chore(version): 1.7.4 -> 1.7.5

启动密钥警告误报的修复。

README.md

@@ -24,6 +24,7 @@
 ### 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
@@ -178,6 +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：
@@ -344,7 +420,10 @@ Environment=MALLOC_ARENA_MAX=2
   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。

README_EN.md

@@ -38,6 +38,11 @@ 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
@@ -171,11 +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:

cmd/gui/package-lock.json

@@ -1,12 +1,12 @@
 {
   "name": "modelrouter-gui",
-  "version": "1.4.2",
+  "version": "1.7.5",
   "lockfileVersion": 3,
   "requires": true,
   "packages": {
     "": {
       "name": "modelrouter-gui",
-      "version": "1.4.2",
+      "version": "1.7.5",
       "license": "Proprietary",
       "devDependencies": {
         "electron": "^33.0.0",
@@ -582,7 +582,7 @@
       }
     },
     "node_modules/@peculiar/webcrypto": {
-      "version": "1.7.1",
+      "version": "1.7.5",
       "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": "1.7.5",
       "resolved": "https://registry.npmjs.org/resedit/-/resedit-1.7.2.tgz",
       "integrity": "sha512-vHjcY2MlAITJhC0eRD/Vv8Vlgmu9Sd3LX9zZvtGzU5ZImdTN3+d6e/4mnTyV8vEbyf1sgNIrWxhWlrys52OkEA==",
       "dev": true,

cmd/gui/package.json

@@ -1,6 +1,6 @@
 {
   "name": "modelrouter-gui",
-  "version": "1.4.2",
+  "version": "1.7.5",
   "description": "ModelRouter Desktop — embedded ModelRouter core with tray",
   "author": "ModelRouter",
   "main": "main.js",

cmd/llmsproxy/main.go

@@ -24,6 +24,7 @@ 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
@@ -41,6 +42,20 @@ func main() {
 		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)
@@ -58,19 +73,26 @@ func main() {
 		}
 	}
 
-	gw, err := gateway.New(c, c.GatewayKeys())
+	gw, err := gateway.New(c)
 	if err != nil {
 		log.Fatalf("[llmsproxy] gateway: %v", err)
 	}
 
 	// Ops hygiene: surface the two most common footguns instead of silently
 	// running with them.
-	if keys := c.GatewayKeys(); len(keys) == 0 {
-		log.Printf("[llmsproxy] WARNING: gateway_keys is EMPTY — without a key every request is rejected")
+	//
+	// Read the AUTHORITATIVE source. Auth uses core.ListKeys() (cfg.Keys), not
+	// the legacy cfg.GatewayKeys list: seedKeys copies gateway_keys into keys[]
+	// on first start and the YAML list is ignored for auth afterwards. Checking
+	// GatewayKeys() here made the warning lie — once the legacy list was empty
+	// (e.g. after rotating the starter key away) it reported "every request is
+	// rejected" while seven working keys, one of them admin, were in service.
+	if keys := c.ListKeys(); len(keys) == 0 {
+		log.Printf("[llmsproxy] WARNING: no gateway key configured — every request will be rejected. Add one in the WebUI or set gateway_keys")
 	} else {
 		for _, k := range keys {
-			if k == "sk-gw-local-0001" || k == "sk-local-0001" {
-				log.Printf("[llmsproxy] WARNING: gateway key %q looks like the starter/example key — rotate it before exposing the gateway", k)
+			if k.Key == "sk-gw-local-0001" || k.Key == "sk-local-0001" {
+				log.Printf("[llmsproxy] WARNING: gateway key %q looks like the starter/example key — rotate it before exposing the gateway", k.Key)
 			}
 		}
 	}

deploy.sh

@@ -244,6 +244,63 @@ verify_config() {
     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: $key_file"
+    err "  常见原因：master.key 被重新生成、恢复了不匹配的备份，或换了机器。"
+    return 1
+}
+
 # ---------- 同步适配器 ----------
 sync_adapters() {
     log "同步适配器"
@@ -393,6 +450,10 @@ main() {
     parse_args "$@"
     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

e2e/e2e_test.go

@@ -326,3 +326,144 @@ func statusOf(resp *http.Response) int {
 	}
 	return resp.StatusCode
 }
+
+// The startup warning about key configuration must describe the key set that
+// AUTH actually uses (cfg.Keys, via core.ListKeys), not the legacy
+// gateway_keys YAML list.
+//
+// It used to check GatewayKeys() only. seedKeys copies gateway_keys into keys[]
+// on first start and the YAML list stops being consulted, so once that list was
+// emptied — e.g. after rotating the starter key away — the process logged
+// "every request will be rejected" while seven working keys, one of them admin,
+// were in service. Observed in production: the warning appeared and that same
+// key returned HTTP 200 on /v1/models.
+//
+// The fixture reproduces exactly that shape: gateway_keys EMPTY, with the
+// working key present only in keys[].
+func TestStartupWarningReflectsRealKeysNotLegacyList(t *testing.T) {
+	up := newMockUpstream(t)
+	dir := t.TempDir()
+
+	// Seed a real install once so keys[] gets populated from gateway_keys, then
+	// rewrite the file to look like a rotated install: legacy list cleared, the
+	// seeded key still present in keys[].
+	path := writeConfig(t, dir, "127.0.0.1:0", map[string]*mockUpstream{"good": up})
+	bin := buildBinary(t)
+
+	// First run creates the sealed config with keys[] populated.
+	seedPort := freePort(t)
+	cfg1 := writeConfig(t, dir, "127.0.0.1:"+seedPort, map[string]*mockUpstream{"good": up})
+	g1 := startGateway(t, bin, "127.0.0.1:"+seedPort, cfg1)
+	g1.stop()
+
+	data, err := os.ReadFile(cfg1)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !strings.Contains(string(data), "keys:") {
+		t.Fatalf("first run did not populate keys[]:\n%s", data)
+	}
+
+	// Now clear the legacy gateway_keys list, keeping keys[] intact. The
+	// rewritten file may use different indentation, so match loosely.
+	cleared, n := clearGatewayKeys(string(data))
+	if n == 0 {
+		t.Fatalf("could not find a gateway_keys entry to clear in:\n%s", data)
+	}
+	if err := os.WriteFile(cfg1, []byte(cleared), 0o644); err != nil {
+		t.Fatal(err)
+	}
+
+	// Start again and inspect the log. The config pins its listen address, so
+	// rewrite that field to the port we will actually poll.
+	port2 := freePort(t)
+	rehosted, ok := replaceListen(string(cleared), "127.0.0.1:"+port2)
+	if !ok {
+		t.Fatalf("could not rewrite listen in:\n%s", cleared)
+	}
+	cfg2 := filepath.Join(dir, "config2.yaml")
+	if err := os.WriteFile(cfg2, []byte(rehosted), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	g2 := startGateway(t, bin, "127.0.0.1:"+port2, cfg2)
+	defer g2.stop()
+
+	// Sanity: the key really does authenticate, so the warning would be a lie.
+	resp, _ := g2.do("GET", "/v1/models", "", true)
+	if resp == nil || resp.StatusCode != http.StatusOK {
+		t.Fatalf("key from keys[] did not authenticate (status %v); test fixture is wrong",
+			resp)
+	}
+
+	logs := g2.logString()
+	if strings.Contains(logs, "gateway_keys is EMPTY") {
+		t.Errorf("startup logged 'gateway_keys is EMPTY ... every request is rejected' while "+
+			"a working admin key was in service — the warning must read the authoritative "+
+			"key set (core.ListKeys), not the legacy YAML list.\nlog:\n%s", logs)
+	}
+	_ = path
+	_ = g1
+}
+
+func freePort(t *testing.T) string {
+	t.Helper()
+	l, err := net.Listen("tcp", "127.0.0.1:0")
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer l.Close()
+	_, port, _ := net.SplitHostPort(l.Addr().String())
+	return port
+}
+
+// stop terminates the gateway and waits, so its log buffer is complete.
+func (g *gatewayUnderTest) stop() {
+	if g == nil || g.cmd == nil || g.cmd.Process == nil {
+		return
+	}
+	_ = g.cmd.Process.Kill()
+	_, _ = g.cmd.Process.Wait()
+}
+
+// logString returns everything the gateway has written so far.
+func (g *gatewayUnderTest) logString() string { return g.log.String() }
+
+// clearGatewayKeys empties the gateway_keys list in a config, tolerating any
+// indentation the YAML writer chose. Returns the new text and how many entries
+// it removed.
+func clearGatewayKeys(src string) (string, int) {
+	lines := strings.Split(src, "\n")
+	var out []string
+	removed := 0
+	inList := false
+	for _, l := range lines {
+		trimmed := strings.TrimSpace(l)
+		if strings.HasPrefix(l, "gateway_keys:") {
+			out = append(out, "gateway_keys: []")
+			inList = true
+			removed++
+			continue
+		}
+		if inList {
+			if strings.HasPrefix(trimmed, "- ") {
+				removed++
+				continue
+			}
+			inList = false
+		}
+		out = append(out, l)
+	}
+	return strings.Join(out, "\n"), removed
+}
+
+// replaceListen rewrites the top-level listen: value.
+func replaceListen(src, addr string) (string, bool) {
+	lines := strings.Split(src, "\n")
+	for i, l := range lines {
+		if strings.HasPrefix(l, "listen:") {
+			lines[i] = "listen: " + addr
+			return strings.Join(lines, "\n"), true
+		}
+	}
+	return src, false
+}

