chore(gui): 版本号 1.9.5 → 1.9.6

## 症状

生产 Quota 页面上 17 个源里 13 个显示「有数据」徽标 —— 而它们实际上什么都
没返回，只是刚刚被退避跳过。

## 根因（三层，逐层都静默）

**① 退避记录缺 `skipped` 标记**
`due_sources` 与 `collectSource` 返回退避记录时没有 `skipped = true`，页面的
`filter(r => !r.skipped)` 就把它们当正常记录放进卡片。卡片渲染匹配已知字段，
一条只有 `reason` 的记录什么都匹配不上，于是落到兜底分支。

**② 兜底分支文案是「有数据」**
那是个断言 —— 它在说「探测有应答」。落到这里的记录其实是：一条退避通知、
一种我们不认识的错误形状、或供应商改了响应格式。把「没有证据」说成「有数据」，
十三个死上游就被渲染成健康的了。改成「未探测」。

**③ 退避与「没有余额接口」被混为一谈**
页脚原本统一说「未暴露余额接口」。这是对上游的另一种断言，而且对死上游是错的
—— 它有接口，只是暂时没探到。

## 「只改了一处」这个坑

`render()` 改成按 `backing_off` 标记分组之后，两个调用点都必须传**记录**而不是
名字。我先改了刷新路径就部署，结果是页脚显示「未列出 13 个源：、、、……」——
症状看起来像渲染 bug，实际是两个调用点与同一个函数之间的契约不一致。

快照路径那行 `.map(function(r){ return r.name; })` 是后来才补的。
判据 `TestQuotaRenderCallSitesPassRecordsNotNames` 现在要求两处都在。

## 顺带修掉的两个真 bug

**空 Lua 表在 Go 侧解码成 `{}` 而不是 `[]`**
`plan.backing_off || []` 里 `{}` 是 truthy，空表会原样进入 `out.concat()` 并抛错，
整次刷新报失败。改用 `Array.isArray` —— 它与插件实际返回的类型匹配。

**分组靠文案匹配是个定时炸弹**
原先用 `reason.indexOf("backing off") === 0` 判断退避。理由字符串一旦本地化，
退避就会被并进「未暴露余额接口」。改用显式 `backing_off: true` 标记，
判据里也改成按标记匹配 —— 变异把文案改回英文立刻变红。

## 判据

新增两条：
- `TestQuotaBackoffRecordIsSkippedAndFlagged` —— 退避记录必须带 `skipped` 和
  `backing_off` 标记，且必须有 reason（否则卡片是空的）
- `TestQuotaRenderCallSitesPassRecordsNotNames` —— 两个调用点都不得把跳过列表
  映射成名字

变异自证四次全部变红：去掉 `skipped`、去掉 `backing_off`、文案改回英文、
快照路径改回 map-to-name。

## 生产实测（浏览器真实点击）

```
第1轮 1694ms: 4 个源 · 0 个读取失败
        13 个源正在退避（连续探测失败，等下次重试）：trae、qijiar、toter、…
第2轮 2488ms: 4 个源 · 0 个读取失败
        13 个源正在退避（…）：trae、qijiar、toter、…
```

页面现在只列 4 个真有数据的源，13 个退避中的归到页脚并说明原因。

Co-Authored-By: ModelRouter <noreply@modelrouter.dev>

cmd/gui/package.json

@@ -1,6 +1,6 @@
 {
   "name": "modelrouter-gui",
-  "version": "1.9.3",
+  "version": "1.9.6",
   "description": "ModelRouter Desktop — embedded ModelRouter core with tray",
   "author": "ModelRouter",
   "main": "main.js",

internal/billing/compile.go

@@ -195,6 +195,36 @@ func applyRule(prices map[string]interface{}, rule *config.BillingRule, srcName
 			if t.CacheDiscount != nil {
 				entry["cache_discount"] = *t.CacheDiscount
 			}
+			// reasoning is optional and stays ABSENT (not nil, not 0) when the
+			// rule does not set it, which is what makes the billing plugin skip
+			// the repricing branch entirely. Writing an explicit 0 here would
+			// mean "thinking is free" and silently under-bill every request
+			// from a provider that bills it like ordinary output.
+			if strings.TrimSpace(t.Reasoning) != "" {
+				rp, err := usdPerM(t.Reasoning)
+				if err != nil {
+					return fmt.Errorf("model %q reasoning: %w", modelID, err)
+				}
+				entry["reasoning_price"] = rp
+			}
+			if lc := t.LongContext; lc != nil {
+				if lc.Above <= 0 {
+					return fmt.Errorf("model %q long_context: above must be > 0", modelID)
+				}
+				above, err := longContextRate(lc, prompt)
+				if err != nil {
+					return fmt.Errorf("model %q long_context: %w", modelID, err)
+				}
+				entry["long_context"] = map[string]interface{}{
+					"above": lc.Above,
+					// Marginal, not blended. The plugin charges the tokens past the
+					// threshold at this rate and the rest at the base rate, so the
+					// threshold composes with a cache discount (which applies to the
+					// cached portion) instead of forcing the plugin to guess how
+					// the two interact.
+					"prompt": above,
+				}
+			}
 			if rule.Peak != nil {
 				entry["peak"] = compilePeak(rule.Peak)
 			}
@@ -262,3 +292,32 @@ func compilePeak(p *config.BillingPeak) map[string]interface{} {
 	}
 	return out
 }
+
+// longContextRate resolves the marginal per-token rate for the tokens ABOVE a
+// threshold, given the model's base per-million prompt price.
+//
+// Two ways to express it are supported because providers publish both: an
+// absolute rate (OpenAI quotes \$3.00/M past 200k) and a multiple (DeepSeek
+// quotes exactly 2x past 128k). Neither may be missing: a step with no rate is a
+// silent no-op, and an operator who wrote one meant something.
+func longContextRate(lc *config.BillingLongContext, basePerToken float64) (float64, error) {
+	hasPrompt := strings.TrimSpace(lc.Prompt) != ""
+	hasMult := lc.Multiplier > 0
+	switch {
+	case hasPrompt && hasMult:
+		p, err := usdPerM(lc.Prompt)
+		if err != nil {
+			return 0, err
+		}
+		// Both given: the absolute rate wins, because it is the one the provider
+		// actually quotes. Silently multiplying instead would produce a number
+		// the operator never wrote.
+		return p, nil
+	case hasPrompt:
+		return usdPerM(lc.Prompt)
+	case hasMult:
+		return basePerToken * lc.Multiplier, nil
+	default:
+		return 0, fmt.Errorf("needs either prompt or multiplier")
+	}
+}

internal/billing/wildcard_test.go

@@ -86,3 +86,129 @@ func TestSpecificURLTokenRuleStillMarksItsSource(t *testing.T) {
 		t.Error("an unrelated source was marked priced by a specific-url rule")
 	}
 }
+
+// TestReasoningPriceIsOmittedUnlessSet pins the absence convention for the
+// thinking-token rate.
+//
+// A rule that says nothing about reasoning must produce a price entry with NO
+// reasoning_price key, so the billing plugin skips the repricing branch and
+// bills thinking at the ordinary output rate — which is what every provider
+// actually charges. Writing an explicit 0 instead would read as "thinking is
+// free" and silently under-bill every request from that model.
+func TestReasoningPriceIsOmittedUnlessSet(t *testing.T) {
+	sources := []config.Source{{Name: "up", BaseURL: "https://up.example/v1"}}
+
+	// Case 1: no reasoning declared → the key must not appear.
+	p1 := &config.BillingProfile{ID: "p", Rules: []config.BillingRule{{
+		URL: "https://up.example/v1", Mode: "token",
+		Models: map[string]config.BillingToken{
+			"m": {Prompt: "1.00", Completion: "2.00"},
+		},
+	}}}
+	prices, err := Compile(p1, sources)
+	if err != nil {
+		t.Fatal(err)
+	}
+	entry := prices["models"].(map[string]interface{})["m"].(map[string]interface{})
+	if _, present := entry["reasoning_price"]; present {
+		t.Errorf("reasoning_price present (%v) without a rule; the plugin would treat 0 as free",
+			entry["reasoning_price"])
+	}
+
+	// Case 2: declared → present, and converted to per-token like every other rate.
+	p2 := &config.BillingProfile{ID: "p", Rules: []config.BillingRule{{
+		URL: "https://up.example/v1", Mode: "token",
+		Models: map[string]config.BillingToken{
+			"m": {Prompt: "1.00", Completion: "2.00", Reasoning: "6.00"},
+		},
+	}}}
+	prices2, err := Compile(p2, sources)
+	if err != nil {
+		t.Fatal(err)
+	}
+	e2 := prices2["models"].(map[string]interface{})["m"].(map[string]interface{})
+	rp, ok := e2["reasoning_price"].(float64)
+	if !ok {
+		t.Fatalf("reasoning_price missing or wrong type: %#v", e2["reasoning_price"])
+	}
+	if rp != 6e-6 {
+		t.Errorf("reasoning_price = %v, want 6e-6 (USD per TOKEN)", rp)
+	}
+	// A malformed rate must be rejected, not silently ignored: an operator who
+	// typed "6 USD" expects either that value or an error, never a rule that
+	// looks applied and bills nothing.
+	p3 := &config.BillingProfile{ID: "p", Rules: []config.BillingRule{{
+		URL: "https://up.example/v1", Mode: "token",
+		Models: map[string]config.BillingToken{
+			"m": {Prompt: "1.00", Completion: "2.00", Reasoning: "six dollars"},
+		},
+	}}}
+	if _, err := Compile(p3, sources); err == nil {
+		t.Error("a non-numeric reasoning rate must fail the compile")
+	}
+}
+
+// TestLongContextStepCompiles pins the long-context step's two accepted forms
+// and its rejection of a step that would silently do nothing.
+//
+// Providers publish both shapes — DeepSeek says "2x past 128k", OpenAI quotes
+// an absolute "$3.00/M past 200k" — so both must compile. Neither may be
+// missing: a step with no rate is a rule that LOOKS like it handles long
+// context and does not, which under-bills without any visible sign.
+func TestLongContextStepCompiles(t *testing.T) {
+	sources := []config.Source{{Name: "up", BaseURL: "https://up.example/v1"}}
+
+	compile := func(lc config.BillingLongContext) (map[string]interface{}, error) {
+		p := &config.BillingProfile{ID: "p", Rules: []config.BillingRule{{
+			URL: "https://up.example/v1", Mode: "token",
+			Models: map[string]config.BillingToken{
+				"m": {Prompt: "1.00", Completion: "2.00", LongContext: &lc},
+			},
+		}}}
+		prices, err := Compile(p, sources)
+		if err != nil {
+			return nil, err
+		}
+		e := prices["models"].(map[string]interface{})["m"].(map[string]interface{})
+		return e["long_context"].(map[string]interface{}), nil
+	}
+
+	// Absolute rate (OpenAI's published shape).
+	lc1, err := compile(config.BillingLongContext{Above: 200000, Prompt: "3.00"})
+	if err != nil {
+		t.Fatalf("absolute long_context rate rejected: %v", err)
+	}
+	if lc1["above"] != 200000 || lc1["prompt"] != 3e-6 {
+		t.Errorf("absolute step = %#v, want above=200000 prompt=3e-6", lc1)
+	}
+
+	// Multiplier (DeepSeek's shape): 2x of the 1.00/M base.
+	lc2, err := compile(config.BillingLongContext{Above: 128000, Multiplier: 2})
+	if err != nil {
+		t.Fatalf("multiplier long_context rejected: %v", err)
+	}
+	if lc2["prompt"] != 2e-6 {
+		t.Errorf("multiplier step = %#v, want prompt=2e-6", lc2)
+	}
+
+	// Both given: the absolute rate wins, because that is the number the
+	// provider quotes. Multiplying instead would produce a figure the operator
+	// never wrote.
+	lc3, err := compile(config.BillingLongContext{Above: 128000, Prompt: "5.00", Multiplier: 2})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if lc3["prompt"] != 5e-6 {
+		t.Errorf("both given = %#v, want the quoted absolute rate 5e-6", lc3)
+	}
+
+	// Neither: a no-op step must fail, not compile into a rule that silently
+	// bills long context at the base rate.
+	if _, err := compile(config.BillingLongContext{Above: 128000}); err == nil {
+		t.Error("a step with neither prompt nor multiplier must be rejected")
+	}
+	// A zero threshold would make every request "long".
+	if _, err := compile(config.BillingLongContext{Above: 0, Multiplier: 2}); err == nil {
+		t.Error("above=0 must be rejected")
+	}
+}

internal/config/config.go

@@ -39,6 +39,15 @@ type Config struct {
 	// so peak traffic was billed at off-peak rates — no error, the bill just
 	// quietly halved. A schema-checked config file catches both at load.
 	BillingDSL *BillingDSL `yaml:"billing,omitempty"`
+	// PluginVisibility overrides the audience a plugin declared for itself,
+	// keyed by plugin name: "admin" hides a plugin (and its state) from user
+	// keys, "all" widens one that shipped admin-only, "user,admin" names roles
+	// explicitly. An absent name leaves the plugin's own declaration in force.
+	//
+	// This exists because a plugin author cannot know the operator's policy, and
+	// the operator should not have to edit a bundled .lua to change one — the
+	// next upgrade overwrites it.
+	PluginVisibility map[string]string `yaml:"plugin_visibility,omitempty" json:"plugin_visibility,omitempty"`
 	// box seals credentials (sources' api_key/headers, keys' key) at rest.
 	// In-memory values are always plaintext; only the bytes on disk are sealed.
 	// Wired by AttachSecretBox — Load leaves it nil so `-check` and tests stay
@@ -444,6 +453,19 @@ type BillingToken struct {
 	Prompt        string   `yaml:"prompt" json:"prompt"`         // USD per million tokens, e.g. "0.15"
 	Completion    string   `yaml:"completion" json:"completion"` // USD per million tokens
 	CacheDiscount *float64 `yaml:"cache_discount,omitempty" json:"cache_discount,omitempty"`
+	// Reasoning is the USD-per-million rate for thinking tokens, for the rare
+	// provider that charges them differently from ordinary output. Omit it and
+	// reasoning is billed at Completion — which is what OpenAI, DeepSeek and
+	// Anthropic all do, and what every source on this deployment reports.
+	//
+	// It is a rate for a SUBSET of completion tokens, never an extra charge: the
+	// billing plugin applies only the difference between this and Completion, so
+	// setting it to a multiple of Completion reprices the thinking share rather
+	// than billing it twice.
+	Reasoning string `yaml:"reasoning,omitempty" json:"reasoning,omitempty"`
+	// LongContext is the prompt-price step above a token threshold. Nil = one
+	// flat rate, which is what a rule means when it says nothing.
+	LongContext *BillingLongContext `yaml:"long_context,omitempty" json:"long_context,omitempty"`
 }
 
 // BillingPeak declares peak-hour pricing.
@@ -455,6 +477,31 @@ type BillingPeak struct {
 	Hours [][2]int `yaml:"hours,omitempty" json:"hours,omitempty"`
 }
 
+// BillingLongContext is a prompt-price step for long inputs.
+//
+// Every major provider charges the input tokens ABOVE a threshold at a higher
+// rate — DeepSeek doubles past 128k, OpenAI and Anthropic step up as well. A
+// flat prompt price therefore under-bills exactly the traffic an agent workload
+// generates, and it under-bills it silently: the number is plausible, the bill
+// adds up, and nothing says the rate was wrong.
+//
+// The step applies to the PROMPT only. Output and cache reads are billed at
+// their own rates regardless of input length, which is how providers publish
+// it — the higher rate is for the extra context you made the model read, not for
+// what it wrote back.
+type BillingLongContext struct {
+	// Above is the token count the step starts at. A request with exactly this
+	// many prompt tokens is still at the base rate; the step covers the tokens
+	// BEYOND it. Inclusive-below matches how the providers document it.
+	Above int `yaml:"above" json:"above"`
+	// Prompt is the USD-per-million rate for the tokens above Above.
+	Prompt string `yaml:"prompt" json:"prompt"`
+	// Multiplier is an alternative to Prompt: a factor applied to the base rate
+	// (DeepSeek's published step is exactly 2x). When both are set Prompt wins,
+	// and setting neither is a validation error rather than a silent no-op.
+	Multiplier float64 `yaml:"multiplier,omitempty" json:"multiplier,omitempty"`
+}
+
 // Validate rejects a BillingDSL that could not work as written.
 func (b *BillingDSL) Validate() error {
 	if b == nil {

internal/core/core.go

@@ -11,6 +11,7 @@ import (
 	"os"
 	"path/filepath"
 	"sort"
+	"net/url"
 	"strings"
 	"sync"
 	"sync/atomic"
@@ -68,6 +69,29 @@ func NewFromConfig(cfg *config.Config) (*Core, error) {
 	// to boot the gateway because a third-party .lua has a typo would be the
 	// wrong trade.
 	c.plugins = lua.NewPlugins(c.vm, cfg.PluginDir)
+	// Plugin outbound HTTP is scoped to the hosts of configured sources. Set
+	// before LoadDir so a plugin that requests on its first hook already sees
+	// the real allowlist rather than the deny-all placeholder.
+	c.plugins.SetOutboundHosts(sourceHosts(cfg))
+	// Let the plugin layer validate access specs against the keys that exist.
+	// A stored allowlist naming a key that does not exist refuses every caller,
+	// which looks exactly like a broken endpoint.
+	lua.SetKeyNameLookup(func() []string {
+		// ListKeys is the structured view (bare `gateway_keys:` entries are
+		// expanded into records with names); cfg.GatewayKeys is still []string.
+		keys := c.ListKeys()
+		out := make([]string, 0, len(keys))
+		for _, k := range keys {
+			if k.Name != "" {
+				out = append(out, k.Name)
+			}
+		}
+		return out
+	})
+	// Operator overrides win over the plugin's own declared audience.
+	for name, aud := range cfg.PluginVisibility {
+		c.plugins.SetPluginVisibility(name, aud)
+	}
 	if err := c.plugins.SeedBundled(); err != nil {
 		log.Printf("[core] seed bundled plugins: %v", err)
 	}
@@ -615,9 +639,35 @@ func (c *Core) rebuildRegistry() error {
 	c.buildAutoChain()
 	c.buildAutoImageChain()
 	c.rebuildKeyAutoChains()
+	// A source edit can add or remove an upstream host, which changes what a
+	// plugin is allowed to reach. Keeping this beside the chain rebuilds means
+	// one call site owns "config changed, refresh derived state".
+	c.plugins.SetOutboundHosts(sourceHosts(c.cfg))
 	return nil
 }
 
+// sourceHosts extracts the host of every configured source's base_url — the
+// allowlist for plugin outbound HTTP. Unparseable URLs are skipped rather than
+// failing startup: a malformed base_url is already rejected when the provider is
+// built, and it must not take plugins down with it.
+func sourceHosts(cfg *config.Config) []string {
+	if cfg == nil {
+		return nil
+	}
+	out := make([]string, 0, len(cfg.Sources))
+	for _, s := range cfg.Sources {
+		if s.BaseURL == "" {
+			continue
+		}
+		u, err := url.Parse(s.BaseURL)
+		if err != nil || u.Hostname() == "" {
+			continue
+		}
+		out = append(out, u.Hostname())
+	}
+	return out
+}
+
 // buildAutoChain rebuilds the AUTO chain snapshot from config.yaml rules
 // against the current providers. Slot model ids are normalized to the exact
 // configured spelling (case-insensitive match), otherwise ModelFor would fall

internal/gateway/api.go

@@ -464,6 +464,13 @@ func (g *Gateway) handleStatsAPI(w http.ResponseWriter, r *http.Request) {
 			"by_status": out.Status,
 			"buckets":   out.Bucket,
 			"truncated": out.Truncated,
+			// "Active requests" is a live gauge (requests in flight right now),
+			// not a period aggregate. Omitting it from the period response made
+			// the KPI read `st.active || 0` → 0 on every day/week/month view while
+			// the all-time view showed the truth, which read as "the gateway is
+			// idle" precisely when it was busiest. A calendar window cannot
+			// filter it, so it travels with every response unfiltered.
+			"active": g.stats.Active(),
 			"records":   page.Records,
 			// The paging cursor has to travel with the first page, otherwise
 			// the scroll handler has nothing to page from and the table is

internal/gateway/chat.go

@@ -789,6 +789,18 @@ func recordChatUsage(rec *Req, req *types.ChatRequest, resp *types.UnifiedRespon
 	// Cache accounting: whichever source format the adapter normalized into
 	// (prompt_tokens_details.cached_tokens or legacy prompt_cache_hit_tokens),
 	// read it back so the request record carries the hit/miss split.
+	// reasoning_tokens is a SUBSET of completion_tokens and is billed at the
+	// completion rate by every provider measured, so it must NOT be added to
+	// Compl. Recording it separately is what lets a price rule split thinking
+	// from output for the providers that do price them differently, and what
+	// keeps the billing plugin from guessing.
+	//
+	// Leaving it unrecorded is worse than useless: the adapters already parse it
+	// and forward it (opencodego/gemini/anthropic all do), so the number exists
+	// in the response and is thrown away at the one place that could keep it.
+	if d := resp.TokenUsage.CompletionTokensDetails; d != nil && d.ReasoningTokens > 0 {
+		rec.Reasoning = int64(d.ReasoningTokens)
+	}
 	if d := resp.TokenUsage.PromptTokensDetails; d != nil && d.CachedTokens > 0 {
 		rec.CacheHit = int64(d.CachedTokens)
 	}
@@ -1044,6 +1056,9 @@ func (g *Gateway) fireEnd(rec *Req) {
 		// carried this (rec.CacheReported); the plugin payload did not, so any
 		// plugin-level cache accounting had to guess.
 		"cache_reported": rec.CacheReported,
+		// 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,

internal/gateway/chat_reasoning_test.go

@@ -0,0 +1,71 @@
+package gateway
+
+import (
+	"testing"
+
+	"llmsproxy/internal/types"
+)
+
+// TestRecordChatUsageKeepsReasoningTokens pins the reasoning-token plumbing.
+//
+// The adapters already parse completion_tokens_details.reasoning_tokens and
+// forward it (opencodego, gemini, anthropic all do), and types.go has carried
+// the field since it was defined — but nothing ever READ it. The number arrived
+// in the response and was discarded at the one place that could have kept it,
+// so a pricing rule had no way to split thinking from output.
+//
+// The property that matters most: reasoning is a SUBSET of completion. Every
+// provider measured bills it at the completion rate, so a billing plugin that
+// added it to completion_tokens would double-charge every thinking token. The
+// test asserts completion is left alone.
+func TestRecordChatUsageKeepsReasoningTokens(t *testing.T) {
+	rec := &Req{}
+	resp := &types.UnifiedResponse{
+		TokenUsage: types.TokenUsage{
+			Prompt:     100,
+			Completion: 500,
+			CompletionTokensDetails: &types.CompletionTokensDetails{
+				ReasoningTokens: 320,
+			},
+		},
+	}
+	recordChatUsage(rec, &types.ChatRequest{}, resp)
+
+	if rec.Reasoning != 320 {
+		t.Errorf("Reasoning = %d, want 320", rec.Reasoning)
+	}
+	// The subset must not be folded into completion — it is already counted there.
+	if rec.Compl != 500 {
+		t.Errorf("Compl = %d, want 500 (reasoning is a subset, never additive)", rec.Compl)
+	}
+	if rec.Prompt != 100 {
+		t.Errorf("Prompt = %d, want 100", rec.Prompt)
+	}
+}
+
+// TestRecordChatUsageReasoningAbsentStaysZero: "not reported" and "reported as
+// zero" must both leave the field at 0, and neither may invent a number. A
+// plugin reading Reasoning cannot tell them apart, which is why the field is
+// documented as "0 means not reported" rather than "0 means no thinking".
+func TestRecordChatUsageReasoningAbsentStaysZero(t *testing.T) {
+	for _, tc := range []struct {
+		name string
+		u    types.TokenUsage
+	}{
+		{"no details object", types.TokenUsage{Prompt: 10, Completion: 20}},
+		{"details without reasoning", types.TokenUsage{
+			Prompt:                  10,
+			Completion:              20,
+			CompletionTokensDetails: &types.CompletionTokensDetails{},
+		}},
+	} {
+		rec := &Req{}
+		recordChatUsage(rec, &types.ChatRequest{}, &types.UnifiedResponse{TokenUsage: tc.u})
+		if rec.Reasoning != 0 {
+			t.Errorf("%s: Reasoning = %d, want 0", tc.name, rec.Reasoning)
+		}
+		if rec.Compl != 20 {
+			t.Errorf("%s: Compl = %d, want 20", tc.name, rec.Compl)
+		}
+	}
+}

