diff --git a/internal/gateway/api.go b/internal/gateway/api.go
index f7996d9..4dc75f0 100644
--- a/internal/gateway/api.go
+++ b/internal/gateway/api.go
@@ -470,8 +470,8 @@ func (g *Gateway) handleStatsAPI(w http.ResponseWriter, r *http.Request) {
 			// 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,
+			"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
 			// stuck at one screen forever.
diff --git a/internal/gateway/discover.go b/internal/gateway/discover.go
index 3c5db64..7a5efa4 100644
--- a/internal/gateway/discover.go
+++ b/internal/gateway/discover.go
@@ -535,11 +535,11 @@ type modelCaps struct {
 }
 
 type upstreamModel struct {
-	ID       string `json:"id"`
-	Object   string `json:"object"`
-	Kind     string `json:"kind"`
-	Type     string `json:"type"`
-	Modality string `json:"modality"`
+	ID           string `json:"id"`
+	Object       string `json:"object"`
+	Kind         string `json:"kind"`
+	Type         string `json:"type"`
+	Modality     string `json:"modality"`
 	Architecture struct {
 		InputModalities  []string `json:"input_modalities"`
 		OutputModalities []string `json:"output_modalities"`
@@ -699,8 +699,8 @@ func modelMeta(m upstreamModel) map[string]interface{} {
 		}
 	}
 
-	copyFields(m.Metadata)      // nested first
-	copyFields(mustJSON(m))     // top level wins
+	copyFields(m.Metadata)  // nested first
+	copyFields(mustJSON(m)) // top level wins
 	if len(meta) == 0 {
 		return nil
 	}
diff --git a/internal/gateway/discover_api.go b/internal/gateway/discover_api.go
index 0840954..257dc77 100644
--- a/internal/gateway/discover_api.go
+++ b/internal/gateway/discover_api.go
@@ -5,10 +5,10 @@ import (
 	"encoding/json"
 	"errors"
 	"io"
+	"llmsproxy/internal/config"
 	"net/http"
 	"strings"
 	"time"
-	"llmsproxy/internal/config"
 )
 
 // HTTP surface for model discovery.
@@ -68,9 +68,9 @@ type discoverResponse struct {
 	OK bool `json:"ok"`
 
 	// Normalized source fields, ready to be echoed back to /api/sources.
-	BaseURL string `json:"base_url"`
-	Adapter string `json:"adapter"`
-	Models  []string `json:"models"`
+	BaseURL     string   `json:"base_url"`
+	Adapter     string   `json:"adapter"`
+	Models      []string `json:"models"`
 	ImageModels []string `json:"image_models"`
 
 	// Endpoint is set only when discovery had to move a version segment.
@@ -152,18 +152,18 @@ func (g *Gateway) handleSourceDiscoverAPI(w http.ResponseWriter, r *http.Request
 	}
 
 	out := &discoverResponse{
-		OK:          true,
-		BaseURL:     res.BaseURL,
-		Adapter:     req.Adapter,
-		Models:      chat,
-		ImageModels: image,
+		OK:            true,
+		BaseURL:       res.BaseURL,
+		Adapter:       req.Adapter,
+		Models:        chat,
+		ImageModels:   image,
 		Endpoint:      res.Endpoint,
 		ImageEndpoint: res.ImageEndpoint,
 		ModelsPath:    res.ModelsPath,
-		Tried:       res.Tried,
-		LatencyMS:   res.LatencyMS,
-		Note:        res.Note,
-		Warnings:    warns,
+		Tried:         res.Tried,
+		LatencyMS:     res.LatencyMS,
+		Note:          res.Note,
+		Warnings:      warns,
 	}
 	if len(chat) == 0 && len(image) == 0 {
 		out.Warnings = append(out.Warnings,
@@ -207,14 +207,14 @@ func (g *Gateway) handleSourceDiscoverAPI(w http.ResponseWriter, r *http.Request
 	// keep the stored one" rule, and going through a second decoder here would
 	// duplicate (and eventually diverge from) that logic.
 	src := config.Source{
-		Name:        name,
-		BaseURL:     res.BaseURL,
-		APIKey:      g.resolveAPIKey(name, req.APIKey, exists),
-		Adapter:     req.Adapter,
+		Name:          name,
+		BaseURL:       res.BaseURL,
+		APIKey:        g.resolveAPIKey(name, req.APIKey, exists),
+		Adapter:       req.Adapter,
 		Endpoint:      res.Endpoint,
 		ImageEndpoint: res.ImageEndpoint,
 		Models:        modelsToConfig(chat, image),
-		Temperature: req.Temperature,
+		Temperature:   req.Temperature,
 	}
 	if req.MaxConcurrent > 0 {
 		src.MaxConcurrent = req.MaxConcurrent
diff --git a/internal/gateway/discover_api_test.go b/internal/gateway/discover_api_test.go
index 1fa6983..71d6f03 100644
--- a/internal/gateway/discover_api_test.go
+++ b/internal/gateway/discover_api_test.go
@@ -32,12 +32,12 @@ func catalogUpstream(t *testing.T) *httptest.Server {
 
 func TestDefaultSourceName(t *testing.T) {
 	cases := map[string]string{
-		"https://api.vendor.com/v1":   "vendor.com",
-		"https://api.openai.com/v1":   "openai.com",
-		"https://www.vendor.com/v1":   "vendor.com",
-		"http://127.0.0.1:8080/v1":    "127.0.0.1",
+		"https://api.vendor.com/v1":          "vendor.com",
+		"https://api.openai.com/v1":          "openai.com",
+		"https://www.vendor.com/v1":          "vendor.com",
+		"http://127.0.0.1:8080/v1":           "127.0.0.1",
 		"https://amr-link.open-design.ai/v1": "amr-link.open-design.ai",
-		"not a url at all":             "",
+		"not a url at all":                   "",
 	}
 	for in, want := range cases {
 		if got := defaultSourceName(in); got != want {
@@ -122,16 +122,16 @@ func TestSelectModels_WarnsThatImagesAreNotInAUTO(t *testing.T) {
 // operator previews one set of values and saves another.
 func TestDiscoveryResultAndConfigRoundTrip(t *testing.T) {
 	res := &discoveryResult{
-		BaseURL:    "https://h",
-		Endpoint:   "/v1/chat/completions",
-		Models:     []string{"m1", "m2"},
+		BaseURL:     "https://h",
+		Endpoint:    "/v1/chat/completions",
+		Models:      []string{"m1", "m2"},
 		ImageModels: []string{"img1"},
 	}
 	src := config.Source{
-		Name:      "x",
-		BaseURL:   res.BaseURL,
-		Endpoint:  res.Endpoint,
-		Models:    modelsToConfig(res.Models, res.ImageModels),
+		Name:     "x",
+		BaseURL:  res.BaseURL,
+		Endpoint: res.Endpoint,
+		Models:   modelsToConfig(res.Models, res.ImageModels),
 	}
 	if len(src.Models) != 3 {
 		t.Fatalf("models = %d, want 3", len(src.Models))
diff --git a/internal/gateway/discover_endpoint_test.go b/internal/gateway/discover_endpoint_test.go
index 8865921..4801b6a 100644
--- a/internal/gateway/discover_endpoint_test.go
+++ b/internal/gateway/discover_endpoint_test.go
@@ -31,10 +31,10 @@ func TestComposeImageEndpoint_NoVersionBaseLeavesDefault(t *testing.T) {
 // TestDiscovery_BareHostNeedsVersionOnEndpointWhenCatalogIsVersioned
 // distinguishes two cases that look identical from the outside.
 //
-//   bare host + catalog at "/models"       → chat is "/chat/completions" (the
-//     adapter default, no endpoint written)
-//   bare host + catalog at "/v1/models"    → chat is "/v1/chat/completions",
-//     so an endpoint MUST be written
+//	bare host + catalog at "/models"       → chat is "/chat/completions" (the
+//	  adapter default, no endpoint written)
+//	bare host + catalog at "/v1/models"    → chat is "/v1/chat/completions",
+//	  so an endpoint MUST be written
 //
 // The catalog path is the only evidence of whether the upstream wants the
 // version segment, and getting this backwards sends every chat request to a
@@ -105,15 +105,15 @@ func TestDiscovery_BareHostNeedsVersionOnEndpointWhenCatalogIsVersioned(t *testi
 // silent, expensive mistakes, so they are asserted literally.
 func TestComposeEndpoints_BothFormsProduceWorkingURLs(t *testing.T) {
 	cases := []struct {
-		name        string
-		base        string
-		wantChat    string
-		wantImage   string
+		name      string
+		base      string
+		wantChat  string
+		wantImage string
 	}{
 		{
-			name:      "bare host: adapter and image defaults compose",
-			base:      "https://h",
-			wantChat:  "https://h/chat/completions",
+			name:     "bare host: adapter and image defaults compose",
+			base:     "https://h",
+			wantChat: "https://h/chat/completions",
 			// openai's adapter endpoint is /chat/completions (verified in
 			// internal/lua/adapters/openai.lua: adapter.endpoint).
 			wantImage: "https://h/v1/images/generations",
diff --git a/internal/gateway/discover_test.go b/internal/gateway/discover_test.go
index d465b43..2d18d0d 100644
--- a/internal/gateway/discover_test.go
+++ b/internal/gateway/discover_test.go
@@ -356,10 +356,10 @@ func TestDiscoveryCandidates_DedupesAndPrefersTyped(t *testing.T) {
 
