feat(keys): 每个密钥可配独立 AUTO 链，管理员代配

此前 AUTO 链是全局单值（cfg.Auto + Core.AutoChain()），所有用户共用一条链。管理员无法为某个用户单独指定调度链。按 per-key 覆盖 + 全局兜底实现： - config.GWKey 增加 Auto 字段与 HasOwnAuto()。未配置即继承全局链，存量部署零改动，新密钥天然继承全局链。 - Core 把 buildAutoChain 的归一化逻辑抽成 chainForRules，全局链、生图链、per-key 链共用同一套编译，避免两处漂移。 - Core 增加 keyAutoChains 缓存 + AutoChainFor(key)。请求路径读缓存不加锁，与全局链查询一致。缓存整表原子替换，不会看到半成品。 - 请求侧 chat.go 改用 AutoChainFor(reqKey)。冷却仍在 Provider 上按 model+source 共享：两条链指向同一个 slot 时共用冷却，与今天单链行为相同，也避免为 per-key 维度重构冷却而改变现有可观测语义。 - API：GET/PUT/DELETE /api/keys/{key}/auto（admin），GET /api/keys/me/auto（任意角色，只能读自己的）。空 PUT 与 DELETE 等价于 "恢复继承"，无法持久化一条会 503 的空链。写入时回报解析出的槽位数，让管理员当场看到模型名写错，而不是等用户下次请求 503。 - 源变更时一并重编译 per-key 链，加源后无需重启即可生效。判据 18 条，5 个变异全部被抓住：AutoChainFor 忽略 key、清空后不重建缓存、源变更不重建、空 PUT 落盘成空链、me/auto 误要求 admin。 UI 判据做变异时发现漏放：只查函数定义存在，删掉按钮后仍通过。已改为断言 keyCanvasHtml 内的调用点。端到端实测（真实 HTTP + 两个 mock 上游）：admin 与 bob 初始同为 m-fast，给 bob 配 m-cheap 后两者分流，重启后仍分流，DELETE 后 bob 回到 m-fast。 Co-Authored-By: ModelRouter <noreply@modelrouter.dev>
2026-10-03 23:54:06 +00:00 · 2026-10-03 07:45:54 +08:00
parent ba01da937b
commit 22decf2bd3
8 changed files with 897 additions and 12 deletions
--- a/internal/config/config.go
+++ b/internal/config/config.go
@ -384,15 +384,25 @@ type SourceTemplate struct {
 // (full management) or "user" (sees only its own key); Models is the allowed
 // model scope with per-model token quota (0 = unlimited).
 type GWKey struct {
-	Key       string       `yaml:"key" json:"key"`
-	Role      string       `yaml:"role" json:"role"`
-	Name      string       `yaml:"name,omitempty" json:"name,omitempty"`
-	Models    []ModelScope `yaml:"models,omitempty" json:"models,omitempty"`
-	Note      string       `yaml:"note,omitempty" json:"note,omitempty"`
-	CreatedAt int64        `yaml:"created_at,omitempty" json:"created_at,omitempty"`
-	Seed      bool         `yaml:"seed,omitempty" json:"seed,omitempty"` // true if migrated from config gateway_keys
+	Key   string `yaml:"key" json:"key"`
+	Role  string `yaml:"role" json:"role"`
+	Name  string `yaml:"name,omitempty" json:"name,omitempty"`
+	// Auto is this key's own AUTO scheduling chain. It overrides the global
+	// cfg.Auto for requests authenticated with this key; nil (or an empty
+	// list) means "inherit the global chain", so keys created before per-key
+	// chains existed keep working untouched.
+	Auto           []ModelScope `yaml:"auto,omitempty" json:"auto,omitempty"`
+	Models         []ModelScope `yaml:"models,omitempty" json:"models,omitempty"`
+	Note           string       `yaml:"note,omitempty" json:"note,omitempty"`
+	CreatedAt      int64        `yaml:"created_at,omitempty" json:"created_at,omitempty"`
+	Seed           bool         `yaml:"seed,omitempty" json:"seed,omitempty"` // true if migrated from config gateway_keys
 }

+// HasOwnAuto reports whether the key declares its own AUTO chain rather than
+// inheriting the global one. An explicitly empty list still counts as "no own
+// chain" so a user can clear an override back to the global default.
+func (k GWKey) HasOwnAuto() bool { return len(k.Auto) > 0 }
+
 // BillingDSL holds declarative per-URL pricing profiles for the billing plugin.
 //
 // Profiles are the "let the user pick" axis: the same upstream URL can be
--- a/internal/core/core.go
+++ b/internal/core/core.go
@ -37,6 +37,12 @@ type Core struct {
 	registry       *provider.Registry
 	autoChain      atomic.Pointer[scheduler.Chain] // chat AUTO chain
 	autoImageChain atomic.Pointer[scheduler.Chain] // image-generation AUTO chain
+
+	// keyAutoChains caches per-key AUTO chains, keyed by the gateway key.
+	// Entries are rebuilt by rebuildKeyAutoChains whenever sources or key
+	// config change; AutoChainFor reads them without taking c.mu so request
+	// handling stays lock-free like the global chain lookup.
+	keyAutoChains atomic.Pointer[map[string]*scheduler.Chain]
 }

 // New builds the core from a config file plus runtime overlay.
@ -257,6 +263,68 @@ func (c *Core) Registry() *provider.Registry { return c.registry }
 // AutoChain returns the current AUTO scheduling chain.
 func (c *Core) AutoChain() *scheduler.Chain { return c.autoChain.Load() }

+// AutoChainFor returns the AUTO chain to use for a request authenticated with
+// key. A key that declares its own chain gets that chain; every other key
+// inherits the global chain, so keys created before per-key chains existed
+// keep their current behaviour with no configuration change.
+//
+// It returns (chain, ok). ok is false only when the key has its own chain but
+// it compiled to nothing usable (e.g. every slot names a model that no longer
+// exists) — the caller must surface that rather than silently downgrading the
+// user to the global chain, which would be a confusing, invisible switch.
+func (c *Core) AutoChainFor(key string) (*scheduler.Chain, bool) {
+	if m := c.keyAutoChains.Load(); m != nil {
+		if ch, ok := (*m)[key]; ok {
+			return ch, true
+		}
+	}
+	return c.autoChain.Load(), true
+}
+
+// rebuildKeyAutoChains recompiles every per-key AUTO chain from the current
+// config and provider registry, then swaps the cache in one shot so requests
+// never observe a half-built map. Keys without their own chain are absent
+// from the map and fall back to the global chain.