internal/gateway/plugin_call_api_test.go

@@ -0,0 +1,150 @@
+package gateway
+
+// The /api/plugins/{name}/call/{fn} endpoint splits its path with string
+// surgery, and the split point is easy to get wrong in a way that still
+// "works": keeping the "call/" segment in the function name produced
+// `does not export "call/collect_account"` — the plugin was correct, the router
+// was not, and every call 502'd while the exports list looked perfectly fine.
+//
+// These tests drive the real router, so a refactor that reintroduces the
+// off-by-one is caught here rather than in production.
+
+import (
+	"encoding/json"
+	"net/http"
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+
+	"llmsproxy/internal/config"
+	"llmsproxy/internal/core"
+)
+
+// callPlugin exposes exactly one export plus a function that is NOT exported.
+// If the exports gate is not honoured, secretFn is reachable and with it the
+// plugin's private state — for a plugin holding an API key that means handing
+// the key to anyone who can POST to the gateway.
+const callPlugin = `
+local P = {}
+P.name = "calltest"
+P.version = "1.0.0"
+P.exports = { "add" }
+P.secret = "do-not-leak"
+function P.add(a)
+  return { result = (tonumber(a and a.x) or 0) + 1 }
+end
+function P.secretFn()
+  return { leaked = P.secret }
+end
+return P
+`
+
+func gatewayWithCallPlugin(t *testing.T) *Gateway {
+	t.Helper()
+	dir := t.TempDir()
+	cfgPath := filepath.Join(dir, "config.yaml")
+	body := "listen: :0\n" +
+		"adapter_dir: " + filepath.Join(dir, "adapters") + "\n" +
+		"plugin_dir: " + filepath.Join(dir, "plugins") + "\n" +
+		"runtime_file: " + filepath.Join(dir, "runtime.json") + "\n" +
+		"gateway_keys:\n  - sk-test\n"
+	if err := os.WriteFile(cfgPath, []byte(body), 0600); err != nil {
+		t.Fatal(err)
+	}
+	cfg, err := config.Load(cfgPath)
+	if err != nil {
+		t.Fatal(err)
+	}
+	c, err := core.NewFromConfig(cfg)
+	if err != nil {
+		t.Fatalf("core: %v", err)
+	}
+	t.Cleanup(c.Close)
+	if err := c.Plugins().LoadSource("calltest", callPlugin); err != nil {
+		t.Fatalf("load calltest: %v", err)
+	}
+	g, err := New(c)
+	if err != nil {
+		t.Fatalf("gateway: %v", err)
+	}
+	return g
+}
+
+func TestPluginCallSplitsFunctionNameCorrectly(t *testing.T) {
+	g := gatewayWithCallPlugin(t)
+	rr := doReq(t, g, http.MethodPost,
+		"/api/plugins/calltest/call/add", `{"x":41}`)
+	if rr.Code != http.StatusOK {
+		t.Fatalf("call failed: HTTP %d %s", rr.Code, rr.Body.String())
+	}
+	var out struct {
+		Result struct {
+			Result int `json:"result"`
+		} `json:"result"`
+	}
+	if err := json.Unmarshal(rr.Body.Bytes(), &out); err != nil {
+		t.Fatalf("decode: %v (body=%s)", err, rr.Body.String())
+	}
+	// 41 + 1 == 42. Reaching 42 proves BOTH halves of the contract: the "/call/"
+	// segment did not leak into the function name, and the payload arrived.
+	if out.Result.Result != 42 {
+		t.Fatalf("got %d, want 42 — either the /call/ segment leaked into the "+
+			"function name or the argument table never reached Lua", out.Result.Result)
+	}
+}
+
+func TestPluginCallRefusesUndeclaredFunction(t *testing.T) {
+	g := gatewayWithCallPlugin(t)
+	rr := doReq(t, g, http.MethodPost,
+		"/api/plugins/calltest/call/secretFn", `{}`)
+	if rr.Code == http.StatusOK {
+		t.Fatalf("an unexported function must not be callable, got 200: %s",
+			rr.Body.String())
+	}
+	if strings.Contains(rr.Body.String(), "do-not-leak") {
+		t.Fatal("the refusal leaked the plugin's private state")
+	}
+}
+
+// A plugin that declares no exports at all has no callable surface. Reading the
+// missing list as "allow everything" would turn every plugin into a remote
+// evaluator the moment this endpoint ships.
+func TestPluginCallWithoutExportsHasNoCallableSurface(t *testing.T) {
+	g := gatewayWithCallPlugin(t)
+	const bare = `
+local P = {}
+P.name = "bare"
+P.version = "1.0.0"
+function P.anything()
+  return { ok = true }
+end
+return P
+`
+	if err := g.core.Plugins().LoadSource("bare", bare); err != nil {
+		t.Fatalf("load bare: %v", err)
+	}
+	rr := doReq(t, g, http.MethodPost,
+		"/api/plugins/bare/call/anything", `{}`)
+	if rr.Code == http.StatusOK {
+		t.Fatalf("a plugin with no plugin.exports must expose nothing, got 200: %s",
+			rr.Body.String())
+	}
+}
+
+func TestPluginCallIsAdminOnly(t *testing.T) {
+	g := gatewayWithCallPlugin(t)
+	// A real user-role key, not a bare one from config: sk-test/sk-user in the
+	// config both default to admin, so a test using them would pass the gate for
+	// the wrong reason and prove nothing.
+	rec, err := g.core.CreateKey("viewer", "user", nil, "")
+	if err != nil {
+		t.Fatal(err)
+	}
+	rr := doReqAs(t, g, rec.Key, http.MethodPost,
+		"/api/plugins/calltest/call/add", `{"x":1}`)
+	if rr.Code != http.StatusForbidden {
+		t.Fatalf("a non-admin key reached the call endpoint: HTTP %d %s",
+			rr.Code, rr.Body.String())
+	}
+}

internal/gateway/plugin_http_routes.go

@@ -0,0 +1,207 @@
+package gateway
+
+// Mount for plugin-registered HTTP endpoints.
+//
+// The gateway owns the listener, the authentication, the role check and the
+// response framing; the plugin owns the logic. That split is the point: a plugin
+// cannot open a socket, cannot bypass auth, and cannot choose a path outside
+// /api/plugins/http/.
+//
+// Why this exists rather than "plugins can call each other": some things are
+// worth querying from OUTSIDE the browser — a CI job asking whether the upstream
+// balance is above zero, a monitoring probe, another tool on the box. billing
+// never wanted this (it only receives hooks); a quota plugin does. Hard-coding a
+// route per plugin would mean a gateway change for every such plugin, and an
+// arbitrary plugin's code in the main mux is not something the routing table
+// should ever grow.
+
+import (
+	"net/http"
+	"strconv"
+	"strings"
+
+	"llmsproxy/internal/lua"
+)
+
+// pluginHTTPPrefix is the mount point. It is deliberately narrow: a plugin's
+// routes live under one prefix, which makes them greppable, makes them easy to
+// revoke wholesale, and means a plugin cannot shadow /api/keys or /api/stats.
+const pluginHTTPPrefix = "/api/plugins/http/"
+
+// handlePluginHTTP dispatches a request to whichever plugin registered the
+// longest matching path.
+//
+// Longest-match rather than first-match: a plugin may register /health and
+// another /health/deep, and both must work. With registration order deciding it,
+// the outcome would depend on load order and change when a plugin is added.
+func (g *Gateway) handlePluginHTTP(w http.ResponseWriter, r *http.Request) {
+	rest := strings.TrimPrefix(r.URL.Path, pluginHTTPPrefix)
+	ps := g.core.Plugins()
+	if ps == nil {
+		writeError(w, http.StatusNotFound, "not_found", "no plugins loaded")
+		return
+	}
+
+	matches := ps.HTTPHandlers()
+	var chosen *lua.HTTPHandler
+	var params map[string]string
+	for i := range matches {
+		h := matches[i]
+		if p, ok := matchHandlerPath(h.Path, rest); ok {
+			if chosen == nil || segmentCount(h.Path) > segmentCount(chosen.Path) {
+				chosen, params = &h, p
+			}
+		}
+	}
+	if chosen == nil {
+		writeError(w, http.StatusNotFound, "not_found",
+			"no plugin serves this path")
+		return
+	}
+
+	if !strings.EqualFold(r.Method, chosen.Method) {
+		w.Header().Set("Allow", chosen.Method)
+		writeError(w, http.StatusMethodNotAllowed, "method_not_allowed",
+			"this endpoint accepts "+chosen.Method+" only")
+		return
+	}
+
+	role := reqRole(r.Context())
+	if !chosen.Allows(role, reqKeyName(r.Context())) {
+		writeError(w, http.StatusForbidden, "forbidden",
+			handlerAccessHint(*chosen))
+		return
+	}
+
+	body, err := readBody(r)
+	if err != nil {
+		writeError(w, http.StatusBadRequest, "invalid_request", err.Error())
+		return
+	}
+	req := &lua.HandlerRequest{
+		Method:  r.Method,
+		Path:    rest,
+		Query:   flattenQuery(r.URL.Query()),
+		Params:  params,
+		Headers: flattenHeader(r.Header),
+		Body:    body,
+		Role:    role,
+		Key:     reqKey(r.Context()),
+	}
+	resp, err := ps.InvokeHandler(handlerOwner(chosen), chosen.Handler, req)
+	if err != nil {
+		// A plugin error is a 502, not a 500: the gateway is fine, the plugin is
+		// not, and the operator needs to tell those apart in their logs.
+		writeError(w, http.StatusBadGateway, "plugin_error", err.Error())
+		return
+	}
+	lua.WriteHandlerResponse(w, resp)
+}
+
+// handlerAccessHint explains, in the refusal, exactly which declaration would
+// open this endpoint. A 403 with no reason is the kind of thing an operator
+// opens the source to work out, and then the fix is a guess.
+func handlerAccessHint(h lua.HTTPHandler) string {
+	switch {
+	case len(h.Keys) > 0:
+		return "this endpoint allows only the named keys (" +
+			strings.Join(h.Keys, ", ") + "); authenticate with one of them"
+	case len(h.Roles) > 0:
+		return "this endpoint requires role " + strings.Join(h.Roles, " or ") +
+			" (widen it in the plugin's http declaration, or set public=true)"
+	default:
+		return "this endpoint declares no roles, so it is admin-only; " +
+			"set roles=… or keys=… in the plugin's http declaration"
+	}
+}
+
+// handlerOwner returns the plugin name that registered h. The gateway stores the
+// owner alongside the handler precisely so this does not have to be guessed from
+// the path prefix.
+func handlerOwner(h *lua.HTTPHandler) string { return h.Plugin }
+
+// segmentCount is how deep a route is, used for longest-match. A path with more
+// segments wins, so /health/deep beats /health when both are registered.
+func segmentCount(p string) int { return strings.Count(p, "/") }
+
+// matchHandlerPath matches a route pattern against a concrete path, capturing
+// ":name" segments.
+//
+// Returns ok=false on a segment-count mismatch or a missing value for a
+// param — a route must not match with an empty capture, or two different URLs
+// would reach the same handler indistinguishably.
+func matchHandlerPath(pattern, path string) (map[string]string, bool) {
+	pp := strings.Split(strings.Trim(pattern, "/"), "/")
+	sp := strings.Split(strings.Trim(path, "/"), "/")
+	if len(pp) != len(sp) {
+		return nil, false
+	}
+	var params map[string]string
+	for i := range pp {
+		if strings.HasPrefix(pp[i], ":") {
+			name := pp[i][1:]
+			if sp[i] == "" {
+				return nil, false
+			}
+			if params == nil {
+				params = map[string]string{}
+			}
+			params[name] = sp[i]
+			continue
+		}
+		if pp[i] != sp[i] {
+			return nil, false
+		}
+	}
+	return params, true
+}
+
+func flattenQuery(q map[string][]string) map[string]string {
+	out := make(map[string]string, len(q))
+	for k, v := range q {
+		if len(v) > 0 {
+			out[k] = v[0]
+		}
+	}
+	return out
+}
+
+func flattenHeader(h http.Header) map[string]string {
+	out := make(map[string]string, len(h))
+	for k := range h {
+		out[strings.ToLower(k)] = h.Get(k)
+	}
+	return out
+}
+
+func readBody(r *http.Request) (string, error) {
+	if r.Body == nil {
+		return "", nil
+	}
+	defer r.Body.Close()
+	// A cap, because the handler runs inside the gateway process: an unbounded
+	// body from an authenticated caller is a memory exhaustion vector.
+	const maxBody = 4 << 20
+	buf := make([]byte, 0, 4096)
+	tmp := make([]byte, 32*1024)
+	for {
+		n, err := r.Body.Read(tmp)
+		if n > 0 {
+			if len(buf)+n > maxBody {
+				return "", &bodyTooLarge{}
+			}
+			buf = append(buf, tmp[:n]...)
+		}
+		if err != nil {
+			break
+		}
+	}
+	return string(buf), nil
+}
+
+type bodyTooLarge struct{}
+
+func (bodyTooLarge) Error() string { return "request body exceeds 4 MiB" }
+
+// limitBody is kept for the numeric form some callers want in a log line.
+var _ = strconv.Itoa

internal/gateway/plugins_api.go

@@ -93,6 +93,44 @@ func (g *Gateway) handlePluginsAPI(w http.ResponseWriter, r *http.Request) {
 		return
 	}
 
+	// /api/plugins/{name}/http/{path}/access — the operator's view of, and edit
+	// surface for, which keys an endpoint answers.
+	if i := strings.Index(path, "/http/"); i >= 0 {
+		name := path[:i]
+		rest := path[i+len("/http/"):]
+		if strings.HasSuffix(rest, "/access") {
+			epPath := strings.TrimSuffix(rest, "/access")
+			g.handleEndpointAccess(w, r, name, epPath)
+			return
+		}
+	}
+
+	// /api/plugins/{name}/call/{fn}
+	if i := strings.Index(path, "/call/"); i >= 0 {
+		name := path[:i]
+		fn := path[i+len("/call/"):]
+		if name == "" || fn == "" {
+			writeError(w, http.StatusBadRequest, "invalid_request", "call needs a plugin and a function name")
+			return
+		}
+		g.handlePluginCall(w, r, name, fn)
+		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
@@ -106,6 +144,11 @@ func (g *Gateway) handlePluginsAPI(w http.ResponseWriter, r *http.Request) {
 				"on_disk":     ps.OnDisk(),
 				"hook_errors": ps.HookErrors(),
 				"plugin_dir":  g.pluginDir(),
+				// 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(),
+				"http_access": ps.EndpointAccess(),
 			})
 		case http.MethodPost:
 			var body struct {
@@ -211,15 +254,33 @@ func (g *Gateway) handlePluginState(w http.ResponseWriter, r *http.Request, name
 	ps := g.core.Plugins()
 	switch r.Method {
 	case http.MethodGet:
-		// Any role may read: plugin state is reporting data (cost, counts),
-		// and the caller has already been authenticated. Admin-only would stop
-		// a user key's own billing widget from rendering.
+		// The plugin's audience gates its DATA as well as its page. Hiding the UI
+		// while leaving GET /state open would put the numbers one curl away, which
+		// is the leak the audience field exists to close — and billing's state is
+		// a global ledger, not the caller's own usage.
+		//
+		// A plugin that declares no audience is unchanged (every role may read
+		// it), so a single-purpose widget keeps working without edits.
+		role := reqRole(r.Context())
+		if !ps.StateVisibleTo(name, role) {
+			writeError(w, http.StatusForbidden, "forbidden",
+				"plugin "+name+" is not visible to role "+role)
+			return
+		}
 		// prices rides along: the billing rules editor must show the table
 		// that is actually in effect, not a reconstruction from config. It is
 		// absent for plugins with no prices (every plugin but billing).
+		state := ps.State(name)
+		// A plugin may SPLIT its state: global buckets stay deployment-wide while
+		// per-caller buckets are narrowed to the authenticated key. Role
+		// visibility alone cannot express this — the same role is allowed to read
+		// the plugin, just not another caller's row.
+		if scope := ps.StateScopeFor(name); scope != nil && role != "admin" {
+			state = splitStateForKey(state, scope, reqKey(r.Context()))
+		}
 		payload := map[string]interface{}{
 			"plugin": name,
-			"state":  ps.State(name),
+			"state":  state,
 		}
 		if pr := ps.Prices(name); pr != nil {
 			payload["prices"] = pr
@@ -245,6 +306,107 @@ func (g *Gateway) handlePluginState(w http.ResponseWriter, r *http.Request, name
 	}
 }
 
+// Call invokes a named function on a plugin's Lua state and returns its result.
+//
+// This exists because plugin pages otherwise have NO way to reach their own
+// plugin code: pluginAPI offers state read/write and a couple of callbacks, and
+// that is enough for billing (whose numbers arrive on hooks during request
+// forwarding) but useless for a plugin that must do work on demand. A quota
+// page has to ask the vendor when the operator clicks Refresh, and no hook
+// fires at that moment.
+//
+// It is deliberately narrow:
+//   - The function must be one the plugin explicitly lists in plugin.exports.
+//     Anything else is refused, so this endpoint cannot become a generic "run
+//     arbitrary Lua with my credentials" oracle over a plugin's private state.
+//   - The caller must be admin, and the result is JSON — no Lua value is ever
+//     handed back to Go as a live closure.
+//   - Errors come back as (nil, message) from the plugin, matching every other
+//     plugin entry point, and surface as a 502 rather than raising.
+func (g *Gateway) handlePluginCall(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 args map[string]interface{}
+	if err := json.NewDecoder(r.Body).Decode(&args); err != nil {
+		writeError(w, http.StatusBadRequest, "invalid_request", "invalid json: "+err.Error())
+		return
+	}
+	result, err := g.core.Plugins().Call(name, fn, args)
+	if err != nil {
+		writeError(w, http.StatusBadGateway, "plugin_error", err.Error())
+		return
+	}
+	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
@@ -282,3 +444,176 @@ func (g *Gateway) removePlugin(name string) error {
 	_ = os.Remove(filepath.Join(g.pluginDir(), name+".lua"))
 	return g.core.Plugins().Unload(name)
 }
+
+// handleEndpointAccess serves the operator's view and edit surface for one
+// plugin endpoint's reachable keys.
+//
+//	GET    …/http/{path}/access   the declared rules plus any override
+//	PUT    …/http/{path}/access   body {"access": "admin,@homeagent"}
+//	DELETE …/http/{path}/access   clear the override
+//
+// The spec string is a comma list: a bare word is a role, "@name" is a gateway
+// key's configured name, and the literal "public" opens it to any
+// authenticated key. That is the same grammar the visibility override uses, and
+// keeping one grammar means an operator learns it once.
+//
+// Key names are validated against the configured keys BEFORE being stored. An
+// allowlist naming a key that does not exist is a silent no-op — every caller
+// is refused — and the operator would have no way to tell that apart from the
+// endpoint simply working.
+func (g *Gateway) handleEndpointAccess(w http.ResponseWriter, r *http.Request, name, epPath string) {
+	if reqRole(r.Context()) != "admin" {
+		writeError(w, http.StatusForbidden, "forbidden", "admin role required")
+		return
+	}
+	ps := g.core.Plugins()
+
+	// epPath "*" is the WHOLE-PLUGIN visibility override, not an endpoint. It
+	// shares this route (and therefore this store) on purpose: an operator
+	// setting "who can see this plugin" and "who can call this endpoint" is
+	// asking the same question at two granularities, and two stores would let
+	// them disagree. SetEndpointAccess cannot serve it — it validates that the
+	// endpoint exists, which "*" by definition does not.
+	if epPath == pluginWholeOverride {
+		g.handlePluginVisibility(w, r, name)
+		return
+	}
+
+	switch r.Method {
+	case http.MethodGet:
+		for _, e := range ps.EndpointAccess() {
+			if e.Plugin == name && e.Path == epPath {
+				writeJSON(w, http.StatusOK, map[string]interface{}{"endpoint": e})
+				return
+			}
+		}
+		writeError(w, http.StatusNotFound, "not_found", "no such endpoint")
+	case http.MethodPut, http.MethodDelete:
+		spec := ""
+		if r.Method == http.MethodPut {
+			var body struct {
+				Access string `json:"access"`
+			}
+			if err := json.NewDecoder(r.Body).Decode(&body); err != nil {
+				writeError(w, http.StatusBadRequest, "invalid_request", "invalid json: "+err.Error())
+				return
+			}
+			spec = body.Access
+		}
+		if err := ps.ValidateAccessSpec(spec); err != nil {
+			writeError(w, http.StatusBadRequest, "invalid_request", err.Error())
+			return
+		}
+		if err := ps.SetEndpointAccess(name, epPath, spec); err != nil {
+			writeError(w, http.StatusBadRequest, "invalid_request", err.Error())
+			return
+		}
+		writeJSON(w, http.StatusOK, map[string]interface{}{"ok": true, "access": spec})
+	default:
+		writeError(w, http.StatusMethodNotAllowed, "method_not_allowed", "")
+	}
+}
+
+// splitStateForKey rewrites a plugin state so that a non-admin caller sees only
+// the buckets that belong to them.
+//
+// The plugin declares which buckets are global and which are per-key. A per-key
+// bucket is keyed by the MASKED key ("***a1b2c3" — the same value the audit log
+// uses), so the caller's own row is found by masking the presented secret and
+// matching the last six characters.
+//
+// Keys that do not match are DROPPED, not zeroed: a zeroed row would render as
+// "you spent nothing" and, worse, would leak that a row exists for someone else.
+func splitStateForKey(state interface{}, scope *lua.StateScope, key string) interface{} {
+	m, ok := state.(map[string]interface{})
+	if !ok || scope == nil || key == "" {
+		return state
+	}
+	masked := maskKey(key)
+	out := make(map[string]interface{}, len(m))
+	for k, v := range m {
+		if !scope.IsGlobal(k) {
+			// Per-key bucket: keep only the caller's rows, at every nesting level
+			// the scope mentions (by_day_key is {day: {masked: bucket}}).
+			out[k] = filterKeyBuckets(v, masked, scope.DepthOf(k))
+			continue
+		}
+		out[k] = v
+	}
+	return out
+}
+
+// filterKeyBuckets narrows a keyed structure to one masked key.
+//
+// depth 1 is {masked: bucket}; depth 2 is {day: {masked: bucket}}. The level is
+// declared rather than sniffed, because both shapes are dicts and sniffing would
+// confuse an empty bucket with a nested one.
+func filterKeyBuckets(v interface{}, masked string, depth int) interface{} {
+	switch depth {
+	case 0:
+		return v // not a per-key bucket; leave alone
+	case 1:
+		m, ok := v.(map[string]interface{})
+		if !ok {
+			return v
+		}
+		if row, ok := m[masked]; ok {
+			return map[string]interface{}{masked: row}
+		}
+		return map[string]interface{}{}
+	default:
+		m, ok := v.(map[string]interface{})
+		if !ok {
+			return v
+		}
+		out := make(map[string]interface{}, len(m))
+		for day, inner := range m {
+			im, ok := inner.(map[string]interface{})
+			if !ok {
+				continue
+			}
+			if row, ok := im[masked]; ok {
+				out[day] = map[string]interface{}{masked: row}
+			}
+		}
+		return out
+	}
+}
+
+// pluginWholeOverride is the epPath that addresses a plugin's overall visibility
+// rather than one of its endpoints.
+const pluginWholeOverride = "*"
+
+// handlePluginVisibility is the whole-plugin half of the access API. The
+// audience string uses the same vocabulary as an endpoint spec (roles, @key
+// names, public) so the UI can drive both from one input.
+func (g *Gateway) handlePluginVisibility(w http.ResponseWriter, r *http.Request, name string) {
+	ps := g.core.Plugins()
+	switch r.Method {
+	case http.MethodGet:
+		writeJSON(w, http.StatusOK, map[string]interface{}{
+			"plugin":     name,
+			"visibility": ps.VisibilityOverrides()[name],
+		})
+	case http.MethodPut, http.MethodDelete:
+		spec := ""
+		if r.Method == http.MethodPut {
+			var body struct {
+				Access string `json:"access"`
+			}
+			if err := json.NewDecoder(r.Body).Decode(&body); err != nil {
+				writeError(w, http.StatusBadRequest, "invalid_request", "invalid json: "+err.Error())
+				return
+			}
+			spec = body.Access
+		}
+		if err := ps.ValidateAccessSpec(spec); err != nil {
+			writeError(w, http.StatusBadRequest, "invalid_request", err.Error())
+			return
+		}
+		ps.SetPluginVisibility(name, spec)
+		writeJSON(w, http.StatusOK, map[string]interface{}{"ok": true, "visibility": spec})
+	default:
+		writeError(w, http.StatusMethodNotAllowed, "method_not_allowed", r.Method+" is not allowed here")
+	}
+}

internal/gateway/plugins_api_test.go

@@ -210,3 +210,54 @@ func TestPluginStatePUTIsAdminOnly(t *testing.T) {
 		t.Errorf("admin PUT state = %d: %s", adm.Code, adm.Body.String())
 	}
 }
+
+// TestPluginsListCarriesAccessTables: the plugin page renders visibility
+// overrides and the endpoint access table from /api/plugins. Without them in the
+// response the operator has to open a .lua or config.yaml to learn who can see
+// what, which is exactly what the UI is meant to avoid.
+func TestPluginsListCarriesAccessTables(t *testing.T) {
+	g := gatewayWithCallPlugin(t)
+	rr := doReq(t, g, http.MethodGet, "/api/plugins", "")
+	if rr.Code != http.StatusOK {
+		t.Fatalf("GET /api/plugins = %d", rr.Code)
+	}
+	var out struct {
+		Visibility map[string]string     `json:"visibility"`
+		HTTPAccess []map[string]interface{} `json:"http_access"`
+	}
+	if err := json.Unmarshal(rr.Body.Bytes(), &out); err != nil {
+		t.Fatalf("decode: %v (body=%s)", err, rr.Body.String())
+	}
+	// The calltest plugin registers no endpoints, so the table is present but
+	// empty — the KEY is that the field exists at all. A missing field decodes to
+	// nil and the UI silently renders nothing, which looks identical to "this
+	// plugin has no endpoints".
+	if !strings.Contains(rr.Body.String(), "http_access") {
+		t.Error("/api/plugins response lacks the http_access field")
+	}
+	if !strings.Contains(rr.Body.String(), "visibility") {
+		t.Error("/api/plugins response lacks the visibility field")
+	}
+}
+
+// TestPluginWholeVisibilityViaAccessAPI covers the "*" path: the plugin page
+// edits a plugin's overall audience through the same route it edits endpoints
+// with. Before this existed the UI had a visibility selector whose save 404'd,
+// because SetEndpointAccess validates that the endpoint exists and "*" does not.
+func TestPluginWholeVisibilityViaAccessAPI(t *testing.T) {
+	g := gatewayWithCallPlugin(t)
+	rr := doReq(t, g, http.MethodPut, "/api/plugins/calltest/http/*/access",
+		`{"access":"admin, user"}`)
+	if rr.Code != http.StatusOK {
+		t.Fatalf("PUT */access = %d: %s", rr.Code, rr.Body.String())
+	}
+	if !strings.Contains(rr.Body.String(), `"ok":true`) {
+		t.Errorf("unexpected body: %s", rr.Body.String())
+	}
+	// A user role is a legitimate audience; a nonsense one is not.
+	bad := doReq(t, g, http.MethodPut, "/api/plugins/calltest/http/*/access",
+		`{"access":"../../etc"}`)
+	if bad.Code != http.StatusBadRequest {
+		t.Errorf("traversal-shaped audience = %d, want 400: %s", bad.Code, bad.Body.String())
+	}
+}

internal/gateway/server.go

@@ -235,6 +235,10 @@ func (g *Gateway) routes(w http.ResponseWriter, r *http.Request) {
 		// treats anything after the name as a plugin sub-resource and would
 		// 404 on "rules" instead of reaching the billing editor.
 		g.handleBillingRules(w, r)
+	case strings.HasPrefix(r.URL.Path, pluginHTTPPrefix):
+		// Before the generic plugin router: these paths belong to a PLUGIN, not to
+		// the plugin manager, and "http" is a valid sub-resource name otherwise.
+		g.handlePluginHTTP(w, r)
 	case r.URL.Path == "/api/plugins" || strings.HasPrefix(r.URL.Path, "/api/plugins/"):
 		g.handlePluginsAPI(w, r)
 	case r.URL.Path == "/api/chat":
@@ -305,20 +309,43 @@ func (g *Gateway) auth(next http.Handler) http.Handler {
 			http.Redirect(w, r, "/login?continue="+url.QueryEscape(r.URL.Path), http.StatusFound)
 			return
 		}
-		next.ServeHTTP(w, r.WithContext(withAuth(r.Context(), key, rec.Role)))
+		next.ServeHTTP(w, r.WithContext(withAuthNamed(r.Context(), key, rec.Role, rec.Name)))
 	})
 }
 