internal/config/config.go

@@ -27,6 +27,11 @@ type Config struct {
 	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.
@@ -192,10 +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
@@ -253,11 +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)
 	}
@@ -361,15 +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"`
 }
@@ -388,6 +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"`
 	}
@@ -398,6 +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/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/core/core.go

@@ -7,6 +7,7 @@ import (
 	"crypto/rand"
 	"encoding/hex"
 	"fmt"
+	"log"
 	"os"
 	"path/filepath"
 	"sort"
@@ -53,6 +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)
 	}
@@ -74,6 +97,12 @@ func NewFromConfig(cfg *config.Config) (*Core, error) {
 	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
 }
 
@@ -238,10 +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
@@ -254,9 +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
@@ -264,10 +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 != "" {
@@ -439,6 +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 != "" {
@@ -447,7 +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
 }
@@ -655,7 +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.
@@ -756,3 +803,20 @@ func (c *Core) Close() {
 		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/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

@@ -7,6 +7,8 @@ import (
 	"fmt"
 	"log"
 	"net/http"
+	"regexp"
+	"strconv"
 	"strings"
 	"sync/atomic"
 	"time"
@@ -28,6 +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.
@@ -38,6 +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 {
@@ -62,6 +71,10 @@ type ChatChunk struct {
 	// 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 {
@@ -115,13 +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
 		}
 	}
@@ -152,20 +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{}
@@ -179,48 +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" /
@@ -229,18 +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"), ""
@@ -287,6 +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
@@ -301,8 +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()
@@ -317,6 +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)),
@@ -354,8 +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()
@@ -372,6 +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)),
@@ -517,21 +609,115 @@ func upstreamErrStatus(err error) int {
 	return http.StatusBadGateway
 }
 
+// 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
+		}
+	}
+	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 {
-			parts = append(parts, types.OneLine(fmt.Sprintf("%s/%s: %v", t.Source, t.Model, t.Err), 80))
+			// 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, 80))
+			parts = append(parts, types.OneLine(sk, 160))
 		}
 		msg := strings.Join(parts, "; ")
 		if len(msg) > 300 {
@@ -606,6 +792,7 @@ func writeChatCompletion(w http.ResponseWriter, resp *types.UnifiedResponse, mod
 		Created: time.Now().Unix(),
 		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
@@ -725,11 +912,15 @@ func (g *Gateway) pumpStream(w http.ResponseWriter, rec *Req, chunks <-chan type
 	// 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: modelName,
 		}
@@ -766,6 +957,19 @@ func (g *Gateway) pumpStream(w http.ResponseWriter, rec *Req, chunks <-chan type
 	// 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
@@ -796,6 +1000,14 @@ func (g *Gateway) pumpStream(w http.ResponseWriter, rec *Req, chunks <-chan type
 			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")
@@ -903,8 +1115,8 @@ func (g *Gateway) handleImage(w http.ResponseWriter, r *http.Request) {
 	}
 	if isAuto(model) {
 		if chain := g.core.AutoImageChain(); chain != nil && len(chain.Tiers) > 0 {
-			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
 			}
 			done := g.stats.Begin()
@@ -926,7 +1138,10 @@ func (g *Gateway) handleImage(w http.ResponseWriter, r *http.Request) {
 			}
 			rec.OK = true
 			rec.Status = http.StatusOK
-			rec.Compl = int64(len(resp.ImageData))
+			// 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(),
@@ -950,8 +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()
@@ -975,7 +1190,8 @@ func (g *Gateway) handleImage(w http.ResponseWriter, r *http.Request) {
 	}
 	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/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

@@ -58,7 +58,7 @@ func newTestGateway(t *testing.T, srcs ...config.Source) *Gateway {
 		t.Fatalf("core: %v", err)
 	}
 	t.Cleanup(c.Close)
-	g, err := New(c, []string{"sk-test"})
+	g, err := New(c)
 	if err != nil {
 		t.Fatalf("gateway: %v", err)
 	}
@@ -670,8 +670,11 @@ 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")
 	}
 }
 

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)
+	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)
+	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)
+	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)
+	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

@@ -48,13 +48,7 @@ func (g *Gateway) handleKeysAPI(w http.ResponseWriter, r *http.Request) {
 			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())
@@ -76,6 +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())

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

@@ -115,7 +115,7 @@ func (g *Gateway) loginRecord(ip string, ok bool, now int64) {
 	}
 }
 
-func New(c *core.Core, gatewayKeys []string) (*Gateway, error) {
+func New(c *core.Core) (*Gateway, error) {
 	sub, err := fs.Sub(uiFS, "ui")
 	if err != nil {
 		return nil, err
@@ -222,6 +222,8 @@ 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/"):

internal/gateway/stats.go

@@ -48,6 +48,10 @@ type Req struct {
 	Status int `json:"status"`
 	// Err short error message
 	Err string `json:"error,omitempty"`
+	// ImageCount is how many images an image-generation request produced.
+	// Kept separate from Compl/Prompt: image generation has no token concept,
+	// so counting images as "completion tokens" would corrupt the token totals.
+	ImageCount int `json:"image_count,omitempty"`
 }
 