 func TestVersionSegment(t *testing.T) {
 	cases := map[string]string{
-		"/v1/models":   "/v1",
+		"/v1/models":     "/v1",
 		"/v1beta/models": "/v1beta",
-		"/models":      "",
-		"/api/models":  "",
+		"/models":        "",
+		"/api/models":    "",
 		"/vision/models": "",
 	}
 	for in, want := range cases {
@@ -371,19 +371,19 @@ func TestVersionSegment(t *testing.T) {
 
 func TestTrimVersionSuffix(t *testing.T) {
 	cases := map[string]string{
-		"https://h/v1":          "https://h",
-		"https://h/v2":          "https://h",
-		"https://h/v1beta":      "https://h",
-		"https://h/api/v1":      "https://h/api",
-		"https://h/vision":      "https://h/vision", // not numeric — a real path segment
-		"https://h":             "https://h",
-		"https://h/openai":      "https://h/openai",
-		"https://h/v1/extra":    "https://h/v1/extra",
+		"https://h/v1":       "https://h",
+		"https://h/v2":       "https://h",
+		"https://h/v1beta":   "https://h",
+		"https://h/api/v1":   "https://h/api",
+		"https://h/vision":   "https://h/vision", // not numeric — a real path segment
+		"https://h":          "https://h",
+		"https://h/openai":   "https://h/openai",
+		"https://h/v1/extra": "https://h/v1/extra",
 		// A segment that merely STARTS with v must not be stripped: stripping
 		// it would rewrite base "https://h/vision" to "https://h" and, worse,
 		// would treat a real vendor path as a version.
-		"https://h/vertex":      "https://h/vertex",
-		"https://h/volcengine":  "https://h/volcengine",
+		"https://h/vertex":       "https://h/vertex",
+		"https://h/volcengine":   "https://h/volcengine",
 		"https://h/vertex-ai/v1": "https://h/vertex-ai",
 	}
 	for in, want := range cases {
@@ -463,7 +463,7 @@ func TestComposeEndpoint_ExactlyOneVersionSegment(t *testing.T) {
 			// Whichever way it resolved, base + endpoint must never carry the
 			// same version segment twice.
 			onBase := trimVersionSuffix(tc.persistedBase) != tc.persistedBase
-			onEndpoint := versionSegment(strings.TrimSuffix(got, "/chat/completions") + "/models") != ""
+			onEndpoint := versionSegment(strings.TrimSuffix(got, "/chat/completions")+"/models") != ""
 			if onBase && onEndpoint {
 				t.Errorf("version segment duplicated: base %q + endpoint %q", tc.persistedBase, got)
 			}
diff --git a/internal/gateway/key_auto_api_test.go b/internal/gateway/key_auto_api_test.go
index c2ab771..7bb62cd 100644
--- a/internal/gateway/key_auto_api_test.go
+++ b/internal/gateway/key_auto_api_test.go
@@ -4,8 +4,8 @@ import (
 	"encoding/json"
 	"net/http"
 	"net/http/httptest"
-	"path/filepath"
 	"os"
+	"path/filepath"
 	"strings"
 	"testing"
 
@@ -33,11 +33,11 @@ func keyAutoGateway(t *testing.T) *Gateway {
 		t.Fatal(err)
 	}
 	cfg := &config.Config{
-		Path:        cfgPath,
-		AdapterDir:  filepath.Join(td, "adapters"),
-		RuntimeFile: filepath.Join(td, "runtime.json"),
+		Path:         cfgPath,
+		AdapterDir:   filepath.Join(td, "adapters"),
+		RuntimeFile:  filepath.Join(td, "runtime.json"),
 		DefaultModel: "AUTO",
-		GatewayKeys: []string{"sk-admin"},
+		GatewayKeys:  []string{"sk-admin"},
 		Keys: []config.GWKey{
 			{Key: "sk-admin", Role: "admin", Name: "admin"},
 			{Key: "sk-user", Role: "user", Name: "user"},
@@ -97,7 +97,7 @@ func TestKeyAutoAPICrudAndInherit(t *testing.T) {
 