-// authCtx carries the authenticated gateway key and its role.
+// authCtx carries the authenticated gateway key, its role, and the key's
+// configured NAME.
+//
+// The name is here because plugin HTTP endpoints authorise by key name rather
+// than by secret, and looking the name up from the key at the point of use would
+// mean reading config on every request. It is "" for a key configured without
+// one, which such an endpoint can never match.
 type authCtx struct {
 	key  string
 	role string
+	name string
 }
 
 type authCtxKeyT struct{}
 
 func withAuth(ctx context.Context, key, role string) context.Context {
-	return context.WithValue(ctx, authCtxKeyT{}, authCtx{key: key, role: role})
+	return withAuthNamed(ctx, key, role, "")
+}
+
+// withAuthNamed is withAuth plus the key's configured name.
+func withAuthNamed(ctx context.Context, key, role, name string) context.Context {
+	return context.WithValue(ctx, authCtxKeyT{}, authCtx{key: key, role: role, name: name})
+}
+
+// reqKeyName returns the authenticated key's configured name, or "" if it has
+// none. Plugin endpoint authorisation is written against this, not against the
+// secret: a secret is rotated without notice, so a rule keyed on it would
+// silently stop matching.
+func reqKeyName(ctx context.Context) string {
+	if a, ok := ctx.Value(authCtxKeyT{}).(authCtx); ok {
+		return a.name
+	}
+	return ""
 }
 
 // reqKey returns the authenticated gateway key id (masked suffix for display).

internal/gateway/stats.go

@@ -43,6 +43,16 @@ type Req struct {
 	// accounting at all (even when the hit count is 0). The WebUI shows
 	// "0%" instead of "—" for such rows.
 	CacheReported bool `json:"cache_reported,omitempty"`
+	// Reasoning is completion_tokens_details.reasoning_tokens.
+	//
+	// It is a SUBSET of Compl, not an addition: reasoning tokens are billed at
+	// the completion rate and already counted there. This field exists so a
+	// pricing rule can split them when a provider prices thinking tokens
+	// differently (most do NOT — OpenAI and DeepSeek bill them as ordinary
+	// output), and so a reader can see what fraction of a bill is thinking.
+	// 0 means "not reported", which is NOT the same as "no reasoning happened" —
+	// three of the sources measured return no such field at all.
+	Reasoning int64 `json:"reasoning_tokens,omitempty"`
 	OK            bool `json:"ok"`
 	// Status http status code
 	Status int `json:"status"`
@@ -1211,3 +1221,16 @@ func (s *Stats) Snapshot(limit int, key string) map[string]interface{} {
 		"replay_partial": s.replayPartial,
 	}
 }
+
+// Active reports requests in flight right now.
+//
+// It is a live gauge, not an aggregate: there is no calendar period that scopes
+// it, which is why the period-scoped /api/stats response has to carry it
+// explicitly. Before it did, the day/week/month views fed `st.active || 0` a
+// missing key and showed a hard 0 — the same gateway that was serving hundreds
+// of requests a second looked idle in three of the four views.
+func (s *Stats) Active() int64 {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	return s.active
+}

internal/gateway/stats_period_test.go

@@ -401,3 +401,34 @@ func TestPeriodWeekTotalCanExceedMonthTotal(t *testing.T) {
 		t.Fatal("today must be contained in both the week and the month totals")
 	}
 }
+
+// TestPeriodStatsCarryActiveGauge pins the live "active requests" gauge onto
+// every period response.
+//
+// It is a live gauge, not an aggregate, so no calendar window can scope it — but
+// the period response is what the day/week/month views render from, and it used
+// to omit the field entirely. The UI then evaluated `st.active || 0` and printed
+// a hard 0 in three of four views, while the all-time view (whose response did
+// carry the field) showed the truth. Reading "0 active" off a busy gateway is
+// worse than reading nothing: it is a confident wrong answer.
+func TestPeriodStatsCarryActiveGauge(t *testing.T) {
+	for _, period := range []string{"day", "week", "month", "all"} {
+		g := newTestGateway(t)
+		rr := doReq(t, g, http.MethodGet, "/api/stats?period="+period, "")
+		if rr.Code != http.StatusOK {
+			t.Fatalf("period=%s: %d", period, rr.Code)
+		}
+		var out map[string]interface{}
+		if err := json.Unmarshal(rr.Body.Bytes(), &out); err != nil {
+			t.Fatalf("period=%s decode: %v", period, err)
+		}
+		v, ok := out["active"]
+		if !ok {
+			t.Errorf("period=%s response has no active field; the UI renders 0", period)
+			continue
+		}
+		if _, ok := v.(float64); !ok {
+			t.Errorf("period=%s active is %T, want a number", period, v)
+		}
+	}
+}

internal/gateway/ui/index.html

@@ -815,8 +815,21 @@
           plTitle: "插件",
           plState: "状态",
           plStages: "阶段",
-          plActive: "启用中",
-          plDisabled: "已禁用",
+
+            plVisibility: "可见性",
+            plEndpoints: "端点",
+            plAccessTitle: "访问控制",
+            plAccessHint: "填写允许访问的角色或密钥（@名称）。留空=仅 admin。填 public=任何人。指定 @密钥 时只有该密钥能访问，不受角色规则影响。",
+            plWholePlugin: "以下端点",
+            plKeyHint: "可用密钥名",
+            plAccessSave: "保存访问控制",          plActive: "启用中",
+
+            tVisDefault: "跟随插件",
+            tVisAdmin: "仅 admin",
+            tVisRoles: "admin + user",
+            tVisAll: "所有人",
+            plOverridden: "已覆盖插件声明",
+            plEpsN: "%d 个端点",          plDisabled: "已禁用",
           plBroken: "加载失败",
           plBuiltin: "内置",
           plHooks: "个阶段",
@@ -1090,8 +1103,21 @@
           plTitle: "Plugins",
           plState: "State",
           plStages: "Stages",
-          plActive: "Active",
-          plDisabled: "Disabled",
+
+          plVisibility: "Visibility",
+          plEndpoints: "Endpoints",
+          plAccessTitle: "Access control",
+          plAccessHint: "Roles or keys (@name) allowed to reach each endpoint. Empty = admin only. public = anyone. Listing @keys replaces the role rules entirely.",
+          plWholePlugin: "endpoints below",
+          plKeyHint: "Key names",
+          plAccessSave: "Save access",          plActive: "Active",
+
+          tVisDefault: "follow plugin",
+          tVisAdmin: "admin only",
+          tVisRoles: "admin + user",
+          tVisAll: "everyone",
+          plOverridden: "overrides plugin",
+          plEpsN: "%d endpoints",          plDisabled: "Disabled",
           plBroken: "Failed to load",
           plBuiltin: "Built-in",
           plHooks: "stages",
@@ -1994,7 +2020,12 @@
             const e = $(sel);
             if (e) e.textContent = txt;
           };
-          kset('[data-kpi="active"]', st.active || 0);
+          // No `|| 0` fallback: that fallback is what turned a MISSING field
+          // into a confident "0 active" on three of four views. If the server
+          // ever stops sending it again, an em dash says "unknown" and the
+          // regression is visible, instead of the gauge quietly claiming the
+          // gateway is idle while it serves traffic.
+          kset('[data-kpi="active"]', st.active === undefined ? "—" : fmtN(st.active));
           kset('[data-kpi="reqs"]', fmtN(tot.reqs));
           const rsub = $('[data-kpi="reqs-sub"]');
           if (rsub)
@@ -3648,13 +3679,20 @@ recsState.keyNames = keyNames || {};
         sortState.origin = JSON.stringify(sortState.lanes);
         const kindToggle =
           `<div class="kind-toggle"><button class="${sortState.kind === "chat" ? "" : "ghost"} small" onclick="switchSortKind('chat')">${t("kChat") || "聊天"}</button><button class="${sortState.kind === "image" ? "" : "ghost"} small" onclick="switchSortKind('image')">${t("kImage") || "生图"}</button></div>`;
+        // renderSortEditor replaces its mount wholesale, so the mount must be
+        // the canvas slot ALONE. Passing the .card (which holds the h2 with the
+        // save/reset/add buttons) used to wipe the toolbar on every open — the
+        // lane canvas survived, the buttons did not. The per-key modal never hit
+        // this because its mount only ever contained the canvas.
         $("#tab-sort").innerHTML = `
-    <div class="card"><h2><span>${t("sortTitle")}</span><span class="grow"></span>${kindToggle}
+    <div class="card">
+      <h2><span>${t("sortTitle")}</span><span class="grow"></span>${kindToggle}
       <button class="ghost small" onclick="scrAddModal()">+ ${t("sortAdd")}</button>
       <button class="ghost small" onclick="sortReset()">${t("sortReset")}</button>
-<button class="small" onclick="saveSort()">${t("sortSave")}</button></h2>
+      <button class="small" onclick="saveSort()">${t("sortSave")}</button></h2>
+      <div id="sort-mount"></div>
     </div>`;
-        renderSortEditor($("#tab-sort").querySelector(".card"), {
+        renderSortEditor($("#sort-mount"), {
           canvasId: "scr-canvas",
         });
       }
@@ -4466,6 +4504,8 @@ function afterChainEdit() {
               <td><b>${esc(p.name)}</b>${builtin}<div class="muted">${esc(p.description || p.version || "")}</div>${err}</td>
               <td>${state}</td>
               <td>${hooks}</td>
+                <td>${plVisCell(j, p)}</td>
+                <td>${plEpsCell(j, p)}</td>
               <td style="white-space:nowrap">
                 <button class="small" onclick="editPlugin('${escAttr(p.name)}')">${t("plEdit")}</button>
                 ${toggle}
@@ -4487,8 +4527,9 @@ function afterChainEdit() {
         $("#tab-plugins").innerHTML = `
     <div class="card"><h2>${t("plTitle")} (${(j.on_disk || []).length})</h2>
       <div class="tbl-wrap"><table><tr>
-        <th>${t("tName")}</th><th>${t("plState")}</th><th>${t("plStages")}</th><th></th>
-      </tr>${rows || `<tr><td colspan="4" class="empty">${t("plEmpty")}</td></tr>`}</table></div>
+        <th>${t("tName")}</th><th>${t("plState")}</th><th>${t("plStages")}</th><th>${t("plVisibility")}</th><th>${t("plEndpoints")}</th><th></th>
+      </tr>${rows || `<tr><td colspan="6" class="empty">${t("plEmpty")}</td></tr>`}</table></div>
+        <div id="pl-access"></div>
       ${errList ? `<p class="muted" style="color:var(--danger)">${t("plHookErr")}</p><ul style="font-size:12px">${errList}</ul>` : ""}
       <p class="muted">${t("plDir")} <code>${esc(j.plugin_dir)}</code></p>
     </div>
@@ -4501,6 +4542,161 @@ function afterChainEdit() {
       <p><button onclick="installPlugin()">${t("plInstallBtn")}</button></p>
     </div>`;
         bindDropzone("pl-dz", "pl-file", "#pl-name", "#pl-code");
+        renderPluginAccess(j);
+      }
+
+
+      const PL_VIS = [
+        ["", "tVisDefault"],
+        ["admin", "tVisAdmin"],
+        ["user,admin", "tVisRoles"],
+        ["all", "tVisAll"],
+      ];
+
+      // plVisCell shows the plugin's own declaration next to any operator
+      // override, because "admin-only" means something different depending on
+      // whether the plugin chose it or the deployment imposed it. Editing goes
+      // through the same access API as endpoints so there is one place where
+      // visibility is stored.
+      function plVisCell(j, p) {
+        const decl = (p.ui && p.ui.visibility) || "";
+        const ov = (j.visibility || {})[p.name] || "";
+        const cur = ov || decl;
+        // Compare as SETS. The server stores whatever order the operator typed
+        // and we round-trip it verbatim, so "admin, user" must still light up a
+        // choice — otherwise the selector silently shows "follow plugin", and
+        // saving from there writes an empty override that WIDENS access back to
+        // the plugin's own declaration.
+        const norm = (v) =>
+          (v || "")
+            .split(",")
+            .map((x) => x.trim())
+            .filter(Boolean)
+            .sort()
+            .join(",");
+        const curN = norm(cur);
+        // An audience with no matching choice (one naming a @key, say) is offered
+        // as itself so the value in force is never hidden from the operator.
+        const opts = PL_VIS.some((c) => norm(c[0]) === curN)
+          ? PL_VIS
+          : PL_VIS.concat([[cur, "plCustomVis"]]);
+        return `<select data-plvis="${escAttr(p.name)}" style="font-size:11.5px;max-width:180px">
+          ${opts
+            .map(
+              (c) =>
+                `<option value="${c[0]}"${norm(c[0]) === curN ? " selected" : ""}>${
+                  c[1] === "plCustomVis" ? "自定义: " + cur : t(c[1])
+                }</option>`,
+            )
+            .join("")}
+        </select>${
+          ov && decl && ov !== decl
+            ? `<div class="muted" style="font-size:10.5px" title="${esc(decl)}">${t("plOverridden")}</div>`
+            : ""
+        }`;
+      }
+
+      function plEpsCell(j, p) {
+        const n = (j.http_access || []).filter((e) => e.plugin === p.name).length;
+        return n
+          ? `<a href="#" onclick="document.getElementById('pl-access').scrollIntoView({behavior:'smooth'});return false">${t("plEpsN").replace("%d", n)}</a>`
+          : `<span class="muted">—</span>`;
+      }
+
+      // ---- visibility & endpoint access ----
+      //
+      // The plugin page is where an operator edits who can see what. The
+      // alternatives are editing a bundled .lua — overwritten on the next
+      // upgrade — or hand-editing config.yaml, with no feedback about what is
+      // actually in force. Each row shows the rule next to whether it was
+      // DECLARED by the plugin or OVERRIDDEN here, because "admin-only" means
+      // something different depending on which.
+      function renderPluginAccess(j) {
+        const host = $("#pl-access");
+        if (!host) return;
+        const eps = j.http_access || [];
+        if (!eps.length) { host.innerHTML = ""; return; }
+        const keys = (window.__gwKeyRows || []).map((k) => k.name).filter(Boolean);
+        const byPlugin = {};
+        eps.forEach((e) => { (byPlugin[e.plugin] = byPlugin[e.plugin] || []).push(e); });
+        let h = `<div class="card" style="margin-top:12px"><h2>${t("plAccessTitle")}</h2>
+          <p class="muted" style="font-size:12px">${t("plAccessHint")}</p>`;
+        Object.keys(byPlugin).forEach((pn) => {
+          h += `<div style="margin:0 0 14px">
+            <div style="display:flex;align-items:center;gap:8px;flex-wrap:wrap;margin:0 0 6px">
+              <b style="font-size:13px">${esc(pn)}</b>
+              <span class="muted" style="font-size:11.5px">${t("plWholePlugin")}</span>
+            </div>`;
+          (byPlugin[pn] || []).forEach((e) => {
+            const ovr = (e.override_keys && e.override_keys.length) ||
+                        (e.override_roles && e.override_roles.length);
+            const cur = ovr
+              ? [...(e.override_keys || []).map((k) => "@" + k), ...(e.override_roles || [])].join(",")
+              : [...(e.keys || []).map((k) => "@" + k), ...(e.roles || []), e.public ? "public" : ""]
+                  .filter(Boolean).join(",") || "admin";
+            h += `<div style="margin:6px 0 0;padding-left:10px;border-left:2px solid var(--bd)">
+              <div style="display:flex;align-items:center;gap:6px;flex-wrap:wrap">
+                <code style="font-size:11.5px">${esc(e.method)} ${esc(e.full_path)}</code>
+                <span class="muted" style="font-size:11px">&rarr; ${esc(e.handler)}()</span>
+                ${e.description ? `<span class="muted" style="font-size:11px">${esc(e.description)}</span>` : ""}
+                <span class="grow"></span>
+                <input data-ep="${esc(pn)}|${esc(e.path)}" value="${esc(cur)}"
+                  style="max-width:340px;font-size:11.5px"
+                  placeholder="admin, user, @homeagent, public">
+              </div>
+              ${keys.length ? `<div style="font-size:11px;opacity:.7;margin-top:3px">${t("plKeyHint")}: ${esc(keys.join(", "))}</div>` : ""}
+            </div>`;
+          });
+          h += `</div>`;
+        });
+        h += `<p><button id="pl-access-save" onclick="savePluginAccess(this)">${t("plAccessSave")}</button></p></div>`;
+        host.innerHTML = h;
+      }
+
+      async function savePluginAccess(btn) {
+        btn.disabled = true;
+        try {
+          // Whole-plugin visibility first. Sent through the SAME access API as
+          // endpoints (path "*") so visibility has exactly one storage path.
+          const viss = Array.prototype.slice.call(
+            document.querySelectorAll("[data-plvis]"));
+          for (const sel of viss) {
+            const pn = sel.getAttribute("data-plvis");
+            const url =
+              "/api/plugins/" + encodeURIComponent(pn) + "/http/*/access";
+            const v = sel.value;
+            if (v) {
+              await api(url, { method: "PUT",
+                headers: { "Content-Type": "application/json" },
+                body: JSON.stringify({ access: v }) });
+            } else {
+              await api(url, { method: "DELETE" });
+            }
+          }
+          const inputs = Array.prototype.slice.call(
+            document.querySelectorAll("#pl-access [data-ep]"));
+          for (const inp of inputs) {
+            const raw = inp.getAttribute("data-ep");
+            const sep = raw.indexOf("|");
+            const pn = raw.slice(0, sep), ep = raw.slice(sep + 1);
+            const url = "/api/plugins/" + encodeURIComponent(pn) + "/http/" +
+                        encodeURIComponent(ep) + "/access";
+            const v = inp.value.trim();
+            if (v) {
+              await api(url, { method: "PUT",
+                headers: { "Content-Type": "application/json" },
+                body: JSON.stringify({ access: v }) });
+            } else {
+              await api(url, { method: "DELETE" });
+            }
+          }
+          toast(t("toastSaved"));
+          renderPlugins();
+        } catch (e) {
+          toast(tFmt("toastSaveFail", e.message));
+        } finally {
+          btn.disabled = false;
+        }
       }
 
       async function togglePlugin(name, disabled) {
@@ -5563,6 +5759,60 @@ toast(t("kAutoChainGlobal"));
           if (!r.ok) throw new Error((await r.json().catch(() => ({}))).error?.message || r.status);
           return true;
         },
+        // call invokes one of the plugin's OWN exported functions. Billing never
+        // needs this — its numbers arrive on hooks while the gateway forwards a
+        // request — but a plugin that must do work on demand (asking a vendor
+        // for a balance when the operator clicks Refresh) has no other way to
+        // reach its own code: no hook fires on a click.
+        //
+        // Admin-only server side, and the plugin must have listed the function in
+        // plugin.exports, so this cannot be used to reach a plugin's internals.
+        async call(name, fn, args) {
+          const r = await fetch(
+            "/api/plugins/" + encodeURIComponent(name) + "/call/" + encodeURIComponent(fn),
+            {
+              method: "POST",
+              credentials: "same-origin",
+              headers: { "Content-Type": "application/json" },
+              body: JSON.stringify(args || {}),
+            },
+          );
+          const body = await r.json().catch(() => ({}));
+          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_key_auto_test.go

@@ -393,3 +393,46 @@ func TestAutoScopeConflictMessageComesFromCheckQuota(t *testing.T) {
 		t.Fatal("checkQuota's AUTO branch must name the scope as the cause")
 	}
 }
+
+// renderSortEditor replaces its mount's innerHTML wholesale. The global
+// priority page used to hand it the .card that also holds the h2 toolbar, so
+// every open deleted the save/reset/add buttons: the lanes still rendered and
+// the drag worked, but there was no way to persist a change. The per-key modal
+// never showed this because its mount held only the canvas.
+//
+// The regression is invisible in the DOM when you only count lane elements, so
+// this asserts on the structure: the toolbar lives in a sibling of the mount,
+// not inside it.
+func TestGlobalSortToolbarSurvivesCanvasRender(t *testing.T) {
+	src := uiSource(t)
+
+	// The global editor must mount into a dedicated host, not the .card.
+	i := strings.Index(src, `$("#tab-sort").innerHTML = `)
+	if i < 0 {
+		t.Fatal("global sort markup not found")
+	}
+	chunk := src[i:min(len(src), i+1400)]
+	if !strings.Contains(chunk, `id="sort-mount"`) {
+		t.Fatal("global sort page must provide a dedicated #sort-mount for the canvas")
+	}
+	// The mount handed to renderSortEditor must be that host.
+	if !strings.Contains(src, `renderSortEditor($("#sort-mount")`) {
+		t.Fatal("renderSortEditor must be given #sort-mount; passing the .card " +
+			"wipes the h2 toolbar that holds the save button")
+	}
+	// The toolbar (save in particular) must be OUTSIDE #sort-mount.
+	close := strings.Index(chunk, "</div>`;")
+	if close < 0 {
+		t.Fatal("malformed #tab-sort template")
+	}
+	head := chunk[:close]
+	mountAt := strings.Index(head, `id="sort-mount"`)
+	saveAt := strings.Index(head, `onclick="saveSort()"`)
+	if saveAt < 0 {
+		t.Fatal("the global sort page has no save button at all")
+	}
+	if mountAt >= 0 && saveAt > mountAt {
+		t.Fatal("the save button sits inside the canvas mount and will be " +
+			"erased the moment the canvas renders")
+	}
+}

internal/gateway/ui_plugin_test.go

@@ -388,3 +388,26 @@ func TestPluginSidebarEntryCarriesATitle(t *testing.T) {
 	t.Errorf("plugin sidebar button renders the icon only; a label is required. "+
 		"got: %.200s", stmt)
 }
+
+// TestPluginPageHasAccessEditor: the plugin page must let an operator configure
+// endpoint access. The APIs existed with zero UI entry point, which meant the
+// only way to change who could reach a plugin endpoint was to hand-edit
+// config.yaml — the thing the UI is meant to replace.
+func TestPluginPageHasAccessEditor(t *testing.T) {
+	src := uiSource(t)
+	for _, needle := range []string{
+		`renderPluginAccess`,    // called after rendering the plugin list
+		`savePluginAccess`,      // wired to a button
+		`/access`,               // hits the per-endpoint access API
+		`id="pl-access"`,        // container exists
+		`http_access`,           // reads the table from /api/plugins
+		`data-ep=`,              // one input per endpoint
+		`data-plvis=`,           // one visibility selector per plugin
+		`/http/*/access`,        // whole-plugin visibility shares the access API
+		`plVisCell`,             // renders the declared-vs-override distinction
+	} {
+		if !strings.Contains(src, needle) {
+			t.Errorf("plugin page lacks %q", needle)
+		}
+	}
+}

internal/lua/billing_test.go

@@ -2,12 +2,25 @@ package lua
 
 import (
 	"encoding/json"
+	"fmt"
 	"os"
 	"path/filepath"
 	"strings"
 	"testing"
 )
 
+// pluginLoaded reports whether a plugin with this name is present, ignoring any
+// LoadError: a plugin that is present but broken still counts as "loaded here",
+// and the caller decides whether the error matters.
+func pluginLoaded(ps *Plugins, name string) bool {
+	for _, p := range ps.List() {
+		if fmt.Sprint(p["name"]) == name {
+			return true
+		}
+	}
+	return false
+}
+
 // The billing plugin ships with the gateway, so its arithmetic is a contract:
 // a wrong price silently produces wrong money. These tests drive it through the
 // real hook path and check the NUMBERS, not merely that it loads.
@@ -447,8 +460,16 @@ return plugin`)
 	}
 
