Merge pull request #5668 from Wei-Shaw/fix/codex-turn-state-and-fingerprint-optin

fix(openai): Codex 回合状态回传、v2 压缩探测与指纹收敛改为 opt-in

backend/internal/handler/openai_gateway_handler.go

@@ -300,6 +300,11 @@ func (h *OpenAIGatewayHandler) Responses(c *gin.Context) {
 	}
 	legacyCompact := service.IsOpenAIResponsesCompactPath(c)
 	nativeV2 := isBareOpenAIResponsesPath(c) && isOpenAIRemoteCompactionV2Request(body)
+	if nativeV2 {
+		// 原生 v2 压缩出站前补注 x-codex-beta-features: remote_compaction_v2，
+		// 与真实 Codex 线型一致（网关链剥头后本级负责恢复，#5586）。
+		service.MarkOpenAINativeCompactionV2(c)
+	}
 	// body-signal compact：上游 unary 等待期间向下游发 SSE 注释行心跳，防止
 	// 反向代理空闲超时掐断长压缩连接（#3887）。首拍延迟一个心跳间隔，快速
 	// 失败仍走 JSON+状态码链路；未标记客户端流式或间隔为 0 时是 no-op。

backend/internal/service/account_test_service.go

@@ -610,10 +610,11 @@ func (s *AccountTestService) testOpenAIAccountConnection(c *gin.Context, account
 	}
 
 	// Align test routing with gateway behavior: OpenAI accounts apply normal
-	// account model mapping, and compact mode applies compact-only mapping on top.
+	// account model mapping. Native remote compaction v2 rides the ordinary
+	// /responses wire and does NOT apply the legacy compact-only mapping
+	// (post-#5641 semantics: compact_model_mapping is /responses/compact-only).
 	testModelID = account.GetMappedModel(testModelID)
 	if mode == AccountTestModeCompact {
-		testModelID = resolveOpenAICompactForwardModel(account, testModelID)
 		return s.testOpenAICompactConnection(c, account, testModelID)
 	}
 
@@ -1976,8 +1977,10 @@ func (s *AccountTestService) testOpenAIChatCompletionsConnection(
 	return s.processOpenAIChatCompletionsStream(c, resp.Body)
 }
 
-// testOpenAICompactConnection probes /responses/compact and persists the
-// resulting capability state on the account.
+// testOpenAICompactConnection probes native remote compaction v2 (streaming
+// /responses with a compaction_trigger input item) and persists the resulting
+// capability state on the account. The legacy unary /responses/compact
+// endpoint has been sunset upstream (404, #5598/#5624) and is no longer probed.
 func (s *AccountTestService) testOpenAICompactConnection(c *gin.Context, account *Account, testModelID string) error {
 	ctx := c.Request.Context()
 	credentialAccount := account
@@ -2002,7 +2005,7 @@ func (s *AccountTestService) testOpenAICompactConnection(c *gin.Context, account
 		if authToken == "" && !credentialAccount.IsOpenAIAgentIdentity() {
 			return s.sendErrorAndEnd(c, "No access token available")
 		}
-		apiURL = chatgptCodexAPIURL + "/compact"
+		apiURL = chatgptCodexAPIURL
 	case account.Type == AccountTypeAPIKey:
 		authToken = account.GetOpenAIApiKey()
 		if authToken == "" {
@@ -2016,7 +2019,7 @@ func (s *AccountTestService) testOpenAICompactConnection(c *gin.Context, account
 		if err != nil {
 			return s.sendErrorAndEnd(c, fmt.Sprintf("Invalid base URL: %s", err.Error()))
 		}
-		apiURL = appendOpenAIResponsesRequestPathSuffix(buildOpenAIResponsesURL(normalizedBaseURL), "/compact")
+		apiURL = buildOpenAIResponsesURL(normalizedBaseURL)
 	default:
 		return s.sendErrorAndEnd(c, fmt.Sprintf("Unsupported account type: %s", account.Type))
 	}
@@ -2027,7 +2030,11 @@ func (s *AccountTestService) testOpenAICompactConnection(c *gin.Context, account
 	c.Writer.Header().Set("X-Accel-Buffering", "no")
 	c.Writer.Flush()
 
-	payloadBytes, _ := json.Marshal(createOpenAICompactProbePayload(testModelID))
+	// 原生 v2 走普通 /responses 线：OAuth 与真实转发一致做上游模型归一化。
+	if isOAuth {
+		testModelID = normalizeOpenAIModelForUpstream(credentialAccount, testModelID)
+	}
+	payloadBytes, _ := json.Marshal(createOpenAICompactProbePayload(testModelID, isOAuth))
 	if !agentIdentityTaskRecoveryWasTried(ctx) {
 		s.sendEvent(c, TestEvent{Type: "test_start", Model: testModelID})
 	}
@@ -2039,7 +2046,9 @@ func (s *AccountTestService) testOpenAICompactConnection(c *gin.Context, account
 	req = req.WithContext(WithHTTPUpstreamProfile(req.Context(), HTTPUpstreamProfileOpenAI))
 
 	req.Header.Set("Content-Type", "application/json")
-	req.Header.Set("Accept", "application/json")
+	// v2 探测是流式请求；同时补注协商头，与真实 codex 出站线型一致。
+	req.Header.Set("Accept", "text/event-stream")
+	ensureOpenAIRemoteCompactionV2BetaFeature(req.Header)
 	if credentialAccount.IsOpenAIAgentIdentity() {
 		authHeaders, authErr := buildAgentIdentityAuthenticationHeaders(ctx, s.accountRepo, s.agentIdentityWS, &s.agentIdentityTaskMu, credentialAccount)
 		if authErr != nil {
@@ -2061,6 +2070,12 @@ func (s *AccountTestService) testOpenAICompactConnection(c *gin.Context, account
 	if isOAuth {
 		req.Host = "chatgpt.com"
 		setOpenAIChatGPTAccountHeaders(req.Header, credentialAccount)
+		// 指纹收敛：探测与真实转发走同一个 /responses 端点，身份也必须同构，
+		// 否则探测流量会以「缺 x-codex-installation-id + 非收敛 session」的
+		// 形态暴露在上游眼里。账号关闭收敛（off）时返回 nil，探测保持原样。
+		if fpIDs := resolveCodexFingerprintIDsFromRequest(account, req.Header); fpIDs != nil {
+			applyCodexFingerprintHeaders(req.Header, fpIDs)
+		}
 	}
 
 	// 账号级请求头覆写：测试请求与真实转发保持一致的最终头
@@ -2074,7 +2089,7 @@ func (s *AccountTestService) testOpenAICompactConnection(c *gin.Context, account
 	resp, err := s.httpUpstream.DoWithTLS(req, proxyURL, account.ID, account.Concurrency, s.tlsFPProfileService.ResolveTLSProfile(account))
 	if err != nil {
 		if s.accountRepo != nil {
-			updates := buildOpenAICompactProbeExtraUpdates(nil, nil, err, time.Now())
+			updates := buildOpenAICompactProbeExtraUpdates(nil, nil, err, false, time.Now())
 			_ = s.accountRepo.UpdateExtra(ctx, account.ID, updates)
 			mergeAccountExtra(account, updates)
 		}
@@ -2093,8 +2108,9 @@ func (s *AccountTestService) testOpenAICompactConnection(c *gin.Context, account
 		return s.testOpenAICompactConnection(c, account, testModelID)
 	}
 
+	compactionFound := openAICompactProbeFoundCompactionItem(body)
 	if s.accountRepo != nil {
-		updates := buildOpenAICompactProbeExtraUpdates(resp, body, nil, time.Now())
+		updates := buildOpenAICompactProbeExtraUpdates(resp, body, nil, compactionFound, time.Now())
 		if codexUpdates, err := extractOpenAICodexProbeUpdates(resp); err == nil && len(codexUpdates) > 0 {
 			updates = mergeExtraUpdates(updates, codexUpdates)
 		}
@@ -2116,7 +2132,11 @@ func (s *AccountTestService) testOpenAICompactConnection(c *gin.Context, account
 		return s.sendErrorAndEnd(c, fmt.Sprintf("API returned %d: %s", resp.StatusCode, string(body)))
 	}
 
-	s.sendEvent(c, TestEvent{Type: "content", Text: "Compact probe succeeded"})
+	if !compactionFound {
+		return s.sendErrorAndEnd(c, "Upstream returned 2xx without a compaction output item (native remote compaction v2 unsupported on this chain)")
+	}
+
+	s.sendEvent(c, TestEvent{Type: "content", Text: "Compact probe succeeded (native remote compaction v2)"})
 	s.sendEvent(c, TestEvent{Type: "test_complete", Success: true})
 	return nil
 }

backend/internal/service/account_test_service_openai_compact_test.go

@@ -10,10 +10,16 @@ import (
 
 	"github.com/Wei-Shaw/sub2api/internal/config"
 	"github.com/gin-gonic/gin"
+	"github.com/google/uuid"
 	"github.com/stretchr/testify/require"
 	"github.com/tidwall/gjson"
 )
 
+// compactProbeSSESuccessBody 是原生 v2 压缩成功的最小 SSE 形态：
+// output_item.done 携带 compaction item + response.completed。
+const compactProbeSSESuccessBody = "data: {\"type\":\"response.output_item.done\",\"item\":{\"type\":\"compaction\",\"id\":\"cmp_probe\",\"encrypted_content\":\"blob\"}}\n\n" +
+	"data: {\"type\":\"response.completed\",\"response\":{\"id\":\"resp_probe\",\"output\":[]}}\n\n"
+
 func TestAccountTestService_TestAccountConnection_OpenAICompactOAuthSuccessPersistsSupport(t *testing.T) {
 	gin.SetMode(gin.TestMode)
 
@@ -38,8 +44,8 @@ func TestAccountTestService_TestAccountConnection_OpenAICompactOAuthSuccessPersi
 	}
 	upstream := &httpUpstreamRecorder{resp: &http.Response{
 		StatusCode: http.StatusOK,
-		Header:     http.Header{"Content-Type": []string{"application/json"}, "x-request-id": []string{"rid-probe"}},
-		Body:       io.NopCloser(strings.NewReader(`{"id":"cmp_probe","status":"completed"}`)),
+		Header:     http.Header{"Content-Type": []string{"text/event-stream"}, "x-request-id": []string{"rid-probe"}},
+		Body:       io.NopCloser(strings.NewReader(compactProbeSSESuccessBody)),
 	}}
 	svc := &AccountTestService{
 		accountRepo:  repo,
@@ -53,16 +59,22 @@ func TestAccountTestService_TestAccountConnection_OpenAICompactOAuthSuccessPersi
 	err := svc.TestAccountConnection(c, account.ID, "gpt-5.4", "", AccountTestModeCompact)
 	require.NoError(t, err)
 
-	require.Equal(t, chatgptCodexAPIURL+"/compact", upstream.lastReq.URL.String())
+	// 原生 v2：探测普通 /responses 线，不再打已下线的 /responses/compact。
+	require.Equal(t, chatgptCodexAPIURL, upstream.lastReq.URL.String())
 	require.Equal(t, "chatgpt.com", upstream.lastReq.Host)
-	require.Equal(t, "application/json", upstream.lastReq.Header.Get("Accept"))
-	require.Equal(t, codexCLIVersion, upstream.lastReq.Header.Get("Version"))
+	require.Equal(t, "text/event-stream", upstream.lastReq.Header.Get("Accept"))
+	require.Contains(t, upstream.lastReq.Header.Get("x-codex-beta-features"), "remote_compaction_v2")
 	require.NotEmpty(t, upstream.lastReq.Header.Get("Session_Id"))
 	require.Equal(t, HTTPUpstreamProfileOpenAI, HTTPUpstreamProfileFromContext(upstream.lastReq.Context()))
 	require.Equal(t, codexCLIUserAgent, upstream.lastReq.Header.Get("User-Agent"))
 	require.Equal(t, "chatgpt-acc", upstream.lastReq.Header.Get("chatgpt-account-id"))
 	require.Equal(t, "true", upstream.lastReq.Header.Get("x-openai-fedramp"))
 	require.Equal(t, "gpt-5.4", gjson.GetBytes(upstream.lastBody, "model").String())
+	require.True(t, gjson.GetBytes(upstream.lastBody, "stream").Bool())
+	require.False(t, gjson.GetBytes(upstream.lastBody, "store").Bool())
+	inputItems := gjson.GetBytes(upstream.lastBody, "input").Array()
+	require.NotEmpty(t, inputItems)
+	require.Equal(t, "compaction_trigger", inputItems[len(inputItems)-1].Get("type").String())
 
 	updates := <-updateCalls
 	require.Equal(t, true, updates["openai_compact_supported"])
@@ -114,7 +126,7 @@ func TestAccountTestService_TestAccountConnection_OpenAICompactOAuth404MarksUnsu
 	require.Contains(t, rec.Body.String(), `"type":"error"`)
 }
 
-func TestAccountTestService_TestAccountConnection_OpenAICompactAPIKeyUsesCompactPath(t *testing.T) {
+func TestAccountTestService_TestAccountConnection_OpenAICompactAPIKeyUsesNativeResponsesPath(t *testing.T) {
 	gin.SetMode(gin.TestMode)
 
 	updateCalls := make(chan map[string]any, 1)
@@ -127,8 +139,10 @@ func TestAccountTestService_TestAccountConnection_OpenAICompactAPIKeyUsesCompact
 		Schedulable: true,
 		Concurrency: 1,
 		Credentials: map[string]any{
-			"api_key":               "sk-test",
-			"base_url":              "https://example.com/v1",
+			"api_key":  "sk-test",
+			"base_url": "https://example.com/v1",
+			// post-#5641：compact_model_mapping 仅作用于 legacy /responses/compact，
+			// 原生 v2 探测不应用它。
 			"compact_model_mapping": map[string]any{"gpt-5.4": "gpt-5.4-openai-compact"},
 		},
 	}
@@ -138,8 +152,8 @@ func TestAccountTestService_TestAccountConnection_OpenAICompactAPIKeyUsesCompact
 	}
 	upstream := &httpUpstreamRecorder{resp: &http.Response{
 		StatusCode: http.StatusOK,
-		Header:     http.Header{"Content-Type": []string{"application/json"}},
-		Body:       io.NopCloser(strings.NewReader(`{"id":"cmp_probe_apikey","status":"completed"}`)),
+		Header:     http.Header{"Content-Type": []string{"text/event-stream"}},
+		Body:       io.NopCloser(strings.NewReader(compactProbeSSESuccessBody)),
 	}}
 	svc := &AccountTestService{
 		accountRepo:  repo,
@@ -154,14 +168,16 @@ func TestAccountTestService_TestAccountConnection_OpenAICompactAPIKeyUsesCompact
 	err := svc.TestAccountConnection(c, account.ID, "gpt-5.4", "", AccountTestModeCompact)
 	require.NoError(t, err)
 
-	require.Equal(t, "https://example.com/v1/responses/compact", upstream.lastReq.URL.String())
+	require.Equal(t, "https://example.com/v1/responses", upstream.lastReq.URL.String())
 	requireOpenAICodexProbeHeaders(t, upstream.lastReq.Header)
-	require.Equal(t, "gpt-5.4-openai-compact", gjson.GetBytes(upstream.lastBody, "model").String())
+	require.Contains(t, upstream.lastReq.Header.Get("x-codex-beta-features"), "remote_compaction_v2")
+	require.Equal(t, "gpt-5.4", gjson.GetBytes(upstream.lastBody, "model").String(),
+		"原生 v2 探测不应用 compact_model_mapping")
 	updates := <-updateCalls
 	require.Equal(t, true, updates["openai_compact_supported"])
 }
 
