d (floor %d); a metered source would be written off", got, prefMin)
+	}
+	if s.FailCount() != 0 {
+		t.Errorf("quota exhaustion must not inflate failCount, got %d", s.FailCount())
+	}
+}
+
+// TestRateLimitPenaltyIsMild covers the sibling case: a 429 means "too fast",
+// not "broken", so a popular-but-healthy slot must not sink for being throttled.
+func TestRateLimitPenaltyIsMild(t *testing.T) {
+	s := &ModelState{}
+	s.RecordRateLimit()
+	if got := s.Pref(); got != -prefQuotaStep {
+		t.Errorf("429 penalty = %d, want -%d (same class as a quota event)", got, prefQuotaStep)
+	}
+}
+
+// TestSuccessRecoversFasterThanBefore pins the penalty/reward asymmetry. Reward
+// used to be +1 against a -5 penalty, so a slot at -10 needed TEN consecutive
+// successes just to reach neutral — in practice it never got there, because a
+// low score makes the slot unattractive to the scheduler in the first place.
+func TestSuccessRecoversFasterThanBefore(t *testing.T) {
+	s := &ModelState{}
+	s.RecordFailure(false)
+	s.RecordFailure(false) // -10, the score from the bug report
+
+	n := 0
+	for s.Pref() < 0 {
+		s.RecordSuccess()
+		n++
+		if n > 50 {
+			t.Fatal("score never recovered")
+		}
+	}
+	// with +2 per success, -10 needs 5; the old +1 needed 10
+	if n > 5 {
+		t.Errorf("recovery from -10 took %d successes, want <= 5", n)
+	}
+	// a real failure must still outweigh a single success, or failures stop
+	// meaning anything
+	if prefSuccessStep >= prefFailStep {
+		t.Errorf("success reward %d must stay below the failure penalty %d", prefSuccessStep, prefFailStep)
+	}
+}
+
+// TestPrefDecayRehabilitatesIdleSlot closes the starvation loop: a penalised
+// slot is unattractive, so the scheduler stops picking it, so it never earns the
+// successes that would rehabilitate it. Without decay the score stays negative
+// forever even though the upstream may be perfectly healthy.
+func TestPrefDecayRehabilitatesIdleSlot(t *testing.T) {
+	s := &ModelState{}
+	s.RecordFailure(false)
+	s.RecordFailure(false)
+	if s.Pref() != -2*prefFailStep {
+		t.Fatalf("setup: pref=%d", s.Pref())
+	}
+
+	// backdate the last touch by three decay intervals
+	step := int64(prefDecayAfter / time.Second)
+	s.prefTouched.Store(time.Now().Unix() - 3*step)
+
+	got := s.Pref()
+	if got != -2*prefFailStep+3 {
+		t.Errorf("after 3 idle intervals pref=%d, want %d", got, -2*prefFailStep+3)
+	}
+
+	// decay must never overshoot into positive territory
+	s.prefTouched.Store(time.Now().Unix() - 1000*step)
+	if got := s.Pref(); got != 0 {
+		t.Errorf("decay overshot to %d, want to stop at 0", got)
+	}
+
+	// a positive score must not be dragged down for being idle
+	good := &ModelState{}
+	good.RecordSuccess()
+	good.RecordSuccess()
+	before := good.Pref()
+	good.prefTouched.Store(time.Now().Unix() - 1000*step)
+	if after := good.Pref(); after != before {
+		t.Errorf("idle decay must not touch a positive score: %d -> %d", before, after)
+	}
+}
+
+// TestDecayLiftsSlotOffTheFloor verifies decay actually restores schedulability:
+// a slot at prefMin is refused by ModelAvailable, and before decay existed only
+// a probe could rescue it.
+func TestDecayLiftsSlotOffTheFloor(t *testing.T) {
+	s := &ModelState{}
+	for i := 0; i < 10; i++ {
+		s.RecordFailure(false)
+	}
+	if s.Pref() != prefMin {
+		t.Fatalf("setup: want floor %d, got %d", prefMin, s.Pref())
+	}
+	// clear the cooldown so only the score gates the slot
+	s.cooldownUntil.Store(0)
+	s.cooldownFrom.Store(0)
+
+	step := int64(prefDecayAfter / time.Second)
+	s.prefTouched.Store(time.Now().Unix() - 2*step)
+	if got := s.Pref(); got <= prefMin {
+		t.Errorf("decay failed to lift the slot off the floor: %d", got)
+	}
+}

internal/provider/registry.go

@@ -211,6 +211,18 @@ type SourceStatus struct {
 	FailCount     int      `json:"fail_count,omitempty"`
 	BackoffUntil  int64    `json:"backoff_until,omitempty"`
 	Permanent     bool     `json:"permanent,omitempty"`
+	// RecentOK / RecentErr count real gateway requests served by this source
+	// within the last 300s (drives the status column so a source that is
+	// actually serving traffic can never be shown as down just because a
+	// probe was rate-limited by the upstream).
+	RecentOK  int64 `json:"recent_ok,omitempty"`
+	RecentErr int64 `json:"recent_err,omitempty"`
+	// AvgFirstByteMs is the mean time-to-first-byte (ms) for recently
+	// successful requests served by this source (0 when unmeasured).
+	AvgFirstByteMs int64 `json:"avg_first_byte_ms,omitempty"`
+	// AvgTokPerS is the aggregate completion throughput (tokens/s) for
+	// recently successful requests served by this source (0 when no data).
+	AvgTokPerS int64 `json:"avg_tok_per_s,omitempty"`
 }
 
 // ProbeAll runs a live reachability check for every provider (in parallel).
@@ -260,4 +272,4 @@ func (r *Registry) Status() []SourceStatus {
 		out = append(out, s)
 	}
 	return out
-}
+}

internal/scheduler/scheduler.go

@@ -42,6 +42,12 @@ type Provider interface {
 	Name() string
 	ModelFor(reqModel string) string
 	ModelAvailable(model string) bool
+	// ModelSchedulable is the probe-aware availability gate: ok reports
+	// whether the model may take a request now, isProbe marks that it is only
+	// allowed as a cooldown probe (the caller must release the permit with
+	// ProbeDone once the attempt finished).
+	ModelSchedulable(model string) (ok bool, isProbe bool)
+	ProbeDone(model string)
 	Pref(model string) int64
 	Chat(ctx context.Context, req *types.ChatRequest) (*types.UnifiedResponse, error)
 	ChatStream(ctx context.Context, req *types.ChatRequest) (<-chan types.UnifiedChunk, error)
@@ -176,18 +182,84 @@ type tierResult struct {
 	hard   []TierError // hard failures seen in this pass (nil = none)
 }
 