+func (c *Core) rebuildKeyAutoChains() {
+	m := make(map[string]*scheduler.Chain)
+	for _, k := range c.cfg.Keys {
+		if !k.HasOwnAuto() {
+			continue
+		}
+		m[k.Key] = c.chainForRules(k.Auto, false)
+	}
+	c.keyAutoChains.Store(&m)
+}
+
+// SaveKeyAuto persists one key's own AUTO chain and recompiles the per-key
+// chain cache. Passing an empty list clears the override so the key inherits
+// the global chain again.
+func (c *Core) SaveKeyAuto(key string, entries []config.ModelScope) error {
+	c.mu.Lock()
+	defer c.mu.Unlock()
+	idx := -1
+	for i, k := range c.cfg.Keys {
+		if k.Key == key {
+			idx = i
+			break
+		}
+	}
+	if idx < 0 {
+		return fmt.Errorf("key not found")
+	}
+	if err := ValidateScopeQuotas(entries); err != nil {
+		return err
+	}
+	c.cfg.Keys[idx].Auto = cleanScopes(entries)
+	if err := c.saveConfig(); err != nil {
+		return err
+	}
+	// Rebuild the whole cache so a cleared override stops shadowing the
+	// global chain for this key.
+	c.rebuildKeyAutoChains()
+	return nil
+}
+
 // AutoImageChain returns the persisted image-generation AUTO chain snapshot.
 func (c *Core) AutoImageChain() *scheduler.Chain { return c.autoImageChain.Load() }

@ -546,6 +614,7 @@ func (c *Core) rebuildRegistry() error {
 	}
 	c.buildAutoChain()
 	c.buildAutoImageChain()
+	c.rebuildKeyAutoChains()
 	return nil
 }

@ -555,17 +624,28 @@ func (c *Core) rebuildRegistry() error {
 // back to the source's best chat model and cooldown/quota bookkeeping would
 // key on a name that never matches.
 func (c *Core) buildAutoChain() {
+	c.autoChain.Store(c.chainForRules(c.cfg.Auto, false))
+}
+
+// chainForRules compiles a chat AUTO chain from a rule list. kindFilter is
+// false for the chat chain (image models are skipped) and true for the image
+// chain (only image models participate). Per-key chains always compile as
+// chat chains: a user's key configures the models *they* talk to.
+func (c *Core) chainForRules(rules []config.ModelScope, imageOnly bool) *scheduler.Chain {
 	prov := func(model, source string) scheduler.Provider {
 		p := c.registry.ProviderForSlot(model, source)
 		if p == nil {
 			return nil
 		}
-		if m := p.ModelByID(model); m != nil && m.Kind == "image" {
+		m := p.ModelByID(model)
+		if imageOnly && (m == nil || m.Kind != "image") {
+			return nil
+		}
+		if !imageOnly && m != nil && m.Kind == "image" {
 			return nil
 		}
 		return p
 	}
-	rules := c.cfg.Auto
 	sr := make([]scheduler.Rule, 0, len(rules))
 	for _, e := range rules {
 		model, source := e.Model, e.Source
@ -574,7 +654,11 @@ func (c *Core) buildAutoChain() {
 				model = exact
 			}
 			if m := p.ModelByID(model); m != nil && m.Kind == "image" {
-				continue // image-kind models never join the chat AUTO chain
+				if !imageOnly {
+					continue // image-kind models never join the chat AUTO chain
+				}
+			} else if imageOnly {
+				continue
 			}
 			if source == "" {
 				source = p.Name()
@ -589,7 +673,7 @@ func (c *Core) buildAutoChain() {
 			Hours:  e.Hours,
 		})
 	}
-	c.autoChain.Store(scheduler.BuildChain(sr, prov))
+	return scheduler.BuildChain(sr, prov)
 }

 // buildAutoImageChain rebuilds the image-generation AUTO chain snapshot.
--- a/internal/core/key_auto_test.go
+++ b/internal/core/key_auto_test.go
@ -0,0 +1,221 @@
+package core
+
+import (
+	"path/filepath"
+	"testing"
+
+	"llmsproxy/internal/config"
+	"llmsproxy/internal/scheduler"
+)
+
+// Per-key AUTO chains let an admin give one user their own scheduling chain
+// while every other key keeps the global one. These tests pin the three
+// properties that make that safe:
+//
+//  1. a key WITH its own chain gets it;
+//  2. a key WITHOUT one inherits the global chain (backward compatible);
+//  3. the override is actually persisted, so it survives a restart.
+//
+// A test that only checked (1) would pass even if AutoChainFor ignored the key
+// and always returned the global chain whenever the two happened to be equal,
+// so (2) uses deliberately *different* chains.
+
+func perKeyTestConfig(t *testing.T) *config.Config {
+	t.Helper()
+	dir := t.TempDir()
+	return &config.Config{
+		Path:         filepath.Join(dir, "config.yaml"),
+		AdapterDir:   filepath.Join(dir, "adapters"),
+		RuntimeFile:  filepath.Join(dir, "runtime.json"),
+		Listen:       "127.0.0.1:0",
+		DefaultModel: "AUTO",
+		GatewayKeys:  []string{"sk-gw-admin"},
+		Sources: []config.Source{
+			{Name: "s1", BaseURL: "http://127.0.0.1:1/v1", Adapter: "openai",
+				Models: []config.Model{{ID: "gpt-4o", Priority: 10}, {ID: "gpt-4o-mini", Priority: 5}}},
+		},
+		// Global chain points at gpt-4o.
+		Auto: []config.ModelScope{{Model: "gpt-4o", Source: "s1", Tier: 1}},
+		Keys: []config.GWKey{
+			{Key: "sk-gw-admin", Role: "admin"},
+			{Key: "sk-gw-inherits", Role: "user", Name: "inherits"},
+			// This key's own chain points at a DIFFERENT model, so returning
+			// the global chain by mistake is detectable.
