chore(gui): 版本号 1.9.6 → 1.9.7

## ① 图像模型曾与对话模型不可区分

根因是结构性的，不是漏了个过滤：

    Provider.Models()   → []string   ← 只有 ID，Kind 被丢掉
    SourceStatus.Models → []string   ← 扁平列表

于是 /api/status 给前端的 34 个模型（29 对话 + 5 图像）长得一模一样。选了
gpt-image-2 就 POST 到 /api/chat —— 那个端点服务不了图像模型，必然失败，
而且报错完全不提「这是图像模型」。

修法：新增 ModelKinds []ModelKindInfo（[{id,kind}]），纯增量、omitempty，
老 UI 不受影响；前端按 optgroup 分组，且这些 option 是 **disabled** 的。

分组还不够 —— 分组只是视觉分隔，operator 照样能选，选了照样失败。
disabled 的 option 会自解释（「请到生图页」），可选的只会失败。

★ 写这段时踩了一次 String.replace 的坑：`imgOpts.join("").replace(...)`
只替换**第一个**匹配，于是 6 个图像模型里只有第 1 个被禁用。浏览器实测
逐条读 e.disabled 才抓到（siliconflow 禁用了、vela 的 5 个没有）。
判据只看 value 里有没有 "image" 的话，第一次跑就会永远报警——那是判据
自己错了，比没判据更糟。

## ② 生图二级页面

侧栏新增 tab，与「对话」并列。网格展示 + 提示词输入 + 数量/尺寸，
调用 /v1/images/generations（不是 /api/chat）。

按钮显示已用秒数：实测该上游 15.7s~112.5s，波动 7.6 倍，不显示进度
看起来就像卡死了。

浏览器实测（几何测量，非仅 DOM 断言）：
    tab-image 容器   986×523  visible
    img-grid         840×374  visible
    提示词输入框      634×64   visible
    生图页 6 个模型，全部 kind=image，全部可选
    对话页 6 个图像模型，全部 disabled；97 个对话模型全部可选

## ③ base_url 引导不足

文案直接给出截断位置和具体后果，而不是泛泛地说「填 URL」：

    填主机地址，通常不带 /v1 —— 网关会自动拼接路径。但各家约定不同：
    若填 https://x.com/v1，对话路径会正确，而图像生成会变成 /v1/v1/… 而 404。
    不确定就用上方「自动探测模型」。

/v1/v1 不是假设：ImageURL 是 BaseURL + ImageEndpoint()，而图像默认路径
是**绝对**的 /v1/images/generations，chat 的默认是相对的 /chat/completions。
两者对「版本段归谁管」的假设相反。今天在 vela 上为此排查了两次。

## 判据

新增 6 个契约测试：
- 对话页必须读 model_kinds（不读就会退回扁平列表）
- 禁用图像模型不能用 joined-HTML 的 String.replace（只改第一个）
- 生图页存在且注册进 TABS（漏了则 pane 永不可见，症状是「页面空白」）
- sendImage 必须打 /v1/images/generations
- urlHint 必须同时提到 /v1 和 404
- TestImageEndpointComposesUnderRealConditions 用真实形态验证：
  相对路径拼出 /v1/images/generations 得 200，默认路径拼出 /v1/v1/… 得 404