+// candidate is one schedulable slot in a tier pass. probe marks a slot that is
+// still cooling but past its half-cooldown mark and holding the probe permit:
+// it is tried only after every normal slot, so probe traffic is what the tier
+// falls back to instead of what it prefers.
+type candidate struct {
+	slot  *Slot
+	probe bool
+}
+
+// collectCands partitions a tier's slots into normal and probe candidates,
+// normal first. Quota-exhausted slots are dropped outright. Every probe
+// candidate returned holds a probe permit, so the caller MUST call
+// releaseProbes on the result exactly once.
+func collectCands(slots []*Slot, exhausted func(*Slot) bool) []candidate {
+	var normal, probes []candidate
+	for _, sl := range slots {
+		if exhausted != nil && exhausted(sl) {
+			continue
+		}
+		ok, isProbe := sl.Prov.ModelSchedulable(sl.Model)
+		if !ok {
+			continue
+		}
+		if isProbe {
+			probes = append(probes, candidate{slot: sl, probe: true})
+			continue
+		}
+		normal = append(normal, candidate{slot: sl})
+	}
+	return append(normal, probes...)
+}
+
+// releaseProbes hands every claimed probe permit back, whether or not the probe
+// slot was actually used.
+func releaseProbes(cands []candidate) {
+	for _, c := range cands {
+		if c.probe {
+			c.slot.Prov.ProbeDone(c.slot.Model)
+		}
+	}
+}
+
+// normalCount is how many leading candidates are normal (non-probe). The
+// round-robin cursor rotates only over those: probe slots are a strictly
+// ordered tail, never a rotation target.
+func normalCount(cands []candidate) int {
+	for i, c := range cands {
+		if c.probe {
+			return i
+		}
+	}
+	return len(cands)
+}
+
 // runTier executes one tier pass starting at the round-robin base index.
 // Cooldown is the only hard skip (re-verified per slot); a busy slot is
 // skipped without any penalty; a hard failure is recorded and the pass moves
 // on to the next slot (plan 2.3: "单请求内不重试已失败槽" — the failed slot is
 // not retried, the others still are). hard == nil and no success means every
 // candidate was merely busy/cooling, so the caller may wait a bounded time.
-func runTier(ctx context.Context, tn *TierNode, cands []*Slot, base int64, req *types.ChatRequest, stream bool) tierResult {
+//
+// Normal candidates rotate by base; probe candidates form a fixed tail tried
+// only after every normal slot failed or was busy.
+func runTier(ctx context.Context, tn *TierNode, cands []candidate, base int64, req *types.ChatRequest, stream bool) tierResult {
 	n := len(cands)
+	norm := normalCount(cands)
 	var hard []TierError
 	for i := 0; i < n; i++ {
-		sl := cands[(int(base)+i)%n]
-		if !sl.Prov.ModelAvailable(sl.Model) {
+		var c candidate
+		if i < norm {
+			c = cands[(int(base)+i)%norm] // rotate within the normal head
+		} else {
+			c = cands[i] // probe tail keeps its order
+		}
+		sl := c.slot
+		// A probe candidate is intentionally NOT re-checked here: it is cooling
+		// by definition, and its permit was already claimed.
+		if !c.probe && !sl.Prov.ModelAvailable(sl.Model) {
 			continue
 		}
 		r := *req
@@ -233,76 +305,40 @@ func (s *Scheduler) chainDrive(ctx context.Context, chain *Chain, req *types.Cha
 	}
 	var ce ChainErr
 	for _, tn := range chain.Tiers {
-		// initial filter: quota-exhausted and cooling slots are dropped
-		var cands []*Slot
-		for _, sl := range tn.Slots {
-			if exhausted != nil && exhausted(sl) {
-				continue
-			}
-			if !sl.Prov.ModelAvailable(sl.Model) {
-				continue
-			}
-			cands = append(cands, sl)
-		}
+		// initial filter: quota-exhausted slots are dropped, cooling slots are
+		// dropped unless they qualify as half-cooldown probes (appended last).
+		cands := collectCands(tn.Slots, exhausted)
 		if len(cands) == 0 {
 			ce.Skipped = append(ce.Skipped, fmt.Sprintf("tier %d: no schedulable slot (cooling or quota exhausted)", tn.Tier))
 			continue
 		}
-		// preference orders a same-tier run; stable so equal prefs keep order
-		sort.SliceStable(cands, func(i, j int) bool {
-			return cands[i].Prov.Pref(cands[i].Model) > cands[j].Prov.Pref(cands[j].Model)
-		})
+		// No Pref sort: load balancing is done by round-robin cursor.
+		// Persistently failing slots are excluded by ModelSchedulable
+		// (which checks Pref > prefMin).
 		base := tn.NextStart()
 		res := runTier(ctx, tn, cands, base, req, stream)
 		if res.resp != nil || res.chunks != nil {
+			releaseProbes(cands)
 			return res.resp, res.chunks, res.src, res.model, nil
 		}
 		if ctx.Err() != nil {
+			releaseProbes(cands)
 			return nil, nil, "", "", ctx.Err()
 		}
 		if len(res.hard) > 0 {
 			ce.Tiers = append(ce.Tiers, res.hard...)
+			releaseProbes(cands)
 			continue // hard failures: fall through to the next tier, no waiting
 		}
 		// every candidate was busy or cooling: bounded poll before downgrading
-		deadline := time.Now().Add(busyWait)
-		timer := time.NewTimer(busyPoll)
-		defer timer.Stop()
-		for {
-			select {
-			case <-ctx.Done():
-				return nil, nil, "", "", ctx.Err()
-			case <-timer.C:
+		if err := s.pollBusyTier(ctx, tn, cands, base, req, stream, &ce); err != nil {
+			releaseProbes(cands)
+			if r, ok := err.(*tierSuccess); ok {
+				return r.res.resp, r.res.chunks, r.res.src, r.res.model, nil
 			}
-			done := time.Now().After(deadline)
-			if done {
-				ce.Skipped = append(ce.Skipped, fmt.Sprintf("tier %d: no free slot within %v", tn.Tier, busyWait))
-				break
-			}
-			// refresh candidates: cooldowns may have expired meanwhile
-			var again []*Slot
-			for _, sl := range cands {
-				if sl.Prov.ModelAvailable(sl.Model) {
-					again = append(again, sl)
-				}
-			}
-			if len(again) == 0 {
-				ce.Skipped = append(ce.Skipped, fmt.Sprintf("tier %d: no free slot within %v", tn.Tier, busyWait))
-				break
-			}
-			res = runTier(ctx, tn, again, base, req, stream)
-			if res.resp != nil || res.chunks != nil {
-				return res.resp, res.chunks, res.src, res.model, nil
-			}
-			if ctx.Err() != nil {
-				return nil, nil, "", "", ctx.Err()
-			}
-			if len(res.hard) > 0 {
-				ce.Tiers = append(ce.Tiers, res.hard...)
-				break // hard failure while waiting: stop waiting, fall through
-			}
-			timer.Reset(busyPoll)
+			return nil, nil, "", "", err
 		}
+		releaseProbes(cands)
 	}
 	if len(ce.Tiers) == 0 && len(ce.Skipped) == 0 {
 		return nil, nil, "", "", fmt.Errorf("no auto slot configured")
@@ -310,6 +346,57 @@ func (s *Scheduler) chainDrive(ctx context.Context, chain *Chain, req *types.Cha
 	return nil, nil, "", "", &ce
 }
 
+// tierSuccess carries a successful result out of pollBusyTier through the error
+// return. It is never surfaced to callers of chainDrive.
+type tierSuccess struct{ res tierResult }
+
+func (t *tierSuccess) Error() string { return "tier success" }
+
+// pollBusyTier waits a bounded time for a fully-busy tier to free a slot,
+// retrying the pass while cooldowns expire. It returns nil when the tier should
+// be abandoned (caller falls through to the next tier), a *tierSuccess when a
+// retry succeeded, or a context error.
+func (s *Scheduler) pollBusyTier(ctx context.Context, tn *TierNode, cands []candidate, base int64, req *types.ChatRequest, stream bool, ce *ChainErr) error {
+	deadline := time.Now().Add(busyWait)
+	timer := time.NewTimer(busyPoll)
+	defer timer.Stop()
+	for {
+		select {
+		case <-ctx.Done():
+			return ctx.Err()
+		case <-timer.C:
+		}
+		if time.Now().After(deadline) {
+			ce.Skipped = append(ce.Skipped, fmt.Sprintf("tier %d: no free slot within %v", tn.Tier, busyWait))
+			return nil
+		}
+		// refresh candidates: cooldowns may have expired meanwhile. Probe
+		// candidates keep their already-claimed permit and stay eligible.
+		var again []candidate
+		for _, c := range cands {
+			if c.probe || c.slot.Prov.ModelAvailable(c.slot.Model) {
+				again = append(again, c)
+			}
+		}
+		if len(again) == 0 {
+			ce.Skipped = append(ce.Skipped, fmt.Sprintf("tier %d: no free slot within %v", tn.Tier, busyWait))
+			return nil
+		}
+		res := runTier(ctx, tn, again, base, req, stream)
+		if res.resp != nil || res.chunks != nil {
+			return &tierSuccess{res: res}
+		}
+		if ctx.Err() != nil {
+			return ctx.Err()
+		}
+		if len(res.hard) > 0 {
+			ce.Tiers = append(ce.Tiers, res.hard...)
+			return nil // hard failure while waiting: stop waiting, fall through
+		}
+		timer.Reset(busyPoll)
+	}
+}
+
 // ChainChat runs a non-streaming AUTO request down the chain. exhausted, when
 // non-nil, decides slot token-quota exhaustion. Returns the response, the
 // serving source and the exact model id used; on total failure a *ChainErr
@@ -327,6 +414,65 @@ func (s *Scheduler) ChainChatStream(ctx context.Context, chain *Chain, req *type
 	return chunks, src, model, err
 }
 