 	// Initially inherits.
 	var got struct {
-		Inherits bool               `json:"inherits"`
+		Inherits bool                `json:"inherits"`
 		Auto     []config.ModelScope `json:"auto"`
 	}
 	rr := doReqAs(t, g, "sk-admin", "GET", "/api/keys/sk-user/auto", "")
@@ -254,4 +254,4 @@ func TestKeyAutoAPIDoesNotShadowKeyRoutes(t *testing.T) {
 	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)
 	}
-}
\ No newline at end of file
+}
diff --git a/internal/gateway/messages.go b/internal/gateway/messages.go
new file mode 100644
index 0000000..0201831
--- /dev/null
+++ b/internal/gateway/messages.go
@@ -0,0 +1,322 @@
+package gateway
+
+import (
+	"context"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"net/http"
+	"strings"
+	"time"
+
+	"llmsproxy/internal/provider"
+	"llmsproxy/internal/scheduler"
+	"llmsproxy/internal/types"
+)
+
+// handleMessages serves the Anthropic Messages protocol (POST /v1/messages).
+//
+// The point of this endpoint is NOT to add a second scheduling path. It reuses
+// everything /v1/chat/completions does — key auth, model scope, per-model
+// quotas, the AUTO chain, cooldowns, audit records, plugin hooks — and differs
+// in exactly two places:
+//
+//  1. the request body is parsed from Anthropic's shape into the unified format
+//     (messages with content blocks, a top-level `system`, `max_tokens`
+//     mandatory, `thinking` optional);
+//  2. the response is rendered by the SERVING adapter's serialize_* hooks, so
+//     the client sees Anthropic's shape back.
+//
+// The kernel therefore contains no Anthropic field names on this path: parsing
+// is table-driven, and serialization belongs to the adapter.
+//
+// A source whose adapter does not declare serialize_* falls back to the
+// OpenAI shape, which is what the kernel guarantees for an unconfigured
+// deployment. That asymmetry is deliberate (see serializer.go): an Anthropic
+// client talking to a source that cannot speak Anthropic gets a clear
+// OpenAI-shaped answer rather than a malformed Anthropic one.
+func (g *Gateway) handleMessages(w http.ResponseWriter, r *http.Request) {
+	if r.Method != http.MethodPost {
+		writeError(w, http.StatusMethodNotAllowed, "method_not_allowed", "use POST")
+		return
+	}
+	var raw anthropicRequest
+	if err := json.NewDecoder(r.Body).Decode(&raw); err != nil {
+		writeError(w, http.StatusBadRequest, "invalid_request", "invalid json: "+err.Error())
+		return
+	}
+	req, err := raw.toUnified(g.core.DefaultModel())
+	if err != nil {
+		writeError(w, http.StatusBadRequest, "invalid_request", err.Error())
+		return
+	}
+
+	model := req.Model
+	ctx := r.Context()
+	g.fireStart(ctx, &chatRequest{Messages: req.Messages, Tools: nil},
+		chatKind(req.Stream), model, len(req.Messages), 0)
+
+	if isAuto(model) {
+		chain, _ := g.core.AutoChainFor(reqKey(ctx))
+		if chain == nil || len(chain.Tiers) == 0 {
+			writeError(w, http.StatusServiceUnavailable, "no_provider", "no auto slot configured")
+			return
+		}
+		if q := g.checkQuota(ctx, "AUTO"); q != nil {
+			g.writeReject(w, q)
+			return
+		}
+		g.serveUnified(ctx, w, chain, req, nil)
+		return
+	}
+	if allow := g.allowedModels(ctx); allow != nil && !g.hasScopeModel(allow, model) {
+		writeError(w, http.StatusForbidden, "model_not_allowed",
+			"model \""+model+"\" is not allowed for this key")
+		return
+	}
+	cands, effective := g.resolveCands(ctx, &chatRequest{Model: model})
+	if len(cands) == 0 {
+		writeError(w, http.StatusNotFound, "model_not_found",
+			"model \""+model+"\" is not configured")
+		return
+	}
+	if effective == "" {
+		effective = firstModel(cands[0])
+	}
+	if q := g.checkQuota(ctx, effective); q != nil {
+		g.writeReject(w, q)
+		return
+	}
+	g.serveUnified(ctx, w, nil, req, &directRoute{cands: cands, effective: effective})
+}
+
+// directRoute carries the pre-resolved candidate list for a model-pinned
+// request, so serveUnified can run both routes (direct / AUTO) from one body.
+type directRoute struct {
+	cands     []*provider.Provider
+	effective string
+}
+
+// serveUnified runs the request through the scheduler and writes the reply via
+// the serving adapter's serializer. It is the shared tail of handleChat and
+// handleMessages: by this point every gate (auth, scope, quota) has already run,
+// so the only remaining questions are "which source serves it" and "what shape
+// does the client get".
+func (g *Gateway) serveUnified(ctx context.Context, w http.ResponseWriter, chain *scheduler.Chain, req *types.ChatRequest, direct *directRoute) {
+	rec := &Req{Key: keyID(reqKey(ctx)), Type: "chat", OK: false}
+	if direct != nil {
+		rec.Model = direct.effective
+		rec.Source = firstSource(direct.cands)
+	} else {
+		rec.Type = "stream"
+	}
+	if req.Stream {
+		rec.Type = "stream"
+	}
+
+	done := g.stats.Begin()
+	defer done()
+
+	var (
+		resp               *types.UnifiedResponse
+		chunks             <-chan types.UnifiedChunk
+		usedSrc, usedModel string
+		err                error
+	)
+	t0 := time.Now()
+	if direct != nil {
+		cands := toScheduler(direct.cands)
+		if req.Stream {
+			chunks, usedSrc, usedModel, err = g.core.Scheduler().ChatStream(ctx, cands, req)
+		} else {
+			resp, usedSrc, usedModel, err = g.core.Scheduler().Chat(ctx, cands, req)
+		}
+	} else {
+		if req.Stream {
+			chunks, usedSrc, usedModel, err = g.core.Scheduler().ChainChatStream(ctx, chain, req, nil, nil)
+		} else {
+			resp, usedSrc, usedModel, err = g.core.Scheduler().ChainChat(ctx, chain, req, nil, nil)
+		}
+	}
+	rec.LatMs = time.Since(t0).Milliseconds()
+	if err != nil {
+		g.failChat(w, rec, err)
+		g.writeRec(rec)
+		return
+	}
+	rec.OK = true
+	rec.Status = http.StatusOK
+	if usedModel != "" {
+		rec.Model = usedModel
+	}
+	rec.Source = usedSrc
+	ser := g.serializerFor(usedSrc, directProviderList(direct))
+
+	if chunks != nil {
+		g.fireRouted(ctx, "stream", usedSrc, usedModel, tierForRoute(direct), true)
+		rec.Prompt = estimatePromptTokens(req)
+		g.pumpStream(w, ser, rec, chunks, modelLabel(req, usedModel), t0)
+		return
+	}
+	recordChatUsage(rec, req, resp)
+	g.fireRouted(ctx, "chat", usedSrc, usedModel, tierForRoute(direct), false)
+	rec.FirstByteMs = rec.LatMs
+	g.writeRec(rec)
+	writeChatCompletion(w, ser, resp, modelLabel(req, usedModel))
+}
+
+// ---- Anthropic request shape (in-bound only; the kernel holds no outbound
+// Anthropic field names) ----
+
+// anthropicRequest is the client-facing parse target for POST /v1/messages.
+//
+// Only the fields the gateway must understand to schedule and account for a
+// request appear here. Anything else the client sent is preserved verbatim in
+// raw so an adapter can read it (the same "kernel does not interpret" rule that
+// governs tools).
+type anthropicRequest struct {
+	Model     string             `json:"model"`
+	System    json.RawMessage    `json:"system,omitempty"`
+	Messages  []anthropicMessage `json:"messages"`
+	MaxTokens int                `json:"max_tokens"`
+	Stream    bool               `json:"stream,omitempty"`
+
+	Temperature *float64               `json:"temperature,omitempty"`
+	TopP        *float64               `json:"top_p,omitempty"`
+	TopK        int                    `json:"top_k,omitempty"`
+	StopSeqs    []string               `json:"stop_sequences,omitempty"`
+	Tools       json.RawMessage        `json:"tools,omitempty"`
+	ToolChoice  json.RawMessage        `json:"tool_choice,omitempty"`
+	Thinking    *types.ThinkingConfig  `json:"thinking,omitempty"`
+	Metadata    map[string]interface{} `json:"metadata,omitempty"`
+
+	// raw is the untouched body, so an adapter can read fields the gateway
+	// has no reason to know (cache_control, service_tier, …).
+	raw json.RawMessage
+}
+
+// anthropicMessage accepts Anthropic's content as either a plain string or an
+// array of typed blocks; both normalize to the unified message shape.
+type anthropicMessage struct {
+	Role    string          `json:"role"`
+	Content json.RawMessage `json:"content"`
+}
+
+// toUnified converts the parsed request into the internal format.
+func (a *anthropicRequest) toUnified(defaultModel string) (*types.ChatRequest, error) {
+	if len(a.Messages) == 0 {
+		return nil, errors.New("messages is required")
+	}
+	msgs := make([]types.ChatMessage, 0, len(a.Messages))
+	for i, m := range a.Messages {
+		u := types.ChatMessage{Role: m.Role, Content: normalizeAnthropicContent(m.Content)}
+		if u.Role == "" {
+			return nil, fmt.Errorf("messages[%d].role is required", i)
+		}
+		msgs = append(msgs, u)
+	}
+	model := a.Model
+	if model == "" {
+		model = defaultModel
+	}
+	th := a.Thinking
+	if th == nil {
+		// Anthropic's default is thinking OFF, and `thinking: {enabled:false}`
+		// is the explicit spelling of that. Emitting it keeps the unified
+		// request honest about the client's intent instead of leaving the
+		// field absent and letting each adapter guess.
+		no := false
+		th = &types.ThinkingConfig{Enabled: &no}
+	}
+	return &types.ChatRequest{
+		Model:         model,
+		Messages:      msgs,
+		Temperature:   a.Temperature,
+		MaxTokens:     a.MaxTokens,
+		Stream:        a.Stream,
+		TopP:          a.TopP,
+		TopK:          a.TopK,
+		StopSequences: a.StopSeqs,
+		Tools:         a.Tools,
+		ToolChoice:    a.ToolChoice,
+		System:        systemToString(a.System),
+		Thinking:      th,
+		ExtraBody:     map[string]interface{}{},
+	}, nil
+}
+
+// normalizeAnthropicContent turns Anthropic content into the unified shape.
+//
+// A string stays a string. An array of blocks is preserved as-is: the unified
+// format deliberately carries multimodal/tool blocks verbatim (ChatMessage
+// .Content is json.RawMessage) rather than flattening them, because only an
+// adapter knows which block types its upstream accepts.
+func normalizeAnthropicContent(raw json.RawMessage) json.RawMessage {
+	trimmed := strings.TrimSpace(string(raw))
+	if trimmed == "" {
+		return types.StringContent("")
+	}
+	if strings.HasPrefix(trimmed, `"`) {
+		return raw
+	}
+	return raw
+}
+
+// systemToString accepts Anthropic's two spellings: a plain string, or an array
+// of {type:"text", text} blocks.
+func systemToString(raw json.RawMessage) string {
+	trimmed := strings.TrimSpace(string(raw))
+	if trimmed == "" || trimmed == "null" {
+		return ""
+	}
+	if strings.HasPrefix(trimmed, `"`) {
+		var s string
+		if err := json.Unmarshal(raw, &s); err == nil {
+			return s
+		}
+		return ""
+	}
+	var blocks []struct {
+		Type string `json:"type"`
+		Text string `json:"text"`
+	}
+	if err := json.Unmarshal(raw, &blocks); err != nil {
+		return ""
+	}
+	var sb strings.Builder
+	for _, b := range blocks {
+		if b.Type == "text" {
+			sb.WriteString(b.Text)
+		}
+	}
+	return sb.String()
+}
+
+func chatKind(stream bool) string {
+	if stream {
+		return "stream"
+	}
+	return "chat"
+}
+
+// directProviderList hands the serving-source resolver its candidate list.
+func directProviderList(direct *directRoute) []*provider.Provider {
+	if direct == nil {
+		return nil
+	}
+	return direct.cands
+}
+
+func tierForRoute(direct *directRoute) int {
+	if direct != nil {
+		return -1
+	}
+	return -2
+}
+
+func modelLabel(req *types.ChatRequest, usedModel string) string {
+	if usedModel != "" {
+		return usedModel
+	}
+	return req.Model
+}
diff --git a/internal/gateway/messages_test.go b/internal/gateway/messages_test.go
new file mode 100644
index 0000000..ec4a0e5
--- /dev/null
+++ b/internal/gateway/messages_test.go
@@ -0,0 +1,303 @@
+package gateway
+
+import (
+	"context"
+	"encoding/json"
+	"fmt"
+	"io"
+	"net/http"
+	"net/http/httptest"
+	"path/filepath"
+	"strings"
+	"testing"
+
+	"llmsproxy/internal/config"