 	// Order must be honoured so the sidebar is predictable.
-	if len(ui.Pages) != 4 {
-		t.Fatalf("expected all 4 pages to merge, got %d: %v", len(ui.Pages), got)
+	//
+	// The page COUNT is not asserted: it used to be 4, and it grew every time a
+	// bundled plugin added a page. A hard count here reads as "the merge is
+	// broken" whenever the product gains a feature, which trains the reader to
+	// bump a number instead of reading the failure. What matters is that all
+	// three pages this test contributed merged, and that they came out sorted.
+	for _, want := range []string{"third-a", "third-b", "other-page"} {
+		if !got[want] {
+			t.Errorf("page %s did not merge: %v", want, got)
+		}
 	}
 	for i := 1; i < len(ui.Pages); i++ {
 		if ui.Pages[i].Order < ui.Pages[i-1].Order {
@@ -462,8 +483,15 @@ return plugin`)
 // configuration, since seeding only happens on a fresh plugin dir.
 func TestBillingPluginLoadedByDefault(t *testing.T) {
 	ps, pdir := billingVM(t)
-	if ps.Count() != 1 {
-		t.Fatalf("expected 1 bundled plugin, got %d", ps.Count())
+	// billing must be among the bundled plugins, not the only one. The count
+	// grows with every shipped plugin and asserting it makes this test a
+	// changelog detector.
+	if !pluginLoaded(ps, "billing") {
+		var names []string
+		for _, p := range ps.List() {
+			names = append(names, fmt.Sprint(p["name"]))
+		}
+		t.Fatalf("billing is not loaded; bundled plugins are %v", names)
 	}
 	if _, err := os.Stat(filepath.Join(pdir, "billing.lua")); err != nil {
 		t.Errorf("billing.lua was not written to the plugin dir: %v", err)
@@ -966,3 +994,245 @@ func TestBillingCacheHitClampedInStats(t *testing.T) {
 			tot["cache_fresh_tokens"])
 	}
 }
+
+// TestUnpricedIsTrackedPerDay pins the day-keyed unpriced tally.
+//
+// The lifetime counters cannot answer the question an operator actually has —
+// "is it STILL happening?" — because they keep the debt from before a rule was
+// added. After pricing a mispriced source the lifetime number never moves, so a
+// dashboard reading "unpriced: 6376" is indistinguishable from a system that is
+// still billing nothing. Day buckets are what make a pricing fix observable: the
+// offending day keeps its count and the next day starts at zero for it.
+func TestUnpricedIsTrackedPerDay(t *testing.T) {
+	ps, _ := billingVM(t)
+	// Nothing is priced in the default config, so every request is unpriced.
+	// Two requests on two different UTC days.
+	for _, ms := range []int64{1790000000000, 1790100000000} {
+		ps.Fire(StageRequestEnd, map[string]interface{}{
+			"model": "mystery-model", "source": "nowhere", "key": "***aaaaaa",
+			"ok": true, "prompt_tokens": 10, "completion_tokens": 5,
+			"time": ms,
+		})
+	}
+	st := stateOf(t, ps)
+	byDay, ok := st["unpriced_by_day"].(map[string]interface{})
+	if !ok {
+		t.Fatalf("state has no unpriced_by_day map: %T", st["unpriced_by_day"])
+	}
+	if len(byDay) != 2 {
+		t.Fatalf("expected 2 day buckets, got %d: %v", len(byDay), byDay)
+	}
+	var sum float64
+	for day, bucket := range byDay {
+		b := bucket.(map[string]interface{})
+		sum += b["reqs"].(float64)
+		models := b["models"].(map[string]interface{})
+		if models["mystery-model"].(float64) != 1 {
+			t.Errorf("%s: model tally = %v, want 1", day, models["mystery-model"])
+		}
+	}
+	if sum != 2 {
+		t.Errorf("day buckets sum to %v, want 2 (the lifetime counter must not be double-counted)", sum)
+	}
+	// The lifetime counters still exist and still agree — this is additive, not
+	// a replacement.
+	if st["unpriced_reqs"].(float64) != 2 {
+		t.Errorf("unpriced_reqs = %v, want 2", st["unpriced_reqs"])
+	}
+}
+
+// TestReasoningBilledAtCompletionRateByDefault is the property that keeps the
+// standard formula matching a real invoice.
+//
+// reasoning_tokens is a subset of completion_tokens and every provider measured
+// (OpenAI, DeepSeek, Anthropic) charges it at the ordinary output rate. A rule
+// with no reasoning_price must therefore produce EXACTLY the number the
+// pre-reasoning formula produced — the field exists for a provider that prices
+// thinking differently, not as a new charge.
+func TestReasoningBilledAtCompletionRateByDefault(t *testing.T) {
+	ps, _ := billingVM(t)
+	if err := ps.SetState("billing", map[string]interface{}{
+		"prices": map[string]interface{}{
+			"currency": "USD",
+			"models": map[string]interface{}{
+				// USD PER TOKEN (compile.go divides the per-million figure by 1e6).
+				// prompt 0.15/M, completion 0.60/M. No reasoning_price anywhere.
+				"m": map[string]interface{}{"prompt": 0.15e-6, "completion": 0.60e-6},
+			},
+		},
+	}); err != nil {
+		t.Fatal(err)
+	}
+	ps.Fire(StageRequestEnd, map[string]interface{}{
+		"model": "m", "source": "s", "key": "***aaaaaa", "ok": true,
+		"prompt_tokens": 1000, "completion_tokens": 1000,
+		"reasoning_tokens": 800, // all of it thinking; still billed as output
+		"time": 1750000000000,
+	})
+	st := stateOf(t, ps)
+	cost := st["total"].(map[string]interface{})["cost"].(float64)
+	approx(t, "reasoning at completion rate", cost, 0.00075)
+}
+
+// TestReasoningPriceOverridesOnlyWhenSet: with a rule that DOES price thinking
+// differently, only the difference is applied. Charging the reasoning portion a
+// second time in full is the obvious wrong implementation and it doubles the
+// completion line.
+func TestReasoningPriceOverridesOnlyWhenSet(t *testing.T) {
+	ps, _ := billingVM(t)
+	if err := ps.SetState("billing", map[string]interface{}{
+		"prices": map[string]interface{}{
+			"currency": "USD",
+			"models": map[string]interface{}{
+				"m": map[string]interface{}{
+					"prompt": 0.15e-6, "completion": 0.60e-6,
+					// Thinking costs 3x ordinary output: 1.80/M.
+					"reasoning_price": 1.80e-6,
+				},
+			},
+		},
+	}); err != nil {
+		t.Fatal(err)
+	}
+	ps.Fire(StageRequestEnd, map[string]interface{}{
+		"model": "m", "source": "s", "key": "***aaaaaa", "ok": true,
+		"prompt_tokens": 1000, "completion_tokens": 1000, "reasoning_tokens": 800,
+		"time": 1750000000000,
+	})
+	st := stateOf(t, ps)
+	cost := st["total"].(map[string]interface{})["cost"].(float64)
+	// base = 1000*0.15e-6 + 1000*0.60e-6 = 0.00075
+	//      + 800 * (1.80-0.60)e-6   = 0.00096  →  0.00171
+	approx(t, "reasoning repriced to 1.80/M", cost, 0.00171)
+}
+
+// TestReasoningClampedToCompletion: a provider claiming MORE reasoning than
+// completion is malformed. Without a clamp the visible-output count goes
+// negative and the delta exceeds the whole request, so the bill silently
+// explodes on bad data instead of degrading.
+func TestReasoningClampedToCompletion(t *testing.T) {
+	ps, _ := billingVM(t)
+	if err := ps.SetState("billing", map[string]interface{}{
+		"prices": map[string]interface{}{
+			"currency": "USD",
+			"models": map[string]interface{}{
+				"m": map[string]interface{}{
+					"prompt": 0.15e-6, "completion": 0.60e-6, "reasoning_price": 1.80e-6,
+				},
+			},
+		},
+	}); err != nil {
+		t.Fatal(err)
+	}
+	ps.Fire(StageRequestEnd, map[string]interface{}{
+		"model": "m", "source": "s", "key": "***aaaaaa", "ok": true,
+		"prompt_tokens": 1000, "completion_tokens": 1000, "reasoning_tokens": 99999,
+		"time": 1750000000000,
+	})
+	st := stateOf(t, ps)
+	cost := st["total"].(map[string]interface{})["cost"].(float64)
+	// reasoning is clamped from 99999 down to completion (1000):
+	//   base  = 1000*0.15e-6 + 1000*0.60e-6            = 0.00075
+	//   delta = 1000 * (1.80-0.60)e-6                  = 0.00120
+	//   total                                          = 0.00195
+	// Without the clamp this would be 0.18 — 92x the whole request, and a
+	// number plausible enough to survive a glance at a dashboard.
+	approx(t, "reasoning clamped to completion", cost, 0.00195)
+}
+
+// TestLongContextStepIsMarginal: the threshold reprices only the tokens PAST
+// it, at a marginal rate.
+//
+// A flat prompt price under-bills exactly the traffic an agent workload
+// produces — on this deployment 88% of requests exceed 128k with a median of
+// 247k — and it under-bills silently, because the number stays plausible. The
+// step must apply to the excess only; charging the WHOLE prompt at the higher
+// rate is the other obvious implementation and it overcharges short requests by
+// the same factor it was meant to fix long ones.
+func TestLongContextStepIsMarginal(t *testing.T) {
+
+	// Short request: below the threshold, so the step must not touch it at all.
+	ps, _ := billingVM(t)
+	if err := ps.SetState("billing", map[string]interface{}{
+		"prices": map[string]interface{}{
+			"currency": "USD",
+			"models": map[string]interface{}{
+				"m": map[string]interface{}{
+					"prompt": 1e-6, "completion": 0,
+					"long_context": map[string]interface{}{"above": 128000, "prompt": 2e-6},
+				},
+			},
+		},
+	}); err != nil {
+		t.Fatal(err)
+	}
+	ps.Fire(StageRequestEnd, map[string]interface{}{
+		"model": "m", "source": "s", "key": "***aaaaaa", "ok": true,
+		"prompt_tokens": 100000, "completion_tokens": 0, "time": 1750000000000,
+	})
+	got := stateOf(t, ps)["total"].(map[string]interface{})["cost"].(float64)
+	approx(t, "short request unaffected", got, 0.1)
+
+	// Long request: 200k prompt with a 128k threshold → 72k tokens at the step.
+	ps2, _ := billingVM(t)
+	if err := ps2.SetState("billing", map[string]interface{}{
+		"prices": map[string]interface{}{
+			"currency": "USD",
+			"models": map[string]interface{}{
+				"m": map[string]interface{}{
+					"prompt": 1e-6, "completion": 0,
+					"long_context": map[string]interface{}{"above": 128000, "prompt": 2e-6},
+				},
+			},
+		},
+	}); err != nil {
+		t.Fatal(err)
+	}
+	ps2.Fire(StageRequestEnd, map[string]interface{}{
+		"model": "m", "source": "s", "key": "***aaaaaa", "ok": true,
+		"prompt_tokens": 200000, "completion_tokens": 0, "time": 1750000000000,
+	})
+	got2 := stateOf(t, ps2)["total"].(map[string]interface{})["cost"].(float64)
+	// 128000*1e-6 + 72000*2e-6 = 0.128 + 0.144 = 0.272
+	approx(t, "long request marginal step", got2, 0.272)
+}
+
+// TestLongContextComposesWithCacheDiscount pins the interaction, which is the
+// whole reason for computing the step as a marginal rate.
+//
+// On this deployment the cache hit rate is 88% and long prompts dominate, so
+// "cached tokens beyond the threshold" is the common case, not an edge case. A
+// cache read already bills at prompt*discount; the long-context rate applies on
+// top of THAT, not on top of the full prompt price. Computing it any other way
+// silently changes the price of the majority of traffic.
+func TestLongContextComposesWithCacheDiscount(t *testing.T) {
+	ps, _ := billingVM(t)
+	if err := ps.SetState("billing", map[string]interface{}{
+		"prices": map[string]interface{}{
+			"currency": "USD", "cache_discount": 0.1,
+			"models": map[string]interface{}{
+				"m": map[string]interface{}{
+					"prompt": 1e-6, "completion": 0,
+					"long_context": map[string]interface{}{"above": 100000, "prompt": 2e-6},
+				},
+			},
+		},
+	}); err != nil {
+		t.Fatal(err)
+	}
+	// 200k prompt, 150k of it cached, 50k fresh. Excess past 100k = 100k, split
+	// 25% fresh / 75% cached.
+	ps.Fire(StageRequestEnd, map[string]interface{}{
+		"model": "m", "source": "s", "key": "***aaaaaa", "ok": true,
+		"prompt_tokens": 200000, "completion_tokens": 0,
+		"cache_hit_tokens": 150000, "cache_reported": true,
+		"time": 1750000000000,
+	})
+	got := stateOf(t, ps)["total"].(map[string]interface{})["cost"].(float64)
+	// base:  fresh 50000*1e-6            = 0.05
+	//        cache 150000*1e-6*0.1        = 0.015
+	// step:  fresh excess 25000*(2-1)e-6 = 0.025
+	//        cache excess 75000*(2-0.1)e-6 = 0.1425
+	// total                             = 0.2325
+	approx(t, "step composes with cache discount", got, 0.2325)
+}

internal/lua/billing_ui_test.go

@@ -414,3 +414,86 @@ func TestRuleTableLocksColumnWidthsInAColgroup(t *testing.T) {
 			"control and must never be squeezed by the flexible inputs")
 	}
 }
+
+// TestBillingUIShowsRecentUnpriced: the page must be able to distinguish a
+// settled pricing debt from a live leak. The lifetime unpriced counter answers
+// neither — it keeps every request from before a rule was added, so it reads the
+// same number in both situations. The day buckets exist to make the difference
+// visible, and a bucket nothing renders is a bucket nobody benefits from.
+func TestBillingUIShowsRecentUnpriced(t *testing.T) {
+	raw, err := os.ReadFile("plugins/billing.lua")
+	if err != nil {
+		t.Fatal(err)
+	}
+	src := string(raw)
+	for _, needle := range []string{
+		"unpriced_by_day",  // the tally is written
+		"recentUnpriced",   // it is reduced to a window
+		"unpriced_recent",  // it reaches the view
+		"unpriced7d",       // the card labels it
+		"unpricedClean",    // and says so when the window is clear
+	} {
+		if !strings.Contains(src, needle) {
+			t.Errorf("billing UI lacks %q", needle)
+		}
+	}
+}
+
+// TestBillingRulesUIExposesReasoningPrice: the thinking-token rate must be
+// editable in the rules editor, not only settable by hand-editing config.yaml.
+// A pricing capability that lives only in the config file is one the operator
+// will assume does not exist.
+func TestBillingRulesUIExposesReasoningPrice(t *testing.T) {
+	raw, err := os.ReadFile("plugins/billing.lua")
+	if err != nil {
+		t.Fatal(err)
+	}
+	src := string(raw)
+	for _, needle := range []string{
+		`class='m-r'`,       // the input, in both the existing rows and the "+ model" row
+		`models[n].reasoning`, // collected back out of the DOM on save
+		`reasoningHint`,     // and it says what leaving it blank means
+	} {
+		if !strings.Contains(src, needle) {
+			t.Errorf("rules editor lacks %q", needle)
+		}
+	}
+	// Two occurrences: an existing model's row and the template for a new one.
+	// One would mean the "+ model" button produces a row that silently drops
+	// the rate on the next save.
+	if n := strings.Count(src, `class='m-r'`); n < 2 {
+		t.Errorf("m-r appears %d time(s); both the row template and the editor need it", n)
+	}
+}
+
+// TestBillingRulesUIExposesLongContextStep: the long-context step must be
+// editable in the rules editor.
+//
+// This is not a niche control. On an agent workload 88% of requests carry more
+// than 128k prompt tokens with a median of 247k, so a rule that omits the step
+// under-bills the MAJORITY of traffic while every number on the dashboard stays
+// plausible. Leaving the capability in config.yaml alone would mean the operator
+// never discovers it needs it.
+func TestBillingRulesUIExposesLongContextStep(t *testing.T) {
+	raw, err := os.ReadFile("plugins/billing.lua")
+	if err != nil {
+		t.Fatal(err)
+	}
+	src := string(raw)
+	for _, needle := range []string{
+		`class='m-lc-a'`,        // threshold
+		`class='m-lc-p'`,        // absolute rate
+		`class='m-lc-m'`,        // multiplier
+		`models[n].long_context`, // collected back out on save
+		`lcHint`,                // explains that blank = no step
+	} {
+		if !strings.Contains(src, needle) {
+			t.Errorf("rules editor lacks %q", needle)
+		}
+	}
+	// Both the existing-row template and the "+ model" template need the three
+	// inputs; one would make a newly added model lose its step on the next save.
+	if n := strings.Count(src, `class='m-lc-a'`); n < 2 {
+		t.Errorf("long-context inputs appear %d time(s); both row templates need them", n)
+	}
+}

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_persist.go

@@ -0,0 +1,115 @@
+package lua
+
+import (
+	"encoding/json"
+	"os"
+	"sort"
+)
+
+// Access overrides are OPERATOR policy, not plugin data. They live in their own
+// file rather than in each plugin's state because:
+//
+//   - they describe the deployment, not the plugin, so a plugin upgrade that
+//     replaces its .state.json must not silently widen access again;
+//   - they must survive a restart. An in-memory map made "restrict billing to
+//     admin" look like it worked right up to the next deploy, at which point the
+//     restriction evaporated with nothing in the logs to say why;
+//   - the plugin_dir is the one place an operator already looks when asking
+//     "what does this plugin actually get to see", so the answer lives beside it.
+//
+// The file is a flat map of "name" (whole plugin) or "name#path" (one endpoint)
+// to an access spec, which is the same shape the in-memory store uses, so load
+// and save cannot disagree about what a key means.
+
+const accessFileName = ".access.json"
+
+func (ps *Plugins) accessFile() string {
+	if ps.dir == "" {
+		return ""
+	}
+	return ps.dir + string(os.PathSeparator) + accessFileName
+}
+
+// LoadAccessOverrides reads the persisted overrides into memory. A missing file
+// is not an error: no operator has restricted anything yet, which is the
+// default state of a fresh install.
+//
+// A corrupt file is also not fatal. Losing a restriction is bad, but refusing to
+// start the gateway is worse, and an empty map plus a logged error leaves the
+// operator with something to fix.
+func (ps *Plugins) LoadAccessOverrides() error {
+	path := ps.accessFile()
+	if path == "" {
+		return nil
+	}
+	b, err := os.ReadFile(path)
+	if os.IsNotExist(err) {
+		return nil
+	}
+	if err != nil {
+		return err
+	}
+	m := map[string]string{}
+	if err := json.Unmarshal(b, &m); err != nil {
+		return err
+	}
+	ps.mu.Lock()
+	defer ps.mu.Unlock()
+	if ps.visibility == nil {
+		ps.visibility = visibilityOverrides{}
+	}
+	for k, v := range m {
+		ps.visibility[k] = v
+	}
+	return nil
+}
+
+// saveAccessOverridesLocked flushes the whole override map. It is called on
+// every change, and unlike plugin state the map is tiny and changes only when an
+// operator clicks Save — so a synchronous write costs nothing that matters, and
+// it means the change is durable the moment the API returns 200.
+func (ps *Plugins) saveAccessOverridesLocked() {
+	path := ps.accessFile()
+	if path == "" {
+		return
+	}
+	out := make(map[string]string, len(ps.visibility))
+	for k, v := range ps.visibility {
+		if v != "" {
+			out[k] = v
+		}
+	}
+	if len(out) == 0 {
+		// Leave no empty file behind: its absence is the honest encoding of
+		// "nothing is overridden".
+		_ = os.Remove(path)
+		return
+	}
+	writeJSONAtomic(path, out)
+}
+
+// AccessOverrides returns a copy of every override in force, for the UI.
+func (ps *Plugins) accessOverridesSnapshot() map[string]string {
+	ps.mu.RLock()
+	defer ps.mu.RUnlock()
+	if len(ps.visibility) == 0 {
+		return nil
+	}
+	out := make(map[string]string, len(ps.visibility))
+	for k, v := range ps.visibility {
+		out[k] = v
+	}
+	return out
+}
+
+// sortedOverrideKeys exists only so a debug dump is stable; map iteration order
+// would otherwise make two identical states look different.
+func (ps *Plugins) sortedOverrideKeys() []string {
+	m := ps.accessOverridesSnapshot()
+	keys := make([]string, 0, len(m))
+	for k := range m {
+		keys = append(keys, k)
+	}
+	sort.Strings(keys)
+	return keys
+}

internal/lua/plugin_access_persist_test.go

@@ -0,0 +1,110 @@
+package lua
+
+import (
+	"os"
+	"path/filepath"
+	"testing"
+)
+
+// TestAccessOverrideSurvivesReload is the test that matters for an access
+// restriction. An in-memory map makes "billing is admin-only" look like it
+// worked right up to the next restart, and then the restriction is simply gone
+// with nothing in the logs to explain it. Both the plugin-level audience and
+// the per-endpoint spec must come back.
+func TestAccessOverrideSurvivesReload(t *testing.T) {
+	dir := t.TempDir()
+	src := `local P = {}
+P.name = "api"
+P.version = "1.0.0"
+P.http = {
+  { path = "health", handler = "health", roles = { "admin" } },
+}
+function P.health(req) return { json = { ok = true } } end
+return P
+`
+	first := NewPlugins(NewVM(dir), dir)
+	if err := first.LoadSource("api", src); err != nil {
+		t.Fatal(err)
+	}
+	first.SetPluginVisibility("api", "admin, user")
+	if err := first.SetEndpointAccess("api", "health", "@homeagent"); err != nil {
+		t.Fatal(err)
+	}
+	if _, err := os.Stat(filepath.Join(dir, accessFileName)); err != nil {
+		t.Fatalf("override was not written: %v", err)
+	}
+
+	// A fresh Plugins over the same dir is what a restart looks like.
+	second := NewPlugins(NewVM(dir), dir)
+	// The plugin source is on disk too, so this second instance sees both the
+	// plugin and the overrides — the same pair a restart sees. Loading only the
+	// overrides would leave no endpoints to walk, and the assertion below would
+	// pass vacuously.
+	if err := second.LoadSource("api", src); err != nil {
+		t.Fatal(err)
+	}
+	if err := second.LoadAccessOverrides(); err != nil {
+		t.Fatalf("reload overrides: %v", err)
+	}
+	if got := second.VisibilityOverride("api"); got != "admin, user" {
+		t.Errorf("plugin visibility lost across reload: %q", got)
+	}
+	found := false
+	for _, e := range second.EndpointAccess() {
+		if e.Path == "health" && len(e.OverrideKeys) == 1 && e.OverrideKeys[0] == "homeagent" {
+			found = true
+		}
+	}
+	if !found {
+		t.Error("endpoint override lost across reload")
+	}
+}
+
+// TestAccessOverrideClearedRemovesFile keeps the "nothing overridden" state
+// distinguishable from "overridden with something".
+func TestAccessOverrideClearedRemovesFile(t *testing.T) {
+	dir := t.TempDir()
+	ps := NewPlugins(NewVM(dir), dir)
+	ps.SetPluginVisibility("api", "admin")
+	path := filepath.Join(dir, accessFileName)
+	if _, err := os.Stat(path); err != nil {
+		t.Fatalf("expected an access file: %v", err)
+	}
+	ps.SetPluginVisibility("api", "")
+	if _, err := os.Stat(path); !os.IsNotExist(err) {
+		t.Errorf("clearing the last override should remove the file, got %v", err)
+	}
+}
+
+// TestVisibilityOverridesExcludeEndpointKeys keeps endpoint policy out of the
+// plugin-audience listing. Both live in one map, and reporting "probe#hello"
+// under visibility is how the plugin page ends up offering to save an audience
+// of that name — which nothing reads as an audience.
+func TestVisibilityOverridesExcludeEndpointKeys(t *testing.T) {
+	dir := t.TempDir()
+	src := `local P = {}
+P.name = "api"
+P.version = "1.0.0"
+P.http = { { path = "health", handler = "health", roles = { "admin" } } }
+function P.health(req) return { json = { ok = true } } end
+return P
+`
+	ps := NewPlugins(NewVM(dir), dir)
+	if err := ps.LoadSource("api", src); err != nil {
+		t.Fatal(err)
+	}
+	ps.SetPluginVisibility("api", "admin, user")
+	if err := ps.SetEndpointAccess("api", "health", "@homeagent"); err != nil {
+		t.Fatal(err)
+	}
+	got := ps.VisibilityOverrides()
+	if got["api"] != "admin, user" {
+		t.Errorf("plugin audience missing: %+v", got)
+	}
+	if _, leaked := got["api#health"]; leaked {
+		t.Errorf("endpoint override leaked into the audience map: %+v", got)
+	}
+	if len(got) != 1 {
+		t.Errorf("expected exactly the plugin audience, got %+v", got)
+	}
+}

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_http.go

@@ -0,0 +1,248 @@
+package lua
+
+// Outbound HTTP for plugins.
+//
+// The billing plugin never needed this: everything it reports arrives on a hook
+// payload that the gateway produced while forwarding a request. A quota plugin
+// is the first thing that must REACH OUT on its own — Command Code exposes
+// /alpha/billing/credits and friends, and there is no way to learn a plan's
+// remaining balance without asking.
+//
+// That capability is also the sharpest edge in the plugin model. A plugin is a
+// single .lua file the WebUI installs from a textarea: untrusted, browser-facing
+// code that runs in the gateway process. Handing it a general-purpose HTTP
+// client would make every install a standing SSRF primitive against whatever
+// the gateway can reach — the cloud metadata endpoint, the cluster control
+// plane, whatever else sits on 127.0.0.1 in this deployment. Reading the
+// environment for credentials is a second door out.
+//
+// So the bridge is deliberately narrow:
+//
+//   - Destination must match a host that a CONFIGURED SOURCE already talks to.
+//     Those are operator-declared, not plugin-declared: a plugin cannot widen
+//     its own reach by inventing a URL, and an existing source's host is
+//     exactly the set of upstreams this deployment already trusts with API keys.
+//   - http:// on a non-loopback host is refused. Credentials belong in HTTPS.
+//   - Loopback needs no explicit opt-in: it is the only way a plugin can read
+//     the gateway's own API (the billing rules UI does exactly that), and it
+//     cannot leave the machine.
+//   - Every request is bounded by a timeout and the response is size-capped, so
+//     a plugin cannot wedge a worker goroutine or exhaust memory.
+//   - Credentials are NEVER read from the environment here. A plugin that needs
+//     a key must be given it (state, or the caller's own config), which keeps
+//     the grant visible to the operator instead of silently ambient.
+//
+// Redirects are NOT followed: the allowlist is checked per hop, and a redirect
+// to an unlisted host would otherwise walk straight past it. A 3xx is returned
+// as-is so the plugin can see it.
+
+import (
+	"fmt"
+	"io"
+	"net"
+	"net/http"
+	"net/url"
+	"strings"
+	"time"
+
+	golua "github.com/aarzilli/golua/lua"
+)
+
+// httpDefaultTimeout bounds a whole request; httpMaxResponseBytes caps the body
+// a plugin can pull into the gateway's heap.
+const (
+	httpDefaultTimeout = 15 * time.Second
+	httpMaxTimeout     = 60 * time.Second
+	httpMaxResponse    = 4 << 20 // 4 MiB
+)
+
+// httpClient is shared by every plugin request. It deliberately has no
+// ProxyFromEnvironment: an operator who proxies upstreams configures the proxy
+// per source, and a plugin request should not silently inherit one.
+var httpClient = &http.Client{
+	Timeout: httpDefaultTimeout,
+	Transport: &http.Transport{
+		Proxy:                 nil,
+		DialContext:           (&net.Dialer{Timeout: 10 * time.Second, KeepAlive: 30 * time.Second}).DialContext,
+		ForceAttemptHTTP2:     true,
+		MaxIdleConns:          16,
+		IdleConnTimeout:       60 * time.Second,
+		TLSHandshakeTimeout:   10 * time.Second,
+		ExpectContinueTimeout: time.Second,
+	},
+	// See above: a redirect is a hop the allowlist has not seen.
+	CheckRedirect: func(*http.Request, []*http.Request) error {
+		return http.ErrUseLastResponse
+	},
+}
+
+// httpGlobal is the Lua name of the outbound bridge.
+const httpGlobal = "__llmsproxy_http"
+
+// httpAllowlist is the set of hosts plugins may contact: the host of every
+// configured source's base_url.
+type httpAllowlist struct {
+	hosts map[string]bool
+}
+
+func newHTTPAllowlist(hosts []string) *httpAllowlist {
+	m := make(map[string]bool, len(hosts))
+	for _, h := range hosts {
+		if h != "" {
+			m[strings.ToLower(h)] = true
+		}
+	}
+	return &httpAllowlist{hosts: m}
+}
+
+// allows reports whether host:port may be contacted. Ports 80/443 are implied
+// by the bare host so a source configured as https://api.example.com/v1 does not
+// have to spell the port.
+func (a *httpAllowlist) allows(host string, port string) bool {
+	host = strings.ToLower(host)
+	if host == "" {
+		return false
+	}
+	if a.hosts[host] {
+		return true
+	}
+	switch port {
+	case "", "80", "443":
+		return a.hosts[host]
+	}
+	return a.hosts[net.JoinHostPort(host, port)]
+}
+
+// checkURL validates a plugin-supplied URL against the allowlist. It returns
+// the parsed URL plus a reason when it is refused.
+func (a *httpAllowlist) checkURL(raw string) (*url.URL, error) {
+	u, err := url.Parse(strings.TrimSpace(raw))
+	if err != nil {
+		return nil, fmt.Errorf("bad url: %v", err)
+	}
+	switch strings.ToLower(u.Scheme) {
+	case "https":
+	case "http":
+		// Plain HTTP is only tolerated on loopback, where it stays on this
+		// machine and is how the gateway's own plain-HTTP dev port is reached.
+		if !isLoopbackHost(u.Hostname()) {
+			return nil, fmt.Errorf("refusing plain http to non-loopback host %q: use https", u.Hostname())
+		}
+	default:
+		return nil, fmt.Errorf("unsupported scheme %q", u.Scheme)
+	}
+	if !a.allows(u.Hostname(), u.Port()) && !isLoopbackHost(u.Hostname()) {
+		// Loopback needs no entry: it cannot leave the machine, and the billing
+		// rules UI reads the gateway's own API over plain HTTP in dev setups.
+		return nil, fmt.Errorf("host %q is not an allowed outbound destination "+
+			"(allowed: the hosts of configured sources, plus loopback)", u.Hostname())
+	}
+	return u, nil
+}
+
+func isLoopbackHost(h string) bool {
+	if h == "localhost" {
+		return true
+	}
+	ip := net.ParseIP(strings.Trim(h, "[]"))
+	return ip != nil && ip.IsLoopback()
+}
+
+// httpRequestSpec is the decoded argument table Lua passes to __llmsproxy_http.
+type httpRequestSpec struct {
+	URL     string
+	Method  string
+	Headers map[string]string
+	Body    string
+	Timeout time.Duration
+}
+
+// luaHTTPFn builds the golua closure. It takes one Lua table argument and
+// returns (responseTable, nil) or (nil, errorMessage) — never raises, so a
+// plugin mistake cannot take the gateway with it.
+func luaHTTPFn(L *golua.State, allow *httpAllowlist) golua.LuaGoFunction {
+	return func(L *golua.State) int {
+		spec, err := readHTTPRequestSpec(L)
+		if err != nil {
+			L.PushNil()
+			L.PushString(err.Error())
+			return 2
+		}
+		u, err := allow.checkURL(spec.URL)
+		if err != nil {
+			L.PushNil()
+			L.PushString(err.Error())
+			return 2
+		}
+		resp, err := performHTTP(u, spec)
+		if err != nil {
+			L.PushNil()
+			L.PushString(err.Error())
+			return 2
+		}
+		pushHTTPResponse(L, resp)
+		L.PushNil()
+		return 2
+	}
+}
+
+func performHTTP(u *url.URL, spec httpRequestSpec) (*httpResult, error) {
+	method := strings.ToUpper(spec.Method)
+	if method == "" {
+		method = http.MethodGet
+	}
+	timeout := spec.Timeout
+	if timeout <= 0 || timeout > httpMaxTimeout {
+		timeout = httpDefaultTimeout
+	}
+
+	var body io.Reader
+	if spec.Body != "" {
+		body = strings.NewReader(spec.Body)
+	}
+	req, err := http.NewRequest(method, u.String(), body)
+	if err != nil {
+		return nil, fmt.Errorf("build request: %v", err)
+	}
+	for k, v := range spec.Headers {
+		req.Header.Set(k, v)
+	}
+	if req.Header.Get("User-Agent") == "" {
+		req.Header.Set("User-Agent", "ModelRouter-plugin")
+	}
+
+	client := *httpClient
+	client.Timeout = timeout
+	resp, err := client.Do(req)
+	if err != nil {
+		return nil, fmt.Errorf("request failed: %v", err)
+	}
+	defer resp.Body.Close()
+
+	data, err := io.ReadAll(io.LimitReader(resp.Body, httpMaxResponse+1))
+	if err != nil {
+		return nil, fmt.Errorf("read body: %v", err)
+	}
+	if len(data) > httpMaxResponse {
+		return nil, fmt.Errorf("response exceeds %d byte cap", httpMaxResponse)
+	}
+	return &httpResult{
+		Status:  resp.StatusCode,
+		Headers: flattenHeaders(resp.Header),
+		Body:    string(data),
+	}, nil
+}
+
+type httpResult struct {
+	Status  int
+	Headers map[string]string
+	Body    string
+}
+
+func flattenHeaders(h http.Header) map[string]string {
+	out := make(map[string]string, len(h))
+	for k := range h {
+		out[strings.ToLower(k)] = h.Get(k)
+	}
+	return out
+}