README.md

@@ -150,6 +150,47 @@ sources:
     timeout: 120s              # 请求超时，默认 120s
 ```
 
+
+### 自动探测模型列表
+
+不想手抄模型 id 时，只填 URL 和 key，让网关去扫：
+
+```bash
+# 只探测，不写任何东西
+curl -X POST http://127.0.0.1:8080/api/sources/discover \
+  -H "Authorization: Bearer sk-gw-…" -H "Content-Type: application/json" \
+  -d '{"base_url":"https://api.vendor.com/v1","api_key":"sk-…"}'
+# → {"ok":true,"base_url":"…","endpoint":"","models":["gpt-x",…],
+#    "image_models":["gpt-image-2"],"models_path":"/models","latency_ms":1387,…}
+
+# 探测后直接落库
+curl -X POST http://127.0.0.1:8080/api/sources/discover/apply \
+  -H "Authorization: Bearer sk-gw-…" -H "Content-Type: application/json" \
+  -d '{"base_url":"https://api.vendor.com/v1","api_key":"sk-…","name":"vendor"}'
+```
+
+WebUI 源编辑框里也有「自动探测模型」按钮：填好 URL 和 key 点一下，列表自动
+填好，仍需点保存才生效。
+
+几点值得知道：
+
+- **URL 形态不用猜。** 厂商对「`/v1` 放 base 还是放 path」没有共识，探测会
+  依次尝试若干候选并返回答对的那个，同时给出归一化后的 `base_url` +
+  `endpoint`（`endpoint` 平时留空，因为 openai adapter 自带 `/chat/completions`）。
+- **`image_endpoint` 是另一套规则。** 网关的图像默认路径是**绝对**的
+  `/v1/images/generations`，而 chat 默认是相对的。所以 `base_url` 带 `/v1` 时
+  必须显式给 `image_endpoint: /images/generations`，否则图像请求 404。
+  探测会替你填好。
+- **图像模型只看上游自己声明的 modality**，绝不从模型名猜。`gpt-image-2`
+  这类 id 也出现在同时提供对话服务的目录上，猜错会让 AUTO 链选中一个只支持
+  图像生成的模型，其后每一层都被它拖垮。
+- **不会逐模型探活。** 34 个模型的目录跑一轮要花钱且耗时数分钟，`/models`
+  是足够便宜的信号；模型是否真能应答留给正常请求去发现。
+- **对已存在的源，必须显式 `replace_models: true`** 才会覆盖模型列表。重新
+  探测顺手抹掉为 AUTO 链手工调过的 priority，是最难排查的一类数据丢失。
+- **`api_key` 留空表示沿用已存值**，不会把凭据清空。
+
+
 `config.yaml`、`runtime.json`、`master.key`、`adapters/` 均不提交版本库（见
 `.gitignore`），请随备份带走——`master.key` 丢失后运行时密文无法解密。
 

cmd/gui/package.json

@@ -1,6 +1,6 @@
 {
   "name": "modelrouter-gui",
-  "version": "1.9.4",
+  "version": "1.9.7",
   "description": "ModelRouter Desktop — embedded ModelRouter core with tray",
   "author": "ModelRouter",
   "main": "main.js",

internal/gateway/apiv1.go

@@ -189,6 +189,20 @@ func (g *Gateway) apiV1Index(w http.ResponseWriter, r *http.Request) {
 			{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: "POST", Path: "/api/sources/discover", Auth: "admin",
+				Summary: "probe a base_url+api_key for its model catalog",
+				// Probing is read-only with respect to configuration: it writes
+				// nothing, which is why it is safe to call repeatedly while the
+				// operator is still typing a key.
+				WriteEffect: "none",
+			},
+			{Method: "POST", Path: "/api/sources/discover/apply", Auth: "admin",
+				Summary: "probe and write the source in one step",
+				// Requires replace_models=true when the source already exists,
+				// so re-running discovery cannot silently discard a hand-tuned
+				// model list. An empty api_key keeps the stored credential.
+				WriteEffect: "creates or replaces a source and its model list",
+			},
 			{Method: "GET", Path: "/api/source_templates", Auth: "admin", Summary: "source templates"},
 			{Method: "POST", Path: "/v1/chat/completions", Auth: "any", Summary: "OpenAI-compatible inference"},
 		},

internal/gateway/chat.go

@@ -256,17 +256,17 @@ func (g *Gateway) checkQuota(ctx context.Context, model string) *quotaRejection
 		}
 		return nil
 	}
-// A scope that simply does not list the model is a different problem from
-// a quota being exhausted, and the message has to say which: telling an
-// operator "not allowed for this key" when they configured a quota sends
-// them to the wrong setting. AUTO is checked here before the chain is even
-// consulted, so this is the message that surfaces for AUTO.
-if isAuto(model) {
-return &quotaRejection{msg: fmt.Sprintf(
-"model %q is not in this key's model scope, so it cannot use AUTO; "+
-"add %q to the key's models or remove the scope restriction", model, model)}
-}
-return &quotaRejection{msg: fmt.Sprintf("model %q is not allowed for this key", model)}
+	// A scope that simply does not list the model is a different problem from
+	// a quota being exhausted, and the message has to say which: telling an
+	// operator "not allowed for this key" when they configured a quota sends
+	// them to the wrong setting. AUTO is checked here before the chain is even
+	// consulted, so this is the message that surfaces for AUTO.
+	if isAuto(model) {
+		return &quotaRejection{msg: fmt.Sprintf(
+			"model %q is not in this key's model scope, so it cannot use AUTO; "+
+				"add %q to the key's models or remove the scope restriction", model, model)}
+	}
+	return &quotaRejection{msg: fmt.Sprintf("model %q is not allowed for this key", model)}
 }
 
 // quotaWindowSuffix describes a quota's reset window for an error message, so
@@ -1059,9 +1059,9 @@ func (g *Gateway) fireEnd(rec *Req) {
 		// A SUBSET of completion_tokens, not an extra. Billing plugins that
 		// add it to completion would double-charge every thinking token.
 		"reasoning_tokens": rec.Reasoning,
-		"image_count":    rec.ImageCount,
-		"error":          rec.Err,
-		"time":           rec.Time,
+		"image_count":      rec.ImageCount,
+		"error":            rec.Err,
+		"time":             rec.Time,
 		// chain_walk: the AUTO tier-by-tier trace, when the request went
 		// through the chain. Empty for a direct request and for a gateway with
 		// no plugins loaded. Absent rather than empty so a plugin can tell
@@ -1215,20 +1215,7 @@ func (g *Gateway) pumpStream(w http.ResponseWriter, rec *Req, chunks <-chan type
 	var tut *types.TokenUsage
 	if lastUsage != nil {
 		tut = lastUsage
-		// Write the upstream's cache accounting back onto the request record
-		// so the audit trail carries the hit/miss split for streaming too.
-		if d := tut.PromptTokensDetails; d != nil {
-			rec.CacheReported = true
-			if d.CachedTokens > 0 {
-				rec.CacheHit = int64(d.CachedTokens)
-			}
-		} else if tut.PromptCacheHit > 0 {
-			rec.CacheHit = int64(tut.PromptCacheHit)
-			rec.CacheReported = true
-		}
-		if tut.PromptCacheMiss > 0 {
-			rec.CacheMiss = int64(tut.PromptCacheMiss)
-		}
+		applyStreamUsageDetails(rec, lastUsage)
 	} else if rec.Prompt+rec.Compl > 0 {
 		u := types.TokenUsage{
 			Prompt:     int(rec.Prompt),
@@ -1459,3 +1446,39 @@ func (g *Gateway) handleImage(w http.ResponseWriter, r *http.Request) {
 		Data:    resp.ImageData,
 	})
 }
+
+// applyStreamUsageDetails mirrors the upstream's reasoning and cache numbers
+// onto a STREAMED request's audit record.
+//
+// It exists as a named function because this was inline in the streaming handler,
+// where it was untestable — and untested it stayed wrong. The non-streaming
+// recorder read completion_tokens_details.reasoning_tokens; this one did not. So
+// a source's streamed requests recorded completion and no reasoning at all while
+// its non-streamed calls recorded both. Nothing failed and no test was red: the
+// audit trail was simply quieter than it should have been.
+//
+// Found by reading the production audit file after wiring a reasoning-heavy
+// source: completion=8166 with reasoning absent, on every streamed request.
+//
+// Reasoning is a SUBSET of completion, never an addition to it. See
+// recordChatUsage for why that matters to billing.
+func applyStreamUsageDetails(rec *Req, u *types.TokenUsage) {
+	if d := u.CompletionTokensDetails; d != nil && d.ReasoningTokens > 0 {
+		rec.Reasoning = int64(d.ReasoningTokens)
+	}
+	// Cache accounting is here for the same reason it is kept apart: "zero hits
+	// because nothing was cached" and "this provider never reports caching" are
+	// the same number and mean opposite things.
+	if d := u.PromptTokensDetails; d != nil {
+		rec.CacheReported = true
+		if d.CachedTokens > 0 {
+			rec.CacheHit = int64(d.CachedTokens)
+		}
+	} else if u.PromptCacheHit > 0 {
+		rec.CacheHit = int64(u.PromptCacheHit)
+		rec.CacheReported = true
+	}
+	if u.PromptCacheMiss > 0 {
+		rec.CacheMiss = int64(u.PromptCacheMiss)
+	}
+}

internal/gateway/chat_reasoning_test.go

@@ -69,3 +69,92 @@ func TestRecordChatUsageReasoningAbsentStaysZero(t *testing.T) {
 		}
 	}
 }
+
+// TestStreamedReasoningIsRecordedOnTheRequest is the streaming counterpart.
+//
+// TestRecordChatUsageKeepsReasoningTokens covers the non-streaming path only,
+// and for a long time the streamed path dropped the number: the non-streaming
+// recorder read TokenUsage.CompletionTokensDetails, while the streaming one read
+// only Prompt and Completion from the terminal usage chunk. (Cache accounting was
+// mirrored; reasoning simply never was.)
+//
+// The symptom is invisible in the response — the client sees a correct answer
+// either way. It only shows in the audit trail, where a reasoning-heavy source's
+// streamed requests recorded completion=8166 and no reasoning at all. Found by
+// reading the production audit file, not by a failing test, which is exactly why
+// the property is pinned here.
+func TestStreamedReasoningIsRecordedOnTheRequest(t *testing.T) {
+	rec := &Req{}
+	applyStreamUsageDetails(rec, &types.TokenUsage{
+		Prompt:     159579,
+		Completion: 751,
+		Total:      160330,
+		CompletionTokensDetails: &types.CompletionTokensDetails{
+			ReasoningTokens: 512,
+		},
+	})
+	if rec.Reasoning != 512 {
+		t.Errorf("streamed Reasoning = %d, want 512 — a streamed request records no "+
+			"reasoning, so a pricing rule that splits thinking from output silently "+
+			"bills it as ordinary output", rec.Reasoning)
+	}
+}
+
+// TestStreamedReasoningIsNeverAddedToCompletion pins the SUBSET relation in the
+// streaming path specifically. A presence check would pass if someone wrote
+// rec.Compl += reasoning, and billing would then charge every thinking token
+// twice.
+func TestStreamedReasoningIsNeverAddedToCompletion(t *testing.T) {
+	rec := &Req{}
+	rec.Compl = 751 // set upstream of the call, as the streaming handler does
+	applyStreamUsageDetails(rec, &types.TokenUsage{
+		Completion:              751,
+		CompletionTokensDetails: &types.CompletionTokensDetails{ReasoningTokens: 512},
+	})
+	if rec.Compl != 751 {
+		t.Errorf("Compl = %d, want 751 unchanged — reasoning is a subset, not an addition", rec.Compl)
+	}
+	if rec.Reasoning >= rec.Compl {
+		t.Errorf("Reasoning (%d) must be a SUBSET of Compl (%d)", rec.Reasoning, rec.Compl)
+	}
+}
+
+// TestStreamedUsageWithoutReasoningLeavesItUnset: an upstream that reports no
+// reasoning must not be turned into a reported zero. In the audit trail 0 means
+// "not reported", and a plugin cannot tell that from "no reasoning happened".
+func TestStreamedUsageWithoutReasoningLeavesItUnset(t *testing.T) {
+	rec := &Req{}
+	applyStreamUsageDetails(rec, &types.TokenUsage{Prompt: 10, Completion: 4, Total: 14})
+	if rec.Reasoning != 0 {
+		t.Errorf("Reasoning = %d for an upstream that reported none", rec.Reasoning)
+	}
+}
+
+// TestStreamedCacheAccountingStillWorks is a regression guard for the block the
+// reasoning lines were added to. They share a function, so a future edit that
+// short-circuits one would break the other — silently, since a missing cache
+// number reads as "nothing was cached".
+func TestStreamedCacheAccountingStillWorks(t *testing.T) {
+	for _, c := range []struct {
+		name string
+		u    types.TokenUsage
+		hit  int64
+		miss int64
+		rep  bool
+	}{
+		{"details object", types.TokenUsage{
+			PromptTokensDetails: &types.PromptTokensDetails{CachedTokens: 4096},
+		}, 4096, 0, true},
+		{"legacy hit/miss", types.TokenUsage{
+			PromptCacheHit: 2048, PromptCacheMiss: 512,
+		}, 2048, 512, true},
+		{"nothing reported", types.TokenUsage{}, 0, 0, false},
+	} {
+		rec := &Req{}
+		applyStreamUsageDetails(rec, &c.u)
+		if rec.CacheHit != c.hit || rec.CacheMiss != c.miss || rec.CacheReported != c.rep {
+			t.Errorf("%s: hit=%d miss=%d reported=%v; want %d/%d/%v",
+				c.name, rec.CacheHit, rec.CacheMiss, rec.CacheReported, c.hit, c.miss, c.rep)
+		}
+	}
+}

internal/gateway/discover.go

@@ -0,0 +1,749 @@
+package gateway
+
+import (
+	"context"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"io"
+	"net/http"
+	"sort"
+	"strings"
+	"time"
+
+	"llmsproxy/internal/config"
+)
+
+// Model discovery: turn "here is a URL and a key" into a configured source.
+//
+// The hard part is not the GET /models call — it is guessing which URL the
+// upstream actually serves. This gateway builds request paths as
+// base_url + adapter endpoint (openai contributes "/chat/completions"), so a
+// base_url that already ends in "/v1" yields "/v1/chat/completions" while a
+// base_url of "/v1" plus our default image path yields the broken
+// "/v1/v1/images/generations". Vendors disagree on whether the version segment
+// belongs to the base or to the path, and there is no header that says which.
+//
+// So discovery does not trust the URL it is given. It tries a small, ordered
+// set of candidates, records which one answered, and reports that back as the
+// normalized base_url + adapter endpoint pair. A caller who typed the wrong
+// form gets a working source anyway, plus the correction needed to make it
+// permanent.
+//
+// What is deliberately NOT here:
+//
+//   - No guessing of model capabilities from the name. "vision" in an id does
+//     not make a model accept images, and marking an image-only endpoint as
+//     kind:"chat" makes every AUTO request fail on it. Kind stays "chat" unless
+//     the upstream's own metadata says otherwise (see discoveredModelKind).
+//   - No live probe per model. A single-token completion per model would cost
+//     real money across a 34-model catalog and takes minutes. /models is the
+//     cheap, authoritative-enough signal for "this id is served here"; whether
+//     it actually answers stays for the normal request path to discover.
+//
+// The one thing that IS verified live is reachability, because a source that
+// 404s is not worth writing to disk.
+
+const (
+	// discoveryTimeout bounds the whole probe. Vendors are routinely slow on
+	// the first request to an uncached endpoint; 20s is enough for a healthy
+	// one and short enough that the UI is not left hanging.
+	discoveryTimeout = 20 * time.Second
+
+	// discoveryModelsPath is the OpenAI-compatible listing endpoint. Nearly
+	// every vendor proxies it verbatim, including the ones whose chat path
+	// differs.
+	discoveryModelsPath = "/v1/models"
+
+	// discoveryMaxBody caps the listing response. A catalog with thousands of
+	// models is still far under this; anything larger is a captive portal or
+	// an error page, not a model list.
+	discoveryMaxBody = 8 << 20
+)
+
+// discoveredModel is one entry of a parsed /models response, kept in the
+// upstream's own terms. The upstream is the only authority on what it serves.
+type discoveredModel struct {
+	ID   string
+	Kind string // "chat" | "image"
+	Meta map[string]interface{}
+}
+
+// discoveryResult reports what a probe found, including the case where nothing
+// worked — the failure text is what the operator needs, so it must name every
+// URL that was tried rather than just the last one.
+type discoveryResult struct {
+	BaseURL  string `json:"base_url"`
+	Endpoint string `json:"endpoint"`
+	// ImageEndpoint is reported separately because the gateway composes it
+	// differently: Provider.ImageURL is BaseURL + ImageEndpoint(), and the
+	// default ImageEndpoint is the ABSOLUTE path "/v1/images/generations".
+	// Chat's default is the relative "/chat/completions". So a base_url that
+	// already ends in /v1 produces a working chat URL and a broken image URL
+	// ("/v1/v1/images/generations") — which is exactly the 404 seen on the
+	// first vela configuration, and which discovery has to correct.
+	ImageEndpoint string   `json:"image_endpoint"`
+	Adapter       string   `json:"adapter"`
+	ModelsPath    string   `json:"models_path"`
+	Models        []string `json:"models"`
+	ImageModels   []string `json:"image_models"`
+	Tried         []string `json:"tried"`
+	LatencyMS     int64    `json:"latency_ms"`
+	Note          string   `json:"note,omitempty"`
+}
+
+// MarshalJSON guarantees the two list fields are always JSON arrays, never
+// null. The WebUI concatenates them directly, and "null" renders as the string
+// "undefined" in every row — a failure that only shows up in a browser and not
+// in a JSON schema check.
+//
+// Relying on the constructor to initialize the slices is not enough: the zero
+// value of this struct is a legitimate thing for a caller to build, and a
+// future code path that forgets the initializer would ship a UI bug silently.
+func (r discoveryResult) MarshalJSON() ([]byte, error) {
+	type alias discoveryResult // avoid recursion
+	out := alias(r)
+	if out.Models == nil {
+		out.Models = []string{}
+	}
+	if out.ImageModels == nil {
+		out.ImageModels = []string{}
+	}
+	if out.Tried == nil {
+		out.Tried = []string{}
+	}
+	return json.Marshal(out)
+}
+
+// discoveryCandidate is one (base_url, models-path) guess. Order matters: the
+// first candidate is the one reported as normalized when several work, because
+// it is the one that keeps base_url free of a version segment and therefore
+// composes correctly with the adapter's chat path.
+type discoveryCandidate struct {
+	baseURL    string
+	modelsPath string
+	// prefer ranks candidates that are more likely to be the vendor's real
+	// path. A 200 from a lower-preference candidate is still accepted, but
+	// only after every higher-preference one failed — see DiscoverModels.
+	prefer int
+}
+
+// discoveryFailure carries the HTTP status the caller should answer with.
+//
+// The status is part of the error rather than left to the handler because the
+// whole point of classifying a failure is that the operator must not have to
+// guess whether the key, the endpoint, or the vendor is at fault.
+type discoveryFailure struct {
+	msg    string
+	status int
+}
+
+func (e *discoveryFailure) Error() string { return e.msg }
+
+// StatusOrDefault reports the intended HTTP status.
+func (e *discoveryFailure) StatusOrDefault() int {
+	if e.status == 0 {
+		return http.StatusBadGateway
+	}
+	return e.status
+}
+
+// discoveryCandidates enumerates the URL forms worth trying for a raw input.
+//
+// The ranking encodes a real observation rather than a guess: the amr-link
+// endpoint serves its catalog at "/v1/models" (the form an SDK example shows)
+// from a base the operator pasted WITH the "/v1", while several other vendors
+// mount it at the host root. Both are probed; the one that answers is reported
+// back as a normalized base/endpoint pair.
+//
+// What we are really deciding is the base_url to persist. The safest one keeps
+// the version segment on the ENDPOINT, because openai's adapter endpoint is
+// "/chat/completions" and the two are concatenated: base "https://h" +
+// endpoint "/v1/chat/completions" is correct, while leaving "/v1" on the base
+// doubles the default image path into "/v1/v1/images/generations" — which is
+// exactly the 404 observed on the first vela configuration attempt.
+func discoveryCandidates(raw string) []discoveryCandidate {
+	raw = strings.TrimRight(strings.TrimSpace(raw), "/")
+	if raw == "" {
+		return nil
+	}
+	// Deduplicate on the URL that will actually be requested, not on the
+	// (base, path) pair. base "https://h/v1" + path "/models" and base "https://h"
+	// + path "/v1/models" both request the same address, and listing it twice in
+	// an error message makes the operator think two things were tried.
+	seenURL := map[string]bool{}
+	seenPair := map[string]bool{}
+	var out []discoveryCandidate
+	add := func(base, path string, prefer int) {
+		pair := base + "|" + path
+		if seenPair[pair] || seenURL[base+path] {
+			return
+		}
+		seenPair[pair] = true
+		seenURL[base+path] = true
+		out = append(out, discoveryCandidate{baseURL: base, modelsPath: path, prefer: prefer})
+	}
+
+	stripped := trimVersionSuffix(raw)
+	hasVersion := stripped != raw
+
+	// Both catalog paths are always tried, whichever form the operator typed.
+	//
+	// The earlier version offered "/models" only when the base already carried a
+	// version segment. That left a bare host ("https://api.vendor.com") with a
+	// SINGLE candidate and therefore no fallback at all — so a vendor serving
+	// its catalog at the bare "/models" was reported as unreachable. Every
+	// test used a "/v1"-suffixed base and so never saw it; siliconflow's
+	// "/v1/models" happens to answer, which is why the gap survived review.
+	//
+	// Preference only decides WHICH URL is reported when several work, never
+	// which ones are attempted.
+	//
+	// Order matters: a base that already ends in a version segment must be
+	// paired with the RELATIVE "/models" first, because pairing it with
+	// "/v1/models" produces "…/v1/v1/models" — a URL that is 404 by construction.
+	// Trying it first would spend a round trip on a certain miss and, worse,
+	// make the reported "tried" list lead with nonsense.
+	if hasVersion {
+		add(raw, "/models", 0)
+		add(raw, discoveryModelsPath, 1) // the doubled form; kept only as a last resort
+		add(stripped, "/models", 2)
+		add(stripped, discoveryModelsPath, 3)
+	} else {
+		add(raw, discoveryModelsPath, 0)
+		add(raw, "/models", 1)
+	}
+
+	// Stable insertion sort by preference; add() above is not ordered.
+	for i := 1; i < len(out); i++ {
+		for j := i; j > 0 && out[j].prefer < out[j-1].prefer; j-- {
+			out[j], out[j-1] = out[j-1], out[j]
+		}
+	}
+	return out
+}
+
+// trimVersionSuffix removes a trailing "/v1" (or /v2, /v1beta, ...) so the
+// host root can be probed on its own.
+func trimVersionSuffix(base string) string {
+	i := strings.LastIndex(base, "/")
+	if i < 0 {
+		return base
+	}
+	if !isVersionSegment(base[i+1:]) {
+		return base
+	}
+	return base[:i]
+}
+
+// versionSegment returns the leading path segment of a models path when that
+// segment is a version marker ("/v1/models" → "/v1", "/models" → ""). It is the
+// piece that must move between base_url and endpoint so the two compose.
+func versionSegment(modelsPath string) string {
+	parts := strings.Split(strings.Trim(modelsPath, "/"), "/")
+	if len(parts) != 2 || !isVersionSegment(parts[0]) {
+		return ""
+	}
+	return "/" + parts[0]
+}
+
+// isVersionSegment reports whether a path segment looks like "v1", "v2",
+// "v1beta" — and not like a real path name such as "vision" or a wildcard
+// path such as "v1x".
+//
+// The rule is: a leading "v", then digits, then an OPTIONAL suffix drawn from
+// the prerelease markers vendors actually use. A free-form alphabetic tail was
+// tried and rejected — it classifies "v1x" as a version, and that is how a
+// templated path segment would get stripped off a real base URL.
+func isVersionSegment(seg string) bool {
+	if len(seg) < 2 || seg[0] != 'v' {
+		return false
+	}
+	i := 1
+	digits := 0
+	for ; i < len(seg); i++ {
+		if seg[i] < '0' || seg[i] > '9' {
+			break
+		}
+		digits++
+	}
+	if digits == 0 {
+		return false
+	}
+	suffix := strings.ToLower(seg[i:])
+	if suffix == "" {
+		return true
+	}
+	for _, marker := range []string{"alpha", "beta", "preview", "dev", "rc"} {
+		if suffix == marker || strings.HasPrefix(suffix, "-"+marker) {
+			return true
+		}
+	}
+	return false
+}
+
+// composeEndpoint derives the chat endpoint to persist, given the models path
+// that actually answered, the base_url being persisted, and the base the
+// operator typed.
+//
+// The rule is a single invariant: exactly one version segment must end up in
+// base_url + endpoint. The adapter contributes "/chat/completions" and the
+// default image path contributes "/v1/images/generations", so a version left on
+// BOTH breaks one of them — and one left on NEITHER sends chat to the wrong
+// path on vendors that require the segment.
+func composeEndpoint(modelsPath, persistedBase, typedBase string) string {
+	// The version segment currently present on the persisted base.
+	onBase := ""
+	if b := trimVersionSuffix(persistedBase); b != persistedBase {
+		onBase = persistedBase[len(b):]
+	}
+	// The version segment present in the models path that answered.
+	inPath := versionSegment(modelsPath)
+
+	// The version segment actually WORKED: prefer the models path's, then the
+	// one the operator typed.
+	version := inPath
+	if version == "" && onBase == "" {
+		if b := trimVersionSuffix(typedBase); b != typedBase && typedBase != "" {
+			version = typedBase[len(b):]
+		}
+	}
+	if version == "" {
+		return "" // adapter default already composes
+	}
+	return version + "/chat/completions"
+}
+
+// composeImageEndpoint derives the image endpoint to persist.
+//
+// Provider.ImageURL is `base_url + ImageEndpoint()`, and the built-in default
+// is the absolute path "/v1/images/generations" — unlike chat's relative
+// "/chat/completions". So the two defaults do NOT agree about who owns the
+// version segment, which is the trap this function exists to close:
+//
+//	base "https://h"     + "/v1/images/generations"  → correct
+//	base "https://h/v1"  + "/v1/images/generations"  → "/v1/v1/…" 404
+//	base "https://h/v1"  + "/images/generations"     → "/v1/images/generations"  correct
+//
+// The second row is not a theoretical guess. Measured against the real
+// amr-link endpoint with base_url "https://amr-link.open-design.ai/v1":
+//
+//	…/v1/images/generations   → 200 (the image is then generated; that call
+//	                              takes minutes, which is the upstream's own
+//	                              latency and not a routing problem)
+//	…/v1/v1/images/generations → 404
+//	…/images/generations      → 404
+//
+// Leaving the default in place when the base already ends in a version is
+// therefore always wrong for such an upstream, and the version-relative form
+// is the one that composes.
+func composeImageEndpoint(persistedBase string) string {
+	if b := trimVersionSuffix(persistedBase); b != persistedBase {
+		return "/images/generations"
+	}
+	return ""
+}
+
+// DiscoverModels probes rawURL for a model catalog. apiKey is sent as a bearer
+// token; some gateways accept an unauthenticated listing, so an empty key is
+// allowed and simply omits the header.
+//
+// It returns on the first candidate that yields a parseable catalog, so the
+// reported latency is that of the working path, not the sum of the failures.
+func DiscoverModels(ctx context.Context, rawURL, apiKey string) (*discoveryResult, error) {
+	cands := discoveryCandidates(rawURL)
+	if len(cands) == 0 {
+		return nil, fmt.Errorf("base url required")
+	}
+
+	client := &http.Client{Timeout: discoveryTimeout}
+	var tried []string
+	var firstErr string
+	var authErr string
+	typedBase := strings.TrimRight(strings.TrimSpace(rawURL), "/")
+
+	var lastErr string
+	for _, c := range cands {
+		full := c.baseURL + c.modelsPath
+		tried = append(tried, full)
+
+		req, err := http.NewRequestWithContext(ctx, http.MethodGet, full, nil)
+		if err != nil {
+			firstErr = err.Error()
+			continue
+		}
+		req.Header.Set("Accept", "application/json")
+		if apiKey != "" {
+			req.Header.Set("Authorization", "Bearer "+apiKey)
+		}
+
+		start := time.Now()
+		resp, err := client.Do(req)
+		if err != nil {
+			lastErr = classifyTransport(err)
+			if lastErr == "" {
+				lastErr = err.Error()
+			}
+			continue
+		}
+		body, readErr := io.ReadAll(io.LimitReader(resp.Body, discoveryMaxBody))
+		resp.Body.Close()
+		took := time.Since(start)
+		if readErr != nil {
+			firstErr = readErr.Error()
+			continue
+		}
+		if resp.StatusCode != http.StatusOK {
+			// An auth rejection is the operator's key being wrong, not the
+			// upstream being down. Keep it: it is the single most common
+			// failure and 502 would send them to the wrong place.
+			if af := classifyAuthStatus(resp.StatusCode, body); af != "" {
+				authErr = af
+				continue
+			}
+			firstErr = fmt.Sprintf("%s returned %d", full, resp.StatusCode)
+			continue
+		}
+
+		models, err := parseModelCatalog(body)
+		if err != nil {
+			firstErr = fmt.Sprintf("%s: %v", full, err)
+			continue
+		}
+
+		res := &discoveryResult{
+			BaseURL:     c.baseURL,
+			ModelsPath:  c.modelsPath,
+			Adapter:     "openai",
+			Models:      []string{},
+			ImageModels: []string{},
+			Tried:       append([]string{}, tried...),
+			LatencyMS:   took.Milliseconds(),
+		}
+		// Persist a base/endpoint pair that composes correctly. The adapter's
+		// own endpoint is "/chat/completions", so any version segment has to be
+		// carried on the endpoint rather than left on the base (where it doubles
+		// the default image path into "/v1/v1/images/generations" — the exact
+		// 404 seen on the first vela configuration attempt).
+		res.Endpoint = composeEndpoint(c.modelsPath, c.baseURL, typedBase)
+		res.ImageEndpoint = composeImageEndpoint(c.baseURL)
+
+		// If the operator typed a version segment and we are persisting a base
+		// that no longer has it, the endpoint above is what restores it. If the
+		// base kept the segment, the adapter default is already correct.
+
+		seen := map[string]bool{}
+		for _, m := range models {
+			if m.ID == "" || seen[m.ID] {
+				continue
+			}
+			seen[m.ID] = true
+			if m.Kind == "image" {
+				res.ImageModels = append(res.ImageModels, m.ID)
+			} else {
+				res.Models = append(res.Models, m.ID)
+			}
+		}
+		sort.Strings(res.Models)
+		sort.Strings(res.ImageModels)
+		if len(res.Models) == 0 && len(res.ImageModels) == 0 {
+			firstErr = fmt.Sprintf("%s: catalog contained no usable model ids", full)
+			continue
+		}
+		if c.baseURL != typedBase {
+			res.Note = fmt.Sprintf("base url normalized from %s to %s; endpoint set to %q so it composes with the adapter",
+				typedBase, c.baseURL, res.Endpoint)
+		}
+		return res, nil
+	}
+	return nil, buildDiscoveryError(tried, authErr, firstErr, lastErr)
+}
+
+// classifyAuthStatus turns a 401/403 from the catalog endpoint into a message
+// that names the credential as the cause.
+//
+// This distinction is the whole point: a wrong key and an unreachable vendor
+// are the two most common discovery failures, and they need opposite actions
+// (re-type the key vs. go check the vendor). Reporting both as "upstream
+// error" makes the operator guess.
+func classifyAuthStatus(status int, body []byte) string {
+	if status != http.StatusUnauthorized && status != http.StatusForbidden {
+		return ""
+	}
+	snippet := strings.TrimSpace(string(body))
+	// Vendors put the reason in wildly different shapes; a compact quote is
+	// more useful than a parse attempt that fails on most of them.
+	snippet = strings.Join(strings.Fields(snippet), " ")
+	if len(snippet) > 200 {
+		snippet = snippet[:200] + "…"
+	}
+	return fmt.Sprintf("upstream rejected the api key (HTTP %d): %s — "+
+		"check the key, not the endpoint", status, snippet)
+}
+
+// classifyTransport turns low-level network failures into something an
+// operator can act on. An empty result means "no specific advice", and the
+// caller keeps the raw error.
+func classifyTransport(err error) string {
+	if err == nil {
+		return ""
+	}
+	msg := err.Error()
+	lower := strings.ToLower(msg)
+	switch {
+	case strings.Contains(lower, "no such host"):
+		return "hostname does not resolve: " + msg
+	case strings.Contains(lower, "certificate"), strings.Contains(lower, "tls"):
+		return "TLS verification failed: " + msg
+	case strings.Contains(lower, "connection refused"):
+		return "connection refused: " + msg
+	}
+	return ""
+}
+
+// buildDiscoveryError assembles the final error, preferring the most specific
+// diagnosis available. Every URL tried is listed: an operator staring at a
+// failed probe needs to know the search actually happened, not just the last
+// attempt.
+func buildDiscoveryError(tried []string, authErr, firstErr, lastErr string) error {
+	triedPart := "no model catalog found"
+	if len(tried) > 0 {
+		triedPart = "no model catalog found; tried " + strings.Join(tried, ", ")
+	}
+	if authErr != "" {
+		return &discoveryFailure{status: http.StatusUnauthorized, msg: authErr + " (" + triedPart + ")"}
+	}
+	detail := firstErr
+	if detail == "" {
+		detail = lastErr
+	}
+	if detail == "" {
+		return errors.New(triedPart)
+	}
+	if lastErr != "" && lastErr != firstErr {
+		detail += " (last: " + lastErr + ")"
+	}
+	return &discoveryFailure{status: http.StatusBadGateway, msg: triedPart + "; " + detail}
+}
+
+// upstreamModel is the subset of a /models entry this code reads. Vendors add
+// fields freely and rename the nesting; only id is universal, and the modality
+// hints below are read opportunistically.
+type modelCaps struct {
+	ImageGeneration bool `json:"image_generation"`
+	Vision          bool `json:"vision"`
+}
+
+type upstreamModel struct {
+	ID       string `json:"id"`
+	Object   string `json:"object"`
+	Kind     string `json:"kind"`
+	Type     string `json:"type"`
+	Modality string `json:"modality"`
+	Architecture struct {
+		InputModalities  []string `json:"input_modalities"`
+		OutputModalities []string `json:"output_modalities"`
+	} `json:"architecture"`
+	Capabilities modelCaps `json:"capabilities"`
+
+	// Metadata is kept as raw JSON on purpose. The real amr-link catalog nests
+	// cost/capability/context_limit here, and a typed struct would silently
+	// discard every key this build does not know about — the operator would
+	// then lose the context limit their vendor published. Raw also lets
+	// input_modalities/output_modalities be read without a second type.
+	Metadata json.RawMessage `json:"metadata"`
+}
+
+// metadataFields lists the keys worth surfacing in the UI. They are copied
+// through verbatim and nothing is inferred from them; an upstream that sends
+// no context limit must leave that absent rather than have this code write 0.
+var metadataFields = []string{
+	"context_length", "context_window", "max_context_length",
+	"context_limit", "context_budget", "cost", "capability",
+}
+
+// parseModelCatalog reads the OpenAI list shape {"data":[...]} and the bare
+// array shape some gateways return. A response that is valid JSON but has
+// neither shape is an error, not an empty catalog: reporting "0 models" for a
+// login page or a rate-limit JSON would silently create an empty source.
+func parseModelCatalog(body []byte) ([]discoveredModel, error) {
+	// "null" unmarshals into a nil slice without error, and an HTML login page
+	// fails to unmarshal at all. Both must be errors.
+	trimmed := strings.TrimSpace(string(body))
+	if trimmed == "" || trimmed == "null" {
+		return nil, fmt.Errorf("empty response")
+	}
+
+	var wrapper struct {
+		Data   []upstreamModel `json:"data"`
+		Models []upstreamModel `json:"models"`
+	}
+	if err := json.Unmarshal(body, &wrapper); err == nil {
+		if len(wrapper.Data) > 0 {
+			return toDiscovered(wrapper.Data), nil
+		}
+		if len(wrapper.Models) > 0 {
+			return toDiscovered(wrapper.Models), nil
+		}
+		// A well-formed but empty catalog is a legitimate answer from a key
+		// that has no models provisioned. Distinguish it from "not a catalog"
+		// by checking the shape actually parsed.
+		var probe map[string]json.RawMessage
+		if json.Unmarshal(body, &probe) == nil {
+			if _, ok := probe["data"]; ok {
+				return nil, nil
+			}
+			if _, ok := probe["models"]; ok {
+				return nil, nil
+			}
+		}
+	}
+	var bare []upstreamModel
+	if err := json.Unmarshal(body, &bare); err == nil {
+		return toDiscovered(bare), nil
+	}
+	return nil, fmt.Errorf("response is not a model catalog")
+}
+
+func toDiscovered(in []upstreamModel) []discoveredModel {
+	out := make([]discoveredModel, 0, len(in))
+	for _, m := range in {
+		out = append(out, discoveredModel{
+			ID:   m.ID,
+			Kind: discoveredModelKind(m),
+			Meta: modelMeta(m),
+		})
+	}
+	return out
+}
+
+// discoveredModelKind trusts the upstream's own modality fields and nothing
+// else. It never infers from the model NAME: ids like "gpt-image-1" or
+// "ideogram-v2" appear on catalogs that also serve chat, and a wrong kind
+// poisons the AUTO chain (an image-only id selected for a chat request fails
+// every tier behind it).
+func discoveredModelKind(m upstreamModel) string {
+	modIn := append(append([]string{}, m.Architecture.InputModalities...),
+		modalityList(m.Metadata, "input_modalities")...)
+	modOut := append(append([]string{}, m.Architecture.OutputModalities...),
+		modalityList(m.Metadata, "output_modalities")...)
+
+	has := func(list []string, want string) bool {
+		for _, v := range list {
+			if strings.EqualFold(strings.TrimSpace(v), want) {
+				return true
+			}
+		}
+		return false
+	}
+	if m.Capabilities.ImageGeneration || has(modOut, "image") {
+		return "image"
+	}
+	if has(modIn, "image") || has(modIn, "file") || has(modIn, "audio") || m.Capabilities.Vision {
+		return "chat" // multimodal input still means a chat model
+	}
+	// Some catalogs encode modality as a top-level string instead of a list.
+	// Only the image-ish spellings count: "chat" here means the upstream said
+	// chat, and treating that as "unknown" would ignore its statement.
+	switch strings.ToLower(strings.TrimSpace(m.Kind + " " + m.Modality + " " + m.Type)) {
+	case "image", "text-to-image", "image_generation", "text_to_image":
+		return "image"
+	}
+	return "chat"
+}
+
+// modalityList reads one array-of-strings field out of a raw JSON object,
+// tolerating both absence and a non-array value.
+func modalityList(raw json.RawMessage, field string) []string {
+	if len(raw) == 0 {
+		return nil
+	}
+	var obj map[string]json.RawMessage
+	if err := json.Unmarshal(raw, &obj); err != nil {
+		return nil
+	}
+	v, ok := obj[field]
+	if !ok {
+		return nil
+	}
+	var out []string
+	if err := json.Unmarshal(v, &out); err != nil {
+		return nil
+	}
+	return out
+}
+
+// modelMeta carries through the fields an operator wants to see in the UI
+// without this code inventing them. The upstream nests most of them under
+// "metadata" (the real amr-link catalog does) while some put the same keys at
+// the top level, so both are read and top level wins on a tie. Only keys the
+// upstream actually sent are copied, so an absent context limit stays absent
+// instead of becoming 0.
+func modelMeta(m upstreamModel) map[string]interface{} {
+	meta := map[string]interface{}{}
+
+	copyFields := func(raw json.RawMessage) {
+		if len(raw) == 0 {
+			return
+		}
+		var obj map[string]json.RawMessage
+		if err := json.Unmarshal(raw, &obj); err != nil {
+			return
+		}
+		for _, k := range metadataFields {
+			if v, ok := obj[k]; ok {
+				if dec := rawOrString(v); dec != nil {
+					meta[k] = dec
+				}
+			}
+		}
+	}
+
+	copyFields(m.Metadata)      // nested first
+	copyFields(mustJSON(m))     // top level wins
+	if len(meta) == 0 {
+		return nil
+	}
+	return meta
+}
+
+// rawOrString keeps numeric metadata as a number and everything else as its
+// decoded string, so the UI can sort on a cost or context length instead of
+// receiving "1048000" and having to coerce it.
+func rawOrString(raw json.RawMessage) interface{} {
+	var num json.Number
+	if err := json.Unmarshal(raw, &num); err == nil {
+		if f, err := num.Float64(); err == nil {
+			return f
+		}
+	}
+	var s string
+	if err := json.Unmarshal(raw, &s); err == nil {
+		return s
+	}
+	var any interface{}
+	if err := json.Unmarshal(raw, &any); err == nil {
+		return any
+	}
+	return nil
+}
+
+func mustJSON(v interface{}) []byte {
+	b, _ := json.Marshal(v)
+	return b
+}
+
+// modelsToConfig converts discovery output into config.Model entries, applying
+// the same priority convention the UI uses: chat models the operator is
+// likely to want first get a lower number, and image models sort last because
+// they are unreachable from /v1/chat/completions.
+func modelsToConfig(chat, image []string) []config.Model {
+	out := make([]config.Model, 0, len(chat)+len(image))
+	for _, id := range chat {
+		out = append(out, config.Model{ID: id, Priority: 50, Kind: "chat"})
+	}
+	for _, id := range image {
+		out = append(out, config.Model{ID: id, Priority: 70, Kind: "image"})
+	}
+	return out
+}

internal/gateway/discover_api.go

@@ -0,0 +1,403 @@
+package gateway
+
+import (
+	"context"
+	"encoding/json"
+	"errors"
+	"io"
+	"net/http"
+	"strings"
+	"time"
+	"llmsproxy/internal/config"
+)
+
+// HTTP surface for model discovery.
+//
+// Two endpoints, deliberately separate:
+//
+//	POST /api/sources/discover        — probe a URL+key, return the catalog.
+//	                                   Nothing is written.
+//	POST /api/sources/discover/apply  — probe, then write the source.
+//
+// The split exists because discovery mutates nothing and is therefore safe to
+// call repeatedly while the operator is still typing a key, whereas apply
+// rewrites credentials and the AUTO chain's inputs. Making them one endpoint
+// with a flag would mean a "just check it" click can silently create a source.
+//
+// Both are admin-only, matching /api/sources: the request body carries an API
+// key, and a model catalog reveals which upstreams an operator uses.
+
+// discoverRequest is the body of both endpoints.
+type discoverRequest struct {
+	// BaseURL is what the operator pasted. It may or may not include a
+	// version segment; discovery works that out rather than making the
+	// caller know this gateway's path-composition rules.
+	BaseURL string `json:"base_url"`
+	APIKey  string `json:"api_key"`
+
+	// Name is required for apply, ignored for discover. It defaults from the
+	// base URL's host when omitted, because "https://api.vendor.com/v1" is a
+	// better starting name than "new-source".
+	Name string `json:"name"`
+
+	// Adapter is optional. Discovery assumes "openai" because that is what a
+	// /v1/models catalog implies; a non-OpenAI upstream (Anthropic, Gemini)
+	// has no such endpoint, so discovery will legitimately fail for those and
+	// the operator configures them by hand.
+	Adapter string `json:"adapter"`
+
+	// Temperature and MaxConcurrent are carried into the written source.
+	Temperature   float64 `json:"temperature"`
+	MaxConcurrent int     `json:"max_concurrent"`
+
+	// ReplaceModels drops the source's existing model list before applying the
+	// discovered one. It defaults to true for a NEW source and is required to
+	// be spelled out for an existing one, so a re-discovery cannot silently
+	// delete hand-tuned priorities.
+	ReplaceModels *bool `json:"replace_models,omitempty"`
+
+	// Select narrows what gets written. Empty means everything discovered.
+	// Ids that are not in the catalog are an error rather than a silent skip:
+	// a typo here would produce a source that 404s on every request.
+	Include []string `json:"include,omitempty"`
+	Exclude []string `json:"exclude,omitempty"`
+}
+
+// discoverResponse is the JSON the WebUI binds to.
+type discoverResponse struct {
+	OK bool `json:"ok"`
+
+	// Normalized source fields, ready to be echoed back to /api/sources.
+	BaseURL string `json:"base_url"`
+	Adapter string `json:"adapter"`
+	Models  []string `json:"models"`
+	ImageModels []string `json:"image_models"`
+
+	// Endpoint is set only when discovery had to move a version segment.
+	Endpoint string `json:"endpoint"`
+	// ImageEndpoint is required whenever base_url carries a version segment:
+	// the gateway's default image path is the ABSOLUTE "/v1/images/generations"
+	// while chat's default is the relative "/chat/completions", so the same
+	// base_url that yields a working chat URL yields a 404 image URL.
+	ImageEndpoint string `json:"image_endpoint"`
+
+	// Diagnostics for the operator.
+	ModelsPath string   `json:"models_path"`
+	Tried      []string `json:"tried"`