-func TestAccountTestService_TestAccountConnection_OpenAICompactAPIKeyDefaultBaseURLUsesV1Path(t *testing.T) {
+func TestAccountTestService_TestAccountConnection_OpenAICompactAPIKeyDefaultBaseURLUsesResponsesPath(t *testing.T) {
 	gin.SetMode(gin.TestMode)
 
 	updateCalls := make(chan map[string]any, 1)
@@ -183,8 +199,8 @@ func TestAccountTestService_TestAccountConnection_OpenAICompactAPIKeyDefaultBase
 	}
 	upstream := &httpUpstreamRecorder{resp: &http.Response{
 		StatusCode: http.StatusOK,
-		Header:     http.Header{"Content-Type": []string{"application/json"}},
-		Body:       io.NopCloser(strings.NewReader(`{"id":"cmp_probe_apikey_default","status":"completed"}`)),
+		Header:     http.Header{"Content-Type": []string{"text/event-stream"}},
+		Body:       io.NopCloser(strings.NewReader(compactProbeSSESuccessBody)),
 	}}
 	svc := &AccountTestService{
 		accountRepo:  repo,
@@ -198,6 +214,112 @@ func TestAccountTestService_TestAccountConnection_OpenAICompactAPIKeyDefaultBase
 
 	err := svc.TestAccountConnection(c, account.ID, "gpt-5.4", "", AccountTestModeCompact)
 	require.NoError(t, err)
-	require.Equal(t, "https://api.openai.com/v1/responses/compact", upstream.lastReq.URL.String())
+	require.Equal(t, "https://api.openai.com/v1/responses", upstream.lastReq.URL.String())
 	<-updateCalls
 }
+
+func TestAccountTestService_TestAccountConnection_OpenAICompact2xxWithoutItemMarksUnsupported(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+
+	updateCalls := make(chan map[string]any, 1)
+	account := Account{
+		ID:          5,
+		Name:        "openai-oauth-no-item",
+		Platform:    PlatformOpenAI,
+		Type:        AccountTypeOAuth,
+		Status:      StatusActive,
+		Schedulable: true,
+		Concurrency: 1,
+		Credentials: map[string]any{
+			"access_token":       "oauth-token",
+			"chatgpt_account_id": "chatgpt-acc",
+		},
+	}
+	repo := &snapshotUpdateAccountRepo{
+		stubOpenAIAccountRepo: stubOpenAIAccountRepo{accounts: []Account{account}},
+		updateExtraCalls:      updateCalls,
+	}
+	// 200 但流里没有 compaction item：链路吞掉了 compaction_trigger 的形态
+	//（#5478 的 "got 0 items"），必须判定为不支持。
+	noItemBody := "data: {\"type\":\"response.output_item.done\",\"item\":{\"type\":\"message\",\"id\":\"msg_1\"}}\n\n" +
+		"data: {\"type\":\"response.completed\",\"response\":{\"id\":\"resp_x\",\"output\":[]}}\n\n"
+	upstream := &httpUpstreamRecorder{resp: &http.Response{
+		StatusCode: http.StatusOK,
+		Header:     http.Header{"Content-Type": []string{"text/event-stream"}},
+		Body:       io.NopCloser(strings.NewReader(noItemBody)),
+	}}
+	svc := &AccountTestService{
+		accountRepo:  repo,
+		httpUpstream: upstream,
+	}
+
+	rec := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(rec)
+	c.Request = httptest.NewRequest(http.MethodPost, "/api/v1/admin/accounts/5/test", bytes.NewReader(nil))
+
+	err := svc.TestAccountConnection(c, account.ID, "gpt-5.4", "", AccountTestModeCompact)
+	require.Error(t, err)
+
+	updates := <-updateCalls
+	require.Equal(t, false, updates["openai_compact_supported"])
+	require.Contains(t, rec.Body.String(), `"type":"error"`)
+}
+
+// 探测与真实转发走同一 /responses 端点，出站身份必须与真实 Codex 同构：
+// session/thread 为 UUID、携带 x-codex-installation-id（收敛账号用收敛值）。
+func TestAccountTestService_TestAccountConnection_OpenAICompactProbeIdentityMatchesRealTraffic(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+
+	updateCalls := make(chan map[string]any, 1)
+	account := Account{
+		ID:          6,
+		Name:        "openai-oauth-identity",
+		Platform:    PlatformOpenAI,
+		Type:        AccountTypeOAuth,
+		Status:      StatusActive,
+		Schedulable: true,
+		Concurrency: 1,
+		Credentials: map[string]any{
+			"access_token":       "oauth-token",
+			"chatgpt_account_id": "chatgpt-acc",
+		},
+		// 收敛是显式 opt-in（#5610），这里显式开启以验证探测身份与真实流量同构。
+		Extra: map[string]any{"codex_fingerprint_mode": "session"},
+	}
+	repo := &snapshotUpdateAccountRepo{
+		stubOpenAIAccountRepo: stubOpenAIAccountRepo{accounts: []Account{account}},
+		updateExtraCalls:      updateCalls,
+	}
+	upstream := &httpUpstreamRecorder{resp: &http.Response{
+		StatusCode: http.StatusOK,
+		Header:     http.Header{"Content-Type": []string{"text/event-stream"}},
+		Body:       io.NopCloser(strings.NewReader(compactProbeSSESuccessBody)),
+	}}
+	svc := &AccountTestService{accountRepo: repo, httpUpstream: upstream}
+
+	rec := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(rec)
+	c.Request = httptest.NewRequest(http.MethodPost, "/api/v1/admin/accounts/6/test", bytes.NewReader(nil))
+
+	require.NoError(t, svc.TestAccountConnection(c, account.ID, "gpt-5.4", "", AccountTestModeCompact))
+
+	// 显式 session 收敛模式：出站身份 = 账号级收敛值
+	converged := resolveConvergedSessionID(&account)
+	require.Equal(t, converged, upstream.lastReq.Header.Get("session-id"))
+	require.Equal(t, converged, upstream.lastReq.Header.Get("session_id"))
+	require.Equal(t, resolveConvergedInstallationID(&account), upstream.lastReq.Header.Get("x-codex-installation-id"),
+		"真实 Codex 每个请求必带 installation-id，探测不得缺失")
+	require.NotContains(t, upstream.lastReq.Header.Get("session-id"), "probe_compact",
+		"探测标识不得是可被上游一眼识别的字面量")
+	<-updateCalls
+}
+
+func TestCompactProbeSessionID_IsUUIDShaped(t *testing.T) {
+	for _, id := range []int64{0, 1, 987654} {
+		got := compactProbeSessionID(id)
+		_, err := uuid.Parse(got)
+		require.NoError(t, err, "探测会话标识必须是 UUID 形态: %s", got)
+	}
+	require.Equal(t, compactProbeSessionID(7), compactProbeSessionID(7), "同账号应稳定复用同一会话")
+	require.NotEqual(t, compactProbeSessionID(7), compactProbeSessionID(8))
+}

backend/internal/service/openai_agent_identity_compat_test.go

@@ -40,7 +40,7 @@ func TestAccountTestServiceOpenAICompactAgentIdentityUsesFreshAssertion(t *testi
 	upstream := &httpUpstreamRecorder{resp: &http.Response{
 		StatusCode: http.StatusOK,
 		Header:     http.Header{"Content-Type": []string{"application/json"}},
-		Body:       io.NopCloser(strings.NewReader(`{"id":"compact-agent","status":"completed"}`)),
+		Body:       io.NopCloser(strings.NewReader(`{"id":"compact-agent","status":"completed","output":[{"type":"compaction","id":"cmp_agent_fresh","encrypted_content":"blob"}]}`)),
 	}}
 	svc := &AccountTestService{accountRepo: repo, httpUpstream: upstream}
 
@@ -87,7 +87,7 @@ func TestAccountTestServiceOpenAICompactAgentIdentityRecoversInvalidTaskOnce(t *
 
 	upstream := &httpUpstreamRecorder{responses: []*http.Response{
 		{StatusCode: http.StatusUnauthorized, Header: http.Header{"Content-Type": []string{"application/json"}}, Body: io.NopCloser(strings.NewReader(`{"error":{"code":"invalid_task_id"}}`))},
-		{StatusCode: http.StatusOK, Header: http.Header{"Content-Type": []string{"application/json"}}, Body: io.NopCloser(strings.NewReader(`{"id":"compact-agent","status":"completed"}`))},
+		{StatusCode: http.StatusOK, Header: http.Header{"Content-Type": []string{"application/json"}}, Body: io.NopCloser(strings.NewReader(`{"id":"compact-agent","status":"completed","output":[{"type":"compaction","id":"cmp_agent","encrypted_content":"blob"}]}`))},
 	}}
 	invalidator := &agentIdentityWSInvalidationRecorder{}
 	svc := &AccountTestService{accountRepo: repo, httpUpstream: upstream, agentIdentityWS: invalidator}

backend/internal/service/openai_codex_fingerprint.go

@@ -9,9 +9,43 @@ import (
 	"strings"
 	"time"
 
+	"github.com/gin-gonic/gin"
 	"github.com/google/uuid"
+	"github.com/tidwall/gjson"
+	"github.com/tidwall/sjson"
 )
 
+// codexFingerprintIDsContextKey 是暂存在 gin context 的收敛 ID 集合键。
+// 由 Forward（非透传）或 forwardOpenAIPassthrough（透传）解析后写入，请求
+// 构造器读取用于出站头改写——请求体与出站头必须共享同一份 IDs，保证
+// turn_id 等随机字段一致。
+const codexFingerprintIDsContextKey = "codex_fingerprint_ids"
+
+// stageCodexFingerprintIDs 将本 attempt 解析出的收敛 ID 暂存到 gin context。
+// 必须无条件覆写（含 nil）：failover 从收敛账号切到 off 账号时，上一账号的
+// IDs 不得残留并被误应用到新账号的出站头（typed-nil 由应用侧 nil 守卫吸收）。
+func stageCodexFingerprintIDs(c *gin.Context, ids *codexFingerprintIDs) {
+	if c != nil {
+		c.Set(codexFingerprintIDsContextKey, ids)
+	}
+}
+
+// applyStagedCodexFingerprintHeaders 读取 context 暂存的收敛 ID 并改写出站头。
+// 非透传与透传两个请求构造器共用本函数，防止应用语义漂移。仅 OAuth 账号
+// 生效（stale 键在账号类型混合 failover 下由该门挡住）。
+func applyStagedCodexFingerprintHeaders(c *gin.Context, account *Account, h http.Header) {
+	if c == nil || account == nil || account.Type != AccountTypeOAuth {
+		return
+	}
+	value, ok := c.Get(codexFingerprintIDsContextKey)
+	if !ok {
+		return
+	}
+	if ids, ok := value.(*codexFingerprintIDs); ok {
+		applyCodexFingerprintHeaders(h, ids)
+	}
+}
+
 // codexFingerprintMode 控制 OAuth 账号出站请求的设备指纹收敛强度。
 // 多人共享同一 OAuth 账号时，每个用户的 Codex 客户端会携带各自不同的
 // installation_id / session_id / thread_id，上游据此判定设备数和会话数。
@@ -19,7 +53,8 @@ import (
 type codexFingerprintMode string
 
 const (
-	// codexFingerprintOff 不做任何收敛，原样透传客户端标识（默认行为）。
+	// codexFingerprintOff 不做任何收敛，原样透传客户端标识。
+	// 这是默认值：收敛是显式 opt-in 的（见 GetCodexFingerprintMode）。
 	codexFingerprintOff codexFingerprintMode = "off"
 	// codexFingerprintDevice 仅收敛 installation_id 为账号级恒定值。
 	// 上游看到 1 台设备 + 多会话（每用户各自的 session）。
@@ -36,7 +71,16 @@ const (
 const codexFingerprintModeExtraKey = "codex_fingerprint_mode"
 
 // GetCodexFingerprintMode 从账号 extra JSON 读取指纹收敛模式。
-// 未设置时默认 session（设备+会话收敛），显式设为 "off" 才关闭。
+//
+// **收敛是显式 opt-in**：未设置、空值或非法值一律按 off 处理，只有管理员
+// 明确配置 device / session / full 才收敛。
+//
+// 历史：v0.1.175（#5553）把缺省值当作 session，导致升级后存量 OAuth 账号
+// （普遍没有这个 extra 键）的每个非透传请求都被静默改写 installation /
+// session / thread / turn / window 五类标识；#5555、#5556、#5582 报告的额度
+// 缩水都卡在该版本边界，并有"回退 v0.1.173 即恢复"与"新账号开收敛后降额"
+// 的 A/B 实测。上游的配额判定策略不可观测，因此这里取兼容安全的一侧：
+// 不显式 opt-in 就保持 v0.1.175 之前的客户端身份（#5610）。
 func (a *Account) GetCodexFingerprintMode() codexFingerprintMode {
 	if a == nil || !a.IsOpenAIOAuth() {
 		return codexFingerprintOff
@@ -46,7 +90,7 @@ func (a *Account) GetCodexFingerprintMode() codexFingerprintMode {
 	case codexFingerprintOff, codexFingerprintDevice, codexFingerprintSession, codexFingerprintFull:
 		return codexFingerprintMode(raw)
 	default:
-		return codexFingerprintSession
+		return codexFingerprintOff
 	}
 }
 
@@ -245,6 +289,21 @@ func applyCodexFingerprintClientMetadata(reqBody map[string]any, ids *codexFinge
 		existing = make(map[string]any)
 	}
 
+	if !applyCodexFingerprintToClientMetadataMap(existing, ids) {
+		return false
+	}
+	reqBody["client_metadata"] = existing
+	return true
+}
+
+// applyCodexFingerprintToClientMetadataMap 是 client_metadata 改写的共享核心，
+// map 版（非透传，body 已解码）与 raw 字节版（透传热路径）都经由它，保证两条
+// 路径的收敛语义永不漂移。
+func applyCodexFingerprintToClientMetadataMap(existing map[string]any, ids *codexFingerprintIDs) bool {
+	if existing == nil || ids == nil {
+		return false
+	}
+
 	modified := false
 
 	if ids.installationID != "" {
@@ -256,9 +315,6 @@ func applyCodexFingerprintClientMetadata(reqBody map[string]any, ids *codexFinge
 		rewriteClientMetadataEmbeddedTurnMetadata(existing, map[string]any{
 			"installation_id": ids.installationID,
 		})
-		if modified {
-			reqBody["client_metadata"] = existing
-		}
 		return modified
 	}
 
@@ -276,11 +332,47 @@ func applyCodexFingerprintClientMetadata(reqBody map[string]any, ids *codexFinge
 		"window_id":               ids.windowID,
 		"turn_started_at_unix_ms": time.Now().UnixMilli(),
 	})
-
-	reqBody["client_metadata"] = existing
 	return true
 }
 