internal/lua/plugin_http_args.go

@@ -0,0 +1,122 @@
+package lua
+
+import (
+	"fmt"
+	"strings"
+	"time"
+
+	golua "github.com/aarzilli/golua/lua"
+)
+
+// httpArgIdx is the stack slot holding the caller's table. Verified against
+// golua: a PushGoClosure function is invoked with its arguments already on top
+// and nothing of itself below them.
+const httpArgIdx = 1
+
+// readHTTPRequestSpec decodes the Lua argument table:
+//
+//	__llmsproxy_http({
+//	  url = "https://host/path",
+//	  method = "GET",              -- optional, default GET
+//	  headers = { ["X-A"] = "1" }, -- optional
+//	  body = "...",                -- optional
+//	  timeout_ms = 8000,           -- optional, clamped to 60s
+//	})
+//
+// The table is at index 1: measured, not assumed. A Go closure registered with
+// PushGoClosure is NOT on the stack when it runs — GetTop() is 1 and slot 1 is
+// the caller's table (type LUA_TTABLE). An earlier version assumed the closure
+// occupied slot 1 and read from slot 2, which then reported "expected a table
+// argument" for a perfectly good call.
+//
+// luaTableToGoValue already turns the table into a map/slice, so the whole
+// argument is decoded in one pass instead of walking the table field by field.
+func readHTTPRequestSpec(L *golua.State) (httpRequestSpec, error) {
+	var spec httpRequestSpec
+	if L.GetTop() < httpArgIdx {
+		return spec, fmt.Errorf("__llmsproxy_http: expected a table argument")
+	}
+	if L.Type(httpArgIdx) != golua.LUA_TTABLE {
+		return spec, fmt.Errorf("__llmsproxy_http: expected a table, got %s",
+			normalizeString(L, httpArgIdx))
+	}
+	arg, ok := luaTableToGoValue(L, httpArgIdx).(map[string]interface{})
+	if !ok {
+		return spec, fmt.Errorf("__llmsproxy_http: argument must be a keyed table")
+	}
+	for k, v := range arg {
+		switch strings.ToLower(k) {
+		case "url", "uri", "endpoint":
+			spec.URL = asString(v)
+		case "method", "verb":
+			spec.Method = asString(v)
+		case "body", "data", "payload":
+			spec.Body = asString(v)
+		case "headers", "header":
+			if m, ok := v.(map[string]interface{}); ok {
+				spec.Headers = make(map[string]string, len(m))
+				for hk, hv := range m {
+					spec.Headers[hk] = asString(hv)
+				}
+			}
+		case "timeout_ms", "timeoutms":
+			if n, ok := asFloat(v); ok {
+				spec.Timeout = time.Duration(n) * time.Millisecond
+			}
+		}
+	}
+	if spec.URL == "" {
+		return spec, fmt.Errorf("__llmsproxy_http: missing url")
+	}
+	return spec, nil
+}
+
+// pushHTTPResponse converts the result into two Lua values: a table
+// {status, headers, body} and nil. The caller pushes the error slot itself.
+func pushHTTPResponse(L *golua.State, r *httpResult) {
+	L.NewTable()
+
+	L.PushString("status")
+	L.PushNumber(float64(r.Status))
+	L.SetTable(-3)
+
+	L.PushString("headers")
+	pushStringMap(L, r.Headers)
+	L.SetTable(-3)
+
+	L.PushString("body")
+	L.PushString(r.Body)
+	L.SetTable(-3)
+}
+
+func pushStringMap(L *golua.State, m map[string]string) {
+	L.NewTable()
+	for k, v := range m {
+		L.PushString(v)
+		L.SetField(-2, k)
+	}
+}
+
+func asString(v interface{}) string {
+	switch t := v.(type) {
+	case string:
+		return t
+	case nil:
+		return ""
+	default:
+		return fmt.Sprintf("%v", t)
+	}
+}
+
+func asFloat(v interface{}) (float64, bool) {
+	switch t := v.(type) {
+	case float64:
+		return t, true
+	case int:
+		return float64(t), true
+	case int64:
+		return float64(t), true
+	default:
+		return 0, false
+	}
+}

internal/lua/plugin_http_handler.go

@@ -0,0 +1,530 @@
+package lua
+
+// Plugin-registered HTTP endpoints.
+//
+// The gateway mounts what a plugin declares in plugin.http and calls back into
+// the plugin's Lua state to produce each response. This is the piece billing
+// never needed: it only receives hooks, so it has nothing to expose. A plugin
+// that computes something an operator wants to query from outside the browser —
+// a health probe, a CI check, another tool — needs a route, and hard-coding one
+// per plugin would mean a gateway change for every such plugin.
+//
+// The request and response are both plain tables crossing the Go/Lua boundary:
+//
+//	request  { method, path, query{}, params{}, headers{}, body, role, key }
+//	response { status, headers{}, body }  or  { status, json = <any table> }
+//
+// A handler returns (response, nil) or (nil, errorMessage). Errors never raise:
+// an escaping Lua error inside a hook has taken the whole gateway down before
+// (see hookGuardSrc), so the same discipline applies here.
+//
+// Two properties are enforced on the Go side rather than trusted to the plugin:
+//
+//   - Method and status are checked against an allowlist before the response is
+//     written. A plugin returning status 0 or a redirect with an injected
+//     Location must not be able to steer the gateway's own routing.
+//   - Path traversal cannot escape the mount: params are substituted into a
+//     fixed segment count, and the handler's own Path is rejected at
+//     registration if it contains "..".
+
+import (
+	"encoding/json"
+	"fmt"
+	"net/http"
+	"strings"
+)
+
+// allowedMethods is deliberately short. An arbitrary verb would need a case in
+// the router for no benefit, and a plugin that wants every verb can register the
+// five that exist.
+var allowedMethods = map[string]bool{
+	http.MethodGet: true, http.MethodPost: true, http.MethodPut: true,
+	http.MethodDelete: true, http.MethodPatch: true,
+}
+
+// normalizeHandlerMethod validates and defaults a declared method.
+func normalizeHandlerMethod(m string) (string, error) {
+	m = strings.ToUpper(strings.TrimSpace(m))
+	if m == "" {
+		return http.MethodGet, nil
+	}
+	if !allowedMethods[m] {
+		return "", fmt.Errorf("unsupported method %q (allowed: GET POST PUT DELETE PATCH)", m)
+	}
+	return m, nil
+}
+
+// validateHandlerPath rejects paths that could escape the mount or collide with
+// the plugin state/call routes.
+func validateHandlerPath(p string) error {
+	p = strings.TrimSpace(p)
+	if p == "" {
+		return fmt.Errorf("empty path")
+	}
+	if strings.Contains(p, "..") {
+		return fmt.Errorf("path %q contains .. ", p)
+	}
+	if strings.Contains(p, "//") {
+		return fmt.Errorf("path %q contains an empty segment", p)
+	}
+	for _, seg := range strings.Split(strings.Trim(p, "/"), "/") {
+		if seg == "" {
+			continue
+		}
+		if strings.HasPrefix(seg, ":") {
+			if len(seg) == 1 {
+				return fmt.Errorf("path %q has a nameless param segment", p)
+			}
+			continue
+		}
+		// A literal segment must not look like a wildcard, or the route table
+		// gets two patterns that match the same URL.
+		if strings.ContainsAny(seg, "*?") {
+			return fmt.Errorf("path %q uses a wildcard; use :name for a param", p)
+		}
+	}
+	return nil
+}
+
+// handlerRequest is what a Lua handler sees.
+type handlerRequest struct {
+	Method  string
+	Path    string
+	Query   map[string]string
+	Params  map[string]string
+	Headers map[string]string
+	Body    string
+	Role    string
+	Key     string
+}
+
+// handlerResponse is what a Lua handler returns.
+type handlerResponse struct {
+	Status  int
+	Headers map[string]string
+	Body    string
+	// JSON is set when the handler returned {json = <table>}; it is encoded by
+	// the kernel so a plugin does not hand-roll escaping.
+	JSON interface{}
+}
+
+// writeHTTPHandlerResponse renders a handler's reply onto the gateway's writer.
+//
+// The status is validated even though it came from a plugin: a handler
+// returning 0 or 999 would make net/http panic, and one returning 3xx with a
+// Location the plugin chose could be used to bounce a caller at a URL of the
+// plugin's picking. Anything outside 100..599 that is not a redirect becomes 500
+// with an error body rather than a panic.
+func writeHTTPHandlerResponse(w http.ResponseWriter, r *handlerResponse) {
+	status := r.Status
+	if status == 0 {
+		status = http.StatusOK
+	}
+	if status < 100 || status > 599 {
+		status = http.StatusInternalServerError
+	}
+	for k, v := range r.Headers {
+		// A plugin cannot set Content-Length (the body is written by the kernel)
+		// or hop-by-hop headers (Connection, Transfer-Encoding) which would
+		// corrupt the gateway's own response framing.
+		switch http.CanonicalHeaderKey(k) {
+		case "Content-Length", "Connection", "Transfer-Encoding", "Upgrade":
+			continue
+		}
+		w.Header().Set(k, v)
+	}
+	if r.JSON != nil {
+		w.Header().Set("Content-Type", "application/json; charset=utf-8")
+		w.WriteHeader(status)
+		// Encoded here rather than by the gateway layer: this file is in the lua
+		// package and importing the gateway would be a cycle.
+		if enc, err := json.Marshal(r.JSON); err == nil {
+			_, _ = w.Write(enc)
+		} else {
+			_, _ = w.Write([]byte(`{"error":"handler json could not be encoded"}`))
+		}
+		return
+	}
+	if w.Header().Get("Content-Type") == "" {
+		w.Header().Set("Content-Type", "text/plain; charset=utf-8")
+	}
+	w.WriteHeader(status)
+	if r.Body != "" {
+		_, _ = w.Write([]byte(r.Body))
+	}
+}
+
+// parseHandlerResponse reads a Lua handler's return value.
+//
+// It accepts three shapes, because a plugin should not have to think about it:
+//
+//	{ status = 200, body = "text" }
+//	{ status = 201, json = { any = "table" } }
+//	{ status = 204 }                       -- no body
+//
+// Anything else — a string, a number, a table without status — is a handler bug
+// and becomes an error rather than an empty 200, so the plugin author sees it.
+func parseHandlerResponse(v interface{}) (*handlerResponse, error) {
+	m, ok := v.(map[string]interface{})
+	if !ok {
+		return nil, fmt.Errorf("handler must return a table, got %T", v)
+	}
+	out := &handlerResponse{Status: http.StatusOK}
+	if s, ok := m["status"]; ok {
+		switch n := s.(type) {
+		case float64:
+			out.Status = int(n)
+		case int:
+			out.Status = n
+		default:
+			return nil, fmt.Errorf("status must be a number, got %T", s)
+		}
+	}
+	if b, ok := m["body"].(string); ok {
+		out.Body = b
+	}
+	if h, ok := m["headers"].(map[string]interface{}); ok {
+		out.Headers = make(map[string]string, len(h))
+		for k, hv := range h {
+			out.Headers[k] = asString(hv)
+		}
+	}
+	if j, ok := m["json"]; ok && j != nil {
+		out.JSON = j
+	}
+	// An empty table is a handler bug — it means the author forgot to say what to
+	// return. A table carrying ONLY a status is not: `{status=204}` is the normal
+	// way to answer with no body, and treating it as an error would make the
+	// correct spelling unusable.
+	if out.Body == "" && out.JSON == nil && out.Headers == nil && len(m) == 0 {
+		return nil, fmt.Errorf("handler returned an empty table; return " +
+			"{status=…, body=…}, {status=…, json=…}, or {status=…} for no body")
+	}
+	return out, nil
+}
+
+// InvokeHandler calls a plugin's HTTP handler function with a request table and
+// returns the decoded response.
+//
+// It reuses the same guard-in-Lua contract as the hooks: the function is called
+// through __llmsproxy_call_hook, so a plugin error comes back as (nil, message)
+// and never as a C-level abort.
+func (ps *Plugins) InvokeHandler(plugin, fnName string, req *handlerRequest) (*handlerResponse, error) {
+	p := ps.find(plugin)
+	if p == nil {
+		return nil, fmt.Errorf("plugin %s not found", plugin)
+	}
+	p.mu.Lock()
+	w := p.state
+	p.mu.Unlock()
+	if w == nil {
+		return nil, fmt.Errorf("plugin %s has no state", plugin)
+	}
+
+	// The whole call happens under the plugin lock, network waits included —
+	// same property as Call, and for the same reason: one Lua state.
+	p.mu.Lock()
+	defer p.mu.Unlock()
+	L := w.L
+	L.SetTop(0)
+	defer L.SetTop(0)
+
+	L.GetGlobal(pluginGlobal)
+	if L.IsNil(-1) {
+		return nil, fmt.Errorf("plugin table missing")
+	}
+	plug := L.GetTop()
+	L.GetGlobal(hookGuardName)
+	if L.IsNil(-1) {
+		L.Pop(1)
+	}
+	L.GetField(plug, fnName)
+	if !L.IsFunction(-1) {
+		return nil, fmt.Errorf("handler %s not found in plugin %s", fnName, plugin)
+	}
+
+	pushGoValue(L, plainForLua(map[string]interface{}{
+		"method":  req.Method,
+		"path":    req.Path,
+		"query":   stringMapToAny(req.Query),
+		"params":  stringMapToAny(req.Params),
+		"headers": stringMapToAny(req.Headers),
+		"body":    req.Body,
+		"role":    req.Role,
+		"key":     req.Key,
+	}))
+	if err := L.Call(2, 2); err != nil {
+		return nil, err
+	}
+	errMsg := ""
+	if L.GetTop() >= 2 {
+		if !L.IsNil(-1) {
+			errMsg = L.ToString(-1)
+		}
+		L.Pop(1)
+	}
+	if errMsg != "" {
+		return nil, fmt.Errorf("plugin %s: %s", plugin, errMsg)
+	}
+	if L.GetTop() < 1 || L.IsNil(-1) {
+		return nil, fmt.Errorf("handler %s returned nothing", fnName)
+	}
+	var raw interface{}
+	if err := luaToJSON(L, -1, &raw); err != nil {
+		return nil, err
+	}
+	return parseHandlerResponse(raw)
+}
+
+func stringMapToAny(m map[string]string) map[string]interface{} {
+	if len(m) == 0 {
+		return map[string]interface{}{}
+	}
+	out := make(map[string]interface{}, len(m))
+	for k, v := range m {
+		out[k] = v
+	}
+	return out
+}
+
+// HandlerRequest is the request table handed to a plugin HTTP handler.
+// Exported because the gateway package constructs it.
+type HandlerRequest = handlerRequest
+
+// HandlerResponse is what a handler returns, after decoding.
+type HandlerResponse = handlerResponse
+
+// WriteHandlerResponse renders a handler response onto the gateway's writer.
+func WriteHandlerResponse(w http.ResponseWriter, r *HandlerResponse) {
+	writeHTTPHandlerResponse(w, r)
+}
+
+// HTTPHandlers returns the registered endpoints across all plugins, each tagged
+// with its owner so the dispatcher knows which Lua state to enter.
+func (ps *Plugins) HTTPHandlers() []HTTPHandler {
+	ps.mu.RLock()
+	defer ps.mu.RUnlock()
+	var out []HTTPHandler
+	for _, p := range ps.plugins {
+		if p.http == nil || p.LoadError != "" {
+			continue
+		}
+		for _, declared := range p.http {
+			// effectiveAccess applies the operator's per-endpoint override, so the
+			// dispatcher enforces what the WebUI shows rather than what the .lua
+			// declared.
+			h := ps.effectiveAccess(p, declared)
+			h.Plugin = p.Info.Name
+			out = append(out, h)
+		}
+	}
+	return out
+}
+
+// Allows is the exported form of the access rule, so the gateway dispatcher does
+// not need a duplicate copy of it in its own package.
+func (h HTTPHandler) Allows(role, keyName string) bool { return h.allows(role, keyName) }
+
+// HTTPEndpointAccess is the operator's view of one endpoint's permissions,
+// returned by /api/plugins so the WebUI can render what is exposed without
+// parsing .lua files.
+type HTTPEndpointAccess struct {
+	Plugin      string   `json:"plugin"`
+	Path        string   `json:"path"`
+	FullPath    string   `json:"full_path"`
+	Method      string   `json:"method"`
+	Handler     string   `json:"handler"`
+	Roles       []string `json:"roles,omitempty"`
+	Keys        []string `json:"keys,omitempty"`
+	Public      bool     `json:"public"`
+	Description string   `json:"description,omitempty"`
+	// OverrideRoles / OverrideKeys carry the operator's own restriction, which
+	// REPLACES whatever the plugin declared. Empty means "use the plugin's".
+	//
+	// This is the mechanism the admin uses to narrow (or widen) an endpoint's
+	// reachable keys without editing the plugin: a bundled .lua is overwritten on
+	// the next upgrade, and a hand-installed one is not something the operator
+	// should have to maintain.
+	OverrideRoles []string `json:"override_roles,omitempty"`
+	OverrideKeys  []string `json:"override_keys,omitempty"`
+}
+
+// endpointAccess returns the access table for every registered endpoint, with
+// operator overrides folded in.
+func (ps *Plugins) endpointAccess() []HTTPEndpointAccess {
+	ps.mu.RLock()
+	defer ps.mu.RUnlock()
+	return ps.endpointAccessLocked()
+}
+
+// endpointAccessLocked is the body of endpointAccess for callers that already
+// hold ps.mu — VisibilityOverrides walks the same map and must not re-lock.
+func (ps *Plugins) endpointAccessLocked() []HTTPEndpointAccess {
+	var out []HTTPEndpointAccess
+	for _, p := range ps.plugins {
+		if p.http == nil || p.LoadError != "" {
+			continue
+		}
+		for _, h := range p.http {
+			// The override key is per ENDPOINT (plugin#path), not per plugin.
+			// Reading the plugin-level key here reported the plugin's own
+			// audience override as if it governed every endpoint, and reported
+			// no endpoint override at all — so the UI drew the DECLARED roles
+			// after an operator had saved a change, and re-saving from that
+			// display silently reverted the restriction.
+			ovRoles, ovKeys := splitAccessSpec(ps.visibility[endpointOverrideKey(p.Info.Name, h.Path)])
+			out = append(out, HTTPEndpointAccess{
+				Plugin: p.Info.Name, Path: h.Path, FullPath: h.routePath(),
+				Method: h.Method, Handler: h.Handler,
+				Roles: h.Roles, Keys: h.Keys, Public: h.Public,
+				Description:   h.Description,
+				OverrideRoles: ovRoles, OverrideKeys: ovKeys,
+			})
+		}
+	}
+	return out
+}
+
+// EndpointAccess exposes the table to the gateway package.
+func (ps *Plugins) EndpointAccess() []HTTPEndpointAccess { return ps.endpointAccess() }
+
+// SetEndpointAccess narrows or widens one endpoint's reachable keys/roles.
+//
+// It is an OVERRIDE, applied on top of what the plugin declared, and stored in
+// the same per-plugin visibility map the UI audience uses. An empty value clears
+// it. Overriding per endpoint rather than per plugin is what the operator
+// actually wants: /health is fine for every key while /balance is not, and
+// those two live in the same plugin.
+func (ps *Plugins) SetEndpointAccess(plugin, path, spec string) error {
+	p := ps.find(plugin)
+	if p == nil {
+		return fmt.Errorf("plugin %s not found", plugin)
+	}
+	found := false
+	for _, h := range p.http {
+		if h.Path == path {
+			found = true
+			break
+		}
+	}
+	if !found {
+		return fmt.Errorf("plugin %s has no endpoint %q", plugin, path)
+	}
+	ps.mu.Lock()
+	defer ps.mu.Unlock()
+	if ps.visibility == nil {
+		ps.visibility = visibilityOverrides{}
+	}
+	key := endpointOverrideKey(plugin, path)
+	if strings.TrimSpace(spec) == "" {
+		delete(ps.visibility, key)
+	} else {
+		ps.visibility[key] = spec
+	}
+	ps.saveAccessOverridesLocked()
+	return nil
+}
+
+// endpointOverrideKey namespaces a per-endpoint override so it cannot collide
+// with a whole-plugin audience override of the same plugin name.
+func endpointOverrideKey(plugin, path string) string {
+	return plugin + "#" + path
+}
+
+// effectiveAccess applies an endpoint override to the declared handler.
+func (ps *Plugins) effectiveAccess(p *Plugin, h HTTPHandler) HTTPHandler {
+	ps.mu.RLock()
+	ov := ps.visibility[endpointOverrideKey(p.Info.Name, h.Path)]
+	ps.mu.RUnlock()
+	if ov == "" {
+		return h
+	}
+	out := h
+	out.Roles, out.Keys = nil, nil
+	out.Public = false
+	for _, part := range strings.Split(ov, ",") {
+		part = strings.TrimSpace(part)
+		switch {
+		case part == "":
+		case part == "public":
+			out.Public = true
+		case strings.HasPrefix(part, "@"):
+			out.Keys = append(out.Keys, strings.TrimPrefix(part, "@"))
+		default:
+			out.Roles = append(out.Roles, part)
+		}
+	}
+	return out
+}
+
+// knownKeyNames is set by the gateway during wiring so ValidateAccessSpec can
+// refuse an allowlist naming a key that does not exist.
+//
+// Without the check, an allowlist naming a typo ("homeagent" vs "home_agent")
+// is stored, saved, shown in the WebUI as configured — and refuses every caller,
+// including the one the operator meant. The failure looks like the endpoint being
+// broken rather than the name being wrong.
+var knownKeyNames func() []string
+
+// SetKeyNameLookup registers the source of configured gateway key names.
+func SetKeyNameLookup(fn func() []string) { knownKeyNames = fn }
+
+// ValidateAccessSpec checks an access spec against the configured keys and roles.
+//
+// Roles are checked against the two that exist (admin, user) rather than left
+// free-form: a role that does not exist produces an endpoint nobody can reach,
+// which is the same silent failure as a bad key name.
+func (ps *Plugins) ValidateAccessSpec(spec string) error {
+	spec = strings.TrimSpace(spec)
+	if spec == "" {
+		return nil
+	}
+	known := map[string]bool{"public": true, "admin": true, "user": true}
+	if knownKeyNames != nil {
+		for _, n := range knownKeyNames() {
+			if n != "" {
+				known[strings.ToLower(n)] = true
+			}
+		}
+	}
+	for _, part := range strings.Split(spec, ",") {
+		part = strings.TrimSpace(part)
+		if part == "" {
+			continue
+		}
+		name := part
+		if strings.HasPrefix(name, "@") {
+			name = strings.TrimPrefix(name, "@")
+		}
+		if name == "" {
+			return fmt.Errorf("access entry %q names no key", part)
+		}
+		if !known[strings.ToLower(name)] {
+			return fmt.Errorf("access entry %q is neither a role (admin/user/public) "+
+				"nor a configured key name", part)
+		}
+	}
+	return nil
+}
+
+// splitAccessSpec turns an override string into role and key lists. "@name" is a
+// key; a bare word is a role. Shared by the reporting path and the enforcement
+// path so the two can never disagree about what an override means.
+func splitAccessSpec(spec string) (roles, keys []string) {
+	if strings.TrimSpace(spec) == "" {
+		return nil, nil
+	}
+	for _, part := range strings.Split(spec, ",") {
+		part = strings.TrimSpace(part)
+		switch {
+		case part == "", part == "public":
+			// "public" is an enforcement concern, not a listing; it is reported
+			// via the Public field.
+		case strings.HasPrefix(part, "@"):
+			keys = append(keys, strings.TrimPrefix(part, "@"))
+		default:
+			roles = append(roles, part)
+		}
+	}
+	return roles, keys
+}