 // Stat aggregates counters for one dimension row.
@@ -84,6 +88,21 @@ type Stats struct {
 	auditPath     string
 	replayPartial bool                       // aggregates built from a bounded audit tail
 	modelHour     map[string]map[int64]int64 // model -> unix-hour bucket -> tokens
+	// keyModelHour buckets the same tokens as modelHour but keyed by
+	// (gateway key id, model) so per-key model quotas are isolated from each
+	// other. modelHour stays key-blind on purpose: it backs the AUTO slot
+	// quota, which is a gateway-wide limit on a slot, not a per-key one.
+	keyModelHour map[string]map[string]map[int64]int64 // key -> model -> unix-hour -> tokens
+	// keyHour buckets a key's total tokens per unix hour, backing the
+	// whole-key quota (all models of one key share one budget).
+	keyHour map[string]map[int64]int64
+	// keyReqHour buckets a key's request count per unix hour, backing the
+	// key-wide request quota. Counted for every request type (chat, stream,
+	// image) including failed ones, so a failing client cannot loop for free.
+	keyReqHour map[string]map[int64]int64
+	// pinnedWanted records which (key, model) pairs have a quota that pins a
+	// source, so the pinned bucket is maintained only where it is read.
+	pinnedWanted map[string]bool
 }
 
 const hourSec = 3600
@@ -116,14 +135,18 @@ func NewStats(maxRecords int) *Stats {
 		maxRecords = defaultRingSize
 	}
 	return &Stats{
-		byKey:      map[string]*Stat{},
-		byModel:    map[string]*Stat{},
-		bySrc:      map[string]*Stat{},
-		byKeyModel: map[string]map[string]*Stat{},
-		byKeySrc:   map[string]map[string]*Stat{},
-		byStatus:   map[int]*Stat{},
-		modelHour:  map[string]map[int64]int64{},
-		maxRecs:    maxRecords,
+		byKey:        map[string]*Stat{},
+		byModel:      map[string]*Stat{},
+		bySrc:        map[string]*Stat{},
+		byKeyModel:   map[string]map[string]*Stat{},
+		byKeySrc:     map[string]map[string]*Stat{},
+		byStatus:     map[int]*Stat{},
+		modelHour:    map[string]map[int64]int64{},
+		keyModelHour: map[string]map[string]map[int64]int64{},
+		keyHour:      map[string]map[int64]int64{},
+		keyReqHour:   map[string]map[int64]int64{},
+		pinnedWanted: map[string]bool{},
+		maxRecs:      maxRecords,
 	}
 }
 
@@ -318,6 +341,9 @@ func (s *Stats) Record(r Req) {
 // aggregateLocked folds r into every aggregate row and the quota window
 // bucket. Caller must hold s.mu.
 func (s *Stats) aggregateLocked(r Req) {
+	if r.Key != "" {
+		s.addKeyReqLocked(r.Key, (r.Time/1000)/hourSec, 1)
+	}
 	inc(s.byKey, r.Key, r)
 	if r.Model != "" {
 		inc(s.byModel, r.Model, r)
@@ -362,6 +388,12 @@ func (s *Stats) aggregateLocked(r Req) {
 			s.modelHour[key] = hm
 		}
 		hm[h] += tok
+		// per-key buckets: same token split, but isolated per key so one
+		// key's quota cannot be drained by another key's traffic
+		if r.Key != "" {
+			s.addKeyTokenLocked(r.Key, r.Model, r.Source, h, tok)
+			s.addKeyHourLocked(r.Key, h, tok)
+		}
 		// retention: 24*40 = 960 hourly buckets ≈ 40 days of history (covers
 		// the longest "month" quota window)
 		if len(hm) > 24*40 {
@@ -374,6 +406,185 @@ func (s *Stats) aggregateLocked(r Req) {
 	}
 }
 
+// quotaRetentionHours is how much hourly history the per-key quota buckets
+// keep. It matches the modelHour retention (40 days) so the longest "month"
+// window is fully covered after a restart, and it is deliberately applied per
+// key so an idle key's buckets are reclaimed instead of pinning memory.
+const quotaRetentionHours = 24 * 40
+
+// addKeyTokenLocked adds tok to one key's (model[, source]) hourly bucket.
+//
+// The bare-model bucket is always written: it is what a scope entry without a
+// source pin reads, and writing it is what makes that quota work at all (every
+// recorded request carries a resolved source, so a source::model-only bucket
+// would leave an unpinned quota reading an empty bucket forever).
+//
+// The pinned "source::model" bucket is written only for (key, model) pairs
+// some scope entry actually pins a source for. Maintaining it unconditionally
+// doubles 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 budget — to
+// serve a lookup that almost nobody performs.
+func (s *Stats) addKeyTokenLocked(key, model, source string, h, tok int64) {
+	if model == "" {
+		return
+	}
+	byModel := s.keyModelHour[key]
+	if byModel == nil {
+		byModel = map[string]map[int64]int64{}
+		s.keyModelHour[key] = byModel
+	}
+	addBucket(byModel, model, h, tok)
+	if source == "" {
+		return
+	}
+	pinned := source + "::" + model
+	if pinned == model || !s.pinnedWanted[key+"\x00"+model] {
+		return
+	}
+	addBucket(byModel, pinned, h, tok)
+}
+
+// addBucket adds tok to one named hourly bucket and enforces retention.
+func addBucket(byModel map[string]map[int64]int64, name string, h, tok int64) {
+	hm := byModel[name]
+	if hm == nil {
+		hm = map[int64]int64{}
+		byModel[name] = hm
+	}
+	hm[h] += tok
+	if len(hm) > quotaRetentionHours {
+		for k := range hm {
+			if k < h-quotaRetentionHours {
+				delete(hm, k)
+			}
+		}
+	}
+}
+
+// wantPinnedBuckets starts maintaining the pinned (source::model) bucket for a
+// key's model, so a quota that pins a source counts traffic from the moment it
+// is configured. Usage recorded before that cannot be attributed to a source
+// after the fact, so a pinned quota may under-count its first window.
+func (s *Stats) wantPinnedBuckets(key, model string) {
+	if key == "" || model == "" {
+		return
+	}
+	s.pinnedWanted[key+"\x00"+model] = true
+}
+
+// addKeyHourLocked adds tok to one key's all-model hourly total.
+func (s *Stats) addKeyHourLocked(key string, h, tok int64) {
+	hm := s.keyHour[key]
+	if hm == nil {
+		hm = map[int64]int64{}
+		s.keyHour[key] = hm
+	}
+	hm[h] += tok
+	if len(hm) > quotaRetentionHours {
+		for k := range hm {
+			if k < h-quotaRetentionHours {
+				delete(hm, k)
+			}
+		}
+	}
+}
+
+// KeyWindowReqs returns how many requests one key issued within sec seconds;
+// sec <= 0 means all retained history.
+func (s *Stats) KeyWindowReqs(key string, sec int64) int64 {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	return sumBuckets(s.keyReqHour[key], time.Now().Unix(), sec)
+}
+
+// addKeyReqLocked adds n requests to a key's hourly count bucket.
+func (s *Stats) addKeyReqLocked(key string, h, n int64) {
+	hm := s.keyReqHour[key]
+	if hm == nil {
+		hm = map[int64]int64{}
+		s.keyReqHour[key] = hm
+	}
+	hm[h] += n
+	if len(hm) > quotaRetentionHours {
+		for k := range hm {
+			if k < h-quotaRetentionHours {
+				delete(hm, k)
+			}
+		}
+	}
+}
+
+// sumBuckets totals the hourly buckets inside the window ending at `now`
+// (unix seconds). sec <= 0 means "all retained history" (no reset).
+//
+// It only visits the hour buckets a window can actually cover. The naive full
+// scan measured 5.9us/op at the full 40-day retention, and a per-key quota
+// check runs on every request (2-3 of these), so the scan cost lands directly
+// on the request path; visiting at most sec/3600 + 1 buckets makes it
+// proportional to the window instead of to the retained history. The map keys
+// are whole unix hours, so the covered range is exact rather than a sample.
+func sumBuckets(hm map[int64]int64, now, sec int64) int64 {
+	if len(hm) == 0 {
+		return 0
+	}
+	if sec <= 0 {
+		var total int64
+		for _, v := range hm {
+			total += v
+		}
+		return total
+	}
+	nowHour := now / hourSec
+	// A bucket at hour H covers [H*3600, (H+1)*3600). The pre-existing
+	// WindowTokens counts a bucket when h*3600 >= now-sec, so the first hour
+	// that qualifies is ceil((now-sec)/3600) — NOT floor, which would pull in
+	// one bucket too many (caught by TestSumBucketsMatchesFullScan: a 30-day
+	// window reported 8878 vs 8649).
+	firstHour := (now - sec + hourSec - 1) / hourSec
+	if firstHour < 0 {
+		firstHour = 0
+	}
+	var total int64
+	for h := firstHour; h <= nowHour; h++ {
+		total += hm[h]
+	}
+	return total
+}
+
+// KeyWindowModelTokens returns the tokens one key consumed for one model
+// (optionally pinned to a source) within sec seconds; sec <= 0 means all
+// retained history. Unlike WindowTokens this is isolated per key, which is
+// what a per-key model quota needs.
+//
+// A source pin matches the pinned bucket exactly. Without a pin the call
+// totals the key's bare-model bucket, which counts traffic on every source —
+// a quota on "this model" should not stop applying just because the request
+// happened to be served by a different upstream.
+func (s *Stats) KeyWindowModelTokens(key, model, source string, sec int64) int64 {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	byModel := s.keyModelHour[key]
+	if len(byModel) == 0 {
+		return 0
+	}
+	mk := model
+	if source != "" {
+		mk = source + "::" + model
+		// from now on this pair keeps a pinned bucket, so a quota configured
+		// against this source counts the traffic it is meant to cap
+		s.wantPinnedBuckets(key, model)
+	}
+	return sumBuckets(byModel[mk], time.Now().Unix(), sec)
+}
+
+// KeyWindowTokens returns one key's total tokens across all models within sec
+// seconds; sec <= 0 means all retained history. It backs the whole-key quota.
+func (s *Stats) KeyWindowTokens(key string, sec int64) int64 {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	return sumBuckets(s.keyHour[key], time.Now().Unix(), sec)
+}
+
 // rotateAuditLocked renames the audit file to <path>.<unix>.old once it
 // exceeds auditRotateBytes and prunes old files beyond auditKeepOld, keeping
 // the newest ones. Caller must hold s.mu.
@@ -464,6 +675,27 @@ func AutoPeriodSeconds(period string, hours int64) int64 {
 	return 0
 }
 