+// applyCodexFingerprintClientMetadataRaw 在原始 JSON 字节上改写 client_metadata，
+// 供透传路径使用——透传是热路径，禁止对可能高达数十 MB 的 body 做全量
+// Unmarshal（见 forwardOpenAIPassthrough 的轻量提取注释）。实现为：gjson 提取
+// client_metadata 小对象单独解码，经共享核心改写后 sjson 一次性拼回，body
+// 其余字节原样保留。语义与 applyCodexFingerprintClientMetadata 逐点一致
+// （含"非对象值整体替换为收敛集合"的行为）。
+func applyCodexFingerprintClientMetadataRaw(body []byte, ids *codexFingerprintIDs) ([]byte, bool, error) {
+	if len(body) == 0 || ids == nil {
+		return body, false, nil
+	}
+	// 非 JSON 对象的 body（数组/标量/畸形）没有 client_metadata 语义，
+	// sjson 在这类根上写字段会改写整体结构，直接放行保持原样。
+	if !gjson.ParseBytes(body).IsObject() {
+		return body, false, nil
+	}
+
+	existing := map[string]any{}
+	if cm := gjson.GetBytes(body, "client_metadata"); cm.IsObject() {
+		if err := json.Unmarshal([]byte(cm.Raw), &existing); err != nil {
+			return body, false, fmt.Errorf("decode client_metadata for fingerprint: %w", err)
+		}
+	}
+
+	if !applyCodexFingerprintToClientMetadataMap(existing, ids) {
+		return body, false, nil
+	}
+
+	raw, err := json.Marshal(existing)
+	if err != nil {
+		return body, false, fmt.Errorf("encode converged client_metadata: %w", err)
+	}
+	next, err := sjson.SetRawBytes(body, "client_metadata", raw)
+	if err != nil {
+		return body, false, fmt.Errorf("splice converged client_metadata: %w", err)
+	}
+	return next, true, nil
+}
+
 // rewriteClientMetadataEmbeddedTurnMetadata 改写 client_metadata 中内嵌的
 // x-codex-turn-metadata JSON 字符串里的指定字段。
 func rewriteClientMetadataEmbeddedTurnMetadata(clientMetadata map[string]any, fields map[string]any) {

backend/internal/service/openai_codex_fingerprint_test.go

@@ -1,10 +1,13 @@
 package service
 
 import (
+	"context"
 	"encoding/json"
 	"net/http"
+	"net/http/httptest"
 	"testing"
 
+	"github.com/gin-gonic/gin"
 	"github.com/google/uuid"
 	"github.com/stretchr/testify/assert"
 	"github.com/stretchr/testify/require"
@@ -51,9 +54,11 @@ func TestGetCodexFingerprintMode(t *testing.T) {
 	}{
 		{"nil 账号", nil, codexFingerprintOff},
 		{"非 OAuth 账号", &Account{Platform: PlatformOpenAI, Type: "api_key"}, codexFingerprintOff},
-		{"无 extra 默认 session", newTestOAuthAccount(1, nil), codexFingerprintSession},
-		{"空值默认 session", newTestOAuthAccount(1, map[string]any{codexFingerprintModeExtraKey: ""}), codexFingerprintSession},
-		{"非法值默认 session", newTestOAuthAccount(1, map[string]any{codexFingerprintModeExtraKey: "invalid"}), codexFingerprintSession},
+		// 收敛是显式 opt-in：缺省/空/非法一律 off（#5610）。存量账号普遍没有这个
+		// extra 键，升级不得把它们静默切进收敛。
+		{"无 extra 默认 off", newTestOAuthAccount(1, nil), codexFingerprintOff},
+		{"空值默认 off", newTestOAuthAccount(1, map[string]any{codexFingerprintModeExtraKey: ""}), codexFingerprintOff},
+		{"非法值默认 off", newTestOAuthAccount(1, map[string]any{codexFingerprintModeExtraKey: "invalid"}), codexFingerprintOff},
 		{"显式 off", newTestOAuthAccount(1, map[string]any{codexFingerprintModeExtraKey: "off"}), codexFingerprintOff},
 		{"device", newTestOAuthAccount(1, map[string]any{codexFingerprintModeExtraKey: "device"}), codexFingerprintDevice},
 		{"session", newTestOAuthAccount(1, map[string]any{codexFingerprintModeExtraKey: "session"}), codexFingerprintSession},
@@ -116,13 +121,24 @@ func TestResolveCodexFingerprintIDsFromRequest_ExplicitOff(t *testing.T) {
 	assert.Nil(t, ids, "显式 off 模式应返回 nil")
 }
 
-func TestResolveCodexFingerprintIDsFromRequest_DefaultIsSession(t *testing.T) {
+// 未显式配置的存量账号不得被收敛（#5610）：默认返回 nil，出站身份保持
+// v0.1.175 之前的客户端原值。
+func TestResolveCodexFingerprintIDsFromRequest_DefaultIsOff(t *testing.T) {
 	account := newTestOAuthAccount(1, nil)
-	ids := resolveCodexFingerprintIDsFromRequest(account, nil)
-	require.NotNil(t, ids, "无 extra 默认 session 模式，应返回非 nil")
-	assert.Equal(t, codexFingerprintSession, ids.mode)
-	assert.NotEmpty(t, ids.sessionID)
-	assert.NotEmpty(t, ids.turnID)
+	assert.Nil(t, resolveCodexFingerprintIDsFromRequest(account, nil), "无 extra 应视为 off")
+}
+
+// 管理员显式 opt-in 的账号行为不变。
+func TestResolveCodexFingerprintIDsFromRequest_ExplicitOptInHonored(t *testing.T) {
+	for _, mode := range []string{"device", "session", "full"} {
+		t.Run(mode, func(t *testing.T) {
+			account := newTestOAuthAccount(1, map[string]any{codexFingerprintModeExtraKey: mode})
+			ids := resolveCodexFingerprintIDsFromRequest(account, nil)
+			require.NotNil(t, ids, "显式配置必须生效")
+			assert.Equal(t, codexFingerprintMode(mode), ids.mode)
+			assert.NotEmpty(t, ids.installationID)
+		})
+	}
 }
 
 // --- applyCodexFingerprintHeaders: off 模式 ---
@@ -458,3 +474,188 @@ func TestExtractClientSessionID(t *testing.T) {
 		})
 	}
 }
+
+// --- 透传路径：raw 字节版 client_metadata 改写 ---
+
+// rawVsMapClientMetadata 用同一份 ids 分别跑 map 版与 raw 字节版，
+// 返回两侧最终的 client_metadata 解码结果。
+func rawVsMapClientMetadata(t *testing.T, body []byte, ids *codexFingerprintIDs) (map[string]any, map[string]any) {
+	t.Helper()
+
+	var decoded map[string]any
+	require.NoError(t, json.Unmarshal(body, &decoded))
+	applyCodexFingerprintClientMetadata(decoded, ids)
+	mapCM, _ := decoded["client_metadata"].(map[string]any)
+
+	rawBody, changed, err := applyCodexFingerprintClientMetadataRaw(body, ids)
+	require.NoError(t, err)
+	require.True(t, changed)
+	var rawDecoded map[string]any
+	require.NoError(t, json.Unmarshal(rawBody, &rawDecoded))
+	rawCM, _ := rawDecoded["client_metadata"].(map[string]any)
+	return mapCM, rawCM
+}
+
+func TestApplyCodexFingerprintClientMetadataRaw_MatchesMapVariant(t *testing.T) {
+	embedded := `{\"installation_id\":\"real-install\",\"session_id\":\"real-session\",\"sandbox\":\"seatbelt\"}`
+	bodies := map[string]string{
+		"no_client_metadata": `{"model":"gpt-5.6-sol","input":[],"stream":true}`,
+		"object_with_extras": `{"model":"gpt-5.6-sol","client_metadata":{"session_id":"client-session","traceparent":"00-abc-def-01","x-codex-turn-metadata":"` + embedded + `"},"stream":true}`,
+		"non_object_value":   `{"model":"gpt-5.6-sol","client_metadata":"bogus","stream":true}`,
+	}
+	for _, mode := range []codexFingerprintMode{codexFingerprintDevice, codexFingerprintSession, codexFingerprintFull} {
+		account := newTestOAuthAccount(4242, nil)
+		ids := resolveCodexFingerprintIDs(account, "client-sess-raw", mode)
+		require.NotNil(t, ids)
+		for name, body := range bodies {
+			t.Run(string(mode)+"/"+name, func(t *testing.T) {
+				mapCM, rawCM := rawVsMapClientMetadata(t, []byte(body), ids)
+				assert.Equal(t, mapCM, rawCM, "raw 字节版与 map 版的 client_metadata 结果必须逐点一致")
+			})
+		}
+	}
+}
+
+func TestApplyCodexFingerprintClientMetadataRaw_PreservesUnrelatedFields(t *testing.T) {
+	account := newTestOAuthAccount(4243, nil)
+	ids := resolveCodexFingerprintIDs(account, "client-sess-preserve", codexFingerprintSession)
+	require.NotNil(t, ids)
+
+	body := []byte(`{"model":"gpt-5.6-sol","input":[{"type":"message","role":"user","content":"hi"}],"stream":true,"prompt_cache_key":"pck-1"}`)
+	out, changed, err := applyCodexFingerprintClientMetadataRaw(body, ids)
+	require.NoError(t, err)
+	require.True(t, changed)
+
+	var decoded map[string]any
+	require.NoError(t, json.Unmarshal(out, &decoded))
+	assert.Equal(t, "gpt-5.6-sol", decoded["model"])
+	assert.Equal(t, "pck-1", decoded["prompt_cache_key"])
+	assert.Equal(t, true, decoded["stream"])
+	cm, _ := decoded["client_metadata"].(map[string]any)
+	require.NotNil(t, cm)
+	assert.Equal(t, ids.sessionID, cm["session_id"])
+	assert.Equal(t, ids.turnID, cm["turn_id"])
+}
+
+func TestApplyCodexFingerprintClientMetadataRaw_Noop(t *testing.T) {
+	body := []byte(`{"model":"gpt-5.6-sol"}`)
+	out, changed, err := applyCodexFingerprintClientMetadataRaw(body, nil)
+	require.NoError(t, err)
+	assert.False(t, changed)
+	assert.Equal(t, body, out)
+
+	out, changed, err = applyCodexFingerprintClientMetadataRaw(nil, &codexFingerprintIDs{mode: codexFingerprintSession, installationID: "x"})
+	require.NoError(t, err)
+	assert.False(t, changed)
+	assert.Nil(t, out)
+}
+
+// --- context 暂存与出站头应用（透传/非透传共用 seam）---
+
+func newFingerprintStageTestContext(t *testing.T) *gin.Context {
+	t.Helper()
+	gin.SetMode(gin.TestMode)
+	c, _ := gin.CreateTestContext(httptest.NewRecorder())
+	c.Request = httptest.NewRequest(http.MethodPost, "/v1/responses", nil)
+	return c
+}
+
+func TestStageCodexFingerprintIDs_NilOverwritesPreviousAccount(t *testing.T) {
+	c := newFingerprintStageTestContext(t)
+	accountA := newTestOAuthAccount(1001, nil)
+	idsA := resolveCodexFingerprintIDs(accountA, "sess-x", codexFingerprintSession)
+	require.NotNil(t, idsA)
+	stageCodexFingerprintIDs(c, idsA)
+
+	// failover 切到 off 模式账号：无条件覆写为 nil，上一账号 IDs 不得残留
+	stageCodexFingerprintIDs(c, nil)
+
+	h := http.Header{}
+	h.Set("session_id", "isolated-session")
+	accountB := newTestOAuthAccount(1002, map[string]any{"codex_fingerprint_mode": "off"})
+	applyStagedCodexFingerprintHeaders(c, accountB, h)
+	assert.Equal(t, "isolated-session", h.Get("session_id"), "off 账号不得应用上一账号的收敛 ID")
+	assert.Empty(t, h.Get("x-codex-installation-id"))
+}
+
+func TestApplyStagedCodexFingerprintHeaders_SkipsNonOAuthAccount(t *testing.T) {
+	c := newFingerprintStageTestContext(t)
+	oauthIDs := resolveCodexFingerprintIDs(newTestOAuthAccount(1003, nil), "sess-y", codexFingerprintSession)
+	require.NotNil(t, oauthIDs)
+	stageCodexFingerprintIDs(c, oauthIDs)
+
+	h := http.Header{}
+	apiKeyAccount := &Account{ID: 1004, Platform: PlatformOpenAI, Type: AccountTypeAPIKey}
+	applyStagedCodexFingerprintHeaders(c, apiKeyAccount, h)
+	assert.Empty(t, h.Get("x-codex-installation-id"), "stale 收敛 ID 不得应用到非 OAuth 账号")
+}
+
+func TestBuildUpstreamRequestOpenAIPassthrough_AppliesStagedFingerprint(t *testing.T) {
+	svc := &OpenAIGatewayService{}
+	// 收敛是显式 opt-in（#5610）：显式开启后验证透传路径的出站头收敛。
+	account := newTestOAuthAccount(2001, map[string]any{
+		"openai_oauth_passthrough": true,
+		"codex_fingerprint_mode":   "session",
+	})
+
+	c := newFingerprintStageTestContext(t)
+	c.Request.Header.Set("session_id", "real-client-session")
+	c.Request.Header.Set("User-Agent", "codex_cli_rs/0.144.1 (Ubuntu 22.4.0; x86_64) xterm-256color")
+	c.Request.Header.Set("originator", "codex_cli_rs")
+	c.Request.Header.Set("x-codex-turn-metadata", `{"installation_id":"real-install","session_id":"real-session","sandbox":"seatbelt"}`)
+
+	// 复刻 forwardOpenAIPassthrough 的解析+暂存 seam（默认 session 模式）
+	ids := resolveCodexFingerprintIDsFromRequest(account, c.Request.Header)
+	require.NotNil(t, ids)
+	stageCodexFingerprintIDs(c, ids)
+
+	body := []byte(`{"model":"gpt-5.6-sol","input":[],"stream":true}`)
+	req, err := svc.buildUpstreamRequestOpenAIPassthrough(context.Background(), c, account, body, "test-token")
+	require.NoError(t, err)
+
+	assert.Equal(t, ids.sessionID, req.Header.Get("session_id"), "session 模式下出站 session_id 应为账号级收敛值")
+	assert.Equal(t, ids.installationID, req.Header.Get("x-codex-installation-id"))
+	assert.Equal(t, ids.windowID, req.Header.Get("x-codex-window-id"))
+	assert.Equal(t, ids.threadID, req.Header.Get("x-client-request-id"))
+	turnMetadata := req.Header.Get("x-codex-turn-metadata")
+	require.NotEmpty(t, turnMetadata)
+	assert.Contains(t, turnMetadata, ids.sessionID, "turn-metadata JSON 中的 session_id 应被收敛")
+	assert.Contains(t, turnMetadata, `"sandbox":"seatbelt"`, "turn-metadata 未指定字段应原样保留")
+}
+
+func TestBuildUpstreamRequestOpenAIPassthrough_OffModeKeepsIsolatedSession(t *testing.T) {
+	svc := &OpenAIGatewayService{}
+	account := newTestOAuthAccount(2002, map[string]any{
+		"openai_oauth_passthrough": true,
+		"codex_fingerprint_mode":   "off",
+	})
+
+	c := newFingerprintStageTestContext(t)
+	c.Request.Header.Set("session_id", "real-client-session")
+	c.Request.Header.Set("originator", "codex_cli_rs")
+
+	ids := resolveCodexFingerprintIDsFromRequest(account, c.Request.Header)
+	require.Nil(t, ids)
+	stageCodexFingerprintIDs(c, ids)
+
+	body := []byte(`{"model":"gpt-5.6-sol","input":[],"stream":true}`)
+	req, err := svc.buildUpstreamRequestOpenAIPassthrough(context.Background(), c, account, body, "test-token")
+	require.NoError(t, err)
+
+	assert.NotEmpty(t, req.Header.Get("session_id"))
+	assert.NotEqual(t, resolveConvergedSessionID(account), req.Header.Get("session_id"), "off 模式不得收敛 session_id")
+	assert.Empty(t, req.Header.Get("x-codex-window-id"))
+}
+
+func TestApplyCodexFingerprintClientMetadataRaw_NonObjectBodyUntouched(t *testing.T) {
+	account := newTestOAuthAccount(4244, nil)
+	ids := resolveCodexFingerprintIDs(account, "client-sess-nonobj", codexFingerprintSession)
+	require.NotNil(t, ids)
+
+	for _, body := range []string{`[1,2,3]`, `"plain string"`, `not json at all`} {
+		out, changed, err := applyCodexFingerprintClientMetadataRaw([]byte(body), ids)
+		require.NoError(t, err)
+		assert.False(t, changed, "非 JSON 对象 body 不应被改写: %s", body)
+		assert.Equal(t, []byte(body), out)
+	}
+}