+			{Key: "sk-gw-own", Role: "user", Name: "own",
+				Auto: []config.ModelScope{{Model: "gpt-4o-mini", Source: "s1", Tier: 1}}},
+		},
+	}
+}
+
+func firstSlotModel(t *testing.T, c *Core, key string) string {
+	t.Helper()
+	ch, _ := c.AutoChainFor(key)
+	if ch == nil || len(ch.Tiers) == 0 {
+		t.Fatalf("key %s: no chain tiers", key)
+	}
+	slots := ch.Tiers[0].Slots
+	if len(slots) == 0 {
+		t.Fatalf("key %s: tier 1 has no slots", key)
+	}
+	return slots[0].Model
+}
+
+// A key that declares its own chain must be scheduled by it, not the global.
+func TestAutoChainForUsesKeyOwnChain(t *testing.T) {
+	c := newTestCore(t, perKeyTestConfig(t))
+
+	if got := firstSlotModel(t, c, "sk-gw-own"); got != "gpt-4o-mini" {
+		t.Fatalf("key with own chain must use it, got %q (want gpt-4o-mini)", got)
+	}
+	// Sanity: the global chain really is different, so the assertion above
+	// cannot pass by accident.
+	if got := firstSlotModel(t, c, "sk-gw-admin"); got != "gpt-4o" {
+		t.Fatalf("admin must use global chain, got %q (want gpt-4o)", got)
+	}
+}
+
+// A key with no own chain inherits the global one. This is what keeps every
+// pre-existing deployment working with no config change.
+func TestAutoChainForInheritsGlobalWhenUnset(t *testing.T) {
+	c := newTestCore(t, perKeyTestConfig(t))
+
+	if got := firstSlotModel(t, c, "sk-gw-inherits"); got != "gpt-4o" {
+		t.Fatalf("key without own chain must inherit global, got %q (want gpt-4o)", got)
+	}
+}
+
+// A key that is not in the key table at all (e.g. a seed key) must also
+// inherit the global chain rather than erroring or returning nil.
+func TestAutoChainForUnknownKeyInheritsGlobal(t *testing.T) {
+	c := newTestCore(t, perKeyTestConfig(t))
+
+	ch, ok := c.AutoChainFor("sk-gw-never-configured")
+	if !ok || ch == nil || len(ch.Tiers) == 0 {
+		t.Fatalf("unknown key must fall back to global chain, got ok=%v chain=%v", ok, ch)
+	}
+}
+
+// The per-key chain must survive a save/reload round trip: an admin configuring
+// a chain in the WebUI and restarting the gateway must not silently lose it.
+func TestSaveKeyAutoPersistsAndApplies(t *testing.T) {
+	cfg := perKeyTestConfig(t)
+	c := newTestCore(t, cfg)
+
+	if err := c.SaveKeyAuto("sk-gw-inherits", []config.ModelScope{
+		{Model: "gpt-4o-mini", Source: "s1", Tier: 1},
+	}); err != nil {
+		t.Fatalf("SaveKeyAuto: %v", err)
+	}
+	if got := firstSlotModel(t, c, "sk-gw-inherits"); got != "gpt-4o-mini" {
+		t.Fatalf("saved chain must apply immediately, got %q", got)
+	}
+
+	// Reload from disk exactly like a restart does.
+	reloaded, err := config.Load(cfg.Path)
+	if err != nil {
+		t.Fatalf("reload config: %v", err)
+	}
+	c2 := newTestCore(t, reloaded)
+	if got := firstSlotModel(t, c2, "sk-gw-inherits"); got != "gpt-4o-mini" {
+		t.Fatalf("per-key chain must survive restart, got %q", got)
+	}
+	// The other key must be unaffected by its neighbour's override.
+	if got := firstSlotModel(t, c2, "sk-gw-admin"); got != "gpt-4o" {
+		t.Fatalf("override must not leak to other keys, got %q", got)
+	}
+}
+
+// Clearing the override must return the key to the global chain, both in
+// memory and after a restart. This is what the WebUI's "use global" toggle
+// does, so a stale shadowing entry would silently pin the user to their old
+// chain forever.
+func TestSaveKeyAutoClearRestoresGlobal(t *testing.T) {
+	cfg := perKeyTestConfig(t)
+	c := newTestCore(t, cfg)
+
+	if err := c.SaveKeyAuto("sk-gw-own", nil); err != nil {
+		t.Fatalf("clear: %v", err)
+	}
+	if got := firstSlotModel(t, c, "sk-gw-own"); got != "gpt-4o" {
+		t.Fatalf("cleared key must inherit global again, got %q", got)
+	}
+
+	reloaded, err := config.Load(cfg.Path)
+	if err != nil {
+		t.Fatalf("reload: %v", err)
+	}
+	c2 := newTestCore(t, reloaded)
+	if got := firstSlotModel(t, c2, "sk-gw-own"); got != "gpt-4o" {
+		t.Fatalf("clear must survive restart, got %q", got)
+	}
+}
+
+// Saving an unknown key must fail loudly rather than creating a shadow entry
+// that no request can ever match.
+func TestSaveKeyAutoUnknownKeyFails(t *testing.T) {
+	c := newTestCore(t, perKeyTestConfig(t))
+
+	if err := c.SaveKeyAuto("sk-gw-nope", []config.ModelScope{{Model: "gpt-4o", Source: "s1"}}); err == nil {
+		t.Fatal("SaveKeyAuto on unknown key must return an error")
+	}
+}
+
+// Per-key quota validation must run, exactly like the model scope path: an
+// override is another place an operator can write a bad budget.
+func TestSaveKeyAutoRejectsBadQuota(t *testing.T) {
+	c := newTestCore(t, perKeyTestConfig(t))
+
+	err := c.SaveKeyAuto("sk-gw-inherits", []config.ModelScope{
+		{Model: "gpt-4o", Source: "s1", TokenQuota: -5},
+	})
+	if err == nil {
+		t.Fatal("negative token quota in a per-key auto chain must be rejected")
+	}
+}
+
+// Source changes must recompile per-key chains. A key whose chain names a
+// model that does not exist yet compiles to an empty chain (BuildChain drops
+// slots it cannot resolve — sending the request into a black hole would be
+// worse), and must start working as soon as the source appears.
+func TestKeyAutoChainsRebuildOnSourceChange(t *testing.T) {
+	cfg := perKeyTestConfig(t)
+	cfg.Keys = append(cfg.Keys, config.GWKey{
+		Key: "sk-gw-late", Role: "user", Name: "late",
+		Auto: []config.ModelScope{{Model: "new-model", Source: "s2", Tier: 1}},
+	})
+	c := newTestCore(t, cfg)
+
+	// s2 does not exist yet, so the chain has no usable slot.