+	LatencyMS  int64    `json:"latency_ms"`
+	Note       string   `json:"note,omitempty"`
+	Warnings   []string `json:"warnings,omitempty"`
+
+	// Set by apply only.
+	SourceName string `json:"source_name,omitempty"`
+	Applied    bool   `json:"applied,omitempty"`
+}
+
+func (g *Gateway) handleSourceDiscoverAPI(w http.ResponseWriter, r *http.Request) {
+	if reqRole(r.Context()) != "admin" {
+		writeError(w, http.StatusForbidden, "forbidden", "admin role required")
+		return
+	}
+	apply := strings.HasSuffix(r.URL.Path, "/apply")
+	if r.Method != http.MethodPost {
+		w.Header().Set("Allow", "POST")
+		writeError(w, http.StatusMethodNotAllowed, "method_not_allowed",
+			"use POST with {\"base_url\":…,\"api_key\":…}")
+		return
+	}
+
+	body, err := io.ReadAll(io.LimitReader(r.Body, 1<<20))
+	if err != nil {
+		writeError(w, http.StatusBadRequest, "invalid_request", "cannot read body")
+		return
+	}
+	var req discoverRequest
+	if err := json.Unmarshal(body, &req); err != nil {
+		writeError(w, http.StatusBadRequest, "invalid_request", "invalid json: "+err.Error())
+		return
+	}
+	req.BaseURL = strings.TrimSpace(req.BaseURL)
+	if req.BaseURL == "" {
+		writeError(w, http.StatusBadRequest, "invalid_request", "base_url required")
+		return
+	}
+	if req.Adapter == "" {
+		req.Adapter = "openai"
+	}
+
+	// Bounded so a hung upstream cannot pin the handler; discovery has its own
+	// per-request timeout on top.
+	ctx, cancel := context.WithTimeout(r.Context(), 60*time.Second)
+	defer cancel()
+
+	res, err := DiscoverModels(ctx, req.BaseURL, req.APIKey)
+	if err != nil {
+		// A rejected key must not be reported as an upstream outage: the two
+		// need opposite actions from the operator, and 502 sends them to the
+		// wrong one.
+		status := http.StatusBadGateway
+		var f *discoveryFailure
+		if errors.As(err, &f) {
+			status = f.StatusOrDefault()
+		}
+		writeError(w, status, "discover_failed", err.Error())
+		return
+	}
+
+	chat, image, warns, err := selectModels(res.Models, res.ImageModels, req.Include, req.Exclude)
+	if err != nil {
+		// 400: the operator asked for a model the catalog does not have.
+		writeError(w, http.StatusBadRequest, "invalid_request", err.Error())
+		return
+	}
+
+	out := &discoverResponse{
+		OK:          true,
+		BaseURL:     res.BaseURL,
+		Adapter:     req.Adapter,
+		Models:      chat,
+		ImageModels: image,
+		Endpoint:      res.Endpoint,
+		ImageEndpoint: res.ImageEndpoint,
+		ModelsPath:    res.ModelsPath,
+		Tried:       res.Tried,
+		LatencyMS:   res.LatencyMS,
+		Note:        res.Note,
+		Warnings:    warns,
+	}
+	if len(chat) == 0 && len(image) == 0 {
+		out.Warnings = append(out.Warnings,
+			"every discovered model was excluded; applying this would create a source with no models")
+	}
+
+	if !apply {
+		writeJSON(w, http.StatusOK, out)
+		return
+	}
+
+	name := strings.TrimSpace(req.Name)
+	if name == "" {
+		name = defaultSourceName(req.BaseURL)
+	}
+	if name == "" {
+		writeError(w, http.StatusBadRequest, "invalid_request",
+			"name required: cannot derive one from base_url")
+		return
+	}
+	if strings.ContainsAny(name, "/\\ \t") || name == "." || name == ".." {
+		writeError(w, http.StatusBadRequest, "invalid_request",
+			"source name must not contain slashes or whitespace")
+		return
+	}
+
+	exists := g.sourceExists(name)
+	if exists && req.ReplaceModels == nil {
+		// Refusing here is deliberate. Re-running discovery on an existing
+		// source and having it silently replace a hand-tuned model list —
+		// including priorities chosen for the AUTO chain — is the kind of
+		// "helpful" behaviour that loses an afternoon.
+		writeError(w, http.StatusConflict, "conflict",
+			"source "+name+" already exists; set replace_models=true to overwrite its model list, "+
+				"or choose a different name")
+		return
+	}
+
+	// Build the source directly rather than round-tripping through the
+	// /api/sources payload: resolveAPIKey carries the "an empty api_key means
+	// keep the stored one" rule, and going through a second decoder here would
+	// duplicate (and eventually diverge from) that logic.
+	src := config.Source{
+		Name:        name,
+		BaseURL:     res.BaseURL,
+		APIKey:      g.resolveAPIKey(name, req.APIKey, exists),
+		Adapter:     req.Adapter,
+		Endpoint:      res.Endpoint,
+		ImageEndpoint: res.ImageEndpoint,
+		Models:        modelsToConfig(chat, image),
+		Temperature: req.Temperature,
+	}
+	if req.MaxConcurrent > 0 {
+		src.MaxConcurrent = req.MaxConcurrent
+	}
+
+	// An existing source may carry fields discovery does not know about
+	// (proxy_url, headers, api_key_env, timeout). Rebuilding from scratch would
+	// drop them, so they are inherited unless the request overrides them.
+	if exists {
+		old := g.sourceByName(name)
+		if old != nil {
+			src.ProxyURL = old.ProxyURL
+			src.APIKeyEnv = old.APIKeyEnv
+			src.Timeout = old.Timeout
+			src.QueueTimeout = old.QueueTimeout
+			src.Headers = old.Headers
+			src.Meta = old.Meta
+			src.MaxTokens = old.MaxTokens
+			src.RPM = old.RPM
+			// image_endpoint is NOT inherited when discovery supplied one:
+			// the stored value was correct for the OLD base_url, and the base
+			// is exactly what may have changed. Inheriting it would keep a
+			// stale value that 404s against the new base.
+			if res.ImageEndpoint == "" {
+				src.ImageEndpoint = old.ImageEndpoint
+			}
+			if src.Adapter == "" {
+				src.Adapter = old.Adapter
+			}
+			if req.MaxConcurrent <= 0 {
+				src.MaxConcurrent = old.MaxConcurrent
+			}
+			// replace_models=false keeps the curated list. Discovery still
+			// refreshed base_url/endpoint/credential, which is the point of
+			// re-running it.
+			if req.ReplaceModels != nil && !*req.ReplaceModels {
+				src.Models = old.Models
+				out.Warnings = append(out.Warnings,
+					"model list left unchanged (replace_models=false)")
+			}
+		}
+	}
+
+	if err := g.core.AddSource(src); err != nil {
+		writeError(w, http.StatusBadRequest, "source_error", err.Error())
+		return
+	}
+
+	out.SourceName = name
+	out.Applied = true
+	writeJSON(w, http.StatusOK, out)
+}
+
+// selectModels applies the include/exclude filters.
+//
+// An include entry the catalog does not have is an error: silently dropping it
+// produces a source that looks configured and 404s at request time, which is
+// far harder to diagnose than a rejected apply. An exclude entry that is not
+// present is fine — the operator is allowed to be defensive.
+func selectModels(chat, image, include, exclude []string) ([]string, []string, []string, error) {
+	have := map[string]bool{}
+	for _, m := range chat {
+		have[m] = true
+	}
+	for _, m := range image {
+		have[m] = true
+	}
+
+	if len(include) > 0 {
+		var missing []string
+		for _, want := range include {
+			if !have[want] {
+				missing = append(missing, want)
+			}
+		}
+		if len(missing) > 0 {
+			return nil, nil, nil, &discoverError{
+				msg: "these models are not in the upstream catalog: " + strings.Join(missing, ", "),
+			}
+		}
+		keep := map[string]bool{}
+		for _, want := range include {
+			keep[want] = true
+		}
+		var c, i []string
+		for _, m := range chat {
+			if keep[m] {
+				c = append(c, m)
+			}
+		}
+		for _, m := range image {
+			if keep[m] {
+				i = append(i, m)
+			}
+		}
+		chat, image = c, i
+	}
+
+	if len(exclude) > 0 {
+		drop := map[string]bool{}
+		for _, m := range exclude {
+			drop[m] = true
+		}
+		var c, i []string
+		for _, m := range chat {
+			if !drop[m] {
+				c = append(c, m)
+			}
+		}
+		for _, m := range image {
+			if !drop[m] {
+				i = append(i, m)
+			}
+		}
+		chat, image = c, i
+	}
+
+	// chat and image must stay non-nil: the JSON contract forbids null here.
+	if chat == nil {
+		chat = []string{}
+	}
+	if image == nil {
+		image = []string{}
+	}
+
+	var warns []string
+	if len(chat) == 0 && len(image) == 0 {
+		// Caller can only get here by excluding everything, or from a catalog
+		// that was already empty. Either way "silence" would read as success.
+		warns = append(warns,
+			"no models selected; applying this would create a source that cannot serve any request")
+		return chat, image, warns, nil
+	}
+	if len(image) > 0 && len(chat) == 0 {
+		warns = append(warns,
+			"every model is an image model; AUTO chat requests have nothing to select")
+	}
+	if len(chat) > 0 && len(image) > 0 {
+		warns = append(warns,
+			"image models are only reachable through /v1/images/generations, not AUTO chat")
+	}
+	return chat, image, warns, nil
+}
+
+// discoverError is a client mistake rather than an upstream failure, which is
+// why it gets its own type instead of fmt.Errorf.
+type discoverError struct{ msg string }
+
+func (e *discoverError) Error() string { return e.msg }
+
+// defaultSourceName derives a source name from a base URL host, stripping the
+// version segment and any "api." prefix. It returns "" for input that is not a
+// URL at all — the caller must not fall back to echoing garbage into a source
+// name, since the name becomes a config key and an API path segment.
+func defaultSourceName(baseURL string) string {
+	s := strings.TrimSpace(baseURL)
+	if !strings.Contains(s, "://") {
+		return ""
+	}
+	s = s[strings.Index(s, "://")+3:]
+	if i := strings.IndexAny(s, "/?#"); i >= 0 {
+		s = s[:i]
+	}
+	if i := strings.LastIndex(s, ":"); i > 0 && !strings.Contains(s[i+1:], "]") {
+		s = s[:i] // strip port
+	}
+	// A host must still look like one after trimming the cosmetic prefixes.
+	s = strings.TrimPrefix(s, "api.")
+	s = strings.TrimPrefix(s, "www.")
+	if s == "" || strings.ContainsAny(s, " \t\"'<>") {
+		return ""
+	}
+	for _, r := range s {
+		switch {
+		case r >= 'a' && r <= 'z', r >= 'A' && r <= 'Z', r >= '0' && r <= '9':
+		case r == '.' || r == '-' || r == '_':
+		default:
+			return ""
+		}
+	}
+	if !strings.Contains(s, ".") {
+		// A bare label is far more likely to be a mistake than a hostname.
+		return ""
+	}
+	return s
+}

internal/gateway/discover_api_test.go

@@ -0,0 +1,314 @@
+package gateway
+
+import (
+	"context"
+	"encoding/json"
+	"errors"
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"testing"
+
+	"llmsproxy/internal/config"
+)
+
+// upstream that answers /v1/models with the real captured catalog
+func catalogUpstream(t *testing.T) *httptest.Server {
+	t.Helper()
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		if r.URL.Path != "/v1/models" {
+			w.WriteHeader(http.StatusNotFound)
+			return
+		}
+		if got := r.Header.Get("Authorization"); got != "Bearer sk-secret" {
+			w.WriteHeader(http.StatusUnauthorized)
+			return
+		}
+		_, _ = w.Write([]byte(realCatalogBody))
+	}))
+	t.Cleanup(srv.Close)
+	return srv
+}
+
+func TestDefaultSourceName(t *testing.T) {
+	cases := map[string]string{
+		"https://api.vendor.com/v1":   "vendor.com",
+		"https://api.openai.com/v1":   "openai.com",
+		"https://www.vendor.com/v1":   "vendor.com",
+		"http://127.0.0.1:8080/v1":    "127.0.0.1",
+		"https://amr-link.open-design.ai/v1": "amr-link.open-design.ai",
+		"not a url at all":             "",
+	}
+	for in, want := range cases {
+		if got := defaultSourceName(in); got != want {
+			t.Errorf("defaultSourceName(%q) = %q, want %q", in, got, want)
+		}
+	}
+}
+
+func TestSelectModels_IncludeUnknownIsAnError(t *testing.T) {
+	chat := []string{"a", "b"}
+	image := []string{"img"}
+	// Silently dropping a typo'd id would create a source that looks
+	// configured and 404s on every request — much worse than a rejection.
+	_, _, _, err := selectModels(chat, image, []string{"a", "typo"}, nil)
+	if err == nil {
+		t.Fatal("expected an error for an include id that is not in the catalog")
+	}
+	if !strings.Contains(err.Error(), "typo") {
+		t.Errorf("error should name the offending id, got %q", err)
+	}
+}
+
+func TestSelectModels_IncludeNarrows(t *testing.T) {
+	chat, image, _, err := selectModels([]string{"a", "b", "c"}, []string{"img1", "img2"}, []string{"b", "img1"}, nil)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(chat) != 1 || chat[0] != "b" {
+		t.Errorf("chat = %v, want [b]", chat)
+	}
+	if len(image) != 1 || image[0] != "img1" {
+		t.Errorf("image = %v, want [img1]", image)
+	}
+}
+
+func TestSelectModels_ExcludeUnknownIsFine(t *testing.T) {
+	chat, _, _, err := selectModels([]string{"a"}, nil, nil, []string{"never-existed"})
+	if err != nil {
+		t.Errorf("excluding a model that is not present should be allowed: %v", err)
+	}
+	if len(chat) != 1 {
+		t.Errorf("chat = %v, want [a] unchanged", chat)
+	}
+}
+
+func TestSelectModels_NeverReturnsNil(t *testing.T) {
+	chat, image, warns, err := selectModels(nil, nil, nil, nil)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if chat == nil || image == nil {
+		t.Fatal("nil slices would serialize as JSON null and break the UI")
+	}
+	if len(warns) == 0 {
+		t.Error("an empty catalog deserves a warning, not silence")
+	}
+}
+
+func TestSelectModels_WarnsAboutImageOnlyCatalog(t *testing.T) {
+	_, _, warns, err := selectModels(nil, []string{"gpt-image-2"}, nil, nil)
+	if err != nil {
+		t.Fatal(err)
+	}
+	joined := strings.Join(warns, " ")
+	if !strings.Contains(joined, "AUTO") {
+		t.Errorf("warnings %v should say AUTO has nothing to select", warns)
+	}
+}
+
+func TestSelectModels_WarnsThatImagesAreNotInAUTO(t *testing.T) {
+	_, _, warns, err := selectModels([]string{"a"}, []string{"img"}, nil, nil)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !strings.Contains(strings.Join(warns, " "), "images/generations") {
+		t.Errorf("warnings %v should point at the image endpoint", warns)
+	}
+}
+
+// TestDiscoveryResultAndConfigRoundTrip checks the fields the WebUI will echo
+// back actually survive into a config.Source — a mismatch here means the
+// operator previews one set of values and saves another.
+func TestDiscoveryResultAndConfigRoundTrip(t *testing.T) {
+	res := &discoveryResult{
+		BaseURL:    "https://h",
+		Endpoint:   "/v1/chat/completions",
+		Models:     []string{"m1", "m2"},
+		ImageModels: []string{"img1"},
+	}
+	src := config.Source{
+		Name:      "x",
+		BaseURL:   res.BaseURL,
+		Endpoint:  res.Endpoint,
+		Models:    modelsToConfig(res.Models, res.ImageModels),
+	}
+	if len(src.Models) != 3 {
+		t.Fatalf("models = %d, want 3", len(src.Models))
+	}
+	if src.Endpoint != "/v1/chat/completions" {
+		t.Errorf("endpoint lost: %q", src.Endpoint)
+	}
+	var img int
+	for _, m := range src.Models {
+		if m.Kind == "image" {
+			img++
+		}
+	}
+	if img != 1 {
+		t.Errorf("image models = %d, want 1", img)
+	}
+}
+
+// TestDiscoverModels_SendsAuthorizationHeader pins the only way the feature
+// works against a real vendor: without the bearer token every paid /models
+// endpoint returns 401.
+func TestDiscoverModels_SendsAuthorizationHeader(t *testing.T) {
+	var seen string
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		seen = r.Header.Get("Authorization")
+		_, _ = w.Write([]byte(`{"data":[{"id":"m"}]}`))
+	}))
+	defer srv.Close()
+
+	if _, err := DiscoverModels(context.Background(), srv.URL+"/v1", "sk-secret"); err != nil {
+		t.Fatal(err)
+	}
+	if seen != "Bearer sk-secret" {
+		t.Errorf("Authorization = %q, want %q", seen, "Bearer sk-secret")
+	}
+}
+
+// TestDiscoverModels_WrongKeyIsUnauthorizedNotBadGateway pins the distinction
+// that decides whether the operator knows what to do next.
+//
+// A rejected key and an unreachable vendor are the two most common discovery
+// failures, and they need opposite actions: re-type the key, versus go check the
+// vendor's status page. Both used to surface as one generic error, so the
+// operator had to guess.
+func TestDiscoverModels_WrongKeyIsUnauthorizedNotBadGateway(t *testing.T) {
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.Header().Set("Content-Type", "application/json")
+		w.WriteHeader(http.StatusUnauthorized)
+		_, _ = w.Write([]byte(`{"error":{"message":"invalid api key","code":"invalid_api_key"}}`))
+	}))
+	defer srv.Close()
+
+	_, err := DiscoverModels(context.Background(), srv.URL+"/v1", "sk-wrong")
+	if err == nil {
+		t.Fatal("expected a failure against a 401 endpoint")
+	}
+	var f *discoveryFailure
+	if !errors.As(err, &f) {
+		t.Fatalf("error should be classified, got %T: %v", err, err)
+	}
+	if got := f.StatusOrDefault(); got != http.StatusUnauthorized {
+		t.Errorf("status = %d, want %d — a wrong key must not read as an upstream outage",
+			got, http.StatusUnauthorized)
+	}
+	// The message has to name which credential is at fault, or the status code
+	// alone still sends the operator to the wrong place.
+	if !strings.Contains(strings.ToLower(f.Error()), "api key") {
+		t.Errorf("message %q should name the api key as the cause", f.Error())
+	}
+	if !strings.Contains(f.Error(), "/v1/models") {
+		t.Errorf("message %q should list what was tried", f.Error())
+	}
+}
+
+// TestDiscoverModels_UnreachableStaysBadGateway: the classification must not
+// over-correct — an unresolvable host is still the vendor's or the network's
+// problem, not the key's.
+func TestDiscoverModels_UnreachableStaysBadGateway(t *testing.T) {
+	_, err := DiscoverModels(context.Background(), "https://no-such-host-xyz.invalid/v1", "k")
+	if err == nil {
+		t.Fatal("expected a failure for an unresolvable host")
+	}
+	var f *discoveryFailure
+	if !errors.As(err, &f) {
+		t.Fatalf("error should be classified, got %T", err)
+	}
+	if got := f.StatusOrDefault(); got != http.StatusBadGateway {
+		t.Errorf("status = %d, want %d", got, http.StatusBadGateway)
+	}
+	if !strings.Contains(strings.ToLower(f.Error()), "resolve") {
+		t.Errorf("message %q should say the hostname does not resolve", f.Error())
+	}
+}
+
+// TestDiscoveryCandidates_NoDuplicateRequests: base "https://h/v1" with path
+// "/models" and base "https://h" with path "/v1/models" request the SAME URL.
+// Listing it twice in a failure message reads as "two things were tried" when
+// only one was.
+func TestDiscoveryCandidates_NoDuplicateRequests(t *testing.T) {
+	seen := map[string]bool{}
+	for _, c := range discoveryCandidates("https://h/v1") {
+		url := c.baseURL + c.modelsPath
+		if seen[url] {
+			t.Errorf("candidate %q would be requested twice", url)
+		}
+		seen[url] = true
+	}
+	if len(seen) < 2 {
+		t.Errorf("only %d distinct URLs; the search collapsed", len(seen))
+	}
+}
+
+// TestDiscoverModels_FailureListsEachURLOnce is the end-to-end form of the same
+// rule, on the error text an operator actually reads.
+func TestDiscoverModels_FailureListsEachURLOnce(t *testing.T) {
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.WriteHeader(http.StatusNotFound)
+	}))
+	defer srv.Close()
+
+	_, err := DiscoverModels(context.Background(), srv.URL+"/v1", "k")
+	if err == nil {
+		t.Fatal("expected failure")
+	}
+	msg := err.Error()
+	needle := srv.URL + "/v1/models"
+	if n := strings.Count(msg, needle); n > 1 {
+		t.Errorf("URL %s appears %d times in %q", needle, n, msg)
+	}
+}
+
+func TestDiscoverModels_EmptyKeyOmitsHeader(t *testing.T) {
+	var present bool
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		_, present = r.Header["Authorization"]
+		_, _ = w.Write([]byte(`{"data":[{"id":"m"}]}`))
+	}))
+	defer srv.Close()
+
+	if _, err := DiscoverModels(context.Background(), srv.URL+"/v1", ""); err != nil {
+		t.Fatal(err)
+	}
+	if present {
+		t.Error("an empty key should omit the header, not send an empty bearer")
+	}
+}
+
+// TestDiscoverModels_RealCatalogEndToEnd walks the actual captured response
+// through the whole path and asserts the numbers an operator would see.
+func TestDiscoverModels_RealCatalogEndToEnd(t *testing.T) {
+	srv := catalogUpstream(t)
+	res, err := DiscoverModels(context.Background(), srv.URL+"/v1", "sk-secret")
+	if err != nil {
+		t.Fatalf("discovery failed: %v", err)
+	}
+	if res.BaseURL != srv.URL+"/v1" {
+		t.Errorf("base = %q, want the typed URL unchanged", res.BaseURL)
+	}
+	if res.Endpoint != "" {
+		t.Errorf("endpoint = %q, want empty (base already carries /v1)", res.Endpoint)
+	}
+	if len(res.Models) != 2 || len(res.ImageModels) != 2 {
+		t.Errorf("chat/image = %d/%d, want 2/2", len(res.Models), len(res.ImageModels))
+	}
+	if res.Adapter != "openai" {
+		t.Errorf("adapter = %q, want openai", res.Adapter)
+	}
+
+	// The response must be JSON the UI can consume directly.
+	b, _ := json.Marshal(res)
+	var round map[string]interface{}
+	if err := json.Unmarshal(b, &round); err != nil {
+		t.Fatal(err)
+	}
+	if round["ok"] != nil {
+		// discoverResponse adds ok; discoveryResult does not. Both marshal
+		// fine, so this is only a shape note.
+		_ = round
+	}
+}

internal/gateway/discover_endpoint_test.go

@@ -0,0 +1,153 @@
+package gateway
+
+import (
+	"context"
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"testing"
+)
+
+// TestComposeImageEndpoint_NoVersionBaseLeavesDefault is the counterpart of the
+// /v1 case: when the base carries no version segment, discovery must emit "" so
+// the gateway's own default "/v1/images/generations" composes.
+//
+// This is the shape every currently-working source uses — siliconflow, the only
+// source in production with kind:image, has base_url
+// "https://api.siliconflow.cn" — so writing a non-empty value here would break
+// exactly the configuration that is known to work today.
+func TestComposeImageEndpoint_NoVersionBaseLeavesDefault(t *testing.T) {
+	if got := composeImageEndpoint("https://api.siliconflow.cn"); got != "" {
+		t.Errorf("composeImageEndpoint = %q, want \"\" so the built-in default composes", got)
+	}
+	// And the composition it implies must be the vendor's real path.
+	base := "https://api.siliconflow.cn"
+	composed := base + "/v1/images/generations"
+	if !strings.HasSuffix(composed, "/v1/images/generations") {
+		t.Errorf("composed URL %q lost the image path", composed)
+	}
+}
+
+// TestDiscovery_BareHostNeedsVersionOnEndpointWhenCatalogIsVersioned
+// distinguishes two cases that look identical from the outside.
+//
+//   bare host + catalog at "/models"       → chat is "/chat/completions" (the
+//     adapter default, no endpoint written)
+//   bare host + catalog at "/v1/models"    → chat is "/v1/chat/completions",
+//     so an endpoint MUST be written
+//
+// The catalog path is the only evidence of whether the upstream wants the
+// version segment, and getting this backwards sends every chat request to a
+// path the vendor does not serve.
+func TestDiscovery_BareHostNeedsVersionOnEndpointWhenCatalogIsVersioned(t *testing.T) {
+	cases := []struct {
+		name         string
+		catalogPath  string
+		wantEndpoint string
+		wantImage    string
+	}{
+		{
+			name:         "catalog at the bare /models needs no endpoint",
+			catalogPath:  "/models",
+			wantEndpoint: "",
+			wantImage:    "",
+		},
+		{
+			name:         "catalog at /v1/models requires /v1 on the chat endpoint",
+			catalogPath:  "/v1/models",
+			wantEndpoint: "/v1/chat/completions",
+			wantImage:    "",
+		},
+	}
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+				if r.URL.Path != tc.catalogPath {
+					w.WriteHeader(http.StatusNotFound)
+					return
+				}
+				_, _ = w.Write([]byte(`{"data":[{"id":"m1"},{"id":"m2"}]}`))
+			}))
+			defer srv.Close()
+
+			res, err := DiscoverModels(context.Background(), srv.URL, "k")
+			if err != nil {
+				t.Fatalf("discovery failed: %v", err)
+			}
+			if res.BaseURL != srv.URL {
+				t.Errorf("base = %q, want %q", res.BaseURL, srv.URL)
+			}
+			if res.Endpoint != tc.wantEndpoint {
+				t.Errorf("endpoint = %q, want %q", res.Endpoint, tc.wantEndpoint)
+			}
+			if res.ImageEndpoint != tc.wantImage {
+				t.Errorf("image_endpoint = %q, want %q", res.ImageEndpoint, tc.wantImage)
+			}
+			// Whatever it wrote, the chat URL must reach the same version
+			// prefix the catalog was served under.
+			ep := res.Endpoint
+			if ep == "" {
+				ep = "/chat/completions"
+			}
+			if got := res.BaseURL + ep; strings.Contains(got, "/v1/v1") {
+				t.Errorf("chat URL %q has a doubled version", got)
+			}
+		})
+	}
+}
+
+// TestComposeEndpoints_BothFormsProduceWorkingURLs is the end-to-end
+// invariant across the two shapes, stated as the URLs the gateway will build.
+//
+// A base with a version and no image_endpoint produces "/v1/v1/…" — the exact
+// 404 observed before composeImageEndpoint existed. A base without one and a
+// non-empty image_endpoint produces a path missing the version. Both are
+// silent, expensive mistakes, so they are asserted literally.
+func TestComposeEndpoints_BothFormsProduceWorkingURLs(t *testing.T) {
+	cases := []struct {
+		name        string
+		base        string
+		wantChat    string
+		wantImage   string
+	}{
+		{
+			name:      "bare host: adapter and image defaults compose",
+			base:      "https://h",
+			wantChat:  "https://h/chat/completions",
+			// openai's adapter endpoint is /chat/completions (verified in
+			// internal/lua/adapters/openai.lua: adapter.endpoint).
+			wantImage: "https://h/v1/images/generations",
+		},
+		{
+			name:      "versioned base: the version moves onto both endpoints",
+			base:      "https://h/v1",
+			wantChat:  "https://h/v1/chat/completions",
+			wantImage: "https://h/v1/images/generations",
+		},
+	}
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			ep := composeEndpoint("/models", tc.base, tc.base)
+			if ep == "" {
+				ep = "/chat/completions" // the adapter default
+			}
+			iep := composeImageEndpoint(tc.base)
+			if iep == "" {
+				iep = "/v1/images/generations" // the gateway default
+			}
+			gotChat := tc.base + ep
+			gotImage := tc.base + iep
+			if gotChat != tc.wantChat {
+				t.Errorf("chat URL = %q, want %q", gotChat, tc.wantChat)
+			}
+			if gotImage != tc.wantImage {
+				t.Errorf("image URL = %q, want %q", gotImage, tc.wantImage)
+			}
+			for _, u := range []string{gotChat, gotImage} {
+				if strings.Contains(u, "/v1/v1") {
+					t.Errorf("%q has a doubled version segment", u)
+				}
+			}
+		})
+	}
+}

internal/gateway/discover_mutate_test.go

@@ -0,0 +1,303 @@
+package gateway
+
+// Real mutation injection: rewrites the IMPLEMENTATION on disk, re-runs the
+// suite, and requires it to go red.
+//
+// Why this exists and why it is a Go file rather than a shell script:
+//
+// The first version was a shell script. Its -run filter contained a regex that
+// the shell mangled, so every mutant reported "SURVIVED" while go test printed
+// "[no tests to run]". Four mutants looked like judgement failures and were
+// actually never executed. A judgement that cannot tell "tests passed" from
+// "no tests ran" is not a judgement.
+//
+// This version therefore refuses to report anything unless it has first proved
+// the filter selects a non-trivial number of tests (see baseline below), and it
+// treats "[no tests to run]" as a hard error rather than a green run.
+//
+// Run: go test -tags luajit ./internal/gateway/ -run TestRealMutants -v
+
+import (
+	"os"
+	"os/exec"
+	"path/filepath"
+	"sort"
+	"strings"
+	"testing"
+)
+
+// mutationFilter must select every assertion that guards the code below. It is
+// deliberately broad: a narrow filter is how mutants get to survive for free.
+const mutationFilter = "Test(ParseModelCatalog|DiscoveredModelKind|Discover|TrimVersion|VersionSegment|ComposeEndpoint|ComposeImageEndpoint|ComposeEndpoints|IsVersionSegment|ModelsToConfig|DiscoveryResult|DiscoveryCandidates|Discovery_BareHost)"
+
+type mutation struct {
+	name string
+	from string
+	to   string
+}
+
+func TestRealMutants(t *testing.T) {
+	if testing.Short() {
+		t.Skip("mutation campaign compiles the package repeatedly")
+	}
+	wd, err := os.Getwd()
+	if err != nil {
+		t.Fatal(err)
+	}
+	target := filepath.Join(wd, "discover.go")
+	orig, err := os.ReadFile(target)
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer func() {
+		if werr := os.WriteFile(target, orig, 0o644); werr != nil {
+			t.Fatalf("restore discover.go: %v", werr)
+		}
+	}()
+
+	// --- prove the harness is armed before trusting any verdict ---
+	// A filter that selects nothing exits 0, which is indistinguishable from a
+	// real pass. The only way to tell them apart is to compare against a filter
+	// KNOWN to be empty: if that empty filter produces the same result as the
+	// real one, the real one is probably empty too.
+	//
+	// Note go test does NOT print "[no tests to run]" for a wholly unmatched
+	// filter (it prints that only when the package has other matching tests), so
+	// the check has to compare outputs rather than look for a marker string.
+	// Getting this backwards made the harness refuse to run at all.
+	emptyOut := runSuite(wd, "TestZZZNoSuchTestExists")
+	realOut := runSuite(wd, mutationFilter)
+	if countTests(emptyOut) != 0 {
+		t.Fatalf("a filter matching nothing selected %d tests; the harness's own probe is broken",
+			countTests(emptyOut))
+	}
+	// Kept for the no-op check below: a mutant whose output matches the
+	// baseline never changed behaviour, so it is an invalid mutation rather
+	// than a gap in the tests.
+	baseOut := realOut
+	if strings.Contains(realOut, "[no tests to run]") {
+		t.Fatalf("baseline selected no tests:\n%s", realOut)
+	}
+	baseCount := countTests(realOut)
+	if baseCount < 25 {
+		t.Fatalf("baseline ran only %d tests; the filter is too narrow to judge anything", baseCount)
+	}
+	if strings.Contains(realOut, "FAIL") {
+		t.Logf("note: baseline is not green; kill verdicts remain valid")
+	}
+	t.Logf("baseline: %d tests selected, empty filter selects %d",
+		baseCount, countTests(emptyOut))
+
+	muts := []mutation{
+		{
+			// Written as a condition that still READS modOut, because a
+			// mutation that leaves a variable unused fails to compile — and a
+			// build failure is not a judgement (see the harness note above).
+			name: "image kind ignores upstream output modality",
+			from: `if m.Capabilities.ImageGeneration || has(modOut, "image") {`,
+			to:   "if m.Capabilities.ImageGeneration || (len(modOut) > 0 && false) {",
+		},
+		{
+			name: "kind defaults to image (guessing from names)",
+			from: "\t}\n\treturn \"chat\"\n}\n\n// modalityList",
+			to:   "\t}\n\treturn \"image\"\n}\n\n// modalityList",
+		},
+		{
+			name: "nested metadata no longer read",
+			from: "\tcopyFields(m.Metadata)      // nested first",
+			to:   "\tcopyFields(json.RawMessage(nil)) // mutant: nested metadata dropped",
+		},
+		{
+			name: "non-catalog bodies accepted",
+			from: "\tif trimmed == \"\" || trimmed == \"null\" {\n\t\treturn nil, fmt.Errorf(\"empty response\")\n\t}",
+			to:   "\tif false && (trimmed == \"\" || trimmed == \"null\") {\n\t\treturn nil, fmt.Errorf(\"empty response\")\n\t}",
+		},
+		{
+			// Removing the bare "/models" candidate must be observable: a vendor
+			// serving its catalog at the bare path would otherwise look
+			// unreachable. The `len(...) >= 0` is a tautology that only exists to
+			// keep the expression syntactically valid after the call is dropped.
+			name: "URL fallback search removed",
+			from: "\t\tadd(raw, \"/models\", 0)",
+			to:   "\t\tif len(\"/models\") >= 0 {\n\t\t\t_ = hasVersion\n\t\t}",
+		},
+		{
+			name: "version segment never moved onto the chat endpoint",
+			from: "\tif version == \"\" {\n\t\treturn \"\" // adapter default already composes\n\t}",
+			to:   "\tif version == \"\" || len(version) > 0 {\n\t\treturn \"\"\n\t}",
+		},
+		{
+			name: "image endpoint never corrected (/v1/v1 regression)",
+			from: "\tif b := trimVersionSuffix(persistedBase); b != persistedBase {\n\t\treturn \"/images/generations\"\n\t}",
+			to:   "\tif b := trimVersionSuffix(persistedBase); b == b && b != b {\n\t\treturn \"/images/generations\"\n\t}",
+		},
+		{
+			name: "version test loosened so 'vision' counts as a version",
+			from: "\tsuffix := strings.ToLower(seg[i:])",
+			to:   "\tsuffix := strings.ToLower(seg[i:])\n\tif len(suffix) >= 0 {\n\t\tsuffix = \"beta\"\n\t}",
+		},
+		{
+			// The guard must be removable WITHOUT the mutant becoming a
+			// semantic no-op. Two earlier attempts failed:
+			//
+			//   `&& out.BaseURL != ""`  — the guard this test exercises uses a
+			//     zero-value struct whose BaseURL is empty, so the added
+			//     condition made the mutant behave identically to the original.
+			//   `&& len(out.BaseURL) >= 0` — always true for any string, so the
+			//     mutant was STILL an identity transform.
+			//
+			// Reading the variable into a blank assignment keeps the compiler
+			// happy without changing behaviour, which is what makes the
+			// mutant observable.
+			name: "nil slice guard removed from JSON output",
+			from: "\tif out.ImageModels == nil {\n\t\tout.ImageModels = []string{}\n\t}",
+			to:   "\tif out.ImageModels == nil {\n\t\t_ = out.ImageModels\n\t}",
+		},
+		{
+			name: "auth rejection no longer classified",
+			from: "\tif status != http.StatusUnauthorized && status != http.StatusForbidden {\n\t\treturn \"\"\n\t}",
+			to:   "\tif status != http.StatusUnauthorized && status != http.StatusForbidden || len(body) >= 0 {\n\t\treturn \"\"\n\t}",
+		},
+		{
+			name: "duplicate candidate URLs no longer collapsed",
+			from: "\t\tif seenPair[pair] || seenURL[base+path] {",
+			to:   "\t\tif len(pair) > 0 && (seenPair[pair] || seenURL[base+path]) && false {",
+		},
+	}
+
+	survived := 0
+	for _, m := range muts {
+		t.Run(m.name, func(t *testing.T) {
+			if !strings.Contains(string(orig), m.from) {
+				t.Fatalf("anchor not found in discover.go — this mutation would be a no-op:\n%q", m.from)
+			}
+			mutated := strings.Replace(string(orig), m.from, m.to, 1)
+			if mutated == string(orig) {
+				t.Fatal("replacement changed nothing")
+			}
+			if werr := os.WriteFile(target, []byte(mutated), 0o644); werr != nil {
+				t.Fatal(werr)
+			}
+			defer os.WriteFile(target, orig, 0o644)
+
+			out := runSuite(wd, mutationFilter)
+			// A build failure kills EVERY mutant, so it would mask a broken
+			// harness as a perfect score. Distinguish it from a real test
+			// failure and fail loudly instead.
+			if strings.Contains(out, "[build failed]") {
+				t.Fatalf("mutant broke the build — that is not a judgement, it is a broken harness:\n%s", firstLines(out, 20))
+			}
+			if countTests(out) == 0 {
+				t.Fatalf("filter selected nothing — the verdict would be meaningless")
+			}
+			if !strings.Contains(out, "FAIL") {
+				// The mutant compiled and the suite stayed green. Before calling
+				// this a survivor, check the mutation was not a no-op: an earlier
+				// version reported four "survivors" that were really invalid
+				// mutations whose added guard made them behaviourally identical
+				// to the original. A silent no-op is a broken mutation, and
+				// conflating the two sends the next reader hunting for a missing
+				// test that does not exist.
+				if behaviourOnly(out) == behaviourOnly(baseOut) {
+					t.Fatalf("mutant is a no-op: the suite output is identical to the "+
+						"baseline, so this mutation never changed behaviour. Fix the "+
+						"mutation rather than looking for a missing test.\nmutant: %q", m.to)
+				}
+				survived++
+				t.Errorf("SURVIVED %s — no test noticed this wrong implementation", m.name)
+				return
+			}
+			t.Logf("KILL  %s", m.name)
+		})
+	}
+	if survived > 0 {
+		t.Errorf("%d of %d mutants survived", survived, len(muts))
+	}
+}
+
+// runSuite invokes the go tool on this package with the given -run filter.
+//
+// The luajit build tag is mandatory: without it the package falls back to plain
+// Lua 5.1, billing.lua's `goto` fails to compile, and the whole package fails
+// to build — which the harness would misread as "every mutant killed".
+func runSuite(wd, filter string) string {
+	tool := os.Getenv("GO_TEST_GO")
+	if tool == "" {
+		tool = "go"
+	}
+	cmd := exec.Command(tool, "test",
+		"-tags", "luajit",
+		".", "-run", filter, "-count=1", "-v")
+	cmd.Dir = wd
+	cmd.Env = os.Environ()
+	out, _ := cmd.CombinedOutput()
+	return string(out)
+}
+
+func countTests(out string) int {
+	n := 0
+	for _, line := range strings.Split(out, "\n") {
+		// Subtests are indented; only count top-level ones so a single test
+		// with a table of cases cannot inflate the count.
+		if strings.HasPrefix(line, "=== RUN   Test") && !strings.Contains(line, "/") {
+			n++
+		}
+	}
+	return n
+}
+
+// behaviourOnly strips the parts of `go test -v` output that vary between runs
+// (timings, goroutine ids, temp paths) so two runs can be compared for equality.
+// If a mutant produces output identical to the baseline under this projection,
+// the mutation did not change observable behaviour and is therefore invalid.
+func behaviourOnly(out string) string {
+	var kept []string
+	for _, line := range strings.Split(out, "\n") {
+		trimmed := strings.TrimSpace(line)
+		switch {
+		// Keep only the outcome lines. Test NAMES are deliberately dropped:
+		// two runs of the same suite can legitimately order or cache tests
+		// differently, and keeping the names made every mutant look like it had
+		// "changed behaviour" — which defeats the no-op detector entirely.
+		case strings.HasPrefix(trimmed, "--- PASS"), strings.HasPrefix(trimmed, "--- FAIL"),
+			strings.HasPrefix(trimmed, "--- SKIP"),
+			strings.HasPrefix(trimmed, "ok  "), trimmed == "FAIL",
+			strings.Contains(trimmed, ".go:"):
+			kept = append(kept, trimmed)
+		}
+	}
+	// Counting outcomes is stronger than comparing the sequence: a mutation
+	// that swaps which test fails still changes the multiset, and one that
+	// changes nothing cannot.
+	sort.Strings(kept)
+	return strings.Join(kept, "\n")
+}
+
+// firstLines trims output for an error message; a failing nested run can be
+// thousands of lines and the cause is always near the top.
+func firstLines(s string, n int) string {
+	lines := strings.Split(s, "\n")
+	if len(lines) > n {
+		lines = lines[:n]
+	}
+	return strings.Join(lines, "\n")
+}
+
+// TestMutationFilterCoversEveryGuard keeps the filter honest. Dropping a group
+// from it silently disarms that area's mutants.
+func TestMutationFilterCoversEveryGuard(t *testing.T) {
+	for _, group := range []string{
+		"ParseModelCatalog",   // nested metadata + non-catalog rejection
+		"DiscoveredModelKind", // kind follows upstream metadata
+		"Discover",            // URL fallback, auth classification
+		"ComposeEndpoint",     // version segment exactly once (chat)
+		"ComposeImageEndpoint",
+		"ComposeEndpoints", // the composed URLs, both base shapes
+		"IsVersionSegment", // "v1beta" vs "vision"
+		"DiscoveryResult",  // nil slices are arrays
+	} {
+		if !strings.Contains(mutationFilter, group) {
+			t.Errorf("mutation filter omits %s; mutants there survive silently", group)
+		}
+	}
+}

internal/gateway/discover_mutation_test.go

@@ -0,0 +1,171 @@
+package gateway
+
+// Mutation harness for the discovery code. Kept in the tree (not in a shell
+// script) so the -run filter is a Go string literal rather than something the
+// shell can mangle into "[no tests to run]" — which is exactly how an earlier
+// version of this harness reported four surviving mutants that were simply
+// never executed.
+//
+// Each mutant flips one behaviour; TestMutationHarness_Fails proves the
+// harness itself is armed.
+
+import (
+	"encoding/json"
+	"net/http"
+	"strings"
+	"testing"
+)
+
+// discoveryTestFilter must cover every assertion that guards the behaviour
+// below. Narrowing it is how mutants get to "survive" for free.
+const discoveryTestFilter = "Test(ParseModelCatalog|DiscoveredModelKind|Discover|TrimVersion|VersionSegment|ComposeEndpoint|IsVersionSegment|ModelsToConfig|DiscoveryResult|DiscoveryCandidates)"
+
+func TestDiscoveryFilterRunsRealTests(t *testing.T) {
+	if discoveryTestFilter == "" {
+		t.Fatal("filter is empty; every mutant would survive trivially")
+	}
+	// The whole point: the catalog parser test must be inside the filter, or a
+	// mutant that stops reading nested metadata survives with a green run.
+	for _, group := range []string{
+		"ParseModelCatalog",   // nested metadata + non-catalog rejection
+		"DiscoveredModelKind", // kind must follow upstream metadata
+		"Discover",            // URL fallback + endpoint composition
+		"ComposeEndpoint",     // version segment exactly once
+		"IsVersionSegment",    // "v1beta" vs "vision"
+	} {
+		if !strings.Contains(discoveryTestFilter, group) {
+			t.Errorf("filter %q skips %s tests; mutants in that area survive", discoveryTestFilter, group)
+		}
+	}
+}
+
+// TestMutantsAreDetectable is the harness's own regression guard: it runs the
+// filter's worth of assertions against deliberately broken inputs and requires
+// each to be rejected. If a future edit makes one of these pass, the
+// corresponding protection is gone.
+func TestMutantsAreDetectable(t *testing.T) {
+	t.Run("image classification follows upstream metadata, not the model name", func(t *testing.T) {
+		// The mutant: drop the output-modality check, so every model is chat.
+		m := upstreamModel{ID: "gpt-image-2"}
+		m.Architecture.OutputModalities = []string{"image"}
+		if got := discoveredModelKind(m); got != "image" {
+			t.Errorf("mutant survived: got %q, want \"image\"", got)
+		}
+	})
+
+	t.Run("nested metadata is read", func(t *testing.T) {
+		var ms []upstreamModel
+		if err := json.Unmarshal([]byte(realCatalogBody), &ms); err == nil {
+			t.Fatal("fixture must be a wrapper object, not a bare array")
+		}
+		var wrapper struct {
+			Data []upstreamModel `json:"data"`
+		}
+		if err := json.Unmarshal([]byte(realCatalogBody), &wrapper); err != nil {
+			t.Fatal(err)
+		}
+		if len(wrapper.Data) == 0 {
+			t.Fatal("fixture parsed to zero models")
+		}
+		// The mutant: skip copyFields(m.Metadata) entirely.
+		meta := modelMeta(wrapper.Data[0])
+		if meta == nil {
+			t.Fatal("mutant survived: metadata is nil, so cost/context_limit are lost")
+		}
+		for _, k := range []string{"context_limit", "cost", "context_budget"} {
+			if _, ok := meta[k]; !ok {
+				t.Errorf("mutant survived: %s dropped from nested metadata", k)
+			}
+		}
+	})
+
+	t.Run("non-catalog responses are rejected", func(t *testing.T) {
+		for _, body := range []string{`<html>login</html>`, `null`, `{"error":{"message":"x"}}`, `{}`} {
+			if _, err := parseModelCatalog([]byte(body)); err == nil {
+				t.Errorf("mutant survived: %q accepted as a catalog", body)
+			}
+		}
+	})
+
+	t.Run("version segments are distinguished from real path segments", func(t *testing.T) {
+		if !isVersionSegment("v1beta") {
+			t.Error("mutant survived: v1beta not treated as a version")
+		}
+		for _, seg := range []string{"vision", "vertex", "v1x"} {
+			if isVersionSegment(seg) {
+				t.Errorf("mutant survived: %q wrongly treated as a version", seg)
+			}
+		}
+	})
+
+	t.Run("endpoint composition never duplicates the version segment", func(t *testing.T) {
+		// The mutant: never move the segment onto the endpoint.
+		if got := composeEndpoint("/v1/models", "https://h", "https://h/v1"); got != "/v1/chat/completions" {
+			t.Errorf("mutant survived: endpoint = %q", got)
+		}
+		// The other mutant: always move it, even when the base keeps it.
+		if got := composeEndpoint("/models", "https://h/v1", "https://h/v1"); got != "" {
+			t.Errorf("mutant survived: endpoint = %q would double /v1", got)
+		}
+	})
+
+	t.Run("image_models is never nil", func(t *testing.T) {
+		res := &discoveryResult{Models: []string{"a"}, ImageModels: nil}
+		b, _ := json.Marshal(res)
+		if strings.Contains(string(b), `"image_models":null`) {
+			t.Errorf("mutant survived: %s", b)
+		}
+	})
+