+// ChainImage runs an image-generation AUTO request down the chain: tiers
+// ascending (tier 1 highest priority), per-tier round-robin, same-tier order
+// by preference. Each slot's model is pinned to its own image id (ModelFor),
+// so a fallback switches per source. Cooling-down slots are skipped. Returns
+// the response, serving source and the exact model id used; on total failure
+// a *ChainErr summarizing every tier.
+func (s *Scheduler) ChainImage(ctx context.Context, chain *Chain, req *types.ImageGenRequest) (*types.UnifiedResponse, string, string, error) {
+	if chain == nil || len(chain.Tiers) == 0 {
+		return nil, "", "", fmt.Errorf("no image auto slot configured")
+	}
+	var ce ChainErr
+	for _, tn := range chain.Tiers {
+		cands := collectCands(tn.Slots, nil)
+		if len(cands) == 0 {
+			ce.Skipped = append(ce.Skipped, fmt.Sprintf("tier %d: no schedulable slot (cooling)", tn.Tier))
+			continue
+		}
+		// No Pref sort: load balancing is done by round-robin cursor.
+		base := tn.NextStart()
+		norm := normalCount(cands)
+		var hard []TierError
+		for i := 0; i < len(cands); i++ {
+			var c candidate
+			if i < norm {
+				c = cands[(int(base)+i)%norm]
+			} else {
+				c = cands[i]
+			}
+			sl := c.slot
+			if !c.probe && !sl.Prov.ModelAvailable(sl.Model) {
+				continue
+			}
+			r := *req
+			r.Model = sl.Model
+			resp, err := sl.Prov.Image(ctx, &r)
+			if ctx.Err() != nil {
+				releaseProbes(cands)
+				return nil, "", "", ctx.Err()
+			}
+			if err == nil {
+				releaseProbes(cands)
+				return resp, sl.Source, sl.Model, nil
+			}
+			if errors.Is(err, types.ErrBusy) {
+				continue
+			}
+			hard = append(hard, TierError{Tier: tn.Tier, Source: sl.Source, Model: sl.Model, Err: err})
+		}
+		releaseProbes(cands)
+		if len(hard) > 0 {
+			ce.Tiers = append(ce.Tiers, hard...)
+		}
+	}
+	if len(ce.Tiers) == 0 && len(ce.Skipped) == 0 {
+		return nil, "", "", fmt.Errorf("no image auto slot configured")
+	}
+	return nil, "", "", &ce
+}
+
 // ---- direct scheduling ----
 
 // Chat runs a chat request across cands, falling back on failure. Each
@@ -345,11 +491,15 @@ func (s *Scheduler) Chat(ctx context.Context, cands []Provider, req *types.ChatR
 		p := cands[i]
 		r := *req
 		r.Model = p.ModelFor(req.Model)
-		if !p.ModelAvailable(r.Model) {
+		ok, isProbe := p.ModelSchedulable(r.Model)
+		if !ok {
 			lastErr = fmt.Errorf("provider %s: model %q cooling down", p.Name(), r.Model)
 			continue
 		}
 		resp, err := p.Chat(ctx, &r)
+		if isProbe {
+			p.ProbeDone(r.Model)
+		}
 		if ctx.Err() != nil {
 			return nil, "", "", ctx.Err()
 		}
@@ -377,11 +527,15 @@ func (s *Scheduler) ChatStream(ctx context.Context, cands []Provider, req *types
 		p := cands[i]
 		r := *req
 		r.Model = p.ModelFor(req.Model)
-		if !p.ModelAvailable(r.Model) {
+		ok, isProbe := p.ModelSchedulable(r.Model)
+		if !ok {
 			lastErr = fmt.Errorf("provider %s: model %q cooling down", p.Name(), r.Model)
 			continue
 		}
 		resp, err := p.ChatStream(ctx, &r)
+		if isProbe {
+			p.ProbeDone(r.Model)
+		}
 		if err == nil {
 			return resp, p.Name(), r.Model, nil
 		}
@@ -400,11 +554,16 @@ func (s *Scheduler) Image(ctx context.Context, cands []Provider, req *types.Imag
 	var lastErr error
 	for i := 0; i < attempts && i < len(cands); i++ {
 		p := cands[i]
-		if !p.ModelAvailable(p.ModelFor(req.Model)) {
-			lastErr = fmt.Errorf("provider %s: model %q cooling down", p.Name(), p.ModelFor(req.Model))
+		im := p.ModelFor(req.Model)
+		ok, isProbe := p.ModelSchedulable(im)
+		if !ok {
+			lastErr = fmt.Errorf("provider %s: model %q cooling down", p.Name(), im)
 			continue
 		}
 		resp, err := p.Image(ctx, req)
+		if isProbe {
+			p.ProbeDone(im)
+		}
 		if err == nil {
 			return resp, p.Name(), nil
 		}

internal/scheduler/scheduler_test.go

@@ -22,6 +22,14 @@ type fakeProvider struct {
 	busy      atomic.Bool
 	fail      atomic.Bool
 	chatHits  atomic.Int64
+	// probeable makes an unavailable provider eligible as a half-cooldown
+	// probe candidate. probePermit is the single-token permit; probeClaims
+	// counts how many times it was handed out and probeDones how many times it
+	// was returned, so tests can assert the permit is always released.
+	probeable   atomic.Bool
+	probePermit atomic.Bool
+	probeClaims atomic.Int64
+	probeDones  atomic.Int64
 }
 
 func fakeProv(name, model string) *fakeProvider {
@@ -36,6 +44,25 @@ func (f *fakeProvider) ModelFor(reqModel string) string { return f.model }
 
 func (f *fakeProvider) ModelAvailable(model string) bool { return f.available.Load() }
 