+	"llmsproxy/internal/core"
+)
+
+// 本文件端到端验证 Phase 2 的核心主张：**对外协议由适配器决定，内核只调度**。
+//
+// 具体做法：起一个真实的 core + gateway，源挂 anthropic 适配器，上游是一个
+// 只懂 Anthropic Messages 的假服务器。然后：
+//
+//	1. Anthropic 客户端形状的请求 → 拿到 Anthropic 形状的响应
+//	2. 调度/配额/审计与 /v1/chat/completions 走同一条路径
+//	3. 上游若不支持该模型，内核如实回 503，不伪造格式
+
+// anthropicUpstream 是一个只实现 Anthropic Messages 的假上游。
+func anthropicUpstream(t *testing.T) *httptest.Server {
+	return httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		if !strings.HasSuffix(r.URL.Path, "/messages") {
+			// 网关把请求发到了非 Anthropic 路径 —— 显式暴露这个 bug
+			t.Errorf("请求打到了非 Anthropic 路径: %s", r.URL.Path)
+			w.WriteHeader(http.StatusNotFound)
+			return
+		}
+		body, _ := io.ReadAll(r.Body)
+		var probe map[string]interface{}
+		_ = json.Unmarshal(body, &probe)
+		if _, ok := probe["messages"]; !ok {
+			t.Errorf("上游收到没有 messages 的请求体: %s", body)
+		}
+		if _, ok := probe["system"]; !ok && probe["system"] == nil {
+			// system 是可选的，仅在客户端给了的时候才应出现；不强制
+			_ = ok
+		}
+		if _, ok := probe["max_tokens"]; !ok {
+			t.Errorf("Anthropic 要求 max_tokens，上游请求体却缺这个字段: %s", body)
+		}
+
+		// Streaming requests must get a real Anthropic SSE sequence; a JSON body
+		// here would leave the adapter with zero frames and the gateway would
+		// (correctly) reject the stream as empty.
+		if stream, _ := probe["stream"].(bool); stream {
+			w.Header().Set("Content-Type", "text/event-stream")
+			fl, _ := w.(http.Flusher)
+			frames := []string{
+				`event: message_start` + "\n" + `data: {"type":"message_start","message":{"id":"msg_mock","type":"message","role":"assistant","model":"` + probeModel(probe) + `","content":[],"usage":{"input_tokens":3,"output_tokens":0}}}`,
+				`event: content_block_start` + "\n" + `data: {"type":"content_block_start","index":0,"content_block":{"type":"text","text":""}}`,
+				`event: content_block_delta` + "\n" + `data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"pong"}}`,
+				`event: content_block_stop` + "\n" + `data: {"type":"content_block_stop","index":0}`,
+				`event: message_delta` + "\n" + `data: {"type":"message_delta","delta":{"stop_reason":"end_turn"},"usage":{"output_tokens":5}}`,
+				`event: message_stop` + "\n" + `data: {"type":"message_stop"}`,
+			}
+			for _, f := range frames {
+				fmt.Fprint(w, f+"\n\n")
+				if fl != nil {
+					fl.Flush()
+				}
+			}
+			return
+		}
+		w.Header().Set("Content-Type", "application/json")
+		fmt.Fprint(w, `{
+			"id":"msg_mock","type":"message","role":"assistant","model":"`+probeModel(probe)+`",
+			"content":[{"type":"thinking","thinking":"let me think"},{"type":"text","text":"pong"}],
+			"stop_reason":"end_turn",
+			"usage":{"input_tokens":3,"output_tokens":5}
+		}`)
+	}))
+}
+
+func probeModel(m map[string]interface{}) string {
+	if s, ok := m["model"].(string); ok {
+		return s
+	}
+	return ""
+}
+
+// messagesGateway 起一个真实网关，源挂 anthropic 适配器。
+func messagesGateway(t *testing.T) (*httptest.Server, string) {
+	t.Helper()
+	up := anthropicUpstream(t)
+	t.Cleanup(up.Close)
+
+	dir := t.TempDir()
+	// adapterDir 故意不预先创建：VM 首启会把内置适配器 seed 进去。
+	// 预先创建会被当成"目录已存在即为权威"，于是没有任何内置适配器被加载，
+	// 表现为 "adapter anthropic not loaded"。
+	cfg := &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: "claude-test",
+		GatewayKeys:  []string{"sk-e2e-0001"},
+		Sources: []config.Source{{
+			Name: "anth", BaseURL: up.URL, Adapter: "anthropic",
+			APIKey: "k", MaxConcurrent: 4,
+			Models: []config.Model{{ID: "claude-test", Kind: "chat", Priority: 100}},
+		}},
+	}
+
+	c, err := core.NewFromConfig(cfg)
+	if err != nil {
+		t.Fatalf("core: %v", err)
+	}
+	t.Cleanup(c.Close)
+	gw, err := New(c)
+	if err != nil {
+		t.Fatalf("gateway: %v", err)
+	}
+	ts := httptest.NewServer(gw.Handler())
+	t.Cleanup(ts.Close)
+	return ts, "sk-e2e-0001"
+}
+
+func postJSON(t *testing.T, url, key, body string) (int, string) {
+	t.Helper()
+	req, _ := http.NewRequest(http.MethodPost, url, strings.NewReader(body))
+	req.Header.Set("Content-Type", "application/json")
+	req.Header.Set("Authorization", "Bearer "+key)
+	resp, err := http.DefaultClient.Do(req)
+	if err != nil {
+		t.Fatalf("request: %v", err)
+	}
+	defer resp.Body.Close()
+	b, _ := io.ReadAll(resp.Body)
+	return resp.StatusCode, string(b)
+}
+
+// TestMessagesEndpointSpeaksAnthropic 是本文件的主断言：Anthropic 客户端
+// 打 /v1/messages，拿到 Anthropic 形状。
+func TestMessagesEndpointSpeaksAnthropic(t *testing.T) {
+	ts, key := messagesGateway(t)
+	code, body := postJSON(t, ts.URL+"/v1/messages", key,
+		`{"model":"claude-test","max_tokens":64,"messages":[{"role":"user","content":"hi"}]}`)
+	if code != 200 {
+		t.Fatalf("status = %d: %s", code, body)
+	}
+	var m map[string]interface{}
+	if err := json.Unmarshal([]byte(body), &m); err != nil {
+		t.Fatalf("响应不是 JSON: %v (%s)", err, body)
+	}
+	if m["type"] != "message" {
+		t.Errorf("type = %v, want message（内核不该退回 OpenAI 的 chat.completion）", m["type"])
+	}
+	if m["stop_reason"] != "end_turn" {
+		t.Errorf("stop_reason = %v, want end_turn", m["stop_reason"])
+	}
+	blocks, _ := m["content"].([]interface{})
+	if len(blocks) != 2 {
+		t.Fatalf("content blocks = %v, want [thinking, text]", m["content"])
+	}
+	if b0 := blocks[0].(map[string]interface{}); b0["type"] != "thinking" || b0["thinking"] != "let me think" {
+		t.Errorf("首个块应为 thinking: %v", b0)
+	}
+	if b1 := blocks[1].(map[string]interface{}); b1["type"] != "text" || b1["text"] != "pong" {
+		t.Errorf("第二个块应为 text/pong: %v", b1)
+	}
+	u := m["usage"].(map[string]interface{})
+	if u["input_tokens"] != float64(3) || u["output_tokens"] != float64(5) {
+		t.Errorf("usage 应为 Anthropic 的 input/output 命名: %v", u)
+	}
+	// 内核不得混入 OpenAI 形状
+	for _, forbidden := range []string{`"object"`, `"choices"`, `"chat.completion"`, `"prompt_tokens"`} {
+		if strings.Contains(body, forbidden) {
+			t.Errorf("OpenAI 字段 %s 混入 Anthropic 响应: %s", forbidden, body)
+		}
+	}
+}
+
+// TestMessagesStreamSpeaksAnthropic 验证流式也是 Anthropic 事件序列，
+// 而不是 OpenAI SSE。
+func TestMessagesStreamSpeaksAnthropic(t *testing.T) {
+	ts, key := messagesGateway(t)
+	req, _ := http.NewRequest(http.MethodPost, ts.URL+"/v1/messages",
+		strings.NewReader(`{"model":"claude-test","max_tokens":64,"stream":true,"messages":[{"role":"user","content":"hi"}]}`))
+	req.Header.Set("Content-Type", "application/json")
+	req.Header.Set("Authorization", "Bearer "+key)
+	resp, err := http.DefaultClient.Do(req)
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer resp.Body.Close()
+	body, _ := io.ReadAll(resp.Body)
+	s := string(body)
+
+	if !strings.Contains(s, "message_start") {
+		t.Errorf("Anthropic 流应以 message_start 开头: %s", s)
+	}
+	if !strings.Contains(s, "message_stop") {
+		t.Errorf("Anthropic 流应以 message_stop 收尾: %s", s)
+	}
+	// 内核不得擅自给 Anthropic 流加 OpenAI 的 [DONE]
+	if strings.Contains(s, "[DONE]") {
+		t.Errorf("Anthropic 流里出现了 OpenAI 的 [DONE] 终止标记: %s", s)
+	}
+	if strings.Contains(s, "chat.completion.chunk") {
+		t.Errorf("流里混入 OpenAI chunk 形状: %s", s)
+	}
+}
+
+// TestMessagesRejectsBadRequest 确认入向校验（Anthropic 要求 max_tokens）
+// 与 OpenAI 路径一样在入口拒绝，不占用上游槽位。
+func TestMessagesRejectsBadRequest(t *testing.T) {
+	ts, key := messagesGateway(t)
+	// Anthropic 允许省略 max_tokens（上游会给默认），所以「有 messages 无
+	// max_tokens」必须被接受；真正该拒的是没有 messages。
+	code, body := postJSON(t, ts.URL+"/v1/messages", key,
+		`{"model":"claude-test","messages":[{"role":"user","content":"hi"}]}`)
+	if code != 200 {
+		t.Fatalf("有 messages 的请求应被接受（max_tokens 可省），却得 %d: %s", code, body)
+	}
+	code, body = postJSON(t, ts.URL+"/v1/messages", key, `{"model":"claude-test"}`)
+	if code != 400 {
+		t.Errorf("无 messages 的请求 status = %d, want 400: %s", code, body)
+	}
+	// 坏 JSON
+	code, _ = postJSON(t, ts.URL+"/v1/messages", key, `{"model":`)
+	if code != 400 {
+		t.Errorf("坏 JSON status = %d, want 400", code)
+	}
+}
+
+// TestMessagesUnconfiguredFallsBackToOpenai 验证 R5 的兜底在新增端点上同样
+// 生效：源没有声明序列化钩子时（此例用 openai 适配器），回 OpenAI 形状。
+func TestMessagesUnconfiguredFallsBackToOpenai(t *testing.T) {
+	up := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.Header().Set("Content-Type", "application/json")
+		fmt.Fprint(w, `{"choices":[{"message":{"content":"pong"},"finish_reason":"stop"}],
+			"usage":{"prompt_tokens":3,"completion_tokens":1,"total_tokens":4}}`)
+	}))
+	defer up.Close()
+
+	dir := t.TempDir()
+	// adapterDir 故意不预先创建：VM 首启会把内置适配器 seed 进去。
+	// 预先创建会被当成"目录已存在即为权威"，于是没有任何内置适配器被加载，
+	// 表现为 "adapter anthropic not loaded"。
+	cfg := &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: "gpt-test",
+		GatewayKeys:  []string{"sk-e2e-0001"},
+		Sources: []config.Source{{
+			Name: "oai", BaseURL: up.URL, Adapter: "openai", APIKey: "k", MaxConcurrent: 4,
+			Models: []config.Model{{ID: "gpt-test", Kind: "chat"}},
+		}},
+	}
+	c, err := core.NewFromConfig(cfg)
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer c.Close()
+	gw, err := New(c)
+	if err != nil {
+		t.Fatal(err)
+	}
+	ts := httptest.NewServer(gw.Handler())
+	defer ts.Close()
+
+	code, body := postJSON(t, ts.URL+"/v1/messages", "sk-e2e-0001",
+		`{"model":"gpt-test","max_tokens":64,"messages":[{"role":"user","content":"hi"}]}`)
+	if code != 200 {
+		t.Fatalf("status = %d: %s", code, body)
+	}
+	if !strings.Contains(body, "chat.completion") {
+		t.Errorf("无序列化钩子的源应回落 OpenAI 兜底形状: %s", body)
+	}
+}
+
+// TestMessagesRequiresAuth 确认新端点走同一套鉴权（不因为是新协议而放行）。
+func TestMessagesRequiresAuth(t *testing.T) {
+	ts, _ := messagesGateway(t)
+	code, _ := postJSON(t, ts.URL+"/v1/messages", "wrong-key",
+		`{"model":"claude-test","max_tokens":8,"messages":[{"role":"user","content":"hi"}]}`)
+	if code != 401 {
+		t.Errorf("错误 key status = %d, want 401", code)
+	}
+	// 无 auth
+	req, _ := http.NewRequest(http.MethodPost, ts.URL+"/v1/messages", strings.NewReader("{}"))
+	req.Header.Set("Content-Type", "application/json")
+	resp, err := http.DefaultClient.Do(req)
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer resp.Body.Close()
+	if resp.StatusCode != 401 {
+		t.Errorf("无 key status = %d, want 401", resp.StatusCode)
+	}
+	_ = context.Background()
+}
diff --git a/internal/gateway/plugin_call_api_test.go b/internal/gateway/plugin_call_api_test.go
index 1f9f328..7f4afa1 100644
--- a/internal/gateway/plugin_call_api_test.go
+++ b/internal/gateway/plugin_call_api_test.go
@@ -147,4 +147,4 @@ func TestPluginCallIsAdminOnly(t *testing.T) {
 		t.Fatalf("a non-admin key reached the call endpoint: HTTP %d %s",
 			rr.Code, rr.Body.String())
 	}
-}
\ No newline at end of file
+}
diff --git a/internal/gateway/plugin_http_routes.go b/internal/gateway/plugin_http_routes.go
index 71e6e46..d7f8ce1 100644
--- a/internal/gateway/plugin_http_routes.go
+++ b/internal/gateway/plugin_http_routes.go
@@ -204,4 +204,4 @@ 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
\ No newline at end of file
+var _ = strconv.Itoa
diff --git a/internal/gateway/plugins_api_test.go b/internal/gateway/plugins_api_test.go
index 85cbcd7..8c4e69f 100644
--- a/internal/gateway/plugins_api_test.go
+++ b/internal/gateway/plugins_api_test.go
@@ -222,7 +222,7 @@ func TestPluginsListCarriesAccessTables(t *testing.T) {
 		t.Fatalf("GET /api/plugins = %d", rr.Code)
 	}
 	var out struct {
-		Visibility map[string]string     `json:"visibility"`
+		Visibility map[string]string        `json:"visibility"`
 		HTTPAccess []map[string]interface{} `json:"http_access"`
 	}
 	if err := json.Unmarshal(rr.Body.Bytes(), &out); err != nil {
diff --git a/internal/gateway/serializer_test.go b/internal/gateway/serializer_test.go
index 6b4403b..ca49326 100644
--- a/internal/gateway/serializer_test.go
+++ b/internal/gateway/serializer_test.go
@@ -32,15 +32,17 @@ func luaVMFor(t *testing.T) *lua.VM {
 	return vm
 }
 