+// AutoSecondsToReset returns how many seconds remain until the quota window
+// rolls over, for the Retry-After header on a 429. Buckets are whole unix
+// hours, so the value is rounded up to the next hour boundary and never
+// exceeds one window. 0 means "unknown" (never resets, or no period set), in
+// which case the caller should not advertise a retry time.
+func AutoSecondsToReset(period string, hours int64) int64 {
+	sec := AutoPeriodSeconds(period, hours)
+	if sec <= 0 {
+		return 0
+	}
+	now := time.Now().Unix()
+	// window = the last sec seconds, bucketed by whole hours; the oldest
+	// bucket still inside the window expires at its own hour boundary
+	elapsed := now % hourSec
+	left := sec - elapsed
+	if left <= 0 || left > sec {
+		left = sec
+	}
+	return left
+}
+
 // WindowTokens returns the tokens consumed for one model (optionally pinned
 // to a single source) within the window; sec <= 0 means all time. Buckets are
 // whole unix hours, so a sliding window overcounts by up to one hour — an

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/index.html

@@ -818,6 +818,10 @@
           kAnySrc: "任意源",
           kQuotaB: "Token 配额",
           kQuotaHintB: "0 / 留空 = 无限",
+          kReqQuota: "请求数配额",
+          kReqQuotaHint: "限制周期内的请求次数。0 / 留空 = 无限。",
+          kQuotaPerModelHint:
+            "配额按模型单独设置：创建后点「+ 添加模型」，逐个模型配 token 配额与周期。一个模型用满只影响该模型，同一密钥的其它模型照常。",
           kPeriodB: "重置周期",
           kPerNothing: "不限",
           kPerHour: "每 小时",
@@ -980,6 +984,7 @@
           thErr: "失败",
           thPrompt: "输入 Tokens",
           thCompl: "输出 Tokens",
+          thImgs: "张",
           thAvgLat: "平均延迟",
           thMaxLat: "最长延迟",
           thTime: "时间",
@@ -1043,6 +1048,10 @@
           kAnySrc: "any source",
           kQuotaB: "Token quota",
           kQuotaHintB: "0 / empty = unlimited",
+          kReqQuota: "Request quota",
+          kReqQuotaHint: "Caps requests per window. 0 / empty = unlimited.",
+          kQuotaPerModelHint:
+            "Quotas are per model: after creating the key, use \"Add model\" to give each model its own token budget and reset period. One model running out affects only that model; the key's other models keep working.",
           kPeriodB: "Reset period",
           kPerNothing: "Never",
           kPerHour: "Every hour",
@@ -1206,6 +1215,7 @@
           thErr: "Err",
           thPrompt: "Prompt Tokens",
           thCompl: "Completion Tokens",
+          thImgs: "imgs",
           thAvgLat: "Avg latency",
           thMaxLat: "Max latency",
           thTime: "Time",
@@ -1692,7 +1702,20 @@
   </div>`;
         document.body.appendChild(wrap);
       }
-      function downloadStatsCsv(from, to) {
+      // closeTopModal removes the topmost dialog. Several dialogs share the id
+      // "modal-wrap" and more than one can be open at once (the seed-key notice
+      // sits on top of the keys page), so `$("#modal-wrap")` resolves to
+      // whichever comes FIRST in the document — not necessarily the one the
+      // user is looking at. Handlers that receive the clicked button resolve
+      // their own dialog with .closest(); handlers with no button reference
+      // use this, which matches what the user sees.
+      function closeTopModal() {
+        const all = [...document.querySelectorAll("#modal-wrap")];
+        const w = all[all.length - 1];
+        if (w) w.remove();
+        return w;
+      }
+      function downloadStatsCsv(from, to, btn) {
         const q = new URLSearchParams({
           export: "csv",
           from: String(Math.floor(from)),
@@ -1700,8 +1723,7 @@
         });
         if (statsKeyF) q.set("key", statsKeyF);
         location.href = "/api/stats?" + q.toString();
-        const w = $("#modal-wrap");
-        if (w) w.remove();
+        (btn ? btn.closest("#modal-wrap") : null) || closeTopModal();
       }
       // downloadStatsCsvFromForm reads the custom date range from the export
       // modal's date inputs and forwards to downloadStatsCsv. (Previously this
@@ -1719,12 +1741,11 @@
         if (!isFinite(from) || from <= 0) return toast(t("expRange"));
         downloadStatsCsv(from, to);
       }
-      function downloadKeysCsv() {
+      function downloadKeysCsv(btn) {
         location.href =
           "/api/stats?export=keys-csv" +
           (statsKeyF ? "&key=" + encodeURIComponent(statsKeyF) : "");
-        const w = $("#modal-wrap");
-        if (w) w.remove();
+        (btn ? btn.closest("#modal-wrap") : null) || closeTopModal();
       }
       async function paintStats() {
         try {
@@ -2174,7 +2195,7 @@
       <td class="t-tag">${fmtTime(r.time)}</td>
       <td class="num">${r.ok ? `<span class="tag tag-green">${r.status || 200}</span>` : `<span class="tag tag-red" title="${esc(r.error || "")}">${r.status || 500}</span>`}</td>
       <td>${esc(keyNames[r.key] ? keyNames[r.key] + " · " + r.key : r.key)}</td><td class="t-tag">${esc(r.type)}</td><td>${esc(r.model)}</td><td>${esc(r.source || "")}</td>
-      <td class="num">${fmtTok(r.prompt_tokens)}</td><td class="num">${fmtTok(r.completion_tokens)}</td>
+      <td class="num">${fmtTok(r.prompt_tokens)}</td><td class="num">${r.type === "image" ? fmtN(r.image_count || 0) + " " + t("thImgs") : fmtTok(r.completion_tokens)}</td>
       <td class="num">${cacheCell(r)}</td>
       <td class="num">${fmtMs(r.latency_ms)}</td></tr>`;
       }
@@ -2772,7 +2793,10 @@
 