+	if ch, _ := c.AutoChainFor("sk-gw-late"); chainSlotsOf(ch) != 0 {
+		t.Fatalf("chain naming an unknown model must compile empty, got %d slots",
+			chainSlotsOf(ch))
+	}
+	// The other keys are unaffected by the broken one.
+	if got := firstSlotModel(t, c, "sk-gw-admin"); got != "gpt-4o" {
+		t.Fatalf("broken per-key chain must not affect others, got %q", got)
+	}
+
+	// Add the source; the per-key chain must pick it up without a restart.
+	cfg.Sources = append(cfg.Sources, config.Source{
+		Name: "s2", BaseURL: "http://127.0.0.1:2/v1", Adapter: "openai",
+		Models: []config.Model{{ID: "new-model", Priority: 10}},
+	})
+	if err := c.Reload(); err != nil {
+		t.Fatalf("Reload: %v", err)
+	}
+	if got := firstSlotModel(t, c, "sk-gw-late"); got != "new-model" {
+		t.Fatalf("per-key chain must resolve after the source is added, got %q", got)
+	}
+}
+
+// chainSlotsOf counts usable slots, tolerating a nil chain.
+func chainSlotsOf(ch *scheduler.Chain) int {
+	if ch == nil {
+		return 0
+	}
+	n := 0
+	for _, tn := range ch.Tiers {
+		n += len(tn.Slots)
+	}
+	return n
+}
--- a/internal/gateway/chat.go
+++ b/internal/gateway/chat.go
@ -400,7 +400,8 @@ func (g *Gateway) handleChat(w http.ResponseWriter, r *http.Request) {
 	// TestHooksFireOnRealDirectChat, not by reading the code.
 	g.fireStart(r.Context(), &req, "chat", model, len(req.Messages), len(req.Tools))
 	if isAuto(model) {
-		chain := g.core.AutoChain()
+		// The key's own chain wins; keys without one inherit the global chain.
+		chain, _ := g.core.AutoChainFor(reqKey(r.Context()))
 		if chain == nil || len(chain.Tiers) == 0 {
 			writeError(w, http.StatusServiceUnavailable, "no_provider", "no auto slot configured")
 			return
--- a/internal/gateway/key_auto_api_test.go
+++ b/internal/gateway/key_auto_api_test.go
@ -0,0 +1,257 @@
+package gateway
+
+import (
+	"encoding/json"
+	"net/http"
+	"net/http/httptest"
+	"path/filepath"
+	"os"
+	"strings"
+	"testing"
+
+	"llmsproxy/internal/config"
+	"llmsproxy/internal/core"
+)
+
+// Per-key AUTO chain admin API. The behaviours pinned here are the ones that
+// are easy to get wrong and invisible when they are:
+//
+//   - /api/keys/me/auto must be reachable by a NON-admin (it is their own
+//     chain); routing it through the generic {key}/auto handler would 403 it.
+//   - /api/keys/{key}/auto must 403 a non-admin.
+//   - an empty PUT and a DELETE mean the same thing ("inherit global"), so the
+//     WebUI cannot persist a chain that would then 503 with no slots.
+//   - a PUT that resolves to zero slots is reported back, so an admin who
+//     mistyped a model learns it now instead of from the user's next 503.
+
+// keyAutoGateway builds a gateway with one admin key and one user key.
+func keyAutoGateway(t *testing.T) *Gateway {
+	t.Helper()
+	td := t.TempDir()
+	cfgPath := filepath.Join(td, "config.yaml")
+	if err := os.WriteFile(cfgPath, []byte("listen: :0"), 0644); err != nil {
+		t.Fatal(err)
+	}
+	cfg := &config.Config{
+		Path:        cfgPath,
+		AdapterDir:  filepath.Join(td, "adapters"),
+		RuntimeFile: filepath.Join(td, "runtime.json"),
+		DefaultModel: "AUTO",
+		GatewayKeys: []string{"sk-admin"},
+		Keys: []config.GWKey{
+			{Key: "sk-admin", Role: "admin", Name: "admin"},
+			{Key: "sk-user", Role: "user", Name: "user"},
+		},
+		Sources: []config.Source{
+			{Name: "s1", BaseURL: "http://127.0.0.1:1/v1", Adapter: "openai",
+				Models: []config.Model{{ID: "gpt-4o", Priority: 10}, {ID: "gpt-4o-mini", Priority: 5}}},
+		},
+	}
+	if err := cfg.ApplyDefaults(); err != nil {
+		t.Fatal(err)
+	}
+	c, err := core.NewFromConfig(cfg)
+	if err != nil {
+		t.Fatalf("core: %v", err)
+	}
+	t.Cleanup(c.Close)
+	g, err := New(c)
+	if err != nil {
+		t.Fatalf("gateway: %v", err)
+	}
+	return g
+}
+
+// doReqAs issues an authenticated request as a specific key.
+func doReqAs(t *testing.T, g *Gateway, key, method, path, body string) *httptest.ResponseRecorder {
+	t.Helper()
+	req, _ := http.NewRequest(method, path, strings.NewReader(body))
+	req.Header.Set("Authorization", "Bearer "+key)
+	if body != "" {
+		req.Header.Set("Content-Type", "application/json")
+	}
+	rr := httptest.NewRecorder()
+	g.Handler().ServeHTTP(rr, req)
+	return rr
+}
+
+func TestKeyAutoAPIPermissions(t *testing.T) {
+	g := keyAutoGateway(t)
+
+	// A user may read their own chain...
+	if rr := doReqAs(t, g, "sk-user", "GET", "/api/keys/me/auto", ""); rr.Code != http.StatusOK {
+		t.Fatalf("GET /api/keys/me/auto as user = %d, want 200 (body %s)", rr.Code, rr.Body.String())
+	}
+	// ...but not another key's.
+	if rr := doReqAs(t, g, "sk-user", "GET", "/api/keys/sk-admin/auto", ""); rr.Code != http.StatusForbidden {
+		t.Fatalf("GET other key's auto as user = %d, want 403", rr.Code)
+	}
+	// ...and not their own via the admin route either.