-// TestFallbackUsedWhenAdapterDeclaresNothing 是「兜底」本身的测试：
-// 没有任何内置适配器声明 serialize_*，所以默认路径必须是 OpenAI。
+// TestFallbackUsedWhenAdapterDeclaresNothing 是「兜底」本身的测试。
+//
+// 注：anthropic 现在**确实**声明了 serialize_*（Phase 2 给了它 Anthropic
+// 形状），所以本测试只在仍未声明钩子的那些内置适配器上验证兜底行为。
+// 这正是设计意图：谁声明谁接管，没声明的继续用 OpenAI 兜底。
 func TestFallbackUsedWhenAdapterDeclaresNothing(t *testing.T) {
 	vm := luaVMFor(t)
-	// 先证明前提：内置适配器确实没声明这些钩子（否则下面测的是别的东西）
-	for _, name := range []string{"openai", "anthropic", "gemini", "ollama", "deepseek"} {
+	for _, name := range []string{"openai", "gemini", "ollama", "deepseek", "github", "groq"} {
 		for _, hook := range []string{"serialize_response", "serialize_chunk"} {
 			if vm.HasFn(name, hook) {
-				t.Fatalf("内置适配器 %s 声明了 %s —— 本测试的前提（默认走兜底）已不成立", name, hook)
+				t.Errorf("内置适配器 %s 声明了 %s —— 兜底测试的前提已不成立", name, hook)
 			}
 		}
 	}
@@ -48,6 +50,11 @@ func TestFallbackUsedWhenAdapterDeclaresNothing(t *testing.T) {
 	if _, isFallback := ser.(openAISerializer); !isFallback {
 		t.Fatalf("未声明钩子的适配器应得到兜底序列化器，得到 %T", ser)
 	}
+
+	// 对照：声明了钩子的适配器不走兜底
+	if _, isFallback := serializerFor(vm, "anthropic").(openAISerializer); isFallback {
+		t.Error("anthropic 声明了 serialize_response，不应走兜底")
+	}
 }
 