       /* ---------- sources tab ---------- */
       async function renderSources() {
-        const j = await api("/api/sources");
+        // Read through /api/v1: the UI is a client of the same public API an
+        // agent uses, so there is one contract rather than a private path.
+        // The list view needs no credentials, so the masked view is enough.
+        const j = await api("/api/v1/sources");
         const rows = j.sources
           .map(
             (
@@ -2794,13 +2818,23 @@
       function editSource(name) {
         const wrap = document.createElement("div");
         wrap.id = "modal-wrap";
-        Promise.all([api("/api/sources"), api("/api/status")])
-          .then(([src, st]) => {
-            const s = src.sources.find((x) => x.name === name) || {
-              name: name,
-              models: [{ id: "", priority: 0, kind: "chat" }],
-            };
-            const cur = s.adapter || "openai";
+            // The edit dialog round-trips the whole source, so it must load the
+            // real api_key — otherwise saving an unrelated field would blank
+            // it. That is what ?reveal=credentials is for (admin-only).
+            const srcPath = name
+              ? "/api/v1/sources/" + encodeURIComponent(name) + "?reveal=credentials"
+              : "/api/v1/sources";
+            Promise.all([api(srcPath), api("/api/status")])
+              .then(([src, st]) => {
+                // Editing: the single-source shape. Creating: an empty form.
+                const s = name
+                  ? src.source
+                  : { name: "", models: [{ id: "", priority: 0, kind: "chat" }] };
+                if (!s) {
+                  toast(t("toastSave"));
+                  return;
+                }
+                const cur = s.adapter || "openai";
             const apps = ["", ...(st.adapters || []).map((a) => a.name)];
             if (cur && !apps.includes(cur)) apps.push(cur);
             const adSel =
@@ -2905,9 +2939,10 @@
             body: JSON.stringify(payload),
           });
           toast(t("toastSaved"));
-          const w = $("#modal-wrap");
+          const w = btn && btn.closest("#modal-wrap");
           if (w) w.remove();
           else if (window._modal) window._modal.remove();
+          else closeTopModal();
           renderSources();
         } catch (e) {
           toast(tFmt("toastSaveFail", e.message));
@@ -3064,11 +3099,11 @@
         ? `<div class="tbl-wrap"><table><tr><th>${t("tName")}</th><th>${t("tURL")}</th><th>${t("tAdapter")}</th><th>${t("tModels")}</th><th></th></tr>` +
             tpls
               .map(
-                (t) =>
-                  `<tr><td><b>${esc(t.name)}</b></td><td>${esc(t.base_url)}</td><td>${esc(t.adapter)}</td>
-          <td><div class="src-models">${(t.models || []).map((m) => `<span class="tag tag-blue">${esc(m.id)}</span>`).join("")}</div></td>
-          <td><button class="ghost small" onclick="editTemplate(${JSON.stringify(t.name).replace(/["<>]/g, (c) => ({ '"': "&quot;", "<": "&lt;", ">": "&gt;" }[c]))})">${t("srcEdit")}</button>
-              <button class="danger small" onclick="delTemplate('${escAttr(t.name)}')">${t("srcDel")}</button></td></tr>`,
+                (tpl) =>
+                  `<tr><td><b>${esc(tpl.name)}</b></td><td>${esc(tpl.base_url)}</td><td>${esc(tpl.adapter)}</td>
+          <td><div class="src-models">${(tpl.models || []).map((m) => `<span class="tag tag-blue">${esc(m.id)}</span>`).join("")}</div></td>
+          <td><button class="ghost small" onclick="editTemplate(${JSON.stringify(tpl.name).replace(/["<>]/g, (c) => ({ '"': "&quot;", "<": "&lt;", ">": "&gt;" }[c]))})">${t("srcEdit")}</button>
+              <button class="danger small" onclick="delTemplate('${escAttr(tpl.name)}')">${t("srcDel")}</button></td></tr>`,
               )
               .join("") +
             "</table></div>"
@@ -3168,8 +3203,9 @@
             body: JSON.stringify({ name: nm, ...body }),
           });
           toast(t("toastSaved"));
-          const w = $("#modal-wrap");
+          const w = btn && btn.closest("#modal-wrap");
           if (w) w.remove();
+          else closeTopModal();
           openTemplateModal();
         } catch (e) {
           toast(e.message);
@@ -3215,7 +3251,7 @@
       const sortLanes = { chat: [], image: [] };
       let sortStateMap = new Map();
       async function renderSort() {
-        const j = await api("/api/sources");
+        const j = await api("/api/v1/sources");
         let autoR = [];
         let autoImg = [];
         try {
@@ -3875,7 +3911,7 @@
       </select>
       <div id="a-hours-box" style="display:none"><label>${t("kPerNHint")}</label>
         <input id="a-hours" type="number" min="1" step="1" value="24"></div>
-      <p><button onclick="scrAddFromForm()">${t("kSaveScope")}</button>
+      <p><button onclick="scrAddFromForm(this)">${t("kSaveScope")}</button>
          <button class="ghost" onclick="this.closest('#modal-wrap').remove()">${t("mCancel")}</button></p>
     </div>`;
         document.body.appendChild(wrap);
@@ -3885,7 +3921,7 @@
         });
         $("#a-model").focus();
       }
-      function scrAddFromForm() {
+      function scrAddFromForm(btn) {
         const raw = $("#a-model").value.trim();
         if (!raw) {
           toast(t("kName"));
@@ -3908,14 +3944,14 @@
           ],
         });
         const li = sortState.lanes.length - 1;
-        const w = $("#modal-wrap");
-        if (w) w.remove();
+        if (btn) btn.closest("#modal-wrap")?.remove();
+        else closeTopModal();
         paintSort([li]);
         persistAuto()
           .then(() => toast(t("sortHintSave")))
           .catch((e) => toast(e.message));
       }
-      function sortScopeSave(li, ji) {
+      function sortScopeSave(li, ji, btn) {
         const it = sortState.lanes[li].models[ji];
         if (!it) return;
         let q = parseInt($("#q-quota").value);
@@ -3924,8 +3960,7 @@
         let h = parseInt($("#q-hours").value);
         if (isNaN(h) || h < 1) h = 1;
         it.meta = { quota: q, period: p, hours: p || q ? h : 0 };
-        const w = $("#modal-wrap");
-        if (w) w.remove();
+        (btn ? btn.closest("#modal-wrap") : null) || closeTopModal();
         paintSort([li]);
         persistAuto()
           .then(() => toast(t("sortHintSave")))