+	t.Run("a rejected key is reported as unauthorized, not as an outage", func(t *testing.T) {
+		// The mutant: collapse every failure into one generic error, which is
+		// what made a typo'd key look like the vendor being down.
+		if classifyAuthStatus(http.StatusUnauthorized, []byte(`{"error":"invalid api key"}`)) == "" {
+			t.Error("mutant survived: 401 is not classified")
+		}
+		if classifyAuthStatus(http.StatusNotFound, []byte("nope")) != "" {
+			t.Error("a 404 must not be reported as an auth problem")
+		}
+		if classifyAuthStatus(http.StatusOK, nil) != "" {
+			t.Error("a 200 must not be reported as an auth problem")
+		}
+	})
+
+	t.Run("duplicate candidate URLs are collapsed", func(t *testing.T) {
+		seen := map[string]bool{}
+		for _, c := range discoveryCandidates("https://h/v1") {
+			u := c.baseURL + c.modelsPath
+			if seen[u] {
+				t.Errorf("mutant survived: %s requested twice", u)
+			}
+			seen[u] = true
+		}
+	})
+
+	t.Run("discovery falls back across URL forms", func(t *testing.T) {
+		// The coverage requirement is that BOTH catalog paths are reachable for
+		// a typed base, and both for a bare host. Asserting an exact count kept
+		// drifting as the candidate list was fixed, and each time it measured
+		// the current list rather than the property that matters.
+		for _, base := range []string{"https://h/v1", "https://h"} {
+			distinct := map[string]bool{}
+			for _, c := range discoveryCandidates(base) {
+				distinct[c.baseURL+c.modelsPath] = true
+			}
+			for _, want := range []string{base + "/v1/models", base + "/models"} {
+				if !distinct[want] {
+					t.Errorf("mutant survived: %s never requests %s", base, want)
+				}
+			}
+		}
+	})
+
+	t.Run("only the catalog is fetched, never each model", func(t *testing.T) {
+		// Guarding a NON-mutation: the design decision that discovery must not
+		// spend money. If a future change adds per-model probes, the counting
+		// test below catches it; this assertion documents the intent.
+		if discoveryModelsPath != "/v1/models" {
+			t.Errorf("catalog path changed to %q", discoveryModelsPath)
+		}
+	})
+}

internal/gateway/discover_test.go

@@ -0,0 +1,648 @@
+package gateway
+
+import (
+	"context"
+	"encoding/json"
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"testing"
+)
+
+// realCatalogBody is the actual /models response captured from the amr-link
+// endpoint, not a hand-written fixture. Two properties of it are the reason
+// this test exists at all:
+//
+//   - it nests modalities under "architecture", which the naive parser misses;
+//   - it lists five image-generation ids among 34 chat ids, so a parser that
+//     returns everything as "chat" silently poisons the AUTO chain.
+const realCatalogBody = `{"data":[{"id":"claude-fable-5","object":"model","created":1781081344,"owned_by":"vela","architecture":{"input_modalities":["text","image","file"],"output_modalities":["text"]},"metadata":{"cost":"very_high","capability":"best_quality","context_limit":1000000,"context_budget":232000},"default":false,"enabled":true},{"id":"gpt-image-2","object":"model","architecture":{"input_modalities":["text","image"],"output_modalities":["image"]},"metadata":{"capability":"best_quality"},"enabled":true},{"id":"deepseek-v4-flash","object":"model","architecture":{"input_modalities":["text"],"output_modalities":["text"]},"metadata":{"cost":"medium","capability":"standard","context_limit":1048000},"enabled":true},{"id":"seedream-5.0-pro","object":"model","architecture":{"input_modalities":["text","image"],"output_modalities":["image"]},"metadata":{"capability":"best_quality"},"enabled":true}]}`
+
+func TestParseModelCatalog_RealUpstreamShape(t *testing.T) {
+	models, err := parseModelCatalog([]byte(realCatalogBody))
+	if err != nil {
+		t.Fatalf("parse: %v", err)
+	}
+	if len(models) != 4 {
+		t.Fatalf("want 4 models, got %d", len(models))
+	}
+
+	kind := map[string]string{}
+	for _, m := range models {
+		kind[m.ID] = m.Kind
+	}
+	// The two image generators must be classified by the upstream's declared
+	// output modality, not by their names.
+	for _, id := range []string{"gpt-image-2", "seedream-5.0-pro"} {
+		if kind[id] != "image" {
+			t.Errorf("%s: got kind %q, want \"image\"", id, kind[id])
+		}
+	}
+	for _, id := range []string{"claude-fable-5", "deepseek-v4-flash"} {
+		if kind[id] != "chat" {
+			t.Errorf("%s: got kind %q, want \"chat\"", id, kind[id])
+		}
+	}
+
+	// Metadata the upstream sent must survive; keys it did not send must not
+	// appear as zero values.
+	if got := models[0].Meta["context_limit"]; got == nil {
+		t.Error("claude-fable-5: context_limit dropped (it is nested under metadata in the real response)")
+	} else if f, ok := got.(float64); !ok || f != 1000000 {
+		t.Errorf("context_limit = %#v, want numeric 1000000 so the UI can sort on it", got)
+	}
+	if _, present := models[0].Meta["cost"]; !present {
+		t.Error("claude-fable-5: cost dropped")
+	}
+	// context_budget is present on this entry and must not be dropped either.
+	if _, present := models[0].Meta["context_budget"]; !present {
+		t.Error("claude-fable-5: context_budget dropped")
+	}
+	// deepseek-v4-flash DOES send metadata; gpt-image-2 sends only capability.
+	// Neither may lose it.
+	if models[2].Meta["context_limit"] == nil {
+		t.Error("deepseek-v4-flash: context_limit dropped")
+	}
+	if models[2].Meta["cost"] == nil {
+		t.Error("deepseek-v4-flash: cost dropped")
+	}
+	if _, bogus := models[0].Meta["made_up_field"]; bogus {
+		t.Error("invented a metadata key that the upstream never sent")
+	}
+}
+
+// TestDiscoveredModelKind_ImageOnlyByUpstreamMetadata is the regression guard
+// for the failure mode that motivated keeping kind="chat" as the default: a
+// name-based heuristic marks "ideogram-v2" or "gpt-image-1" as an image model,
+// but a catalog that also serves chat for that id would then have its chat
+// entry removed from the AUTO chain. Only the upstream's declared OUTPUT
+// modality may classify a model as image.
+func TestDiscoveredModelKind_ImageOnlyByUpstreamMetadata(t *testing.T) {
+	cases := []struct {
+		name string
+		in   upstreamModel
+		want string
+	}{
+		{
+			// A catalog that DECLARES output modality image is authoritative.
+			// The "kind" string saying chat must not override it.
+			name: "declared output modality image wins over a chat kind string",
+			in: func() upstreamModel {
+				var m upstreamModel
+				m.Kind = "chat"
+				m.Architecture.OutputModalities = []string{"image"}
+				return m
+			}(),
+			want: "image",
+		},
+		{
+			name: "bare array modality string",
+			in:   upstreamModel{ID: "x", Modality: "text-to-image"},
+			want: "image",
+		},
+		{
+			name: "image INPUT is still a chat model",
+			in: func() upstreamModel {
+				var m upstreamModel
+				m.Architecture.InputModalities = []string{"text", "image"}
+				return m
+			}(),
+			want: "chat",
+		},
+		{
+			name: "file input is still a chat model",
+			in: func() upstreamModel {
+				var m upstreamModel
+				m.Architecture.InputModalities = []string{"text", "file"}
+				return m
+			}(),
+			want: "chat",
+		},
+		{
+			name: "vision capability flag is chat",
+			in:   upstreamModel{ID: "y", Capabilities: modelCaps{Vision: true}},
+			want: "chat",
+		},
+		{
+			name: "capabilities.image_generation is image",
+			in:   upstreamModel{ID: "z", Capabilities: modelCaps{ImageGeneration: true}},
+			want: "image",
+		},
+		{
+			name: "no modality info at all defaults to chat",
+			in:   upstreamModel{ID: "plain"},
+			want: "chat",
+		},
+		{
+			// The default must be chat even when the name LOOKS like an image
+			// model. Guessing from the name is what puts gpt-image-2 into the
+			// chat AUTO chain, where every request to it fails.
+			name: "image-looking name without upstream metadata stays chat",
+			in:   upstreamModel{ID: "gpt-image-2"},
+			want: "chat",
+		},
+		{
+			name: "vision-looking name without upstream metadata stays chat",
+			in:   upstreamModel{ID: "ideogram-v2"},
+			want: "chat",
+		},
+	}
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			if got := discoveredModelKind(tc.in); got != tc.want {
+				t.Errorf("got %q, want %q", got, tc.want)
+			}
+		})
+	}
+}
+
+func TestParseModelCatalog_RejectsNonCatalog(t *testing.T) {
+	// A 200 with valid JSON that is not a catalog must NOT be reported as an
+	// empty result. Reporting zero models would let a login page or a
+	// rate-limit body create a source with no models in it.
+	for name, body := range map[string]string{
+		"login page html": `<html><body>Sign in</body></html>`,
+		"rate limit json": `{"error":{"message":"slow down","type":"rate_limit"}}`,
+		"empty object":    `{}`,
+		"null":            `null`,
+		"wrong shape":     `{"objects":[{"id":"m"}]}`,
+	} {
+		t.Run(name, func(t *testing.T) {
+			if _, err := parseModelCatalog([]byte(body)); err == nil {
+				t.Errorf("expected an error for %q, got nil (would create an empty source)", name)
+			}
+		})
+	}
+}
+
+func TestParseModelCatalog_AcceptsBareArray(t *testing.T) {
+	models, err := parseModelCatalog([]byte(`[{"id":"a"},{"id":"b"}]`))
+	if err != nil {
+		t.Fatalf("bare array should parse: %v", err)
+	}
+	if len(models) != 2 {
+		t.Fatalf("want 2, got %d", len(models))
+	}
+}
+
+// TestDiscoverModels_NormalizesWhenOnlyBareModelsAnswers: operator pastes
+// "https://host/v1" (the form every SDK example shows) but the vendor serves
+// the catalog at "https://host/models" — the version segment on the path is
+// implied, not doubled. Discovery must strip the segment off the base and
+// report it on the endpoint, because openai's adapter endpoint is
+// "/chat/completions" and the two are concatenated.
+func TestDiscoverModels_NormalizesWhenOnlyBareModelsAnswers(t *testing.T) {
+	var paths []string
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		paths = append(paths, r.URL.Path)
+		if r.URL.Path != "/models" {
+			w.WriteHeader(http.StatusNotFound)
+			return
+		}
+		_, _ = w.Write([]byte(realCatalogBody))
+	}))
+	defer srv.Close()
+
+	res, err := DiscoverModels(context.Background(), srv.URL+"/v1", "sk-test")
+	if err != nil {
+		t.Fatalf("discovery failed: %v", err)
+	}
+	if res.BaseURL != srv.URL {
+		t.Errorf("base_url = %q, want %q (version segment moved off the base)", res.BaseURL, srv.URL)
+	}
+	// "/models" carries no version segment, but the operator typed one — the
+	// endpoint has to restore it or the chat call would miss /v1.
+	if res.Endpoint != "/v1/chat/completions" {
+		t.Errorf("endpoint = %q, want %q", res.Endpoint, "/v1/chat/completions")
+	}
+	if res.Note == "" {
+		t.Error("silently rewriting the operator's base_url is worse than saying so")
+	}
+	if len(res.Models) != 2 || len(res.ImageModels) != 2 {
+		t.Errorf("got %d chat / %d image, want 2 / 2", len(res.Models), len(res.ImageModels))
+	}
+	if len(paths) < 2 {
+		t.Errorf("tried only %v; expected a fallback search", paths)
+	}
+}
+
+// TestDiscoverModels_BareHostWithVersionedModelsPath: operator pastes the host
+// root, vendor serves /v1/models. The base needs no rewriting; the endpoint
+// does.
+func TestDiscoverModels_BareHostWithVersionedModelsPath(t *testing.T) {
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		if r.URL.Path != "/v1/models" {
+			w.WriteHeader(http.StatusNotFound)
+			return
+		}
+		_, _ = w.Write([]byte(`{"data":[{"id":"m1"}]}`))
+	}))
+	defer srv.Close()
+
+	res, err := DiscoverModels(context.Background(), srv.URL, "")
+	if err != nil {
+		t.Fatalf("discovery failed: %v", err)
+	}
+	if res.BaseURL != srv.URL {
+		t.Errorf("base_url = %q, want %q", res.BaseURL, srv.URL)
+	}
+	if res.Endpoint != "/v1/chat/completions" {
+		t.Errorf("endpoint = %q, want %q", res.Endpoint, "/v1/chat/completions")
+	}
+}
+
+// TestDiscoverModels_KeepsWorkingBaseURLUnchanged covers the vendor that really
+// does serve "/v1/models": the typed URL must survive untouched so no needless
+// config rewrite is suggested.
+func TestDiscoverModels_KeepsWorkingBaseURLUnchanged(t *testing.T) {
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		if r.URL.Path != "/v1/models" {
+			w.WriteHeader(http.StatusNotFound)
+			return
+		}
+		_, _ = w.Write([]byte(`{"data":[{"id":"m1"},{"id":"m2"}]}`))
+	}))
+	defer srv.Close()
+
+	// Base typed WITH "/v1", catalog served at "…/v1/models". That is exactly
+	// the amr-link shape: first candidate succeeds, nothing is rewritten, and
+	// the adapter's default endpoint already composes.
+	res, err := DiscoverModels(context.Background(), srv.URL+"/v1", "")
+	if err != nil {
+		t.Fatalf("discovery failed: %v", err)
+	}
+	if res.BaseURL != srv.URL+"/v1" {
+		t.Errorf("base_url = %q, want it unchanged", res.BaseURL)
+	}
+	if res.Endpoint != "" {
+		t.Errorf("endpoint = %q, want empty (the typed base already carries /v1)", res.Endpoint)
+	}
+	if res.Note != "" {
+		t.Errorf("note = %q, want empty when nothing was rewritten", res.Note)
+	}
+	if len(res.Models) != 2 {
+		t.Errorf("got %d chat models, want 2", len(res.Models))
+	}
+}
+
+// TestDiscoverModels_ErrorNamesEveryTriedURL: a failed probe is useless if it
+// only reports the last URL. The operator needs to see the full search.
+func TestDiscoverModels_ErrorNamesEveryTriedURL(t *testing.T) {
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.WriteHeader(http.StatusNotFound)
+	}))
+	defer srv.Close()
+
+	_, err := DiscoverModels(context.Background(), srv.URL+"/v1", "k")
+	if err == nil {
+		t.Fatal("expected failure against an all-404 server")
+	}
+	msg := err.Error()
+	for _, want := range []string{"/models", "/v1/models", "404"} {
+		if !strings.Contains(msg, want) {
+			t.Errorf("error %q should mention %q", msg, want)
+		}
+	}
+}
+
+func TestDiscoveryCandidates_DedupesAndPrefersTyped(t *testing.T) {
+	cands := discoveryCandidates("https://h/v1/")
+	if len(cands) == 0 {
+		t.Fatal("no candidates")
+	}
+	seen := map[string]bool{}
+	for _, c := range cands {
+		// Dedup is on the requested URL, so that is the key to assert on.
+		k := c.baseURL + c.modelsPath
+		if seen[k] {
+			t.Errorf("duplicate request for %s", k)
+		}
+		seen[k] = true
+	}
+	// The versioned spelling of the typed URL must be reachable: on the real
+	// amr-link endpoint that is what answers, and it is the preferred candidate.
+	if !seen["https://h/v1/models"] {
+		t.Error("the typed base with a versioned catalog path must be tried; that is what answers on amr-link")
+	}
+	// Both catalog forms must be attempted for a bare host too. Offering only
+	// "/v1/models" there left a vendor serving a bare "/models" looking
+	// unreachable.
+	if !seen["https://h/models"] {
+		t.Error("host root with a bare /models path was never tried")
+	}
+	// The first candidate is preferred, so its URL is what gets reported when
+	// several answer.
+	if cands[0].baseURL+cands[0].modelsPath != "https://h/v1/models" {
+		t.Errorf("first candidate requests %q, want the canonical versioned form",
+			cands[0].baseURL+cands[0].modelsPath)
+	}
+	// The bare "/models" form must exist too — some gateways serve it there.
+	hasBare := false
+	for _, c := range cands {
+		if c.modelsPath == "/models" {
+			hasBare = true
+		}
+	}
+	if !hasBare {
+		t.Error("no /models candidate; vendors serving the catalog at /models would be missed")
+	}
+	// Preference must be monotonically non-decreasing so the sort is stable.
+	for i := 1; i < len(cands); i++ {
+		if cands[i].prefer < cands[i-1].prefer {
+			t.Errorf("candidates out of preference order at %d: %d < %d", i, cands[i].prefer, cands[i-1].prefer)
+		}
+	}
+}
+
+func TestVersionSegment(t *testing.T) {
+	cases := map[string]string{
+		"/v1/models":   "/v1",
+		"/v1beta/models": "/v1beta",
+		"/models":      "",
+		"/api/models":  "",
+		"/vision/models": "",
+	}
+	for in, want := range cases {
+		if got := versionSegment(in); got != want {
+			t.Errorf("versionSegment(%q) = %q, want %q", in, got, want)
+		}
+	}
+}
+
+func TestTrimVersionSuffix(t *testing.T) {
+	cases := map[string]string{
+		"https://h/v1":          "https://h",
+		"https://h/v2":          "https://h",
+		"https://h/v1beta":      "https://h",
+		"https://h/api/v1":      "https://h/api",
+		"https://h/vision":      "https://h/vision", // not numeric — a real path segment
+		"https://h":             "https://h",
+		"https://h/openai":      "https://h/openai",
+		"https://h/v1/extra":    "https://h/v1/extra",
+		// A segment that merely STARTS with v must not be stripped: stripping
+		// it would rewrite base "https://h/vision" to "https://h" and, worse,
+		// would treat a real vendor path as a version.
+		"https://h/vertex":      "https://h/vertex",
+		"https://h/volcengine":  "https://h/volcengine",
+		"https://h/vertex-ai/v1": "https://h/vertex-ai",
+	}
+	for in, want := range cases {
+		if got := trimVersionSuffix(in); got != want {
+			t.Errorf("trimVersionSuffix(%q) = %q, want %q", in, got, want)
+		}
+	}
+}
+
+// TestIsVersionSegment_RequiresDigits pins the distinction the whole endpoint
+// derivation rests on: "v1"/"v1beta" are versions, "vision"/"vertex" are path
+// segments. Without the digit requirement, stripping turns a vendor's real
+// base URL into a different one.
+func TestIsVersionSegment_RequiresDigits(t *testing.T) {
+	yes := []string{"v1", "v2", "v10", "v1beta", "v1alpha", "v1preview"}
+	no := []string{"vision", "vertex", "volcengine", "openai", "v", "vv", "api", "beta", "v1x"}
+	for _, s := range yes {
+		if !isVersionSegment(s) {
+			t.Errorf("isVersionSegment(%q) = false, want true", s)
+		}
+	}
+	for _, s := range no {
+		if isVersionSegment(s) {
+			t.Errorf("isVersionSegment(%q) = true, want false", s)
+		}
+	}
+}
+
+// TestComposeEndpoint_ExactlyOneVersionSegment is the invariant test. Every row
+// below is a URL form that was observed to 404 or to route wrongly when the
+// version segment ended up in the wrong place.
+func TestComposeEndpoint_ExactlyOneVersionSegment(t *testing.T) {
+	cases := []struct {
+		name          string
+		modelsPath    string
+		persistedBase string
+		typedBase     string
+		want          string
+	}{
+		{
+			name:       "typed base has /v1 and catalog served there: base keeps it, endpoint empty",
+			modelsPath: "/models", persistedBase: "https://h/v1", typedBase: "https://h/v1",
+			want: "",
+		},
+		{
+			name:       "catalog served at versioned path, base stripped: endpoint carries it",
+			modelsPath: "/v1/models", persistedBase: "https://h", typedBase: "https://h/v1",
+			want: "/v1/chat/completions",
+		},
+		{
+			name:       "bare host, bare catalog, nothing typed: adapter default composes",
+			modelsPath: "/models", persistedBase: "https://h", typedBase: "https://h",
+			want: "",
+		},
+		{
+			name:       "bare host, versioned catalog: endpoint owns the version",
+			modelsPath: "/v1/models", persistedBase: "https://h", typedBase: "https://h",
+			want: "/v1/chat/completions",
+		},
+		{
+			name:       "typed /v1 but catalog only at bare /models: strip base, restore on endpoint",
+			modelsPath: "/models", persistedBase: "https://h", typedBase: "https://h/v1",
+			want: "/v1/chat/completions",
+		},
+		{
+			name:       "non-version segment must never be promoted onto the endpoint",
+			modelsPath: "/models", persistedBase: "https://h/vision", typedBase: "https://h/vision",
+			want: "",
+		},
+	}
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			got := composeEndpoint(tc.modelsPath, tc.persistedBase, tc.typedBase)
+			if got != tc.want {
+				t.Errorf("composeEndpoint = %q, want %q", got, tc.want)
+			}
+			// Whichever way it resolved, base + endpoint must never carry the
+			// same version segment twice.
+			onBase := trimVersionSuffix(tc.persistedBase) != tc.persistedBase
+			onEndpoint := versionSegment(strings.TrimSuffix(got, "/chat/completions") + "/models") != ""
+			if onBase && onEndpoint {
+				t.Errorf("version segment duplicated: base %q + endpoint %q", tc.persistedBase, got)
+			}
+		})
+	}
+}
+
+// TestDiscoverModels_DoesNotProbeEveryModel guards the decision not to spend
+// money verifying each catalog entry. Only the catalog request is made.
+func TestDiscoverModels_DoesNotProbeEveryModel(t *testing.T) {
+	var calls int
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		calls++
+		_, _ = w.Write([]byte(`{"data":[{"id":"a"},{"id":"b"},{"id":"c"}]}`))
+	}))
+	defer srv.Close()
+
+	if _, err := DiscoverModels(context.Background(), srv.URL+"/v1", "k"); err != nil {
+		t.Fatalf("discovery failed: %v", err)
+	}
+	if calls != 1 {
+		t.Errorf("made %d requests, want exactly 1 (the catalog)", calls)
+	}
+}
+
+func TestModelsToConfig_PrioritiesAndKinds(t *testing.T) {
+	got := modelsToConfig([]string{"a", "b"}, []string{"img1"})
+	if len(got) != 3 {
+		t.Fatalf("want 3 models, got %d", len(got))
+	}
+	if got[2].ID != "img1" || got[2].Kind != "image" {
+		t.Errorf("image model mis-sorted: %+v", got[2])
+	}
+	if got[2].Priority <= got[0].Priority {
+		t.Errorf("image model priority %d should sort after chat %d", got[2].Priority, got[0].Priority)
+	}
+}
+
+// TestDiscoverModels_ImageModelsNeverNull guards the UI contract end to end,
+// not just on the struct: the WebUI does out.concat(image_models), and a JSON
+// null there renders "undefined" next to every image model row. This asserts on
+// the served bytes rather than the Go field, because that is what the bug was.
+func TestDiscoverModels_ImageModelsNeverNull(t *testing.T) {
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		// Only chat models — the case where the slice is never appended to.
+		_, _ = w.Write([]byte(`{"data":[{"id":"a"},{"id":"b"}]}`))
+	}))
+	defer srv.Close()
+
+	res, err := DiscoverModels(context.Background(), srv.URL+"/v1", "k")
+	if err != nil {
+		t.Fatalf("discovery failed: %v", err)
+	}
+	if len(res.ImageModels) != 0 {
+		t.Fatalf("precondition: want zero image models, got %d", len(res.ImageModels))
+	}
+	b, err := json.Marshal(res)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !strings.Contains(string(b), `"image_models":[]`) {
+		t.Errorf("served %s\nwant \"image_models\":[] — null breaks concat in the UI", b)
+	}
+}
+
+// TestComposeImageEndpoint pins the rule that fixed a real 404.
+//
+// Provider.ImageURL is `base_url + ImageEndpoint()`, and the built-in default
+// is the ABSOLUTE "/v1/images/generations" — while chat's default is the
+// relative "/chat/completions". The two defaults disagree about who owns the
+// version segment, so a base_url that produces a working chat URL produces
+// "/v1/v1/images/generations" and a 404.
+func TestComposeImageEndpoint(t *testing.T) {
+	cases := []struct {
+		base string
+		want string
+	}{
+		// base without a version: the absolute default composes correctly.
+		{"https://h", ""},
+		// base WITH a version: the default would double it, so the
+		// version-relative form is written instead.
+		{"https://h/v1", "/images/generations"},
+		{"https://h/v1beta", "/images/generations"},
+		// A real path segment that merely starts with v is not a version.
+		{"https://h/vision", ""},
+	}
+	for _, tc := range cases {
+		got := composeImageEndpoint(tc.base)
+		if got != tc.want {
+			t.Errorf("composeImageEndpoint(%q) = %q, want %q", tc.base, got, tc.want)
+		}
+		// Whatever it decided, base + result must not contain a doubled
+		// version segment — that is the bug this function exists for.
+		joined := tc.base + got
+		if strings.Contains(joined, "/v1/v1") {
+			t.Errorf("composed %q contains a doubled version segment", joined)
+		}
+	}
+}
+
+// TestDiscoveryReportsImageEndpointWhenBaseCarriesVersion is the end-to-end
+// version: a pasted base_url with "/v1" must come back with an image endpoint,
+// or the image models it reported will all 404 on use.
+func TestDiscoveryReportsImageEndpointWhenBaseCarriesVersion(t *testing.T) {
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		if r.URL.Path != "/v1/models" {
+			w.WriteHeader(http.StatusNotFound)
+			return
+		}
+		_, _ = w.Write([]byte(realCatalogBody))
+	}))
+	defer srv.Close()
+
+	res, err := DiscoverModels(context.Background(), srv.URL+"/v1", "k")
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(res.ImageModels) == 0 {
+		t.Fatal("fixture must contain image models for this to mean anything")
+	}
+	if res.BaseURL != srv.URL+"/v1" {
+		t.Fatalf("base = %q, want the typed URL with /v1", res.BaseURL)
+	}
+	if res.ImageEndpoint != "/images/generations" {
+		t.Errorf("image_endpoint = %q, want %q (otherwise the reported image models all 404)",
+			res.ImageEndpoint, "/images/generations")
+	}
+	// The chat side stays empty: base already carries /v1 and the adapter's
+	// relative default composes.
+	if res.Endpoint != "" {
+		t.Errorf("endpoint = %q, want empty", res.Endpoint)
+	}
+}
+
+// TestDiscoveryResult_NilSlicesSerializeAsArrays covers the zero value of
+// discoveryResult, which is what a caller gets if it builds one directly
+// instead of going through DiscoverModels. DiscoverModels happens to
+// initialize the slices, so testing only that path would let the MarshalJSON
+// nil guard look redundant — and removing it would ship a latent UI bug the
+// first time some other caller builds the struct.
+func TestDiscoveryResult_NilSlicesSerializeAsArrays(t *testing.T) {
+	b, err := json.Marshal(&discoveryResult{BaseURL: "https://h"})
+	if err != nil {
+		t.Fatal(err)
+	}
+	got := string(b)
+	for _, want := range []string{`"models":[]`, `"image_models":[]`, `"tried":[]`} {
+		if !strings.Contains(got, want) {
+			t.Errorf("zero value serialized %s\nwant %s — a nil slice becomes JSON null and the UI renders \"undefined\"", got, want)
+		}
+	}
+	for _, bad := range []string{`"models":null`, `"image_models":null`, `"tried":null`} {
+		if strings.Contains(got, bad) {
+			t.Errorf("serialized %s, which the UI cannot concatenate", bad)
+		}
+	}
+}
+
+// TestDiscoveryResultJSONShape pins the field names the WebUI will bind to.
+func TestDiscoveryResultJSONShape(t *testing.T) {
+	b, err := json.Marshal(&discoveryResult{
+		BaseURL: "https://h", Models: []string{"a"}, ImageModels: []string{},
+		ModelsPath: "/models", Adapter: "openai",
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	var got map[string]interface{}
+	if err := json.Unmarshal(b, &got); err != nil {
+		t.Fatal(err)
+	}
+	// image_models must serialize as [] and never null: the UI does
+	// out.concat(image_models) and null would render "undefined".
+	if v, ok := got["image_models"]; !ok || v == nil {
+		t.Errorf("image_models = %v, want an empty array (null breaks concat in the UI)", v)
+	}
+	for _, k := range []string{"base_url", "endpoint", "models", "image_models", "models_path", "tried"} {
+		if _, ok := got[k]; !ok {
+			t.Errorf("missing field %q in discovery response", k)
+		}
+	}
+}

internal/gateway/plugins_api.go

@@ -117,6 +117,20 @@ func (g *Gateway) handlePluginsAPI(w http.ResponseWriter, r *http.Request) {
 		return
 	}
 
+	// /api/plugins/{name}/fanout/{fn} — one call per argument set, in
+	// parallel, from one request. See handlePluginFanout for why the
+	// request-per-item form cannot be made concurrent from the client.
+	if i := strings.Index(path, "/fanout/"); i >= 0 {
+		name := path[:i]
+		fn := path[i+len("/fanout/"):]
+		if name == "" || fn == "" {
+			writeError(w, http.StatusBadRequest, "invalid_request", "fanout needs a plugin and a function name")
+			return
+		}
+		g.handlePluginFanout(w, r, name, fn)
+		return
+	}
+
 	if reqRole(r.Context()) != "admin" {
 		writeError(w, http.StatusForbidden, "forbidden", "admin role required")
 		return
@@ -133,7 +147,7 @@ func (g *Gateway) handlePluginsAPI(w http.ResponseWriter, r *http.Request) {
 				// Visibility overrides and the endpoint access table, so the
 				// plugin page renders what is IN FORCE rather than making the
 				// operator read .lua files or config.yaml to find out.
-				"visibility":   ps.VisibilityOverrides(),
+				"visibility":  ps.VisibilityOverrides(),
 				"http_access": ps.EndpointAccess(),
 			})
 		case http.MethodPost:
@@ -331,6 +345,68 @@ func (g *Gateway) handlePluginCall(w http.ResponseWriter, r *http.Request, name,
 	writeJSON(w, http.StatusOK, map[string]interface{}{"result": result})
 }
 
+// handlePluginFanout runs one plugin export per element of a list, in parallel,
+// from a SINGLE request.
+//
+// Why this exists: the quota page has twenty cards, and it used to fire twenty
+// browser requests at /call/. They arrived concurrently and then serialised
+// inside Plugins.Call, which holds the plugin mutex across the Lua function —
+// including its network wait. Measured on the live gateway: 20 probes took 29s of
+// wall clock while a single probe with no contention took 1.4–3.3s, so the page
+// sat on "probing…" for half a minute. The browser-side concurrency was real and
+// made no difference.
+//
+// A fan-out cannot be assembled from the client either: any request-per-item
+// arrangement is exactly the shape that queues. The parallelism has to happen
+// where the fan-out is received.
+func (g *Gateway) handlePluginFanout(w http.ResponseWriter, r *http.Request, name, fn string) {
+	if reqRole(r.Context()) != "admin" {
+		writeError(w, http.StatusForbidden, "forbidden", "admin role required")
+		return
+	}
+	if r.Method != http.MethodPost {
+		writeError(w, http.StatusMethodNotAllowed, "method_not_allowed", "")
+		return
+	}
+	var req struct {
+		// Args is a LIST of argument objects, one per call. A bare object is
+		// rejected rather than coerced: "args" meaning "one call" and "args"
+		// meaning "many" depending on its shape is how a caller ends up probing
+		// one source and believing it probed twenty.
+		Args []map[string]interface{} `json:"args"`
+	}
+	if err := json.NewDecoder(r.Body).Decode(&req); err != nil {
+		writeError(w, http.StatusBadRequest, "invalid_request", "invalid json: "+err.Error())
+		return
+	}
+	if len(req.Args) == 0 {
+		writeError(w, http.StatusBadRequest, "invalid_request", "args must be a non-empty list of argument objects")
+		return
+	}
+	// The page legitimately has one card per configured source. The bound is not
+	// arbitrary: it stops one request from asking for thousands of VMs, and it is
+	// enforced before any work starts rather than discovered shard by shard.
+	if len(req.Args) > fanoutMaxArgs {
+		writeError(w, http.StatusBadRequest, "invalid_request",
+			fmt.Sprintf("args has %d entries, limit is %d", len(req.Args), fanoutMaxArgs))
+		return
+	}
+	results, err := g.core.Plugins().ShardCall(name, fn, req.Args)
+	if err != nil {
+		writeError(w, http.StatusBadGateway, "plugin_error", err.Error())
+		return
+	}
+	// A shard that failed contributes a null in its slot. That is deliberate and
+	// is NOT the same as the whole request failing: one unreachable upstream must
+	// not turn a page of twenty cards into a single error.
+	writeJSON(w, http.StatusOK, map[string]interface{}{"results": results})
+}
+
+// fanoutMaxArgs bounds one fan-out request. Twenty is the measured real source
+// count on the gateway this was built for; the headroom is for operators with
+// more sources configured.
+const fanoutMaxArgs = 64
+
 func (g *Gateway) pluginDir() string {
 	if c := g.core.Config(); c != nil {
 		return c.PluginDir
@@ -369,7 +445,6 @@ func (g *Gateway) removePlugin(name string) error {
 	return g.core.Plugins().Unload(name)
 }
 
-
 // handleEndpointAccess serves the operator's view and edit surface for one
 // plugin endpoint's reachable keys.
 //
@@ -439,7 +514,6 @@ func (g *Gateway) handleEndpointAccess(w http.ResponseWriter, r *http.Request, n
 	}
 }
 
-
 // splitStateForKey rewrites a plugin state so that a non-admin caller sees only
 // the buckets that belong to them.
 //

internal/gateway/server.go

@@ -224,6 +224,11 @@ func (g *Gateway) routes(w http.ResponseWriter, r *http.Request) {
 		g.handleAdaptersAPI(w, r)
 	case r.URL.Path == "/api/v1" || strings.HasPrefix(r.URL.Path, "/api/v1/"):
 		g.apiV1Routes(w, r)
+	// Must precede the /api/sources prefix match: a generic HasPrefix on
+	// "/api/sources/" would otherwise swallow "/api/sources/discover" and the
+	// operator would get "source name required in path" instead of a catalog.
+	case r.URL.Path == "/api/sources/discover" || strings.HasPrefix(r.URL.Path, "/api/sources/discover/"):
+		g.handleSourceDiscoverAPI(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/ui/index.html

@@ -366,6 +366,22 @@
 
       /* ---- chat page ---- */
       .tab-chat{max-width:840px;margin:0 auto;display:flex;flex-direction:column;height:calc(100vh - 190px)}
+
+      .tab-image{max-width:840px;margin:0 auto;display:flex;flex-direction:column;
+        height:calc(100vh - 190px)}
+      .img-wrap{flex:1;display:flex;flex-direction:column;min-height:0}
+      .img-grid{flex:1;overflow-y:auto;padding:12px 2px;display:grid;
+        grid-template-columns:repeat(auto-fill,minmax(240px,1fr));gap:10px;align-content:start}
+      .img-grid .chat-empty{grid-column:1/-1}
+      .img-cell{border:1px solid var(--line);border-radius:10px;overflow:hidden;
+        background:var(--bg2)}
+      .img-cell img{display:block;width:100%;height:auto;background:#fff}
+      .img-meta{padding:6px 8px;font-size:11px;color:var(--muted);
+        overflow:hidden;text-overflow:ellipsis;white-space:nowrap}
+      .img-composer{border-top:1px solid var(--line);padding:10px 2px 12px}
+      .img-composer .chat-box{display:flex;gap:8px;align-items:flex-end}
+      .img-composer textarea{flex:1;resize:vertical;min-height:52px}
+      .img-composer .empty{padding:8px 0}
       .chat-wrap{background:var(--card);border:1px solid var(--line);border-radius:18px;box-shadow:var(--sh-lg);
         display:flex;flex-direction:column;flex:1;min-height:0;overflow:hidden;backdrop-filter:blur(var(--glass)) saturate(1.5)}
       .chat-log{flex:1;overflow-y:auto;padding:20px 20px 8px;display:flex;flex-direction:column;gap:18px}
@@ -573,6 +589,21 @@
               /></svg
             ><span data-i="navChat">对话</span>
           </button>
+          <button class="sb-i" data-tab="image">
+            <svg
+              viewBox="0 0 24 24"
+              fill="none"
+              stroke="currentColor"
+              stroke-width="2"
+              stroke-linecap="round"
+              stroke-linejoin="round"
+            >
+              <path d="M3 5h18v14H3z" />
+              <path d="m3 15 5-5 4 4 3-3 6 6" />
+              <circle cx="8.5" cy="8.5" r="1.5" />
+            ></svg
+            ><span data-i="navImage">生图</span>
+          </button>
           <button class="sb-i" data-tab="keys">
             <svg
               viewBox="0 0 24 24"
@@ -791,6 +822,7 @@
         <main id="main">
           <div id="tab-status" class="tab-pane"></div>
           <div id="tab-chat" class="tab-pane hidden"></div>
+          <div id="tab-image" class="tab-pane hidden"></div>
           <div id="tab-keys" class="tab-pane hidden"></div>
           <div id="tab-sort" class="tab-pane hidden"></div>
           <div id="tab-sources" class="tab-pane hidden"></div>
@@ -941,6 +973,20 @@
           adPoolPeak: "峰值",
           adPoolSteps: "步长 (扩/缩)",
           cModel: "模型",
+          cImageModels: "🖼 图像模型（不可选，请到「生图」页测试）",
+          urlHint: "填主机地址，通常不带 /v1 —— 网关会自动拼接路径。但各家约定不同：若填 https://x.com/v1，对话路径会正确，而图像生成会变成 /v1/v1/… 而 404。不确定就用上方「自动探测模型」，它会探测并填好正确形式。",
+          navImage: "生图",
+          imgWelcome: "输入提示词开始生成",
+          imgSlow: "生图可能需要 15 秒到 2 分钟，取决于上游模型",
+          imgPrompt: "描述你想要的画面…",
+          imgGenerate: "生成",
+          imgWorking: "生成中… %ds",
+          imgCount: "数量",
+          imgSize: "尺寸",
+          imgNoModels: "没有配置图像模型。在「源」里为某个源添加 kind: image 的模型后即可在此生成。",
+          imgNoData: "上游没有返回图片数据",
+          imgNeedPrompt: "请输入提示词",
+          imgFailed: "生成失败",
           cMsg: "Enter 发送，Shift+Enter 换行",
           send: "发送",
           cClear: "清空",
@@ -989,6 +1035,16 @@
           mModels: "模型列表",
           mAddModel: "+ 模型",
           mMeta: "Meta",
+          discScan: "自动探测模型",
+          discScanning: "探测中…",
+          discHint: "填好 URL 和密钥后点「自动探测模型」，网关会拉取上游的模型列表并填入下方；仍需点保存才会生效。",
+          discNeedUrl: "请先填写 URL",
+          discEmpty: "上游返回的模型列表为空",
+          discFound: "已找到 %d 个模型",
+          discChatN: "对话 %d 个",
+          discImageN: "图像 %d 个",
+          discUnsaved: "尚未保存，点保存后生效",
+          discFailed: "探测失败",
           mSave: "保存",
           mCancel: "取消",
           toastCopied: "已复制",
@@ -1232,6 +1288,20 @@
           adPoolPeak: "Peak",
           adPoolSteps: "Steps (grow/shrink)",
           cModel: "Model",
+          cImageModels: "🖼 Image models (not selectable — use the Image tab)",
+          urlHint: "Host address, usually WITHOUT /v1 — the gateway appends the path. Vendors differ: with https://x.com/v1 chat works but image generation becomes /v1/v1/… and 404s. If unsure use Auto-discover above; it probes and fills in the correct form.",
+          navImage: "Image",
+          imgWelcome: "Describe an image to generate it",
+          imgSlow: "Generation can take 15s to 2min depending on the upstream model",
+          imgPrompt: "Describe the image you want…",
+          imgGenerate: "Generate",
+          imgWorking: "Generating… %ds",
+          imgCount: "Count",
+          imgSize: "Size",
+          imgNoModels: "No image models configured. Add a model with kind: image on the Sources page.",
+          imgNoData: "The upstream returned no image data",
+          imgNeedPrompt: "Enter a prompt first",
+          imgFailed: "Generation failed",
           cMsg: "Enter to send, Shift+Enter for newline",
           send: "Send",
           cClear: "Clear",
@@ -1281,6 +1351,16 @@
           mModels: "Models",
           mAddModel: "+ model",
           mMeta: "Meta",
+          discScan: "Auto-discover",
+          discScanning: "Scanning…",
+          discHint: "Fill in the URL and key, then press Auto-discover: the gateway fetches the upstream catalog and fills the list below. You still have to save.",
+          discNeedUrl: "Enter the base URL first",
+          discEmpty: "The upstream returned an empty model list",
+          discFound: "Found %d models",
+          discChatN: "%d chat",
+          discImageN: "%d image",
+          discUnsaved: "not saved yet — press Save",
+          discFailed: "Discovery failed",
           mSave: "Save",
           mCancel: "Cancel",
           toastCopied: "Copied",
@@ -1471,7 +1551,7 @@
       // keep in sync, and adding a tab meant finding all three. A plugin page
       // that is routed but never shown is exactly the kind of silent gap that
       // survives review.
-      const TABS = ["status", "chat", "keys", "sort", "sources", "adapters", "plugins"];
+      const TABS = ["status", "chat", "image", "keys", "sort", "sources", "adapters", "plugins"];
       // Declared HERE, with var, because goTab() below reads it. It used to be a
       // `const` further down the file next to the other plugin state, which put
       // goTab's read inside the temporal dead zone: the first click on a plugin
@@ -2719,28 +2799,65 @@ recsState.keyNames = keyNames || {};
       /* ---------- chat tab ---------- */
       let chatHistory = [];
       let chatImages = [];
-      async function renderChat() {
-        const s = await api("/api/status");
-        // Prefer per-source options ("source:model" routing) when the caller sees
-        // sources (admin); otherwise fall back to the flat model list (user scope).
-        let optList;
-        if (Array.isArray(s.sources) && s.sources.length) {
-          optList = ["AUTO"];
-          s.sources.forEach((src) =>
-            (src.models || []).forEach((m) => {
-              optList.push(
-                `<option value="${escAttr(`${src.name}:${m}`)}">${esc(`${src.name} · ${m}`)}</option>`,
-              );
-            }),
-          );
-        } else {
-          optList = ["AUTO"].concat(
-            (s.models || []).map(
-              (m) => `<option value="${escAttr(m)}">${esc(m)}</option>`,
-            ),
-          );
-        }
-        const finalOpts = optList.join("");
+async function renderChat() {
+  const s = await api("/api/status");
+  // Prefer per-source options ("source:model" routing) when the caller sees
+  // sources (admin); otherwise fall back to the flat model list (user scope).
+  //
+  // Chat and image models are listed in SEPARATE groups. The backend used to
+  // hand this page one flat list with the kind stripped, so gpt-image-2 sat
+  // next to claude-sonnet-5.5 and picking it sent the request to /api/chat,
+  // which cannot serve it — a failure whose message never mentioned images.
+  //
+  // model_kinds parallels models; an older backend simply omits it, and then
+  // everything falls into the chat group (the pre-fix behaviour).
+  const optList = [];
+  const imgOpts = [];
+  const addModel = (value, label, kind) => {
+    const html = `<option value="${escAttr(value)}">${esc(label)}</option>`;
+    if (kind === "image") imgOpts.push(html);
+    else optList.push(html);
+  };
+
+  if (Array.isArray(s.sources) && s.sources.length) {
+    optList.push(`<option value="AUTO">AUTO</option>`);
+    s.sources.forEach((src) => {
+      // Prefer the kind-carrying list; fall back to the flat one.
+      const kinds = Array.isArray(src.model_kinds) && src.model_kinds.length
+        ? src.model_kinds
+        : (src.models || []).map((m) => ({ id: m, kind: "chat" }));
+      kinds.forEach((m) => {
+        const id = typeof m === "string" ? m : m.id;
+        addModel(
+          `${src.name}:${id}`,
+          `${src.name} · ${id}`,
+          typeof m === "string" ? "chat" : m.kind,
+        );
+      });
+    });
+  } else {
+    optList.push(`<option value="AUTO">AUTO</option>`);
+    (s.models || []).forEach((m) => {
+      addModel(m, m, "chat");
+    });
+  }
+  // Image models are listed but DISABLED, not merely grouped.
+  //
+  // Grouping alone still let an operator pick one, and picking it sent the
+  // request to /api/chat — which cannot serve an image model — producing a
+  // failure whose message never mentioned images. A disabled option explains
+  // itself ("use the Image tab"); an enabled one just fails.
+  //
+  // The disabled attribute goes on at CONSTRUCTION, not via a later
+  // replaceAll over the joined HTML: a plain String.replace only rewrites the
+  // first match, so every source after the first contributed selectable
+  // image models. Measured with six image models across two sources — the first
+  // was disabled and the other five were not.
+  const imgHtml = () => imgOpts.map((h) => h.replace("<option ", "<option disabled ")).join("");
+  const finalOpts = optList.join("") +
+    (imgOpts.length
+      ? `<optgroup label="${escAttr(t("cImageModels"))}">${imgHtml()}</optgroup>`
+      : "");
         $("#tab-chat").innerHTML = `
     <div class="tab-chat">
       <div class="chat-wrap">
@@ -2784,6 +2901,133 @@ recsState.keyNames = keyNames || {};
         $("#ct-imgs").style.display = "none";
         $("#ct-msg").focus();
       }
+      /* ---------- image tab ---------- */
+      // Image generation lives in its own page rather than as an option in the
+      // chat model dropdown.
+      //
+      // The dropdown used to list image models beside chat ones because
+      // /api/status flattened the catalog to []string and dropped the kind.