 // TestAdapterSerializerWinsOverFallback 锁定核心规则：适配器声明钩子即接管，
@@ -323,19 +330,30 @@ func TestSerializerForConcurrent(t *testing.T) {
 					errs <- "nil serializer"
 					return
 				}
-				// 结果必须自洽：同一适配器每次拿到同一类型
+				// 结果必须自洽：同一适配器每次拿到同一类型，且形状与该适配器
+				// 声明的协议一致（openai→OpenAI，anthropic→Anthropic）。
 				_, isFallback := ser.(openAISerializer)
-				if w%2 == 0 && !isFallback {
-					errs <- "openai 未走兜底"
-					return
+				wantShape := "chat.completion"
+				switch name {
+				case "anthropic":
+					if isFallback {
+						errs <- "anthropic 并发下退回兜底"
+						return
+					}
+					wantShape = `"type":"message"`
+				default:
+					if !isFallback {
+						errs <- name + " 并发下未走兜底"
+						return
+					}
 				}
 				body, err := ser.NonStream(&types.UnifiedResponse{Content: "x"}, "m", "id", 1)
 				if err != nil {
 					errs <- "serialize: " + err.Error()
 					return
 				}
-				if !strings.Contains(string(body), "chat.completion") {
-					errs <- "并发下输出形状不一致"
+				if !strings.Contains(string(body), wantShape) {
+					errs <- "并发下输出形状不一致: " + name + " -> " + string(body)
 					return
 				}
 			}
diff --git a/internal/gateway/server.go b/internal/gateway/server.go
index 3fe9cb8..32bc1d4 100644
--- a/internal/gateway/server.go
+++ b/internal/gateway/server.go
@@ -216,6 +216,14 @@ func (g *Gateway) routes(w http.ResponseWriter, r *http.Request) {
 	switch {
 	case r.URL.Path == "/v1/chat/completions":
 		g.handleChat(w, r)
+	case r.URL.Path == "/v1/messages":
+		// Anthropic Messages protocol. Scheduling, quotas, cooldowns, AUTO
+		// chain and accounting are EXACTLY the ones /v1/chat/completions uses —
+		// only the wire shapes differ, and those come from the serving
+		// adapter's serialize_* hooks. The kernel holds no Anthropic field
+		// names (see serializer.go: the fallback is OpenAI, and any adapter
+		// that declares serialize_* replaces it wholesale).
+		g.handleMessages(w, r)
 	case r.URL.Path == "/v1/images/generations":
 		g.handleImage(w, r)
 	case r.URL.Path == "/v1/models":
diff --git a/internal/gateway/stats.go b/internal/gateway/stats.go
index 8aeff2a..68d6f1b 100644
--- a/internal/gateway/stats.go
+++ b/internal/gateway/stats.go
@@ -53,7 +53,7 @@ type Req struct {
 	// 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"`
+	OK        bool  `json:"ok"`
 	// Status http status code
 	Status int `json:"status"`
 	// Err short error message
diff --git a/internal/gateway/stats_period_records_test.go b/internal/gateway/stats_period_records_test.go
index 912e08f..465ebf5 100644
--- a/internal/gateway/stats_period_records_test.go
+++ b/internal/gateway/stats_period_records_test.go
@@ -74,7 +74,7 @@ func TestStatsPeriodRecordsMatchPager(t *testing.T) {
 	rr := doReq(t, g, "GET", "/api/stats?period=day&limit=100", "")
 	var st struct {
 		Records []struct {
-			Time  int64 `json:"time"`
+			Time  int64  `json:"time"`
 			Model string `json:"model"`
 		} `json:"records"`
 	}
@@ -84,7 +84,7 @@ func TestStatsPeriodRecordsMatchPager(t *testing.T) {
 	rr2 := doReq(t, g, "GET", "/api/stats/records?limit=100", "")
 	var pg struct {
 		Records []struct {
-			Time  int64 `json:"time"`
+			Time  int64  `json:"time"`
 			Model string `json:"model"`
 		} `json:"records"`
 	}
diff --git a/internal/gateway/stats_unknown_key_test.go b/internal/gateway/stats_unknown_key_test.go
index 30a281b..62eff24 100644
--- a/internal/gateway/stats_unknown_key_test.go
+++ b/internal/gateway/stats_unknown_key_test.go
@@ -137,4 +137,4 @@ func TestStatsAPIUnknownKeyReturnsJSON(t *testing.T) {
 	if resp.StatusCode != http.StatusOK {
 		t.Fatalf("status = %d, want 200", resp.StatusCode)
 	}
-}
\ No newline at end of file
+}
diff --git a/internal/gateway/ui_image_page_test.go b/internal/gateway/ui_image_page_test.go
index a1b6bfe..5f533ba 100644
--- a/internal/gateway/ui_image_page_test.go
+++ b/internal/gateway/ui_image_page_test.go
@@ -48,11 +48,11 @@ func TestUI_HasDedicatedImagePage(t *testing.T) {
 	ui := uiSource(t)
 