backend/internal/service/openai_codex_turn_state.go

@@ -0,0 +1,165 @@
+package service
+
+import (
+	"net/http"
+	"strconv"
+	"strings"
+	"time"
+
+	"github.com/gin-gonic/gin"
+)
+
+// openAICodexTurnStateHeader 是 Codex 的回合状态头。上游在响应头中铸造该
+// 不透明 blob，客户端在同一回合的后续请求中原样回带（codex-rs 侧从
+// /responses SSE、/responses/compact JSON 与 WS 握手三种响应中捕获，见
+// codex-api/src/sse/responses.rs 与 endpoint/compact.rs）。
+const openAICodexTurnStateHeader = "x-codex-turn-state"
+
+// turn-state blob 是上游在"出站身份"（含 #5553 指纹收敛改写后的
+// installation/session/thread 标识）下铸造的，同账号回放自洽；跨账号回放
+// （failover 换号后客户端仍回带旧账号的 blob）是代理链独有、真实 Codex
+// 永远不会产生的矛盾信号。溯源表记录每个下游会话最近一次铸造该 blob 的
+// 账号，出站守卫据此剥离已知异账号的回带值。
+type openAICodexTurnStateOrigin struct {
+	accountID int64
+	expiresAt time.Time
+}
+
+// openAICodexTurnStateSeed 返回溯源表键：API Key + 客户端原始会话标识。
+// 客户端会话标识取自请求头（与指纹收敛的 thread 派生同源，见
+// extractClientSessionID），确保同一下游会话的记录/守卫两侧使用同一键。
+// 无会话标识时返回空串，表示不做跟踪（保持透传现状）。
+func openAICodexTurnStateSeed(c *gin.Context) string {
+	if c == nil || c.Request == nil {
+		return ""
+	}
+	sessionID := extractClientSessionID(c.Request.Header)
+	if sessionID == "" {
+		return ""
+	}
+	return strconv.FormatInt(getAPIKeyIDFromContext(c), 10) + "\x00" + sessionID
+}
+
+// relayOpenAICodexTurnState 将上游响应中的 turn-state 显式写入下游响应头，
+// 并记录铸造账号。必须在响应头提交点调用（WriteHeader 之前、且确认本次
+// 上游响应就是将要写回客户端的响应之后）。上游无该头时主动清除 writer 上
+// 可能残留的上一 failover attempt 的值——否则换号后旧账号的 blob 会粘到
+// 新账号的响应上，这正是本文件要防止的跨账号矛盾。
+func (s *OpenAIGatewayService) relayOpenAICodexTurnState(c *gin.Context, account *Account, upstream http.Header) {
+	if c == nil || c.Writer == nil {
+		return
+	}
+	canonical := http.CanonicalHeaderKey(openAICodexTurnStateHeader)
+	state := extractOpenAICodexTurnState(upstream)
+	if state == "" {
+		c.Writer.Header().Del(canonical)
+		return
+	}
+	c.Writer.Header().Set(canonical, state)
+	s.noteOpenAICodexTurnStateProvenance(c, account)
+}
+
+// stageOpenAICodexTurnState 将上游 turn-state 暂存到延迟提交的响应头集合
+// （首输出守卫路径先缓存头、见到首个输出事件才提交）。此处**不**记录铸造
+// 账号：该 attempt 仍可能在首输出超时后 failover，暂存头会被整体丢弃，
+// 客户端从未收到该 blob。溯源必须在真正提交时记录，见
+// noteStagedOpenAICodexTurnStateCommitted。
+func stageOpenAICodexTurnState(dst *http.Header, upstream http.Header) {
+	if dst == nil {
+		return
+	}
+	canonical := http.CanonicalHeaderKey(openAICodexTurnStateHeader)
+	state := extractOpenAICodexTurnState(upstream)
+	if state == "" {
+		if *dst != nil {
+			dst.Del(canonical)
+		}
+		return
+	}
+	if *dst == nil {
+		*dst = http.Header{}
+	}
+	dst.Set(canonical, state)
+}
+
+// noteStagedOpenAICodexTurnStateCommitted 在暂存响应头真正写入下游时记录
+// 铸造账号——只有此刻客户端才确定收到了该 blob，溯源表才与客户端持有的
+// 值一致（否则被 failover 丢弃的 attempt 会污染溯源，导致后续误剥离）。
+func (s *OpenAIGatewayService) noteStagedOpenAICodexTurnStateCommitted(c *gin.Context, account *Account, staged http.Header) {
+	if staged == nil || strings.TrimSpace(staged.Get(openAICodexTurnStateHeader)) == "" {
+		return
+	}
+	s.noteOpenAICodexTurnStateProvenance(c, account)
+}
+
+func extractOpenAICodexTurnState(upstream http.Header) string {
+	if upstream == nil {
+		return ""
+	}
+	return strings.TrimSpace(upstream.Get(openAICodexTurnStateHeader))
+}
+
+// noteOpenAICodexTurnStateProvenance 记录（下游会话 → 铸造账号）。
+func (s *OpenAIGatewayService) noteOpenAICodexTurnStateProvenance(c *gin.Context, account *Account) {
+	if s == nil || account == nil || account.ID <= 0 {
+		return
+	}
+	seed := openAICodexTurnStateSeed(c)
+	if seed == "" {
+		return
+	}
+	s.openaiCodexTurnStateOrigins.Store(seed, openAICodexTurnStateOrigin{
+		accountID: account.ID,
+		expiresAt: time.Now().Add(s.openAIWSSessionStickyTTL()),
+	})
+	s.sweepOpenAICodexTurnStateOrigins()
+}
+
+// guardOpenAICodexTurnStateEcho 出站守卫：客户端回带的 turn-state 若已知由
+// 其他账号铸造则剥离，同账号或无溯源记录时保持原样。只剥离、不注入——
+// /responses 路径的客户端是真实 Codex，会按自身回合语义自行回带；服务端
+// 注入是 Claude 兼容桥（无法回带的客户端）的专属行为。
+func (s *OpenAIGatewayService) guardOpenAICodexTurnStateEcho(c *gin.Context, account *Account, h http.Header) {
+	if s == nil || h == nil || account == nil {
+		return
+	}
+	if strings.TrimSpace(h.Get(openAICodexTurnStateHeader)) == "" {
+		return
+	}
+	seed := openAICodexTurnStateSeed(c)
+	if seed == "" {
+		return
+	}
+	raw, ok := s.openaiCodexTurnStateOrigins.Load(seed)
+	if !ok {
+		return
+	}
+	origin, ok := raw.(openAICodexTurnStateOrigin)
+	if !ok {
+		s.openaiCodexTurnStateOrigins.Delete(seed)
+		return
+	}
+	if !origin.expiresAt.IsZero() && time.Now().After(origin.expiresAt) {
+		s.openaiCodexTurnStateOrigins.Delete(seed)
+		return
+	}
+	if origin.accountID != account.ID {
+		h.Del(openAICodexTurnStateHeader)
+	}
+}
+
+// sweepOpenAICodexTurnStateOrigins 机会式清扫过期溯源记录：每 256 次写入
+// 全量遍历一轮，防止仅靠读侧惰性删除导致的慢泄漏（会话键无上界）。
+func (s *OpenAIGatewayService) sweepOpenAICodexTurnStateOrigins() {
+	if s.openaiCodexTurnStateWrites.Add(1)%256 != 0 {
+		return
+	}
+	now := time.Now()
+	s.openaiCodexTurnStateOrigins.Range(func(key, value any) bool {
+		origin, ok := value.(openAICodexTurnStateOrigin)
+		if !ok || (!origin.expiresAt.IsZero() && now.After(origin.expiresAt)) {
+			s.openaiCodexTurnStateOrigins.Delete(key)
+		}
+		return true
+	})
+}