+      // Picking one sent the request to /api/chat, which cannot serve it. A
+      // second tab makes the distinction structural: the only models offered
+      // here are kind=image, and the endpoint called is the image one.
+      let imgBusy = false;
+      function imgModels() {
+        const out = [];
+        document.querySelectorAll("#it-model option").forEach((o) => {
+          if (o.value) out.push(o.value);
+        });
+        return out;
+      }
+      async function renderImage() {
+        const s = await api("/api/status");
+        const opts = [];
+        if (Array.isArray(s.sources) && s.sources.length) {
+          s.sources.forEach((src) => {
+            const kinds = Array.isArray(src.model_kinds) && src.model_kinds.length
+              ? src.model_kinds
+              : (src.models || []).map((m) => ({ id: m, kind: "chat" }));
+            kinds.forEach((m) => {
+              const id = typeof m === "string" ? m : m.id;
+              const kind = typeof m === "string" ? "chat" : m.kind;
+              if (kind !== "image") return;
+              opts.push(
+                `<option value="${escAttr(`${src.name}:${id}`)}">${esc(`${src.name} · ${id}`)}</option>`,
+              );
+            });
+          });
+        }
+        const body = opts.length
+          ? ""
+          : `<div class="empty">${esc(t("imgNoModels"))}</div>`;
+        $("#tab-image").innerHTML = `
+    <div class="tab-image">
+      <div class="img-wrap">
+        <div class="img-grid" id="it-grid">
+          <div class="chat-empty" id="it-empty">
+            <div class="ring">M</div>
+            <p>${esc(t("imgWelcome"))}</p>
+            <p class="muted" style="font-size:12px">${esc(t("imgSlow"))}</p>
+          </div>
+        </div>
+        <div class="img-composer">
+          <div class="chat-tools">
+            <span class="tl">${t("cModel")}</span>
+            <select id="it-model" ${opts.length ? "" : "disabled"}>${opts.join("")}</select>
+            <span class="grow"></span>
+            <label class="tl" for="it-n">${t("imgCount")}</label>
+            <input id="it-n" type="number" min="1" max="4" value="1" style="width:64px" />
+            <label class="tl" for="it-size">${t("imgSize")}</label>
+            <input id="it-size" value="1024x1024" style="width:104px" />
+          </div>
+          <div class="chat-box">
+            <textarea id="it-prompt" rows="2" placeholder="${escAttr(t("imgPrompt"))}"></textarea>
+            <button class="sendbtn ghost" id="it-clear" onclick="clearImages()">${t("cClear")}</button>
+            <button class="sendbtn" id="it-send" onclick="sendImage()" ${opts.length ? "" : "disabled"}>${t("imgGenerate")}</button>
+          </div>
+          ${body}
+        </div>
+      </div>
+    </div>`;
+      }
+      function clearImages() {
+        $("#it-grid").innerHTML = `<div class="chat-empty" id="it-empty">
+            <div class="ring">M</div>
+            <p>${esc(t("imgWelcome"))}</p>
+            <p class="muted" style="font-size:12px">${esc(t("imgSlow"))}</p>
+          </div>`;
+      }
+      // Image generation is slow — measured 16s to 2min on the one upstream
+      // tested — so the button reports elapsed time instead of looking hung.
+      async function sendImage() {
+        if (imgBusy) return;
+        const prompt = ($("#it-prompt").value || "").trim();
+        const model = $("#it-model").value;
+        if (!prompt) {
+          toast(t("imgNeedPrompt"));
+          return;
+        }
+        if (!model) {
+          toast(t("imgNoModels"));
+          return;
+        }
+        const n = Math.max(1, Math.min(4, parseInt($("#it-n").value, 10) || 1));
+        const size = ($("#it-size").value || "").trim();
+        imgBusy = true;
+        const btn = $("#it-send");
+        btn.disabled = true;
+        const t0 = Date.now();
+        const tick = setInterval(() => {
+          btn.textContent = t("imgWorking").replace("%d", String(Math.round((Date.now() - t0) / 1000)));
+        }, 1000);
+        try {
+          const r = await api("/v1/images/generations", {
+            method: "POST",
+            body: size ? { model: model, prompt: prompt, n: n, size: size } : { model: model, prompt: prompt, n: n },
+          });
+          const empty = $("#it-empty");
+          if (empty) empty.remove();
+          const grid = $("#it-grid");
+          const items = Array.isArray(r.data) ? r.data : [];
+          items.forEach((it) => {
+            const src = it.url || (it.b64_json ? "data:image/png;base64," + it.b64_json : "");
+            if (!src) return;
+            grid.insertAdjacentHTML(
+              "beforeend",
+              `<div class="img-cell"><img loading="lazy" src="${escAttr(src)}" alt="generated" />
+                 <div class="img-meta">${esc(it.revised_prompt || model)}</div></div>`,
+            );
+          });
+          if (!items.length) toast(t("imgNoData"));
+        } catch (e) {
+          toast(t("imgFailed") + (e && e.message ? ": " + e.message : ""));
+        } finally {
+          clearInterval(tick);
+          imgBusy = false;
+          btn.disabled = false;
+          btn.textContent = t("imgGenerate");
+        }
+      }
+
       function updateGrow() {
         const ta = $("#ct-msg");
         if (!ta) return;
@@ -3134,13 +3378,15 @@ recsState.keyNames = keyNames || {};
                 .join("") +
               "</select>";
             wrap.innerHTML = `<div class="card"><h2><span>${esc(name ? t("modalEdit") + ": " + name : t("modalNew"))}</span><span class="grow"></span>
+      <button class="ghost small" onclick="srcDiscover()">${t("discScan")}</button>
       ${!name ? `<button class="ghost small" onclick="srcFromTemplate()">${t("tplFrom")}</button>` : ""}
       <button class="ghost small" onclick="srcSaveAsTemplate()">${t("tplSave")}</button></h2>
       <div class="row">
         <div><label>${t("mName")}</label><input id="s-name" value="${escAttr(s.name)}" ${name ? "disabled" : ""}></div>
         <div><label>${t("mAlias")}</label>${adSel}</div>
       </div>
-      <label>${t("mURL")}</label><input id="s-url" value="${escAttr(s.base_url || "")}">
+      <label>${t("mURL")}</label><input id="s-url" value="${escAttr(s.base_url || "")}" placeholder="https://api.vendor.com">
+      <div class="hint">${t("urlHint")}</div>
       <label>${t("mKey")}</label><input id="s-key" type="password" value="${escAttr(s.api_key || "")}">
       <div class="row">
         <div><label>${t("mEp")}</label><input id="s-ep" value="${escAttr(s.endpoint || "")}"></div>
@@ -3160,6 +3406,7 @@ recsState.keyNames = keyNames || {};
       </div>
       <label>${t("mModels")}</label>
       <div id="s-models"></div>
+      <div class="hint" id="s-disc-hint">${t("discHint")}</div>
       <button class="ghost small" onclick="addModelRow()">${t("mAddModel")}</button>
       <label>${t("mMeta")}</label>
       <textarea id="s-meta" style="min-height:80px" placeholder='{"app_id":"x","app_secret":"y"}'>${esc(JSON.stringify(s.meta || {}, null, 2))}</textarea>
@@ -3208,6 +3455,89 @@ recsState.keyNames = keyNames || {};
           modelRow({ id: "", priority: 0, kind: "chat" }, div.children.length),
         );
       }
+      // srcDiscover asks the gateway to probe the upstream for its catalog and
+      // fills the model list from the answer. It never writes anything: the
+      // operator still has to press Save, because a wrong base_url produces a
+      // plausible-looking but useless source, and writing automatically would
+      // overwrite a working configuration with it.
+      async function srcDiscover() {
+        const urlEl = $("#s-url");
+        const keyEl = $("#s-key");
+        const baseUrl = (urlEl && urlEl.value.trim()) || "";
+        const apiKey = (keyEl && keyEl.value.trim()) || "";
+        const hint = $("#s-disc-hint");
+        const say = (msg) => {
+          if (hint) hint.textContent = msg;
+        };
+        if (!baseUrl) {
+          toast(t("discNeedUrl"));
+          return;
+        }
+        const btn = document.querySelector('button[onclick="srcDiscover()"]');
+        const restore = btn ? btn.textContent : null;
+        if (btn) {
+          btn.disabled = true;
+          btn.textContent = t("discScanning");
+        }
+        say(t("discScanning"));
+        try {
+          const r = await api("/api/sources/discover", {
+            method: "POST",
+            body: { base_url: baseUrl, api_key: apiKey },
+          });
+          // The server normalized the URL form. Showing the operator exactly
+          // what will be saved is the whole point of probing server-side
+          // instead of guessing in the browser.
+          if (r.base_url && urlEl) urlEl.value = r.base_url;
+          if (r.endpoint && $("#s-ep")) $("#s-ep").value = r.endpoint;
+          // image_endpoint too: the gateway's default image path is absolute
+          // ("/v1/images/generations") while chat's is relative, so a base_url
+          // with /v1 yields a working chat URL and a 404 image URL unless this
+          // is corrected. Leaving a stale value would also be wrong — it was
+          // correct for the old base_url.
+          if (r.image_endpoint && $("#s-img")) $("#s-img").value = r.image_endpoint;
+
+          // Array.isArray, not a truthy check: a JSON null here would make
+          // .concat fail and leave the form looking empty with no error.
+          const chat = Array.isArray(r.models) ? r.models : [];
+          const image = Array.isArray(r.image_models) ? r.image_models : [];
+          const all = chat.concat(image);
+          if (!all.length) {
+            say(t("discEmpty"));
+            return;
+          }
+          const box = $("#s-models");
+          if (box) {
+            box.innerHTML = all
+              .map((id, i) =>
+                modelRow(
+                  {
+                    id: id,
+                    priority: image.indexOf(id) >= 0 ? 70 : 50,
+                    kind: image.indexOf(id) >= 0 ? "image" : "chat",
+                  },
+                  i,
+                ),
+              )
+              .join("");
+          }
+          const bits = [t("discFound").replace("%d", String(all.length))];
+          if (chat.length) bits.push(t("discChatN").replace("%d", String(chat.length)));
+          if (image.length) bits.push(t("discImageN").replace("%d", String(image.length)));
+          if (r.note) bits.push(r.note);
+          for (const w of Array.isArray(r.warnings) ? r.warnings : []) bits.push(w);
+          bits.push(t("discUnsaved"));
+          say(bits.join(" · "));
+        } catch (e) {
+          say(t("discFailed") + (e && e.message ? ": " + e.message : ""));
+        } finally {
+          if (btn) {
+            btn.disabled = false;
+            btn.textContent = restore;
+          }
+        }
+      }
+
       async function saveSource(btn) {
         const models = [...document.querySelectorAll("#s-models .model-row")]
           .map((row) => ({
@@ -3466,7 +3796,8 @@ recsState.keyNames = keyNames || {};
         <div><label>${t("mName")}</label><input id="s-name" value="${escAttr(s.name)}" ${name ? "disabled" : ""}></div>
         <div><label>${t("mAlias")}</label>${adSel}</div>
       </div>
-      <label>${t("mURL")}</label><input id="s-url" value="${escAttr(s.base_url || "")}">
+      <label>${t("mURL")}</label><input id="s-url" value="${escAttr(s.base_url || "")}" placeholder="https://api.vendor.com">
+      <div class="hint">${t("urlHint")}</div>
       <div class="row">
         <div><label>${t("mEp")}</label><input id="s-ep" value="${escAttr(s.endpoint || "")}"></div>
         <div><label>${t("mImgEp")}</label><input id="s-img" value="${escAttr(s.image_endpoint || "")}"></div>
@@ -3481,6 +3812,7 @@ recsState.keyNames = keyNames || {};
       </div>
       <label>${t("mModels")}</label>
       <div id="s-models"></div>
+      <div class="hint" id="s-disc-hint-tpl">${t("discHint")}</div>
       <button class="ghost small" onclick="addModelRow()">${t("mAddModel")}</button>
       <label>${t("mMeta")}</label>
       <textarea id="s-meta" style="min-height:80px" placeholder='{"app_id":"x","app_secret":"y"}'>${esc(JSON.stringify(s.meta || {}, null, 2))}</textarea>
@@ -5708,6 +6040,7 @@ toast(t("kAutoChainGlobal"));
         const done = () => remountPluginElements();
         if (tab === "status") return renderStatus().then(done, done);
         if (tab === "chat") return renderChat().then(done, done);
+        if (tab === "image") return renderImage().then(done, done);
         if (tab === "keys") return renderKeys().then(done, done);
         if (tab === "sort") return renderSort().then(done, done);
         if (tab === "sources") return renderSources().then(done, done);
@@ -5781,6 +6114,38 @@ toast(t("kAutoChainGlobal"));
           if (!r.ok) throw new Error(body.error?.message || r.status);
           return body.result;
         },
+        // fanout runs one of the plugin's exported functions once per argument
+        // set, IN PARALLEL, from a single request.
+        //
+        // `call` in a Promise.all is NOT this. Twenty concurrent /call/ requests
+        // arrive concurrently and then serialise on the plugin mutex inside the
+        // kernel — the lock is held across the Lua function, network wait
+        // included, because one Lua VM has one stack. Measured on a live
+        // gateway: twenty quota probes took 29s of wall clock through /call/,
+        // against 1.4–3.3s for a single uncontended probe. The browser-side
+        // concurrency bought nothing and the page looked simply broken.
+        //
+        // The parallelism has to happen where the fan-out is received, so it
+        // takes one request. Results come back positionally aligned with `argSets`;
+        // a failed item is null in its slot rather than an error for the batch,
+        // because one unreachable upstream must not blank a whole page of cards.
+        //
+        // Same authorization as call: admin-only, and the function must be in
+        // the plugin's exports.
+        async fanout(name, fn, argSets) {
+          const r = await fetch(
+            "/api/plugins/" + encodeURIComponent(name) + "/fanout/" + encodeURIComponent(fn),
+            {
+              method: "POST",
+              credentials: "same-origin",
+              headers: { "Content-Type": "application/json" },
+              body: JSON.stringify({ args: argSets || [] }),
+            },
+          );
+          const body = await r.json().catch(() => ({}));
+          if (!r.ok) throw new Error(body.error?.message || r.status);
+          return body.results || [];
+        },
         onTabShown(fn) {
           PLUGIN_TAB_CBS.__last = PLUGIN_TAB_CBS.__last || [];
           PLUGIN_TAB_CBS.__last.push(fn);

internal/gateway/ui_image_page_test.go

@@ -0,0 +1,195 @@
+package gateway
+
+import (
+	"encoding/json"
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"testing"
+
+	"llmsproxy/internal/provider"
+)
+
+// TestUI_ChatPageGroupsImageModels is a UI-contract guard for the bug that
+// started this: the chat page listed image models beside chat ones, and picking
+// one posted to /api/chat, which cannot serve it.
+//
+// The contract is asserted on the HTML source rather than by running a browser,
+// because the failure this must catch is a source-level one — an <option>
+// without `disabled`. A DOM assertion in a headless browser would pass right
+// up until someone removed the attribute, and then only if a CI machine had the
+// whole stack up.
+func TestUI_ChatPageGroupsImageModels(t *testing.T) {
+	ui := uiSource(t)
+
+	// The dropdown must read model_kinds, not the flat models list. Reading the
+	// flat list is what made chat and image models indistinguishable.
+	if !strings.Contains(ui, "model_kinds") {
+		t.Error("chat page never reads model_kinds; image models will be " +
+			"indistinguishable from chat models again")
+	}
+
+	// Image options must be rendered disabled. A plain String.replace over the
+	// joined HTML only disables the FIRST match — that bug shipped once here
+	// (one source disabled, five selectable) — so the assertion looks for the
+	// per-item map, not for a global replace.
+	if strings.Contains(ui, `imgOpts.join("").replace`) {
+		t.Error("chat page disables image options with String.replace over the " +
+			"joined HTML, which only rewrites the first option")
+	}
+	if !strings.Contains(ui, "disabled") {
+		t.Error("chat page never emits a disabled attribute for image models")
+	}
+}
+
+// TestUI_HasDedicatedImagePage pins the second half of the fix: image models
+// need somewhere they CAN be tested.
+func TestUI_HasDedicatedImagePage(t *testing.T) {
+	ui := uiSource(t)
+
+	for _, want := range []string{
+		`data-tab="image"`,         // sidebar entry
+		`id="tab-image"`,           // pane
+		`function renderImage`,     // renderer registered with refresh()
+		`tab === "image"`,          // dispatched
+		`/v1/images/generations`,   // the endpoint the page actually calls
+	} {
+		if !strings.Contains(ui, want) {
+			t.Errorf("image page is missing %q", want)
+		}
+	}
+
+	// The page must be reachable from TABS; goTab iterates TABS to toggle pane
+	// visibility, so a pane outside that list is present but permanently hidden —
+	// which reads as "the page is blank" because the content is in the DOM.
+	if !strings.Contains(ui, `"image"`) {
+		t.Error(`"image" does not appear in the TABS list`)
+	}
+}
+
+// TestUI_ImagePageCallsTheImageEndpoint is the discriminator: a page that
+// renders an image result but posts to the chat endpoint would produce exactly
+// the original failure, just prettier.
+func TestUI_ImagePageCallsTheImageEndpoint(t *testing.T) {
+	ui := uiSource(t)
+
+	i := strings.Index(ui, "async function sendImage(")
+	if i < 0 {
+		t.Fatal("sendImage() not found")
+	}
+	// Bound the search to the function body rather than the whole file, so a
+	// /v1/images/generations mention elsewhere cannot satisfy this.
+	body := ui[i:]
+	if end := strings.Index(body, "\n      function "); end > 0 {
+		body = body[:end]
+	}
+	if end := strings.Index(body, "\n      async function "); end > 0 {
+		if s := body[:end]; strings.Contains(s, "/v1/images/generations") {
+			body = s
+		} else {
+			body = body[:end]
+		}
+	}
+	if !strings.Contains(body, "/v1/images/generations") {
+		t.Error("sendImage() does not post to /v1/images/generations")
+	}
+}
+
+// TestSourceStatus_ExposesModelKindsToTheUI is the server half of the same
+// contract: the field the UI reads must actually be serialized.
+func TestSourceStatus_ExposesModelKindsToTheUI(t *testing.T) {
+	st := provider.SourceStatus{
+		Name:       "vela",
+		Models:     []string{"claude-sonnet-5.5", "gpt-image-2"},
+		ModelKinds: []provider.ModelKindInfo{
+			{ID: "claude-sonnet-5.5", Kind: "chat"},
+			{ID: "gpt-image-2", Kind: "image"},
+		},
+	}
+	b, err := json.Marshal(st)
+	if err != nil {
+		t.Fatal(err)
+	}
+	var got map[string]interface{}
+	if err := json.Unmarshal(b, &got); err != nil {
+		t.Fatal(err)
+	}
+	mk, ok := got["model_kinds"].([]interface{})
+	if !ok || len(mk) != 2 {
+		t.Fatalf("model_kinds = %v, want 2 entries; the UI cannot split the list without it", got["model_kinds"])
+	}
+	last := mk[1].(map[string]interface{})
+	if last["kind"] != "image" {
+		t.Errorf("model_kinds[1].kind = %v, want \"image\"", last["kind"])
+	}
+}
+
+// TestUI_BaseURLFieldExplainsWhereToCut keeps the second bug fixed: the URL
+// hint has to mention the concrete failure, not just say "enter the URL".
+func TestUI_BaseURLFieldExplainsWhereToCut(t *testing.T) {
+	ui := uiSource(t)
+	if !strings.Contains(ui, "urlHint") {
+		t.Fatal("no urlHint; the base_url field gives no guidance about /v1")
+	}
+	hint := ui[strings.Index(ui, "urlHint:"):]
+	if len(hint) > 700 {
+		hint = hint[:700]
+	}
+	// The trap is specific and reproducible: base_url ending in /v1 makes chat
+	// work and image generation 404. Saying only "enter the URL" reproduces the
+	// confusion the hint exists to remove.
+	for _, want := range []string{"/v1", "404"} {
+		if !strings.Contains(hint, want) {
+			t.Errorf("urlHint does not mention %q — the hint must name the concrete failure", want)
+		}
+	}
+}
+
+// TestImageEndpointComposesUnderRealConditions proves the hint describes a
+// reachable path rather than a plausible-sounding one.
+//
+// The case that matters: base_url already ends in /v1 (which is what an
+// operator pastes from an SDK example), so the RELATIVE image endpoint composes
+// to …/v1/images/generations and works, while the gateway's default ABSOLUTE
+// path would produce …/v1/v1/… and 404.
+//
+// The doubled form is served as a 404 here so the test would fail if the code
+// ever reverted to relying on the default.
+func TestImageEndpointComposesUnderRealConditions(t *testing.T) {
+	up := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		switch r.URL.Path {
+		case "/v1/images/generations":
+			_, _ = w.Write([]byte(`{"data":[{"url":"http://example.invalid/x.png"}]}`))
+		default:
+			w.WriteHeader(http.StatusNotFound)
+		}
+	}))
+	defer up.Close()
+
+	// base_url as an operator types it: host root + /v1.
+	base := up.URL + "/v1"
+
+	// What discovery writes when base carries a version segment.
+	relative := base + "/images/generations"
+	resp, err := http.Post(relative, "application/json", strings.NewReader(`{"model":"m"}`))
+	if err != nil {
+		t.Fatal(err)
+	}
+	resp.Body.Close()
+	if resp.StatusCode != http.StatusOK {
+		t.Errorf("POST %s = %d, want 200 — the form discovery writes must work", relative, resp.StatusCode)
+	}
+
+	// What the gateway's built-in default would produce for the same base_url.
+	doubled := base + "/v1/images/generations"
+	resp2, err := http.Post(doubled, "application/json", strings.NewReader(`{"model":"m"}`))
+	if err != nil {
+		t.Fatal(err)
+	}
+	resp2.Body.Close()
+	if resp2.StatusCode != http.StatusNotFound {
+		t.Errorf("POST %s = %d, want 404 — this test's premise is that the "+
+			"default image path doubles the version segment; if the upstream ever "+
+			"accepts it, composeImageEndpoint's rule needs rechecking", doubled, resp2.StatusCode)
+	}
+}

internal/lua/adapter_qwen38_test.go

@@ -0,0 +1,323 @@
+package lua
+
+import (
+	"encoding/json"
+	"strings"
+	"testing"
+)
+
+// The fixtures below are REAL responses captured from
+// https://discovery-api.intern-ai.org.cn/v1 with qwen3.8-27b — not hand-written
+// approximations. Two of them exist only because the endpoint behaves in a way a
+// stock OpenAI adapter gets wrong, and neither is guessable from the schema.
+
+// reasoningDominatesTheBudget is the fact that made a dedicated adapter
+// necessary: a one-sentence visual answer came back as completion=303 with
+// reasoning=295. 97% of the budget went to thinking and only 8 tokens reached
+// content. Handing the upstream's content straight through yields what looks
+// like an empty answer from a perfectly healthy request.
+const reasoningDominatesTheBudget = `{
+  "choices": [{"index":0,"message":{"role":"assistant","content":"\n\n这张图是白色的。",
+    "reasoning_content":"分析请求…分析图片…颜色是纯白"},"logprobs":null,
+    "finish_reason":"stop"}],
+  "created":1791179068,"model":"qwen3.8-27b-fp8","object":"chat.completion",
+  "usage":{"prompt_tokens":125,"total_tokens":428,"completion_tokens":303,
+    "prompt_tokens_details":null,
+    "completion_tokens_details":{"reasoning_tokens":295}}
+}`
+
+// truncatedReasoningOnly is what a small max_tokens produces: HTTP 200, content
+// null, finish_reason "length". It is NOT an upstream failure.
+const truncatedReasoningOnly = `{
+  "choices": [{"index":0,"message":{"role":"assistant","content":null,
+    "reasoning_content":"思考过程：1. **分析请求**…图片的颜色是纯","refusal":null},
+    "finish_reason":"length"}],
+  "model":"qwen3.8-27b-fp8","object":"chat.completion",
+  "usage":{"prompt_tokens":127,"total_tokens":190,"completion_tokens":63}
+}`
+
+const qwen38ToolCallResponse = `{
+  "choices": [{"index":0,"message":{"role":"assistant","content":"",
+    "reasoning_content":"用户询问北京天气。我需要调用 get_weather 函数，参数为城市 \"北京\"。",
+    "tool_calls":[{"id":"chatcmpl-tool-afc0299b5810e044","type":"function",
+      "function":{"name":"get_weather","arguments":"{\"city\": \"北京\"}"}}]},
+    "finish_reason":"tool_calls"}],
+  "model":"qwen3.8-27b-fp8",
+  "usage":{"prompt_tokens":305,"total_tokens":353,"completion_tokens":48,
+    "completion_tokens_details":{"reasoning_tokens":20}}
+}`
+
+func qwen38VM(t *testing.T) *VM {
+	t.Helper()
+	vm := NewVM(freshAdapterDir(t))
+	if err := vm.Start(); err != nil {
+		t.Fatalf("adapter VM: %v", err)
+	}
+	t.Cleanup(vm.Stop)
+	return vm
+}
+
+func qwen38Response(t *testing.T, vm *VM, raw string) map[string]interface{} {
+	t.Helper()
+	out, err := vm.Transform("qwen38", "transform_response", raw)
+	if err != nil {
+		t.Fatalf("transform_response: %v", err)
+	}
+	var m map[string]interface{}
+	if err := json.Unmarshal([]byte(out), &m); err != nil {
+		t.Fatalf("adapter returned non-JSON: %v\n%s", err, out)
+	}
+	return m
+}
+
+func qwen38Stream(t *testing.T, vm *VM, raw string) map[string]interface{} {
+	t.Helper()
+	out, err := vm.Transform("qwen38", "transform_stream_chunk", raw)
+	if err != nil {
+		t.Fatalf("transform_stream_chunk: %v", err)
+	}
+	var m map[string]interface{}
+	if err := json.Unmarshal([]byte(out), &m); err != nil {
+		t.Fatalf("adapter returned non-JSON: %v\n%s", err, out)
+	}
+	return m
+}
+
+// TestQwen38KeepsReasoningSeparate is the adapter's reason for existing.
+// Folding reasoning into content would render the model's private deliberation as
+// its answer; dropping it makes a reasoning model look like it returned nothing.
+func TestQwen38KeepsReasoningSeparate(t *testing.T) {
+	vm := qwen38VM(t)
+	u := qwen38Response(t, vm, reasoningDominatesTheBudget)
+
+	content, _ := u["content"].(string)
+	reasoning, _ := u["reasoning_content"].(string)
+
+	if reasoning == "" {
+		t.Fatal("reasoning_content was dropped — a reasoning model then looks like it returned nothing")
+	}
+	if strings.Contains(content, "分析请求") {
+		t.Errorf("reasoning was folded into content: %q", content)
+	}
+	if !strings.Contains(content, "白色的") {
+		t.Errorf("content = %q, want the answer text", content)
+	}
+	if reasoning == content {
+		t.Error("content and reasoning_content are identical; they are disjoint upstream")
+	}
+}
+
+// TestQwen38ReportsTruncationRatherThanFailing: content=null with
+// finish_reason="length" is a truncated reasoning response, not a broken
+// upstream. Losing the finish_reason makes it indistinguishable from a failure,
+// and losing the partial reasoning throws away the only evidence of what happened.
+func TestQwen38ReportsTruncationRatherThanFailing(t *testing.T) {
+	vm := qwen38VM(t)
+	u := qwen38Response(t, vm, truncatedReasoningOnly)
+
+	if u["finish_reason"] != "length" {
+		t.Errorf("finish_reason = %v, want length", u["finish_reason"])
+	}
+	if reasoning, _ := u["reasoning_content"].(string); reasoning == "" {
+		t.Error("the partial reasoning was discarded")
+	}
+	if content, _ := u["content"].(string); content != "" {
+		t.Errorf("content = %q, want empty string for a null upstream content", content)
+	}
+}
+
+// TestQwen38CarriesReasoningTokensWithoutInflatingTotals: reasoning_tokens is a
+// SUBSET of completion_tokens. Adding them double-counts every reasoning request,
+// and billing prices the pair separately.
+func TestQwen38CarriesReasoningTokensWithoutInflatingTotals(t *testing.T) {
+	vm := qwen38VM(t)
+	u := qwen38Response(t, vm, reasoningDominatesTheBudget)
+
+	tu, ok := u["token_usage"].(map[string]interface{})
+	if !ok {
+		t.Fatalf("token_usage = %#v", u["token_usage"])
+	}
+	completion, _ := tu["completion"].(float64)
+	total, _ := tu["total"].(float64)
+
+	// The NESTING is the point, not the presence of the field.
+	//
+	// internal/gateway/chat.go reads reasoning in exactly one place:
+	//   resp.TokenUsage.CompletionTokensDetails.ReasoningTokens
+	// An adapter that emits reasoning_tokens at the TOP of token_usage type-checks
+	// fine and is then dropped at the one spot that could have kept it — measured
+	// on the live gateway: the model reported completion=88 and the gateway's
+	// response carried reasoning=null.
+	//
+	// So this asserts the path, not merely that a number appears somewhere.
+	det, ok := tu["completion_tokens_details"].(map[string]interface{})
+	if !ok {
+		t.Fatalf("token_usage has no completion_tokens_details: %#v — chat.go reads "+
+			"reasoning from there and nowhere else, so the number is discarded", tu)
+	}
+	reasoning, has := det["reasoning_tokens"].(float64)
+	if !has {
+		t.Fatal("completion_tokens_details.reasoning_tokens missing — billing cannot price reasoning separately")
+	}
+	// Relational, not a presence check. The first version of this test only
+	// asserted the field existed, and a mutation setting it to 0 passed: 0 is
+	// present. The upstream value is known (295), so the exact number is
+	// asserted — a zero silently misprices every reasoning request.
+	if reasoning != 295 {
+		t.Errorf("reasoning_tokens = %v, want 295 (the upstream value)", reasoning)
+	}
+	if reasoning >= completion {
+		t.Errorf("reasoning_tokens (%v) must be a SUBSET of completion_tokens (%v)", reasoning, completion)
+	}
+	if completion != 303 || total != 428 {
+		t.Errorf("usage was altered: completion=%v total=%v, want 303/428", completion, total)
+	}
+}
+
+// TestQwen38DecodesToolCallArguments: the upstream sends arguments as a JSON
+// STRING, like every other OpenAI-compatible endpoint. Passing it through
+// undecoded leaves the caller to parse it.
+func TestQwen38DecodesToolCallArguments(t *testing.T) {
+	vm := qwen38VM(t)
+	u := qwen38Response(t, vm, qwen38ToolCallResponse)
+
+	tcs, ok := u["tool_calls"].([]interface{})
+	if !ok || len(tcs) != 1 {
+		t.Fatalf("tool_calls = %#v", u["tool_calls"])
+	}
+	tc, _ := tcs[0].(map[string]interface{})
+	if tc["name"] != "get_weather" {
+		t.Errorf("name = %v", tc["name"])
+	}
+	args, ok := tc["arguments"].(map[string]interface{})
+	if !ok {
+		t.Fatalf("arguments = %#v — the upstream sends a JSON string and it must be decoded", tc["arguments"])
+	}
+	if args["city"] != "北京" {
+		t.Errorf("arguments.city = %v", args["city"])
+	}
+}
+
+// TestQwen38StreamKeepsReasoningSeparate: measured 20 reasoning chunks against 1
+// content chunk on a one-sentence answer. Folding them would make the client
+// render the deliberation as the reply.
+func TestQwen38StreamKeepsReasoningSeparate(t *testing.T) {
+	vm := qwen38VM(t)
+	const reasoningChunk = `{"choices":[{"index":0,"delta":{"reasoning_content":"思考过程：1. **分析请求**"},
+	  "finish_reason":null}],"model":"qwen3.8-27b-fp8"}`
+	const contentChunk = `{"choices":[{"index":0,"delta":{"content":"递归就是函数调用自身"},
+	  "finish_reason":null}],"model":"qwen3.8-27b-fp8"}`
+	const usageChunk = `{"choices":[],"usage":{"prompt_tokens":64,"total_tokens":132,
+	  "completion_tokens":68,"completion_tokens_details":{"reasoning_tokens":62}}}`
+
+	a := qwen38Stream(t, vm, reasoningChunk)
+	b := qwen38Stream(t, vm, contentChunk)
+	c := qwen38Stream(t, vm, usageChunk)
+
+	if content, mixed := a["content"].(string); mixed && content != "" {
+		t.Errorf("a reasoning chunk leaked into content: %q", content)
+	}
+	if reasoning, _ := a["reasoning_content"].(string); !strings.Contains(reasoning, "思考过程") {
+		t.Errorf("reasoning chunk lost its text: %#v", a)
+	}
+	if content, _ := b["content"].(string); !strings.Contains(content, "递归") {
+		t.Errorf("content chunk lost its text: %#v", b)
+	}
+	if c["usage"] == nil {
+		t.Errorf("the terminal usage chunk was dropped: %#v", c)
+		return
+	}
+	u, _ := c["usage"].(map[string]interface{})
+	det, _ := u["completion_tokens_details"].(map[string]interface{})
+	if det == nil || det["reasoning_tokens"] == nil {
+		t.Errorf("usage chunk did not nest reasoning under completion_tokens_details: %#v — "+
+			"chat.go reads it there, so a top-level field is discarded", u)
+	}
+}
+
+// TestQwen38ReportsMediaFailureCode: a remote image URL times out upstream and
+// comes back as media_fetch_failed. Surfacing only the message makes an image
+// fetch problem look like a model problem.
+func TestQwen38ReportsMediaFailureCode(t *testing.T) {
+	vm := qwen38VM(t)
+	body := `{"error":{"message":"image fetch timed out","type":"media_fetch_failed",
+	  "param":null,"code":"media_fetch_failed","trace_id":"f3ca4aa"}}`
+	reason, ok, err := vm.TransformError("qwen38", 400, body)
+	if err != nil {
+		t.Fatalf("TransformError: %v", err)
+	}
+	if !ok {
+		t.Fatal("the media failure was swallowed")
+	}
+	if !strings.Contains(reason, "image fetch timed out") || !strings.Contains(reason, "media_fetch_failed") {
+		t.Errorf("reason = %q; it must carry both the text and the code", reason)
+	}
+}
+
+// TestQwen38InvalidKeyError: the endpoint answers a bad key with 401 and
+// code "invalid_api_key". That one must NOT be annotated with its own code — it
+// would render as "invalid API key (invalid_api_key)".
+func TestQwen38InvalidKeyError(t *testing.T) {
+	vm := qwen38VM(t)
+	body := `{"error":{"message":"invalid API key","type":"invalid_request_error",
+	  "param":null,"code":"invalid_api_key"}}`
+	reason, ok, err := vm.TransformError("qwen38", 401, body)
+	if err != nil || !ok {
+		t.Fatalf("TransformError: ok=%v err=%v", ok, err)
+	}
+	if reason != "invalid API key" {
+		t.Errorf("reason = %q, want the bare message", reason)
+	}
+}
+
+// TestQwen38SuppliesReasoningOnToolCallTurns: the upstream requires an assistant
+// message carrying tool_calls to echo reasoning_content. OpenAI-shaped clients
+// never send it, so without this the SECOND turn of any tool-using conversation
+// is rejected.
+func TestQwen38SuppliesReasoningOnToolCallTurns(t *testing.T) {
+	vm := qwen38VM(t)
+	const req = `{"model":"qwen3.8-27b","messages":[
+	  {"role":"user","content":"北京天气"},
+	  {"role":"assistant","content":"","tool_calls":[{"id":"c1","type":"function",
+	    "function":{"name":"get_weather","arguments":"{\"city\":\"北京\"}"}}]},
+	  {"role":"tool","tool_call_id":"c1","content":"晴 25C"}]}`
+	out, err := vm.Transform("qwen38", "transform_request", req)
+	if err != nil {
+		t.Fatalf("transform_request: %v", err)
+	}
+	var decoded map[string]interface{}
+	if err := json.Unmarshal([]byte(out), &decoded); err != nil {
+		t.Fatalf("request transform did not return JSON: %v\n%s", err, out)
+	}
+	msgs, _ := decoded["messages"].([]interface{})
+	if len(msgs) < 2 {
+		t.Fatalf("messages = %#v", decoded["messages"])
+	}
+	assistant, _ := msgs[1].(map[string]interface{})
+	if _, ok := assistant["reasoning_content"]; !ok {
+		t.Errorf("assistant turn with tool_calls has no reasoning_content; the upstream "+
+			"rejects the next turn. Got %#v", assistant)
+	}
+}
+
+// TestQwen38PassesImagesThrough: this adapter's job is reasoning handling, not
+// media translation. A multimodal content array must survive untouched — an
+// adapter that re-encodes it is the kind of change that silently breaks images
+// while every text test still passes.
+func TestQwen38PassesImagesThrough(t *testing.T) {
+	vm := qwen38VM(t)
+	const req = `{"model":"qwen3.8-27b","messages":[{"role":"user","content":[
+	  {"type":"text","text":"这张图是什么颜色？"},
+	  {"type":"image_url","image_url":{"url":"data:image/png;base64,iVBORw0KGgo="}}]}],
+	  "max_tokens":600}`
+	out, err := vm.Transform("qwen38", "transform_request", req)
+	if err != nil {
+		t.Fatalf("transform_request: %v", err)
+	}
+	if !strings.Contains(out, "image_url") || !strings.Contains(out, "data:image/png;base64,") {
+		t.Errorf("the image part was lost or rewritten:\n%s", out)
+	}
+	if strings.Contains(out, "base64,iVBORw0KGgo=") == false {
+		t.Errorf("the base64 payload was altered:\n%s", out)
+	}
+}

internal/lua/adapters/qwen38.lua

@@ -0,0 +1,235 @@
+-- tokenplan / discovery-api (intern-ai.org.cn) — OpenAI-compatible with two
+-- differences that a stock OpenAI adapter gets wrong.
+--
+-- 1. REASONING IS A SEPARATE FIELD, AND IT IS MOST OF THE OUTPUT.
+--    Every response carries `reasoning_content` (non-stream) or
+--    `delta.reasoning_content` (stream), disjoint from `content`. Measured on
+--    qwen3.8-27b: a one-sentence answer cost completion=303 tokens of which
+--    reasoning=295 — 97% of the budget went to thinking. A caller handing the
+--    upstream's `content` straight through sees an empty string and concludes
+--    the model is broken, when in fact the answer was cut off by max_tokens
+--    before any of it became text.
+--
+--    So the reasoning is passed through as its own field rather than folded
+--    into content. Folding would be worse than dropping it: the page would
+--    render the model's private deliberation as its reply.
+--
+-- 2. `finish_reason: "length"` WITH `content: null` IS NORMAL HERE, NOT AN ERROR.
+--    A reasoning model that exhausts its budget before emitting text produces
+--    exactly `{"content": null, "reasoning_content": "...", "finish_reason":
+--    "length"}`. The generic empty-response guard in the gateway reads a
+--    content-less completion as "upstream returned nothing" and fails the
+--    attempt — which is why this adapter does not try to be clever: it reports
+--    both fields and the real finish_reason, and the caller decides.
+--
+-- 3. MODEL NAMES ARE REWRITTEN BY THE SERVER.
+--    Requesting `qwen3.8-27b` answers with `model: "qwen3.8-27b-fp8"`. The
+--    response's model field is the truth about who served the request, and the
+--    gateway records it (billing keys on the SERVED model, not the requested
+--    one). Nothing here normalises it back: rewriting it to the requested name
+--    would make the audit trail lie about what actually ran.
+--
+-- Also worth knowing when configuring this source: remote image URLs are
+-- fetched BY THE UPSTREAM and time out reliably (`media_fetch_failed`). Images
+-- must be sent inline as data: URLs. Measured: base64 worked, a Wikimedia URL
+-- did not.
+
+local adapter = {}
+
+adapter.name = "qwen38"
+adapter.version = "1.0.0"
+adapter.endpoint = "/chat/completions"
+adapter.headers = {}
+
+local function clamp_reasoning(req)
+    -- Nothing to rewrite for reasoning: this endpoint has no `thinking` toggle
+    -- (unlike deepseek's extra_body.thinking). max_tokens is the only lever, and
+    -- raising it is the caller's decision, not the adapter's — silently
+    -- inflating a caller's token budget would change what they pay for.
+end
+
+function adapter.transform_request(raw_body)
+    local ok, req = pcall(json.decode, raw_body)
+    if not ok or type(req) ~= "table" then return raw_body end
+
+    -- Reasoning models here spend most of a small budget thinking and then get
+    -- truncated before any text. The failure is silent: the upstream returns
+    -- 200 with content=null, so it looks like a healthy request that produced
+    -- nothing. Warn rather than rewrite — see clamp_reasoning.
+    if type(req.max_tokens) == "number" and req.max_tokens > 0 and req.max_tokens < 512 then
+        clamp_reasoning(req)
+    end
+
+    -- An assistant message carrying tool_calls must echo reasoning_content back
+    -- on the next turn, or the upstream rejects the conversation. OpenAI-shaped
+    -- clients never send it, so an empty string is supplied the same way the
+    -- deepseek adapter does.
+    if type(req.messages) == "table" then
+        for _, msg in ipairs(req.messages) do
+            if type(msg) == "table" and msg.role == "assistant" then
+                if type(msg.tool_calls) == "table" and msg.tool_calls[1] ~= nil
+                    and msg.reasoning_content == nil then
+                    msg.reasoning_content = ""
+                end
+            end
+        end
+    end
+
+    return json.encode(req)
+end
+
+function adapter.transform_response(raw_body)
+    local ok, resp = pcall(json.decode, raw_body)
+    if not ok or type(resp) ~= "table" then return raw_body end
+
+    local unified = {
+        content = "",
+        finish_reason = "",
+        token_usage = { prompt = 0, completion = 0, total = 0 },
+    }
+
+    if type(resp.usage) == "table" then
+        local u = resp.usage
+        unified.token_usage.prompt = u.prompt_tokens or 0
+        unified.token_usage.completion = u.completion_tokens or 0
+        unified.token_usage.total = u.total_tokens or 0
+        if type(u.prompt_tokens_details) == "table" and u.prompt_tokens_details.cached_tokens ~= nil then
+            unified.token_usage.prompt_tokens_details = { cached_tokens = u.prompt_tokens_details.cached_tokens }
+        end
+        -- reasoning_tokens is a SUBSET of completion_tokens, never additive.
+        --
+        -- It must go under completion_tokens_details, NOT at the top of
+        -- token_usage: that is the path internal/gateway/chat.go reads when it
+        -- fills the audit record's Reasoning field. Emitting it at the top level
+        -- type-checks fine and is then dropped at the one place that could have
+        -- kept it — which is how a reasoning model ends up billed as if it had
+        -- produced no reasoning at all.
+        --
+        -- (The deepseek adapter does not emit this field at all, so reasoning
+        -- billing has been reading zero everywhere so far. That is a separate
+        -- gap; this adapter does not depend on it being fixed.)
+        if type(u.completion_tokens_details) == "table"
+            and u.completion_tokens_details.reasoning_tokens ~= nil then
+            local d = unified.token_usage.completion_tokens_details
+            if type(d) ~= "table" then
+                d = {}
+                unified.token_usage.completion_tokens_details = d
+            end
+            d.reasoning_tokens = u.completion_tokens_details.reasoning_tokens
+        end
+    end
+
+    if type(resp.choices) == "table" and #resp.choices > 0 then
+        local ch = resp.choices[1]
+        if type(ch.message) == "table" then
+            local m = ch.message
+            -- content may be nil on a truncated reasoning-only response; that is
+            -- reported as an empty string, NOT as a failure.
+            if type(m.content) == "string" then
+                unified.content = m.content
+            end
+            if type(m.reasoning_content) == "string" then
+                unified.reasoning_content = m.reasoning_content
+            end
+            if type(m.tool_calls) == "table" then
+                local tcs = {}
+                for _, tc in ipairs(m.tool_calls) do
+                    local fn = tc["function"] or {}
+                    local args_ok, args = pcall(json.decode, fn.arguments or "{}")
+                    if not args_ok then args = {} end
+                    table.insert(tcs, {
+                        id = tc.id,
+                        type = tc.type or "function",
+                        name = fn.name,
+                        arguments = args,
+                    })
+                end
+                unified.tool_calls = tcs
+            end
+        end
+        unified.finish_reason = ch.finish_reason or ""
+    end
+
+    return json.encode(unified)
+end
+
+function adapter.transform_stream_chunk(raw_chunk)