 	for _, want := range []string{
-		`data-tab="image"`,         // sidebar entry
-		`id="tab-image"`,           // pane
-		`function renderImage`,     // renderer registered with refresh()
-		`tab === "image"`,          // dispatched
-		`/v1/images/generations`,   // the endpoint the page actually calls
+		`data-tab="image"`,       // sidebar entry
+		`id="tab-image"`,         // pane
+		`function renderImage`,   // renderer registered with refresh()
+		`tab === "image"`,        // dispatched
+		`/v1/images/generations`, // the endpoint the page actually calls
 	} {
 		if !strings.Contains(ui, want) {
 			t.Errorf("image page is missing %q", want)
@@ -99,8 +99,8 @@ func TestUI_ImagePageCallsTheImageEndpoint(t *testing.T) {
 // contract: the field the UI reads must actually be serialized.
 func TestSourceStatus_ExposesModelKindsToTheUI(t *testing.T) {
 	st := provider.SourceStatus{
-		Name:       "vela",
-		Models:     []string{"claude-sonnet-5.5", "gpt-image-2"},
+		Name:   "vela",
+		Models: []string{"claude-sonnet-5.5", "gpt-image-2"},
 		ModelKinds: []provider.ModelKindInfo{
 			{ID: "claude-sonnet-5.5", Kind: "chat"},
 			{ID: "gpt-image-2", Kind: "image"},
@@ -192,4 +192,4 @@ func TestImageEndpointComposesUnderRealConditions(t *testing.T) {
 			"default image path doubles the version segment; if the upstream ever "+
 			"accepts it, composeImageEndpoint's rule needs rechecking", doubled, resp2.StatusCode)
 	}
-}
\ No newline at end of file
+}
diff --git a/internal/gateway/ui_image_wire_test.go b/internal/gateway/ui_image_wire_test.go
index 56efb32..3364ddf 100644
--- a/internal/gateway/ui_image_wire_test.go
+++ b/internal/gateway/ui_image_wire_test.go
@@ -106,4 +106,4 @@ func sendImageSource(t *testing.T) string {
 		body = body[:end+1]
 	}
 	return body
-}
\ No newline at end of file
+}
diff --git a/internal/gateway/ui_key_auto_test.go b/internal/gateway/ui_key_auto_test.go
index c49fea9..3d131cb 100644
--- a/internal/gateway/ui_key_auto_test.go
+++ b/internal/gateway/ui_key_auto_test.go
@@ -126,6 +126,7 @@ func TestUIStatsShowsPeriodWindow(t *testing.T) {
 		t.Fatalf("i18n key periodRange appears %d time(s), want both zh and en", n)
 	}
 }
+
 // Both canvases can exist in the DOM at once: the modal is mounted on top of
 // the settings tab, whose #scr-canvas stays alive behind it. Resolving the
 // canvas by document order (whichever comes first) painted the per-key chain
diff --git a/internal/gateway/ui_plugin_test.go b/internal/gateway/ui_plugin_test.go
index 9414097..8e258a8 100644
--- a/internal/gateway/ui_plugin_test.go
+++ b/internal/gateway/ui_plugin_test.go
@@ -396,15 +396,15 @@ func TestPluginSidebarEntryCarriesATitle(t *testing.T) {
 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
+		`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)
diff --git a/internal/lua/adapters/anthropic.lua b/internal/lua/adapters/anthropic.lua
index 578ebb8..459eb41 100644
--- a/internal/lua/adapters/anthropic.lua
+++ b/internal/lua/adapters/anthropic.lua
@@ -456,4 +456,124 @@ function adapter.transform_error(status, body)
     return nil
 end
 