internal/lua/plugin_http_handler_test.go

@@ -0,0 +1,430 @@
+package lua
+
+// Plugin-registered HTTP endpoints.
+//
+// Three properties have to hold, and none of them is visible from the plugin's
+// own point of view:
+//
+//  1. Auth and roles are enforced by the GATEWAY. A plugin that forgets to
+//     declare roles gets admin-only, not open — the failure mode of "public by
+//     omission" is an internet-scannable endpoint inside the gateway process.
+//  2. A handler error is a value, never a panic. The pcall guard is the same one
+//     the hooks use, for the same reason.
+//  3. The plugin cannot steer the gateway's own response framing: a nonsense
+//     status becomes a 500, and Content-Length / hop-by-hop headers are dropped.
+//
+// These tests drive the declaration layer and the response writer directly; the
+// router-level dispatch is covered in the gateway package.
+
+import (
+	"net/http"
+	"net/http/httptest"
+	"testing"
+)
+
+const httpPlugin = `
+local P = {}
+P.name = "api"
+P.version = "1.0.0"
+P.exports = {}
+P.http = {
+  {
+    path = "health",
+    handler = "health",
+    roles = { "admin" },
+    description = "liveness",
+  },
+  {
+    path = "echo/:what",
+    method = "POST",
+    handler = "echo",
+    roles = { "admin", "user" },
+  },
+}
+function P.health(req)
+  return { status = 200, json = { ok = true, role = req.role, key = req.key } }
+end
+function P.echo(req)
+  return { status = 201, json = { what = req.params.what, body = req.body } }
+end
+return P
+`
+
+func TestPluginRegistersHTTPHandlers(t *testing.T) {
+	dir := t.TempDir()
+	ps := NewPlugins(NewVM(dir), dir)
+	if err := ps.LoadSource("api", httpPlugin); err != nil {
+		t.Fatal(err)
+	}
+	hs := ps.HTTPHandlers()
+	if len(hs) != 2 {
+		t.Fatalf("expected 2 registered endpoints, got %d: %+v", len(hs), hs)
+	}
+	for _, h := range hs {
+		if h.Plugin != "api" {
+			t.Errorf("handler %q has no owner; the dispatcher could not route it", h.Path)
+		}
+		if h.Method == "" {
+			t.Errorf("handler %q has no method; the gateway would answer 405 forever", h.Path)
+		}
+	}
+}
+
+func TestHandlerRoleRules(t *testing.T) {
+	// A handler with no roles and no public opt-out is ADMIN-ONLY. This is the
+	// decision the whole feature turns on: "I forgot to declare roles" must not
+	// mean "anyone on the internet".
+	open := HTTPHandler{Path: "x", Handler: "h"}
+	if open.allows("user", "") {
+		t.Error("an undeclared handler must default to admin-only")
+	}
+	if !open.allows("admin", "") {
+		t.Error("an undeclared handler must still serve admin")
+	}
+	// The explicit opt-out is how a plugin says "this one really is public".
+	pub := HTTPHandler{Path: "x", Handler: "h", Public: true}
+	if !pub.allows("user", "") || !pub.allows("", "") {
+		t.Error("public=true must admit any role")
+	}
+	// Declared roles.
+	scoped := HTTPHandler{Path: "x", Roles: []string{"admin", "user"}}
+	if !scoped.allows("user", "homeagent") || !scoped.allows("admin", "admin") {
+		t.Error("a listed role must be admitted")
+	}
+	if (HTTPHandler{Path: "x", Roles: []string{"admin"}}).allows("user", "homeagent") {
+		t.Error("an unlisted role must be refused")
+	}
+}
+
+// Naming keys is how the operator grants one collaborator access without opening
+// the endpoint to every user key. Two properties matter and neither is obvious:
+// a named key wins even when the role rule would have refused it, and the named
+// list REPLACES the role rule rather than adding to it.
+func TestHandlerKeyRules(t *testing.T) {
+	h := HTTPHandler{Path: "x", Roles: []string{"admin"}, Keys: []string{"homeagent"}}
+	if !h.allows("user", "homeagent") {
+		t.Error("a named key must be admitted even though its role is not listed")
+	}
+	// The replacement semantics: listing one key must NOT also admit every user
+	// key through the role rule, which is what an OR would have done.
+	if h.allows("user", "club") {
+		t.Error("naming keys must replace the role rule, not add to it")
+	}
+	if h.allows("user", "") {
+		t.Error("an unnamed key must not match a named-key allowlist")
+	}
+	// Admin is always admitted, so a typo in the list never locks the operator
+	// out of the endpoint they are trying to tighten.
+	if !h.allows("admin", "admin") {
+		t.Error("admin must always reach an endpoint")
+	}
+	// Name matching is case-insensitive; a config written by hand varies.
+	if !(HTTPHandler{Path: "x", Keys: []string{"HomeAgent"}}).allows("user", "homeagent") {
+		t.Error("key names should match case-insensitively")
+	}
+}
+
+// The stored override replaces whatever the plugin declared — this is the admin
+// editing access without touching a .lua that an upgrade would overwrite.
+func TestEndpointAccessOverrideReplacesDeclaration(t *testing.T) {
+	dir := t.TempDir()
+	ps := NewPlugins(NewVM(dir), dir)
+	if err := ps.LoadSource("api", httpPlugin); err != nil {
+		t.Fatal(err)
+	}
+	// health is declared admin-only; widen it to one named key.
+	if err := ps.SetEndpointAccess("api", "health", "@homeagent"); err != nil {
+		t.Fatal(err)
+	}
+	var found *HTTPHandler
+	for _, h := range ps.HTTPHandlers() {
+		if h.Path == "health" {
+			hh := h
+			found = &hh
+		}
+	}
+	if found == nil {
+		t.Fatal("health endpoint disappeared after an access override")
+	}
+	if len(found.Roles) != 0 {
+		t.Errorf("the declared roles should be replaced, got %v", found.Roles)
+	}
+	if len(found.Keys) != 1 || found.Keys[0] != "homeagent" {
+		t.Errorf("override keys = %v, want [homeagent]", found.Keys)
+	}
+	if !found.allows("user", "homeagent") || found.allows("user", "club") {
+		t.Error("the override is not in force")
+	}
+	// Clearing restores the plugin's own declaration.
+	if err := ps.SetEndpointAccess("api", "health", ""); err != nil {
+		t.Fatal(err)
+	}
+	for _, h := range ps.HTTPHandlers() {
+		if h.Path == "health" {
+			if len(h.Keys) != 0 {
+				t.Errorf("clearing the override should restore the declaration, got %v", h.Keys)
+			}
+			return
+		}
+	}
+	t.Error("health endpoint vanished")
+}
+
+func TestSetEndpointAccessRejectsUnknownTarget(t *testing.T) {
+	dir := t.TempDir()
+	ps := NewPlugins(NewVM(dir), dir)
+	if err := ps.LoadSource("api", httpPlugin); err != nil {
+		t.Fatal(err)
+	}
+	if err := ps.SetEndpointAccess("api", "nope", "@homeagent"); err == nil {
+		t.Error("setting access on a path the plugin does not expose must fail loudly")
+	}
+	if err := ps.SetEndpointAccess("nosuchplugin", "health", "@homeagent"); err == nil {
+		t.Error("setting access for an unknown plugin must fail loudly")
+	}
+}
+
+// A stored allowlist naming a key that does not exist refuses every caller while
+// looking perfectly configured. ValidateAccessSpec exists to catch that at write
+// time instead.
+func TestValidateAccessSpec(t *testing.T) {
+	SetKeyNameLookup(func() []string { return []string{"homeagent", "club"} })
+	t.Cleanup(func() { SetKeyNameLookup(nil) })
+
+	for _, ok := range []string{"admin", "user", "public", "@homeagent",
+		"admin,@homeagent,@club", " admin , @club "} {
+		if err := (&Plugins{}).ValidateAccessSpec(ok); err != nil {
+			t.Errorf("spec %q should be accepted: %v", ok, err)
+		}
+	}
+	for _, bad := range []string{"@nosuchkey", "superuser", "@", "admin,@typo_key"} {
+		if err := (&Plugins{}).ValidateAccessSpec(bad); err == nil {
+			t.Errorf("spec %q must be refused: it is neither a role nor a configured key", bad)
+		}
+	}
+	if err := (&Plugins{}).ValidateAccessSpec(""); err != nil {
+		t.Errorf("an empty spec clears the override and must be accepted: %v", err)
+	}
+}
+
+func TestHandlerPathValidationRejectsEscapes(t *testing.T) {
+	// ":name" with empty segments around it is legal — the trims below reduce it
+	// to ":name". What must be refused is an empty PARAM NAME, which cannot be
+	// captured into request.params.
+	for _, bad := range []string{"", "  ", "a/../b", "a//b", "wild*card", "q?x", "/:/"} {
+		if err := validateHandlerPath(bad); err == nil {
+			t.Errorf("path %q must be rejected: it can escape the mount or make the "+
+				"route table ambiguous", bad)
+		}
+	}
+	for _, ok := range []string{"health", "a/b", "echo/:what", "v1/things/:id", "/:name/"} {
+		if err := validateHandlerPath(ok); err != nil {
+			t.Errorf("path %q should be accepted: %v", ok, err)
+		}
+	}
+}
+
+func TestHandlerMethodValidation(t *testing.T) {
+	if m, err := normalizeHandlerMethod(""); err != nil || m != http.MethodGet {
+		t.Errorf("an empty method should default to GET, got %q %v", m, err)
+	}
+	if _, err := normalizeHandlerMethod("TRACE"); err == nil {
+		t.Error("TRACE must be refused")
+	}
+	if m, err := normalizeHandlerMethod("post"); err != nil || m != http.MethodPost {
+		t.Errorf("method should normalise to upper case, got %q %v", m, err)
+	}
+}
+
+func TestInvokeHandlerReturnsStructuredData(t *testing.T) {
+	dir := t.TempDir()
+	ps := NewPlugins(NewVM(dir), dir)
+	if err := ps.LoadSource("api", httpPlugin); err != nil {
+		t.Fatal(err)
+	}
+	resp, err := ps.InvokeHandler("api", "health", &HandlerRequest{
+		Method: http.MethodGet, Path: "health", Role: "admin", Key: "sk-test",
+	})
+	if err != nil {
+		t.Fatalf("invoke: %v", err)
+	}
+	if resp.Status != 200 {
+		t.Errorf("status = %d, want 200", resp.Status)
+	}
+	j, ok := resp.JSON.(map[string]interface{})
+	if !ok {
+		t.Fatalf("json = %T, want a table (the handler's nested value must survive)", resp.JSON)
+	}
+	if j["ok"] != true || j["role"] != "admin" {
+		t.Errorf("handler data crossed wrong: %v", j)
+	}
+}
+
+func TestInvokeHandlerParamsAndBodyReachLua(t *testing.T) {
+	dir := t.TempDir()
+	ps := NewPlugins(NewVM(dir), dir)
+	if err := ps.LoadSource("api", httpPlugin); err != nil {
+		t.Fatal(err)
+	}
+	resp, err := ps.InvokeHandler("api", "echo", &HandlerRequest{
+		Method: http.MethodPost, Path: "echo/thing",
+		Params: map[string]string{"what": "thing"},
+		Body:   `{"n":1}`, Role: "user",
+	})
+	if err != nil {
+		t.Fatalf("invoke: %v", err)
+	}
+	j, _ := resp.JSON.(map[string]interface{})
+	if resp.Status != 201 {
+		t.Errorf("status = %d, want 201 (the handler chose it)", resp.Status)
+	}
+	if j["what"] != "thing" {
+		t.Errorf("path param did not reach the handler: %v", j)
+	}
+	if j["body"] != `{"n":1}` {
+		t.Errorf("body did not reach the handler: %v", j)
+	}
+}
+
+// A handler that raises must produce an error, never a process abort — the same
+// SIGSEGV class of bug the hook guard exists for.
+func TestInvokeHandlerErrorIsValueNotPanic(t *testing.T) {
+	dir := t.TempDir()
+	ps := NewPlugins(NewVM(dir), dir)
+	const boom = `
+local P = {}
+P.name = "boom"
+P.version = "1.0.0"
+P.exports = {}
+P.http = { { path = "blow", handler = "blow", roles = { "admin" } } }
+function P.blow(req)
+  error("handler exploded")
+end
+return P
+`
+	if err := ps.LoadSource("boom", boom); err != nil {
+		t.Fatal(err)
+	}
+	if _, err := ps.InvokeHandler("boom", "blow", &HandlerRequest{Role: "admin"}); err == nil {
+		t.Fatal("a raising handler must surface as an error")
+	}
+}
+
+// The plugin does not get to decide how the gateway frames its response.
+func TestHandlerResponseFramingIsKernelOwned(t *testing.T) {
+	cases := []struct {
+		name string
+		in   *HandlerResponse
+		want int
+	}{
+		{"status 0 becomes 200", &HandlerResponse{Status: 0, Body: "x"}, 200},
+		{"status 9999 becomes 500", &HandlerResponse{Status: 9999, Body: "x"}, 500},
+		{"status -1 becomes 500", &HandlerResponse{Status: -1, Body: "x"}, 500},
+		{"explicit 404 is kept", &HandlerResponse{Status: 404, Body: "x"}, 404},
+	}
+	for _, c := range cases {
+		w := httptest.NewRecorder()
+		writeHTTPHandlerResponse(w, c.in)
+		if w.Code != c.want {
+			t.Errorf("%s: status = %d, want %d", c.name, w.Code, c.want)
+		}
+	}
+
+	// A handler must not be able to desync the gateway's own framing.
+	w := httptest.NewRecorder()
+	writeHTTPHandlerResponse(w, &HandlerResponse{
+		Status: 200, Body: "hello",
+		Headers: map[string]string{
+			"Content-Length":   "99999",
+			"Connection":       "close",
+			"Transfer-Encoding": "chunked",
+			"X-Custom":         "kept",
+		},
+	})
+	if got := w.Header().Get("Content-Length"); got != "" {
+		t.Errorf("a plugin set Content-Length (%s); the body length is the kernel's", got)
+	}
+	if got := w.Header().Get("X-Custom"); got != "kept" {
+		t.Errorf("a normal header should survive, got %q", got)
+	}
+}
+
+// A malformed declaration must not take the whole plugin down.
+func TestBadHTTPRouteDoesNotBreakPluginLoad(t *testing.T) {
+	dir := t.TempDir()
+	ps := NewPlugins(NewVM(dir), dir)
+	const mixed = `
+local P = {}
+P.name = "mixed"
+P.version = "1.0.0"
+P.exports = {}
+P.state = { seen = true }
+P.ui = { page = { page_id = "m", title = "M", mount = "<div/>" } }
+P.http = {
+  { path = "../escape", handler = "h", roles = { "admin" } },
+  { path = "fine",      handler = "h", roles = { "admin" } },
+}
+function P.h(req) return { status = 200, body = "ok" } end
+return P
+`
+	if err := ps.LoadSource("mixed", mixed); err != nil {
+		t.Fatalf("a bad route must not fail the plugin load: %v", err)
+	}
+	// The good route survives; the bad one is simply absent.
+	hs := ps.HTTPHandlers()
+	if len(hs) != 1 || hs[0].Path != "fine" {
+		t.Fatalf("expected only the valid route, got %+v", hs)
+	}
+	// And the plugin's own page/state still work.
+	if ps.State("mixed") == nil {
+		t.Error("the plugin's state was lost because of a bad route")
+	}
+}
+
+func TestParseHandlerResponseShapes(t *testing.T) {
+	if _, err := parseHandlerResponse("just a string"); err == nil {
+		t.Error("a non-table return must be an error, not an empty 200")
+	}
+	if _, err := parseHandlerResponse(map[string]interface{}{}); err == nil {
+		t.Error("an empty table should say so rather than returning a blank 200")
+	}
+	r, err := parseHandlerResponse(map[string]interface{}{
+		"status": float64(204), "headers": map[string]interface{}{"X-A": "b"},
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if r.Status != 204 || r.Headers["X-A"] != "b" {
+		t.Errorf("decoded %+v", r)
+	}
+}
+// TestEndpointAccessReportsPerEndpointOverride guards the reporting side of the
+// override store.
+//
+// SetEndpointAccess already keyed overrides by plugin#path, but endpointAccess
+// read the plugin-level key. After an operator saved a restriction on ONE
+// endpoint, /api/plugins kept reporting the plugin's DECLARED roles, so the
+// plugin page drew the pre-change value and a later save from that screen
+// reverted the restriction. Authorization kept working — only the display and
+// the round trip through the UI were broken, which is why it went unnoticed.
+func TestEndpointAccessReportsPerEndpointOverride(t *testing.T) {
+	dir := t.TempDir()
+	ps := NewPlugins(NewVM(dir), dir)
+	if err := ps.LoadSource("api", httpPlugin); err != nil {
+		t.Fatal(err)
+	}
+	// httpPlugin registers "health" and "echo/:what".
+	if err := ps.SetEndpointAccess("api", "health", "user"); err != nil {
+		t.Fatalf("SetEndpointAccess: %v", err)
+	}
+	got := map[string][]string{}
+	for _, e := range ps.EndpointAccess() {
+		got[e.Path] = e.OverrideRoles
+	}
+	if len(got["health"]) != 1 || got["health"][0] != "user" {
+		t.Errorf("health override not reported to the UI: %+v", got["health"])
+	}
+	if len(got["echo/:what"]) != 0 {
+		t.Errorf("echo inherited an override it never had: %+v", got["echo/:what"])
+	}
+}

internal/lua/plugin_http_test.go

@@ -0,0 +1,217 @@
+package lua
+
+// The plugin HTTP bridge is the one capability that lets a .lua file — installed
+// from a textarea in the WebUI — reach out on the gateway's own network. That
+// makes its DENY rules the security boundary, and these tests exist to pin the
+// refusals rather than the happy path.
+//
+// The failure this guards against is concrete: an operator installs a quota
+// plugin that reads a vendor's balance, and a later plugin (or a compromised
+// one) posts that gateway key material to an attacker, or reads the cloud
+// metadata endpoint, because __llmsproxy_http happened to be a general HTTP
+// client.
+
+import (
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"testing"
+	"time"
+
+	golua "github.com/aarzilli/golua/lua"
+)
+
+// newTestState returns a bare Lua state for the argument-decoding tests.
+func newTestState(t *testing.T) *golua.State {
+	t.Helper()
+	L := golua.NewState()
+	t.Cleanup(L.Close)
+	L.OpenLibs()
+	return L
+}
+
+func TestPluginHTTPAllowlistRejectsUnlistedHost(t *testing.T) {
+	allow := newHTTPAllowlist([]string{"api.example.com"})
+	for _, raw := range []string{
+		"https://evil.test/steal",
+		"https://169.254.169.254/latest/meta-data/",
+		"https://api.example.com.evil.test/x", // suffix trick
+		"https://user@evil.test/",
+	} {
+		if _, err := allow.checkURL(raw); err == nil {
+			t.Fatalf("allowlist permitted %q — a plugin must not reach hosts "+
+				"the operator never configured as a source", raw)
+		}
+	}
+}
+
+func TestPluginHTTPAllowlistAcceptsConfiguredHost(t *testing.T) {
+	allow := newHTTPAllowlist([]string{"api.example.com"})
+	for _, raw := range []string{
+		"https://api.example.com/alpha/credits",
+		"https://API.EXAMPLE.COM/x", // host compare must be case-insensitive
+		"https://api.example.com:443/x",
+	} {
+		if _, err := allow.checkURL(raw); err != nil {
+			t.Fatalf("allowlist refused configured host %q: %v", raw, err)
+		}
+	}
+}
+
+func TestPluginHTTPRefusesPlainHTTPOffLoopback(t *testing.T) {
+	allow := newHTTPAllowlist([]string{"api.example.com"})
+	// Allowed host, but plain HTTP would put the API key on the wire in clear.
+	if _, err := allow.checkURL("http://api.example.com/alpha"); err == nil {
+		t.Fatal("plain http to an allowed non-loopback host must be refused")
+	}
+	// Loopback stays allowed: it cannot leave the machine, and the billing rules
+	// UI reads the gateway's own API over plain HTTP in dev setups.
+	for _, raw := range []string{"http://127.0.0.1:8081/api/x", "http://localhost:8081/x"} {
+		if _, err := allow.checkURL(raw); err != nil {
+			t.Fatalf("loopback http must stay usable, got %v for %q", err, raw)
+		}
+	}
+}
+
+func TestPluginHTTPRejectsBadSchemesAndMissingURL(t *testing.T) {
+	allow := newHTTPAllowlist([]string{"api.example.com"})
+	for _, raw := range []string{"file:///etc/passwd", "ftp://api.example.com/x", "gopher://x"} {
+		if _, err := allow.checkURL(raw); err == nil {
+			t.Fatalf("scheme in %q must be refused", raw)
+		}
+	}
+	// An empty allowlist must deny, not fall open.
+	if _, err := newHTTPAllowlist(nil).checkURL("https://api.example.com/x"); err == nil {
+		t.Fatal("an unconfigured gateway must deny outbound HTTP entirely")
+	}
+}
+
+// A redirect is a hop the allowlist never saw. Following it would let an
+// allowed host bounce a plugin's request (and its headers) anywhere.
+func TestPluginHTTPDoesNotFollowRedirects(t *testing.T) {
+	if httpClient.CheckRedirect == nil {
+		t.Fatal("CheckRedirect must be set: following a redirect escapes the allowlist")
+	}
+	req, _ := http.NewRequest(http.MethodGet, "https://api.example.com/x", nil)
+	err := httpClient.CheckRedirect(req, []*http.Request{req})
+	if err != http.ErrUseLastResponse {
+		t.Fatalf("CheckRedirect must return ErrUseLastResponse so the 3xx is "+
+			"handed back unfollowed, got %v", err)
+	}
+}
+
+func TestPluginHTTPTimeoutAndBodyAreClamped(t *testing.T) {
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		_, _ = w.Write([]byte(`{"ok":true}`))
+	}))
+	defer srv.Close()
+	allow := newHTTPAllowlist([]string{"127.0.0.1"})
+	u, err := allow.checkURL(srv.URL)
+	if err != nil {
+		t.Fatalf("loopback test server must be allowed: %v", err)
+	}
+	// An absurd timeout is clamped instead of honoured.
+	res, err := performHTTP(u, httpRequestSpec{Timeout: 24 * time.Hour})
+	if err != nil {
+		t.Fatalf("clamped request failed: %v", err)
+	}
+	if res.Status != 200 || !strings.Contains(res.Body, `"ok"`) {
+		t.Fatalf("unexpected result: status=%d body=%q", res.Status, res.Body)
+	}
+}
+
+func TestPluginHTTPOversizedBodyIsRefused(t *testing.T) {
+	big := strings.Repeat("x", httpMaxResponse+64)
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		_, _ = w.Write([]byte(big))
+	}))
+	defer srv.Close()
+	allow := newHTTPAllowlist([]string{"127.0.0.1"})
+	u, _ := allow.checkURL(srv.URL)
+	// Refusing beats truncating: a truncated body parses as corrupt JSON and the
+	// plugin would report a parse error instead of the real cause.
+	if _, err := performHTTP(u, httpRequestSpec{}); err == nil {
+		t.Fatal("an oversized response must be refused, not silently truncated")
+	}
+}
+
+func TestPluginHTTPErrorPathIsPairNotRaise(t *testing.T) {
+	// Lua-side contract: the bridge returns (nil, message) on every failure so a
+	// plugin mistake never raises through golua (see the SIGSEGV note on
+	// hookGuardSrc — an escaping error can take the gateway down).
+	allow := newHTTPAllowlist([]string{"api.example.com"})
+	if _, err := allow.checkURL("https://evil.test/"); err == nil {
+		t.Fatal("expected refusal")
+	}
+	// readHTTPRequestSpec must reject a non-table without touching the stack.
+	L := newTestState(t)
+	if _, err := readHTTPRequestSpec(L); err == nil {
+		t.Fatal("a missing argument table must be an error, not a panic")
+	}
+}
+// The bridge must be reachable under the name plugins actually call.
+//
+// This exists because the registration once used the CONSTANT'S IDENTIFIER as
+// the Lua global name (`registerFn(L, "httpGlobal", …)`) rather than its VALUE,
+// so the bridge was published as `httpGlobal` while every plugin called
+// `__llmsproxy_http`. The failure only appeared at runtime, as
+// "attempt to call global '__llmsproxy_http' (a nil value)" on the first real
+// request — after the unit tests were green, because they drove the Go-side
+// helpers directly and never went through Lua name resolution.
+//
+// So this test boots a real plugin and calls the bridge the way a plugin does.
+func TestPluginHTTPBridgeIsReachableUnderItsLuaName(t *testing.T) {
+	dir := t.TempDir()
+	ps := NewPlugins(NewVM(dir), dir)
+	ps.SetOutboundHosts([]string{"127.0.0.1"})
+
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.Header().Set("Content-Type", "application/json")
+		_, _ = w.Write([]byte(`{"credits":{"monthlyCredits":1.5}}`))
+	}))
+	defer srv.Close()
+
+	// The plugin asks for a loopback URL, which is allowed without being
+	// configured as a source. It type-checks both the table and arg.url, so a
+	// malformed call is reported by the plugin rather than as a bewildering URL
+	// parse error from the bridge.
+	src := `
+local P = {}
+P.name = "httpprobe"
+P.version = "1.0.0"
+P.exports = { "probe" }
+-- Call passes the WHOLE argument table as a single Lua argument, which is the
+-- same convention the hooks use and what a plugin page's pluginAPI.call ends up
+-- producing. Taking the table (not a bare string) is the correct shape.
+function P.probe(arg)
+  if type(arg) ~= "table" then return { error = "probe takes the args table, got " .. type(arg) } end
+  if type(arg.url) ~= "string" then return { error = "arg.url must be a string, got " .. type(arg.url) } end
+  local res, err = ` + httpGlobal + `({ url = arg.url, method = "GET" })
+  if res == nil then return { error = tostring(err) } end
+  return { status = res.status, body = res.body }
+end
+return P
+`
+	if err := ps.LoadSource("httpprobe", src); err != nil {
+		t.Fatalf("load: %v", err)
+	}
+	got, err := ps.Call("httpprobe", "probe", map[string]interface{}{"url": srv.URL})
+	if err != nil {
+		t.Fatalf("call: %v", err)
+	}
+	m, _ := got.(map[string]interface{})
+	if m == nil {
+		t.Fatalf("probe returned %T, want a table", got)
+	}
+	// A nil global surfaces as {error = "attempt to call global ... (a nil value)"} —
+	// which is a table too, so the error field is what distinguishes it.
+	if e, ok := m["error"]; ok {
+		t.Fatalf("the bridge is not reachable as %s: %v", httpGlobal, e)
+	}
+	if s, _ := m["status"].(float64); int(s) != 200 {
+		t.Fatalf("status=%v body=%v, want 200 from the loopback test server", m["status"], m["body"])
+	}
+	if body, _ := m["body"].(string); !strings.Contains(body, "monthlyCredits") {
+		t.Fatalf("body=%q, want the server's JSON", body)
+	}
+}