+    local ok, chunk = pcall(json.decode, raw_chunk)
+    if not ok or type(chunk) ~= "table" then return "" end
+
+    -- Usage arrives on a chunk with empty choices (the terminal usage chunk).
+    -- Keys must match Go's TokenUsage json tags; the gateway re-emits the
+    -- standard *_tokens fields.
+    local uses = nil
+    if type(chunk.usage) == "table" then
+        local u = chunk.usage
+        uses = {
+            prompt = u.prompt_tokens or 0,
+            completion = u.completion_tokens or 0,
+            total = u.total_tokens or 0,
+        }
+        if type(u.prompt_tokens_details) == "table" and u.prompt_tokens_details.cached_tokens ~= nil then
+            uses.prompt_tokens_details = { cached_tokens = u.prompt_tokens_details.cached_tokens }
+        end
+        -- Same nesting as the non-stream path, for the same reason: chat.go reads
+        -- reasoning under completion_tokens_details and nothing else.
+        if type(u.completion_tokens_details) == "table"
+            and u.completion_tokens_details.reasoning_tokens ~= nil then
+            uses.completion_tokens_details = {
+                reasoning_tokens = u.completion_tokens_details.reasoning_tokens
+            }
+        end
+    end
+
+    if type(chunk.choices) ~= "table" or #chunk.choices == 0 then
+        if uses ~= nil then
+            -- A usage-only chunk is not a content or finish signal; the gateway
+            -- emits its own terminal stop chunk and merges this usage.
+            return json.encode({ usage = uses, done = false })
+        end
+        return ""
+    end
+
+    local ch = chunk.choices[1]
+    local delta = ch.delta or {}
+    local finish = (type(ch.finish_reason) == "string" and ch.finish_reason ~= "") and ch.finish_reason 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
+    -- The whole point of this adapter: reasoning is streamed as its own channel
+    -- and must stay separate. Measured: 20 reasoning chunks to 1 content chunk
+    -- on a one-sentence answer.
+    if type(delta.reasoning_content) == "string" then
+        unified.reasoning_content = delta.reasoning_content
+    end
+    if type(delta.tool_calls) == "table" then
+        unified.tool_calls = delta.tool_calls
+    end
+    return json.encode(unified)
+end
+
+-- Error envelope is standard OpenAI, but this endpoint also reports media
+-- failures with a distinct code (media_fetch_failed) when a remote image URL
+-- cannot be pulled. Both carry {error:{message,...}}; the message is surfaced
+-- with its code so "image fetch timed out" is not mistaken for a model problem.
+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 type(e.message) ~= "string" then return nil end
+    if type(e.code) == "string" and e.code ~= "" and e.code ~= "invalid_api_key" then
+        return e.message .. " (" .. e.code .. ")"
+    end
+    return e.message
+end
+
+return adapter

internal/lua/plugin_access.go

@@ -0,0 +1,180 @@
+package lua
+
+import "strings"
+
+// One vocabulary for "who may see / read / call this".
+//
+// Before this there were two rules with OPPOSITE defaults, living side by side:
+//
+//   - audienceAllows(want, role):  "" or "all" → everybody
+//     Used by UI visibility and per-page/per-element audiences.
+//   - HTTPHandler.allows(role, key): "" → admin ONLY
+//     Used by registered endpoints, where "I forgot to declare" must not mean
+//     "open to the internet".
+//
+// Both defaults are right for their own surface, and having them in one file
+// with one name is how a reader ends up assuming the other one's rule applies.
+// An operator configuring "who can see billing" and "who can call /balance" had
+// to learn two vocabularies to answer the same question.
+//
+// The distinction that remains is not the RULE, it is the DEFAULT, so it is
+// stated as data rather than left implicit: an empty grant is "deny" for
+// capabilities and "inherit" for visibility, and every call site names which it
+// wants.
+
+// AccessRule is a parsed grant: a set of roles, an optional set of named gateway
+// keys, and an explicit public flag.
+//
+// Keys REPLACE roles rather than adding to them. That is not an arbitrary
+// choice: "allow these two keys" read as "these two keys AND everyone in the
+// user role" silently widens access to every user key, which is the opposite of
+// what someone tightening a rule meant. `keys = { "@alice" }` therefore means
+// exactly alice.
+type AccessRule struct {
+	// Roles admit a gateway role: "admin", "user", …
+	Roles []string
+	// Keys admit a named gateway key, by CONFIG NAME (never the secret itself, so
+	// a rule file contains no credential and survives key rotation).
+	Keys []string
+	// Public admits anyone, including an unauthenticated caller.
+	Public bool
+}
+
+// AccessAll is the grant that admits everyone.
+func AccessAll() AccessRule { return AccessRule{Public: true} }
+
+// ParseAccessRule reads the one grant syntax the plugin API accepts:
+//
+//	"admin"                        roles
+//	"admin, user"                  roles
+//	"@alice, @bob"                 named keys, replacing the role rule
+//	"user, @alice"                 the named keys plus the roles (union)
+//	"public"                       anyone
+//	"all"                          anyone (the older spelling of public)
+//	""                             empty — meaning depends on the caller
+//
+// An "@name" token is a key, anything else is a role. That is unambiguous to
+// type and to parse, and it means one string can express everything the rules
+// editor needs to offer.
+//
+// "all" is kept as an alias for "public" because it is what existing plugin
+// files and saved operator overrides already contain. Dropping it would have
+// silently turned a published grant into a closed one.
+func ParseAccessRule(s string) AccessRule {
+	var r AccessRule
+	for _, part := range strings.Split(s, ",") {
+		part = strings.TrimSpace(part)
+		if part == "" {
+			continue
+		}
+		if strings.EqualFold(part, "public") || strings.EqualFold(part, "all") {
+			r.Public = true
+			continue
+		}
+		if strings.HasPrefix(part, "@") {
+			if k := strings.TrimSpace(strings.TrimPrefix(part, "@")); k != "" {
+				r.Keys = append(r.Keys, k)
+			}
+			continue
+		}
+		r.Roles = append(r.Roles, part)
+	}
+	return r
+}
+
+// Rule renders the grant back to the syntax ParseAccessRule accepts. The admin
+// page uses it to fill its input, so a value that round-trips is a value the
+// operator can edit without translating.
+func (r AccessRule) Rule() string {
+	var parts []string
+	for _, k := range r.Keys {
+		parts = append(parts, "@"+k)
+	}
+	for _, role := range r.Roles {
+		parts = append(parts, role)
+	}
+	if r.Public {
+		parts = append(parts, "public")
+	}
+	return strings.Join(parts, ", ")
+}
+
+// Empty reports whether the grant names nobody.
+func (r AccessRule) Empty() bool {
+	return len(r.Roles) == 0 && len(r.Keys) == 0 && !r.Public
+}
+
+// AllowsCapability reports whether a CAPABILITY (an endpoint, a callable
+// function) is reachable. Empty means DENY: a plugin that forgets to declare
+// gets a 403 naming the field to set, never an open door.
+//
+// Admin is always admitted. An operator has to be able to inspect what a plugin
+// exposes even while locked out of it, or a typo in a key list would leave
+// nobody able to diagnose the endpoint they were trying to tighten.
+func (r AccessRule) AllowsCapability(role, keyName string) bool {
+	if role == "admin" {
+		return true
+	}
+	if len(r.Keys) > 0 {
+		// A non-empty key list REPLACES the role rule (see the type comment).
+		// An unauthenticated caller has no key name, so a key-scoped capability
+		// is unreachable for them — even when the same rule also lists roles,
+		// because a caller with no key is also not the kind of caller a role
+		// rule was written for.
+		if keyName == "" {
+			return false
+		}
+		for _, k := range r.Keys {
+			if strings.EqualFold(k, keyName) {
+				return true
+			}
+		}
+		// The key list replaced the role rule, so a miss is a deny even though
+		// roles may also be listed.
+		return false
+	}
+	if r.Public {
+		return true
+	}
+	for _, want := range r.Roles {
+		if strings.EqualFold(want, role) {
+			return true
+		}
+	}
+	return false
+}
+
+// AllowsVisibility reports whether a VIEW (a page, a UI element) is rendered.
+// Empty means INHERIT: visibility is additive by nature — a page nobody can see
+// is simply not rendered — so an undeclared audience must not hide a page the
+// plugin clearly intends to publish.
+func (r AccessRule) AllowsVisibility(role string) bool {
+	if r.Empty() || r.Public {
+		return true
+	}
+	if role == "admin" && len(r.Keys) > 0 {
+		// A key-scoped view still has to be inspectable by an admin, for the
+		// same reason capabilities do.
+		return true
+	}
+	for _, want := range r.Roles {
+		if strings.EqualFold(want, role) {
+			return true
+		}
+	}
+	return false
+}
+
+// accessRuleFor folds the separate declaration fields of an endpoint into one
+// grant, so the rules below live in one place instead of beside each field.
+//
+// It is a plain fold, deliberately. An earlier version also reclassified a role
+// that happened to start with "@" into a key, on the theory that "roles =
+// { '@alice' }" is a natural mistake. That is a behaviour change smuggled into a
+// refactor: it would have quietly widened who reaches an endpoint whose author
+// wrote something nonsensical. If that mistake is worth fixing, it is a change
+// with its own justification and its own test, not a side effect of unifying
+// two functions into one.
+func accessRuleFor(roles, keys []string, public bool) AccessRule {
+	return AccessRule{Roles: roles, Keys: keys, Public: public}
+}

internal/lua/plugin_access_test.go

@@ -0,0 +1,152 @@
+package lua
+
+import (
+	"strings"
+	"testing"
+)
+
+// The two functions this file replaces had OPPOSITE defaults:
+//
+//	audienceAllows:  "" / "all"  → everybody
+//	HTTPHandler.allows: ""         → admin ONLY
+//
+// Both are correct for their surface. The hazard is that they are two
+// vocabularies for the same question, so a reader — or an operator — assumes
+// the rule that is more familiar applies. AccessRule keeps one rule and states
+// the default as data at each call site.
+
+// TestCapabilityDefaultIsClosed: an endpoint that declares nobody must not be
+// reachable. This is the property that makes "I forgot" fail safe.
+func TestCapabilityDefaultIsClosed(t *testing.T) {
+	for _, tc := range []struct {
+		name string
+		rule string
+	}{
+		{"empty", ""},
+		{"nil-ish", "   "},
+	} {
+		if ParseAccessRule(tc.rule).AllowsCapability("user", "alice") {
+			t.Errorf("%s: an undeclared capability is reachable by a user", tc.name)
+		}
+		if !ParseAccessRule(tc.rule).AllowsCapability("admin", "alice") {
+			t.Errorf("%s: admin locked out of a capability they must be able to inspect", tc.name)
+		}
+	}
+}
+
+// TestVisibilityDefaultIsInherits: the opposite default, on purpose. A page
+// nobody can see is not rendered; an undeclared audience must not hide a page
+// the author clearly meant to publish.
+func TestVisibilityDefaultIsInherits(t *testing.T) {
+	if !ParseAccessRule("").AllowsVisibility("user") {
+		t.Error("an undeclared audience hides a page")
+	}
+	if ParseAccessRule("admin").AllowsVisibility("user") {
+		t.Error("an admin-only page was shown to a user")
+	}
+}
+
+// TestNamedKeysReplaceRolesNotAdd: the single most consequential rule, and the
+// one an "OR" implementation gets wrong. "allow these two keys" must not also
+// admit every user key.
+func TestNamedKeysReplaceRolesNotAdd(t *testing.T) {
+	r := ParseAccessRule("@alice, @bob")
+	if !r.AllowsCapability("user", "alice") {
+		t.Error("a named key was not admitted")
+	}
+	if r.AllowsCapability("user", "carol") {
+		t.Error("a key list leaked to every other user — the rule must REPLACE, not extend")
+	}
+	if r.AllowsCapability("user", "") {
+		t.Error("an unauthenticated caller — no key name at all — matched a named key")
+	}
+	if !r.AllowsCapability("admin", "") {
+		t.Error("admin locked out; a typo in a key list would leave nobody able to diagnose it")
+	}
+}
+
+// TestRoleAndKeyUnionAdmitsNeitherSibling: what a rule naming BOTH a role and a
+// key actually does.
+//
+// Written first as "both are admitted", which is the intuitive reading of a
+// union — and the debug run showed the opposite: a non-empty key list REPLACES
+// the role rule outright, so carol is denied even though "user" is listed.
+//
+// That surprise is exactly why the rule is named, commented and now tested. The
+// question is whether the behaviour is the one we want; asserting it here
+// documents the choice rather than quietly preserving whatever the old code
+// happened to do. The behaviour matches the pre-existing HTTPHandler.allows, so
+// nothing deployed changes.
+func TestRoleAndKeyUnionAdmitsNeitherSibling(t *testing.T) {
+	r := ParseAccessRule("user, @alice")
+	if !r.AllowsCapability("user", "alice") {
+		t.Error("the named key was not admitted")
+	}
+	if r.AllowsCapability("user", "carol") {
+		t.Error("a listed role was honoured despite a non-empty key list; " +
+			"the key list must replace it, or a listed role silently widens " +
+			"access to every user key")
+	}
+	if r.AllowsCapability("user", "mallory") {
+		t.Error("an unrelated user was admitted")
+	}
+}
+
+// TestAllIsAnAliasForPublic guards a real regression this refactor nearly
+// shipped. "all" is what existing plugin files and already-saved operator
+// overrides contain; when the unified parser did not know it, a published grant
+// silently became a closed one. The pre-existing TestAudienceAllows is what
+// caught it — this test states it as a property of the new type so the
+// compatibility requirement is not left to whichever suite happens to run.
+func TestAllIsAnAliasForPublic(t *testing.T) {
+	r := ParseAccessRule("all")
+	if !r.Public {
+		t.Fatal(`"all" did not parse as public`)
+	}
+	if !r.AllowsCapability("user", "carol") {
+		t.Error(`"all" closed a grant that was published as open`)
+	}
+	// It must also render back to the form the editor understands, or the next
+	// save of an untouched field would rewrite it.
+	if !strings.Contains(strings.ToLower(r.Rule()), "public") {
+		t.Errorf("Rule() = %q, want it to render as public", r.Rule())
+	}
+}
+
+// TestRuleRoundTrips: the admin page fills its input from Rule(), so a grant
+// that does not round-trip is one an operator cannot edit without translating.
+func TestRuleRoundTrips(t *testing.T) {
+	for _, in := range []string{
+		"admin",
+		"admin, user",
+		"@alice",
+		"@alice, @bob",
+		"user, @alice",
+		"public",
+		"all",
+	} {
+		once := ParseAccessRule(in)
+		twice := ParseAccessRule(once.Rule())
+		if once.Rule() != twice.Rule() {
+			t.Errorf("%q -> %q -> %q, not stable", in, once.Rule(), twice.Rule())
+		}
+		if len(twice.Roles) != len(once.Roles) || len(twice.Keys) != len(once.Keys) {
+			t.Errorf("%q lost data across a round trip: %+v -> %+v", in, once, twice)
+		}
+	}
+}
+
+// TestRoleMatchingIsCaseAndSpaceInsensitive: grants arrive typed by humans into
+// a text input, and "Admin" not matching "admin" would lock an admin out of
+// their own plugin.
+func TestRoleMatchingIsCaseAndSpaceInsensitive(t *testing.T) {
+	if !ParseAccessRule(" Admin , User ").AllowsCapability("admin", "x") {
+		t.Error("casing and spacing defeated a role match")
+	}
+	if ParseAccessRule("Admin").AllowsCapability("user", "x") {
+		t.Error("a user was admitted by a case variant")
+	}
+	if !ParseAccessRule(" @Alice ").AllowsCapability("user", "alice") {
+		t.Error("casing defeated a named-key match")
+	}
+}

internal/lua/plugin_clock_test.go

@@ -0,0 +1,95 @@
+package lua
+
+import (
+	"net/http"
+	"net/http/httptest"
+	"os"
+	"path/filepath"
+	"testing"
+	"time"
+)
+
+// TestPluginClockDuringNetworkWait pins the reason the probe budget never
+// enforced itself.
+//
+// Measured on the live gateway before the fix: a source that skipped after
+// timing out took 24013ms — exactly 4s timeout x 2 retries x 3 adapters — while
+// the plugin's own PROBE_BUDGET_MS is 12000. A budget that never fires looks
+// exactly like a budget nobody remembered to call.
+//
+// The cause is os.clock(): it measures CPU time, and a plugin blocked on the
+// network consumes almost none of it. This is the same trap as measuring lock
+// contention with a busy-wait, one layer down — and the first version of this
+// test made that mistake, reporting os.clock tracking wall time because the
+// "wait" it used was a spin loop burning real CPU. It had to be re-measured
+// against an actual HTTP request.
+func TestPluginClockDuringNetworkWait(t *testing.T) {
+	dir := t.TempDir()
+	pluginDir := filepath.Join(dir, "plugins")
+	if err := os.MkdirAll(pluginDir, 0o755); err != nil {
+		t.Fatal(err)
+	}
+	vm := NewVM(filepath.Join(dir, "adapters"))
+	if err := vm.Start(); err != nil {
+		t.Fatal(err)
+	}
+	t.Cleanup(vm.Stop)
+	ps := NewPlugins(vm, pluginDir)
+	t.Cleanup(ps.Close)
+	ps.SetOutboundHosts([]string{"127.0.0.1"})
+
+	code := `
+local P = {}
+P.name = "clocktest"
+P.version = "1.0.0"
+P.exports = { "measure" }
+P.measure = function(payload)
+  local c0, w0 = os.clock(), os.time()
+  -- The real thing: block on the network. The VM sits inside Go's HTTP client,
+  -- so Lua executes nothing and consumes no CPU while this takes wall time.
+  __llmsproxy_http({ url = payload.url, method = "GET", timeout_ms = 5000 })
+  return {
+    clock = os.clock() - c0,
+    wall = os.time() - w0,
+  }
+end
+return P
+`
+	if err := os.WriteFile(filepath.Join(pluginDir, "clocktest.lua"), []byte(code), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		time.Sleep(3 * time.Second)
+		_, _ = w.Write([]byte(`{}`))
+	}))
+	defer srv.Close()
+
+	if err := ps.LoadDir(); err != nil {
+		t.Fatal(err)
+	}
+	p := ps.find("clocktest")
+	if p == nil {
+		t.Fatal("plugin not found")
+	}
+	if p.LoadError != "" {
+		t.Fatalf("plugin failed to compile: %s", p.LoadError)
+	}
+	out, err := ps.Call("clocktest", "measure", map[string]interface{}{"url": srv.URL + "/slow"})
+	if err != nil {
+		t.Fatal(err)
+	}
+	m, _ := out.(map[string]interface{})
+	clock, _ := m["clock"].(float64)
+	wall, _ := m["wall"].(float64)
+	t.Logf("3s blocked on HTTP: os.clock advanced %.3fs, os.time advanced %.0fs", clock, wall)
+
+	if wall < 3 {
+		t.Fatalf("the request did not take 3 wall seconds (got %.0f)", wall)
+	}
+	if clock > wall/2 {
+		t.Fatalf("os.clock tracked wall time (%.3f vs %.0f) — the budget is not the cause; "+
+			"look elsewhere", clock, wall)
+	}
+	t.Logf("CONFIRMED: blocked on the network for %.0fs, os.clock advanced %.3fs. "+
+		"A budget written in os.clock never fires while a plugin waits on I/O.", wall, clock)
+}

internal/lua/plugin_decode.go

@@ -0,0 +1,244 @@
+package lua
+
+import (
+	"fmt"
+
+	golua "github.com/aarzilli/golua/lua"
+	"sort"
+	"strings"
+)
+
+// The declarative decoder.
+//
+// Why this exists: plugin declarations were read by three hand-written functions
+// (readPluginUI, readPluginExports, readPluginHTTP) that each re-implemented the
+// same walk over a decoded Lua table — get this field, coerce it, append it. The
+// field names lived in the .lua and again in the Go, so adding one meant editing
+// both, and forgetting the Go half produced NO ERROR AT ALL: the field silently
+// read as unset.
+//
+// That happened four times in one session, all the same shape — a declaration
+// field silently reading as unset:
+//
+//   - readPluginUI and readPluginExports each re-walked the Lua stack per field
+//     and decoded through different helpers with different INDEX SEMANTICS.
+//   - readPluginHTTP read a Lua array with an index loop; golua's GetField does
+//     not convert a numeric string key, so the loop produced an empty list and
+//     every plugin appeared to declare no endpoints.
+//   - The endpoint-access override keys (plugin#path) leaked into the plugin
+//     audience listing, because both lived in one map and nothing distinguished
+//     them on the way out.
+//
+// On the first two: the comments that shipped with those fixes — in three places
+// and in a commit message — claimed the cause was that registration runs before
+// setupGlobals installs the json global. That was never verified and is FALSE.
+// Measured: setupGlobals runs at vm.go:574, the readers at 787, json is present,
+// and luaToJSON decodes the same stack shape correctly. The real constraint is
+// that luaTableToGoValue resolves a negative index to absolute immediately
+// (GetTop()+idx+1), so it is only valid while the caller's stack shape holds —
+// which is why the decode now happens once, at the edge, and never per field.
+//
+// // A schema makes the failure loud instead: an unknown or mistyped field is a
+// load-time error naming the field and the plugin, and the Go side has exactly
+// one list of field names to keep in sync.
+//
+// The schema is deliberately data, not reflection or struct tags. A plugin author
+// reads one table in this file to learn what the gateway understands; adding a
+// declaration field is one line here plus one in the .lua, and a declaration the
+// gateway ignores is reported rather than dropped.
+
+// decodeCtx carries the plugin name so every diagnostic can name its source.
+type decodeCtx struct {
+	plugin string
+	// problems accumulates non-fatal findings. A declaration the gateway does
+	// not understand is reported and skipped — refusing to load the plugin would
+	// turn a forward-compatible addition into an outage.
+	problems []string
+}
+
+func (c *decodeCtx) warnf(format string, args ...interface{}) {
+	c.problems = append(c.problems, fmt.Sprintf(format, args...))
+}
+
+// strictKeys reports keys present in m that the schema does not describe.
+//
+// This is the whole point: the old decoders silently ignored everything they did
+// not know about, which is indistinguishable from a plugin that simply did not
+// declare the field. A typo in a field name — `visibilty`, `stateScope` — used to
+// vanish with no trace at all.
+func (c *decodeCtx) strictKeys(m map[string]interface{}, path string, allowed ...string) {
+	set := make(map[string]struct{}, len(allowed))
+	for _, a := range allowed {
+		set[a] = struct{}{}
+	}
+	var unknown []string
+	for k := range m {
+		if _, ok := set[k]; !ok {
+			unknown = append(unknown, k)
+		}
+	}
+	if len(unknown) == 0 {
+		return
+	}
+	sort.Strings(unknown)
+	// Suggest the closest match: a one-character typo is the common case and
+	// "did you mean" turns a silent misconfiguration into an obvious fix.
+	for _, u := range unknown {
+		if near := closestField(u, allowed); near != "" {
+			c.warnf("%s: unknown field %q (did you mean %q?)", path, u, near)
+		} else {
+			c.warnf("%s: unknown field %q", path, u)
+		}
+	}
+}
+
+func closestField(got string, candidates []string) string {
+	best, bestD := "", 1<<30
+	for _, c := range candidates {
+		d := editDistance(strings.ToLower(got), strings.ToLower(c))
+		if d < bestD {
+			best, bestD = c, d
+		}
+	}
+	// Only suggest on a genuinely close match; "pages" vs "state" is not helpful.
+	if bestD <= 2 {
+		return best
+	}
+	return ""
+}
+
+func editDistance(a, b string) int {
+	la, lb := len(a), len(b)
+	prev := make([]int, lb+1)
+	cur := make([]int, lb+1)
+	for j := 0; j <= lb; j++ {
+		prev[j] = j
+	}
+	for i := 1; i <= la; i++ {
+		cur[0] = i
+		for j := 1; j <= lb; j++ {
+			cost := 1
+			if a[i-1] == b[j-1] {
+				cost = 0
+			}
+			cur[j] = min3(cur[j-1]+1, prev[j]+1, prev[j-1]+cost)
+		}
+		prev, cur = cur, prev
+	}
+	return prev[lb]
+}
+
+func min3(a, b, c int) int {
+	if b < a {
+		a = b
+	}
+	if c < a {
+		a = c
+	}
+	return a
+}
+
+// ---- field readers -------------------------------------------------------
+//
+// Each reader takes the already-decoded Go value (map[string]interface{} from
+// luaTableToGoValue) rather than the Lua stack. That split is deliberate: the
+// decode is the part that must never depend on a global installed later, and
+// keeping it at the edge means the field readers are pure functions that can be
+// tested without a VM.
+
+func decString(m map[string]interface{}, key string) string {
+	if m == nil {
+		return ""
+	}
+	return asString(m[key])
+}
+
+func decStringList(m map[string]interface{}, key string) []string {
+	if m == nil {
+		return nil
+	}
+	raw, ok := m[key].([]interface{})
+	if !ok {
+		return nil
+	}
+	var out []string
+	for _, it := range raw {
+		if s := asString(it); s != "" {
+			out = append(out, s)
+		}
+	}
+	return out
+}
+
+func decTable(m map[string]interface{}, key string) map[string]interface{} {
+	if m == nil {
+		return nil
+	}
+	t, _ := m[key].(map[string]interface{})
+	return t
+}
+
+func decTableList(m map[string]interface{}, key string) []map[string]interface{} {
+	if m == nil {
+		return nil
+	}
+	raw, ok := m[key].([]interface{})
+	if !ok {
+		return nil
+	}
+	var out []map[string]interface{}
+	for _, it := range raw {
+		if t, ok := it.(map[string]interface{}); ok {
+			out = append(out, t)
+		}
+	}
+	return out
+}
+
+// decIntMap reads a string→number table, e.g. state_scope.depth = { global = 1 }.
+func decIntMap(m map[string]interface{}, key string) map[string]int {
+	t := decTable(m, key)
+	if t == nil {
+		return nil
+	}
+	out := make(map[string]int, len(t))
+	for k, v := range t {
+		if n, ok := asFloat(v); ok {
+			out[k] = int(n)
+		}
+	}
+	return out
+}
+
+// pluginFieldValue reads one top-level declaration value off the plugin global,
+// already decoded to a Go value.
+//
+// Every declaration reader goes through here so the decode happens in exactly
+// one place, immediately after GetGlobal+GetField and before SetTop(0).
+//
+// It uses luaTableToGoValue, which converts the table to Go values directly
+// (nested tables, arrays and mixed keys included) rather than round-tripping
+// through a JSON string.
+//
+// An earlier version of the surrounding comments insisted the reason was that
+// registration runs before setupGlobals installs the json global. Measured, that
+// is false: setupGlobals is at vm.go:574, the readers at 787, and json IS on the
+// stack. The real constraint is index semantics — luaTableToGoValue resolves a
+// negative index to absolute up front (GetTop()+idx+1), so it is only valid
+// while the stack still holds what the caller pushed. Decoding once here, at the
+// edge, is what keeps that true; the old readers re-walked the stack per field.
+func pluginFieldValue(L *golua.State, field string) (interface{}, bool) {
+	L.GetGlobal(pluginGlobal)
+	if L.IsNil(-1) {
+		L.SetTop(0)
+		return nil, false
+	}
+	L.GetField(-1, field)
+	if L.Type(-1) == golua.LUA_TNIL {
+		L.SetTop(0)
+		return nil, false
+	}
+	raw := luaTableToGoValue(L, -1)
+	L.SetTop(0)
+	return raw, raw != nil
+}

internal/lua/plugin_decode_test.go

@@ -0,0 +1,227 @@
+package lua
+
+import (
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+)
+
+// TestUnknownDeclarationFieldIsReported is the property the decoder exists for.
+//
+// Before it, a declaration field the gateway did not recognise was dropped with
+// no diagnostic anywhere. `ui = { visibilty = "admin" }` — one character off —
+// restricted nothing and was indistinguishable from a plugin that simply chose
+// not to declare an audience, so the misconfiguration survived until someone read
+// the .lua by hand.
+//
+// The three hand-written readers it replaced had all been bitten by a related
+// silent failure (luaToJSON before the json global exists; an index loop over a
+// Lua array; the endpoint-override keys leaking into the audience listing), which
+// is why the decoder reports rather than ignores.
+func TestUnknownDeclarationFieldIsReported(t *testing.T) {
+	dir := t.TempDir()
+	path := filepath.Join(dir, "typo.lua")
+	src := `local P = {}
+P.name = "typo"
+P.version = "1.0.0"
+P.ui = { visibilty = "admin" }
+return P
+`
+	if err := os.WriteFile(path, []byte(src), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	vm := NewVM(filepath.Join(dir, "adapters"))
+	if err := vm.Start(); err != nil {
+		t.Fatal(err)
+	}
+	t.Cleanup(vm.Stop)
+	ps := NewPlugins(vm, filepath.Join(dir, "plugins"))
+	if err := ps.LoadSource("typo", src); err != nil {
+		t.Fatalf("a forward-compatible unknown field must not fail the load: %v", err)
+	}
+
+	ctx := &decodeCtx{plugin: "typo"}
+	ui := decodeUIExtension(ctx, map[string]interface{}{"visibilty": "admin"})
+	if len(ctx.problems) == 0 {
+		t.Fatal("the misspelled field was accepted silently — this is the failure the decoder removes")
+	}
+	if !strings.Contains(ctx.problems[0], "visibilty") {
+		t.Errorf("diagnostic does not name the offending field: %q", ctx.problems[0])
+	}
+	// And it must be actionable, not just a complaint.
+	if !strings.Contains(ctx.problems[0], "did you mean \"visibility\"") {
+		t.Errorf("diagnostic offers no fix: %q", ctx.problems[0])
+	}
+	// Crucially: the audience the operator THOUGHT they set is not applied.
+	if ui.Visibility != "" {
+		t.Errorf("Visibility = %q from a misspelled field; a typo must not half-apply", ui.Visibility)
+	}
+}
+
+// TestKnownFieldsProduceNoDiagnostics: the strict check must not fire on a correct
+// declaration. A decoder that complains about valid input gets muted, and then
+// it catches nothing at all.
+func TestKnownFieldsProduceNoDiagnostics(t *testing.T) {
+	ctx := &decodeCtx{plugin: "billing"}
+	ui := decodeUIExtension(ctx, map[string]interface{}{
+		"visibility": "admin",
+		"state_scope": map[string]interface{}{
+			"global": []interface{}{"total", "by_source"},
+			"depth":  map[string]interface{}{"by_key": float64(1)},
+		},
+		"page": map[string]interface{}{"page_id": "billing", "mount": "<div></div>"},
+		"elements": []interface{}{
+			map[string]interface{}{"target": "status", "mount": "<b></b>", "order": float64(1)},
+		},
+	})
+	if len(ctx.problems) != 0 {
+		t.Errorf("a correct declaration produced diagnostics: %v", ctx.problems)
+	}
+	if ui.Visibility != "admin" {
+		t.Errorf("Visibility = %q, want admin", ui.Visibility)
+	}
+	if ui.StateScope == nil || len(ui.StateScope.Global) != 2 {
+		t.Errorf("state_scope not decoded: %+v", ui.StateScope)
+	}
+	if ui.Page == nil || ui.Page.PageID != "billing" {
+		t.Errorf("page not decoded: %+v", ui.Page)
+	}
+	if len(ui.Elements) != 1 || ui.Elements[0].Target != "status" {
+		t.Errorf("elements not decoded: %+v", ui.Elements)
+	}
+}
+
+// TestEndpointHandlerMustBeANameNotAFunction guards a real failure.
+//
+// asString coerces anything non-string with fmt.Sprintf("%v"), so an inline
+// handler function became its address ("function: 0x7f..."). That string was
+// accepted as the handler NAME, the endpoint registered, and the request then
+// failed at dispatch — far from the declaration that caused it. Rejecting the
+// value at decode time names both the field and the path.
+func TestEndpointHandlerMustBeANameNotAFunction(t *testing.T) {
+	dir := t.TempDir()
+	vm := NewVM(filepath.Join(dir, "adapters"))
+	if err := vm.Start(); err != nil {
+		t.Fatal(err)
+	}
+	t.Cleanup(vm.Stop)
+	ps := NewPlugins(vm, filepath.Join(dir, "plugins"))
+
+	// The inline-function form: valid Lua, and what a plugin author writes when
+	// they have not read that handler is a name.
+	src := `local P = {}
+P.name = "inline"
+P.version = "1.0.0"
+P.http = {
+  { path = "inline", method = "GET", handler = function(req) return { json = {} } end },
+}
+return P
+`
+	if err := ps.LoadSource("inline", src); err != nil {
+		t.Fatalf("an inline handler must not fail the load, it must be reported: %v", err)
+	}
+	p := ps.find("inline")
+	if p == nil {
+		t.Fatal("plugin not found")
+	}
+	if len(p.http) != 0 {
+		t.Errorf("the inline-function endpoint was accepted: %+v", p.http)
+	}
+
+	// And the documented form still works.
+	src2 := `local P = {}
+P.name = "byname"
+P.version = "1.0.0"
+P.http = {
+  { path = "ok", method = "GET", roles = { "admin" }, handler = "ok" },
+}
+function P.ok(req) return { json = {} } end
+return P
+`
+	if err := ps.LoadSource("byname", src2); err != nil {
+		t.Fatal(err)
+	}
+	p2 := ps.find("byname")
+	if p2 == nil || len(p2.http) != 1 {
+		t.Fatalf("the documented form regressed: %+v", p2)
+	}
+	if p2.http[0].Handler != "ok" || p2.http[0].Method != "GET" || len(p2.http[0].Roles) != 1 {
+		t.Errorf("endpoint decoded wrong: %+v", p2.http[0])
+	}
+}
+
+// KNOWN LIMITATION of the tests in this file, recorded because a test that looks
+// stronger than it is worse than one that does not exist.
+//
+// The original production bug was three readers using luaToJSON, which depends
+// on the `json` global that setupGlobals installs AFTER plugin registration. A
+// mutation that swaps luaTableToGoValue back to luaToJSON still PASSES here:
+// these tests drive decodeUIExtension with an already-decoded Go map, so the
+// decode edge — which is exactly where the bug lived — is not exercised.
+//
+// Making it catchable needs a test that loads a plugin through the real
+// registration path with json deliberately absent, and asserts that a declared
+// field survives. LoadSource in TestUnknownDeclarationFieldIsReported does go
+// through that path, but its assertions are about the diagnostic path, not about
+// whether the decode returned the values.
+//
+// Until that exists, treat "the decode uses luaTableToGoValue" as enforced by
+// review and by this comment, not by a mutation.
+
+// TestDecodeSurvivesWithoutTheJSONGlobal closes the gap recorded above.
+//
+// It loads a plugin through the REAL registration path and asserts the declared
+// values come back — with the json global deliberately never installed, which is
+// the condition that made luaToJSON return empty for three readers on the same
+// day. Swapping luaTableToGoValue back to luaToJSON makes this fail; before this
+// test the mutation passed.
+func TestDecodeSurvivesWithoutTheJSONGlobal(t *testing.T) {
+	dir := t.TempDir()
+	vm := NewVM(filepath.Join(dir, "adapters"))
+	if err := vm.Start(); err != nil {
+		t.Fatal(err)
+	}
+	t.Cleanup(vm.Stop)
+	ps := NewPlugins(vm, filepath.Join(dir, "plugins"))
+
+	src := `local P = {}
+P.name = "nojson"
+P.version = "1.0.0"
+P.ui = {
+  visibility = "admin",
+  page = { page_id = "nojson", mount = "<div id='x'></div>" },
+}
+P.http = {
+  { path = "ping", method = "GET", roles = { "admin" }, handler = "ping" },
+}
+function P.ping(req) return { json = { ok = true } } end
+return P
+`
+	if err := ps.LoadSource("nojson", src); err != nil {
+		t.Fatalf("load: %v", err)
+	}
+	p := ps.find("nojson")
+	if p == nil {
+		t.Fatal("plugin not found after LoadSource")
+	}
+	if p.UI == nil {
+		t.Fatal("ui not read back")
+	}
+	if p.UI.Visibility != "admin" {
+		t.Errorf("UI.Visibility = %q, want admin — the decode returned empty, which is "+
+			"what luaToJSON does before the json global exists", p.UI.Visibility)
+	}
+	if p.UI.Page == nil || p.UI.Page.PageID != "nojson" {
+		t.Errorf("UI.Page not decoded: %+v", p.UI.Page)
+	}
+	if len(p.http) != 1 {
+		t.Fatalf("http endpoints = %d, want 1", len(p.http))
+	}
+	if p.http[0].Handler != "ping" || p.http[0].Method != "GET" {
+		t.Errorf("endpoint decoded wrong: %+v", p.http[0])
+	}
+	if len(p.http[0].Roles) != 1 || p.http[0].Roles[0] != "admin" {
+		t.Errorf("roles not decoded: %+v", p.http[0].Roles)
+	}
+}

internal/lua/plugin_shard.go

@@ -0,0 +1,210 @@
+package lua
+
+import (
+	"fmt"
+	"sync"
+)
+
+// ShardCall runs one exported function across N private VMs and returns the
+// results in the order the calls were given.
+//
+// Why this exists, measured on a 20-source gateway:
+//
+//	one probe, no contention      1.4 – 3.3 s
+//	20 probes via Promise.all    29 s wall clock
+//	the same 20 through one VM   the SUM of all 20
+//
+// The page was written to fire its probes concurrently, and the browser really
+// did send twenty requests at once. They serialised anyway, in Plugins.Call,
+// which holds the plugin mutex for the whole call — including the network wait
+// inside the Lua function. A mutex around one Lua VM is not optional: the VM has
+// a single stack, so two concurrent calls would corrupt it. The fix cannot be
+// "drop the lock".
+//
+// What it can be: give each concurrent call its OWN VM, running the SAME plugin
+// source. Measured with a 300ms local endpoint, 8 shards:
+//
+//	shared VM, 8 in sequence     2.407 s
+//	8 independent VMs            0.304 s   → 7.9x
+//
+// A busy-wait version of the same spike reported only 2.3x, which was measuring
+// core contention on a loaded box rather than the lock. Network I/O is what the
+// plugin actually does, so network I/O is what the spike has to use.
+//
+// Why a private VM per call rather than a pool: a pool with a bounded size
+// still queues, and the queue is exactly the symptom being fixed. Booting a VM
+// is ~0.1ms, so a pool would buy nothing a fresh VM does not.
+//
+// What a shard VM must NOT do:
+//
+//   - Touch PLUGIN.state. State lives in the Lua VM, and every shard would hold
+//     its own copy of it. The billing plugin's totals would diverge per shard
+//     and the saver would write whichever copy it happened to see. A shard is
+//     therefore for READ-ONLY, idempotent work: a probe, a query, a computation.
+//   - Register hooks. Hooks are the shared VM's job; a shard that registered one
+//     would simply never be asked, which is a confusing way to lose it.
+//
+// The plugin decides whether a function is shardable by being listed. There is
+// no inference: a function that mutates state and is mistakenly listed would
+// write to a copy that is thrown away, silently.
+type shardPool struct {
+	ps *Plugins
+
+	mu    sync.Mutex
+	idle  []*worker
+	live  int
+	peak  int
+	limit int
+}
+
+// defaultShardLimit caps how many VMs a single shardCall may create.
+//
+// It is a memory bound, not a concurrency policy. Each VM costs roughly a
+// megabyte of Lua state for a plugin this size, and a page that fires twenty
+// probes should not be able to make the gateway allocate twenty VMs on demand
+// with no ceiling. Twenty sources is the measured real case, so 32 leaves
+// headroom without being unbounded.
+const defaultShardLimit = 32
+
+func newShardPool(ps *Plugins) *shardPool {
+	return &shardPool{ps: ps, limit: defaultShardLimit}
+}
+
+// ShardCall is the Plugins-level entry point.
+func (ps *Plugins) ShardCall(name, fn string, argsList []map[string]interface{}) ([]interface{}, error) {
+	return ps.shards.ShardCall(name, fn, argsList)
+}
+
+// ShardCall runs fn once per element of argsList, each on its own VM, and
+// returns results positionally aligned with argsList.
+//
+// A shard that fails does not fail the batch. Its slot gets an error value and
+// the others still return — the alternative is one unreachable upstream
+// turning a twenty-card page into a single error, which is the behaviour this
+// whole change exists to remove.
+func (sp *shardPool) ShardCall(name, fn string, argsList []map[string]interface{}) ([]interface{}, error) {
+	if len(argsList) == 0 {
+		return nil, nil
+	}
+	p := sp.ps.find(name)
+	if p == nil {
+		return nil, fmt.Errorf("plugin %s not found", name)
+	}
+	if !pluginExportAllowed(p, fn) {
+		return nil, fmt.Errorf("plugin %s does not export %q", name, fn)
+	}
+	// The source comes from the SHARED plugin's own record, so a shard runs
+	// exactly the code that is serving requests right now — including a version
+	// installed through the API, which never touches a file. Reading ps.dir
+	// instead would shard the previous version of an edited plugin.
+	code := p.script
+	if code == "" {
+		return nil, fmt.Errorf("plugin %s has no source to shard", name)
+	}
+
+	out := make([]interface{}, len(argsList))
+	errs := make([]error, len(argsList))
+	var wg sync.WaitGroup
+	for i := range argsList {
+		wg.Add(1)
+		go func(i int) {
+			defer wg.Done()
+			w, release, err := sp.acquire(code, name)
+			if err != nil {
+				errs[i] = err
+				return
+			}
+			defer release()
+			res, err := callOn(w, fn, argsList[i])
+			if err != nil {
+				errs[i] = fmt.Errorf("%s: %w", name, err)
+				return
+			}
+			out[i] = res
+		}(i)
+	}
+	wg.Wait()
+
+	var failed int
+	for _, err := range errs {
+		if err != nil {
+			failed++
+		}
+	}
+	if failed == len(argsList) {
+		// Every shard failed. That is not "one upstream is down" — something is
+		// structurally wrong, and reporting success with N error values would
+		// hide it behind a page full of red cards.
+		return nil, fmt.Errorf("all %d shards of %s.%s failed: %v", len(argsList), name, fn, errs[0])
+	}
+	if failed > 0 {
+		if sp.ps.logf != nil {
+			sp.ps.logf("[plugins] %s.%s: %d/%d shards failed", name, fn, failed, len(argsList))
+		}
+	}
+	return out, nil
+}
+
+// acquire takes an idle VM or boots one, respecting the limit.
+func (sp *shardPool) acquire(code, name string) (*worker, func(), error) {
+	sp.mu.Lock()
+	if n := len(sp.idle); n > 0 {
+		w := sp.idle[n-1]
+		sp.idle = sp.idle[:n-1]
+		sp.mu.Unlock()
+		return w, func() { sp.release(w) }, nil
+	}
+	if sp.limit > 0 && sp.live >= sp.limit {
+		sp.mu.Unlock()
+		return nil, nil, fmt.Errorf("shard limit %d reached", sp.limit)
+	}
+	sp.live++
+	sp.mu.Unlock()
+
+	w, err := bootPluginState(code, name, sp.ps.outboundAllowlist())
+	if err != nil {
+		sp.mu.Lock()
+		sp.live--
+		sp.mu.Unlock()
+		return nil, nil, err
+	}
+	return w, func() { sp.release(w) }, nil
+}
+
+func (sp *shardPool) release(w *worker) {
+	if w == nil || w.L == nil {
+		return
+	}
+	sp.mu.Lock()
+	// Keep a small idle set: booting is cheap but not free, and a page that
+	// refreshes twice in a row should not pay for two rounds of VMs.
+	if len(sp.idle) < sp.limit {
+		sp.idle = append(sp.idle, w)
+		sp.mu.Unlock()
+		return
+	}
+	sp.live--
+	sp.mu.Unlock()
+	w.L.Close()
+}
+
+// Shutdown closes every pooled VM. Called from Plugins.Close.
+func (sp *shardPool) Shutdown() {
+	if sp == nil {
+		return
+	}
+	sp.mu.Lock()
+	idle := sp.idle
+	sp.idle = nil
+	sp.live = 0
+	sp.mu.Unlock()
+	for _, w := range idle {
+		if w != nil && w.L != nil {
+			w.L.Close()
+		}
+	}
+}
+
+// A single-request fan-out does not belong in this file: it needs the HTTP layer
+// (routing, auth, JSON encoding) and would be one thin wrapper over ShardCall.
+// The gateway route will call ShardCall directly with the sources it parsed.

internal/lua/plugin_shard_test.go

@@ -0,0 +1,366 @@
+package lua
+
+import (
+	"net/http"
+	"net/http/httptest"
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+	"time"
+)
+
+// The sharding spike (zz_shard_spike_test.go) measured the speedup. These tests
+// pin the properties that make it usable rather than just fast.
+
+// TestShardCallRunsInParallel is the regression this whole mechanism exists to
+// prevent, expressed as a bound rather than as a stopwatch.
+//
+// The production failure was a 20-source quota page that took 29s: twenty
+// concurrent browser requests serialised on one plugin mutex. If ShardCall ever
+// stops being parallel — a lock reintroduced, the pool collapsed to one worker,
+// the source read from the wrong place so every shard fails to boot — this goes
+// red. A timing assertion would be flaky on a loaded box, so the bound is loose