+	if rr := doReqAs(t, g, "sk-user", "PUT", "/api/keys/sk-user/auto", `{"auto":[{"model":"gpt-4o"}]}`); rr.Code != http.StatusForbidden {
+		t.Fatalf("user PUT own auto = %d, want 403 (only admin configures)", rr.Code)
+	}
+}
+
+func TestKeyAutoAPICrudAndInherit(t *testing.T) {
+	g := keyAutoGateway(t)
+
+	// Initially inherits.
+	var got struct {
+		Inherits bool               `json:"inherits"`
+		Auto     []config.ModelScope `json:"auto"`
+	}
+	rr := doReqAs(t, g, "sk-admin", "GET", "/api/keys/sk-user/auto", "")
+	if rr.Code != http.StatusOK {
+		t.Fatalf("GET = %d: %s", rr.Code, rr.Body.String())
+	}
+	if err := json.Unmarshal(rr.Body.Bytes(), &got); err != nil {
+		t.Fatal(err)
+	}
+	if !got.Inherits || len(got.Auto) != 0 {
+		t.Fatalf("fresh key must inherit global, got inherits=%v auto=%v", got.Inherits, got.Auto)
+	}
+
+	// Admin sets a chain; the response reports the resolved slot count.
+	rr = doReqAs(t, g, "sk-admin", "PUT", "/api/keys/sk-user/auto",
+		`{"auto":[{"model":"gpt-4o-mini","source":"s1","tier":1}]}`)
+	if rr.Code != http.StatusOK {
+		t.Fatalf("PUT = %d: %s", rr.Code, rr.Body.String())
+	}
+	var put struct {
+		Inherits bool `json:"inherits"`
+		Slots    int  `json:"slots"`
+	}
+	if err := json.Unmarshal(rr.Body.Bytes(), &put); err != nil {
+		t.Fatal(err)
+	}
+	if put.Inherits {
+		t.Fatal("after a non-empty PUT the key must no longer inherit")
+	}
+	if put.Slots != 1 {
+		t.Fatalf("resolved slots = %d, want 1", put.Slots)
+	}
+
+	// And it reads back as an own chain.
+	rr = doReqAs(t, g, "sk-admin", "GET", "/api/keys/sk-user/auto", "")
+	if err := json.Unmarshal(rr.Body.Bytes(), &got); err != nil {
+		t.Fatal(err)
+	}
+	if got.Inherits || len(got.Auto) != 1 || got.Auto[0].Model != "gpt-4o-mini" {
+		t.Fatalf("own chain must read back, got inherits=%v auto=%v", got.Inherits, got.Auto)
+	}
+
+	// DELETE restores inheritance.
+	if rr := doReqAs(t, g, "sk-admin", "DELETE", "/api/keys/sk-user/auto", ""); rr.Code != http.StatusOK {
+		t.Fatalf("DELETE = %d: %s", rr.Code, rr.Body.String())
+	}
+	rr = doReqAs(t, g, "sk-admin", "GET", "/api/keys/sk-user/auto", "")
+	if err := json.Unmarshal(rr.Body.Bytes(), &got); err != nil {
+		t.Fatal(err)
+	}
+	if !got.Inherits {
+		t.Fatal("after DELETE the key must inherit the global chain again")
+	}
+}
+
+// An empty PUT must behave like DELETE. Otherwise a WebUI that submits an
+// empty list (every slot deleted in the editor) would persist an empty chain
+// and the next AUTO request would fail with no_provider.
+func TestKeyAutoAPIEmptyPutClearsOverride(t *testing.T) {
+	g := keyAutoGateway(t)
+
+	if rr := doReqAs(t, g, "sk-admin", "PUT", "/api/keys/sk-user/auto",
+		`{"auto":[{"model":"gpt-4o","source":"s1","tier":1}]}`); rr.Code != http.StatusOK {
+		t.Fatalf("seed PUT = %d: %s", rr.Code, rr.Body.String())
+	}
+	rr := doReqAs(t, g, "sk-admin", "PUT", "/api/keys/sk-user/auto", `{"auto":[]}`)
+	if rr.Code != http.StatusOK {
+		t.Fatalf("empty PUT = %d: %s", rr.Code, rr.Body.String())
+	}
+	var got struct {
+		Inherits bool `json:"inherits"`
+	}
+	if rr := doReqAs(t, g, "sk-admin", "GET", "/api/keys/sk-user/auto", ""); rr.Code != http.StatusOK {
+		t.Fatal(rr.Body.String())
+	}
+	if err := json.Unmarshal(rr.Body.Bytes(), &got); err != nil {
+		t.Fatal(err)
+	}
+	if !got.Inherits {
+		t.Fatal("an empty PUT must clear the override, not persist an empty chain")
+	}
+}
+
+// A chain naming a model that does not exist compiles to zero slots. The API
+// must say so at write time instead of letting the user's next request 503.
+func TestKeyAutoAPIReportsUnresolvableSlots(t *testing.T) {
+	g := keyAutoGateway(t)
+
+	rr := doReqAs(t, g, "sk-admin", "PUT", "/api/keys/sk-user/auto",
+		`{"auto":[{"model":"does-not-exist","source":"s1","tier":1}]}`)
+	if rr.Code != http.StatusOK {
+		t.Fatalf("PUT = %d: %s", rr.Code, rr.Body.String())
+	}
+	var put struct {
+		Inherits bool `json:"inherits"`
+		Slots    int  `json:"slots"`
+	}
+	if err := json.Unmarshal(rr.Body.Bytes(), &put); err != nil {
+		t.Fatal(err)
+	}
+	if put.Slots != 0 {
+		t.Fatalf("unknown model must resolve to 0 slots, got %d", put.Slots)
+	}
+}
+
+func TestKeyAutoAPIUnknownKey404(t *testing.T) {
+	g := keyAutoGateway(t)
+
+	if rr := doReqAs(t, g, "sk-admin", "GET", "/api/keys/sk-nope/auto", ""); rr.Code != http.StatusNotFound {
+		t.Fatalf("GET unknown key auto = %d, want 404", rr.Code)
+	}
+	if rr := doReqAs(t, g, "sk-admin", "PUT", "/api/keys/sk-nope/auto", `{"auto":[{"model":"gpt-4o"}]}`); rr.Code != http.StatusNotFound {
+		t.Fatalf("PUT unknown key auto = %d, want 404", rr.Code)
+	}
+}
+
+// A bad quota must be rejected with 400, not persisted and not silently
+// clamped — same rule as the model scope path.