backend/internal/service/openai_codex_turn_state_test.go

@@ -0,0 +1,381 @@
+package service
+
+import (
+	"context"
+	"net/http"
+	"net/http/httptest"
+	"testing"
+	"time"
+
+	"github.com/gin-gonic/gin"
+	"github.com/stretchr/testify/require"
+)
+
+func newTurnStateTestContext(t *testing.T, apiKeyID int64, sessionID string) (*gin.Context, *httptest.ResponseRecorder) {
+	t.Helper()
+	gin.SetMode(gin.TestMode)
+	rec := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(rec)
+	c.Request = httptest.NewRequest(http.MethodPost, "/v1/responses", nil)
+	if sessionID != "" {
+		c.Request.Header.Set("session_id", sessionID)
+	}
+	if apiKeyID > 0 {
+		c.Set("api_key", &APIKey{ID: apiKeyID})
+	}
+	return c, rec
+}
+
+func TestOpenAICodexTurnStateSeed(t *testing.T) {
+	c, _ := newTurnStateTestContext(t, 7, "sess-1")
+	require.Equal(t, "7\x00sess-1", openAICodexTurnStateSeed(c))
+
+	// 连字符形式优先（Codex CLI 标准头）
+	c.Request.Header.Set("session-id", "sess-hyphen")
+	require.Equal(t, "7\x00sess-hyphen", openAICodexTurnStateSeed(c))
+
+	// 无会话标识 → 不跟踪
+	cNoSession, _ := newTurnStateTestContext(t, 7, "")
+	require.Empty(t, openAICodexTurnStateSeed(cNoSession))
+
+	require.Empty(t, openAICodexTurnStateSeed(nil))
+}
+
+func TestRelayOpenAICodexTurnState_SetsHeaderAndRecordsProvenance(t *testing.T) {
+	svc := &OpenAIGatewayService{}
+	account := &Account{ID: 42}
+	c, _ := newTurnStateTestContext(t, 7, "sess-relay")
+
+	upstream := http.Header{}
+	upstream.Set("x-codex-turn-state", "blob-A")
+	svc.relayOpenAICodexTurnState(c, account, upstream)
+
+	require.Equal(t, "blob-A", c.Writer.Header().Get("X-Codex-Turn-State"))
+
+	raw, ok := svc.openaiCodexTurnStateOrigins.Load("7\x00sess-relay")
+	require.True(t, ok)
+	origin, ok := raw.(openAICodexTurnStateOrigin)
+	require.True(t, ok)
+	require.Equal(t, int64(42), origin.accountID)
+	require.True(t, origin.expiresAt.After(time.Now()))
+}
+
+func TestRelayOpenAICodexTurnState_ClearsStaleValueWhenUpstreamAbsent(t *testing.T) {
+	svc := &OpenAIGatewayService{}
+	c, _ := newTurnStateTestContext(t, 7, "sess-stale")
+	// 模拟上一 failover attempt 残留的值
+	c.Writer.Header().Set("X-Codex-Turn-State", "blob-old")
+
+	svc.relayOpenAICodexTurnState(c, &Account{ID: 43}, http.Header{})
+
+	require.Empty(t, c.Writer.Header().Get("X-Codex-Turn-State"))
+	_, ok := svc.openaiCodexTurnStateOrigins.Load("7\x00sess-stale")
+	require.False(t, ok)
+}
+
+func TestStageOpenAICodexTurnState_StagedHeaders(t *testing.T) {
+	svc := &OpenAIGatewayService{}
+	c, _ := newTurnStateTestContext(t, 9, "sess-staged")
+
+	// nil 集合 + 上游有值 → 创建集合并写入，但此刻还不记录溯源
+	var staged http.Header
+	upstream := http.Header{}
+	upstream.Set("x-codex-turn-state", "blob-B")
+	stageOpenAICodexTurnState(&staged, upstream)
+	require.NotNil(t, staged)
+	require.Equal(t, "blob-B", staged.Get("X-Codex-Turn-State"))
+	_, noted := svc.openaiCodexTurnStateOrigins.Load("9\x00sess-staged")
+	require.False(t, noted, "暂存阶段不得记录溯源：该 attempt 仍可能 failover 丢弃")
+
+	// 真正提交时才记录
+	svc.noteStagedOpenAICodexTurnStateCommitted(c, &Account{ID: 44}, staged)
+	raw, ok := svc.openaiCodexTurnStateOrigins.Load("9\x00sess-staged")
+	require.True(t, ok)
+	origin, ok := raw.(openAICodexTurnStateOrigin)
+	require.True(t, ok)
+	require.Equal(t, int64(44), origin.accountID)
+
+	// 上游无值 → 清除已暂存的值；nil 集合保持 nil
+	stageOpenAICodexTurnState(&staged, http.Header{})
+	require.Empty(t, staged.Get("X-Codex-Turn-State"))
+	var nilStaged http.Header
+	stageOpenAICodexTurnState(&nilStaged, http.Header{})
+	require.Nil(t, nilStaged)
+}
+
+// 首输出超时导致 attempt 被丢弃时，溯源不得被该 attempt 污染——否则后续
+// 请求会把客户端持有的合法 blob 误判成跨账号回带而剥离。
+func TestStagedTurnState_AbandonedAttemptDoesNotPoisonProvenance(t *testing.T) {
+	svc := &OpenAIGatewayService{}
+	c, _ := newTurnStateTestContext(t, 11, "sess-abandoned")
+
+	// 账号 A 的 attempt 暂存了 blob，但从未提交（首输出超时 → failover）
+	var staged http.Header
+	upstreamA := http.Header{}
+	upstreamA.Set("x-codex-turn-state", "blob-A")
+	stageOpenAICodexTurnState(&staged, upstreamA)
+
+	// 账号 B 接手并真正提交
+	svc.relayOpenAICodexTurnState(c, &Account{ID: 52}, upstreamA)
+
+	// 客户端回带的 blob 来自 B，出站到 B 时不得被剥离
+	h := http.Header{}
+	h.Set("x-codex-turn-state", "blob-A")
+	svc.guardOpenAICodexTurnStateEcho(c, &Account{ID: 52}, h)
+	require.Equal(t, "blob-A", h.Get("x-codex-turn-state"))
+
+	raw, ok := svc.openaiCodexTurnStateOrigins.Load("11\x00sess-abandoned")
+	require.True(t, ok)
+	origin, ok := raw.(openAICodexTurnStateOrigin)
+	require.True(t, ok)
+	require.Equal(t, int64(52), origin.accountID)
+}
+
+func TestNoteStagedOpenAICodexTurnStateCommitted_NoopWithoutState(t *testing.T) {
+	svc := &OpenAIGatewayService{}
+	c, _ := newTurnStateTestContext(t, 12, "sess-nostate")
+
+	svc.noteStagedOpenAICodexTurnStateCommitted(c, &Account{ID: 60}, nil)
+	svc.noteStagedOpenAICodexTurnStateCommitted(c, &Account{ID: 60}, http.Header{"X-Request-Id": []string{"rid"}})
+
+	_, ok := svc.openaiCodexTurnStateOrigins.Load("12\x00sess-nostate")
+	require.False(t, ok)
+}
+
+func TestGuardOpenAICodexTurnStateEcho(t *testing.T) {
+	newOutbound := func(state string) http.Header {
+		h := http.Header{}
+		if state != "" {
+			h.Set("x-codex-turn-state", state)
+		}
+		return h
+	}
+
+	t.Run("same_account_keeps_echo", func(t *testing.T) {
+		svc := &OpenAIGatewayService{}
+		c, _ := newTurnStateTestContext(t, 7, "sess-g1")
+		upstream := http.Header{}
+		upstream.Set("x-codex-turn-state", "blob-A")
+		svc.relayOpenAICodexTurnState(c, &Account{ID: 42}, upstream)
+
+		h := newOutbound("blob-A")
+		svc.guardOpenAICodexTurnStateEcho(c, &Account{ID: 42}, h)
+		require.Equal(t, "blob-A", h.Get("x-codex-turn-state"))
+	})
+
+	t.Run("foreign_account_strips_echo", func(t *testing.T) {
+		svc := &OpenAIGatewayService{}
+		c, _ := newTurnStateTestContext(t, 7, "sess-g2")
+		upstream := http.Header{}
+		upstream.Set("x-codex-turn-state", "blob-A")
+		svc.relayOpenAICodexTurnState(c, &Account{ID: 42}, upstream)
+
+		// failover 换到账号 43：blob 由 42 铸造，必须剥离
+		h := newOutbound("blob-A")
+		svc.guardOpenAICodexTurnStateEcho(c, &Account{ID: 43}, h)
+		require.Empty(t, h.Get("x-codex-turn-state"))
+	})
+
+	t.Run("no_provenance_passthrough", func(t *testing.T) {
+		svc := &OpenAIGatewayService{}
+		c, _ := newTurnStateTestContext(t, 7, "sess-g3")
+		h := newOutbound("blob-unknown")
+		svc.guardOpenAICodexTurnStateEcho(c, &Account{ID: 43}, h)
+		require.Equal(t, "blob-unknown", h.Get("x-codex-turn-state"))
+	})
+
+	t.Run("expired_provenance_passthrough_and_pruned", func(t *testing.T) {
+		svc := &OpenAIGatewayService{}
+		c, _ := newTurnStateTestContext(t, 7, "sess-g4")
+		svc.openaiCodexTurnStateOrigins.Store("7\x00sess-g4", openAICodexTurnStateOrigin{
+			accountID: 42,
+			expiresAt: time.Now().Add(-time.Minute),
+		})
+		h := newOutbound("blob-A")
+		svc.guardOpenAICodexTurnStateEcho(c, &Account{ID: 43}, h)
+		require.Equal(t, "blob-A", h.Get("x-codex-turn-state"))
+		_, ok := svc.openaiCodexTurnStateOrigins.Load("7\x00sess-g4")
+		require.False(t, ok)
+	})
+
+	t.Run("no_session_seed_noop", func(t *testing.T) {
+		svc := &OpenAIGatewayService{}
+		c, _ := newTurnStateTestContext(t, 7, "")
+		h := newOutbound("blob-A")
+		svc.guardOpenAICodexTurnStateEcho(c, &Account{ID: 43}, h)
+		require.Equal(t, "blob-A", h.Get("x-codex-turn-state"))
+	})
+
+	t.Run("no_echo_noop", func(t *testing.T) {
+		svc := &OpenAIGatewayService{}
+		c, _ := newTurnStateTestContext(t, 7, "sess-g5")
+		h := newOutbound("")
+		svc.guardOpenAICodexTurnStateEcho(c, &Account{ID: 43}, h)
+		require.Empty(t, h.Get("x-codex-turn-state"))
+	})
+}
+
+func TestSweepOpenAICodexTurnStateOrigins_PrunesExpiredEntries(t *testing.T) {
+	svc := &OpenAIGatewayService{}
+	svc.openaiCodexTurnStateOrigins.Store("expired", openAICodexTurnStateOrigin{
+		accountID: 1,
+		expiresAt: time.Now().Add(-time.Minute),
+	})
+	svc.openaiCodexTurnStateOrigins.Store("alive", openAICodexTurnStateOrigin{
+		accountID: 2,
+		expiresAt: time.Now().Add(time.Hour),
+	})
+
+	// 计数器推进到触发清扫的边界
+	svc.openaiCodexTurnStateWrites.Store(255)
+	svc.sweepOpenAICodexTurnStateOrigins()
+
+	_, expiredOK := svc.openaiCodexTurnStateOrigins.Load("expired")
+	require.False(t, expiredOK)
+	_, aliveOK := svc.openaiCodexTurnStateOrigins.Load("alive")
+	require.True(t, aliveOK)
+}
+
+func TestWriteOpenAIPassthroughResponseHeaders_RelaysAndClearsTurnState(t *testing.T) {
+	// filter=nil 走 content-type 兜底分支；turn-state 强制放行不依赖 filter。
+	dst := http.Header{}
+	src := http.Header{}
+	src.Set("X-Codex-Turn-State", "blob-P")
+	writeOpenAIPassthroughResponseHeaders(dst, src, nil)
+	require.Equal(t, "blob-P", dst.Get("X-Codex-Turn-State"))
+
+	// 上游缺失时清除残留（failover 换号防串扰）
+	writeOpenAIPassthroughResponseHeaders(dst, http.Header{"Content-Type": []string{"application/json"}}, nil)
+	require.Empty(t, dst.Get("X-Codex-Turn-State"))
+}
+
+func TestEnsureOpenAIRemoteCompactionV2BetaFeature(t *testing.T) {
+	t.Run("absent_sets_feature", func(t *testing.T) {
+		h := http.Header{}
+		ensureOpenAIRemoteCompactionV2BetaFeature(h)
+		require.Equal(t, "remote_compaction_v2", h.Get("x-codex-beta-features"))
+	})
+
+	t.Run("present_unchanged", func(t *testing.T) {
+		h := http.Header{}
+		h.Set("x-codex-beta-features", "responses_websockets_v2, remote_compaction_v2")
+		ensureOpenAIRemoteCompactionV2BetaFeature(h)
+		require.Equal(t, "responses_websockets_v2, remote_compaction_v2", h.Get("x-codex-beta-features"))
+	})
+
+	t.Run("other_tokens_merged", func(t *testing.T) {
+		h := http.Header{}
+		h.Set("x-codex-beta-features", "responses_websockets_v2")
+		ensureOpenAIRemoteCompactionV2BetaFeature(h)
+		require.Equal(t, "responses_websockets_v2,remote_compaction_v2", h.Get("x-codex-beta-features"))
+	})
+
+	t.Run("multi_line_values_merged_single_line", func(t *testing.T) {
+		h := http.Header{}
+		h.Add("x-codex-beta-features", "feature_a")
+		h.Add("x-codex-beta-features", "feature_b")
+		ensureOpenAIRemoteCompactionV2BetaFeature(h)
+		require.Equal(t, []string{"feature_a,feature_b,remote_compaction_v2"}, h.Values("x-codex-beta-features"))
+	})
+}
+
+// 对齐真实 Codex：该头是会话级常量，挂在 OAuth 的每个请求上，而不是只在
+// 压缩回合出现（codex-rs build_model_client_beta_features_header）。
+func TestApplyOpenAICodexBetaFeatures(t *testing.T) {
+	oauthAccount := &Account{ID: 1, Platform: PlatformOpenAI, Type: AccountTypeOAuth}
+	apiKeyAccount := &Account{ID: 2, Platform: PlatformOpenAI, Type: AccountTypeAPIKey}
+
+	t.Run("oauth_plain_request_gets_default_codex_shape", func(t *testing.T) {
+		c, _ := newTurnStateTestContext(t, 7, "sess-beta")
+		h := http.Header{}
+		applyOpenAICodexBetaFeatures(c, oauthAccount, h)
+		require.Equal(t, "remote_compaction_v2", h.Get("x-codex-beta-features"),
+			"OAuth 的普通请求也必须带会话级 beta 头")
+	})
+
+	t.Run("client_declared_header_preserved", func(t *testing.T) {
+		c, _ := newTurnStateTestContext(t, 7, "sess-beta")
+		h := http.Header{}
+		h.Set("x-codex-beta-features", "some_other_feature")
+		applyOpenAICodexBetaFeatures(c, oauthAccount, h)
+		require.Equal(t, "some_other_feature", h.Get("x-codex-beta-features"),
+			"客户端显式声明的能力集不得被网关改写（非空即视为用户已关闭 v2）")
+	})
+
+	t.Run("native_v2_forces_feature_even_when_client_trimmed_it", func(t *testing.T) {
+		c, _ := newTurnStateTestContext(t, 7, "sess-beta")
+		MarkOpenAINativeCompactionV2(c)
+		h := http.Header{}
+		h.Set("x-codex-beta-features", "some_other_feature")
+		applyOpenAICodexBetaFeatures(c, oauthAccount, h)
+		require.Contains(t, h.Get("x-codex-beta-features"), "remote_compaction_v2",
+			"body 带 compaction_trigger 是实锤，必须确保 v2 在列")
+		require.Contains(t, h.Get("x-codex-beta-features"), "some_other_feature")
+	})
+
+	t.Run("native_v2_applies_to_non_oauth_too", func(t *testing.T) {
+		c, _ := newTurnStateTestContext(t, 7, "sess-beta")
+		MarkOpenAINativeCompactionV2(c)
+		h := http.Header{}
+		applyOpenAICodexBetaFeatures(c, apiKeyAccount, h)
+		require.Equal(t, "remote_compaction_v2", h.Get("x-codex-beta-features"))
+	})
+
+	t.Run("non_oauth_plain_request_untouched", func(t *testing.T) {
+		c, _ := newTurnStateTestContext(t, 7, "sess-beta")
+		h := http.Header{}
+		applyOpenAICodexBetaFeatures(c, apiKeyAccount, h)
+		require.Empty(t, h.Get("x-codex-beta-features"),
+			"非 Codex 后端不做会话级注入")
+	})
+
+	t.Run("nil_account_plain_request_untouched", func(t *testing.T) {
+		c, _ := newTurnStateTestContext(t, 7, "sess-beta")
+		h := http.Header{}
+		applyOpenAICodexBetaFeatures(c, nil, h)
+		require.Empty(t, h.Get("x-codex-beta-features"))
+	})
+}
+
+// WS 握手与 HTTP 出站必须给出同一份会话级 beta 头：真实 Codex 的
+// build_websocket_headers 复用 build_responses_headers（client.rs），
+// 两侧不一致还会让预热连接与实际请求落进不同的连接池兼容分桶。
+func TestBuildOpenAIWSHeaders_CarriesSessionBetaFeatures(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+	svc := &OpenAIGatewayService{}
+	decision := OpenAIWSProtocolDecision{Transport: OpenAIUpstreamTransportResponsesWebsocketV2}
+
+	build := func(t *testing.T, account *Account, clientBeta string) http.Header {
+		t.Helper()
+		rec := httptest.NewRecorder()
+		c, _ := gin.CreateTestContext(rec)
+		c.Request = httptest.NewRequest(http.MethodGet, "/v1/responses", nil)
+		if clientBeta != "" {
+			c.Request.Header.Set("x-codex-beta-features", clientBeta)
+		}
+		headers, _, err := svc.buildOpenAIWSHeaders(
+			context.Background(), c, account, "test-token", decision,
+			true, "", "", "", "gpt-5.6-codex", "",
+		)
+		require.NoError(t, err)
+		return headers
+	}
+
+	oauthAccount := &Account{
+		Platform:    PlatformOpenAI,
+		Type:        AccountTypeOAuth,
+		Credentials: map[string]any{"chatgpt_account_id": "test-account"},
+	}
+
+	headers := build(t, oauthAccount, "")
+	require.Equal(t, "remote_compaction_v2", headers.Get("x-codex-beta-features"),
+		"WS 握手也必须带会话级 beta 头")
+
+	declared := build(t, oauthAccount, "some_other_feature")
+	require.Equal(t, []string{"some_other_feature"}, declared.Values("x-codex-beta-features"),
+		"客户端已声明时原样保留")
+
+	apiKeyHeaders := build(t, &Account{Platform: PlatformOpenAI, Type: AccountTypeAPIKey}, "")
+	require.Empty(t, apiKeyHeaders.Get("x-codex-beta-features"),
+		"非 Codex 后端不注入")
+}