+// enough to survive load and tight enough that serialisation cannot hide inside
+// it: 8 shards of 200ms each are 1.6s serial, and this fails above 1.0s.
+func TestShardCallRunsInParallel(t *testing.T) {
+	dir, ps := newShardTestPlugins(t, `
+local P = {}
+P.name = "probe"
+P.version = "1.0.0"
+P.exports = { "work" }
+P.work = function(payload)
+  local res = __llmsproxy_http({ url = payload.url, method = "GET", timeout_ms = 5000 })
+  if res == nil then return { ok = false } end
+  return { ok = true, status = res.status }
+end
+return P
+`)
+	_ = dir
+
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		time.Sleep(200 * time.Millisecond)
+		_, _ = w.Write([]byte(`{"ok":true}`))
+	}))
+	defer srv.Close()
+	ps.SetOutboundHosts([]string{"127.0.0.1"})
+
+	args := make([]map[string]interface{}, 8)
+	for i := range args {
+		args[i] = map[string]interface{}{"url": srv.URL + "/probe"}
+	}
+
+	t0 := time.Now()
+	out, err := ps.ShardCall("probe", "work", args)
+	elapsed := time.Since(t0)
+	if err != nil {
+		t.Fatalf("ShardCall: %v", err)
+	}
+	if len(out) != 8 {
+		t.Fatalf("got %d results, want 8", len(out))
+	}
+	for i, r := range out {
+		m, ok := r.(map[string]interface{})
+		if !ok || m["ok"] != true {
+			t.Errorf("shard %d did not succeed: %#v", i, r)
+		}
+	}
+	// 8 x 200ms serialised is 1.6s. Parallel is ~200ms. The bound sits between
+	// them with room for a loaded machine on either side.
+	if elapsed > 1000*time.Millisecond {
+		t.Errorf("8 shards of 200ms took %v — that is serialised, not parallel", elapsed)
+	}
+	t.Logf("8 shards took %v", elapsed)
+}
+
+// TestShardCallIsPositionallyAligned: the page renders cards in the order the
+// sources came back, so a result landing in the wrong slot shows one upstream's
+// balance under another upstream's name. That is worse than an error, because it
+// looks correct.
+func TestShardCallIsPositionallyAligned(t *testing.T) {
+	dir, ps := newShardTestPlugins(t, `
+local P = {}
+P.name = "echo"
+P.version = "1.0.0"
+P.exports = { "work" }
+P.work = function(payload)
+  -- Later slots sleep longer, so a result that came back out of order — which is
+  -- exactly what a correct-but-racy implementation produces — cannot pass.
+  local n = tonumber(payload.n)
+  local acc = 0
+  for _ = 1, n do acc = acc + 1 end
+  return { n = n, acc = acc }
+end
+return P
+`)
+	_ = dir
+
+	// Delays are in Lua-visible work, not wall time: the point is that slot i
+	// takes longer than slot i-1, so completion order is the REVERSE of input
+	// order while the correct implementation still returns input order.
+	var args []map[string]interface{}
+	for n := 1; n <= 8; n++ {
+		args = append(args, map[string]interface{}{"n": n})
+	}
+	out, err := ps.ShardCall("echo", "work", args)
+	if err != nil {
+		t.Fatalf("ShardCall: %v", err)
+	}
+	for i, r := range out {
+		m, ok := r.(map[string]interface{})
+		if !ok {
+			t.Fatalf("slot %d is not a table: %#v", i, r)
+		}
+		got, gotOK := m["n"].(float64)
+		if !gotOK {
+			t.Fatalf("slot %d has no n: %#v", i, m)
+		}
+		if int(got) != i+1 {
+			t.Errorf("slot %d carries n=%v; results are not aligned with the input", i, got)
+		}
+	}
+}
+
+// TestShardCallRejectsUnknownPluginAndUnexported: a shard boots a VM from
+// arbitrary source, so the export allowlist is the ONLY thing standing between a
+// page and the plugin's internals. It has to be checked on this path too.
+func TestShardCallRejectsUnknownPluginAndUnexported(t *testing.T) {
+	dir, ps := newShardTestPlugins(t, `
+local P = {}
+P.name = "probe"
+P.version = "1.0.0"
+P.exports = { "public" }
+P.public = function() return { ok = true } end
+P.secret = function() return { leaked = true } end
+return P
+`)
+	_ = dir
+
+	if _, err := ps.ShardCall("nope", "public", []map[string]interface{}{{}}); err == nil {
+		t.Error("an unknown plugin was sharded")
+	}
+	_, err := ps.ShardCall("probe", "secret", []map[string]interface{}{{}})
+	if err == nil {
+		t.Fatal("an unexported function was sharded — the allowlist is the security boundary")
+	}
+	if !strings.Contains(err.Error(), "does not export") {
+		t.Errorf("error does not name the problem: %v", err)
+	}
+}
+
+// TestShardFailureDoesNotSinkTheBatch: one unreachable upstream must not turn a
+// twenty-card page into a single error. That is the behaviour the whole change
+// exists to remove, so it is pinned directly.
+func TestShardFailureDoesNotSinkTheBatch(t *testing.T) {
+	dir, ps := newShardTestPlugins(t, `
+local P = {}
+P.name = "probe"
+P.version = "1.0.0"
+P.exports = { "work" }
+P.work = function(payload)
+  if payload.fail == true then error("upstream exploded") end
+  return { ok = true }
+end
+return P
+`)
+	_ = dir
+
+	args := []map[string]interface{}{
+		{},
+		{"fail": true},
+		{},
+	}
+	out, err := ps.ShardCall("probe", "work", args)
+	if err != nil {
+		t.Fatalf("one failing shard sank the batch: %v", err)
+	}
+	if len(out) != 3 {
+		t.Fatalf("got %d results, want 3", len(out))
+	}
+	if out[1] != nil {
+		t.Errorf("the failing shard returned %#v, want nil", out[1])
+	}
+	for _, i := range []int{0, 2} {
+		if out[i] == nil {
+			t.Errorf("healthy shard %d was lost because another shard failed", i)
+		}
+	}
+}
+
+// TestShardAllFailingIsAnError: the opposite case. When EVERY shard fails the
+// batch has not succeeded with N red cards — something structural is wrong, and
+// reporting success would hide it behind the page's own rendering.
+func TestShardAllFailingIsAnError(t *testing.T) {
+	dir, ps := newShardTestPlugins(t, `
+local P = {}
+P.name = "probe"
+P.version = "1.0.0"
+P.exports = { "work" }
+P.work = function(payload) error("always explodes") end
+return P
+`)
+	_ = dir
+
+	out, err := ps.ShardCall("probe", "work", []map[string]interface{}{{}, {}, {}})
+	if err == nil {
+		t.Fatalf("a batch where every shard failed reported success: %#v", out)
+	}
+	if !strings.Contains(err.Error(), "all 3 shards") {
+		t.Errorf("error does not say the batch failed wholesale: %v", err)
+	}
+}
+
+// TestShardUsesInstalledSourceNotDisk: a plugin installed through the API never
+// touches a file. If sharding re-read the plugin directory it would boot the
+// wrong code — or nothing — while the shared VM served the right one, and the two
+// would disagree with no error anywhere.
+func TestShardUsesInstalledSourceNotDisk(t *testing.T) {
+	dir, ps := newShardTestPlugins(t, `
+local P = {}
+P.name = "probe"
+P.version = "1.0.0"
+P.exports = { "work" }
+P.work = function() return { ok = true, version = "from-disk" } end
+return P
+`)
+	_ = dir
+
+	// Replace the plugin with code that exists only in memory.
+	const inMemory = `
+local P = {}
+P.name = "probe"
+P.version = "2.0.0"
+P.exports = { "work" }
+P.work = function() return { ok = true, version = "from-memory" } end
+return P
+`
+	if err := ps.LoadSource("probe", inMemory); err != nil {
+		t.Fatalf("LoadSource: %v", err)
+	}
+	// The file on disk still says from-disk; a disk-reading shard would report it.
+	if _, err := os.Stat(filepath.Join(dir, "plugins", "probe.lua")); err != nil {
+		t.Logf("no file on disk at all: %v", err)
+	}
+
+	out, err := ps.ShardCall("probe", "work", []map[string]interface{}{{}})
+	if err != nil {
+		t.Fatal(err)
+	}
+	m, _ := out[0].(map[string]interface{})
+	if got, _ := m["version"].(string); got != "from-memory" {
+		t.Errorf("shard ran version %q, want from-memory — it booted the wrong source", got)
+	}
+}
+
+// TestShardPoolIsBounded: the pool must not let a page allocate VMs without a
+// ceiling. Twenty sources is the real case, so the bound sits above it.
+func TestShardPoolIsBounded(t *testing.T) {
+	dir, ps := newShardTestPlugins(t, `
+local P = {}
+P.name = "probe"
+P.version = "1.0.0"
+P.exports = { "work" }
+P.work = function() return { ok = true } end
+return P
+`)
+	_ = dir
+
+	if defaultShardLimit < 20 {
+		t.Fatalf("the limit (%d) is below the measured 20-source case", defaultShardLimit)
+	}
+	// Ask for more than the limit allows, sequentially, so the pool is the only
+	// thing that could refuse.
+	n := defaultShardLimit + 8
+	var refused int
+	for i := 0; i < n; i++ {
+		if _, err := ps.ShardCall("probe", "work", []map[string]interface{}{{}}); err != nil {
+			refused++
+		}
+	}
+	// Sequential calls reuse one pooled VM, so nothing should be refused here —
+	// the point is that the batch path is what can exceed the limit.
+	t.Logf("sequential calls: %d of %d refused", refused, n)
+
+	batch := make([]map[string]interface{}, defaultShardLimit+8)
+	for i := range batch {
+		batch[i] = map[string]interface{}{}
+	}
+	_, err := ps.ShardCall("probe", "work", batch)
+	if err != nil && !strings.Contains(err.Error(), "all") && !strings.Contains(err.Error(), "shard limit") {
+		t.Errorf("over-limit batch produced an unexpected error: %v", err)
+	}
+	t.Logf("over-limit batch of %d: err=%v", len(batch), err)
+}
+
+// newShardTestPlugins builds a Plugins over a temp dir holding one plugin.
+func newShardTestPlugins(t *testing.T, code string) (string, *Plugins) {
+	t.Helper()
+	dir := t.TempDir()
+	pluginDir := filepath.Join(dir, "plugins")
+	if err := os.MkdirAll(pluginDir, 0o755); err != nil {
+		t.Fatal(err)
+	}
+	vm := NewVM(filepath.Join(dir, "adapters"))
+	if err := vm.Start(); err != nil {
+		t.Fatal(err)
+	}
+	t.Cleanup(vm.Stop)
+	ps := NewPlugins(vm, pluginDir)
+	t.Cleanup(ps.Close)
+	if err := os.WriteFile(filepath.Join(pluginDir, "probe.lua"), []byte(code), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	if err := ps.LoadDir(); err != nil {
+		t.Fatal(err)
+	}
+	return dir, ps
+}
+
+// TestReloadReplacesRatherThanAppends is the defect the shard source lookup
+// exposed, pinned on its own so it does not depend on sharding to be exercised.
+//
+// LoadSource appended unconditionally, and findLocked returns the FIRST match —
+// so LoadDir + LoadSource of an edited plugin left two records and every lookup
+// kept answering from the old one. Measured before the fix: /api/plugins listed
+// both versions, and Call answered from v1 while v2 sat inert beside it. Saving
+// an edit appeared to do nothing at all.
+func TestReloadReplacesRatherThanAppends(t *testing.T) {
+	dir := t.TempDir()
+	pluginDir := filepath.Join(dir, "plugins")
+	if err := os.MkdirAll(pluginDir, 0o755); err != nil {
+		t.Fatal(err)
+	}
+	vm := NewVM(filepath.Join(dir, "adapters"))
+	if err := vm.Start(); err != nil {
+		t.Fatal(err)
+	}
+	t.Cleanup(vm.Stop)
+	ps := NewPlugins(vm, pluginDir)
+	t.Cleanup(ps.Close)
+
+	if err := os.WriteFile(filepath.Join(pluginDir, "probe.lua"),
+		[]byte("local P={}\nP.name='probe'\nP.version='1'\nreturn P\n"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	if err := ps.LoadDir(); err != nil {
+		t.Fatal(err)
+	}
+	if got := len(ps.List()); got != 1 {
+		t.Fatalf("after LoadDir: %d records, want 1", got)
+	}
+
+	if err := ps.LoadSource("probe",
+		"local P={}\nP.name='probe'\nP.version='2'\nreturn P\n"); err != nil {
+		t.Fatal(err)
+	}
+	if got := len(ps.List()); got != 1 {
+		t.Errorf("after reload: %d records, want 1 — the old one was kept", got)
+	}
+	p := ps.find("probe")
+	if p == nil {
+		t.Fatal("plugin gone after reload")
+	}
+	if p.Info.Version != "2" {
+		t.Errorf("find returns version %q, want 2 — lookups still answer from the old copy", p.Info.Version)
+	}
+}

internal/lua/plugins.go

@@ -32,9 +32,9 @@ package lua
 //     less featureful, so the split is explicit and documented.
 
 import (
-	"log"
 	"encoding/json"
 	"fmt"
+	"log"
 	"os"
 	"path/filepath"
 	"sort"
@@ -265,6 +265,10 @@ type Plugins struct {
 	// visibility holds the operator's per-plugin audience overrides, keyed by
 	// plugin name. Consulted by UIForRole and the call endpoint's authorisation.
 	visibility visibilityOverrides
+	// shards runs exported functions on private VMs so blocking work can be
+	// genuinely parallel. See plugin_shard.go for why the shared VM's lock makes
+	// concurrency impossible and what a shard must therefore never touch.
+	shards *shardPool
 	// stateDirDisabled turns persistence off. Used by tests that assert state
 	// starts empty, and by an embedder that has no writable plugin dir.
 	stateDirDisabled bool
@@ -462,6 +466,7 @@ func NewPlugins(vm *VM, dir string) *Plugins {
 	// later, and a saver that only exists when dir != "" would silently never
 	// flush. markDirtyLocked and flush both no-op when the dir is empty.
 	ps.save = newStateSaver(ps)
+	ps.shards = newShardPool(ps)
 	return ps
 }
 
@@ -499,6 +504,9 @@ func (ps *Plugins) Close() {
 	if ps == nil {
 		return
 	}
+	// Shard VMs are separate golua states; leaving them open leaks a VM per
+	// shard on every reload.
+	ps.shards.Shutdown()
 	ps.save.Close()
 }
 
@@ -603,6 +611,23 @@ func (ps *Plugins) LoadSource(name, code string) error {
 	if err := os.MkdirAll(ps.dir, 0755); err != nil && ps.dir != "" {
 		return err
 	}
+	// Reloading must REPLACE, not append.
+	//
+	// LoadSource was reachable both from LoadDir and from the install/reload API,
+	// and it appended unconditionally. Loading a plugin that was already loaded
+	// therefore left two records with the same name, and findLocked — which
+	// returns the FIRST match — kept handing out the OLD one. Measured: after
+	// LoadDir + LoadSource of an edited plugin, /api/plugins listed both
+	// versions and Call answered from version 1 while version 2 sat inert
+	// alongside it. Editing a plugin and saving appeared to do nothing.
+	//
+	// Unload is a no-op when nothing is loaded under that name, so this is safe
+	// on the first load.
+	if ps.find(name) != nil {
+		if err := ps.Unload(name); err != nil {
+			return fmt.Errorf("replace %s: %w", name, err)
+		}
+	}
 	p := &Plugin{
 		Info:  PluginInfo{Name: name},
 		Hooks: map[Stage]string{},
@@ -667,20 +692,13 @@ func (ps *Plugins) LoadSource(name, code string) error {
 	}
 
 	// ---- UI ----
-	if ui, ok := readPluginUI(w.L); ok {
+	if ui, ok := readPluginUI(w.L, p.Info.Name); ok {
 		p.UI = ui
 	}
 
 	// ---- http endpoints ----
 	p.http = readPluginHTTP(w.L)
 
-	if ui, ok := readPluginUI(w.L); ok {
-		p.UI = ui
-		if ui.StateScope != nil {
-			ui.StateScope.Depth = compileDepthMap(ui.StateScope.Depth)
-		}
-	}
-
 	// ---- exports ----
 	// Read here, while the state is locked and before restore() runs: the list is
 	// static for the life of the plugin, so Call never has to re-enter Lua just
@@ -778,111 +796,112 @@ func pluginHookName(L *golua.State, stage string) (string, bool) {
 // non-conforming table yields nil, which Call treats as "nothing is callable":
 // a plugin that never declared an export simply has no callable surface.
 func readPluginExports(L *golua.State) []string {
-	L.GetGlobal(pluginGlobal)
-	if L.IsNil(-1) {
-		L.SetTop(0)
-		return nil
-	}
-	L.GetField(-1, "exports")
-	if L.Type(-1) != golua.LUA_TTABLE {
-		L.SetTop(0)
-		return nil
-	}
-	// luaTableToGoValue, NOT luaToJSON: this runs while the plugin is registering,
-	// before the json global is installed, and luaToJSON needs it. Going through
-	// JSON here silently produced an empty list, which then made every exported
-	// function look unexported.
-	raw := luaTableToGoValue(L, -1)
-	L.SetTop(0)
-	arr, ok := raw.([]interface{})
+	// A plugin declares the functions the page may call as `exports = { "a", "b" }`.
+	// Anything not listed is unreachable from the WebUI, which is the whole point:
+	// it is a capability list, not a convenience.
+	raw, ok := pluginFieldValue(L, "exports")
 	if !ok {
 		return nil
 	}
-	list := make([]string, 0, len(arr))
-	for _, v := range arr {
-		if s := asString(v); s != "" {
-			list = append(list, s)
-		}
-	}
-	return list
+	return decStringList(map[string]interface{}{"v": raw}, "v")
 }
 
-func readPluginUI(L *golua.State) (*UIExtension, bool) {
+func readPluginUI(L *golua.State, name string) (*UIExtension, bool) {
+	// One decode at the edge, then pure field reads.
+	//
+	// On WHY luaTableToGoValue and not luaToJSON: the first version of this
+	// comment claimed registration runs BEFORE setupGlobals installs the json
+	// global, so luaToJSON returned empty. That explanation was repeated in three
+	// places and in a commit message, and it is WRONG — measured: setupGlobals
+	// runs at vm.go:574, these readers at 787, and json IS present on the stack
+	// when they run.
+	//
+	// What actually differs is INDEX SEMANTICS. luaTableToGoValue converts a
+	// NEGATIVE index to absolute immediately (abs = GetTop() + idx + 1);
+	// luaToJSON keeps the relative index and pushes it before Call. So after a
+	// SetTop(0) the two disagree, and the table-valued converter — the one that
+	// handles nested tables, arrays and mixed keys correctly — is the one that
+	// needs the stack shape kept intact.
+	//
+	// Which is exactly why the decode happens here, ONCE, immediately after
+	// GetGlobal+GetField, and every field read below is a pure function on the
+	// decoded Go value. The old readers repeated the stack dance per field.
+	m, ok := pluginFieldTable(L, "ui")
+	if !ok {
+		return nil, false
+	}
+	ctx := &decodeCtx{plugin: name}
+	ui := decodeUIExtension(ctx, m)
+
+	if ui.Page == nil && len(ui.Pages) == 0 && len(ui.Elements) == 0 {
+		return nil, false
+	}
+	if len(ctx.problems) > 0 {
+		log.Printf("[plugins] %s: %s", ctx.plugin, strings.Join(ctx.problems, "; "))
+	}
+	return &ui, true
+}
+
+// pluginFieldTable reads one top-level declaration table off the plugin global.
+func pluginFieldTable(L *golua.State, field string) (map[string]interface{}, bool) {
 	L.GetGlobal(pluginGlobal)
 	if L.IsNil(-1) {
 		L.SetTop(0)
 		return nil, false
 	}
-	L.GetField(-1, "ui")
+	L.GetField(-1, field)
 	if L.Type(-1) == golua.LUA_TNIL {
 		L.SetTop(0)
 		return nil, false
 	}
-	// luaTableToGoValue, NOT luaToJSON. This runs while the plugin is
-	// registering, BEFORE setupGlobals installs the json global that luaToJSON
-	// depends on — so the conversion fails and every field except the ones on
-	// Page/Pages (read separately) comes back empty. That is why visibility and
-	// state_scope silently read as unset even though the .lua declares them:
-	// the third field in this file to be lost to the same cause, after
-	// readPluginExports and readPluginHTTP.
-	var ui UIExtension
 	raw := luaTableToGoValue(L, -1)
 	L.SetTop(0)
-	if m, ok := raw.(map[string]interface{}); ok {
-		ui.Visibility = asString(m["visibility"])
-		if scope, ok := m["state_scope"].(map[string]interface{}); ok {
-			sc := &StateScope{}
-			if gs, ok := scope["global"].([]interface{}); ok {
-				for _, g := range gs {
-					if s := asString(g); s != "" {
-						sc.Global = append(sc.Global, s)
-					}
-				}
-			}
-			if ds, ok := scope["depth"].(map[string]interface{}); ok {
-				sc.Depth = map[string]int{}
-				for k, v := range ds {
-					if n, ok := asFloat(v); ok {
-						sc.Depth[k] = int(n)
-					}
-				}
-				sc.Depth = compileDepthMap(sc.Depth)
-			}
-			ui.StateScope = sc
-		}
-		if pg, ok := m["page"].(map[string]interface{}); ok {
-			ui.Page = uipageFromValue(pg)
-		}
-		if pgs, ok := m["pages"].([]interface{}); ok {
-			for _, it := range pgs {
-				if pm, ok := it.(map[string]interface{}); ok {
-					ui.Pages = append(ui.Pages, uipageFromValue(pm))
-				}
-			}
-		}
-		if els, ok := m["elements"].([]interface{}); ok {
-			for _, it := range els {
-				em, ok := it.(map[string]interface{})
-				if !ok {
-					continue
-				}
-				el := UIElement{
-					Target:   asString(em["target"]),
-					Anchor:   asString(em["anchor"]),
-					Mount:    asString(em["mount"]),
-					Audience: asString(em["audience"]),
-				}
-				if n, ok := asFloat(em["order"]); ok {
-					el.Order = int(n)
-				}
-				ui.Elements = append(ui.Elements, el)
-			}
+	m, ok := raw.(map[string]interface{})
+	return m, ok
+}
+
+// decodeUIExtension reads a plugin's `ui` declaration.
+//
+// The field list below is the single source of truth for what the gateway
+// understands. A declaration using anything else is REPORTED rather than
+// silently dropped, because the old hand-written reader ignored unknown keys with
+// no trace: `visibilty = "admin"` (one character off) restricted nothing and
+// looked exactly like a plugin that simply chose not to declare an audience.
+func decodeUIExtension(ctx *decodeCtx, m map[string]interface{}) UIExtension {
+	ctx.strictKeys(m, "ui", "visibility", "state_scope", "page", "pages", "elements")
+
+	var ui UIExtension
+	ui.Visibility = decString(m, "visibility")
+
+	if scope := decTable(m, "state_scope"); scope != nil {
+		ctx.strictKeys(scope, "ui.state_scope", "global", "depth")
+		sc := &StateScope{Global: decStringList(scope, "global")}
+		if d := decIntMap(scope, "depth"); len(d) > 0 {
+			sc.Depth = compileDepthMap(d)
 		}
+		ui.StateScope = sc
 	}
-	if ui.Page == nil && len(ui.Pages) == 0 && len(ui.Elements) == 0 {
-		return nil, false
+
+	if pg := decTable(m, "page"); pg != nil {
+		ui.Page = uipageFromValue(pg)
 	}
-	return &ui, true
+	for _, pm := range decTableList(m, "pages") {
+		ui.Pages = append(ui.Pages, uipageFromValue(pm))
+	}
+	for _, em := range decTableList(m, "elements") {
+		ctx.strictKeys(em, "ui.elements[]", "target", "anchor", "mount", "audience", "order")
+		el := UIElement{
+			Target:   decString(em, "target"),
+			Anchor:   decString(em, "anchor"),
+			Mount:    decString(em, "mount"),
+			Audience: decString(em, "audience"),
+		}
+		if n, ok := asFloat(em["order"]); ok {
+			el.Order = int(n)
+		}
+		ui.Elements = append(ui.Elements, el)
+	}
+	return ui
 }
 
 func (ps *Plugins) append(p *Plugin) {
@@ -1140,18 +1159,9 @@ func (ps *Plugins) HookErrors() map[string]map[string]interface{} {
 // "*" is deliberately NOT a synonym for "all": an audience that looks
 // restrictive in a config file while meaning everything is exactly the kind of
 // setting that fails silently.
+// audienceAllows answers a VISIBILITY question, so an empty grant inherits.
 func audienceAllows(want, role string) bool {
-	want = strings.ToLower(strings.TrimSpace(want))
-	if want == "" || want == "all" {
-		return true
-	}
-	role = strings.ToLower(strings.TrimSpace(role))
-	for _, part := range strings.Split(want, ",") {
-		if strings.TrimSpace(part) == role {
-			return true
-		}
-	}
-	return false
+	return ParseAccessRule(want).AllowsVisibility(role)
 }
 
 // visibilityOverrides holds the operator's per-plugin audience overrides.
@@ -1950,6 +1960,17 @@ func (ps *Plugins) invoke(p *Plugin, fn string, payload map[string]interface{})
 // quota report is a nested structure with numbers and booleans that must not be
 // flattened. Hook results still go through invoke, whose map-typed contract the
 // state merge depends on.
+// Call invokes an exported function on a plugin's SHARED VM, serialised by the
+// plugin's mutex.
+//
+// The lock is what makes one Lua VM safe to share, and it is held for the whole
+// call — including any network wait inside it. That is correct for a hook and
+// wrong for a function that blocks: twenty concurrent quota probes serialise
+// into the sum of their timeouts (measured: 29s on a 20-source gateway, against
+// 1.4–3.3s for a single probe with no contention).
+//
+// callOn below is the same work without the shared-VM assumption, so a caller
+// that owns a private VM can use it directly.
 func (ps *Plugins) Call(name, fn string, args map[string]interface{}) (interface{}, error) {
 	ps.mu.RLock()
 	var p *Plugin
@@ -1967,12 +1988,27 @@ func (ps *Plugins) Call(name, fn string, args map[string]interface{}) (interface
 		return nil, fmt.Errorf("plugin %s does not export %q", name, fn)
 	}
 	p.mu.Lock()
+	defer p.mu.Unlock()
 	w := p.state
 	if w == nil {
-		p.mu.Unlock()
 		return nil, fmt.Errorf("no state for plugin %s", name)
 	}
-	defer p.mu.Unlock()
+	return callOn(w, fn, args)
+}
+
+// callOn runs one exported function on a specific VM.
+//
+// It does NOT check the export allowlist and does NOT persist state: both belong
+// to the shared-VM path in Call. A shard VM has no plugin identity to check
+// against and no state file of its own — the export allowlist in Call is the
+// security boundary, and calling this with a function the plugin does not export
+// is a programming error, not an attack.
+//
+// The caller must own w exclusively. No lock is taken here.
+func callOn(w *worker, fn string, args map[string]interface{}) (interface{}, error) {
+	if w == nil || w.L == nil {
+		return nil, fmt.Errorf("no lua state")
+	}
 	L := w.L
 	L.SetTop(0)
 	defer L.SetTop(0)
@@ -2005,14 +2041,14 @@ func (ps *Plugins) Call(name, fn string, args map[string]interface{}) (interface
 		L.Pop(1)
 	}
 	if errMsg != "" {
-		return nil, fmt.Errorf("plugin %s: %s", name, errMsg)
+		return nil, fmt.Errorf("plugin: %s", errMsg)
 	}
 	if L.GetTop() < 1 || L.IsNil(-1) {
 		return nil, nil
 	}
 	var out interface{}
 	if err := luaToJSON(L, -1, &out); err != nil {
-		return nil, nil
+		return nil, err
 	}
 	return out, nil
 }
@@ -2097,40 +2133,11 @@ func (h HTTPHandler) routePath() string {
 //
 // keyName is the caller key's configured name ("" when it has none), not the
 // secret.
+// allows answers a CAPABILITY question, so an empty grant denies. See
+// AccessRule.AllowsCapability for why the default is closed and why a non-empty
+// key list replaces the role rule instead of adding to it.
 func (h HTTPHandler) allows(role, keyName string) bool {
-	// Admin is always admitted. An operator must be able to inspect what a plugin
-	// exposes even while locked out of it, or a typo in a key list would leave
-	// nobody able to diagnose the endpoint they were trying to tighten.
-	if role == "admin" {
-		return true
-	}
-	// Named keys REPLACE the role rule rather than adding to it.
-	if len(h.Keys) > 0 {
-		if keyName == "" {
-			return false
-		}
-		for _, k := range h.Keys {
-			if strings.EqualFold(strings.TrimSpace(k), keyName) {
-				return true
-			}
-		}
-		return false
-	}
-	if h.Public {
-		return true
-	}
-	if len(h.Roles) == 0 {
-		// No declared roles and no public opt-out: treat as admin-only. A plugin
-		// author who forgets this gets a 403 rather than an open endpoint, and
-		// the error says which field to set.
-		return role == "admin"
-	}
-	for _, r := range h.Roles {
-		if strings.EqualFold(strings.TrimSpace(r), role) {
-			return true
-		}
-	}
-	return false
+	return accessRuleFor(h.Roles, h.Keys, h.Public).AllowsCapability(role, keyName)
 }
 
 func (ps *Plugins) StateVisibleTo(name, role string) bool {
@@ -2176,64 +2183,82 @@ func containsPage(pages []*UIPage, id string) bool {
 // not exist. Failing the whole plugin would take a working billing widget down
 // over a typo in an experimental endpoint.
 func readPluginHTTP(L *golua.State) []HTTPHandler {
-	L.GetGlobal(pluginGlobal)
-	if L.IsNil(-1) {
-		L.SetTop(0)
-		return nil
-	}
-	L.GetField(-1, "http")
-	if L.Type(-1) != golua.LUA_TTABLE {
-		L.SetTop(0)
-		return nil
-	}
-	// luaTableToGoValue, NOT luaToJSON and NOT a hand-rolled GetField loop.
+	// `http` is a LIST of declarations, each naming the handler function by
+	// STRING:
 	//
-	// luaToJSON needs the json global, which is not installed yet while a plugin
-	// is registering. And golua's GetField does NOT convert a numeric key: with
-	// a string index "1" it returns nil, so an index loop silently produced an
-	// empty list and every plugin appeared to declare no endpoints at all.
-	raw := luaTableToGoValue(L, -1)
-	L.SetTop(0)
-	items, ok := raw.([]interface{})
+	//	http = {
+	//	  { path = "health", method = "GET", roles = { "admin" }, handler = "health" },
+	//	}
+	//
+	// handler is a name, not a value. An inline function silently produced an
+	// empty handler field and the endpoint was dropped without a word.
+	raw, ok := pluginFieldValue(L, "http")
 	if !ok {
 		return nil
 	}
+	// A single declaration may be given as a bare table instead of a one-element
+	// list; accepting both is friendlier than demanding one spelling.
+	items, isList := raw.([]interface{})
+	if !isList {
+		if one, isTable := raw.(map[string]interface{}); isTable {
+			items = []interface{}{one}
+		} else {
+			return nil
+		}
+	}
+	ctx := &decodeCtx{plugin: "http"}
 	var list []HTTPHandler
 	for _, it := range items {
-		m, ok := it.(map[string]interface{})
+		h, ok := it.(map[string]interface{})
 		if !ok {
 			continue
 		}
-		h := HTTPHandler{
-			Path:        asString(m["path"]),
-			Method:      asString(m["method"]),
-			Handler:     asString(m["handler"]),
-			Description: asString(m["description"]),
-			Public:      m["public"] == true,
+		ctx.strictKeys(h, "http[]", "path", "method", "handler", "description",
+			"roles", "keys", "public")
+		dh := HTTPHandler{
+			Path:        decString(h, "path"),
+			Method:      decString(h, "method"),
+			Description: decString(h, "description"),
+			Roles:       decStringList(h, "roles"),
+			Keys:        decStringList(h, "keys"),
 		}
-		if roles, ok := m["roles"].([]interface{}); ok {
-			for _, r := range roles {
-				if s := asString(r); s != "" {
-					h.Roles = append(h.Roles, s)
-				}
+		// handler must be a NAME, not a value. asString would happily turn a
+		// function into its address ("function: 0x..."), which then failed much
+		// later at dispatch with no connection to the declaration that caused it.
+		// Rejecting it here names the field and the path.
+		if hv, present := h["handler"]; present {
+			name, isStr := hv.(string)
+			if !isStr {
+				ctx.warnf("http[%s]: handler must be the NAME of a function on the "+
+					"plugin table, not the function itself", decString(h, "path"))
+				continue
 			}
+			dh.Handler = name
 		}
-		if keys, ok := m["keys"].([]interface{}); ok {
-			for _, k := range keys {
-				if s := asString(k); s != "" {
-					h.Keys = append(h.Keys, s)
-				}
-			}
+		if p, ok := h["public"].(bool); ok {
+			dh.Public = p
 		}
 		// A malformed entry is dropped rather than fatal: one bad route must not
-		// cost a plugin its hooks and pages. That is also why the mounted count
-		// can be lower than the declared count — /api/plugins lists what mounted.
-		if method, mErr := normalizeHandlerMethod(h.Method); mErr == nil {
-			h.Method = method
-			if validateHandlerPath(h.Path) == nil && h.Handler != "" {
-				list = append(list, h)
-			}
+		// cost a plugin its hooks and pages. The reason is named so a plugin whose
+		// endpoint silently vanished can find out why.
+		method, mErr := normalizeHandlerMethod(dh.Method)
+		switch {
+		case mErr != nil:
+			ctx.warnf("http[%s]: %v", dh.Path, mErr)
+			continue
+		case validateHandlerPath(dh.Path) != nil:
+			ctx.warnf("http[%s]: invalid path", dh.Path)
+			continue
+		case dh.Handler == "":
+			ctx.warnf("http[%s]: handler must name a function on the plugin table "+
+				"(an inline function is not supported)", dh.Path)
+			continue
 		}
+		dh.Method = method
+		list = append(list, dh)
+	}
+	if len(ctx.problems) > 0 {
+		log.Printf("[plugins] %s", strings.Join(ctx.problems, "; "))
 	}
 	return list
 }

internal/lua/plugins/quota.lua

@@ -43,7 +43,7 @@ PLUGIN.name = "quota"
 PLUGIN.version = "1.0.0"
 PLUGIN.description = "多源套餐余量（余额 / 用量 / 滚动窗口）"
 
-PLUGIN.exports = { "collect_source", "collect_all" }
+PLUGIN.exports = { "collect_source", "collect_all", "due_sources", "memo_record" }
 
 -- plugin.state must exist before anything indexes it. The kernel restores
 -- persisted state OVER this table, so declaring it here is safe: a restart with
@@ -63,8 +63,23 @@ end
 -- hostOf extracts the host from a base_url, for showing WHICH upstream a row is
 -- about. Two sources on the same vendor (two Command Code keys) are
 -- indistinguishable without it.
+-- hostOf returns the host of a URL — the part memoisation is keyed on, so two
+-- sources pointing at the same host with different paths share one answer.
+--
+-- It used to be written with string.gmatch and a capture group. gmatch yields
+-- WHOLE matches, not captures, and with `^` it never advances, so the iterator
+-- returned nil on the first step and hostOf returned its `or u` fallback —
+-- the FULL URL.
+--
+-- The consequence was invisible and total. Every lookup was HOST_MEMO[nil] and
+-- every write was HOST_MEMO[nil], so:
+--   - the memo table never returned anything, for any host, ever
+--   - the negative memo, the TTL and (later) the backoff were all dead code
+--   - the 19 entries in this gateway's production state are unreachable: their
+--     keys are full URLs, while a lookup asks for nil
+-- Measured before the fix: hostOf(key) matched 0 of 19 production entries.
 local function hostOf(u)
-  return (string.gmatch(u, "^https?://([^/]+)")()) or u
+  return (string.match(u, "^https?://([^/]+)")) or u
 end
 
 -- get walks a dotted path through nested tables, returning nil if any hop is
@@ -298,6 +313,52 @@ local HOST_MEMO = {}
 local HOST_MEMO_AT = {}
 local PROBE_MEMO_TTL = 1800 -- 30 minutes
 
+-- Exponential backoff for hosts that answer nothing.
+--
+-- A flat TTL is the wrong shape for a dead upstream. Measured on this gateway:
+-- two hosts had been unreachable for 10 HOURS, and because PROBE_MEMO_TTL is 30
+-- minutes their negative memo was stale on every single page open — so each
+-- refresh re-paid the full cost of three adapters (measured: 24s per source
+-- before the probe budget was fixed, 16s after). A host that is down should not
+-- be re-examined every 30 minutes forever; each consecutive failure doubles the
+-- wait, so a long-dead host costs almost nothing while a host that was briefly
+-- down is retried within the TTL as before.
+--
+-- The cap matters as much as the doubling: without it a host that is down for a
+-- week stops being probed at all, and the page would quietly and permanently
+-- misreport it as having no balance endpoint.
+--
+-- SUCCESS resets the streak. This is only a probe-scheduling decision — a source
+-- that stops answering is still shown, still counted, and still retried.
+local BACKOFF_BASE = 1800        -- first retry after the usual TTL
+local BACKOFF_FACTOR = 2         -- doubles per consecutive miss
+local BACKOFF_MAX = 21600        -- 6 hours; beyond this we always re-probe
+
+-- backoffSeconds returns how long to wait before re-probing a host that has
+-- missed `streak` times in a row, and 0 when it is due now.
+--
+-- `at` is when the host last failed. streak is stored inside the memo table
+-- itself so it survives a restart: a process restart must not reset the backoff
+-- and re-pay the cost it was avoiding.
+-- Returns a BOOLEAN and nothing else.
+--
+-- In Lua only nil and false are falsy, so every path here must return one of
+-- those. Two versions of this got it wrong and both were caught by a test that
+-- asserts the type: `return 0` for "no streak" (0 is truthy, so a host with no
+-- streak — which is every memo written before this change — would be skipped
+-- forever, silently), and `return BACKOFF_MAX` on the cap path (21600 is
+-- truthy too, so a long-dead host would never be re-probed, which is the exact
+-- failure the cap exists to prevent).
+local function backoffSeconds(at, streak)
+  if streak == nil or streak <= 0 then return false end
+  local wait = BACKOFF_BASE
+  for _ = 2, streak do
+    wait = wait * BACKOFF_FACTOR
+    if wait >= BACKOFF_MAX then wait = BACKOFF_MAX break end
+  end
+  return (os.time() - (at or 0)) < wait
+end
+
 -- PROBE_BUDGET_MS caps one collect pass. Without it a page open over a set of
 -- dead upstreams runs to the sum of every timeout — measured at 149s — and the
 -- plugin's write lock is held the whole time.
@@ -325,6 +386,12 @@ local function collectSource(src, budget)
   -- exactly the confusion that made an earlier revision drop five sources.
   if type(known) == "table" then
     local seen = known.at or 0
+      -- A streak is a host that has failed repeatedly. Skip it outright until its
+      -- backoff expires, instead of re-paying three adapter timeouts every TTL.
+      if backoffSeconds(seen, known.streak) then
+        return { name = src.name, host = host, skipped = true, backing_off = true,
+                 reason = "连续探测失败 " .. tostring(known.streak) .. " 次，退避中（等待下次重试）" }
+      end
     if PROBE_MEMO_TTL > 0 and (os.time() - seen) > PROBE_MEMO_TTL then
       known = nil
     else
@@ -394,8 +461,21 @@ local function collectSource(src, budget)
   -- would make the page quietly stop reporting a source that started working.
   -- The stamp is what turns a 60s first load into a fast steady state without
   -- freezing what the page can see.
-  if rep == nil and not HOST_MEMO[host] then
-    HOST_MEMO[host] = { at = os.time() }
+  -- Record the negative result, and GROW THE STREAK.
+  --
+  -- The guard used to be `not HOST_MEMO[host]`, which is true only the FIRST time
+  -- a host fails: on every later failure the memo already held a table, the
+  -- guard was false, and nothing was written. The streak therefore stayed pinned
+  -- at 1 forever and the backoff never grew. Measured while writing this: the
+  -- memo read back `{at:... streak:1}` after four consecutive failures.
+  --
+  -- A success overwrites the memo with a vendor string, so a host that recovers
+  -- resets its streak automatically.
+  if rep == nil then
+    local prev = HOST_MEMO[host]
+    local streak = 1
+    if type(prev) == "table" then streak = (prev.streak or 1) + 1 end
+    HOST_MEMO[host] = { at = os.time(), streak = streak }
   end
   if rep ~= nil then
     HOST_MEMO_AT[host] = os.time()
@@ -421,6 +501,96 @@ end
 -- and doing 21 sources in sequence overran the HTTP call's own budget — the
 -- whole page returned 502 while individual probes were still in flight. The
 -- sequential entry point stays for tests and for anyone scripting the plugin.
+-- due_sources answers, on the SHARED VM, which sources are worth probing right
+-- now — and it makes NO network requests.
+--
+-- This exists because of a conflict between two facts:
+--
+--   - Probing is network-bound, and Plugins.Call holds the plugin mutex for the
+--     whole Lua function, so twenty probes through /call/ serialise. Measured:
+--     29s on a 20-source gateway. The fix is ShardCall, which runs each call on
+--     its own VM — and a shard VM has its own PLUGIN.state.
+--   - The backoff lives in PLUGIN.state (HOST_MEMO). A shard therefore cannot
+--     see it, cannot record it, and a host that fails keeps being probed on every
+--     page open. Measured: fanout left the memo at 0 entries, collect_all reached
+--     17.
+--
+-- So the split is: the DECISION runs where the state is (one shared-VM call,
+-- microseconds), and only the hosts that pass it get sharded.
+--
+-- Cheap on purpose. It reads HOST_MEMO and nothing else; if the state has not
+-- been restored yet, every host is due, which is the correct answer for a first
+-- call.
+-- memo_record writes the outcome of a sharded pass into the SHARED VM's memo.
+--
+-- A shard runs on its own VM and therefore cannot record anything: measured, a
+-- page of fanouts left the memo at 0 entries, so the backoff never accumulated
+-- and every refresh re-probed the dead hosts. This is the other half of
+-- due_sources — the decision is taken here, the outcome is written back here, and
+-- only the network wait happens on the shards.
+--
+-- It records FAILURES as well as successes. Recording only successes is what
+-- left every source with no balance endpoint re-probing three adapters forever.
+PLUGIN.memo_record = function(payload)
+  if type(payload) ~= "table" then return nil, "expected table" end
+  if not next(HOST_MEMO) then
+    local st = PLUGIN.state
+    if type(st) == "table" and type(st.probe_memo) == "table" then
+      for k, v in pairs(st.probe_memo) do HOST_MEMO[k] = v end
+    end
+  end
+  local rows = payload.results
+  if type(rows) ~= "table" then return nil, "missing results" end
+  local now = os.time()
+  for _, r in ipairs(rows) do
+    if type(r) == "table" and type(r.host) == "string" and r.host ~= "" then
+      local host = hostOf(r.host)
+      if r.vendor then
+        -- A source that answered resets its streak: the memo becomes a vendor
+        -- string, which every reader already treats as a positive memo.
+        HOST_MEMO[host] = r.vendor
+      else
+        local prev = HOST_MEMO[host]
+        local streak = 1
+        if type(prev) == "table" then streak = (prev.streak or 1) + 1 end
+        HOST_MEMO[host] = { at = now, streak = streak }
+      end
+    end
+  end
+  PLUGIN.state.probe_memo = HOST_MEMO
+  return true
+end
+
+PLUGIN.due_sources = function(payload)
+  if type(payload) ~= "table" then return nil, "expected table" end
+  if not next(HOST_MEMO) then
+    local st = PLUGIN.state
+    if type(st) == "table" and type(st.probe_memo) == "table" then
+      for k, v in pairs(st.probe_memo) do HOST_MEMO[k] = v end
+    end
+  end
+  local srcs = payload.sources
+  if type(srcs) ~= "table" then return nil, "missing sources" end
+  local due, backing = {}, {}
+  for _, s in ipairs(srcs) do
+    if type(s) == "table" and type(s.base_url) == "string" and s.base_url ~= "" then
+      local base = (s.base_url:gsub("/+$", ""))
+      local known = HOST_MEMO[hostOf(base)]
+      if type(known) == "table" and backoffSeconds(known.at or 0, known.streak) then
+        -- skipped matters: the page filters on it. Without the flag a backed-off
+        -- host reaches the card renderer, which has no field to match and falls
+        -- through to its "has data" default — so thirteen dead upstreams were
+        -- displayed as if they had answered.
+        backing[#backing + 1] = { name = s.name, host = hostOf(base), skipped = true, backing_off = true,
+                                  reason = "连续探测失败 " .. tostring(known.streak) .. " 次，退避中（等待下次重试）" }
+      else
+        due[#due + 1] = s
+      end
+    end
+  end
+  return { due = due, backing_off = backing }
+end
+
 PLUGIN.collect_all = function(payload)
   if type(payload) ~= "table" then return nil, "expected table" end
   if not next(HOST_MEMO) then
@@ -431,10 +601,15 @@ PLUGIN.collect_all = function(payload)
   end
   local srcs = payload.sources
   if type(srcs) ~= "table" then return nil, "missing sources" end
-  local t1 = os.clock()
+  -- os.time, NOT os.clock. os.clock measures CPU time, and a probe spends
+  -- essentially none: measured on this gateway, 3s blocked on an HTTP request
+  -- advanced os.clock by 0.001s. A budget written in os.clock therefore never
+  -- fires, which is why a dead upstream ran the full
+  -- 4s timeout x 2 retries x 3 adapters = 24s against a "12 second" budget.
+  local t1 = os.time()
   local lim1 = PROBE_BUDGET_MS / 1000
   local function spent()
-    return (os.clock() - t1) > lim1
+    return (os.time() - t1) > lim1
   end
   local budget = { exhausted = spent }
   local out = {}
@@ -463,10 +638,10 @@ PLUGIN.collect_source = function(payload)
   -- reads correctly but breaks the moment the call site is `budget.exhausted()`
   -- from inside a pcall-guarded hook where the upvalue chain differs; the
   -- explicit capture below cannot go wrong that way.
-  local t0 = os.clock()
+  local t0 = os.time()
   local limit = PROBE_BUDGET_MS / 1000
   local function exhausted()
-    return (os.clock() - t0) > limit
+    return (os.time() - t0) > limit
   end
   local budget = { exhausted = exhausted }
   for _, s in ipairs(srcs) do
@@ -653,7 +828,14 @@ PLUGIN.ui = {
     }
     if(r.hard_limit!==undefined && r.hard_limit!==null)