@@ -4155,7 +4190,7 @@
       async function loadModelPairs() {
         let srcs = [];
         try {
-          srcs = (await api("/api/sources")).sources || [];
+          srcs = (await api("/api/v1/sources")).sources || [];
         } catch (e) {}
         const pairs = [];
         (srcs || []).forEach((s) =>
@@ -4284,13 +4319,14 @@
         const comb = scopeComb(m);
         const src = normSrc(m.source);
         const attrs = `data-key="${escAttr(key)}" data-model="${escAttr(comb)}"
-      data-quota="${m.token_quota || 0}" data-period="${escAttr(m.period || "")}" data-hours="${m.hours || 0}"`;
+      data-quota="${m.token_quota || 0}" data-reqquota="${m.req_quota || 0}"
+      data-period="${escAttr(m.period || "")}" data-hours="${m.hours || 0}"`;
         return `<span class="mb" draggable="true" ${attrs} title="${escAttr(t("kBrickH"))}"
       onclick="scopeEdit('${escAttr(key)}','${escAttr(comb)}')"
       oncontextmenu="scopeCtx(event,'${escAttr(key)}','${escAttr(comb)}')">
     <span class="mb-ico"><svg viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" width="10" height="10"><circle cx="12" cy="12" r="10"/></svg></span>
     <span class="mb-name">${esc(m.model)}${src ? `<em class="mb-src">${esc(src)}</em>` : ""}</span>
-    <span class="mb-quota">${esc(quantBadge(m.token_quota, m.period, m.hours))}</span>
+    <span class="mb-quota">${esc(scopeQuotaBadge(m))}</span>
     <span class="copy-b" title="${escAttr(t("kCopyB"))}" onclick="event.stopPropagation();scopeDup('${escAttr(key)}','${escAttr(comb)}')"><svg viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" width="12" height="12"><rect x="9" y="9" width="13" height="13" rx="2" ry="2"/><path d="M5 15H4a2 2 0 0 1-2-2V4a2 2 0 0 1 2-2h9a2 2 0 0 1 2 2v1"/></svg></span>
     <span class="mb-x" title="${escAttr(t("kDelB"))}" onclick="event.stopPropagation();scopeRm('${escAttr(key)}','${escAttr(comb)}')"><svg viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" width="12" height="12"><line x1="18" y1="6" x2="6" y2="18"/><line x1="6" y1="6" x2="18" y2="18"/></svg></span>
   </span>`;
@@ -4322,6 +4358,16 @@
         if (p === "nhour") return "·" + Math.max(1, h) + "h";
         return "";
       }
+      // scopeQuotaBadge shows a scope entry's budgets: token quota and, when
+      // set, the request count. They are per model — a spent budget blocks
+      // only that model, not the whole key.
+      function scopeQuotaBadge(m) {
+        const parts = [];
+        if (+m.token_quota > 0) parts.push(fmtQuota(m.token_quota));
+        if (+m.req_quota > 0) parts.push(fmtQuota(m.req_quota) + "\u00d7");
+        if (!parts.length) return "\u221e";
+        return parts.join(" ") + periodText(m.period, m.hours);
+      }
       function quantBadge(quota, period, hours) {
         quota = +quota || 0;
         period = period || "";
@@ -4336,16 +4382,21 @@
             model: comb[0],
             source: src || undefined,
             token_quota: parseInt(b.dataset.quota) || 0,
+            req_quota: parseInt(b.dataset.reqquota) || 0,
             period: b.dataset.period || "",
             hours: parseInt(b.dataset.hours) || 0,
           };
         });
       }
       async function putScope(key, scopes) {
+        // Each scope entry carries its own quotas, so the whole budget travels
+        // with the models it applies to. There is no separate key-wide total
+        // that could drift out of sync with the model list.
+        const body = { models: scopes };
         await api("/api/keys/" + encodeURIComponent(key), {
           method: "PUT",
           headers: { "Content-Type": "application/json" },
-          body: JSON.stringify({ models: scopes }),
+          body: JSON.stringify(body),
         });
       }
       async function scopePush(key) {
@@ -4448,6 +4499,9 @@
       <label>${t("kQuotaB")} <span class="muted">${t("kQuotaHintB")}</span></label>
       <input id="sc-quota" type="number" min="0" step="1"
         placeholder="${escAttr(t("kQuotaHintB"))}" value="${sc.token_quota ? sc.token_quota : ""}">
+      <label>${t("kReqQuota")} <span class="muted">${t("kReqQuotaHint")}</span></label>
+      <input id="sc-reqs" type="number" min="0" step="1"
+        placeholder="${escAttr(t("kReqQuotaHint"))}" value="${sc.req_quota || ""}">
       <label>${t("kPeriodB")}</label>
       <select id="sc-period">
         <option value="" ${!sc.period ? "selected" : ""}>${t("kPerNothing")}</option>
@@ -4481,6 +4535,8 @@
         const parts = splitCombKey(comb);
         let q = parseInt($("#sc-quota").value);
         if (isNaN(q) || q < 0) q = 0;
+        let rq = parseInt($("#sc-reqs").value);
+        if (isNaN(rq) || rq < 0) rq = 0;
         let hours = parseInt($("#sc-hours").value);
         if (isNaN(hours) || hours < 1) hours = 1;
         const period = $("#sc-period").value;
@@ -4489,6 +4545,7 @@
           model: parts.model,
           source: parts.source || undefined,
           token_quota: q,
+          req_quota: rq,
           period,
           hours,
         };
@@ -4498,8 +4555,9 @@
         if (btn) btn.disabled = true;
         try {
           await putScope(key, scopes);
-          const w = $("#modal-wrap");
+          const w = btn && btn.closest("#modal-wrap");
           if (w) w.remove();
+          else closeTopModal();
           toast(t("kSaved"));
           await loadKeys();
         } catch (e) {
@@ -4648,6 +4706,7 @@
         <option value="user">${t("kRoleUser")}</option>
         <option value="admin">${t("kRoleAdmin")}</option>
       </select>
+      <p class="muted" style="font-size:12px">${t("kQuotaPerModelHint")}</p>
       <label>${t("kNote")}</label>
       <input id="kc-note">
       <p><button onclick="createKey(this)">${t("kCreateBtn")}</button>
@@ -4679,8 +4738,12 @@
           if (btn) btn.disabled = false;
           return;
         }
-        const w = $("#modal-wrap");
+        // Close THIS dialog (resolved from the clicked button), not whichever
+        // #modal-wrap comes first: the seed-key notice may already be open and
+        // would otherwise be the one that gets removed.
+        const w = btn ? btn.closest("#modal-wrap") : null;
         if (w) w.remove();
+        else closeTopModal();
         $("#k-newbox").innerHTML = `
     <div class="key-canvas" style="background:var(--card2)">
       <div class="kc-head">