+func (f *fakeProvider) ModelSchedulable(model string) (bool, bool) {
+	if f.available.Load() {
+		return true, false
+	}
+	if !f.probeable.Load() {
+		return false, false
+	}
+	if f.probePermit.CompareAndSwap(false, true) {
+		f.probeClaims.Add(1)
+		return true, true
+	}
+	return false, false
+}
+
+func (f *fakeProvider) ProbeDone(model string) {
+	f.probeDones.Add(1)
+	f.probePermit.Store(false)
+}
+
 func (f *fakeProvider) Pref(model string) int64 { return f.pref.Load() }
 
 func (f *fakeProvider) Chat(ctx context.Context, req *types.ChatRequest) (*types.UnifiedResponse, error) {
@@ -177,14 +204,21 @@ func TestChainPreferenceSinksButStaysReachable(t *testing.T) {
 	if err != nil {
 		t.Fatalf("chain: %v", err)
 	}
-	// higher pref (good) is tried first and hard-fails; the pass moves on to
-	// the negative-pref slot, which sinks but stays reachable: with the real
-	// provider its success would RecordSuccess (+1 pref, self-heal)
+	// Round-robin starts at index 0 (neg). neg is available (pref=-5 > -20)
+	// and succeeds. good is never tried because neg already won.
 	if src != "neg" || model != "n" || resp.Content != "neg" {
 		t.Fatalf("served src=%q model=%q content=%q", src, model, resp.Content)
 	}
-	if good.chatHits.Load() == 0 {
-		t.Fatal("higher-pref slot must be tried first")
+	if good.chatHits.Load() != 0 {
+		t.Fatal("neg was tried first and succeeded; good must not be attempted")
+	}
+	// Second request: cursor advances. neg wins again (good hard-fails).
+	resp2, src2, _, err2 := s.ChainChat(context.Background(), ch, chatReq(), nil)
+	if err2 != nil {
+		t.Fatalf("second chain: %v", err2)
+	}
+	if src2 != "neg" || resp2.Content != "neg" {
+		t.Fatalf("second req src=%q content=%q", src2, resp2.Content)
 	}
 }
 
@@ -315,3 +349,149 @@ func TestChainResetCooldownAfterSwap(t *testing.T) {
 		t.Fatalf("first start = %d, want 0", base)
 	}
 }
+
+// ---- half-cooldown probe candidates (plan 阶段 1.3) ----
+
+// TestChainProbeIsLastResort pins the ordering rule: a probe candidate must not
+// steal traffic from a healthy slot in the same tier.
+func TestChainProbeIsLastResort(t *testing.T) {
+	healthy := fakeProv("healthy", "h")
+	cooling := fakeProv("cooling", "c")
+	cooling.available.Store(false)
+	cooling.probeable.Store(true)
+	ch := BuildChain([]Rule{
+		{Tier: 0, Model: "c", Source: "cooling"}, // configured FIRST on purpose
+		{Tier: 0, Model: "h", Source: "healthy"},
+	}, bySource(healthy, cooling))
+	s := New(0)
+	for i := 0; i < 3; i++ {
+		_, src, _, err := s.ChainChat(context.Background(), ch, chatReq(), nil)
+		if err != nil {
+			t.Fatalf("iter %d: %v", i, err)
+		}
+		if src != "healthy" {
+			t.Fatalf("iter %d served by %q; a probe slot must never outrank a healthy one", i, src)
+		}
+	}
+	if cooling.chatHits.Load() != 0 {
+		t.Fatalf("probe slot was hit %d times while a healthy slot existed", cooling.chatHits.Load())
+	}
+	if got := cooling.probeDones.Load(); got != cooling.probeClaims.Load() {
+		t.Fatalf("probe permits leaked: claims=%d dones=%d", cooling.probeClaims.Load(), got)
+	}
+}
+
+// TestChainProbeServesWhenNothingElseCan is the recovery path: with every normal
+// slot gone, the cooling slot's probe carries the request instead of the tier
+// failing outright.
+func TestChainProbeServesWhenNothingElseCan(t *testing.T) {
+	cooling := fakeProv("cooling", "c")
+	cooling.available.Store(false)
+	cooling.probeable.Store(true)
+	ch := BuildChain([]Rule{{Tier: 0, Model: "c", Source: "cooling"}}, bySource(cooling))
+	s := New(0)
+	resp, src, model, err := s.ChainChat(context.Background(), ch, chatReq(), nil)
+	if err != nil {
+		t.Fatalf("probe must serve the request: %v", err)
+	}
+	if src != "cooling" || model != "c" {
+		t.Fatalf("served by src=%q model=%q, want cooling/c", src, model)
+	}
+	if resp == nil || resp.Content != "cooling" {
+		t.Fatalf("bad response: %#v", resp)
+	}
+	if cooling.chatHits.Load() != 1 {
+		t.Fatalf("probe must issue exactly one request, got %d", cooling.chatHits.Load())
+	}
+	if cooling.probeClaims.Load() != 1 || cooling.probeDones.Load() != 1 {
+		t.Fatalf("permit accounting: claims=%d dones=%d, want 1/1", cooling.probeClaims.Load(), cooling.probeDones.Load())
+	}
+}
+
+// TestChainProbePermitReleasedOnFailure: a failed probe must still hand its
+// permit back, otherwise the slot could never be probed again.
+func TestChainProbePermitReleasedOnFailure(t *testing.T) {
+	cooling := fakeProv("cooling", "c")
+	cooling.available.Store(false)
+	cooling.probeable.Store(true)
+	cooling.fail.Store(true)
+	ch := BuildChain([]Rule{{Tier: 0, Model: "c", Source: "cooling"}}, bySource(cooling))
+	s := New(0)
+	if _, _, _, err := s.ChainChat(context.Background(), ch, chatReq(), nil); err == nil {
+		t.Fatal("expected the failing probe to surface an error")
+	}
+	if cooling.probeClaims.Load() != 1 || cooling.probeDones.Load() != 1 {
+		t.Fatalf("permit accounting: claims=%d dones=%d, want 1/1", cooling.probeClaims.Load(), cooling.probeDones.Load())
+	}
+	// permit is free again for the next attempt
+	if _, _, _, err := s.ChainChat(context.Background(), ch, chatReq(), nil); err == nil {
+		t.Fatal("expected the second probe to fail too")
+	}
+	if cooling.probeClaims.Load() != 2 {
+		t.Fatalf("permit must be reclaimable, claims=%d", cooling.probeClaims.Load())
+	}
+}
+
+// TestChainProbeDoesNotBlockTierFallthrough: a cooling tier-1 slot that is not
+// probeable must still let the request fall through to tier 2.
+func TestChainProbeDoesNotBlockTierFallthrough(t *testing.T) {
+	cold := fakeProv("cold", "c")
+	cold.available.Store(false) // cooling and NOT probeable
+	backup := fakeProv("backup", "b")
+	ch := BuildChain([]Rule{
+		{Tier: 1, Model: "c", Source: "cold"},
+		{Tier: 2, Model: "b", Source: "backup"},
+	}, bySource(cold, backup))
+	s := New(0)
+	_, src, _, err := s.ChainChat(context.Background(), ch, chatReq(), nil)
+	if err != nil || src != "backup" {
+		t.Fatalf("want fallthrough to backup, got src=%q err=%v", src, err)
+	}
+	if cold.chatHits.Load() != 0 {
+		t.Fatal("non-probeable cooling slot must not be hit")
+	}
+}
+
+// TestDirectProbeReleasesPermit covers the non-AUTO paths (Chat/ChatStream/Image).
+func TestDirectProbeReleasesPermit(t *testing.T) {
+	cooling := fakeProv("cooling", "m")
+	cooling.available.Store(false)
+	cooling.probeable.Store(true)
+	s := New(0)
+	_, src, _, err := s.Chat(context.Background(), []Provider{cooling}, chatReq())
+	if err != nil || src != "cooling" {
+		t.Fatalf("direct probe must serve: src=%q err=%v", src, err)
+	}
+	if cooling.probeClaims.Load() != 1 || cooling.probeDones.Load() != 1 {
+		t.Fatalf("permit accounting: claims=%d dones=%d", cooling.probeClaims.Load(), cooling.probeDones.Load())
+	}
+}
+
+// TestChainProbeRoundRobinUnaffected: adding a probe tail must not disturb the
+// round-robin rotation over the healthy head.
+func TestChainProbeRoundRobinUnaffected(t *testing.T) {
+	a, b := fakeProv("s1", "a"), fakeProv("s2", "b")
+	cooling := fakeProv("s3", "c")
+	cooling.available.Store(false)
+	cooling.probeable.Store(true)
+	ch := BuildChain([]Rule{
+		{Tier: 0, Model: "a", Source: "s1"},
+		{Tier: 0, Model: "b", Source: "s2"},
+		{Tier: 0, Model: "c", Source: "s3"},
+	}, bySource(a, b, cooling))
+	s := New(0)
+	var got []string
+	for i := 0; i < 4; i++ {
+		_, src, _, err := s.ChainChat(context.Background(), ch, chatReq(), nil)
+		if err != nil {
+			t.Fatalf("iter %d: %v", i, err)
+		}
+		got = append(got, src)
+	}
+	want := []string{"s1", "s2", "s1", "s2"}
+	for i := range want {
+		if got[i] != want[i] {
+			t.Fatalf("rr order = %v, want %v (probe tail must not join the rotation)", got, want)
+		}
+	}
+}

internal/types/types.go

@@ -3,8 +3,10 @@
 package types
 
 import (
+	"bytes"
 	"encoding/json"
 	"errors"
+	"strings"
 	"time"
 )
 