+func TestKeyAutoAPIRejectsBadQuota(t *testing.T) {
+	g := keyAutoGateway(t)
+
+	rr := doReqAs(t, g, "sk-admin", "PUT", "/api/keys/sk-user/auto",
+		`{"auto":[{"model":"gpt-4o","source":"s1","tier":1,"token_quota":-1}]}`)
+	if rr.Code != http.StatusBadRequest {
+		t.Fatalf("negative quota PUT = %d, want 400 (body %s)", rr.Code, rr.Body.String())
+	}
+	// Nothing may have been persisted.
+	var got struct {
+		Inherits bool `json:"inherits"`
+	}
+	get := doReqAs(t, g, "sk-admin", "GET", "/api/keys/sk-user/auto", "")
+	if get.Code != http.StatusOK {
+		t.Fatal(get.Body.String())
+	}
+	if err := json.Unmarshal(get.Body.Bytes(), &got); err != nil {
+		t.Fatal(err)
+	}
+	if !got.Inherits {
+		t.Fatal("a rejected PUT must not persist a chain")
+	}
+}
+
+// The other key endpoints must keep working: the {key}/auto route must not
+// shadow /api/keys/{key} itself.
+func TestKeyAutoAPIDoesNotShadowKeyRoutes(t *testing.T) {
+	g := keyAutoGateway(t)
+
+	if rr := doReqAs(t, g, "sk-admin", "GET", "/api/keys", ""); rr.Code != http.StatusOK {
+		t.Fatalf("GET /api/keys = %d, want 200", rr.Code)
+	}
+	if rr := doReqAs(t, g, "sk-admin", "GET", "/api/keys/me", ""); rr.Code != http.StatusOK {
+		t.Fatalf("GET /api/keys/me = %d, want 200", rr.Code)
+	}
+	if rr := doReqAs(t, g, "sk-user", "GET", "/api/keys/me", ""); rr.Code != http.StatusOK {
+		t.Fatalf("GET /api/keys/me as user = %d, want 200", rr.Code)
+	}
+}
--- a/internal/gateway/keys.go
+++ b/internal/gateway/keys.go
@ -7,6 +7,7 @@ import (
 	"strings"

 	"llmsproxy/internal/config"
+	"llmsproxy/internal/scheduler"
 )

 // handleKeysAPI manages gateway keys: GET /api/keys (admin: all keys),
@ -17,10 +18,22 @@ func (g *Gateway) handleKeysAPI(w http.ResponseWriter, r *http.Request) {
 	path = strings.Trim(path, "/")
 	role := reqRole(r.Context())

+	// /api/keys/me/auto — a user's own chain, readable without admin rights.
+	// Must be matched before the generic {key}/auto case below so it does not
+	// ask admin rights of a plain user asking about their own chain.
+	if path == "me/auto" {
+		g.handleKeyMeAuto(w, r)
+		return
+	}
 	if path == "me" {
 		g.handleKeyMe(w, r)
 		return
 	}
+	// /api/keys/{key}/auto — the per-key AUTO chain editor.
+	if i := strings.LastIndex(path, "/auto"); i > 0 && path[i+len("/auto"):] == "" {
+		g.handleKeyAutoAPI(w, r, path[:i])
+		return
+	}
 	if role != "admin" {
 		writeError(w, http.StatusForbidden, "forbidden", "admin role required")
 		return
@ -139,6 +152,101 @@ func (g *Gateway) allowedModels(ctx context.Context) []config.ModelScope {
 	return rec.Models
 }

+// handleKeyAutoAPI manages one key's own AUTO chain:
+// GET /api/keys/{key}/auto returns it, PUT replaces it, DELETE clears the
+// override so the key falls back to the global chain. Admin only — a user can
+// inspect their own chain through GET /api/keys/me/auto.
+//
+// An empty chain and a cleared chain are deliberately the same state: "inherit
+// the global chain". That keeps the WebUI's "use global" toggle a plain DELETE
+// with no sentinel value to carry through config.yaml.
+func (g *Gateway) handleKeyAutoAPI(w http.ResponseWriter, r *http.Request, key string) {
+	if reqRole(r.Context()) != "admin" {
+		writeError(w, http.StatusForbidden, "forbidden", "admin role required")
+		return
+	}
+	rec, ok := g.core.FindKey(key)
+	if !ok {
+		writeError(w, http.StatusNotFound, "not_found", "key not found")
+		return
+	}
+	switch r.Method {
+	case http.MethodGet:
+		writeJSON(w, http.StatusOK, map[string]interface{}{
+			"key":      key,
+			"auto":     rec.Auto,
+			"inherits": !rec.HasOwnAuto(),
+		})
+	case http.MethodPut, http.MethodPatch:
+		var body struct {
+			Auto []config.ModelScope `json:"auto"`
+		}
+		if err := json.NewDecoder(r.Body).Decode(&body); err != nil {
+			writeError(w, http.StatusBadRequest, "invalid_request", "invalid json: "+err.Error())
+			return
+		}
+		if len(body.Auto) == 0 {
+			// Treat an empty PUT as "give up the override" so the endpoint
+			// cannot persist a chain that would 503 with no slots.
+			if err := g.core.SaveKeyAuto(key, nil); err != nil {
+				writeError(w, http.StatusBadRequest, "key_error", err.Error())
+				return
+			}
+			writeJSON(w, http.StatusOK, map[string]interface{}{"ok": true, "inherits": true})
+			return
+		}
+		if err := g.core.SaveKeyAuto(key, body.Auto); err != nil {
+			writeError(w, http.StatusBadRequest, "key_error", err.Error())
+			return
+		}
+		// Report whether the chain actually resolved to usable slots. An
+		// admin who typed a model that no longer exists should learn it here,
+		// not from the user's next 503.
+		chain, _ := g.core.AutoChainFor(key)
+		writeJSON(w, http.StatusOK, map[string]interface{}{
+			"ok": true, "inherits": false, "slots": chainSlots(chain),
+		})
+	case http.MethodDelete:
+		if err := g.core.SaveKeyAuto(key, nil); err != nil {
+			writeError(w, http.StatusBadRequest, "key_error", err.Error())
+			return
+		}
+		writeJSON(w, http.StatusOK, map[string]interface{}{"ok": true, "inherits": true})
+	default:
+		writeError(w, http.StatusMethodNotAllowed, "method_not_allowed", "")
+	}
+}
+
+// chainSlots counts the usable slots in a chain, for API feedback after an
+// edit. A chain that compiled to zero slots is reported so the caller can warn
+// instead of letting the next request fail with no_provider.