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/plugin_visibility.go

@@ -0,0 +1,118 @@
+package lua
+
+// Per-caller state scoping.
+//
+// Role visibility answers "may this role see the plugin at all". It does not
+// answer "which PART of its data", and for the billing plugin those are
+// different questions: the global figures (total spend, every source, every
+// model) are deployment data, while the per-key buckets describe one caller.
+//
+// So a plugin may declare a split, and the gateway then hands a non-admin
+// caller only the buckets that are theirs. The plugin declares it because only
+// the plugin author knows which of its buckets are global.
+
+import "strings"
+
+// StateScope is a plugin's declaration of which state buckets are global.
+type StateScope struct {
+	// Global names the buckets that describe the whole deployment. Anything not
+	// listed is treated as per-caller and filtered.
+	Global []string `json:"global,omitempty"`
+	// Depth gives the nesting level of each PER-CALLER bucket: 1 for
+	// {maskedKey: bucket}, 2 for {day: {maskedKey: bucket}}. It is declared
+	// because both shapes are maps and sniffing would confuse an empty bucket
+	// with a nested one.
+	Depth map[string]int `json:"depth,omitempty"`
+}
+
+// IsGlobal reports whether a top-level state key is deployment-wide.
+func (s *StateScope) IsGlobal(key string) bool {
+	if s == nil {
+		return true
+	}
+	for _, g := range s.Global {
+		if g == key {
+			return true
+		}
+	}
+	return false
+}
+
+// StateScopeFor returns a plugin's declared split, or nil when it declares none.
+func (ps *Plugins) StateScopeFor(name string) *StateScope {
+	ps.mu.RLock()
+	defer ps.mu.RUnlock()
+	for _, p := range ps.plugins {
+		if p.Info.Name != name || p.UI == nil {
+			continue
+		}
+		return p.UI.StateScope
+	}
+	return nil
+}
+
+// compileDepthMap normalizes the declared depth table, defaulting to 1 for a
+// per-caller bucket the author did not annotate.
+func compileDepthMap(in map[string]int) map[string]int {
+	if len(in) == 0 {
+		return nil
+	}
+	out := make(map[string]int, len(in))
+	for k, v := range in {
+		if v < 1 {
+			v = 1
+		}
+		out[strings.TrimSpace(k)] = v
+	}
+	return out
+}
+
+// DepthOf returns the declared nesting level for a per-caller bucket, 1 by
+// default. A bucket the author did not annotate is assumed flat, which is the
+// common shape ({maskedKey: row}).
+func (s *StateScope) DepthOf(key string) int {
+	if s == nil || s.Depth == nil {
+		return 1
+	}
+	if d, ok := s.Depth[key]; ok {
+		return d
+	}
+	return 1
+}
+
+// VisibilityOverrides returns every operator override in force, keyed by plugin
+// name. An absent name means the plugin's own declaration applies.
+//
+// The WebUI needs this to show which visibility is IMPOSED, not merely which is
+// in effect: an operator comparing two deployments should not have to open a
+// .lua to learn whether a page is admin-only by the plugin's choice or by their
+// own override.
+func (ps *Plugins) VisibilityOverrides() map[string]string {
+	ps.mu.RLock()
+	defer ps.mu.RUnlock()
+	if len(ps.visibility) == 0 {
+		return nil
+	}
+	epKeys := make(map[string]struct{})
+	for _, e := range ps.endpointAccessLocked() {
+		epKeys[endpointOverrideKey(e.Plugin, e.Path)] = struct{}{}
+	}
+	out := make(map[string]string, len(ps.visibility))
+	for k, v := range ps.visibility {
+		// Per-endpoint overrides live in the SAME map under a "name#path" key.
+		// They are endpoint policy, not a plugin audience, so leaking them here
+		// both mislabels them and lets the plugin page write an audience of the
+		// form "probe#hello" that no code ever reads as one.
+		// Per-endpoint overrides share this map under a "name#path" key. They
+		// are endpoint policy, not a plugin audience: reporting them here both
+		// mislabels them and invites the plugin page to save an audience of the
+		// form "probe#hello" that nothing reads as an audience. The set of
+		// endpoint keys is already computed, so ask it rather than re-deriving
+		// the shape from the key.
+		if _, isEndpoint := epKeys[k]; isEndpoint {
+			continue
+		}
+		out[k] = v
+	}
+	return out
+}

internal/lua/plugin_visibility_test.go

@@ -0,0 +1,193 @@
+package lua
+
+// Plugin visibility: a plugin page is not automatically public data.
+//
+// The billing page renders GLOBAL cost across every source and the quota page
+// renders upstream balances. Both were served to any authenticated key because
+// /api/ui-inject applied no role filter — measured, not theorised: a user-role
+// key pulled the 18KB billing state and read the source names and the $8.86
+// total straight out of it.
+//
+// These tests pin the DENIALS. A visibility feature whose tests only assert that
+// an admin still sees everything passes while the leak is wide open.
+
+import (
+	"testing"
+)
+
+func TestAudienceAllows(t *testing.T) {
+	cases := []struct {
+		want, role string
+		ok         bool
+		why        string
+	}{
+		{"", "user", true, "unset means everyone"},
+		{"all", "user", true, "explicit all"},
+		{"admin", "admin", true, "exact match"},
+		{"admin", "user", false, "admin-only excludes user"},
+		{"user", "user", true, "user-only admits user"},
+		{"user", "admin", false, "user-only excludes admin"},
+		{"admin,user", "user", true, "list admits a listed role"},
+		{" admin , user ", "user", true, "whitespace is not significant"},
+		{"ADMIN", "user", false, "case is normalised"},
+		{"*", "user", false, `"*" must NOT be a synonym for all`},
+	}
+	for _, c := range cases {
+		if got := audienceAllows(c.want, c.role); got != c.ok {
+			t.Errorf("audienceAllows(%q, %q) = %v, want %v (%s)",
+				c.want, c.role, got, c.ok, c.why)
+		}
+	}
+}
+
+func TestPluginVisibilityHidesPageFromUserRole(t *testing.T) {
+	dir := t.TempDir()
+	ps := NewPlugins(NewVM(dir), dir)
+
+	// One admin-only plugin, one open plugin: the filter must drop exactly one
+	// and leave the other intact, or a plugin author has no way to ship a mixed
+	// set.
+	const adminOnly = `
+local P = {}
+P.name = "adminonly"
+P.version = "1.0.0"
+P.exports = {}
+P.ui = { visibility = "admin", page = { page_id = "secret", title = "Secret", mount = "<div>secret</div>" } }
+return P
+`
+	const open = `
+local P = {}
+P.name = "openpage"
+P.version = "1.0.0"
+P.exports = {}
+P.ui = { page = { page_id = "open", title = "Open", mount = "<div>open</div>" } }
+return P
+`
+	for name, src := range map[string]string{"adminonly": adminOnly, "openpage": open} {
+		if err := ps.LoadSource(name, src); err != nil {
+			t.Fatalf("load %s: %v", name, err)
+		}
+	}
+
+	admin := ps.UIForRole("admin")
+	if len(admin.Pages) != 2 {
+		t.Errorf("admin should see both pages, got %d", len(admin.Pages))
+	}
+
+	user := ps.UIForRole("user")
+	for _, pg := range user.Pages {
+		if pg.PageID == "secret" {
+			t.Fatal("a user key can see the admin-only page — this is the leak")
+		}
+	}
+	if len(user.Pages) != 1 || user.Pages[0].PageID != "open" {
+		t.Fatalf("user should see exactly the open page, got %v", pageIDs(user.Pages))
+	}
+}
+
+// The operator must be able to impose visibility the plugin did not choose —
+// that is the whole point of the override, since a bundled .lua cannot be edited
+// without being overwritten on the next upgrade.
+func TestPluginVisibilityOverrideBeatsDeclaration(t *testing.T) {
+	dir := t.TempDir()
+	ps := NewPlugins(NewVM(dir), dir)
+	const open = `
+local P = {}
+P.name = "openpage"
+P.version = "1.0.0"
+P.exports = {}
+P.ui = { page = { page_id = "open", title = "Open", mount = "<div>open</div>" } }
+return P
+`
+	if err := ps.LoadSource("openpage", open); err != nil {
+		t.Fatal(err)
+	}
+	if len(ps.UIForRole("user").Pages) != 1 {
+		t.Fatal("precondition: the page should start visible to user")
+	}
+
+	ps.SetPluginVisibility("openpage", "admin")
+	if got := ps.UIForRole("user").Pages; len(got) != 0 {
+		t.Fatalf("the operator override was ignored: user still sees %v", pageIDs(got))
+	}
+	if ps.VisibilityOverride("openpage") != "admin" {
+		t.Error("the override should be reportable so /api/plugins can show it")
+	}
+
+	// Clearing restores the plugin's own declaration.
+	ps.SetPluginVisibility("openpage", "")
+	if got := ps.UIForRole("user").Pages; len(got) != 1 {
+		t.Fatalf("clearing the override should restore visibility, got %v", pageIDs(got))
+	}
+}
+
+// A page may narrow visibility below the plugin default, so one plugin can ship
+// a public summary beside an admin-only detail page.
+func TestPluginPageCanNarrowBelowPluginDefault(t *testing.T) {
+	dir := t.TempDir()
+	ps := NewPlugins(NewVM(dir), dir)
+	const mixed = `
+local P = {}
+P.name = "mixed"
+P.version = "1.0.0"
+P.exports = {}
+P.ui = {
+  pages = {
+    { page_id = "summary", title = "Summary", mount = "<div>s</div>" },
+    { page_id = "details", title = "Details", mount = "<div>d</div>", audience = "admin" },
+  },
+}
+return P
+`
+	if err := ps.LoadSource("mixed", mixed); err != nil {
+		t.Fatal(err)
+	}
+	user := ps.UIForRole("user")
+	if len(user.Pages) != 1 || user.Pages[0].PageID != "summary" {
+		t.Fatalf("user should see only the summary, got %v", pageIDs(user.Pages))
+	}
+	if len(ps.UIForRole("admin").Pages) != 2 {
+		t.Error("admin should see both pages")
+	}
+}
+
+func TestPluginStateVisibleTo(t *testing.T) {
+	dir := t.TempDir()
+	ps := NewPlugins(NewVM(dir), dir)
+	const adminOnly = `
+local P = {}
+P.name = "adminonly"
+P.version = "1.0.0"
+P.exports = {}
+P.state = { secret = "x" }
+P.ui = { visibility = "admin", page = { page_id = "s", title = "S", mount = "<div/>" } }
+return P
+`
+	if err := ps.LoadSource("adminonly", adminOnly); err != nil {
+		t.Fatal(err)
+	}
+	// The data has to follow the page: a hidden UI with an open state endpoint is
+	// the same leak with one more step.
+	if ps.StateVisibleTo("adminonly", "user") {
+		t.Error("user must not read an admin-only plugin's state")
+	}
+	if !ps.StateVisibleTo("adminonly", "admin") {
+		t.Error("admin must still read it")
+	}
+	// Unknown plugin: unchanged behaviour, so a widget that is not installed does
+	// not start answering 403.
+	if !ps.StateVisibleTo("nosuchplugin", "user") {
+		t.Error("an unknown plugin name must not become forbidden; that would change " +
+			"behaviour for plugins that simply are not installed")
+	}
+}
+
+func pageIDs(pages []*UIPage) []string {
+	out := make([]string, 0, len(pages))
+	for _, p := range pages {
+		if p != nil {
+			out = append(out, p.PageID)
+		}
+	}
+	return out
+}

internal/lua/plugins.go

@@ -34,6 +34,7 @@ package lua
 import (
 	"encoding/json"
 	"fmt"
+	"log"
 	"os"
 	"path/filepath"
 	"sort"
@@ -126,6 +127,13 @@ type UIExtension struct {
 	// Elements are snippets injected into EXISTING pages, keyed by target page
 	// id (e.g. "status", "keys"). Order within a target is plugin load order.
 	Elements []UIElement `json:"elements,omitempty"`
+	// Visibility is the plugin-wide default audience, applied to every page and
+	// element this plugin contributes. A per-page Audience narrows it further.
+	Visibility string `json:"visibility,omitempty"`
+	// StateScope splits the plugin's published state into global and per-caller
+	// buckets, for a plugin that is readable by more than one role but must not
+	// hand one caller another's data.
+	StateScope *StateScope `json:"state_scope,omitempty"`
 }
 
 // UIPage is a plugin-provided page.
@@ -134,6 +142,20 @@ type UIPage struct {
 	Title  string `json:"title"`   // sidebar label
 	Icon   string `json:"icon"`    // optional inline SVG or short glyph
 	Order  int    `json:"order"`   // sidebar sort key (default 100)
+	// Audience limits which API-key roles may see this page. "" and "all" mean
+	// every authenticated role; "admin" hides it from user keys; a comma list
+	// ("admin,user") names the roles allowed.
+	//
+	// This exists because a plugin page is not automatically public data. The
+	// billing page shows GLOBAL cost across every source, and the quota page
+	// shows upstream balances — neither belongs to whoever happens to hold a
+	// user key, yet both were served to them because /api/ui-inject applied no
+	// role filter at all.
+	//
+	// The plugin declares its own default and the gateway may override it (see
+	// SetPluginVisibility), because the plugin author cannot know the operator's
+	// policy but the operator must be able to impose one.
+	Audience string `json:"audience,omitempty"`
 	// Mount is the page body. It may contain <script> and <style>; the kernel
 	// executes scripts AFTER injecting the HTML so the DOM exists, and exposes
 	// `pluginAPI` to them (see docs/plugins.md).
@@ -145,6 +167,10 @@ type UIElement struct {
 	// Target is the id of the pane to inject into: "status", "chat", "keys",
 	// "sort", "sources" or "adapters".
 	Target string `json:"target"`
+	// Audience is the same role filter as UIPage.Audience. An injected tile can
+	// leak exactly what a whole page would — the billing status tile shows total
+	// cost — so the field exists here too rather than only on pages.
+	Audience string `json:"audience,omitempty"`
 	// Anchor is where in the target pane: "top", "bottom" or "before:<sel>" /
 	// "after:<sel>" for a CSS selector. Empty = "bottom".
 	Anchor string `json:"anchor,omitempty"`
@@ -179,6 +205,14 @@ type Plugin struct {
 	// reasoning as a systemd unit being masked rather than removed.
 	Disabled bool
 	dir      string
+	// http lists endpoints this plugin mounts under /api/plugins/http/. Read at
+	// load time, never mutated afterwards.
+	http []HTTPHandler
+	// exports lists the functions this plugin allows the WebUI to call through
+	// POST /api/plugins/{name}/call/{fn}. It is read once at load from
+	// plugin.exports and never changes, so Call can check it without taking a
+	// Lua read on the hot path.
+	exports []string
 	// script is the plugin's source, kept so a reload can rebuild its state.
 	script string
 	// Builtin marks a plugin that shipped with the gateway. The UI shows it as
@@ -228,9 +262,20 @@ type Plugins struct {
 	// direct log call so the plugin package stays free of a logging dependency
 	// and tests can capture the warnings.
 	logf func(format string, args ...interface{})
+	// 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
+	// allowHTTP bounds plugin outbound requests to operator-declared hosts.
+	// nil means deny-all: an unconfigured gateway must not silently grant a
+	// plugin unrestricted network access.
+	allowHTTP *httpAllowlist
 	// save coalesces state writes into one background goroutine.
 	//
 	// A hook must NOT write synchronously: the billing plugin mutates state on
@@ -412,11 +457,16 @@ func NewPlugins(vm *VM, dir string) *Plugins {
 		stageFuncs: map[Stage][]hookCall{},
 		hookErr:    newHookErrors(),
 		dir:        dir,
+		// Deny-by-default. NewPlugins is called during wiring, before any
+		// source list is known; SetOutboundHosts replaces this with the real
+		// set. A plugin loaded in that window simply cannot make requests.
+		allowHTTP: newHTTPAllowlist(nil),
 	}
 	// The saver is started even with no dir: a gateway can be given a plugin dir
 	// 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
 }
 
@@ -429,10 +479,34 @@ func NewPlugins(vm *VM, dir string) *Plugins {
 // DisableStatePersistence turns persistence off (tests/benchmarks only).
 func (ps *Plugins) DisableStatePersistence() { ps.stateDirDisabled = true }
 
+// SetOutboundHosts declares which hosts plugins may contact via
+// __llmsproxy_http. The gateway passes the host of every configured source:
+// those are operator-declared upstreams, and reusing that set keeps a plugin's
+// reach no wider than the deployment already trusts. An empty list means no
+// outbound host is allowed at all — plugins then still load, they just cannot
+// make requests.
+func (ps *Plugins) SetOutboundHosts(hosts []string) {
+	ps.mu.Lock()
+	defer ps.mu.Unlock()
+	ps.allowHTTP = newHTTPAllowlist(hosts)
+}
+
+// outboundAllowlist returns the current allowlist. bootPluginState reads it
+// before any concurrent reload can swap it, and a nil result is treated as
+// "deny everything" rather than "allow everything".
+func (ps *Plugins) outboundAllowlist() *httpAllowlist {
+	ps.mu.RLock()
+	defer ps.mu.RUnlock()
+	return ps.allowHTTP
+}
+
 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()
 }
 
@@ -446,6 +520,12 @@ func (ps *Plugins) LoadDir() error {
 	if ps.dir == "" {
 		return nil
 	}
+	// Overrides first: they are keyed by plugin name and must be in place before
+	// anything reads them, and a reload must not blank a restriction the
+	// operator set earlier in this same process.
+	if err := ps.LoadAccessOverrides(); err != nil {
+		log.Printf("[core] plugin access overrides not loaded: %v", err)
+	}
 	if _, err := os.Stat(ps.dir); os.IsNotExist(err) {
 		return nil
 	}
@@ -496,10 +576,17 @@ func (ps *Plugins) SeedBundled() error {
 // bootPluginState compiles one plugin into a fresh Lua state and stores its
 // returned table under pluginGlobal. It is the plugin counterpart of
 // adapterPool.boot, minus the pooling: a plugin keeps exactly one state.
-func bootPluginState(code, name string) (*worker, error) {
+func bootPluginState(code, name string, allow *httpAllowlist) (*worker, error) {
 	L := golua.NewState()
 	L.OpenLibs()
 	setupGlobals(L)
+	// Only plugins get __llmsproxy_http. Adapters are operator-supplied config
+	// and go through setupGlobals too, so installing the bridge here — not in
+	// setupGlobals — is what keeps the capability scoped to installed .lua
+	// files.
+	if allow != nil {
+		registerFn(L, httpGlobal, luaHTTPFn(L, allow))
+	}
 	if err := L.DoString(code); err != nil {
 		L.Close()
 		return nil, fmt.Errorf("compile plugin %s: %w", name, err)
@@ -524,6 +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{},
@@ -537,7 +641,7 @@ func (ps *Plugins) LoadSource(name, code string) error {
 	// One state, held for the plugin's lifetime (see Plugin.state): a hook that
 	// accumulates into plugin.state would otherwise split its totals across
 	// whichever worker it happened to run on.
-	w, err := bootPluginState(code, name)
+	w, err := bootPluginState(code, name, ps.outboundAllowlist())
 	if err != nil {
 		p.LoadError = err.Error()
 		ps.append(p)
@@ -588,10 +692,19 @@ 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)
+
+	// ---- 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
+	// to decide whether a call is allowed.
+	p.exports = readPluginExports(w.L)
+
 	// Restore persisted state AFTER the plugin compiled and registered its
 	// hooks, so the restore overwrites the compiled-in defaults instead of being
 	// overwritten by them. Doing it earlier would mean a restart resets the
@@ -675,26 +788,120 @@ func pluginHookName(L *golua.State, stage string) (string, bool) {
 }
 
 // readPluginUI reads the optional ui extension block.
-func readPluginUI(L *golua.State) (*UIExtension, bool) {
-	L.GetGlobal(pluginGlobal)
-	if L.IsNil(-1) {
+// readPluginExports collects plugin.exports, the list of functions the WebUI is
+// allowed to invoke through POST /api/plugins/{name}/call/{fn}.
+//
+// It accepts either an array of names or a map whose KEYS are the names — both
+// shapes read naturally in Lua and authors reach for either. A missing or
+// 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 {
+	// 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
+	}
+	return decStringList(map[string]interface{}{"v": raw}, "v")
+}
+
+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
 	}
-	L.GetField(-1, "ui")
+	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, field)
 	if L.Type(-1) == golua.LUA_TNIL {
 		L.SetTop(0)
 		return nil, false
 	}
-	var ui UIExtension
-	if err := luaToJSON(L, -1, &ui); err != nil {
-		L.SetTop(0)
-		return nil, false
-	}
+	raw := luaTableToGoValue(L, -1)
 	L.SetTop(0)
-	if ui.Page == nil && len(ui.Pages) == 0 && len(ui.Elements) == 0 {
-		return nil, false
+	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
 	}
-	return &ui, true
+
+	if pg := decTable(m, "page"); pg != nil {
+		ui.Page = uipageFromValue(pg)
+	}
+	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) {
@@ -938,6 +1145,131 @@ func (ps *Plugins) HookErrors() map[string]map[string]interface{} {
 }
 