@@ -18,15 +20,19 @@ var ErrBusy = errors.New("provider busy")
 // ---- OpenAI wire request (gateway input) ----
 
 type ChatRequest struct {
-	Model            string             `json:"model"`
-	Messages         []ChatMessage      `json:"messages"`
-	Temperature      *float64           `json:"temperature,omitempty"`
-	MaxTokens        int                `json:"max_tokens,omitempty"`
-	Stream           bool               `json:"stream,omitempty"`
-	Tools            []interface{}      `json:"tools,omitempty"`
-	ToolChoice       interface{}        `json:"tool_choice,omitempty"`
-	DisableThinking  bool               `json:"disable_thinking"`
-	ExtraBody        map[string]interface{} `json:"-"`
+	Model           string                 `json:"model"`
+	Messages        []ChatMessage          `json:"messages"`
+	Temperature     *float64               `json:"temperature,omitempty"`
+	MaxTokens       int                    `json:"max_tokens,omitempty"`
+	Stream          bool                   `json:"stream,omitempty"`
+	Tools           []interface{}          `json:"tools,omitempty"`
+	ToolChoice      interface{}            `json:"tool_choice,omitempty"`
+	DisableThinking bool                   `json:"disable_thinking"`
+	ExtraBody       map[string]interface{} `json:"-"`
+
+	// ClientSession 是客户端自带的会话标识（请求头 x-session-affinity 等）。
+	// 不参与序列化：它只用于让适配器生成稳定的上游会话，不应泄漏给上游 body。
+	ClientSession string `json:"-"`
 }
 
 func (r *ChatRequest) MarshalJSON() ([]byte, error) {
@@ -60,6 +66,39 @@ type ChatMessage struct {
 
 func StringContent(s string) json.RawMessage { b, _ := json.Marshal(s); return b }
 
+// isEmptyJSONArray reports whether raw is the literal empty JSON array [].
+func isEmptyJSONArray(raw json.RawMessage) bool {
+	t := bytes.TrimSpace(raw)
+	return len(t) == 2 && t[0] == '[' && t[1] == ']'
+}
+
+// UnmarshalJSON decodes a chat message, normalising an empty-array content
+// ("content":[]) to an empty string and dropping an empty tool_calls array.
+//
+// Why this exists: Lua adapters json.decode the request and re-encode it, and
+// an empty Lua table is indistinguishable from an empty JSON array — the
+// encoder emits {} for both. Agent clients serialise an assistant turn that
+// carries tool_calls and no text as content:[], so every pass-through adapter
+// turned it into content:{} — a shape that is not valid OpenAI (content is
+// string | array of parts | null) and that real upstreams reject with
+// "400 invalid arguments". Normalising at the decode boundary fixes every
+// adapter at once, including ones added later.
+func (m *ChatMessage) UnmarshalJSON(b []byte) error {
+	type alias ChatMessage
+	var a alias
+	if err := json.Unmarshal(b, &a); err != nil {
+		return err
+	}
+	*m = ChatMessage(a)
+	if isEmptyJSONArray(m.Content) {
+		m.Content = StringContent("")
+	}
+	if isEmptyJSONArray(m.ToolCalls) {
+		m.ToolCalls = nil
+	}
+	return nil
+}
+
 type ToolCall struct {
 	ID        string                 `json:"id"`
 	Type      string                 `json:"type"`
@@ -70,6 +109,7 @@ type ToolCall struct {
 // ---- Unified internal representation (what adapters produce) ----
 
 type UnifiedResponse struct {
+	Model            string     `json:"model,omitempty"`
 	Content          string     `json:"content"`
 	ReasoningContent string     `json:"reasoning_content,omitempty"`
 	FinishReason     string     `json:"finish_reason,omitempty"`
@@ -77,12 +117,83 @@ type UnifiedResponse struct {
 	ToolCalls        []ToolCall `json:"tool_calls,omitempty"`
 	// ImageData used by image-generation adapters.
 	ImageData []ImageData `json:"image_data,omitempty"`
+	// Cost is the upstream-reported charge for this request, passed through
+	// verbatim (OpenCode Zen/Go put a decimal **string** at the response top
+	// level). Deliberately not parsed or summed by the gateway: it is an
+	// upstream fact, and only some upstreams report it at all.
+	Cost string `json:"cost,omitempty"`
 }
 
 type TokenUsage struct {
 	Prompt     int `json:"prompt"`
 	Completion int `json:"completion"`
 	Total      int `json:"total"`
+	// PromptTokensDetails mirrors the OpenAI v2 usage.prompt_tokens_details
+	// object so cache-hit counts reported by OpenAI-compatible upstreams
+	// (and by adapters that normalize their own cache fields into it) pass
+	// through to clients that read it — dsh reads cached_tokens from here.
+	PromptTokensDetails *PromptTokensDetails `json:"prompt_tokens_details,omitempty"`
+	// PromptCacheHit / PromptCacheMiss carry the DeepSeek-legacy standalone
+	// fields; dsh falls back to prompt_cache_hit_tokens when
+	// prompt_tokens_details.cached_tokens is absent.
+	PromptCacheHit  int `json:"prompt_cache_hit_tokens,omitempty"`
+	PromptCacheMiss int `json:"prompt_cache_miss_tokens,omitempty"`
+	// CompletionTokensDetails mirrors the OpenAI v2 usage.completion_tokens_details
+	// object. OpenCode Zen/Go report how much of the completion was thinking
+	// tokens here; without it clients cannot tell how much of the billed output
+	// was reasoning rather than visible answer.
+	CompletionTokensDetails *CompletionTokensDetails `json:"completion_tokens_details,omitempty"`
+}
+
+// CompletionTokensDetails is the OpenAI v2 completion_tokens_details object.
+type CompletionTokensDetails struct {
+	ReasoningTokens int `json:"reasoning_tokens"`
+}
+
+// PromptTokensDetails is the OpenAI v2 prompt_tokens_details object. Only
+// CachedTokens is emitted (omitempty drops the whole object when zero).
+type PromptTokensDetails struct {
+	// CachedTokens is always emitted (even 0) so clients can distinguish
+	// "upstream reports cache, this request missed" from "no cache data".
+	CachedTokens int `json:"cached_tokens"`
+}
+
+// MarshalJSON emits both the legacy short keys (prompt/completion/total, used
+// by the internal unified representation and older clients) and the OpenAI
+// standard keys (prompt_tokens/completion_tokens/total_tokens). Standard
+// clients such as DSH and DevEco Code read the *_tokens fields.
+func (t TokenUsage) MarshalJSON() ([]byte, error) {
+	// Auto-generate prompt_tokens_details from legacy DeepSeek fields when
+	// the upstream adapter only set the standalone hit count (openai.lua
+	// does this in Lua, but other adapters or the standardSSEChunk fallback
+	// may not). dsh reads prompt_tokens_details.cached_tokens first.
+	pdetails := t.PromptTokensDetails
+	if pdetails == nil && t.PromptCacheHit > 0 {
+		pdetails = &PromptTokensDetails{CachedTokens: t.PromptCacheHit}
+	}
+	return json.Marshal(struct {
+		PromptTokens          int                      `json:"prompt_tokens"`
+		CompletionTokens      int                      `json:"completion_tokens"`
+		TotalTokens           int                      `json:"total_tokens"`
+		Prompt                int                      `json:"prompt"`
+		Completion            int                      `json:"completion"`
+		Total                 int                      `json:"total"`
+		PromptTokensDetails   *PromptTokensDetails     `json:"prompt_tokens_details,omitempty"`
+		PromptCacheHitTokens  int                      `json:"prompt_cache_hit_tokens,omitempty"`
+		PromptCacheMissTokens int                      `json:"prompt_cache_miss_tokens,omitempty"`
+		CompletionDetails     *CompletionTokensDetails `json:"completion_tokens_details,omitempty"`
+	}{
+		PromptTokens:          t.Prompt,
+		CompletionTokens:      t.Completion,
+		TotalTokens:           t.Total,
+		Prompt:                t.Prompt,
+		Completion:            t.Completion,
+		Total:                 t.Total,
+		PromptTokensDetails:   pdetails,
+		PromptCacheHitTokens:  t.PromptCacheHit,
+		PromptCacheMissTokens: t.PromptCacheMiss,
+		CompletionDetails:     t.CompletionTokensDetails,
+	})
 }
 