internal/gateway/ui_contract_test.go

@@ -347,3 +347,60 @@ func cssBlock(src, selector string) (string, bool) {
 	}
 	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

@@ -5,10 +5,12 @@ adapter.version = "1.0.0"
 adapter.endpoint = "/chat/completions"
 adapter.headers = {}
 
--- AgentRouter 的 WAF 会按客户端指纹白名单校验，只放行官方客户端。
--- 这里通过 build_headers 注入官方客户端 User-Agent 以通过校验。
--- 默认 QwenCode 指纹（实测可通）；可在 source.meta.user_agent 覆盖成 Claude Code 指纹。
-
+-- 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
@@ -65,6 +67,13 @@ function adapter.build_headers(meta)
         ["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
 

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

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
 

internal/lua/adapters/opencode.lua

@@ -34,8 +34,10 @@ 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。
@@ -52,8 +54,14 @@ function adapter.transform_request(raw_body)
     if not ok then return raw_body end
     req.disable_thinking = nil
     req.extra_body = nil
-    if type(req.stream_options) ~= "table" then req.stream_options = {} end
-    req.stream_options.include_usage = true
+    -- 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
@@ -72,7 +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

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/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_test.go

@@ -221,6 +221,70 @@ func TestOpenCodeStripsMultiModalParts(t *testing.T) {
 	}
 }
 
+// TestOpenCodeKeepsToolCallWithEmptyContent pins the fix for an infinite
+// "repeated tool call" loop: an assistant turn whose content serialises to an
+// empty ARRAY while carrying tool_calls must survive the multimodal strip.
+// Dropping it orphans the following tool result, so the model never sees the
+// call it already made and re-issues it on every turn.
+func TestOpenCodeKeepsToolCallWithEmptyContent(t *testing.T) {
+	vm := NewVM(freshAdapterDir(t))
+	if err := vm.Start(); err != nil {
+		t.Fatal(err)
+	}
+	defer vm.Stop()
+
+	body := `{"model":"x","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"},
+		{"role":"user","content":"now summarize"}
+	]}`
+	out, err := vm.Transform("opencode", "transform_request", body)
+	if err != nil {
+		t.Fatalf("transform_request: %v", err)
+	}
+
+	var req struct {
+		Messages []struct {
+			Role      string          `json:"role"`
+			Content   json.RawMessage `json:"content"`
+			ToolCalls []struct {
+				ID string `json:"id"`
+			} `json:"tool_calls"`
+			ToolCallID string `json:"tool_call_id"`
+		} `json:"messages"`
+	}
+	if err := json.Unmarshal([]byte(out), &req); err != nil {
+		t.Fatalf("unmarshal: %v (%s)", err, out)
+	}
+	if len(req.Messages) != 4 {
+		t.Fatalf("expected 4 messages (tool call must not be dropped), got %d: %s", len(req.Messages), out)
+	}
+	asst := req.Messages[1]
+	if asst.Role != "assistant" || len(asst.ToolCalls) != 1 || asst.ToolCalls[0].ID != "call_abc" {
+		t.Fatalf("assistant tool_calls lost: %s", out)
+	}
+	if string(asst.Content) != `""` {
+		t.Fatalf("empty content must be rewritten to \"\", got %s", asst.Content)
+	}
+	if req.Messages[2].Role != "tool" || req.Messages[2].ToolCallID != "call_abc" {
+		t.Fatalf("tool result must stay paired with its call: %s", out)
+	}
+
+	// Negative control: a message with neither text nor tool calls is still dropped.
+	only := `{"model":"x","messages":[
+		{"role":"user","content":[{"type":"image_url","image_url":{"url":"https://ex.com/a.png"}}]},
+		{"role":"user","content":"keep me"}
+	]}`
+	out2, err := vm.Transform("opencode", "transform_request", only)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if got := strings.Count(out2, `"role"`); got != 1 {
+		t.Fatalf("image-only message without tool calls must still be dropped, got %d: %s", got, out2)
+	}
+}
+
 func TestBuildHeadersCustomHook(t *testing.T) {
 	vm := NewVM(freshAdapterDir(t))
 	if err := vm.Start(); err != nil {

internal/provider/provider.go

@@ -707,7 +707,7 @@ func (p *Provider) probeModels(ctx context.Context) (bool, string) {
 	if err != nil {
 		return false, ""
 	}
-	if hdrs, herr := p.buildHeaders("{}", u); herr == nil {
+	if hdrs, herr := p.buildHeaders("{}", u, ""); herr == nil {
 		req.Header = hdrs
 	}
 	resp, err := p.client.Do(req)
@@ -750,7 +750,7 @@ func (p *Provider) probeChat(ctx context.Context) (bool, string) {
 		body, err := json.Marshal(probe)
 		if err == nil {
 			var hdr http.Header
-			if hdrs, herr := p.buildHeaders(string(body), p.URL()); herr == nil {
+			if hdrs, herr := p.buildHeaders(string(body), p.URL(), ""); herr == nil {
 				hdr = hdrs
 			}
 			raw, status, derr := p.do(ctx, p.URL(), string(body), hdr)
@@ -1031,13 +1031,15 @@ func (p *Provider) Throttle(ctx context.Context) error {
 
 // ---- request construction ----
 
-func (p *Provider) buildHeaders(body, url string) (http.Header, error) {
+func (p *Provider) buildHeaders(body, url, clientSession string) (http.Header, error) {
 	meta := map[string]interface{}{
 		"url":       url,
 		"method":    http.MethodPost,
 		"body":      body,
 		"api_key":   p.cfg.APIKey,
 		"timestamp": types.Now(),
+		// client_session 是客户端自带的会话标识（请求头），为空时适配器自行推导。
+		"client_session": clientSession,
 		"source": map[string]interface{}{
 			"name": p.cfg.Name,
 			"meta": p.cfg.Meta,
@@ -1082,7 +1084,7 @@ func (p *Provider) Chat(ctx context.Context, req *types.ChatRequest) (*types.Uni
 	if err != nil {
 		return nil, err
 	}
-	hdrs, err := p.buildHeaders(body, p.URL())
+	hdrs, err := p.buildHeaders(body, p.URL(), req.ClientSession)
 	if err != nil {
 		return nil, err
 	}
@@ -1142,7 +1144,7 @@ func (p *Provider) ChatStream(ctx context.Context, req *types.ChatRequest) (<-ch
 		p.Release()
 		return nil, err
 	}
-	hdrs, err := p.buildHeaders(body, p.URL())
+	hdrs, err := p.buildHeaders(body, p.URL(), req.ClientSession)
 	if err != nil {
 		p.Release()
 		return nil, err
@@ -1411,7 +1413,7 @@ func (p *Provider) Image(ctx context.Context, req *types.ImageGenRequest) (*type
 		// fall back to passthrough adapter (openai-style)
 		transformed = string(b)
 	}
-	hdrs, err := p.buildHeaders(transformed, p.ImageURL())
+	hdrs, err := p.buildHeaders(transformed, p.ImageURL(), "")
 	if err != nil {
 		return nil, err
 	}
@@ -1516,12 +1518,24 @@ func standardSSEChunk(data string) string {
 			PromptTokensDetails *struct {
 				CachedTokens int `json:"cached_tokens"`
 			} `json:"prompt_tokens_details"`
+			CompletionTokensDetails *struct {
+				ReasoningTokens int `json:"reasoning_tokens"`
+			} `json:"completion_tokens_details"`
 		} `json:"usage"`
+		// Cost is a top-level decimal string on OpenCode Zen/Go responses.
+		Cost json.RawMessage `json:"cost"`
 	}
 	if err := json.Unmarshal([]byte(data), &raw); err != nil {
 		return ""
 	}
 	if len(raw.Choices) == 0 && raw.UpstreamUsage.Total == 0 && raw.UpstreamUsage.TotalTokens == 0 {
+		// Cost-only chunk: OpenCode sends the charge on its own frame
+		// {"choices":[],"cost":"0"} with no usage at all. Dropping it here
+		// would lose the only cost signal the client can see.
+		if c := rawCostString(raw.Cost); c != "" {
+			out, _ := json.Marshal(types.UnifiedChunk{Cost: c})
+			return string(out)
+		}
 		return ""
 	}
 	var usage *types.TokenUsage
@@ -1539,6 +1553,11 @@ func standardSSEChunk(data string) string {
 				CachedTokens: pu.PromptTokensDetails.CachedTokens,
 			}
 		}
+		if pu.CompletionTokensDetails != nil {
+			usage.CompletionTokensDetails = &types.CompletionTokensDetails{
+				ReasoningTokens: pu.CompletionTokensDetails.ReasoningTokens,
+			}
+		}
 	}
 	finish := ""
 	done := false
@@ -1563,6 +1582,24 @@ func standardSSEChunk(data string) string {
 	return string(out)
 }
 