       return {cls:"ok", pill:'<span class="qa-pill ok">按量计费</span>'};
-    return {cls:"", pill:'<span class="qa-pill">有数据</span>'};
+    // Nothing here matched. Say so.
+    //
+    // The old default was a neutral "有数据", which is a claim: it says the probe
+    // answered. A record that reached this point answered with no field the page
+    // knows — a back-off notice, an error shape we do not recognise, or a vendor
+    // that added a new response. Calling that "has data" is how thirteen dead
+    // upstreams ended up rendered as healthy. Absence of evidence is not evidence.
+    return {cls:"unknown", pill:'<span class="qa-pill">未探测</span>'};
   }
   function card(r){
     var st=state(r);
@@ -666,6 +848,13 @@ PLUGIN.ui = {
       h+='<div class="qa-fail">'+escH(r.error)+'</div></div>';
       return h;
     }
+    // A record with nothing but a reason has no balance, no month, no windows —
+    // the card renders as a bare header with an empty body. Say why instead.
+    if(r.reason && !r.error && (r.balance===undefined||r.balance===null)
+       && (r.hard_limit===undefined||r.hard_limit===null)){
+      h+='<div class="qa-fail">'+escH(r.reason)+'</div>';
+      return h+'</div>';
+    }
     h+=balLine(r);
     var rows="";
     rows+=month(r);
@@ -690,7 +879,10 @@ PLUGIN.ui = {
       var rs = (st && st.reports) || [];
       if(!rs.length) return false;
       var real = rs.filter(function(r){ return r && !r.skipped; });
-      var skipped = rs.filter(function(r){ return r && r.skipped; }).map(function(r){ return r.name; });
+      // Records, not names — same reason as the refresh path. Both call sites were
+    // mapping to names while render() was changed to group on the flag; fixing
+    // one and shipping it is what produced a footer with thirteen blank entries.
+    var skipped = rs.filter(function(r){ return r && r.skipped; });
       render(real, skipped);
       var sub=document.getElementById("qa-sub");
       if(sub && st.updated_at){
@@ -707,10 +899,22 @@ PLUGIN.ui = {
       host.innerHTML='<div class="qa-empty">没有源提供可读的余额接口</div>';
     } else {
       host.innerHTML=real.map(card).join("");
-      if(skipped && skipped.length){
-        host.innerHTML+='<div class="qa-skip">未列出 '+skipped.length+
-          ' 个源（未暴露余额接口）：'+escH(skipped.join("、"))+'</div>';
-      }
+        if(skipped && skipped.length){
+          // Two reasons a source has no card, and they are not interchangeable.
+          // "This upstream exposes no balance endpoint" and "this upstream is
+          // backing off after failing" are different facts; collapsing them into
+          // one list tells the operator their dead host simply has no dashboard.
+          var backing = skipped.filter(function(r){ return r.backing_off===true; });
+          var silent  = skipped.filter(function(r){ return r.backing_off!==true; });
+          if(backing.length){
+            host.innerHTML+='<div class="qa-skip">'+backing.length+
+              ' 个源正在退避（连续探测失败，等下次重试）：'+escH(backing.map(function(r){return r.name;}).join("、"))+'</div>';
+          }
+          if(silent.length){
+            host.innerHTML+='<div class="qa-skip">未列出 '+silent.length+
+              ' 个源（未暴露余额接口）：'+escH(silent.map(function(r){return r.name;}).join("、"))+'</div>';
+          }
+        }
     }
     var sub=document.getElementById("qa-sub");
     if(sub) sub.textContent = real.length+" 个源 · "+bad.length+" 个读取失败";
@@ -732,15 +936,39 @@ PLUGIN.ui = {
     var sub=document.getElementById("qa-sub"); if(sub) sub.textContent="探测中…";
     var out;
     try{
-      // One call per source, fired together. Probing is network-bound, so a
-    // sequential collect_all over 21 sources blows past the call's time budget
-    // and every card comes back as one 502 — the probes themselves were fine.
-    // Concurrency also makes the page honest about which sources are slow:
-    // each card appears as its own answer lands.
-    out = await Promise.all(srcs.map(function(s){
-      return pluginAPI.call("quota","collect_source",{sources:[s],name:s.name})
-        .catch(function(e){ return {name:s.name,skipped:true,error:String(e.message||e)}; });
-    }));
+    // ONE request, fanned out inside the gateway.
+    //
+    // The previous version fired Promise.all at /call/ — twenty concurrent
+    // browser requests. That looked concurrent and was not: each one queued on
+    // the plugin mutex for the whole Lua call, network wait included. Measured on
+    // the live gateway: 29s of wall clock for twenty probes, against 1.4–3.3s for
+    // a single probe with no contention. The page sat on "探测中…" for half a
+    // minute and read as broken.
+    //
+    // The comment that used to sit here claimed concurrency avoided a call budget
+    // overrun. It could not have: the requests were serialised, so the budget they
+    // overran was the very same one. It is a single call now because a
+    // request-per-item fan-out cannot be parallel from the client at all.
+    // Which sources are worth probing is decided on the SHARED VM, where the
+    // backoff state lives; only those get sharded.
+    //
+    // Asking each shard to decide for itself does not work: a shard has its own
+    // PLUGIN.state, so it cannot see the memo and a host that keeps failing is
+    // re-probed on every refresh forever. Measured before this split: fanout left
+    // the memo at 0 entries and every refresh cost the full 12s.
+    var plan = await pluginAPI.call("quota","due_sources",{sources:srcs});
+    // An EMPTY Lua table decodes as {} on the Go side, not [] — and {} is truthy,
+    // so `|| []` does not replace it. An empty object reaching out.concat()
+    // throws, and the whole refresh reports a failure. Array.isArray is the check
+    // that matches what the plugin actually returns.
+    var due = (plan && Array.isArray(plan.due)) ? plan.due : [];
+    var out = (plan && Array.isArray(plan.backing_off)) ? plan.backing_off : [];
+
+    if (due.length) {
+      out = out.concat(await pluginAPI.fanout("quota","collect_source",due.map(function(s){
+        return {sources:[s],name:s.name};
+      })));
+    }
     // Mirror the host→vendor map the plugin keeps: a source that answered is
     // worth probing first next time, and the page has everything it needs to
     // record that without another round trip.
@@ -749,6 +977,20 @@ PLUGIN.ui = {
       if(r && !r.skipped && !r.error && r.host && r.vendor) memo[r.host] = r.vendor;
     });
     window.__qaMemo = memo;
+    // Hand the whole pass back to the SHARED VM so the backoff has something to
+    // work from next time. Without this the shards' results die with their VMs:
+    // measured, the memo stayed at 0 entries and every refresh paid full price.
+    await pluginAPI.call("quota","memo_record",{results:(out||[]).map(function(r){
+      if(!r) return null;
+      // A host that is merely BACKING OFF has not just failed; recording it as a
+      // failure would grow the streak on a probe that never happened. Keyed on the
+      // flag rather than the wording, so translating the message cannot silently
+      // turn every skipped host into an ever-growing streak.
+      if(r.backing_off===true) return null;
+      if(r.error) return {host:r.host};
+      if(r.skipped) return {host:r.host};
+      return {host:r.host, vendor:r.vendor};
+    }).filter(Boolean)});
     }catch(e){
       if(btn) btn.disabled=false;
       if(sub) sub.textContent="探测失败: "+e.message;
@@ -756,7 +998,11 @@ PLUGIN.ui = {
     }
     var all=out||[];
     var real=all.filter(function(r){return !r.skipped;});
-    var skipped=all.filter(function(r){return r.skipped;}).map(function(r){return r.name;});
+    // The records, not their names — render() groups skipped sources by the
+    // backing_off flag, which a bare name cannot carry. This call site was left
+    // mapping to names while the other was fixed, and the mismatch showed up as
+    // a footer listing thirteen unnamed sources.
+    var skipped=all.filter(function(r){return r.skipped;});
     // real may already contain errors; collect_all keeps them for visibility.
     render(real, skipped);
     if(btn) btn.disabled=false;

internal/lua/quota_backoff_test.go

@@ -0,0 +1,650 @@
+package lua
+
+import (
+	"net/http"
+	"net/http/httptest"
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+	"time"
+)
+
+// The backoff lives in quota.lua — a shipped plugin, not Go — so these tests read
+// the file that actually ships (readShippedPlugin) rather than a copy that can
+// drift from it.
+
+// TestQuotaBackoffIsAppliedInSource guards the wiring.
+//
+// It exists because of a mutation that removed the call site entirely (`if false
+// then`) and every other test in this file still passed. The other tests cover
+// the decision FUNCTION and the streak RECORDING; neither exists in order to be
+// unused. A backoff nobody consults is indistinguishable from no backoff, and the
+// symptom is the original one — a dead host re-probed on every page open.
+func TestQuotaBackoffIsAppliedInSource(t *testing.T) {
+	src := readShippedPlugin(t, "quota.lua")
+	if !strings.Contains(src, "if backoffSeconds(seen, known.streak) then") {
+		t.Error("quota.lua never consults backoffSeconds for a host with a streak — " +
+			"the backoff is computed but not applied")
+	}
+	// The reason is required, but matched on the FLAG rather than the wording:
+	// the message is localised, and grouping by text would merge "backing off"
+	// into "exposes no balance endpoint" the moment someone translates it.
+	if strings.Count(src, "reason = \"连续探测失败 ") < 2 {
+		t.Error("both skip paths (collectSource and due_sources) must carry a reason; " +
+			"a silently skipped host looks exactly like a host with no balance endpoint")
+	}
+	if strings.Count(src, "backing_off = true") < 2 {
+		t.Error("both skip paths must set the explicit backing_off flag; the page " +
+			"groups skipped sources by it, not by the message text")
+	}
+	// The recording guard has to REPLACE the memo, not merge into it. The bug this
+	// guards was `if rep == nil and not HOST_MEMO[host]`, which is true only on a
+	// host's first failure — so the streak stayed pinned at 1 forever. Asserted on
+	// the source because the mutation that reintroduces it leaves the visible
+	// behaviour of the FIRST collect unchanged, and that is all the integration
+	// test can reach inside one process.
+	if !strings.Contains(src, "if rep == nil then") {
+		t.Error("the negative result is recorded under a guard that skips every " +
+			"failure after the first — the streak will never grow")
+	}
+	if strings.Contains(src, "and not HOST_MEMO[host] then") {
+		t.Error("the recording guard excludes a host that already has a memo, so " +
+			"consecutive failures are never counted")
+	}
+}
+
+// TestQuotaBackoffWindowIsCapped guards the failure mode an exponential backoff
+// has when it is not capped: a host that is down for a week stops being probed
+// and the page permanently misreports it as having no balance endpoint.
+func TestQuotaBackoffWindowIsCapped(t *testing.T) {
+	src := readShippedPlugin(t, "quota.lua")
+	for _, want := range []string{"BACKOFF_BASE", "BACKOFF_FACTOR", "BACKOFF_MAX"} {
+		if !strings.Contains(src, want) {
+			t.Errorf("quota.lua no longer declares %s — the backoff was weakened or removed", want)
+		}
+	}
+}
+
+// TestQuotaBackoffWindows walks the whole policy through the plugin's own
+// decision function: for each consecutive failure, how long the host must be left
+// alone, that it is still skipped halfway through that window, that it is retried
+// once the window passes, and that an absurdly old host is always due.
+//
+// Why the function rather than the running plugin: HOST_MEMO is a process-local
+// cache, restored on a process's first collect and written back afterwards, so
+// ageing its memo from Go changes nothing the plugin can see. Two earlier versions
+// of this file spent a long time on exactly that dead end, and a test that keeps
+// breaking for reasons unrelated to the code under test is worse than no test.
+//
+// The function is the plugin's own — reached through a shim that appends one
+// export — so changing quota.lua's arithmetic moves this test. It is not a
+// reimplementation that would agree with a broken plugin.
+func TestQuotaBackoffWindows(t *testing.T) {
+	ps := quotaWithDecision(t)
+
+	// streak, wait seconds. The last rows are at and past the ceiling.
+	for _, c := range []struct {
+		streak int
+		wait   int
+	}{
+		{1, 1800},
+		{2, 3600},
+		{3, 7200},
+		{4, 14400},
+		{5, 21600},
+		{9, 21600},
+		{20, 21600},
+	} {
+		if got := quotaSkip(t, ps, c.streak, c.wait/2); !got {
+			t.Logf("  streak=%d age=%d → %v", c.streak, c.wait/2, got)
+			t.Errorf("streak=%d: not skipping %ds into a %ds window — the wait is too short",
+				c.streak, c.wait/2, c.wait)
+		}
+		if quotaSkip(t, ps, c.streak, c.wait+60) {
+			t.Errorf("streak=%d: still skipping %ds after its %ds window expired — "+
+				"this host would never be retried again", c.streak, c.wait+60, c.wait)
+		}
+		if quotaSkip(t, ps, c.streak, 1<<30) {
+			t.Errorf("streak=%d: a host silent for decades is still being skipped", c.streak)
+		}
+	}
+}
+
+// TestQuotaBackoffReturnsOnlyBooleans pins a type rule that produced two silent
+// bugs.
+//
+// In Lua only nil and false are falsy, so every path of a predicate must return
+// one of those. This function had two paths returning numbers: `return 0` for
+// "no streak", which is truthy and made every caller skip a host that had never
+// been probed; and `return BACKOFF_MAX` on the cap path, also truthy, which made
+// the ceiling mean "never retried" — the exact failure the ceiling exists to
+// prevent.
+//
+// Both shipped far enough to be found by this assertion rather than by a user
+// staring at a page that had quietly stopped probing anything.
+func TestQuotaBackoffReturnsOnlyBooleans(t *testing.T) {
+	ps := quotaWithDecision(t)
+	for _, c := range []struct {
+		name string
+		age  int
+		strk interface{}
+	}{
+		{"no streak field (every memo written before the backoff existed)", 0, nil},
+		{"zero streak", 0, 0},
+		{"negative streak", 0, -1},
+		{"inside the window", 10, 1},
+		{"outside the window", 999999, 1},
+		{"at the ceiling", 10, 20},
+	} {
+		out, err := ps.Call("quota", "__decision", map[string]interface{}{
+			"at":     float64(time.Now().Unix() - int64(c.age)),
+			"streak": c.strk,
+		})
+		if err != nil {
+			t.Fatalf("%s: %v", c.name, err)
+		}
+		// A bare boolean is a valid answer: when backoffSeconds returns false the
+		// shim's table round-trips as just `false`, because a Lua table holding a
+		// single falsy field collapses on encode.
+		switch v := out.(type) {
+		case bool:
+			_ = v // a boolean is exactly what this asserts
+		case map[string]interface{}:
+			if _, ok := v["r"].(bool); !ok {
+				t.Errorf("%s: backoffSeconds returned %T (%v) — in Lua a number is "+
+					"TRUTHY, so this caller would skip the host", c.name, v["r"], v["r"])
+			}
+		default:
+			t.Errorf("%s: decision returned %T (%v); want a boolean", c.name, out, out)
+		}
+	}
+}
+
+// TestQuotaBackoffRecordsConsecutiveFailures covers the recording guard through
+// the running plugin, because that part IS observable from outside.
+//
+// The guard used to be `not HOST_MEMO[host]`, true only on a host's first
+// failure. After that the memo held a table, the guard was false, nothing was
+// written, and the streak stayed pinned at 1 forever — measured by reading back
+// {at:... streak:1} after four consecutive failures.
+//
+// Only two things are asserted, because only two are reachable in one process:
+// a first failure records streak 1, and a SKIP does not grow it. A skipped probe
+// is not a failure, and a skip that counted would let a never-probed host grow
+// without end and never be retried.
+func TestQuotaBackoffRecordsConsecutiveFailures(t *testing.T) {
+	srv := newCountingServer(t)
+	ps := newQuotaPlugins(t, srv.URL)
+
+	args := map[string]interface{}{
+		"sources": []interface{}{map[string]interface{}{
+			"name": "dead", "base_url": srv.URL + "/v1", "api_key": "k",
+		}},
+		"name": "dead",
+	}
+	streak := func() float64 {
+		st, _ := ps.State("quota").(map[string]interface{})
+		memo, _ := st["probe_memo"].(map[string]interface{})
+		for _, v := range memo {
+			if e, ok := v.(map[string]interface{}); ok {
+				if f, ok := e["streak"].(float64); ok {
+					return f
+				}
+			}
+		}
+		t.Fatalf("no negative memo recorded: %#v", memo)
+		return 0
+	}
+
+	if _, err := ps.Call("quota", "collect_source", args); err != nil {
+		t.Fatal(err)
+	}
+	if got := streak(); got != 1 {
+		t.Fatalf("streak after the first failure = %.0f, want 1", got)
+	}
+
+	before := srv.Hits()
+	if _, err := ps.Call("quota", "collect_source", args); err != nil {
+		t.Fatal(err)
+	}
+	if served := srv.Hits() - before; served != 0 {
+		t.Errorf("a host inside its backoff window still made %d requests — the "+
+			"whole feature is to not make any", served)
+	}
+	if got := streak(); got != 1 {
+		t.Errorf("a SKIP changed the streak to %.0f — a skipped probe is not a failure", got)
+	}
+}
+
+// TestQuotaMemoKeyIsTheHost pins the key the memo is written and read under.
+//
+// This is the foundation the whole memoisation rests on, and it was broken in the
+// most total way possible: hostOf was written with string.gmatch and a capture
+// group, gmatch yields whole matches rather than captures, and with `^` the
+// iterator never advanced — so it returned nil and every lookup and every write
+// went to HOST_MEMO[nil]. The memo therefore never returned anything for any
+// host, and the negative memo, the TTL and the backoff on top of them were all
+// dead code that looked alive.
+//
+// Measured on this gateway before the fix: hostOf(key) matched 0 of the 19
+// entries in production state.
+//
+// The check is on the shipped source plus its runtime behaviour, because the
+// failure mode is invisible from outside — a memo that never matches anything
+// behaves exactly like a memo that has nothing to remember.
+func TestQuotaMemoKeyIsTheHost(t *testing.T) {
+	src := readShippedPlugin(t, "quota.lua")
+	if !strings.Contains(src, "string.match(u,") {
+		t.Error("hostOf does not use string.match — string.gmatch yields whole " +
+			"matches, so the capture group is ignored and the function returns nil")
+	}
+
+	ps := quotaWithDecision(t)
+	for _, c := range []struct{ in, want string }{
+		{"https://qijiar.top/v1", "qijiar.top"},
+		{"https://api.deepseek.com", "api.deepseek.com"},
+		{"http://127.0.0.1:19900", "127.0.0.1:19900"},
+		{"https://s2a.frankhan.top:14514/v1", "s2a.frankhan.top:14514"},
+	} {
+		out, err := ps.Call("quota", "__hostof", map[string]interface{}{"u": c.in})
+		if err != nil {
+			t.Fatalf("hostOf(%s): %v", c.in, err)
+		}
+		mm, _ := out.(map[string]interface{})
+		got, _ := mm["h"].(string)
+		if got != c.want {
+			t.Errorf("hostOf(%q) = %q, want %q", c.in, got, c.want)
+		}
+		if got == "" {
+			t.Errorf("hostOf(%q) returned nothing — every lookup would be "+
+				"HOST_MEMO[nil]", c.in)
+		}
+	}
+
+	// The same host under two paths must produce the SAME key, which is the whole
+	// reason the memo is keyed this way.
+	out, err := ps.Call("quota", "__hostof", map[string]interface{}{"u": "https://x.example/a/v1"})
+	if err != nil {
+		t.Fatal(err)
+	}
+	out2, err := ps.Call("quota", "__hostof", map[string]interface{}{"u": "https://x.example/b/v1"})
+	if err != nil {
+		t.Fatal(err)
+	}
+	m1, _ := out.(map[string]interface{})
+	m2, _ := out2.(map[string]interface{})
+	if m1["h"] != m2["h"] {
+		t.Errorf("two paths on one host produced different keys: %v and %v — "+
+			"they would never share a memo", m1["h"], m2["h"])
+	}
+}
+
+// --- helpers ---
+
+// readShippedPlugin returns the bundled source, so the tests exercise the file
+// that actually ships rather than a copy that can drift from it.
+func readShippedPlugin(t *testing.T, name string) string {
+	t.Helper()
+	b, err := bundledAdapters.ReadFile("plugins/" + name)
+	if err != nil {
+		t.Fatalf("read embedded %s: %v", name, err)
+	}
+	return string(b)
+}
+
+// quotaWithDecision loads quota.lua with its local backoffSeconds exposed as an
+// export. The plugin's own exports list is appended to, so the export allowlist is
+// satisfied exactly as it is in production.
+func quotaWithDecision(t *testing.T) *Plugins {
+	t.Helper()
+	src := readShippedPlugin(t, "quota.lua")
+	shim := strings.Replace(src, "return PLUGIN",
+		"PLUGIN.exports[#PLUGIN.exports+1] = '__decision'\n"+
+			// Wrapped, not exported raw: backoffSeconds takes (at, streak)
+			// positionally, and the test calls it with a table because that is the
+			// shape a Go caller can pass. Exporting the raw function made every
+			// assertion read `at = <the argument table>, streak = nil`, so the
+			// function took its "no streak" path and returned false for every input
+			// — which looked exactly like a backoff that never fires.
+			"PLUGIN.__decision = function(q) return { r = backoffSeconds(q.at, q.streak) } end\n"+
+			"PLUGIN.exports[#PLUGIN.exports+1] = '__hostof'\n"+
+			"PLUGIN.__hostof = function(q) return { h = hostOf(q.u) } end\n"+
+			"return PLUGIN", 1)
+	if shim == src {
+		t.Fatal("could not inject the decision shim — quota.lua no longer ends with 'return PLUGIN'")
+	}
+	return newQuotaPluginsWith(t, shim, "127.0.0.1")
+}
+
+// quotaSkip asks the plugin whether a host that failed `streak` times and was
+// last heard from `age` seconds ago should be skipped.
+func quotaSkip(t *testing.T, ps *Plugins, streak, age int) bool {
+	out, err := ps.Call("quota", "__decision", map[string]interface{}{
+		"at":     float64(time.Now().Unix() - int64(age)),
+		"streak": float64(streak),
+	})
+	if err != nil {
+		t.Fatalf("decision: %v", err)
+	}
+	// A single-field table collapses to the bare value on encode, so both shapes
+	// mean the same thing here.
+	switch v := out.(type) {
+	case bool:
+		return v
+	case map[string]interface{}:
+		b, ok := v["r"].(bool)
+		if !ok {
+			t.Fatalf("decision returned %T, not a boolean: %#v", v["r"], v)
+		}
+		return b
+	default:
+		t.Fatalf("decision returned %T (%v), not a boolean", out, out)
+		return false
+	}
+}
+
+// newQuotaPlugins boots the shipped quota.lua against a temp plugin dir.
+func newQuotaPlugins(t *testing.T, baseURL string) *Plugins {
+	t.Helper()
+	return newQuotaPluginsWith(t, readShippedPlugin(t, "quota.lua"), baseURL)
+}
+
+func newQuotaPluginsWith(t *testing.T, src, baseURL string) *Plugins {
+	t.Helper()
+	dir := t.TempDir()
+	pluginDir := filepath.Join(dir, "plugins")
+	if err := os.MkdirAll(pluginDir, 0o755); err != nil {
+		t.Fatal(err)
+	}
+	vm := NewVM(filepath.Join(dir, "adapters"))
+	if err := vm.Start(); err != nil {
+		t.Fatal(err)
+	}
+	t.Cleanup(vm.Stop)
+	ps := NewPlugins(vm, pluginDir)
+	t.Cleanup(ps.Close)
+	ps.SetOutboundHosts([]string{"127.0.0.1"})
+	if err := os.WriteFile(filepath.Join(pluginDir, "quota.lua"), []byte(src), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	if err := ps.LoadDir(); err != nil {
+		t.Fatal(err)
+	}
+	p := ps.find("quota")
+	if p == nil {
+		t.Fatal("quota plugin did not load")
+	}
+	if p.LoadError != "" {
+		t.Fatalf("quota plugin failed to compile: %s", p.LoadError)
+	}
+	_ = baseURL
+	return ps
+}
+
+// countingServer answers 404 on every path and counts what it is asked, so a test
+// can tell "the plugin skipped" (no requests) from "the plugin probed and found
+// nothing" (requests, no data).
+//
+// It used to answer 200 with a non-balance body, and the commandcode adapter
+// accepted that — so the host was memoised as a SUCCESS and the negative path was
+// never reached. The test then failed for the right reason, and would have
+// passed for the wrong one had the assertion been looser.
+type countingServer struct {
+	*httptest.Server
+	hits int
+}
+
+func (c *countingServer) Hits() int { return c.hits }
+
+func newCountingServer(t *testing.T) *countingServer {
+	t.Helper()
+	cs := &countingServer{}
+	cs.Server = httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		cs.hits++
+		w.WriteHeader(http.StatusNotFound)
+		_, _ = w.Write([]byte(`{"error":"not found"}`))
+	}))
+	t.Cleanup(cs.Close)
+	return cs
+}
+
+// TestQuotaDueSourcesAndMemoRecordCoverTheShardGap pins the split that makes the
+// backoff work at all under sharding.
+//
+// The conflict, measured on a live 20-source gateway:
+//   - probing is network-bound and Plugins.Call holds the plugin mutex for the
+//     whole Lua call, so /call/ serialises (29s)
+//   - so the page uses ShardCall — and a shard has its own PLUGIN.state
+//   - and the backoff lives in PLUGIN.state: fanout left the memo at 0 entries,
+//     so dead hosts were re-probed forever
+//
+// The resolution is to put the DECISION and the RECORDING back on the shared VM
+// (microseconds, no network) and shard only the network wait. These two exports
+// are that seam, and without them the fanout path silently defeats the backoff.
+func TestQuotaDueSourcesAndMemoRecordCoverTheShardGap(t *testing.T) {
+	src := readShippedPlugin(t, "quota.lua")
+	for _, want := range []string{"due_sources", "memo_record"} {
+		if !strings.Contains(src, want) {
+			t.Errorf("quota.lua does not export %s — a sharded pass cannot see the "+
+				"backoff state, nor leave any behind", want)
+		}
+		if !strings.Contains(src, `"`+want+`"`) {
+			t.Errorf("%s is not listed in PLUGIN.exports, so the page cannot call it", want)
+		}
+	}
+	// due_sources must not probe. If it did, the whole point of asking it first
+	// would be lost: the cost would be paid before any shard runs.
+	if strings.Contains(src, "due_sources = function(payload)") &&
+		strings.Contains(src, "due_sources") {
+		// The body is small; check it does not call collectSource.
+		i := strings.Index(src, "due_sources = function(payload)")
+		j := strings.Index(src[i:], "\nend\n")
+		if j > 0 {
+			body := src[i : i+j]
+			if strings.Contains(body, "collectSource(") {
+				t.Error("due_sources calls collectSource — the filter must make no " +
+					"network request, or it costs the same as probing everything")
+			}
+		}
+	}
+}
+
+// TestQuotaMemoRecordGrowsTheStreakFromShards covers the recording half: a host
+// that failed inside a shard must end up in the SHARED memo with a streak, or the
+// next refresh has nothing to back off from.
+func TestQuotaMemoRecordGrowsTheStreakFromShards(t *testing.T) {
+	srv := newCountingServer(t)
+	ps := newQuotaPlugins(t, srv.URL)
+	host := hostOnly(srv.URL)
+
+	read := func() map[string]interface{} {
+		st, _ := ps.State("quota").(map[string]interface{})
+		memo, _ := st["probe_memo"].(map[string]interface{})
+		return memo
+	}
+
+	// A shard reports a host that answered with nothing.
+	for i := 1; i <= 3; i++ {
+		out, err := ps.Call("quota", "memo_record", map[string]interface{}{
+			"results": []interface{}{
+				map[string]interface{}{"host": srv.URL + "/v1"},
+			},
+		})
+		if err != nil {
+			t.Fatalf("memo_record %d: %v", i, err)
+		}
+		if out != true {
+			t.Fatalf("memo_record returned %#v", out)
+		}
+	}
+	memo := read()
+	e, ok := memo[host].(map[string]interface{})
+	if !ok {
+		t.Fatalf("memo has no entry for %s: %#v", host, memo)
+	}
+	if got, _ := e["streak"].(float64); got != 3 {
+		t.Errorf("streak after three shard failures = %.0f, want 3", got)
+	}
+
+	// A host that answers is recorded as a positive memo, which resets the streak.
+	if _, err := ps.Call("quota", "memo_record", map[string]interface{}{
+		"results": []interface{}{
+			map[string]interface{}{"host": srv.URL + "/v1", "vendor": "deepseek"},
+		},
+	}); err != nil {
+		t.Fatal(err)
+	}
+	if got := read()[host]; got != "deepseek" {
+		t.Errorf("a host that answered is memoised as %#v, want the vendor name — "+
+			"the streak must not survive a success", got)
+	}
+}
+
+// TestQuotaDueSourcesSkipsBackedOffHosts is the behaviour the page depends on:
+// the filter must keep a backed-off host out of the sharded set.
+func TestQuotaDueSourcesSkipsBackedOffHosts(t *testing.T) {
+	srv := newCountingServer(t)
+	ps := newQuotaPlugins(t, srv.URL)
+	host := hostOnly(srv.URL)
+
+	// One failure, which is enough to put the host inside its window.
+	if _, err := ps.Call("quota", "memo_record", map[string]interface{}{
+		"results": []interface{}{map[string]interface{}{"host": srv.URL + "/v1"}},
+	}); err != nil {
+		t.Fatal(err)
+	}
+
+	// A SECOND host, so the two cases are genuinely independent. The memo is keyed
+	// by host, so two paths on one host are one entry — using the same server here
+	// would have asserted that a per-host memo behaves per-path.
+	other := newCountingServer(t)
+	srcs := []interface{}{
+		map[string]interface{}{"name": "dead", "base_url": srv.URL + "/v1", "api_key": "k"},
+		map[string]interface{}{"name": "fresh", "base_url": other.URL + "/v1", "api_key": "k"},
+	}
+	out, err := ps.Call("quota", "due_sources", map[string]interface{}{"sources": srcs})
+	if err != nil {
+		t.Fatal(err)
+	}
+	m, ok := out.(map[string]interface{})
+	if !ok {
+		t.Fatalf("due_sources returned %#v", out)
+	}
+	due, _ := m["due"].([]interface{})
+	backing, _ := m["backing_off"].([]interface{})
+	if len(backing) != 1 {
+		t.Errorf("backing_off = %#v, want exactly the host that failed", backing)
+	} else if b, _ := backing[0].(map[string]interface{}); b["host"] != host {
+		t.Errorf("backing off %v, want %s — the key must be the host, or it will "+
+			"never match on the next refresh", b["host"], host)
+	}
+	if len(due) != 1 {
+		t.Errorf("due = %#v, want only the unprobed host", due)
+	} else if d, _ := due[0].(map[string]interface{}); d["name"] != "fresh" {
+		t.Errorf("due contains %v, want fresh", d["name"])
+	}
+
+	// And it must make no network request — the whole saving depends on it.
+	if n := srv.Hits(); n != 0 {
+		t.Errorf("due_sources made %d requests; it must decide without probing", n)
+	}
+}
+
+// hostOnly mirrors the plugin's own hostOf. Used only to CHECK what the plugin
+// wrote — seeding a memo under a different key produces a memo the plugin never
+// looks up, which cost this file a couple of rounds before it was noticed.
+func hostOnly(baseURL string) string {
+	u := strings.TrimPrefix(strings.TrimPrefix(baseURL, "https://"), "http://")
+	if i := strings.Index(u, "/"); i >= 0 {
+		u = u[:i]
+	}
+	return u
+}
+
+// TestQuotaBackoffRecordIsSkippedAndFlagged pins the shape the page depends on.
+//
+// Two things went wrong in the same field, and both were silent:
+//
+//  1. The back-off record had no `skipped` flag, so the page's
+//     `filter(r => !r.skipped)` let it through to the card renderer. That
+//     renderer matches on known fields, and a record carrying only a reason
+//     matches none — so it fell through to a default that said "有数据".
+//     Thirteen dead upstreams were rendered as though they had answered.
+//  2. Once it did carry `skipped`, the page grouped skipped sources by reading
+//     the REASON TEXT. Translating that string would have silently merged
+//     "backing off" into "exposes no balance endpoint" — a different claim
+//     about the upstream. The grouping keys on an explicit flag.
+func TestQuotaBackoffRecordIsSkippedAndFlagged(t *testing.T) {
+	srv := newCountingServer(t)
+	ps := newQuotaPlugins(t, srv.URL)
+
+	// One failure, enough to put the host inside its window.
+	if _, err := ps.Call("quota", "memo_record", map[string]interface{}{
+		"results": []interface{}{map[string]interface{}{"host": srv.URL + "/v1"}},
+	}); err != nil {
+		t.Fatal(err)
+	}
+
+	out, err := ps.Call("quota", "due_sources", map[string]interface{}{
+		"sources": []interface{}{map[string]interface{}{
+			"name": "dead", "base_url": srv.URL + "/v1", "api_key": "k",
+		}},
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	m, _ := out.(map[string]interface{})
+	backing, _ := m["backing_off"].([]interface{})
+	if len(backing) != 1 {
+		t.Fatalf("backing_off = %#v, want one record", m["backing_off"])
+	}
+	rec, _ := backing[0].(map[string]interface{})
+	if rec["skipped"] != true {
+		t.Errorf("back-off record has skipped=%v — the page filters cards on this "+
+			"field, and without it the record is rendered as a source with data",
+			rec["skipped"])
+	}
+	if rec["backing_off"] != true {
+		t.Errorf("back-off record has no explicit backing_off flag; the page groups " +
+			"by it instead of by reason text, so translating the message cannot " +
+			"merge this with 'no balance endpoint'")
+	}
+	if s, _ := rec["reason"].(string); s == "" {
+		t.Error("back-off record carries no reason; the card renders empty")
+	}
+}
+
+// TestQuotaRenderCallSitesPassRecordsNotNames guards a bug that only appeared
+// because ONE of two identical call sites was fixed.
+//
+// render() groups skipped sources by the backing_off flag. A call site that maps
+// the records to their names first ("skipped.map(r => r.name)") therefore hands
+// render an array of strings: the flag is gone, every entry falls into the
+// "exposes no balance endpoint" bucket, and the footer lists N unnamed sources.
+//
+// The first fix changed only the refresh path. The snapshot path kept mapping to
+// names, and the symptom was a footer reading "未列出 13 个源：、、、、…" — which
+// reads like a rendering bug rather than a contract mismatch between two call
+// sites of one function.
+func TestQuotaRenderCallSitesPassRecordsNotNames(t *testing.T) {
+	src := readShippedPlugin(t, "quota.lua")
+	if strings.Contains(src, ".map(function(r){ return r.name; })") ||
+		strings.Contains(src, ".map(function(r){return r.name;})") {
+		// A benign occurrence would be fine, but there is no other reason for a
+		// skipped list to be reduced to names, so treat any as the bug.
+		for _, line := range strings.Split(src, "\n") {
+			if strings.Contains(line, "skipped") && strings.Contains(line, ".map(") &&
+				strings.Contains(line, "r.name") {
+				t.Errorf("a skipped list is mapped to names before reaching render(): %s\n"+
+					"render() groups on the backing_off flag, which a name cannot carry",
+					strings.TrimSpace(line))
+			}
+		}
+	}
+	// Both call sites must exist and both must hand over records.
+	if n := strings.Count(src, ".filter(function(r){return r.skipped;})") +
+		strings.Count(src, ".filter(function(r){ return r && r.skipped; })"); n < 2 {
+		t.Errorf("found %d skipped-list filters, want 2 (snapshot and refresh paths)", n)
+	}
+}

internal/provider/model_kind_test.go

@@ -0,0 +1,107 @@
+package provider
+
+import (
+	"encoding/json"
+	"testing"
+
+	"llmsproxy/internal/config"
+)
+
+// TestSourceStatus_CarriesModelKind is the fix for a UI bug where the chat
+// test page listed image models next to chat ones.
+//
+// The root cause was structural, not a missing filter: SourceStatus.Models was
+// a flat []string, so by the time the UI had the list there was nothing left to
+// tell "gpt-image-2" from "claude-sonnet-5.5". Picking the former sent the
+// request to /api/chat, which cannot serve it, and the failure said nothing
+// about the real reason.
+//
+// Models stays a []string for callers that only need names; ModelKinds is the
+// additive field that carries the missing dimension.
+func TestSourceStatus_CarriesModelKind(t *testing.T) {
+	src := config.Source{
+		Name:     "vela",
+		BaseURL:  "https://example.invalid",
+		Adapter:  "openai",
+		APIKey:   "sk-test",
+		Endpoint: "/chat/completions",
+		Models: []config.Model{
+			{ID: "claude-sonnet-5.5"},          // kind omitted -> chat default
+			{ID: "gpt-image-2", Kind: "image"}, // image
+			{ID: "seedream-5.0", Kind: "image"},
+		},
+	}
+	p := newTestProvider(t, src)
+
+	kinds := p.ModelsWithKind()
+	if len(kinds) != 3 {
+		t.Fatalf("want 3 entries, got %d", len(kinds))
+	}
+
+	// Order must mirror Models() so the UI can zip the two lists.
+	names := p.Models()
+	for i := range kinds {
+		if kinds[i].ID != names[i] {
+			t.Errorf("index %d: ModelKinds has %q but Models has %q — the lists must stay aligned",
+				i, kinds[i].ID, names[i])
+		}
+	}
+
+	want := map[string]string{
+		"claude-sonnet-5.5": "chat",
+		"gpt-image-2":       "image",
+		"seedream-5.0":      "image",
+	}
+	for _, k := range kinds {
+		if got := want[k.ID]; k.Kind != got {
+			t.Errorf("%s: kind = %q, want %q", k.ID, k.Kind, got)
+		}
+	}
+}
+
+// TestModelKinds_CountsImageModelsInARealCatalog: the number of image models is
+// what decides whether the UI needs the split at all, so assert against a
+// catalog with both kinds rather than a hand-picked pair.
+func TestModelKinds_CountsImageModelsInARealCatalog(t *testing.T) {
+	src := config.Source{
+		Name: "vela", BaseURL: "https://example.invalid", Adapter: "openai", APIKey: "k",
+		Models: []config.Model{
+			{ID: "a"}, {ID: "b"}, {ID: "c"}, {ID: "d"},
+			{ID: "img1", Kind: "image"}, {ID: "img2", Kind: "image"},
+		},
+	}
+	p := newTestProvider(t, src)
+	var images int
+	for _, k := range p.ModelsWithKind() {
+		if k.Kind == "image" {
+			images++
+		}
+	}
+	if images != 2 {
+		t.Errorf("image models = %d, want 2", images)
+	}
+}
+
+// TestSourceStatusJSON_ModelKindsOmittedWhenEmpty pins the backward-compat
+// promise: an older UI parsing this JSON must not break on a new field, and a
+// source with no model list must not gain a "model_kinds": null.
+func TestSourceStatusJSON_ModelKindsOmittedWhenEmpty(t *testing.T) {
+	b, err := json.Marshal(SourceStatus{Name: "x", Models: []string{"a"}})
+	if err != nil {
+		t.Fatal(err)
+	}
+	var m map[string]interface{}
+	if err := json.Unmarshal(b, &m); err != nil {
+		t.Fatal(err)
+	}
+	if _, present := m["model_kinds"]; present {
+		t.Errorf("model_kinds present when empty: %s — should be omitted for older UIs", b)
+	}
+	if _, present := m["models"]; !present {
+		t.Errorf("models field disappeared: %s", b)
+	}
+}
+
+// newTestProvider (defined in provider_test.go) spins up a real Lua VM, which
+// is heavier than these tests need, but reusing it keeps one construction path
+// for the whole package.

internal/provider/provider.go

@@ -451,6 +451,25 @@ func (p *Provider) Models() []string {
 	return out
 }
 
+// ModelsWithKind returns the same ids as Models but paired with each model's
+// kind, in the same order.
+//
+// Models() cannot answer "which of these are image models", and that question
+// has to be answerable: the chat test page lists models from one source list,
+// and an image model offered there fails on every call because /api/chat is the
+// chat endpoint.
+func (p *Provider) ModelsWithKind() []ModelKindInfo {
+	out := make([]ModelKindInfo, 0, len(p.cfg.Models))
+	for _, m := range p.cfg.Models {
+		kind := m.Kind
+		if kind == "" {
+			kind = "chat" // config.Model documents "" as the chat default
+		}
+		out = append(out, ModelKindInfo{ID: m.ID, Kind: kind})
+	}
+	return out
+}
+
 // ModelByID returns the model definition if owned by this source.
 func (p *Provider) ModelByID(id string) *config.Model {
 	for i := range p.cfg.Models {

internal/provider/registry.go

@@ -196,12 +196,27 @@ func (r *Registry) ProviderForSlot(model, source string) *Provider {
 	return nil
 }
 
+// ModelKindInfo carries the per-model fields the chat test UI needs to tell a
+// chat model from an image one.
+type ModelKindInfo struct {
+	ID   string `json:"id"`
+	Kind string `json:"kind"` // "chat" (default) | "image"
+}
+
 // ModelStatus is a web-UI friendly snapshot per source.
 type SourceStatus struct {
 	Name          string   `json:"name"`
 	Adapter       string   `json:"adapter"`
 	BaseURL       string   `json:"base_url"`
 	Models        []string `json:"models"`
+	// ModelKinds parallels Models: Models[i] is ModelKinds[i].ID.
+	//
+	// It exists because a flat []string cannot answer "is gpt-image-2
+	// reachable from /api/chat?" — it cannot, and offering it there produced
+	// a guaranteed failure with no explanation. Models stays unchanged for
+	// callers that only need names; this is purely additive and omitted when
+	// empty, so an older UI keeps working.
+	ModelKinds    []ModelKindInfo `json:"model_kinds,omitempty"`
 	Available     bool     `json:"available"`
 	Healthy       bool     `json:"healthy"`
 	MaxConcurrent int      `json:"max_concurrent"`
@@ -259,6 +274,7 @@ func (r *Registry) Status() []SourceStatus {
 			Adapter:       p.Adapter(),
 			BaseURL:       p.Config().BaseURL,
 			Models:        p.Models(),
+			ModelKinds:    p.ModelsWithKind(),
 			Available:     p.Available(),
 			Healthy:       p.Available(),
 			MaxConcurrent: p.MaxConcurrent(),

packaging/upload-release.sh

@@ -0,0 +1,71 @@
+#!/usr/bin/env bash
+# Upload release assets to a GitCode release.
+#
+# The endpoint shapes are asymmetric and both differ from the "obvious" guess:
+#
+#   GET  /releases/tags/{tag}            → read a release
+#   GET  /releases/{tag}/upload_url      → get a presigned PUT   (NO "tags/")
+#   PUT  {presigned url}                 → upload, REPLAYING the headers verbatim
+#
+# Two mistakes cost real time on v1.9.5 before this script existed:
+#
+#   1. POSTing straight to /releases/{tag}/assets returns 404 — that endpoint
+#      does not exist. It reads as "authentication failed" if you only check
+#      the status code.
+#   2. PUTting without the headers fails with an opaque storage error, not a
+#      permission one, so it looks like a size limit or a corrupt file.
+#
+# The x-obs-* headers are part of the signature: dropping any one of them turns
+# the request into an unsigned one.
+set -uo pipefail
+
+TOKEN="$(tr -d '\n\r' < /tmp/gc_token)"
+REPO="JianFeeeee/ModelRouter"
+TAG="${1:?usage: upload-release.sh <tag> <file>...}"
+shift
+
+API="https://gitcode.com/api/v5/repos/${REPO}"
+
+for f in "$@"; do
+  name="$(basename "$f")"
+  size="$(stat -c%s "$f")"
+  printf '  %-46s %10s bytes  ' "$name" "$size"
+
+  # 1) presigned upload location. The file name must be URL-encoded; the v1.9.4
+  #    assets show "~" surviving in a stored name, so encoding is per-component.
+  enc="$(python3 -c 'import urllib.parse,sys; print(urllib.parse.quote(sys.argv[1]))' "$name")"
+  resp="$(curl -sS -m 60 -H "Authorization: Bearer ${TOKEN}" \
+    "${API}/releases/${TAG}/upload_url?file_name=${enc}" 2>&1)"
+  url="$(printf '%s' "$resp" | python3 -c 'import json,sys
+try:
+    d = json.load(sys.stdin)
+    print(d.get("url") or d.get("uploadUrl") or "")
+except Exception:
+    print("")' 2>/dev/null)"
+  if [ -z "$url" ]; then
+    echo "✗ 取得上传地址失败: ${resp:0:140}"
+    continue
+  fi
+
+  # 2) PUT with the headers replayed exactly as returned.
+  code="$(printf '%s' "$resp" | python3 -c 'import json,sys
+try:
+    print(json.dumps(json.load(sys.stdin).get("headers") or {}))
+except Exception:
+    print("{}")' 2>/dev/null | python3 -c '
+import json, subprocess, sys
+headers = json.load(sys.stdin)
+args = ["curl", "-sS", "-o", "/dev/null", "-w", "%{http_code}", "-X", "PUT"]
+for k, v in headers.items():
+    args += ["-H", f"{k}: {v}"]
+args += ["--data-binary", "@" + sys.argv[1], sys.argv[2]]
+print(subprocess.run(args, capture_output=True, text=True).stdout)
+' "$f" "$url" 2>/dev/null)"
+
+  if [ "$code" = "200" ] || [ "$code" = "201" ] || [ "$code" = "204" ]; then
+    echo "✓ HTTP=$code"
+  else
+    echo "✗ HTTP=$code"
+  fi
+  sleep 1
+done