+func chainSlots(ch *scheduler.Chain) int {
+	if ch == nil {
+		return 0
+	}
+	n := 0
+	for _, tn := range ch.Tiers {
+		n += len(tn.Slots)
+	}
+	return n
+}
+
+// handleKeyMeAuto lets a user read their own AUTO chain without admin rights.
+func (g *Gateway) handleKeyMeAuto(w http.ResponseWriter, r *http.Request) {
+	if r.Method != http.MethodGet {
+		writeError(w, http.StatusMethodNotAllowed, "method_not_allowed", "use GET")
+		return
+	}
+	rec, ok := g.core.FindKey(reqKey(r.Context()))
+	if !ok {
+		writeError(w, http.StatusUnauthorized, "invalid_api_key", "key not found")
+		return
+	}
+	writeJSON(w, http.StatusOK, map[string]interface{}{
+		"key": rec.Key, "auto": rec.Auto, "inherits": !rec.HasOwnAuto(),
+	})
+}
+
 // handleAutoAPI manages the AUTO scheduling slots: GET /api/auto returns the
 // current rules; PUT /api/auto replaces them (admin only).
 func (g *Gateway) handleAutoAPI(w http.ResponseWriter, r *http.Request) {
--- a/internal/gateway/ui/index.html
+++ b/internal/gateway/ui/index.html
@ -822,6 +822,17 @@
          keysHint:
            "管理员密钥可查看与管理全部密钥，并可为每个用户密钥配置可用模型范围；用户密钥只能看到自己。",
          kCreate: "新建密钥",
+          kAutoChain: "AUTO 链",
+          kAutoChainOwn: "独立链",
+          kAutoChainGlobal: "跟随全局",
+          kAutoChainTitle: "独立 AUTO 链",
+          kAutoChainHint: "为空则跟随全局 AUTO 链",
+          kAutoChainInherit: "使用全局链",
+          kAutoChainAdd: "+ 添加槽位",
+          kAutoChainSave: "保存",
+          kAutoChainTier: "层级",
+          kAutoChainEmpty: "该密钥使用全局 AUTO 链",
+          kAutoChainBad: "槽位无法解析（模型或源不存在），保存后 AUTO 请求会失败",
          kName: "名称",
          kRole: "角色",
          kRoleAdmin: "管理员",
@ -1081,6 +1092,17 @@
          keysHint:
            "Admin keys can view and manage every key and configure each user key\u0027s allowed models; user keys only see themselves.",
          kCreate: "Create key",
+          kAutoChain: "AUTO chain",
+          kAutoChainOwn: "own",
+          kAutoChainGlobal: "global",
+          kAutoChainTitle: "Per-key AUTO chain",
+          kAutoChainHint: "Leave empty to inherit the global AUTO chain",
+          kAutoChainInherit: "Use global chain",
+          kAutoChainAdd: "+ Add slot",
+          kAutoChainSave: "Save",
+          kAutoChainTier: "Tier",
+          kAutoChainEmpty: "This key uses the global AUTO chain",
+          kAutoChainBad: "Slot does not resolve (unknown model or source); AUTO requests would fail",
          kName: "Name",
          kRole: "Role",
          kRoleAdmin: "Admin",
@ -4674,6 +4696,8 @@
      <button class="ghost small" onclick="copyText('${escAttr(k.key)}')">${t("kCopy")}</button>
      <span class="grow"></span>
      <span class="muted">${fmtCreated(k.created_at)}</span>
+      <button class="ghost small" title="${escAttr(t("kAutoChainTitle"))}"
+        onclick="openKeyAutoModal('${escAttr(k.key)}','${escAttr(k.name || "")}')">${t("kAutoChain")}</button>
      <button class="ghost small errc" onclick="delKey('${escAttr(k.key)}','${escAttr(k.name || "")}')">${t("kDel")}</button>
      <button class="ghost small errc" title="${escAttr(t("kDel"))}" onclick="delKey('${escAttr(k.key)}','${escAttr(k.name || "")}')">×</button>
    </div>
@ -5085,6 +5109,108 @@
        document.body.appendChild(wrap);
        $("#kc-name").focus();
      }
+      // Per-key AUTO chain editor. Kept deliberately simple: a plain table of
+      // slots rather than a drag canvas, because an admin usually either
+      // mirrors the global chain or deletes a few tiers, and the canvas layout
+      // for the model scope list would be overkill here.
+      async function openKeyAutoModal(key, name) {
+        const wrap = document.createElement("div");
+        wrap.id = "modal-wrap";
+        wrap.style.cssText =
+          "position:fixed;inset:0;background:rgba(15,22,44,.45);display:flex;align-items:flex-start;justify-content:center;overflow:auto;padding:48px 20px;z-index:50";
+        wrap.innerHTML = `<div class="card" style="width:720px;max-width:100%">
+    <h2>${esc(t("kAutoChainTitle"))} · ${esc(name || key)}</h2>
+    <div class="muted" style="margin-bottom:10px">${esc(t("kAutoChainHint"))}</div>
+    <div id="ka-rows"></div>
+    <p><button class="ghost" onclick="keyAutoAddRow()">${esc(t("kAutoChainAdd"))}</button>
+       <button class="ghost" onclick="keyAutoClear(this)">${esc(t("kAutoChainInherit"))}</button></p>
+    <p><button onclick="keyAutoSave(this,'${escAttr(key)}')">${esc(t("kAutoChainSave"))}</button>
+       <button class="ghost" onclick="this.closest('#modal-wrap').remove()">${esc(t("mCancel"))}</button></p>
+  </div>`;
+        document.body.appendChild(wrap);
+        try {
+          const j = await api("/api/keys/" + encodeURIComponent(key) + "/auto");
+          const rows = $("#ka-rows");
+          if (!j.auto || !j.auto.length) {
+            rows.innerHTML = `<div class="muted" style="padding:8px 0">${esc(
+              t("kAutoChainEmpty")
+            )}</div>`;
+          } else {
+            j.auto.forEach((r) => keyAutoRow(r));
+          }
+        } catch (e) {
+          toast(String(e));
+        }
+      }
+      function keyAutoRow(r) {
+        const el = document.createElement("div");
+        el.className = "ka-row";
+        el.style.cssText =
+          "display:flex;gap:8px;align-items:center;margin-bottom:8px;flex-wrap:wrap";
+        const opts = (sel) =>
+          (allModels || [])
+            .map((p) => {
+              // loadModelPairs builds {key,label}; the /api/status fallback
+              // only guarantees an id, so derive a comb key from whatever is
+              // present rather than emitting an empty option value.