 type ImageData struct {
@@ -112,10 +223,22 @@ type ImageGenResponse struct {
 // streaming tool_calls array (incremental fragments with an index field), which
 // OpenAI-compatible clients accumulate themselves.
 type UnifiedChunk struct {
-	Content          string          `json:"content"`
-	Done             bool            `json:"done"`
+	Content string `json:"content"`
+	Done    bool   `json:"done"`
+	// FinishReason carries the upstream finish/stop reason ("tool_calls",
+	// "length", ...) when the adapter provides it; the gateway emits it on
+	// the terminating chunk instead of the default "stop".
+	FinishReason     string          `json:"finish_reason,omitempty"`
 	ToolCalls        json.RawMessage `json:"tool_calls,omitempty"`
 	ReasoningContent string          `json:"reasoning_content,omitempty"`
+	// Usage carries the upstream token usage when the stream chunk provides
+	// it (OpenAI-style streams attach usage to some chunks; the final one
+	// often has empty choices). Gateway uses it to emit exact usage in the
+	// final stream chunk instead of estimates.
+	Usage *TokenUsage `json:"usage,omitempty"`
+	// Cost is upstream-reported charge for this request (decimal string),
+	// passed through verbatim; OpenCode sends it on its own stream chunk.
+	Cost string `json:"cost,omitempty"`
 }
 
 // Meta passed to Lua build_headers hook
@@ -128,4 +251,15 @@ type BuildMeta struct {
 	Source    map[string]interface{} `json:"source"`
 }
 
-func Now() int64 { return time.Now().Unix() }
+func Now() int64 { return time.Now().Unix() }
+
+// OneLine flattens an error string to a single line capped at n chars.
+// Shared by the provider layer (upstream error reasons) and the gateway
+// layer (client-facing AUTO-chain summaries).
+func OneLine(s string, n int) string {
+	s = strings.Join(strings.Fields(s), " ")
+	if len(s) > n {
+		s = s[:n] + "..."
+	}
+	return s
+}

internal/types/types_test.go

@@ -0,0 +1,154 @@
+package types
+
+import (
+	"encoding/json"
+	"strings"
+	"testing"
+)
+
+func TestTokenUsageMarshalCacheFields(t *testing.T) {
+	// DeepSeek-style: prompt_cache_hit_tokens + prompt_tokens_details auto-generated
+	u := TokenUsage{
+		Prompt:          100,
+		Completion:      50,
+		Total:           150,
+		PromptCacheHit:  80,
+		PromptCacheMiss: 20,
+	}
+	b, _ := json.Marshal(u)
+	var out map[string]interface{}
+	json.Unmarshal(b, &out)
+	check := func(name string, want interface{}) {
+		got := out[name]
+		if got != want {
+			t.Errorf("%s = %v (want %v)", name, got, want)
+		}
+	}
+	check("prompt_tokens", 100.0)
+	check("completion_tokens", 50.0)
+	check("prompt_cache_hit_tokens", 80.0)
+	check("prompt_cache_miss_tokens", 20.0)
+	details := out["prompt_tokens_details"]
+	if details == nil {
+		t.Error("prompt_tokens_details is missing (should be set from prompt_cache_hit_tokens?)")
+	} else {
+		d := details.(map[string]interface{})
+		if d["cached_tokens"] != 80.0 {
+			t.Errorf("cached_tokens = %v (want 80)", d["cached_tokens"])
+		}
+	}
+
+	// OpenAI-style: prompt_tokens_details.cached_tokens
+	u2 := TokenUsage{
+		Prompt:     200,
+		Completion: 100,
+		Total:      300,
+		PromptTokensDetails: &PromptTokensDetails{
+			CachedTokens: 150,
+		},
+	}
+	b2, _ := json.Marshal(u2)
+	json.Unmarshal(b2, &out)
+	check2 := func(name string, want interface{}) {
+		got := out[name]
+		if got != want {
+			t.Errorf("OpenAI %s = %v (want %v)", name, got, want)
+		}
+	}
+	check2("prompt_tokens", 200.0)
+	details2 := out["prompt_tokens_details"]
+	if details2 == nil {
+		t.Error("OpenAI prompt_tokens_details is missing")
+	} else {
+		d := details2.(map[string]interface{})
+		if d["cached_tokens"] != 150.0 {
+			t.Errorf("OpenAI cached_tokens = %v (want 150)", d["cached_tokens"])
+		}
+	}
+
+	t.Logf("DeepSeek JSON: %s", string(b))
+	t.Logf("OpenAI JSON: %s", string(b2))
+}
+
+// TestChatMessageNormalizesEmptyArrayContent pins the fix for the malformed
+// content:{} that pass-through adapters produced for an assistant turn carrying
+// tool_calls with no text. Lua adapters cannot tell an empty JSON array from an
+// empty object, so the decode boundary normalises it for every adapter at once.
+func TestChatMessageNormalizesEmptyArrayContent(t *testing.T) {
+	var msg ChatMessage
+	raw := `{"role":"assistant","content":[],"tool_calls":[{"id":"call_1","type":"function","function":{"name":"read","arguments":"{}"}}]}`
+	if err := json.Unmarshal([]byte(raw), &msg); err != nil {
+		t.Fatal(err)
+	}
+	if string(msg.Content) != `""` {
+		t.Fatalf("empty-array content must normalise to \"\", got %s", msg.Content)
+	}
+	if len(msg.ToolCalls) == 0 {
+		t.Fatal("tool_calls must survive normalisation")
+	}
+
+	// A non-empty content array must be left byte-identical (multimodal path).
+	var mm ChatMessage
+	multi := `{"role":"user","content":[{"type":"text","text":"hi"}]}`
+	if err := json.Unmarshal([]byte(multi), &mm); err != nil {
+		t.Fatal(err)
+	}
+	if string(mm.Content) != `[{"type":"text","text":"hi"}]` {
+		t.Fatalf("multimodal content must pass through untouched, got %s", mm.Content)
+	}
+
+	// A plain string content is untouched.
+	var sm ChatMessage
+	if err := json.Unmarshal([]byte(`{"role":"user","content":"plain"}`), &sm); err != nil {
+		t.Fatal(err)
+	}
+	if string(sm.Content) != `"plain"` {
+		t.Fatalf("string content must pass through untouched, got %s", sm.Content)
+	}
+
+	// An omitted content stays omitted (omitempty semantics preserved).
+	var om ChatMessage
+	if err := json.Unmarshal([]byte(`{"role":"assistant","tool_calls":[]}`), &om); err != nil {
+		t.Fatal(err)
+	}
+	if om.Content != nil {
+		t.Fatalf("absent content must stay absent, got %s", om.Content)
+	}
+	if om.ToolCalls != nil {
+		t.Fatalf("empty tool_calls array must be dropped, got %s", om.ToolCalls)
+	}
+}
+
+// 上游报告输出里有多少是思考 token；缺少这个字段，客户端无法判断
+// completion_tokens 里多少是可见回答、多少是思考（两者都按输出计费）。
+func TestTokenUsageMarshalsCompletionTokensDetails(t *testing.T) {
+	u := TokenUsage{
+		Prompt: 37, Completion: 40, Total: 77,
+		CompletionTokensDetails: &CompletionTokensDetails{ReasoningTokens: 40},
+	}
+	b, err := json.Marshal(u)
+	if err != nil {
+		t.Fatal(err)
+	}
+	var got map[string]interface{}
+	if err := json.Unmarshal(b, &got); err != nil {
+		t.Fatal(err)
+	}
+	ctd, ok := got["completion_tokens_details"].(map[string]interface{})
+	if !ok {
+		t.Fatalf("completion_tokens_details 未输出: %s", b)
+	}
+	if n, _ := ctd["reasoning_tokens"].(float64); int(n) != 40 {
+		t.Errorf("reasoning_tokens = %v, want 40 (%s)", ctd["reasoning_tokens"], b)
+	}
+	// 标准 OpenAI 字段同时存在，老客户端不受影响。
+	if _, ok := got["completion_tokens"]; !ok {
+		t.Errorf("标准 completion_tokens 丢失: %s", b)
+	}
+
+	// 未设置时不得凭空造字段。
+	b2, _ := json.Marshal(TokenUsage{Prompt: 1, Completion: 2, Total: 3})
+	if strings.Contains(string(b2), "completion_tokens_details") {
+		t.Errorf("无数据时不应输出 completion_tokens_details: %s", b2)
+	}
+}