 // UI returns the merged UI extensions to inject into the WebUI.
+// audienceAllows reports whether a role is permitted by an audience string.
+//
+// The grammar is deliberately small, because it is authored in two places (a
+// .lua file and config.yaml) and a richer language would only give someone a way
+// to express a rule nobody audits:
+//
+//	""  or "all"   everyone
+//	"admin"        admin role only
+//	"user"         user role only
+//	"admin,user"   either
+//
+// "*" 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 {
+	return ParseAccessRule(want).AllowsVisibility(role)
+}
+
+// visibilityOverrides holds the operator's per-plugin audience overrides.
+//
+// A plugin declares its own default and the operator may force one. Both sides
+// need to be able to: the plugin author does not know the deployment's policy,
+// and the operator must not have to edit a bundled .lua — which the next
+// upgrade would overwrite anyway.
+type visibilityOverrides map[string]string
+
+// SetPluginVisibility records an operator override for a plugin's audience. An
+// empty value clears the override and restores the plugin's own default.
+func (ps *Plugins) SetPluginVisibility(plugin, audience string) {
+	ps.mu.Lock()
+	defer ps.mu.Unlock()
+	if ps.visibility == nil {
+		ps.visibility = visibilityOverrides{}
+	}
+	plugin = strings.TrimSpace(plugin)
+	audience = strings.TrimSpace(audience)
+	if audience == "" {
+		delete(ps.visibility, plugin)
+	} else {
+		ps.visibility[plugin] = audience
+	}
+	ps.saveAccessOverridesLocked()
+}
+
+// VisibilityOverride returns the operator's audience override for a plugin,
+// or "" when none is set. Exposed so /api/plugins can show which visibility is
+// imposed rather than only the one in effect.
+func (ps *Plugins) VisibilityOverride(plugin string) string {
+	ps.mu.RLock()
+	defer ps.mu.RUnlock()
+	return ps.visibility[plugin]
+}
+
+// uiAudience resolves the audience in force for a plugin: the operator's
+// override if set, else the plugin's declared default.
+func (ps *Plugins) uiAudience(p *Plugin, declared string) string {
+	ps.mu.RLock()
+	override := ps.visibility[p.Info.Name]
+	ps.mu.RUnlock()
+	if override != "" {
+		return override
+	}
+	return declared
+}
+
+// UIForRole returns the UI a given role may see.
+//
+// Filtering happens HERE, not in the HTTP layer, so every consumer sees the same
+// view — the WebUI boot payload, the Electron shell and the tests. Filtering at
+// one call site means the others quietly keep serving everything.
+//
+// A plugin's page/element can narrow visibility further than its own default
+// (a public summary page beside an admin-only one), so the role check runs on
+// the resolved plugin audience AND on each item's own declaration.
+func (ps *Plugins) UIForRole(role string) *UIExtension {
+	full := ps.UI()
+	if full == nil {
+		return &UIExtension{}
+	}
+	// An admin sees everything, including anything a plugin hides from itself
+	// by mistake: anything else makes an override unverifiable from the browser.
+	if role == "admin" {
+		return full
+	}
+	out := &UIExtension{}
+	// The plugin-level default is per plugin, so pages are filtered by walking
+	// plugins rather than pages: a page id alone does not say who owns it.
+	for _, p := range ps.plugins {
+		if p.UI == nil || p.LoadError != "" {
+			continue
+		}
+		aud := ps.uiAudience(p, p.UI.Visibility)
+		if !audienceAllows(aud, role) {
+			continue
+		}
+		for _, pg := range p.UI.Pages {
+			if pg == nil || !audienceAllows(pg.Audience, role) {
+				continue
+			}
+			out.Pages = append(out.Pages, pg)
+		}
+		for _, el := range p.UI.Elements {
+			if !audienceAllows(el.Audience, role) {
+				continue
+			}
+			out.Elements = append(out.Elements, el)
+		}
+		// Page (singular) is the other half of the same declaration and both
+		// merge in UI(). Handling only Pages made every plugin that used the
+		// single-page form disappear for a user key — silently, with no error,
+		// because "this plugin contributes nothing to you" is indistinguishable
+		// from "this plugin is not installed".
+		if p.UI.Page != nil && audienceAllows(p.UI.Page.Audience, role) {
+			if out.Page == nil {
+				out.Page = p.UI.Page
+			}
+			if !containsPage(out.Pages, p.UI.Page.PageID) {
+				out.Pages = append(out.Pages, p.UI.Page)
+			}
+		}
+	}
+	return out
+}
+
 func (ps *Plugins) UI() *UIExtension {
 	if u := ps.ui.Load(); u != nil {
 		return u
@@ -1614,3 +1946,334 @@ func (ps *Plugins) invoke(p *Plugin, fn string, payload map[string]interface{})
 	ps.markDirtyLocked(p)
 	return out, nil
 }
+
+// Call invokes a plugin function that the plugin declared in plugin.exports.
+//
+// The gate is the declaration, not the caller: a plugin lists the handful of
+// entry points its own UI is allowed to reach, and anything absent is refused.
+// Without that list this endpoint would be a remote Lua evaluator — it could
+// call the plugin's private helpers, hand them crafted arguments, and read
+// whatever the plugin can see, which for a plugin holding an API key means
+// handing that key to anyone who can POST to the gateway.
+//
+// The return value is JSON-shaped rather than restricted to a map, because a
+// 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
+	for _, c := range ps.plugins {
+		if c.Info.Name == name {
+			p = c
+			break
+		}
+	}
+	ps.mu.RUnlock()
+	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)
+	}
+	p.mu.Lock()
+	defer p.mu.Unlock()
+	w := p.state
+	if w == nil {
+		return nil, fmt.Errorf("no state for plugin %s", name)
+	}
+	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)
+
+	L.GetGlobal(pluginGlobal)
+	if L.IsNil(-1) {
+		return nil, fmt.Errorf("plugin table missing")
+	}
+	plug := L.GetTop()
+	L.GetGlobal(hookGuardName)
+	if L.IsNil(-1) {
+		L.Pop(1)
+	}
+	L.GetField(plug, fn)
+	if !L.IsFunction(-1) {
+		return nil, fmt.Errorf("export %s missing", fn)
+	}
+	if args == nil {
+		args = map[string]interface{}{}
+	}
+	pushGoValue(L, plainForLua(args))
+	if err := L.Call(2, 2); err != nil {
+		return nil, err
+	}
+	errMsg := ""
+	if L.GetTop() >= 2 {
+		if !L.IsNil(-1) {
+			errMsg = L.ToString(-1)
+		}
+		L.Pop(1)
+	}
+	if 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, err
+	}
+	return out, nil
+}
+
+// pluginExportAllowed reports whether fn is in the plugin's exports list.
+// An absent list means nothing is callable — the safe reading, since silently
+// allowing all would turn an undeclared plugin into a remote evaluator.
+func pluginExportAllowed(p *Plugin, fn string) bool {
+	for _, e := range p.exports {
+		if e == fn {
+			return true
+		}
+	}
+	return false
+}
+
+// StateVisibleTo reports whether a role may read a plugin's published state.
+//
+// The audience is resolved exactly as for the UI: operator override first, then
+// the plugin's declaration. A plugin with neither is readable by any
+// authenticated role, which is the pre-existing behaviour and the reason a
+// newly installed widget keeps working without an edit.
+
+// HTTPHandler is one endpoint a plugin registers on the gateway's router.
+//
+// A plugin that only receives hooks (billing) has nothing to expose. A plugin
+// that computes something the operator may want to query from OUTSIDE the
+// browser — a health probe, a CI check, another tool — needs a route, and
+// hard-coding one in the gateway would mean every such plugin needs a gateway
+// change. So the plugin declares the path, the method and the role it requires,
+// and the gateway mounts it.
+//
+// The handler is a Lua function taking (request) and returning (response). Both
+// are plain tables: request carries method/path/query/headers/body, response
+// carries status/headers/body/json. Nothing here can reach the Go side except
+// through those two values, so the surface a plugin can express is bounded by
+// what a table can say.
+type HTTPHandler struct {
+	// Path is the route suffix under /api/plugins/http/. It may contain ":name"
+	// segments, captured into request.params.
+	Path string `json:"path"`
+	// Method is the HTTP method; "" or GET. Only GET/POST/PUT/DELETE/PATCH are
+	// accepted — an arbitrary verb would need a switch in the router for no gain.
+	Method string `json:"method,omitempty"`
+	// Roles are the API-key roles allowed to call it. EMPTY MEANS EVERYONE, which
+	// is the wrong default for an endpoint a plugin invented: an unauthenticated
+	// internet scan should not be able to ask it questions. RequireRole exists
+	// so a plugin that really wants public access has to say so.
+	Roles []string `json:"roles,omitempty"`
+	// Keys names SPECIFIC gateway keys allowed to call it, by their configured
+	// name (the `name:` on a gateway_keys entry), never by the secret.
+	//
+	// Names rather than secrets for three reasons: a secret is rotated without
+	// notice and a rule written against it would silently stop matching; putting
+	// secrets in a plugin file scatters credentials across the install; and the
+	// operator reads names. A key with no name cannot be named here.
+	//
+	// When Keys is non-empty it is the WHOLE rule, not an addition to Roles. An
+	// OR would mean an operator writing a two-key allowlist had also granted
+	// every user key, which is the opposite of what they meant.
+	Keys []string `json:"keys,omitempty"`
+	// RequireAuth defaults to true in practice: the gateway refuses a handler
+	// with neither Roles nor an explicit opt-out.
+	Public bool `json:"public,omitempty"`
+	// Handler is the name of the plugin function to invoke.
+	Handler string `json:"handler"`
+	// Description is shown in /api/plugins so an operator can see what a plugin
+	// has exposed without reading the .lua.
+	Description string `json:"description,omitempty"`
+	// Plugin is filled in by HTTPHandlers: which plugin registered this route.
+	// The gateway dispatcher needs it to know which Lua state to enter, and
+	// deriving it from the path would mean trusting the plugin's own prefix.
+	Plugin string `json:"-"`
+}
+
+// routePath renders the handler's full mount path.
+func (h HTTPHandler) routePath() string {
+	return "/api/plugins/http/" + strings.Trim(h.Path, "/")
+}
+
+// allows reports whether a caller may invoke this handler.
+//
+// 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 {
+	return accessRuleFor(h.Roles, h.Keys, h.Public).AllowsCapability(role, keyName)
+}
+
+func (ps *Plugins) StateVisibleTo(name, role string) bool {
+	if role == "admin" {
+		return true
+	}
+	ps.mu.RLock()
+	defer ps.mu.RUnlock()
+	for _, p := range ps.plugins {
+		if p.Info.Name != name {
+			continue
+		}
+		override := ps.visibility[name]
+		if override != "" {
+			return audienceAllows(override, role)
+		}
+		declared := ""
+		if p.UI != nil {
+			declared = p.UI.Visibility
+		}
+		return audienceAllows(declared, role)
+	}
+	// An unknown plugin name: no visibility can be established, and this endpoint
+	// answers with an empty state rather than an error, so refusing here would
+	// change behaviour for a name that may simply not be installed.
+	return true
+}
+
+func containsPage(pages []*UIPage, id string) bool {
+	for _, p := range pages {
+		if p != nil && p.PageID == id {
+			return true
+		}
+	}
+	return false
+}
+
+// readPluginHTTP collects plugin.http, the endpoints the plugin asks the gateway
+// to mount.
+//
+// A malformed entry is DROPPED rather than failing the plugin: a plugin with one
+// bad route should still load its hooks and pages, and the bad route simply does
+// 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 {
+	// `http` is a LIST of declarations, each naming the handler function by
+	// STRING:
+	//
+	//	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 {
+		h, ok := it.(map[string]interface{})
+		if !ok {
+			continue
+		}
+		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"),
+		}
+		// 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 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. 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
+}
+
+// uipageFromValue builds a UIPage from a decoded Lua table.
+func uipageFromValue(m map[string]interface{}) *UIPage {
+	pg := &UIPage{
+		PageID:   asString(m["page_id"]),
+		Title:    asString(m["title"]),
+		Icon:     asString(m["icon"]),
+		Mount:    asString(m["mount"]),
+		Audience: asString(m["audience"]),
+	}
+	if n, ok := asFloat(m["order"]); ok {
+		pg.Order = int(n)
+	}
+	return pg
+}

internal/lua/plugins/billing.lua

@@ -192,6 +192,11 @@ local function priceFor(payload)
   out = {
     prompt = d.prompt or 0, completion = d.completion or 0,
     per_request = 0, cache_discount = d.cache_discount, peak = d.peak,
+      -- nil = bill reasoning at the completion rate, which is what every
+      -- provider measured actually charges. Set per-rule only for a provider
+      -- that publishes a distinct rate for thinking tokens.
+      reasoning_price = d.reasoning_price,
+      long_context = d.long_context,
   }
 
   -- model dimension (a token price overrides the default's token prices)
@@ -202,6 +207,8 @@ local function priceFor(payload)
     if mp.completion ~= nil then out.completion = mp.completion end
     if mp.per_request ~= nil then out.per_request = out.per_request + mp.per_request end
     if mp.cache_discount ~= nil then out.cache_discount = mp.cache_discount end
+      if mp.reasoning_price ~= nil then out.reasoning_price = mp.reasoning_price end
+      if mp.long_context ~= nil then out.long_context = mp.long_context end
     if mp.peak ~= nil then out.peak = mp.peak end
   end
 
@@ -213,6 +220,8 @@ local function priceFor(payload)
     if sp.completion ~= nil then out.completion = sp.completion end
     if sp.per_request ~= nil then out.per_request = out.per_request + sp.per_request end
     if sp.cache_discount ~= nil then out.cache_discount = sp.cache_discount end
+      if sp.reasoning_price ~= nil then out.reasoning_price = sp.reasoning_price end
+      if sp.long_context ~= nil then out.long_context = sp.long_context end
     if sp.peak ~= nil then out.peak = sp.peak end
   end
 
@@ -346,6 +355,60 @@ local function costFor(payload, price)
              + cacheHit * price.prompt * discount
              + completion * price.completion * mult
 
+  -- Long-context step: the tokens PAST a threshold cost more.
+  --
+  -- A marginal rate on the excess, not a blended average — that is how
+  -- providers publish it (DeepSeek doubles past 128k) and what keeps it
+  -- composable: a cache read beyond the threshold gets BOTH the cache discount
+  -- AND the long-context rate, as it is billed. A blended average would have to
+  -- choose one, and choosing wrong under-bills exactly the cache-heavy long
+  -- requests the cache exists to make cheap.
+  --
+  -- The excess splits between cached and fresh in the same proportion as the
+  -- whole prompt, so no token is charged twice and none is missed.
+  local lc = price.long_context
+  if lc then
+    local totalPrompt = tonumber(payload.prompt_tokens) or 0
+    local threshold = tonumber(lc.above) or 0
+    local lcRate = tonumber(lc.prompt)
+    if totalPrompt > threshold and threshold > 0 and lcRate and lcRate > 0 then
+      local excess = totalPrompt - threshold
+      -- Only the FRESH share of the excess carries the step: a cache read is
+      -- already billed at the discounted rate, and re-pricing it at the
+      -- long-context rate would charge the same tokens twice.
+      local freshShare = totalPrompt > 0 and (fresh / totalPrompt) or 0
+      local excessFresh = excess * freshShare
+      local excessCached = excess - excessFresh
+      cost = cost
+          + excessFresh * (lcRate - price.prompt) * mult
+          + excessCached * (lcRate - price.prompt * discount) * mult
+    end
+  end
+
+    -- Reasoning tokens, when the rule prices them separately.
+    --
+    -- reasoning_tokens is a SUBSET of completion_tokens and is already inside
+    -- the term above. This block only matters when a provider charges thinking
+    -- at a DIFFERENT rate — and every provider measured does not: OpenAI,
+    -- DeepSeek and Anthropic all bill reasoning at the ordinary output rate. So
+    -- reasoning_price is nil by default and this is skipped entirely, which is
+    -- what makes the standard formula match the invoice.
+    --
+    -- When it IS set, only the DIFFERENCE is applied: the reasoning portion is
+    -- re-priced from the normal rate to the reasoning rate. Charging it a
+    -- second time in full would double-bill every thinking token.
+    local reasoningPrice = tonumber(price.reasoning_price)
+    if reasoningPrice and reasoningPrice ~= price.completion then
+      local reasoning = tonumber(payload.reasoning_tokens) or 0
+      if reasoning > 0 then
+        -- Defensive clamp: a provider reporting more reasoning than completion
+        -- would otherwise produce a negative visible-output count and a cost
+        -- delta larger than the whole request.
+        if reasoning > completion then reasoning = completion end
+        cost = cost + reasoning * (reasoningPrice - price.completion) * mult
+      end
+    end
+
   local flat = price.per_request
   if not payload.ok and not plugin.count_failures then
     flat = 0
@@ -442,6 +505,21 @@ function plugin.on_request_end(payload)
     s.unpriced_reqs = s.unpriced_reqs + 1
     local m = payload.model or "?"
     s.unpriced_models[m] = (s.unpriced_models[m] or 0) + 1
+
+  -- Day-keyed unpriced tally. The lifetime counters above cannot answer the
+  -- question an operator actually has — "is it STILL happening?" — because they
+  -- keep the debt from before a pricing rule was added. After pricing a source
+  -- that was mispriced, the lifetime number never moves, so "unpriced: 6376" on
+  -- a dashboard is indistinguishable from a system that is still billing
+  -- nothing. Counting by day makes the fix observable: the offending day keeps
+  -- its number, new days start at zero.
+  if s.unpriced_by_day == nil then s.unpriced_by_day = {} end
+  -- dayKey wants epoch SECONDS; payload.time is milliseconds.
+  local uday = dayKey(math.floor((tonumber(payload.time) or (os.time() * 1000)) / 1000))
+  local ubucket = s.unpriced_by_day[uday]
+  if ubucket == nil then ubucket = { reqs = 0, models = {} }; s.unpriced_by_day[uday] = ubucket end
+  ubucket.reqs = ubucket.reqs + 1
+  ubucket.models[m] = (ubucket.models[m] or 0) + 1
   end
   if s.degraded_reqs == nil then s.degraded_reqs = 0 end
   -- Degradation is counted here rather than in the chain_step hook because
@@ -503,6 +581,24 @@ end
 -- A whole page. The kernel injects this HTML and evaluates the <script> after
 -- the DOM exists, and exposes `pluginAPI` for talking to the gateway.
 plugin.ui = {
+  -- Readable by any role, because the state is SPLIT below: a user key sees
+  -- its own spend and nothing else. Marking the page admin-only would make the
+  -- split pointless; marking it "all" without the split is exactly the leak this
+  -- closes — a user key pulled the whole 18KB ledger (16871 requests, $8.86,
+  -- every source name) straight out of GET /state. The gateway's audience
+  -- override can tighten this per plugin without editing this file.
+  visibility = "all",
+  -- Split the state: global figures are deployment data and are withheld from
+  -- non-admin callers, while the per-key buckets describe ONE caller.
+  -- depth: by_key is {maskedKey: bucket}; by_day_key is {day: {maskedKey: …}}.
+  state_scope = {
+    global = {
+      "total", "by_source", "by_model", "by_day", "by_day_src",
+      "by_day_model", "by_tier_served", "skip_reasons", "degraded_reqs",
+      "unpriced_models", "unpriced_reqs", "unpriced_by_day", "started",
+    },
+    depth = { by_key = 1, by_day_key = 2 },
+  },
   page = {
     page_id = "billing",
     title = "Billing",
@@ -575,6 +671,7 @@ plugin.ui = {
     en: {
       total: "Total", requests: "Requests", degraded: "Degraded",
       unpriced: "Unpriced", prompt: "Prompt tokens",
+      unpriced7d: "Unpriced (7d)", unpricedClean: "none in last 7 days",
       completion: "Completion tokens", failures: "Failures",
       cacheRate: "Cache hit rate", cacheTokens: "Cache read tokens",
       perSource: "Per source", perModel: "Per model",
@@ -596,6 +693,7 @@ plugin.ui = {
     zh: {
       total: "总开销", requests: "请求数", degraded: "降级",
       unpriced: "未定价", prompt: "输入 tokens",
+      unpriced7d: "未定价（近 7 天）", unpricedClean: "近 7 天无",
       completion: "输出 tokens", failures: "失败",
       cacheRate: "缓存命中率", cacheTokens: "缓存读取 tokens",
       perSource: "按源", perModel: "按模型",
@@ -716,13 +814,35 @@ plugin.ui = {
     });
     return out;
   }
+  // recentUnpriced counts unpriced requests over the last n days. When the
+  // result is zero THAT is the useful signal: every pricing rule added since
+  // then is holding. Null when there is no day data at all (a state file
+  // written before the buckets existed) so the UI can say "unknown" rather than
+  // claim a clean record it cannot support.
+  function recentUnpriced(st, n) {
+    var byDay = st && st.unpriced_by_day;
+    if (!byDay) return null;
+    var cutoff = Date.now() - n * 86400000;
+    var sum = 0, seen = false;
+    Object.keys(byDay).forEach(function (d) {
+      var t = Date.parse(d + "T00:00:00Z");
+      if (isNaN(t)) return;
+      seen = true;
+      if (t >= cutoff) sum += (byDay[d].reqs || 0);
+    });
+    return seen ? sum : null;
+  }
+
   // rescale returns the state as the selected period presents it. Degraded and
   // unpriced counters are lifetime-only (there is no day-keyed tally for them)
   // so they are returned as null and the UI hides them instead of printing a
   // lifetime number beside a per-window cost — the exact mismatch the period
   // view exists to remove.
   function rescale(st) {
-    if (PERIOD === "all" || !st.by_day) return st;
+    if (PERIOD === "all" || !st.by_day) {
+      if (st) st.unpriced_recent = recentUnpriced(st, 7);
+      return st;
+    }
     var days = daysInWindow(PERIOD, utcToday());
     var total = zeroBucket();
     var byDay = {};
@@ -740,6 +860,10 @@ plugin.ui = {
       by_day: byDay,
       degraded_reqs: null,
       unpriced_reqs: null,
+      // The lifetime unpriced total is null here for the same reason, but the
+      // day buckets are not — they are what lets the card report "0 in the last
+      // 7 days" instead of leaving a settled debt indistinguishable from a leak.
+      unpriced_recent: recentUnpriced(st, 7),
     };
     return out;
   }
@@ -820,6 +944,15 @@ plugin.ui = {
       // print a lifetime number next to a per-window cost.
       view.degraded_reqs != null ? [T.degraded, fmtInt(view.degraded_reqs)] : null,
       view.unpriced_reqs != null ? [T.unpriced, fmtInt(view.unpriced_reqs)] : null,
+      // "Unpriced (7d)" answers what the lifetime number cannot: is it STILL
+      // happening? A settled debt and a live leak print the same lifetime total.
+      view.unpriced_recent != null
+        ? [
+            T.unpriced7d,
+            fmtInt(view.unpriced_recent),
+            view.unpriced_recent === 0 ? T.unpricedClean : null,
+          ]
+        : null,
       [T.prompt, fmtInt(t.prompt_tokens || 0)],
       [T.completion, fmtInt(t.completion_tokens || 0)],
       [T.failures, t.failures || 0],
@@ -931,7 +1064,11 @@ plugin.ui = {
       monthly: "Monthly fee", peakMult: "Peak multiplier",
       peakHours: "Peak hours (UTC)", peakDays: "Peak days (UTC 1=Mon)",
       models: "Per-model prices", addModel: "+ model",
-      mFree: "free", mToken: "token", mSub: "subscription", mUnpriced: "unpriced",
+
+      reasoning: "think $/M",
+
+      lcAbove: "over n tok", lcPrompt: "then $/M", lcMult: "or x",
+      lcHint: "Long-context step: tokens PAST the threshold cost more. Leave all three blank for one flat rate.",      reasoningHint: "Thinking tokens are billed as output unless you set a different rate here.",      mFree: "free", mToken: "token", mSub: "subscription", mUnpriced: "unpriced",
       noProfile: "No profile yet — create one to start pricing.",
       noRules: "No rules in this profile",
       warnNoSource: "matches no configured source",
@@ -953,7 +1090,11 @@ plugin.ui = {
       monthly: "月费", peakMult: "峰段倍数",
       peakHours: "峰段小时 (UTC)", peakDays: "峰段星期 (UTC 1=周一)",
       models: "按模型价格", addModel: "+ 模型",
-      mFree: "free 免费", mToken: "token 按量",
+
+      reasoning: "思考 $/M",
+
+      lcAbove: "超过 n tok", lcPrompt: "此后 $/M", lcMult: "或 ×",
+      lcHint: "长上下文阶梯：超过阈值的部分按更高价计。三项都留空 = 单一价格。",      reasoningHint: "留空=思考 token 按输出价计（OpenAI/DeepSeek/Anthropic 都这样）。只有供应商对思考单独定价时才填。",      mFree: "free 免费", mToken: "token 按量",
       mSub: "subscription 订阅", mUnpriced: "unpriced 只统计",
       noProfile: "还没有方案 —— 新建一个开始定价。",
       noRules: "该方案暂无规则",
@@ -1051,6 +1192,25 @@ plugin.ui = {
         "<input class='m-name' value='" + esc(m) + "' style='flex:1 1 90px;min-width:0'>" +
         "<input class='m-p' value='" + esc(t.prompt == null ? "" : t.prompt) +
           "' placeholder='" + esc(T.prompt) + "' style='flex:0 1 62px;min-width:0'>" +
+          // Left BLANK by default on purpose: empty means "bill thinking at the
+          // completion rate", which is what OpenAI, DeepSeek and Anthropic do.
+          // A 0 here would mean "thinking is free" and would under-bill quietly.
+          "<input class='m-r' value='" + esc(t.reasoning == null ? "" : t.reasoning) +
+            "' placeholder='" + esc(T.reasoning) + "' title='" + esc(T.reasoningHint) +
+            "' style='flex:0 1 62px;min-width:0'>" +
+          // Long-context step: threshold + (absolute rate OR multiplier). All
+          // three blank means NO step, deliberately distinct from a step whose
+          // rate is 0 — three inputs rather than inferring the step from a zero.
+          // On an agent workload 88% of requests exceed 128k prompt tokens, so
+          // omitting this under-bills the MAJORITY while every number stays
+          // plausible. The step is MARGINAL: it reprices the excess only.
+          "<input class='m-lc-a' value='" + esc(t.long_context == null ? "" : t.long_context.above) +
+            "' placeholder='" + esc(T.lcAbove) + "' title='" + esc(T.lcHint) +
+            "' style='flex:0 1 52px;min-width:0'>" +
+          "<input class='m-lc-p' value='" + esc(t.long_context == null ? "" : (t.long_context.prompt || "")) +
+            "' placeholder='" + esc(T.lcPrompt) + "' style='flex:0 1 52px;min-width:0'>" +
+          "<input class='m-lc-m' value='" + esc(t.long_context == null ? "" : (t.long_context.multiplier || "")) +
+            "' placeholder='" + esc(T.lcMult) + "' style='flex:0 1 46px;min-width:0'>" +
         "<input class='m-c' value='" + esc(t.completion == null ? "" : t.completion) +
           "' placeholder='" + esc(T.completion) + "' style='flex:0 1 62px;min-width:0'>" +
         "<button class='ghost small m-del' title='" + esc(T.delRule) + "' style='flex:0 0 auto'>&times;</button></div>";
@@ -1127,6 +1287,39 @@ plugin.ui = {
           return;
         }
         models[n] = { prompt: pv, completion: cv };
+
+          // Optional thinking-token rate, read from the same row. Blank means
+          // "bill reasoning at the completion rate" — what OpenAI, DeepSeek and
+          // Anthropic all do. A 0 would mean "thinking is free", so an empty
+          // field must stay ABSENT from the saved rule rather than becoming 0.
+          var rel = r.querySelector(".m-r");
+          var rtxt = rel ? String(rel.value || "").trim() : "";
+          if (rtxt !== "") {
+            if (isNaN(parseFloat(rtxt))) {
+              bad = n + " (" + T.reasoning + ")";
+              return;
+            }
+            models[n].reasoning = rtxt;
+          }
+          // Long-context step. All three blank means NO step — the key must be
+          // ABSENT from the saved rule rather than present-and-zero, which would
+          // be a step priced at nothing.
+          var lca = r.querySelector(".m-lc-a");
+          var lcp = r.querySelector(".m-lc-p");
+          var lcm = r.querySelector(".m-lc-m");
+          var lcaT = lca ? String(lca.value || "").trim() : "";
+          var lcpT = lcp ? String(lcp.value || "").trim() : "";
+          var lcmT = lcm ? String(lcm.value || "").trim() : "";
+          if (lcaT !== "" || lcpT !== "" || lcmT !== "") {
+            if (lcaT === "" || isNaN(parseFloat(lcaT))) { bad = n + " (" + T.lcAbove + ")"; return; }
+            if (lcpT !== "" && isNaN(parseFloat(lcpT))) { bad = n + " (" + T.lcPrompt + ")"; return; }
+            if (lcmT !== "" && isNaN(parseFloat(lcmT))) { bad = n + " (" + T.lcMult + ")"; return; }
+            if (lcpT === "" && lcmT === "") { bad = n + " (" + T.lcPrompt + "/" + T.lcMult + ")"; return; }
+            var lc = { above: String(parseFloat(lcaT)) };
+            if (lcpT !== "") lc.prompt = lcpT;
+            if (lcmT !== "") lc.multiplier = parseFloat(lcmT);
+            models[n].long_context = lc;
+          }
       });
       if (bad) throw new Error(T.needPrices + ": " + bad);
       if (Object.keys(models).length) out.models = models;
@@ -1317,6 +1510,11 @@ plugin.ui = {
           "<input class='m-name' placeholder='model' style='flex:1;min-width:80px'>" +
           "<input class='m-p' placeholder='" + esc(T.prompt) + "' style='width:96px'>" +
           "<input class='m-c' placeholder='" + esc(T.completion) + "' style='width:96px'>" +
+            "<input class='m-r' placeholder='" + esc(T.reasoning) + "' title='" + esc(T.reasoningHint) +
+              "' style='width:96px'>" +
+            "<input class='m-lc-a' placeholder='" + esc(T.lcAbove) + "' title='" + esc(T.lcHint) + "' style='width:78px'>" +
+            "<input class='m-lc-p' placeholder='" + esc(T.lcPrompt) + "' style='width:78px'>" +
+            "<input class='m-lc-m' placeholder='" + esc(T.lcMult) + "' style='width:70px'>" +
           "<button class='ghost small m-del'>&times;</button>";
         box.appendChild(row);
       };

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)
+	}
+}

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