+              const val = p.key || (p.id ? p.id + (p.source ? "|" + p.source : "") : "");
+              const lab = p.label || p.id || val;
+              return `<option value="${escAttr(val)}"${
+                val === sel ? " selected" : ""
+              }>${esc(lab)}</option>`;
+            })
+            .join("");
+        el.innerHTML = `<select class="ka-model" style="flex:2 1 220px;min-width:0">${opts(
+          r ? scopeComb(r) : ""
+        )}</select>
+    <label class="muted" style="flex:0 0 auto">${esc(
+          t("kAutoChainTier")
+        )}<input class="ka-tier" type="number" min="1" value="${
+          (r && r.tier) || 1
+        }" style="width:64px;margin-left:6px"></label>
+    <button class="ghost small errc" onclick="this.closest('.ka-row').remove()">×</button>`;
+        $("#ka-rows").appendChild(el);
+      }
+      function keyAutoAddRow() {
+        keyAutoRow(null);
+      }
+      function keyAutoClear() {
+        $("#ka-rows").innerHTML = "";
+      }
+      async function keyAutoSave(btn, key) {
+        const rules = [];
+        document.querySelectorAll("#ka-rows .ka-row").forEach((row) => {
+          const comb = row.querySelector(".ka-model").value;
+          if (!comb) return;
+          const { model, source } = splitCombKey(comb);
+          const tier = parseInt(row.querySelector(".ka-tier").value, 10) || 1;
+          rules.push({ model, source, tier });
+        });
+        btn.disabled = true;
+        try {
+          // An empty rule set means "inherit global" server-side, so one PUT
+          // covers both clearing the override and saving a chain.
+          const j = await api("/api/keys/" + encodeURIComponent(key) + "/auto", {
+            method: "PUT",
+            headers: { "Content-Type": "application/json" },
+            body: JSON.stringify({ auto: rules }),
+          });
+          if (j.inherits) {
+            toast(t("kAutoChainGlobal"));
+          } else if (!j.slots) {
+            // Persisted, but nothing resolves — say so instead of letting the
+            // user discover it as a 503 on their next AUTO request.
+            toast(t("kAutoChainBad"));
+          } else {
+            toast(t("kAutoChainSave") + " (" + j.slots + ")");
+          }
+          document.getElementById("modal-wrap").remove();
+          loadKeys();
+        } catch (e) {
+          toast(String(e));
+          btn.disabled = false;
+        }
+      }
      async function createKey(btn) {
        const name = $("#kc-name").value.trim();
        if (!name) {
--- a/internal/gateway/ui_key_auto_test.go
+++ b/internal/gateway/ui_key_auto_test.go
@ -0,0 +1,78 @@
+package gateway
+
+import (
+	"strings"
+	"testing"
+)
+
+// Per-key AUTO chain UI contract. The feature spans three layers (API handler,
+// config field, WebUI) and the WebUI part is plain inline JS, so a rename or a
+// dropped button fails silently at runtime: the admin simply finds no way to
+// configure a key's chain and has no error to chase.
+//
+// These checks pin the wiring that is easy to break by refactor.
+
+func TestUIPerKeyAutoWiring(t *testing.T) {
+	src := uiSource(t)
+
+	// The admin key list must offer the editor, otherwise the whole feature is
+	// unreachable from the UI. Assert the *call site* inside keyCanvasHtml,
+	// not just that the functions exist: removing the button leaves every
+	// definition intact, and that is exactly the regression to catch.
+	kcStart := strings.Index(src, "function keyCanvasHtml")
+	if kcStart < 0 {
+		t.Fatal("keyCanvasHtml is gone — cannot check the AUTO chain button")
+	}
+	kcEnd := strings.Index(src[kcStart:], "\n      }")
+	if kcEnd < 0 {
+		t.Fatal("could not delimit keyCanvasHtml")
+	}
+	kcBody := src[kcStart : kcStart+kcEnd]
+	if !strings.Contains(kcBody, "openKeyAutoModal(") {
+		t.Fatal("per-key AUTO chain button is missing from the key card header")
+	}
+	for _, fn := range []string{
+		"function openKeyAutoModal",
+		"function keyAutoRow",
+		"function keyAutoAddRow",
+		"function keyAutoClear",
+		"function keyAutoSave",
+	} {
+		if !strings.Contains(src, fn) {
+			t.Fatalf("UI is missing %s — the editor cannot function", fn)
+		}
+	}
+
+	// The editor must talk to the endpoint the gateway actually serves.
+	if !strings.Contains(src, "/auto\"") {
+		t.Fatal("UI never calls the /api/keys/{key}/auto endpoint")
+	}
+}
+
+// The editor's empty-row path sends {"auto":[]} and relies on the server
+// treating that as "clear the override". If the UI instead sent a sentinel,
+// a future server change would silently persist a broken empty chain.
+func TestUIPerKeyAutoSendsEmptyListToClear(t *testing.T) {
+	src := uiSource(t)
+	if !strings.Contains(src, "JSON.stringify({ auto: rules })") {
+		t.Fatal("keyAutoSave must PUT the collected rule list verbatim")
+	}
+	if !strings.Contains(src, "keyAutoClear") {
+		t.Fatal("no way to clear a key's own chain from the UI")
+	}
+}
+
+// Both languages need the strings, otherwise a non-Chinese admin sees raw
+// i18n keys in the modal.
+func TestUIPerKeyAutoHasBothLanguages(t *testing.T) {
+	src := uiSource(t)
+	for _, k := range []string{
+		"kAutoChainTitle", "kAutoChainHint", "kAutoChainInherit",
+		"kAutoChainAdd", "kAutoChainSave", "kAutoChainEmpty",
+	} {
+		n := strings.Count(src, k+`: "`)
+		if n < 2 {
+			t.Fatalf("i18n key %s defined %d time(s), want both zh and en", k, n)
+		}
+	}
+}