packaging/core-dist.sh

@@ -0,0 +1,90 @@
+#!/usr/bin/env bash
+# Build the llmsproxy CORE distribution packages (deb + rpm + tar.gz).
+#
+# Unlike the GUI packages (Electron), the core is a single static-ish binary
+# that server admins install with dpkg/rpm. The systemd unit bakes in the
+# memory tuning (GOGC=50, MALLOC_ARENA_MAX=2) measured on the production box.
+#
+# Requires: go, nfpm (PATH), dpkg-deb/rpmbuild for verification
+# Outputs:  cmd/build/dist/llmsproxy_<ver>_amd64.deb
+#           cmd/build/dist/llmsproxy-<ver>.x86_64.rpm
+#           cmd/build/dist/llmsproxy-<ver>-linux-amd64.tar.gz
+set -euo pipefail
+
+ROOT="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)"
+cd "$ROOT"
+
+VERSION="${1:-$(git describe --tags --dirty 2>/dev/null | sed 's/^v//' || echo 0.1.0)}"
+DIST="$ROOT/cmd/build/dist"
+BUILD_BIN="$ROOT/cmd/build/llmsproxy"
+
+log() { printf '\033[1;34m[core-dist]\033[0m %s\n' "$*"; }
+
+# nfpm is installed via `go install` -> lives in GOPATH/bin, which may not be
+# on the caller's PATH (non-interactive shells, cron). Resolve it explicitly.
+NFPM_BIN="$(command -v nfpm || true)"
+[ -z "$NFPM_BIN" ] && [ -n "$(go env GOPATH)" ] && {
+  [ -x "$(go env GOPATH)/bin/nfpm" ] && NFPM_BIN="$(go env GOPATH)/bin/nfpm"
+}
+[ -n "$NFPM_BIN" ] || {
+  echo "[core-dist] nfpm missing. Install: go install github.com/goreleaser/nfpm/v2/cmd/nfpm@v2.41.2"
+  exit 1
+}
+
+mkdir -p "$DIST"
+
+# 1) build the core binary (luajit tag, -s -w like deploy.sh)
+log "building core binary (luajit, stripped)..."
+CGO_ENABLED=1 go build -tags luajit -trimpath -ldflags="-s -w" -o "$BUILD_BIN" ./cmd/llmsproxy
+
+# 2) sanity gates: binary must contain the expected symbols and not be degenerate.
+#    NOTE under `set -o pipefail`, `strings | grep -q` is BROKEN: grep -q exits
+#    on first hit and closes the pipe, strings gets SIGPIPE -> pipeline returns
+#    141 -> the check spuriously fails. Stage strings output in a temp file.
+SYMS_TMP="$(mktemp)"
+strings "$BUILD_BIN" > "$SYMS_TMP"
+for sym in ProbeSlots reqRing shrinkStepLocked mergeUsage; do
+  grep -qF "$sym" "$SYMS_TMP" || {
+    echo "[core-dist] FATAL: binary missing expected symbol '$sym'"; rm -f "$SYMS_TMP"; exit 1
+  }
+done
+rm -f "$SYMS_TMP"
+[ "$(stat -c%s "$BUILD_BIN")" -gt 5000000 ] || {
+  echo "[core-dist] FATAL: core binary implausibly small"; exit 1
+}
+log "  ok: $(stat -c%s "$BUILD_BIN") bytes, symbols present"
+
+# 3) deb + rpm via nfpm. nfpm v2 here does not render {{ .Env.VERSION }}, so
+#    stamp the version into a throwaway copy of the config; -t places the
+#    generated package in the dist dir.
+log "packaging deb + rpm (v$VERSION)..."
+NFPM_CFG="$(mktemp)" && sed "s/__VERSION__/$VERSION/g" "$ROOT/packaging/nfpm.yaml" > "$NFPM_CFG"
+"$NFPM_BIN" package --config "$NFPM_CFG" --packager deb -t "$DIST"
+"$NFPM_BIN" package --config "$NFPM_CFG" --packager rpm -t "$DIST"
+rm -f "$NFPM_CFG"
+
+# 4) tar.gz
+log "packaging tar.gz..."
+TAR_DIR="$DIST/llmsproxy-$VERSION-linux-amd64"
+rm -rf "$TAR_DIR"; mkdir -p "$TAR_DIR"
+cp "$BUILD_BIN" "$TAR_DIR/llmsproxy"
+cp "$ROOT/config.yaml" "$TAR_DIR/config.example.yaml"
+cp "$ROOT/packaging/llmsproxy.service" "$TAR_DIR/llmsproxy.service"
+mkdir -p "$TAR_DIR/adapters"
+cp "$ROOT"/internal/lua/adapters/*.lua "$TAR_DIR/adapters/"
+tar -C "$DIST" -czf "$DIST/llmsproxy-$VERSION-linux-amd64.tar.gz" \
+  "$(basename "$TAR_DIR")"
+rm -rf "$TAR_DIR"
+
+# 5) verify artifacts are healthy (the same gate that would have caught the
+#    264KB exe fiasco)
+log "verifying packages..."
+for f in "$DIST"/llmsproxy_*.deb "$DIST"/llmsproxy-*.rpm "$DIST"/*.tar.gz; do
+  [ -f "$f" ] || continue
+  sz=$(stat -c%s "$f")
+  [ "$sz" -gt 100000 ] || { echo "[core-dist] FATAL: $f suspiciously small ($sz B)"; exit 1; }
+  log "  ✓ $f ($((sz/1024)) KB)"
+done
+dpkg-deb -I "$DIST"/llmsproxy_*.deb 2>/dev/null | grep -E 'Package|Version' | head -2
+
+log "done: $DIST"