backend/internal/service/openai_compact_body_signal.go

@@ -1,6 +1,114 @@
 package service
 
-import "github.com/tidwall/gjson"
+import (
+	"net/http"
+	"strings"
+
+	"github.com/gin-gonic/gin"
+	"github.com/tidwall/gjson"
+)
+
+// openAINativeCompactionV2Key 标记本请求是原生 remote compaction v2
+// （裸 /responses + stream:true + compaction_trigger），由 handler 在判定后
+// 写入，供上游请求构造时补注协商头。
+const openAINativeCompactionV2Key = "openai_native_compaction_v2"
+
+const openAIRemoteCompactionV2Feature = "remote_compaction_v2"
+
+// MarkOpenAINativeCompactionV2 由 handler 在识别出原生 v2 压缩请求时调用。
+func MarkOpenAINativeCompactionV2(c *gin.Context) {
+	if c != nil {
+		c.Set(openAINativeCompactionV2Key, true)
+	}
+}
+
+func isOpenAINativeCompactionV2(c *gin.Context) bool {
+	if c == nil {
+		return false
+	}
+	return c.GetBool(openAINativeCompactionV2Key)
+}
+
+// ensureOpenAIRemoteCompactionV2BetaFeature 确保出站 x-codex-beta-features
+// 头包含 remote_compaction_v2。真实 Codex 发送 compaction_trigger 时总会同时
+// 携带该协商头（codex-rs build_model_client_beta_features_header 对该 feature
+// 特判 advertise）；上游或下游网关链剥掉它后，请求会在依赖该头做门控的
+// 环节被降级（#5586）。这里在原生 v2 请求出站前补齐，使线型与真实 Codex
+// 一致。已存在时保持原样，不重复追加。
+func ensureOpenAIRemoteCompactionV2BetaFeature(h http.Header) {
+	if h == nil {
+		return
+	}
+	tokens := make([]string, 0, 4)
+	for _, value := range h.Values("x-codex-beta-features") {
+		for _, token := range strings.Split(value, ",") {
+			token = strings.TrimSpace(token)
+			if token == "" {
+				continue
+			}
+			if token == openAIRemoteCompactionV2Feature {
+				return
+			}
+			tokens = append(tokens, token)
+		}
+	}
+	tokens = append(tokens, openAIRemoteCompactionV2Feature)
+	h.Set("x-codex-beta-features", strings.Join(tokens, ","))
+}
+
+// hasOpenAICodexBetaFeaturesHeader 报告出站头里是否已存在非空的
+// x-codex-beta-features（即客户端自己声明过能力集）。
+func hasOpenAICodexBetaFeaturesHeader(h http.Header) bool {
+	if h == nil {
+		return false
+	}
+	for _, value := range h.Values("x-codex-beta-features") {
+		if strings.TrimSpace(value) != "" {
+			return true
+		}
+	}
+	return false
+}
+
+// applyOpenAICodexBetaFeatures 按真实 Codex 的会话级行为补注
+// x-codex-beta-features。
+//
+// codex 侧规则（codex-rs：session/mod.rs build_model_client_beta_features_header
+// 组装、client.rs build_responses_headers 附加）：该头是**会话级常量**，挂在
+// /responses、WS 握手、/responses/compact 三处的**每一个**请求上，而不是只挂
+// 压缩回合。其内容是"已启用且需要 advertise 的 feature 列表"，RemoteCompactionV2
+// 被特判 advertise；实测默认安装下没有任何 Experimental 特性默认开启，
+// 因此默认 Codex 的头值恰好就是单个 "remote_compaction_v2"。
+//
+// 网关据此对齐：
+//   - 原生 v2 压缩回合（body 带 compaction_trigger 实锤）：无论账号类型都确保
+//     v2 在列，覆盖中间网关裁剪 token 的情形（#5586）；
+//   - ChatGPT codex 上游（OAuth）的其余请求：客户端**未**声明该头时补成默认
+//     Codex 形态，消除"仅压缩回合才带该头"这种真实 Codex 不会产生的模式；
+//   - 客户端已声明该头：原样保留。非空但不含 v2 表示用户显式关闭了该特性，
+//     网关不得替其改写能力声明；
+//   - 非 OAuth 上游（API Key/第三方兼容网关）：不做会话级注入，只保留压缩回合
+//     的那一条，避免向非 Codex 后端撒 Codex 专属头。
+//
+// 已知无解的歧义：用户关掉 v2 且无其他特性时，真实 Codex 同样不发该头，与"老
+// 客户端"在线型上不可区分，此时按默认形态补注。该用户的 legacy 压缩端点本就
+// 已被上游下线（404），不存在可回退的正确行为。
+func applyOpenAICodexBetaFeatures(c *gin.Context, account *Account, h http.Header) {
+	if h == nil {
+		return
+	}
+	if isOpenAINativeCompactionV2(c) {
+		ensureOpenAIRemoteCompactionV2BetaFeature(h)
+		return
+	}
+	if account == nil || !account.IsOpenAIOAuth() {
+		return
+	}
+	if hasOpenAICodexBetaFeaturesHeader(h) {
+		return
+	}
+	h.Set("x-codex-beta-features", openAIRemoteCompactionV2Feature)
+}
 
 // HasCompactionTriggerInInput detects an input item with
 // type="compaction_trigger". The handler combines this body signal with the

backend/internal/service/openai_compact_probe.go

@@ -10,7 +10,8 @@ import (
 const (
 	// AccountTestModeDefault drives the standard /responses connection test.
 	AccountTestModeDefault = "default"
-	// AccountTestModeCompact drives the /responses/compact compact-probe test.
+	// AccountTestModeCompact drives the remote-compaction probe test
+	// (native v2: streaming /responses with a compaction_trigger input item).
 	AccountTestModeCompact = "compact"
 )
 
@@ -23,8 +24,12 @@ func normalizeAccountTestMode(mode string) string {
 	}
 }
 
-func createOpenAICompactProbePayload(model string) map[string]any {
-	return map[string]any{
+// createOpenAICompactProbePayload 构造原生 remote compaction v2 探测载荷：
+// 流式 /responses + input 末尾 {"type":"compaction_trigger"}。上游已下线
+// legacy unary /responses/compact（v1 形态恒 404，#5598/#5624），现行 codex
+// 默认协议即 v2（RemoteCompactionV2 Stable + default_enabled）。
+func createOpenAICompactProbePayload(model string, isOAuth bool) map[string]any {
+	payload := map[string]any{
 		"model":        strings.TrimSpace(model),
 		"instructions": "You are a helpful coding assistant.",
 		"input": []any{
@@ -33,8 +38,35 @@ func createOpenAICompactProbePayload(model string) map[string]any {
 				"role":    "user",
 				"content": "Respond with OK.",
 			},
+			map[string]any{"type": "compaction_trigger"},
 		},
+		"stream": true,
 	}
+	// ChatGPT internal API 要求 store: false，与真实转发一致。
+	if isOAuth {
+		payload["store"] = false
+	}
+	return payload
+}
+
+// openAICompactProbeFoundCompactionItem 判定探测响应是否产出了 compaction
+// 输出 item——v2 契约的核心（codex 缺它即 fatal "got 0 items"）。三种形态都
+// 认：① SSE 的 output_item.done/added（原生 v2 主形态，codex 只从这里收集
+// item）；② SSE 终态 response.completed 的 response.output[]（部分上游只在
+// 终态给出 item）；③ 整体 JSON 的 output[]（老网关链把请求降级成 unary）。
+func openAICompactProbeFoundCompactionItem(body []byte) bool {
+	if len(body) == 0 {
+		return false
+	}
+	bodyText := string(body)
+	if _, found := findRawCompactionItemFromSSE(bodyText); found {
+		return true
+	}
+	if finalResponse, ok := extractCodexFinalResponse(bodyText); ok &&
+		responsesOutputHasCompactionItem(finalResponse) {
+		return true
+	}
+	return responsesOutputHasCompactionItem(body)
 }
 
 func shouldMarkOpenAICompactUnsupported(status int, body []byte) bool {
@@ -60,7 +92,12 @@ func shouldMarkOpenAICompactUnsupported(status int, body []byte) bool {
 	return false
 }
 
-func buildOpenAICompactProbeExtraUpdates(resp *http.Response, body []byte, probeErr error, now time.Time) map[string]any {
+// buildOpenAICompactProbeExtraUpdates 计算探测结果的账号 extra 更新。
+// compactionFound 是 v2 契约判据：HTTP 2xx 但响应无 compaction item 时同样
+// 记为不支持（链路把 compaction_trigger 吞掉的形态，等价 codex 的 "got 0
+// items" fatal，#5478/#5648）。极端场景（上游链只支持 legacy unary compact）
+// 可用账号级 openai_compact_mode=force_on 人工覆盖。
+func buildOpenAICompactProbeExtraUpdates(resp *http.Response, body []byte, probeErr error, compactionFound bool, now time.Time) map[string]any {
 	updates := map[string]any{
 		"openai_compact_checked_at":  now.Format(time.RFC3339),
 		"openai_compact_last_status": nil,
@@ -84,10 +121,14 @@ func buildOpenAICompactProbeExtraUpdates(resp *http.Response, body []byte, probe
 			errMsg = "HTTP " + strconv.Itoa(resp.StatusCode)
 		}
 		errMsg = truncateString(sanitizeUpstreamErrorMessage(errMsg), 2048)
-		if resp.StatusCode >= 200 && resp.StatusCode < 300 {
+		switch {
+		case resp.StatusCode >= 200 && resp.StatusCode < 300 && compactionFound:
 			updates["openai_compact_supported"] = true
 			updates["openai_compact_last_error"] = ""
-		} else {
+		case resp.StatusCode >= 200 && resp.StatusCode < 300:
+			updates["openai_compact_supported"] = false
+			updates["openai_compact_last_error"] = "upstream returned 2xx without a compaction output item (native remote compaction v2 unsupported)"
+		default:
 			if shouldMarkOpenAICompactUnsupported(resp.StatusCode, body) {
 				updates["openai_compact_supported"] = false
 			}
@@ -112,9 +153,14 @@ func mergeExtraUpdates(base map[string]any, more map[string]any) map[string]any
 	return out
 }
 
+// compactProbeSessionID 返回探测请求使用的会话标识。真实 Codex 的
+// session-id / thread-id 恒为 UUID（codex-protocol ThreadId 是 UUIDv7），
+// 探测既然与真实流量走同一个 /responses 端点，标识形态就必须同构——
+// 否则上游能凭 "probe_compact_5" 这类字面量一眼区分出探测流量。
+// 账号级稳定派生：重复探测复用同一会话，而不是每次新开一个。
 func compactProbeSessionID(accountID int64) string {
 	if accountID <= 0 {
-		return "probe_compact"
+		return deriveStableUUIDv4("sub2api:codex-compact-probe:v1:anonymous")
 	}
-	return "probe_compact_" + strconv.FormatInt(accountID, 10)
+	return deriveStableUUIDv4("sub2api:codex-compact-probe:v1:" + strconv.FormatInt(accountID, 10))
 }

backend/internal/service/openai_compact_probe_test.go

@@ -28,7 +28,7 @@ func TestNormalizeAccountTestMode(t *testing.T) {
 
 func TestBuildOpenAICompactProbeExtraUpdates_SuccessMarksSupported(t *testing.T) {
 	now := time.Date(2026, 4, 10, 10, 0, 0, 0, time.UTC)
-	updates := buildOpenAICompactProbeExtraUpdates(&http.Response{StatusCode: http.StatusOK}, []byte(`{"id":"cmp_1"}`), nil, now)
+	updates := buildOpenAICompactProbeExtraUpdates(&http.Response{StatusCode: http.StatusOK}, []byte(`{"id":"cmp_1"}`), nil, true, now)
 
 	if got := updates["openai_compact_supported"]; got != true {
 		t.Fatalf("openai_compact_supported = %v, want true", got)
@@ -47,7 +47,7 @@ func TestBuildOpenAICompactProbeExtraUpdates_SuccessMarksSupported(t *testing.T)
 func TestBuildOpenAICompactProbeExtraUpdates_404MarksUnsupported(t *testing.T) {
 	now := time.Date(2026, 4, 10, 10, 0, 0, 0, time.UTC)
 	body := []byte(`404 page not found`)
-	updates := buildOpenAICompactProbeExtraUpdates(&http.Response{StatusCode: http.StatusNotFound}, body, nil, now)
+	updates := buildOpenAICompactProbeExtraUpdates(&http.Response{StatusCode: http.StatusNotFound}, body, nil, false, now)
 
 	if got := updates["openai_compact_supported"]; got != false {
 		t.Fatalf("openai_compact_supported = %v, want false", got)
@@ -59,7 +59,7 @@ func TestBuildOpenAICompactProbeExtraUpdates_404MarksUnsupported(t *testing.T) {
 
 func TestBuildOpenAICompactProbeExtraUpdates_502DoesNotMarkUnsupported(t *testing.T) {
 	now := time.Date(2026, 4, 10, 10, 0, 0, 0, time.UTC)
-	updates := buildOpenAICompactProbeExtraUpdates(&http.Response{StatusCode: http.StatusBadGateway}, []byte(`Upstream request failed`), nil, now)
+	updates := buildOpenAICompactProbeExtraUpdates(&http.Response{StatusCode: http.StatusBadGateway}, []byte(`Upstream request failed`), nil, false, now)
 
 	if _, exists := updates["openai_compact_supported"]; exists {
 		t.Fatalf("did not expect openai_compact_supported for 502 response")
@@ -71,7 +71,7 @@ func TestBuildOpenAICompactProbeExtraUpdates_502DoesNotMarkUnsupported(t *testin
 
 func TestBuildOpenAICompactProbeExtraUpdates_RequestErrorDoesNotMarkUnsupported(t *testing.T) {
 	now := time.Date(2026, 4, 10, 10, 0, 0, 0, time.UTC)
-	updates := buildOpenAICompactProbeExtraUpdates(nil, nil, errors.New("dial tcp timeout"), now)
+	updates := buildOpenAICompactProbeExtraUpdates(nil, nil, errors.New("dial tcp timeout"), false, now)
 
 	if _, exists := updates["openai_compact_supported"]; exists {
 		t.Fatalf("did not expect openai_compact_supported for request error")
@@ -86,7 +86,7 @@ func TestBuildOpenAICompactProbeExtraUpdates_RequestErrorDoesNotMarkUnsupported(
 
 func TestBuildOpenAICompactProbeExtraUpdates_NoResponseClearsLastStatus(t *testing.T) {
 	now := time.Date(2026, 4, 10, 10, 0, 0, 0, time.UTC)
-	updates := buildOpenAICompactProbeExtraUpdates(nil, nil, nil, now)
+	updates := buildOpenAICompactProbeExtraUpdates(nil, nil, nil, false, now)
 
 	if got, exists := updates["openai_compact_last_status"]; !exists || got != nil {
 		t.Fatalf("openai_compact_last_status = %v, want nil key", got)
@@ -99,7 +99,7 @@ func TestBuildOpenAICompactProbeExtraUpdates_NoResponseClearsLastStatus(t *testi
 func TestBuildOpenAICompactProbeExtraUpdates_UnknownModelDoesNotMarkUnsupported(t *testing.T) {
 	now := time.Date(2026, 4, 10, 10, 0, 0, 0, time.UTC)
 	body := []byte(`{"error":{"message":"unknown model gpt-5.4-openai-compact"}}`)
-	updates := buildOpenAICompactProbeExtraUpdates(&http.Response{StatusCode: http.StatusBadRequest}, body, nil, now)
+	updates := buildOpenAICompactProbeExtraUpdates(&http.Response{StatusCode: http.StatusBadRequest}, body, nil, false, now)
 
 	if _, exists := updates["openai_compact_supported"]; exists {
 		t.Fatalf("did not expect openai_compact_supported for unknown-model diagnostics")
@@ -111,7 +111,7 @@ func TestBuildOpenAICompactProbeExtraUpdates_UnknownModelDoesNotMarkUnsupported(
 
 func TestBuildOpenAICompactProbeExtraUpdates_EmptyFailureBodyFallsBackToHTTPStatus(t *testing.T) {
 	now := time.Date(2026, 4, 10, 10, 0, 0, 0, time.UTC)
-	updates := buildOpenAICompactProbeExtraUpdates(&http.Response{StatusCode: http.StatusServiceUnavailable}, nil, nil, now)
+	updates := buildOpenAICompactProbeExtraUpdates(&http.Response{StatusCode: http.StatusServiceUnavailable}, nil, nil, false, now)
 