+// rawCostString normalizes an upstream cost field into a plain string. The
+// value is a decimal **string** on OpenCode Zen/Go ("0", "0.0012"), but other
+// upstreams have been seen to send a bare number, so both are accepted.
+// Returns "" when absent, null, or empty.
+func rawCostString(raw json.RawMessage) string {
+	s := strings.TrimSpace(string(raw))
+	if s == "" || s == "null" {
+		return ""
+	}
+	if s[0] == '"' {
+		var unq string
+		if json.Unmarshal(raw, &unq) == nil {
+			return strings.TrimSpace(unq)
+		}
+	}
+	return s
+}
+
 // pickFirst returns a if non-zero, else b (for usage keys that may appear in
 // either standard *_tokens or legacy short form).
 func pickFirst(a, b int) int {

internal/types/types.go

@@ -3,6 +3,7 @@
 package types
 
 import (
+	"bytes"
 	"encoding/json"
 	"errors"
 	"strings"
@@ -28,6 +29,10 @@ type ChatRequest struct {
 	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) {
@@ -61,6 +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"`
@@ -79,6 +117,11 @@ 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 {
@@ -95,6 +138,16 @@ type TokenUsage struct {
 	// 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
@@ -119,15 +172,16 @@ func (t TokenUsage) MarshalJSON() ([]byte, error) {
 		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"`
+		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,
@@ -138,6 +192,7 @@ func (t TokenUsage) MarshalJSON() ([]byte, error) {
 		PromptTokensDetails:   pdetails,
 		PromptCacheHitTokens:  t.PromptCacheHit,
 		PromptCacheMissTokens: t.PromptCacheMiss,
+		CompletionDetails:     t.CompletionTokensDetails,
 	})
 }
 
@@ -181,6 +236,9 @@ type UnifiedChunk struct {
 	// 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

internal/types/types_test.go

@@ -2,6 +2,7 @@ package types
 
 import (
 	"encoding/json"
+	"strings"
 	"testing"
 )
 
@@ -68,3 +69,86 @@ func TestTokenUsageMarshalCacheFields(t *testing.T) {
 	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/config.example.yaml

@@ -0,0 +1,30 @@
+# llmsproxy — 统一 OpenAI 兼容网关配置（示例模板）
+#
+# 这个是打包进发行包、供首次安装使用的模板。它故意不含任何可用的凭据：
+# 首次启动时程序会生成一个随机 admin key 并写入实际配置（WebUI「密钥」页可见）。
+#
+# ⚠ 不要在本文件里填一个写死的具体 key。示例文件是公开的，写进去的任何值
+#    都会成为一个全网皆知的 admin key（旧版示例里的 sk-gw-local-0001 就踩过
+#    这个坑：它被直接抄进生产配置并真实可用）。
+
+# 网关监听地址（默认 :8080，建议绑内网/回环）
+listen: 127.0.0.1:8080
+
+# 客户端访问本网关所需的 API Key（Bearer）。留空数组 = 不鉴权（仅内网）。
+gateway_keys: []
+
+# 默认模型选择：具体模型 id 或 AUTO（按各源模型的 priority 自动选最高可用源）
+default_model: AUTO
+
+# Lua 适配器目录（默认 adapters/，首次启动自动写入内置适配器）
+adapter_dir: adapters
+
+# 运行时持久化文件（WebUI 新增/编辑的源会写入此文件，重启后仍生效）
+runtime_file: runtime.json
+
+# 全局并发上限（0 = 不限）
+max_concurrent: 0
+
+# ---- 上游 LLM 源列表 ----
+# 留空：安装后到 WebUI 添加源，或从内置模板一键创建。
+sources: []

packaging/core-dist.sh

@@ -68,7 +68,12 @@ 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"
+# Use the tracked sanitised template, NOT the local config.yaml: that file is
+# gitignored and carries the operator's real gateway_keys / upstream keys. It
+# was previously copied verbatim as config.example.yaml, which shipped a real,
+# working admin key (sk-gw-local-0001) to every install — and postinst.sh copies
+# the example to /etc when none exists, so it ended up live in production.
+cp "$ROOT/packaging/config.example.yaml" "$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/"

packaging/nfpm.yaml

@@ -28,8 +28,11 @@ contents:
     dst: /etc/systemd/system/llmsproxy.service
     file_info:
       mode: 0644
-  # example config (seeded to /etc/llmsproxy on first install by postinst)
-  - src: ./config.yaml
+  # example config (seeded to /etc/llmsproxy on first install by postinst).
+  # Must be the tracked sanitised template: ./config.yaml is the operator's own
+  # gitignored config and shipping it would leak real credentials — it shipped a
+  # working admin key (sk-gw-local-0001) to every install for a while.
+  - src: ./packaging/config.example.yaml
     dst: /usr/share/llmsproxy/config.example.yaml
     file_info:
       mode: 0644

packaging/postinst.sh

@@ -9,9 +9,19 @@ 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
+    # Do NOT seed a starter config here. Packaging/installing one used to hand
+    # everyone the same file, and while it was copied from the maintainer's
+    # live config it shipped a real, working admin key (sk-gw-local-0001) to
+    # every install.
+    #
+    # Instead, leave the file absent: on first start the binary's EnsureDefault
+    # path creates config.yaml with a FRESH RANDOM admin key and logs it, which
+    # is both unique per install and actually usable. The packaged
+    # config.example.yaml stays in /usr/share as a reference template only.
+    #
+    # (Seeding it with gateway_keys: [] would be worse: the gateway would start
+    # but reject every request, with no way in except editing the file.)
+    echo "llmsproxy: no /etc/llmsproxy/config.yaml — a random admin key will be generated on first start"
 fi
 
 # systemd