packaging/llmsproxy.service

@@ -0,0 +1,24 @@
+[Unit]
+Description=LLMSProxy - unified OpenAI-compatible multi-source LLM gateway
+After=network.target
+
+[Service]
+Type=simple
+# Memory tuning (measured, see README "内存占用"):
+#   MALLOC_ARENA_MAX=2 caps glibc per-thread malloc arenas. LuaJIT allocates
+#     through cgo -> glibc malloc, and glibc defaults to 8*nproc arenas, so every
+#     OS thread that touches malloc reserved its own ~1 MB arena that is never
+#     returned. Measured: 8-12 arenas -> 0.
+#   GOGC=50 halves the Go heap growth target. On its own it does NOT help (the
+#     saved heap is immediately eaten by more glibc arenas); combined with
+#     MALLOC_ARENA_MAX it cut settled RSS by ~19%. This gateway is I/O bound, so
+#     the extra GC cycles are free.
+Environment=GOGC=50
+Environment=MALLOC_ARENA_MAX=2
+ExecStart=/usr/bin/llmsproxy -config /etc/llmsproxy/config.yaml
+WorkingDirectory=/etc/llmsproxy
+Restart=always
+RestartSec=5
+
+[Install]
+WantedBy=multi-user.target

packaging/nfpm.yaml

@@ -0,0 +1,59 @@
+# nfpm config for the llmsproxy core gateway binary.
+# The version placeholder below is substituted by packaging/core-dist.sh (nfpm v2
+# here does not honour {{ .Env.VERSION }} template rendering), so do not edit it
+# by hand.
+name: llmsproxy
+arch: amd64
+platform: linux
+version: __VERSION__
+epoch: "1"
+section: net
+priority: optional
+maintainer: JianFeeeee <dev@modelrouter.local>
+description: |
+  Unified OpenAI-compatible multi-source LLM gateway.
+  Routes requests across multiple upstream LLM providers with Lua adapters,
+  model scheduling, quota windows, an elastic Lua worker pool, and a Web UI.
+vendor: ModelRouter
+homepage: https://gitcode.com/JianFeeeee/ModelRouter
+license: MIT
+contents:
+  # binary
+  - src: ./cmd/build/llmsproxy
+    dst: /usr/bin/llmsproxy
+    file_info:
+      mode: 0755
+  # systemd unit (memory tuning baked in)
+  - src: ./packaging/llmsproxy.service
+    dst: /etc/systemd/system/llmsproxy.service
+    file_info:
+      mode: 0644
+  # example config (seeded to /etc/llmsproxy on first install by postinst)
+  - src: ./config.yaml
+    dst: /usr/share/llmsproxy/config.example.yaml
+    file_info:
+      mode: 0644
+  # built-in Lua adapters (deployed to /etc/llmsproxy/adapters)
+  - src: ./internal/lua/adapters/*.lua
+    dst: /usr/share/llmsproxy/adapters/
+    file_info:
+      mode: 0644
+  # README pointer
+  - src: ./README.md
+    dst: /usr/share/doc/llmsproxy/README.md
+    file_info:
+      mode: 0644
+
+scripts:
+  postinstall: ./packaging/postinst.sh
+  preremove: ./packaging/prerm.sh
+
+rpm:
+  group: Applications/Internet
+  summary: Unified OpenAI-compatible multi-source LLM gateway
+  compression: gzip
+
+deb:
+  compression: gzip
+  fields:
+    Recommends: ca-certificates

packaging/postinst.sh

@@ -0,0 +1,22 @@
+#!/bin/sh
+# llmsproxy core package postinst
+set -e
+
+# Runtime state belongs under /etc/llmsproxy just like the production layout:
+# runtime.json holds keys/created sources and is NOT touched on reinstall.
+mkdir -p /etc/llmsproxy/adapters
+
+if [ -f /etc/llmsproxy/config.yaml ]; then
+    echo "llmsproxy: keeping existing /etc/llmsproxy/config.yaml"
+else
+    # seed a starter config from the packaged example
+    cp -f /usr/share/llmsproxy/config.example.yaml /etc/llmsproxy/config.yaml 2>/dev/null || true
+    chmod 0644 /etc/llmsproxy/config.yaml
+fi
+
+# systemd
+if command -v systemctl >/dev/null 2>&1; then
+    systemctl daemon-reload >/dev/null 2>&1 || true
+    echo "llmsproxy: use 'systemctl enable --now llmsproxy' to start"
+fi
+exit 0

packaging/prerm.sh

@@ -0,0 +1,14 @@
+#!/bin/sh
+# llmsproxy core package prerm
+set -e
+
+if command -v systemctl >/dev/null 2>&1; then
+    # stop the service if this package's unit is the one installed
+    if systemctl list-unit-files llmsproxy.service >/dev/null 2>&1 && \
+       systemctl is-active llmsproxy.service >/dev/null 2>&1; then
+        systemctl stop llmsproxy.service >/dev/null 2>&1 || true
+        systemctl disable llmsproxy.service >/dev/null 2>&1 || true
+    fi
+    systemctl daemon-reload >/dev/null 2>&1 || true
+fi
+exit 0

packaging/verify-dist.sh

@@ -0,0 +1,49 @@
+#!/usr/bin/env bash
+# Verify packaged GUI artifacts are healthy before a release is declared done.
+#
+# Background: on 2026-08-28 (1.3.0) the host wine broke (missing syswow64, so
+# wine -- the NSIS self-extract step -- failed), and electron-builder silently
+# wrote a 264 KB installer shell that did NOT embed the 86 MB payload. No gate
+# caught it, so a broken Windows installer shipped. This script is that gate.
+#
+# It enforces a per-artifact minimum size, so a degenerate installer (or a
+# build that only emitted a shell without its embedded payload) fails loudly.
+# A size check is the right signal here: a healthy NSIS exe is ~86 MB and the
+# broken 1.3.0 one was 264 KB (326x) — electron-builder compresses the embedded
+# app.7z, so a plaintext-payload string probe would be unreliable.
+set -uo pipefail
+
+GUI_DIST="${1:-cmd/build/gui-dist}"
+VERSION="${2:-}"
+
+log() { printf '[verify-dist] %s\n' "$*"; }
+fail() { printf '[verify-dist] FATAL: %s\n' "$*" >&2; exit 1; }
+
+[ -d "$GUI_DIST" ] || fail "no dist dir: $GUI_DIST"
+
+# (glob, min-bytes) — nsis.7z is NOT a required artifact: with `--win nsis` the
+# payload is embedded inside the Setup exe and only the exe is the formal product
+# (the .nsis.7z appears in intermediate/--dir builds only).
+checks=(
+  "ModelRouter*.AppImage:10000000"
+  "modelrouter-*.deb:10000000"
+  "modelrouter-*.rpm:10000000"
+  "ModelRouter Setup*.exe:5000000"
+)
+
+for entry in "${checks[@]}"; do
+  glob="${entry%%:*}"; min="${entry#*:}"
+  # gui-dist accumulates installers from previous versions; check the MOST
+  # RECENT artifact (last in sort) so a stale file never masks a broken build.
+  # `find -name` (not a bare glob) so the space in "ModelRouter Setup *.exe"
+  # is preserved as one file pattern.
+  match="$(find "$GUI_DIST" -maxdepth 1 -name "$glob" -type f | sort -V | tail -1)"
+  [ -z "$match" ] && fail "no artifact matching '$glob' in $GUI_DIST"
+  sz=$(stat -c%s "$match")
+  if [ "$sz" -lt "$min" ]; then
+    fail "$(basename "$match") is only $sz bytes (min $min) — likely a degenerate shell like the 1.3.0 264KB exe"
+  fi
+  log "  ✓ $(basename "$match") ($((sz/1024)) KB)"
+done
+
+log "all artifacts healthy"