 	if got := updates["openai_compact_last_status"]; got != http.StatusServiceUnavailable {
 		t.Fatalf("openai_compact_last_status = %v, want %d", got, http.StatusServiceUnavailable)
@@ -120,3 +120,79 @@ func TestBuildOpenAICompactProbeExtraUpdates_EmptyFailureBodyFallsBackToHTTPStat
 		t.Fatalf("openai_compact_last_error = %v, want HTTP 503", got)
 	}
 }
+
+func TestBuildOpenAICompactProbeExtraUpdates_2xxWithoutCompactionItemMarksUnsupported(t *testing.T) {
+	now := time.Date(2026, 4, 10, 10, 0, 0, 0, time.UTC)
+	updates := buildOpenAICompactProbeExtraUpdates(&http.Response{StatusCode: http.StatusOK}, []byte(`{"id":"resp_1","output":[]}`), nil, false, now)
+
+	if got := updates["openai_compact_supported"]; got != false {
+		t.Fatalf("openai_compact_supported = %v, want false（2xx 无 compaction item = v2 不可用）", got)
+	}
+	if got := updates["openai_compact_last_error"]; got == "" {
+		t.Fatalf("expected openai_compact_last_error to explain the missing compaction item")
+	}
+}
+
+func TestCreateOpenAICompactProbePayload_NativeV2Shape(t *testing.T) {
+	payload := createOpenAICompactProbePayload("gpt-5.6-sol", true)
+	if payload["stream"] != true {
+		t.Fatalf("v2 probe payload must be streaming")
+	}
+	if payload["store"] != false {
+		t.Fatalf("OAuth probe payload must carry store:false")
+	}
+	input, ok := payload["input"].([]any)
+	if !ok || len(input) != 2 {
+		t.Fatalf("expected 2 input items, got %v", payload["input"])
+	}
+	last, ok := input[len(input)-1].(map[string]any)
+	if !ok || last["type"] != "compaction_trigger" {
+		t.Fatalf("last input item must be compaction_trigger, got %v", input[len(input)-1])
+	}
+
+	apiKeyPayload := createOpenAICompactProbePayload("gpt-5.6-sol", false)
+	if _, has := apiKeyPayload["store"]; has {
+		t.Fatalf("API-key probe payload must not force store")
+	}
+}
+
+func TestOpenAICompactProbeFoundCompactionItem(t *testing.T) {
+	sseWithItem := []byte("event: response.output_item.done\ndata: {\"type\":\"response.output_item.done\",\"item\":{\"type\":\"compaction\",\"id\":\"cmp_1\",\"encrypted_content\":\"blob\"}}\n\ndata: {\"type\":\"response.completed\",\"response\":{\"id\":\"resp_1\",\"output\":[]}}\n\n")
+	if !openAICompactProbeFoundCompactionItem(sseWithItem) {
+		t.Fatalf("SSE output_item.done 携带 compaction item 应判定为支持")
+	}
+
+	sseAlias := []byte("data: {\"type\":\"response.output_item.done\",\"item\":{\"type\":\"compaction_summary\",\"id\":\"cmp_2\"}}\n\n")
+	if !openAICompactProbeFoundCompactionItem(sseAlias) {
+		t.Fatalf("compaction_summary 别名应判定为支持")
+	}
+
+	sseNoItem := []byte("data: {\"type\":\"response.output_item.done\",\"item\":{\"type\":\"message\",\"id\":\"msg_1\"}}\n\ndata: {\"type\":\"response.completed\",\"response\":{\"id\":\"resp_2\",\"output\":[]}}\n\n")
+	if openAICompactProbeFoundCompactionItem(sseNoItem) {
+		t.Fatalf("无 compaction item 的流不应判定为支持")
+	}
+
+	jsonWithItem := []byte(`{"id":"resp_3","output":[{"type":"compaction","id":"cmp_3"}]}`)
+	if !openAICompactProbeFoundCompactionItem(jsonWithItem) {
+		t.Fatalf("JSON output[] 携带 compaction item 应判定为支持（降级链兜底）")
+	}
+
+	if openAICompactProbeFoundCompactionItem(nil) {
+		t.Fatalf("空响应不应判定为支持")
+	}
+}
+
+func TestOpenAICompactProbeFoundCompactionItem_TerminalResponseOutput(t *testing.T) {
+	// 部分上游只在终态 response.completed 的 output[] 给出 compaction item，
+	// 事件流里没有 output_item.done——探针同样必须判定为支持。
+	sseTerminalOnly := []byte("data: {\"type\":\"response.output_item.done\",\"item\":{\"type\":\"message\",\"id\":\"msg_1\"}}\n\n" +
+		"data: {\"type\":\"response.completed\",\"response\":{\"id\":\"resp_t\",\"output\":[{\"type\":\"compaction\",\"id\":\"cmp_t\"}]}}\n\n")
+	if !openAICompactProbeFoundCompactionItem(sseTerminalOnly) {
+		t.Fatalf("终态 response.output 中的 compaction item 应判定为支持")
+	}
+
+	sseTerminalEmpty := []byte("data: {\"type\":\"response.completed\",\"response\":{\"id\":\"resp_e\",\"output\":[]}}\n\n")
+	if openAICompactProbeFoundCompactionItem(sseTerminalEmpty) {
+		t.Fatalf("终态 output 为空且无 item 事件时不应判定为支持")
+	}
+}

backend/internal/service/openai_gateway_forward.go

@@ -427,10 +427,10 @@ func (s *OpenAIGatewayService) Forward(ctx context.Context, c *gin.Context, acco
 					markDecodedModified()
 				}
 			}
-			// 将 fpIDs 存入 gin context，供 buildUpstreamRequest 中头改写使用
-			if c != nil && fpIDs != nil {
-				c.Set("codex_fingerprint_ids", fpIDs)
-			}
+			// 将 fpIDs 存入 gin context，供 buildUpstreamRequest 中头改写使用。
+			// 无条件覆写（含 nil）：failover 从收敛账号切到 off 账号时，上一
+			// 账号的 IDs 不得残留（stageCodexFingerprintIDs 注释）。
+			stageCodexFingerprintIDs(c, fpIDs)
 		}
 		if codexResult.NormalizedModel != "" {
 			upstreamModel = codexResult.NormalizedModel
@@ -1094,6 +1094,9 @@ func (s *OpenAIGatewayService) buildUpstreamRequest(ctx context.Context, c *gin.
 			}
 		}
 	}
+	// 客户端回带的 x-codex-turn-state 若已知由其他账号铸造（failover 换号），
+	// 剥离后再出站——异账号 blob 与本账号的（指纹收敛后）出站身份自相矛盾。
+	s.guardOpenAICodexTurnStateEcho(c, account, req.Header)
 	if account.Type == AccountTypeOAuth {
 		compatMessagesBridge := isOpenAICompatMessagesBridgeContext(c) || isOpenAICompatMessagesBridgeBody(body)
 		// 清除客户端透传的 session 头，后续用隔离后的值重新设置，防止跨用户会话碰撞。
@@ -1144,13 +1147,7 @@ func (s *OpenAIGatewayService) buildUpstreamRequest(ctx context.Context, c *gin.
 	}
 
 	// 指纹收敛：使用 Forward() 中预计算的收敛 ID 改写出站头，与请求体使用同一份 IDs。
-	if account.Type == AccountTypeOAuth && c != nil {
-		if fpIDs, ok := c.Get("codex_fingerprint_ids"); ok {
-			if ids, ok := fpIDs.(*codexFingerprintIDs); ok {
-				applyCodexFingerprintHeaders(req.Header, ids)
-			}
-		}
-	}
+	applyStagedCodexFingerprintHeaders(c, account, req.Header)
 
 	// 终态收口：强制统一 OAuth 出站身份（User-Agent / originator / version 同源自洽）。
 	// 客户端自报身份不参与构造，浏览器型 UA 也因此不会再到达上游（原浏览器 UA 兜底已被吸收）。
@@ -1165,6 +1162,9 @@ func (s *OpenAIGatewayService) buildUpstreamRequest(ctx context.Context, c *gin.
 
 	// 账号级请求头覆写（仅 openai api_key 账号启用时生效；OAuth 路径 no-op）
 	account.ApplyHeaderOverrides(req.Header)
+	// x-codex-beta-features：按真实 Codex 的会话级行为补注（在账号级覆写之后，
+	// 保证不被覆盖丢失）。
+	applyOpenAICodexBetaFeatures(c, account, req.Header)
 	setOpenAICodexRoutingHintFromBody(req.Header, account, body)
 	logOpenAIRoutingDiagnosticsFromBody(ctx, account, "http", req.Header, body, "not_applicable")
 

backend/internal/service/openai_gateway_passthrough.go

@@ -80,6 +80,28 @@ func (s *OpenAIGatewayService) forwardOpenAIPassthrough(
 			body = normalizedBody
 		}
 		reqStream = gjson.GetBytes(body, "stream").Bool()
+
+		// 指纹收敛：与非透传路径同门控（仅 OAuth、legacy compact 形态跳过）。
+		// 一次性解析收敛 ID：请求体 client_metadata 在此改写（raw 字节外科
+		// 手术，透传热路径禁全量 Unmarshal），出站头改写由请求构造器读取
+		// context 中的同一份 IDs 完成（turn_id 等随机字段两侧必须一致）。
+		if !isOpenAIResponsesCompactPath(c) {
+			var clientHeaders http.Header
+			if c != nil && c.Request != nil {
+				clientHeaders = c.Request.Header
+			}
+			fpIDs := resolveCodexFingerprintIDsFromRequest(account, clientHeaders)
+			if fpIDs != nil {
+				fpBody, fpChanged, fpErr := applyCodexFingerprintClientMetadataRaw(body, fpIDs)
+				if fpErr != nil {
+					return nil, fpErr
+				}
+				if fpChanged {
+					body = fpBody
+				}
+			}
+			stageCodexFingerprintIDs(c, fpIDs)
+		}
 	}
 
 	sanitizedBody, sanitized, err := sanitizeEmptyBase64InputImagesInOpenAIBody(body)
@@ -239,6 +261,13 @@ func (s *OpenAIGatewayService) forwardOpenAIPassthrough(
 
 	serviceTier := extractOpenAIServiceTierFromBody(body)
 
+	// x-codex-turn-state 溯源：下游回传由 writeOpenAIPassthroughResponseHeaders
+	// 在各 handler 的写头点强制放行，铸造账号在此统一记录，供出站守卫剥离
+	// failover 换号后的跨账号回带（openai_codex_turn_state.go）。
+	if extractOpenAICodexTurnState(resp.Header) != "" {
+		s.noteOpenAICodexTurnStateProvenance(c, account)
+	}
+
 	var usage *OpenAIUsage
 	var firstTokenMs *int
 	responseID := ""
@@ -376,6 +405,10 @@ func (s *OpenAIGatewayService) buildUpstreamRequestOpenAIPassthrough(
 		}
 	}
 
+	// 客户端回带的 x-codex-turn-state 若已知由其他账号铸造（failover 换号），
+	// 剥离后再出站（openai_codex_turn_state.go）。
+	s.guardOpenAICodexTurnStateEcho(c, account, req.Header)
+
 	// 覆盖入站鉴权残留，并注入上游认证
 	req.Header.Del("authorization")
 	req.Header.Del("x-api-key")
@@ -447,6 +480,10 @@ func (s *OpenAIGatewayService) buildUpstreamRequestOpenAIPassthrough(
 	if s.cfg != nil && s.cfg.Gateway.ForceCodexCLI {
 		req.Header.Set("user-agent", codexCLIUserAgent)
 	}
+	// 指纹收敛：使用 forwardOpenAIPassthrough 中预计算的收敛 ID 改写出站头，
+	// 与请求体 client_metadata 共享同一份 IDs（与非透传路径相同的相对位置：
+	// 会话隔离之后、终态身份收口之前）。
+	applyStagedCodexFingerprintHeaders(c, account, req.Header)
 	// 终态收口：透传路径的 OAuth 与非透传完全一致，同样强制统一出站身份
 	// （User-Agent / originator / version 同源自洽），客户端自报身份不会到达上游。
 	if account.Type == AccountTypeOAuth {
@@ -459,6 +496,9 @@ func (s *OpenAIGatewayService) buildUpstreamRequestOpenAIPassthrough(
 
 	// 账号级请求头覆写（仅 openai api_key 账号启用时生效；OAuth 路径 no-op）
 	account.ApplyHeaderOverrides(req.Header)
+	// x-codex-beta-features：按真实 Codex 的会话级行为补注（在账号级覆写之后，
+	// 保证不被覆盖丢失）。
+	applyOpenAICodexBetaFeatures(c, account, req.Header)
 	setOpenAICodexRoutingHintFromBody(req.Header, account, body)
 	logOpenAIRoutingDiagnosticsFromBody(ctx, account, "http_passthrough", req.Header, body, "not_applicable")
 
@@ -1650,4 +1690,13 @@ func writeOpenAIPassthroughResponseHeaders(dst http.Header, src http.Header, fil
 			dst.Add(key, v)
 		}
 	}
+
+	// x-codex-turn-state：Codex 回合状态头，客户端会在同回合后续请求回带。
+	// 与上面的用量头不同，这里在上游缺失时也主动清除——failover 换号后残留
+	// 上一账号的 blob 会构成跨账号矛盾（openai_codex_turn_state.go）。
+	turnStateKey := http.CanonicalHeaderKey(openAICodexTurnStateHeader)
+	dst.Del(turnStateKey)
+	for _, v := range getCaseInsensitiveValues(src, openAICodexTurnStateHeader) {
+		dst.Add(turnStateKey, v)
+	}
 }

backend/internal/service/openai_gateway_response_handling.go

@@ -64,6 +64,14 @@ func (s *OpenAIGatewayService) handleStreamingResponseWithReasoning(ctx context.
 	} else if s.responseHeaderFilter != nil {
 		responseheaders.WriteFilteredHeaders(c.Writer.Header(), resp.Header, s.responseHeaderFilter)
 	}
+	// x-codex-turn-state 不在通用响应头白名单内，按 Codex 协议显式回传：
+	// 客户端会在同回合的后续请求中回带（openai_codex_turn_state.go）。
+	// 首输出守卫模式下只暂存，溯源在 applyAttemptResponseHeaders 真正提交时记录。
+	if guardFirstOutput {
+		stageOpenAICodexTurnState(&attemptResponseHeaders, resp.Header)
+	} else {
+		s.relayOpenAICodexTurnState(c, account, resp.Header)
+	}
 
 	// Set SSE response headers
 	c.Header("Content-Type", "text/event-stream")
@@ -85,6 +93,9 @@ func (s *OpenAIGatewayService) handleStreamingResponseWithReasoning(ctx context.
 				c.Writer.Header().Add(key, value)
 			}
 		}
+		// 暂存头此刻才真正写给客户端：turn-state 溯源在这里记录（见
+		// noteStagedOpenAICodexTurnStateCommitted 的 failover 说明）。
+		s.noteStagedOpenAICodexTurnStateCommitted(c, account, attemptResponseHeaders)
 		// These headers describe this gateway's SSE stream and are stable across
 		// account attempts. Keep them authoritative over upstream values.
 		c.Header("Content-Type", "text/event-stream")
@@ -1219,7 +1230,7 @@ func (s *OpenAIGatewayService) handleNonStreamingResponse(ctx context.Context, r
 	// Some OpenAI-compatible upstreams (including other sub2api instances)
 	// may return SSE even when stream=false was requested.
 	if isEventStreamResponse(resp.Header) {
-		return s.handleSSEToJSON(resp, c, body, originalModel, mappedModel)
+		return s.handleSSEToJSON(resp, c, account, body, originalModel, mappedModel)
 	}
 	// bodyLooksLikeSSE is a line-level heuristic: real SSE framing requires
 	// "data:"/"event:" field names at the very start of a physical line. A
@@ -1235,7 +1246,7 @@ func (s *OpenAIGatewayService) handleNonStreamingResponse(ctx context.Context, r
 	// positives on JSON responses that coincidentally contain "data:" or
 	// "event:" in their text content.
 	if account.Type == AccountTypeOAuth && bodyLooksLikeSSE {
-		return s.handleSSEToJSON(resp, c, body, originalModel, mappedModel)
+		return s.handleSSEToJSON(resp, c, account, body, originalModel, mappedModel)
 	}
 	if account != nil && account.IsGrok() && isOpenAIResponsesCompactPath(c) {
 		body, err = convertGrokResponseToOpenAICompact(body)
@@ -1247,7 +1258,7 @@ func (s *OpenAIGatewayService) handleNonStreamingResponse(ctx context.Context, r
 	usageValue, usageOK := extractOpenAIUsageFromJSONBytes(body)
 	if !usageOK {
 		if bodyLooksLikeSSE {
-			return s.handleSSEToJSON(resp, c, body, originalModel, mappedModel)
+			return s.handleSSEToJSON(resp, c, account, body, originalModel, mappedModel)
 		}
 		return nil, fmt.Errorf("parse response: invalid json response")
 	}
@@ -1266,6 +1277,9 @@ func (s *OpenAIGatewayService) handleNonStreamingResponse(ctx context.Context, r
 		return nil, fmt.Errorf("restore OpenAI namespace response: %w", err)
 	}
 	responseheaders.WriteFilteredHeaders(c.Writer.Header(), resp.Header, s.responseHeaderFilter)
+	// Codex 协议要求 /responses/compact JSON 响应携带 x-codex-turn-state
+	// （codex-api/src/endpoint/compact.rs 从响应头捕获），显式回传。
+	s.relayOpenAICodexTurnState(c, account, resp.Header)
 
 	contentType := "application/json"
 	if s.cfg != nil && !s.cfg.Security.ResponseHeaders.Enabled {
@@ -1309,7 +1323,7 @@ func bodyHasSSEFraming(body []byte) bool {
 	return false
 }
 