+-- ============================================================
+-- 对外序列化（统一格式 → Anthropic 形状）
+-- ============================================================
+--
+-- 内核只持统一格式；对外发什么形状由适配器决定（Phase 2）。
+-- 这几个钩子让以 anthropic 适配器为服务端的源，直接对 Anthropic 协议客户端
+-- 提供 /v1/messages 语义，而内核里不再有任何 Anthropic 字段名。
+--
+-- 与入向的 transform_* 互为逆运算：那里的映射（end_turn→stop）在这里反着
+-- 做回去（stop→end_turn），且默认 stop_sequence 是最保守的选择——它语义上
+-- 最接近"正常结束"。
+
+-- 统一 finish_reason → Anthropic stop_reason
+local function to_stop_reason(finish)
+    if finish == "tool_calls" then return "tool_use" end
+    if finish == "length" then return "max_tokens" end
+    if finish == "content_filter" then return "refusal" end
+    -- stop / 空 / 未知 → end_turn（Anthropic 的"正常说完"）
+    return "end_turn"
+end
+
+-- usage 数字化：内部 TokenUsage 的字段名与 Anthropic 的 input_tokens/output_tokens
+-- 不同名，这里做最后一次翻译。
+local function to_anthropic_usage(u)
+    if type(u) ~= "table" then
+        return { input_tokens = 0, output_tokens = 0 }
+    end
+    return {
+        input_tokens  = u.prompt or 0,
+        output_tokens = u.completion or 0,
+    }
+end
+
+-- 非流式响应：统一结果 → Anthropic Messages 响应对象
+function adapter.serialize_response(payload)
+    local p = json.decode(payload)
+    local blocks = {}
+    -- thinking 放最前，与 Anthropic 自己的块顺序一致
+    if p.reasoning ~= nil and p.reasoning ~= "" then
+        blocks[#blocks + 1] = { type = "thinking", thinking = p.reasoning }
+    end
+    blocks[#blocks + 1] = { type = "text", text = p.content or "" }
+    if type(p.tool_calls) == "table" and #p.tool_calls > 0 then
+        for _, tc in ipairs(p.tool_calls) do
+            blocks[#blocks + 1] = {
+                type = "tool_use",
+                id = tc.id or "",
+                name = tc.name or "",
+                input = tc.arguments or {},
+            }
+        end
+    end
+    return json.encode({
+        id = p.id or "",
+        type = "message",
+        role = "assistant",
+        model = p.model or "",
+        content = blocks,
+        stop_reason = to_stop_reason(p.finish_reason),
+        stop_sequence = nil,
+        usage = to_anthropic_usage(p.token_usage),
+    })
+end
+
+-- 角色前导：Anthropic 的流以 message_start 开头（不是 role delta）
+function adapter.serialize_open_preamble(payload)
+    local p = json.decode(payload)
+    return json.encode({
+        type = "message_start",
+        message = {
+            id = p.id or "",
+            type = "message",
+            role = "assistant",
+            model = p.model or "",
+            content = {},
+            stop_reason = nil,
+            stop_sequence = nil,
+            usage = { input_tokens = 0, output_tokens = 0 },
+        },
+    })
+end
+
+-- 流式增量：统一 chunk → Anthropic content_block_delta
+--
+-- 真实 Anthropic 流需要 content_block_start（开块）→ delta → stop（关块）
+-- 的三段式。这里用与多数 Anthropic 客户端兼容的简化形式：首个文本增量
+-- 带 index=0 的 delta，后续同 index；关闭由 serialize_close 负责。
+function adapter.serialize_chunk(payload)
+    local p = json.decode(payload)
+    if p.done then
+        -- 终止由 serialize_close 统一发，这里不重复
+        return ""
+    end
+    if p.reasoning ~= nil and p.reasoning ~= "" then
+        return json.encode({
+            type = "content_block_delta",
+            index = 0,
+            delta = { type = "thinking_delta", thinking = p.reasoning },
+        })
+    end
+    return json.encode({
+        type = "content_block_delta",
+        index = 0,
+        delta = { type = "text_delta", text = p.content or "" },
+    })
+end
+
+-- 收尾：message_delta（stop_reason + usage）→ message_stop
+function adapter.serialize_close(payload)
+    local p = json.decode(payload)
+    return json.encode({
+        type = "message_delta",
+        delta = {
+            stop_reason = to_stop_reason(p.finish_reason),
+            stop_sequence = nil,
+        },
+        usage = { output_tokens = 0 },
+    }) .. "\n" .. json.encode({ type = "message_stop" })
+end
+
 return adapter
diff --git a/internal/lua/ser_check_test.go b/internal/lua/ser_check_test.go
new file mode 100644
index 0000000..eae02bd
--- /dev/null
+++ b/internal/lua/ser_check_test.go
@@ -0,0 +1,53 @@
+package lua
+
+import (
+	"encoding/json"
+	"path/filepath"
+	"strings"
+	"testing"
+)
+
+func TestAnthropicSerializerShape(t *testing.T) {
+	vm := NewVM(filepath.Join(t.TempDir(), "adapters"))
+	if err := vm.Start(); err != nil {
+		t.Fatal(err)
+	}
+	defer vm.Stop()
+
+	if !vm.HasFn("anthropic", "serialize_response") {
+		t.Fatal("anthropic 未声明 serialize_response")
+	}
+	out, err := vm.Transform("anthropic", "serialize_response",
+		`{"model":"claude-x","id":"msg_1","content":"hello","finish_reason":"stop","token_usage":{"prompt":10,"completion":5,"total":15}}`)
+	if err != nil {
+		t.Fatal(err)
+	}
+	var m map[string]interface{}
+	if err := json.Unmarshal([]byte(out), &m); err != nil {
+		t.Fatalf("not JSON: %s", out)
+	}
+	if m["type"] != "message" || m["role"] != "assistant" {
+		t.Errorf("Anthropic 形状不对: %s", out)
+	}
+	if m["stop_reason"] != "end_turn" {
+		t.Errorf("stop_reason = %v, want end_turn", m["stop_reason"])
+	}
+	blocks, _ := m["content"].([]interface{})
+	if len(blocks) != 1 {
+		t.Fatalf("blocks = %v", m["content"])
+	}
+	b0 := blocks[0].(map[string]interface{})
+	if b0["type"] != "text" || b0["text"] != "hello" {
+		t.Errorf("首个块应为 text/hello: %v", b0)
+	}
+	u := m["usage"].(map[string]interface{})
+	if u["input_tokens"] != float64(10) || u["output_tokens"] != float64(5) {
+		t.Errorf("usage 字段名/值不对: %v", u)
+	}
+	// 内核不得混入 OpenAI 字段
+	for _, k := range []string{"object", "choices", "usage.prompt_tokens"} {
+		if strings.Contains(out, `"`+k+`"`) {
+			t.Errorf("OpenAI 字段 %q 混入 Anthropic 输出: %s", k, out)
+		}
+	}
+}