-func (s *OpenAIGatewayService) handleSSEToJSON(resp *http.Response, c *gin.Context, body []byte, originalModel, mappedModel string) (*openaiNonStreamingResult, error) {
+func (s *OpenAIGatewayService) handleSSEToJSON(resp *http.Response, c *gin.Context, account *Account, body []byte, originalModel, mappedModel string) (*openaiNonStreamingResult, error) {
 	bodyText := string(body)
 	finalResponse, ok := extractCodexFinalResponse(bodyText)
 
@@ -1361,6 +1375,7 @@ func (s *OpenAIGatewayService) handleSSEToJSON(resp *http.Response, c *gin.Conte
 	}
 
 	responseheaders.WriteFilteredHeaders(c.Writer.Header(), resp.Header, s.responseHeaderFilter)
+	s.relayOpenAICodexTurnState(c, account, resp.Header)
 
 	contentType := "application/json; charset=utf-8"
 	if !ok {

backend/internal/service/openai_gateway_service.go

@@ -462,6 +462,11 @@ type OpenAIGatewayService struct {
 	codexModelsManifestCache            codexModelsManifestCache
 	openaiCompatSessionResponses        sync.Map
 	openaiCompatAnthropicDigestSessions sync.Map
+	// openaiCodexTurnStateOrigins: 下游会话 seed → openAICodexTurnStateOrigin，
+	// 记录最近一次向该会话下发 x-codex-turn-state 的铸造账号，供出站守卫
+	// 剥离跨账号回带（openai_codex_turn_state.go）。
+	openaiCodexTurnStateOrigins sync.Map
+	openaiCodexTurnStateWrites  atomic.Uint64
 }
 
 // NewOpenAIGatewayService creates a new OpenAIGatewayService

backend/internal/service/openai_gateway_service_test.go

@@ -3475,7 +3475,7 @@ func TestHandleSSEToJSON_CompletedEventReturnsJSON(t *testing.T) {
 		`data: [DONE]`,
 	}, "\n"))
 
-	usage, err := svc.handleSSEToJSON(resp, c, body, "gpt-4o", "gpt-4o")
+	usage, err := svc.handleSSEToJSON(resp, c, nil, body, "gpt-4o", "gpt-4o")
 	require.NoError(t, err)
 	require.NotNil(t, usage)
 	require.Equal(t, 7, usage.InputTokens)
@@ -3594,7 +3594,7 @@ func TestHandleSSEToJSON_ReconstructsImageGenerationOutputItemDone(t *testing.T)
 		`data: [DONE]`,
 	}, "\n"))
 
-	usage, err := svc.handleSSEToJSON(resp, c, body, "gpt-5.4", "gpt-5.4")
+	usage, err := svc.handleSSEToJSON(resp, c, nil, body, "gpt-5.4", "gpt-5.4")
 	require.NoError(t, err)
 	require.NotNil(t, usage)
 	require.Equal(t, 4, usage.ImageOutputTokens)
@@ -3621,7 +3621,7 @@ func TestHandleSSEToJSON_NoFinalResponseKeepsSSEBody(t *testing.T) {
 		`data: [DONE]`,
 	}, "\n"))
 
-	usage, err := svc.handleSSEToJSON(resp, c, body, "gpt-4o", "gpt-4o")
+	usage, err := svc.handleSSEToJSON(resp, c, nil, body, "gpt-4o", "gpt-4o")
 	require.NoError(t, err)
 	require.NotNil(t, usage)
 	require.Equal(t, 0, usage.InputTokens)
@@ -3645,7 +3645,7 @@ func TestHandleSSEToJSON_ResponseFailedReturnsProtocolError(t *testing.T) {
 		`data: [DONE]`,
 	}, "\n"))
 
-	usage, err := svc.handleSSEToJSON(resp, c, body, "gpt-4o", "gpt-4o")
+	usage, err := svc.handleSSEToJSON(resp, c, nil, body, "gpt-4o", "gpt-4o")
 	require.Nil(t, usage)
 	require.Error(t, err)
 	require.Equal(t, http.StatusBadGateway, rec.Code)

backend/internal/service/openai_ws_forwarder_payload.go

@@ -99,6 +99,12 @@ func (s *OpenAIGatewayService) buildOpenAIWSHeaders(
 			}
 		}
 	}
+	// 真实 Codex 的 WS 握手同样携带会话级 x-codex-beta-features
+	// （client.rs build_websocket_headers 复用 build_responses_headers），
+	// 客户端未声明时补成默认形态，与 HTTP 出站保持一致。放在客户端头拷贝
+	// 之外：该头是账号/会话级属性，不依赖入站请求是否存在，也避免预热与
+	// 实际请求因头差异落进不同的连接池兼容分桶。
+	applyOpenAICodexBetaFeatures(c, account, headers)
 	// OAuth 账号：将 apiKeyID 混入 session 标识符，防止跨用户会话碰撞。
 	if account != nil && account.Type == AccountTypeOAuth {
 		apiKeyID := getAPIKeyIDFromContext(c)

frontend/src/components/account/BulkEditAccountModal.vue

@@ -1535,7 +1535,7 @@ const codexCLIOnlyEnabled = ref(false)
 const codexCLIOnlyAppServerEnabled = ref(false)
 type CodexFingerprintMode = 'off' | 'device' | 'session' | 'full'
 const enableCodexFingerprintMode = ref(false)
-const codexFingerprintMode = ref<CodexFingerprintMode>('session')
+const codexFingerprintMode = ref<CodexFingerprintMode>('off')
 const codexFingerprintModeOptions = computed(() => [
   { value: 'off' as CodexFingerprintMode, label: t('admin.accounts.openai.codexFingerprintOff') },
   { value: 'device' as CodexFingerprintMode, label: t('admin.accounts.openai.codexFingerprintDevice') },
@@ -1829,7 +1829,8 @@ const buildUpdatePayload = (): Record<string, unknown> | null => {
 
   if (enableCodexFingerprintMode.value) {
     const extra = ensureExtra()
-    if (codexFingerprintMode.value !== 'session') {
+    // off = 默认值，清键即可；device/session/full 是显式 opt-in，必须落键（#5610）。
+    if (codexFingerprintMode.value !== 'off') {
       extra.codex_fingerprint_mode = codexFingerprintMode.value
     } else {
       delete extra.codex_fingerprint_mode
@@ -2082,7 +2083,7 @@ watch(
       enableCodexCLIOnly.value = false
       enableCodexCLIOnlyAppServer.value = false
       enableCodexFingerprintMode.value = false
-      codexFingerprintMode.value = 'session'
+      codexFingerprintMode.value = 'off'
       enableOpenAICompactMode.value = false
       enableOpenAICompactModelMapping.value = false
       enableRpmLimit.value = false

frontend/src/components/account/CreateAccountModal.vue

@@ -3855,7 +3855,7 @@ const openaiAPIKeyResponsesWebSocketV2Mode = ref<OpenAIWSMode>(OPENAI_WS_MODE_OF
 const codexCLIOnlyEnabled = ref(false)
 const codexCLIOnlyAppServerEnabled = ref(false)
 type CodexFingerprintMode = 'off' | 'device' | 'session' | 'full'
-const codexFingerprintMode = ref<CodexFingerprintMode>('session')
+const codexFingerprintMode = ref<CodexFingerprintMode>('off')
 const codexFingerprintModeOptions = computed(() => [
   { value: 'off' as CodexFingerprintMode, label: t('admin.accounts.openai.codexFingerprintOff') },
   { value: 'device' as CodexFingerprintMode, label: t('admin.accounts.openai.codexFingerprintDevice') },
@@ -4741,7 +4741,7 @@ const resetForm = () => {
   openaiAPIKeyResponsesWebSocketV2Mode.value = OPENAI_WS_MODE_OFF
   codexCLIOnlyEnabled.value = false
   codexCLIOnlyAppServerEnabled.value = false
-  codexFingerprintMode.value = 'session'
+  codexFingerprintMode.value = 'off'
   anthropicPassthroughEnabled.value = false
   anthropicAPIKeyAuthScheme.value = 'x_api_key'
   webSearchEmulationMode.value = 'default'
@@ -4840,7 +4840,9 @@ const buildOpenAIExtra = (base?: Record<string, unknown>): Record<string, unknow
   } else {
     delete extra.codex_cli_only_allow_app_server
   }
-  if (codexFingerprintMode.value !== 'session') {
+  // 收敛是显式 opt-in：off 即默认值，不落键；device/session/full 必须显式写入，
+  // 否则管理员的选择会被当成默认而丢失（#5610）。
+  if (codexFingerprintMode.value !== 'off') {
     extra.codex_fingerprint_mode = codexFingerprintMode.value
   } else {
     delete extra.codex_fingerprint_mode

frontend/src/components/account/EditAccountModal.vue

@@ -2980,7 +2980,7 @@ const openaiAPIKeyResponsesWebSocketV2Mode = ref<OpenAIWSMode>(OPENAI_WS_MODE_OF
 const codexCLIOnlyEnabled = ref(false)
 const codexCLIOnlyAppServerEnabled = ref(false)
 type CodexFingerprintMode = 'off' | 'device' | 'session' | 'full'
-const codexFingerprintMode = ref<CodexFingerprintMode>('session')
+const codexFingerprintMode = ref<CodexFingerprintMode>('off')
 type CodexImageToolMode = 'inherit' | 'enabled' | 'disabled' | 'block'
 const codexImageToolMode = ref<CodexImageToolMode>('inherit')
 type AnthropicAPIKeyAuthScheme = 'x_api_key' | 'authorization_bearer'
@@ -3440,7 +3440,7 @@ const syncFormFromAccount = (newAccount: Account | null) => {
   openaiAPIKeyResponsesWebSocketV2Mode.value = OPENAI_WS_MODE_OFF
   codexCLIOnlyEnabled.value = false
   codexCLIOnlyAppServerEnabled.value = false
-  codexFingerprintMode.value = 'session'
+  codexFingerprintMode.value = 'off'
   codexImageToolMode.value = 'inherit'
   anthropicPassthroughEnabled.value = false
   anthropicAPIKeyAuthScheme.value = 'x_api_key'
@@ -3494,9 +3494,10 @@ const syncFormFromAccount = (newAccount: Account | null) => {
     }
     if (newAccount.type === 'oauth') {
       const fpMode = extra?.codex_fingerprint_mode as string | undefined
+      // 缺省/非法值按 off 呈现，与后端 GetCodexFingerprintMode 的 opt-in 语义一致（#5610）
       codexFingerprintMode.value = (['off', 'device', 'session', 'full'].includes(fpMode || '')
         ? fpMode as CodexFingerprintMode
-        : 'session')
+        : 'off')
     }
     const credentials = newAccount.credentials as Record<string, unknown> | undefined
     const compactMappings = credentials?.compact_model_mapping as Record<string, string> | undefined
@@ -4843,9 +4844,10 @@ const handleSubmit = async () => {
         }
       }
 
-      // 指纹收敛模式：默认 session，不写入；非默认值显式写入（包括 off）
+      // 指纹收敛模式：默认 off（不写入）；device/session/full 是显式 opt-in，
+      // 必须落键，否则管理员的选择会被后端当作"未设置"而回落到 off（#5610）。
       if (props.account.type === 'oauth') {
-        if (codexFingerprintMode.value !== 'session') {
+        if (codexFingerprintMode.value !== 'off') {
           newExtra.codex_fingerprint_mode = codexFingerprintMode.value
         } else {
           delete newExtra.codex_fingerprint_mode

frontend/src/i18n/locales/en/admin/accounts.ts

@@ -574,10 +574,10 @@ export default {
         codexCLIOnlyAppServerDesc:
           "Effective only when the switch above is on. When enabled, this account also allows third-party clients that embed the Codex engine over the app-server protocol (e.g. Claude Code's codex plugin); they still pass the global engine-fingerprint gate. OR-combined with the global app-server toggle.",
         codexFingerprintMode: 'Codex fingerprint convergence',
-        codexFingerprintModeDesc: 'When multiple users share the same OAuth account, converge device/session identifiers to account-level stable values to reduce upstream-visible device and session count. Off = pass through client identifiers as-is.',
-        codexFingerprintOff: 'Off (passthrough)',
+        codexFingerprintModeDesc: 'When multiple users share the same OAuth account, converge device/session identifiers to account-level stable values to reduce upstream-visible device and session count. Off by default (client identifiers pass through as-is); opt in explicitly when needed. Some accounts reported quota shrinkage after enabling convergence, so choose based on your own measurements.',
+        codexFingerprintOff: 'Off (passthrough, default)',
         codexFingerprintDevice: 'Device only',
-        codexFingerprintSession: 'Device + Session (recommended)',
+        codexFingerprintSession: 'Device + Session',
         codexFingerprintFull: 'Full convergence',
         codexImageTool: 'Codex image bridge policy',
         codexImageToolDesc:

frontend/src/i18n/locales/zh/admin/accounts.ts

@@ -644,10 +644,10 @@ export default {
         codexCLIOnlyAppServer: '允许 Codex app-server 客户端',
         codexCLIOnlyAppServerDesc: '仅在上方开关开启时生效。开启后本账号额外放行内嵌 Codex 引擎、经 app-server 协议接入的第三方客户端（如 Claude Code 的 codex 插件），仍需通过全局引擎指纹门；与全局 app-server 开关取 OR（任一开即放行）。',
         codexFingerprintMode: 'Codex 指纹收敛',
-        codexFingerprintModeDesc: '多人共享同一 OAuth 账号时，将各用户的设备/会话标识收敛为账号级恒定值，减少上游可见的设备数和会话数。关闭时原样透传客户端标识。',
-        codexFingerprintOff: '关闭（透传）',
+        codexFingerprintModeDesc: '多人共享同一 OAuth 账号时，将各用户的设备/会话标识收敛为账号级恒定值，减少上游可见的设备数和会话数。默认关闭（原样透传客户端标识），需要时再显式开启；部分账号开启收敛后出现过额度缩水，请按自己的实测结果选择。',
+        codexFingerprintOff: '关闭（透传，默认）',
         codexFingerprintDevice: '仅设备',
-        codexFingerprintSession: '设备+会话（推荐）',
+        codexFingerprintSession: '设备+会话',
         codexFingerprintFull: '完全收敛',
         codexImageTool: 'Codex 图片桥接策略',
         codexImageToolDesc: