fix: address codex review findings

src/server/proxy-core/runtime/codexWebsocketRuntime.test.ts

@@ -147,6 +147,9 @@ describe('codexWebsocketRuntime', () => {
           id: 'resp-incomplete',
           model: parsed.model || 'gpt-5.4',
           status: 'incomplete',
+          incomplete_details: {
+            reason: 'max_output_tokens',
+          },
         },
       }));
     };
@@ -154,18 +157,34 @@ describe('codexWebsocketRuntime', () => {
     const { createCodexWebsocketRuntime, CodexWebsocketRuntimeError } = await import('./codexWebsocketRuntime.js');
     const runtime = createCodexWebsocketRuntime();
 
-    await expect(runtime.sendRequest({
+    let error: unknown;
-      sessionId: 'exec-session-incomplete',
+    try {
-      requestUrl: upstreamWsUrl,
+      await runtime.sendRequest({
-      headers: {
+        sessionId: 'exec-session-incomplete',
-        Authorization: 'Bearer oauth-access-token',
+        requestUrl: upstreamWsUrl,
-        'OpenAI-Beta': 'responses_websockets=2026-02-06',
+        headers: {
-      },
+          Authorization: 'Bearer oauth-access-token',
-      body: {
+          'OpenAI-Beta': 'responses_websockets=2026-02-06',
-        model: 'gpt-5.4',
+        },
-        input: [],
+        body: {
-      },
+          model: 'gpt-5.4',
-    })).rejects.toBeInstanceOf(CodexWebsocketRuntimeError);
+          input: [],
+        },
+      });
+    } catch (caught) {
+      error = caught;
+    }
+
+    expect(error).toBeInstanceOf(CodexWebsocketRuntimeError);
+    expect(error).toMatchObject({
+      message: 'max_output_tokens',
+      status: 502,
+    });
+    expect((error as CodexWebsocketRuntimeError).events).toEqual([
+      expect.objectContaining({
+        type: 'response.incomplete',
+      }),
+    ]);
   });
 
   it('treats top-level error frames as terminal websocket failures', async () => {
@@ -182,17 +201,36 @@ describe('codexWebsocketRuntime', () => {
     const { createCodexWebsocketRuntime, CodexWebsocketRuntimeError } = await import('./codexWebsocketRuntime.js');
     const runtime = createCodexWebsocketRuntime();
 
-    await expect(runtime.sendRequest({
+    let error: unknown;
-      sessionId: 'exec-session-error',
+    try {
-      requestUrl: upstreamWsUrl,
+      await runtime.sendRequest({
-      headers: {
+        sessionId: 'exec-session-error',
-        Authorization: 'Bearer oauth-access-token',
+        requestUrl: upstreamWsUrl,
-        'OpenAI-Beta': 'responses_websockets=2026-02-06',
+        headers: {
-      },
+          Authorization: 'Bearer oauth-access-token',
-      body: {
+          'OpenAI-Beta': 'responses_websockets=2026-02-06',
-        model: 'gpt-5.4',
+        },
-        input: [],
+        body: {
-      },
+          model: 'gpt-5.4',
-    })).rejects.toBeInstanceOf(CodexWebsocketRuntimeError);
+          input: [],
+        },
+      });
+    } catch (caught) {
+      error = caught;
+    }
+
+    expect(error).toBeInstanceOf(CodexWebsocketRuntimeError);
+    expect(error).toMatchObject({
+      message: 'account mismatch',
+      status: 502,
+    });
+    expect((error as CodexWebsocketRuntimeError).events).toEqual([
+      expect.objectContaining({
+        type: 'error',
+        error: expect.objectContaining({
+          message: 'account mismatch',
+        }),
+      }),
+    ]);
   });
 });

src/server/proxy-core/runtime/codexWebsocketRuntime.ts

@@ -34,6 +34,32 @@ function isFailureTerminalEvent(payload: Record<string, unknown>): boolean {
   return type === 'response.failed' || type === 'response.incomplete' || type === 'error';
 }
 
+function asFiniteNumber(value: unknown): number | undefined {
+  return typeof value === 'number' && Number.isFinite(value) ? value : undefined;
+}
+
+function extractFailureTerminalStatus(payload: Record<string, unknown>): number {
+  const response = isRecord(payload.response) ? payload.response : null;
+  const responseError = response && isRecord(response.error) ? response.error : null;
+  const topLevelError = isRecord(payload.error) ? payload.error : null;
+  const candidates = [
+    payload.status,
+    payload.statusCode,
+    payload.code,
+    topLevelError?.status,
+    topLevelError?.statusCode,
+    topLevelError?.code,
+    responseError?.status,
+    responseError?.statusCode,
+    responseError?.code,
+  ];
+  for (const candidate of candidates) {
+    const status = asFiniteNumber(candidate);
+    if (status !== undefined) return status;
+  }
+  return 502;
+}
+
 function extractTerminalErrorMessage(payload: Record<string, unknown>): string {
   const type = asTrimmedString(payload.type);
   if (type === 'error' && isRecord(payload.error)) {
@@ -243,7 +269,7 @@ async function sendSessionRequest(
           void closeSocket(socket);
           reject(new CodexWebsocketRuntimeError(extractTerminalErrorMessage(parsed), {
             events: [...events],
-            status: typeof parsed.status === 'number' && Number.isFinite(parsed.status) ? parsed.status : undefined,
+            status: extractFailureTerminalStatus(parsed),
             payload: parsed,
           }));
           return;

src/server/proxy-core/surfaces/openAiResponsesSurface.ts

@@ -135,6 +135,19 @@ function carriesResponsesFileUrlInput(value: unknown): boolean {
   return Object.values(value).some((entry) => carriesResponsesFileUrlInput(entry));
 }
 
+function shouldRefreshOauthResponsesRequest(input: {
+  oauthProvider?: string;
+  status: number;
+  response: Response;
+  rawErrText: string;
+}): boolean {
+  if (input.status === 401) return true;
+  if (input.status !== 403 || input.oauthProvider !== 'codex') return false;
+  const authenticate = input.response.headers.get('www-authenticate') || '';
+  const combined = `${authenticate}\n${input.rawErrText || ''}`;
+  return /\b(invalid_token|expired_token|expired|invalid|unauthorized|account mismatch|authentication)\b/i.test(combined);
+}
+
 type UsageSummary = ReturnType<typeof parseProxyUsage>;
 
 export async function handleOpenAiResponsesSurfaceRequest(
@@ -314,7 +327,12 @@ export async function handleOpenAiResponsesSurfaceRequest(
         dispatchRequest,
       });
       const tryRecover = async (ctx: Parameters<NonNullable<typeof endpointStrategy.tryRecover>>[0]) => {
-        if ((ctx.response.status === 401 || ctx.response.status === 403) && oauth) {
+        if (oauth && shouldRefreshOauthResponsesRequest({
+          oauthProvider: oauth.provider,
+          status: ctx.response.status,
+          response: ctx.response,
+          rawErrText: ctx.rawErrText || '',
+        })) {
           try {
             const refreshed = await refreshOauthAccessTokenSingleflight(selected.account.id);
             selected.tokenValue = refreshed.accessToken;

src/server/routes/api/oauth.test.ts

@@ -19,6 +19,16 @@ function buildJwt(payload: Record<string, unknown>) {
   return `${encode({ alg: 'none', typ: 'JWT' })}.${encode(payload)}.signature`;
 }
 
+function createDeferred<T>() {
+  let resolve!: (value: T | PromiseLike<T>) => void;
+  let reject!: (reason?: unknown) => void;
+  const promise = new Promise<T>((res, rej) => {
+    resolve = res;
+    reject = rej;
+  });
+  return { promise, resolve, reject };
+}
+
 describe('oauth routes', { timeout: 15_000 }, () => {
   let app: FastifyInstance;
   let db: DbModule['db'];
@@ -599,6 +609,118 @@ describe('oauth routes', { timeout: 15_000 }, () => {
         idToken: originalJwt,
       },
     });
+
+    const accounts = await db.select().from(schema.accounts).all();
+    expect(accounts).toHaveLength(1);
+    expect(accounts[0]?.oauthAccountKey).toBe('chatgpt-account-existing');
+  });
+
+  it('keeps the existing codex account non-active while rebind model discovery is still pending', async () => {
+    const originalJwt = buildJwt({
+      email: 'codex-existing@example.com',
+      'https://api.openai.com/auth': {
+        chatgpt_account_id: 'chatgpt-account-existing',
+        chatgpt_plan_type: 'team',
+      },
+    });
+    const site = await db.insert(schema.sites).values({
+      name: 'ChatGPT Codex OAuth',
+      url: 'https://chatgpt.com/backend-api/codex',
+      platform: 'codex',
+      status: 'active',
+    }).returning().get();
+    const existing = await db.insert(schema.accounts).values({
+      siteId: site.id,
+      username: 'codex-existing@example.com',
+      accessToken: 'stable-access-token',
+      apiToken: null,
+      status: 'active',
+      oauthProvider: 'codex',
+      oauthAccountKey: 'chatgpt-account-existing',
+      extraConfig: JSON.stringify({
+        credentialMode: 'session',
+        oauth: {
+          provider: 'codex',
+          accountId: 'chatgpt-account-existing',
+          accountKey: 'chatgpt-account-existing',
+          email: 'codex-existing@example.com',
+          planType: 'team',
+          refreshToken: 'stable-refresh-token',
+          idToken: originalJwt,
+        },
+      }),
+    }).returning().get();
+
+    const reboundJwt = buildJwt({
+      email: 'codex-existing@example.com',
+      'https://api.openai.com/auth': {
+        chatgpt_account_id: 'chatgpt-account-rebound',
+        chatgpt_plan_type: 'plus',
+      },
+    });
+    const discoveryGate = createDeferred<ResponseLike>();
+    fetchMock
+      .mockResolvedValueOnce({
+        ok: true,
+        status: 200,
+        json: async () => ({
+          access_token: 'rebound-access-token',
+          refresh_token: 'rebound-refresh-token',
+          id_token: reboundJwt,
+          expires_in: 3600,
+          token_type: 'Bearer',
+        }),
+        text: async () => JSON.stringify({ ok: true }),
+      })
+      .mockImplementationOnce(() => discoveryGate.promise);
+
+    const startResponse = await app.inject({
+      method: 'POST',
+      url: '/api/oauth/providers/codex/start',
+      headers: {
+        host: 'metapi.example',
+        'x-forwarded-proto': 'https',
+      },
+      payload: {
+        accountId: existing.id,
+      },
+    });
+    const startBody = startResponse.json() as { state: string };
+
+    const callbackPromise = app.inject({
+      method: 'POST',
+      url: `/api/oauth/sessions/${encodeURIComponent(startBody.state)}/manual-callback`,
+      payload: {
+        callbackUrl: `http://localhost:1455/auth/callback?state=${encodeURIComponent(startBody.state)}&code=oauth-code-rebind-pending`,
+      },
+    });
+
+    while (fetchMock.mock.calls.length < 2) {
+      await new Promise((resolve) => setTimeout(resolve, 0));
+    }
+
+    const pendingRow = await db.select().from(schema.accounts).where(eq(schema.accounts.id, existing.id)).get();
+    expect(pendingRow?.status).toBe('disabled');
+
+    discoveryGate.resolve({
+      ok: true,
+      status: 200,
+      json: async () => ({
+        models: [{ id: 'gpt-5.4' }],
+      }),
+      text: async () => JSON.stringify({ ok: true }),
+    } as ResponseLike);
+
+    const callbackResponse = await callbackPromise;
+    expect(callbackResponse.statusCode).toBe(200);
+
+    const stored = await db.select().from(schema.accounts).where(eq(schema.accounts.id, existing.id)).get();
+    expect(stored).toMatchObject({
+      id: existing.id,
+      status: 'active',
+      accessToken: 'rebound-access-token',
+      oauthAccountKey: 'chatgpt-account-rebound',
+    });
   });
 
   it('fails codex oauth onboarding when token exchange does not expose chatgpt_account_id', async () => {

src/server/routes/proxy/responses.codex-oauth.test.ts

@@ -255,6 +255,33 @@ describe('responses proxy codex oauth refresh', () => {
     expect(response.json()?.output_text).toContain('ok after codex forbidden refresh');
   });
 
+  it('does not refresh codex oauth token on non-auth 403 responses', async () => {
+    fetchMock.mockResolvedValueOnce(new Response(JSON.stringify({
+      error: { message: 'quota exceeded for workspace', type: 'usage_limit_reached' },
+    }), {
+      status: 403,
+      headers: { 'content-type': 'application/json' },
+    }));
+
+    const response = await app.inject({
+      method: 'POST',
+      url: '/v1/responses',
+      payload: {
+        model: 'gpt-5.2-codex',
+        input: 'hello codex',
+      },
+    });
+
+    expect(response.statusCode).toBe(403);
+    expect(refreshOauthAccessTokenSingleflightMock).not.toHaveBeenCalled();
+    expect(fetchMock).toHaveBeenCalledTimes(1);
+    expect(response.json()).toMatchObject({
+      error: {
+        message: expect.stringContaining('quota exceeded for workspace'),
+      },
+    });
+  });
+
   it('retries oauth responses requests with a normalized upstream URL after refresh', async () => {
     selectChannelMock.mockReturnValue({
       channel: { id: 11, routeId: 22 },

src/server/routes/proxy/responses.websocket.test.ts

@@ -194,6 +194,16 @@ function waitForSocketMessages(socket: WebSocket, count: number, timeoutMs = 100
   });
 }
 
+function createDeferred<T>() {
+  let resolve!: (value: T | PromiseLike<T>) => void;
+  let reject!: (reason?: unknown) => void;
+  const promise = new Promise<T>((res, rej) => {
+    resolve = res;
+    reject = rej;
+  });
+  return { promise, resolve, reject };
+}
+
 function createClientSocket(baseUrl: string, headers: Record<string, string> = {}) {
   return new WebSocket(`${baseUrl}/v1/responses`, {
     headers: {
@@ -203,6 +213,10 @@ function createClientSocket(baseUrl: string, headers: Record<string, string> = {
   });
 }
 
+function createClientSocketForPath(path: string, headers: Record<string, string> = {}) {
+  return new WebSocket(path, { headers });
+}
+
 describe('responses websocket transport', () => {
   const originalCodexResponsesWebsocketBeta = config.codexResponsesWebsocketBeta;
   let app: FastifyInstance;
@@ -545,6 +559,54 @@ describe('responses websocket transport', () => {
     expect(messages[0]?.response?.id).toBe('resp_http_fallback_401');
   });
 
+  it('preserves query parameter auth when websocket transport falls back to the HTTP responses route', async () => {
+    rejectedUpgradeStatus = 401;
+    rejectedUpgradeStatusText = 'Unauthorized';
+    rejectedUpgradeBody = JSON.stringify({
+      error: {
+        message: 'expired token',
+        type: 'invalid_request_error',
+      },
+    });
+    const selectedChannel = createSelectedChannel({
+      siteUrl: rejectedUpgradeSiteUrl,
+    });
+    selectChannelMock.mockReturnValue(selectedChannel);
+    previewSelectedChannelMock.mockResolvedValue(selectedChannel);
+    authorizeDownstreamTokenMock.mockResolvedValueOnce({
+      ok: true,
+      source: 'global',
+      token: 'sk-query-auth',
+      key: null,
+      policy: {
+        supportedModels: [],
+        allowedRouteIds: [],
+        siteWeightMultipliers: {},
+      },
+    });
+    fetchMock.mockResolvedValueOnce(createSseResponse([
+      'event: response.completed\n',
+      'data: {"type":"response.completed","response":{"id":"resp_http_fallback_query","model":"gpt-5.4","status":"completed","output":[],"usage":{"input_tokens":1,"output_tokens":1,"total_tokens":2}}}\n\n',
+      'data: [DONE]\n\n',
+    ]));
+
+    const socket = createClientSocketForPath(`${baseUrl}/v1/responses?key=sk-query-auth`);
+    await waitForSocketOpen(socket);
+    const messagesPromise = waitForSocketMessages(socket, 1);
+
+    socket.send(JSON.stringify({
+      type: 'response.create',
+      model: 'gpt-5.4',
+      input: [],
+    }));
+
+    const messages = await messagesPromise;
+    socket.close();
+
+    expect(fetchMock).toHaveBeenCalledTimes(1);
+    expect(messages[0]?.response?.id).toBe('resp_http_fallback_query');
+  });
+
   it('rejects websocket turns whose model is blocked by the downstream key policy before channel selection', async () => {
     authorizeDownstreamTokenMock.mockResolvedValueOnce({
       ok: true,
@@ -678,6 +740,111 @@ describe('responses websocket transport', () => {
     ]);
   });
 
+  it('serializes websocket messages per connection so follow-up turns wait for the previous HTTP fallback to finish', async () => {
+    const selectedChannel = createSelectedChannel({
+      sitePlatform: 'openai',
+      actualModel: 'gpt-4.1',
+    });
+    selectChannelMock.mockReturnValue(selectedChannel);
+    previewSelectedChannelMock.mockResolvedValue(selectedChannel);
+    const firstResponseGate = createDeferred<Response>();
+    fetchMock
+      .mockImplementationOnce(() => firstResponseGate.promise)
+      .mockResolvedValueOnce(createSseResponse([
+        'event: response.completed\n',
+        'data: {"type":"response.completed","response":{"id":"resp_ws_2","model":"gpt-4.1","status":"completed","output":[{"id":"msg_2","type":"message","role":"assistant","status":"completed","content":[{"type":"output_text","text":"done"}]}],"usage":{"input_tokens":5,"output_tokens":1,"total_tokens":6}}}\n\n',
+        'data: [DONE]\n\n',
+      ]));
+
+    const socket = createClientSocket(baseUrl);
+    await waitForSocketOpen(socket);
+
+    socket.send(JSON.stringify({
+      type: 'response.create',
+      model: 'gpt-4.1',
+      instructions: 'be helpful',
+      input: [
+        {
+          id: 'msg_user_1',
+          type: 'message',
+          role: 'user',
+          content: [{ type: 'input_text', text: 'call the tool' }],
+        },
+      ],
+    }));
+
+    while (fetchMock.mock.calls.length < 1) {
+      await new Promise((resolve) => setTimeout(resolve, 0));
+    }
+
+    socket.send(JSON.stringify({
+      type: 'response.create',
+      previous_response_id: 'resp_ws_1',
+      input: [
+        {
+          id: 'tool_out_1',
+          type: 'function_call_output',
+          call_id: 'call_1',
+          output: 'tool result',
+        },
+      ],
+    }));
+
+    await new Promise((resolve) => setTimeout(resolve, 0));
+    expect(fetchMock).toHaveBeenCalledTimes(1);
+
+    const messagesPromise = waitForSocketMessages(socket, 4);
+    firstResponseGate.resolve(createSseResponse([
+      'event: response.output_item.done\n',
+      'data: {"type":"response.output_item.done","output_index":0,"item":{"id":"fc_1","type":"function_call","call_id":"call_1"}}\n\n',
+      'event: response.output_item.done\n',
+      'data: {"type":"response.output_item.done","output_index":1,"item":{"id":"msg_1","type":"message","role":"assistant","status":"completed","content":[{"type":"output_text","text":"call tool"}]}}\n\n',
+      'event: response.completed\n',
+      'data: {"type":"response.completed","response":{"id":"resp_ws_1","model":"gpt-4.1","status":"completed","output":[{"id":"fc_1","type":"function_call","call_id":"call_1"},{"id":"msg_1","type":"message","role":"assistant","status":"completed","content":[{"type":"output_text","text":"call tool"}]}],"usage":{"input_tokens":3,"output_tokens":1,"total_tokens":4}}}\n\n',
+      'data: [DONE]\n\n',
+    ]));
+
+    const messages = await messagesPromise;
+    socket.close();
+
+    expect(fetchMock).toHaveBeenCalledTimes(2);
+    expect(messages.map((message) => message.type)).toEqual([
+      'response.output_item.done',
+      'response.output_item.done',
+      'response.completed',
+      'response.completed',
+    ]);
+    const [, secondOptions] = fetchMock.mock.calls[1] as [string, RequestInit];
+    const secondBody = JSON.parse(String(secondOptions.body));
+    expect(secondBody.input).toEqual([
+      {
+        id: 'msg_user_1',
+        type: 'message',
+        role: 'user',
+        content: [{ type: 'input_text', text: 'call the tool' }],
+      },
+      {
+        id: 'fc_1',
+        type: 'function_call',
+        call_id: 'call_1',
+        status: 'completed',
+      },
+      {
+        id: 'msg_1',
+        type: 'message',
+        role: 'assistant',
+        status: 'completed',
+        content: [{ type: 'output_text', text: 'call tool' }],
+      },
+      {
+        id: 'tool_out_1',
+        type: 'function_call_output',
+        call_id: 'call_1',
+        output: 'tool result',
+      },
+    ]);
+  });
+
   it('preserves incremental response.create payloads with previous_response_id for websocket-capable upstreams', async () => {
     const selectedChannel = createSelectedChannel({
       siteUrl: upstreamSiteUrl,

src/server/routes/proxy/responsesWebsocket.ts

@@ -339,15 +339,25 @@ async function forwardResponsesRequestViaHttp(input: {
   request: IncomingMessage;
   payload: Record<string, unknown>;
   preserveIncrementalMode: boolean;
+  authToken: string;
 }): Promise<unknown[] | null> {
+  const injectHeaders = {
+    ...buildInjectHeaders(input.request),
+    [RESPONSES_WEBSOCKET_TRANSPORT_HEADER]: '1',
+    ...(input.preserveIncrementalMode ? { [RESPONSES_WEBSOCKET_MODE_HEADER]: 'incremental' } : {}),
+  };
+  if (
+    !headerValueToTrimmedString(injectHeaders.authorization)
+    && !headerValueToTrimmedString(injectHeaders['x-api-key'])
+    && !headerValueToTrimmedString(injectHeaders['x-goog-api-key'])
+  ) {
+    injectHeaders.authorization = `Bearer ${input.authToken}`;
+  }
+
   const response = await input.app.inject({
     method: 'POST',
     url: '/v1/responses',
-    headers: {
+    headers: injectHeaders,
-      ...buildInjectHeaders(input.request),
-      [RESPONSES_WEBSOCKET_TRANSPORT_HEADER]: '1',
-      ...(input.preserveIncrementalMode ? { [RESPONSES_WEBSOCKET_MODE_HEADER]: 'incremental' } : {}),
-    },
     payload: input.payload,
   });
 
@@ -444,7 +454,7 @@ function writeUpgradeHttpError(socket: Duplex, status: number, message: string):
         ? 'Bad Request'
         : 'Error';
   const body = JSON.stringify({ error: message });
-  socket.write(
+  socket.end(
     `HTTP/1.1 ${status} ${statusText}\r\n`
     + 'Content-Type: application/json\r\n'
     + `Content-Length: ${Buffer.byteLength(body)}\r\n`
@@ -452,7 +462,6 @@ function writeUpgradeHttpError(socket: Duplex, status: number, message: string):
     + '\r\n'
     + body,
   );
-  socket.destroy();
 }
 
 async function supportsResponsesWebsocketIncrementalInput(
@@ -483,148 +492,155 @@ async function handleResponsesWebsocketConnection(
   let lastRequest: Record<string, unknown> | null = null;
   let lastResponseOutput: unknown[] = [];
   let selectedChannel: SelectedChannel | null = null;
+  let messageQueue = Promise.resolve();
 
   socket.once('close', () => {
     void codexWebsocketRuntime.closeSession(websocketSessionId);
   });
 
-  socket.on('message', async (raw) => {
+  socket.on('message', (raw) => {
-    try {
+    messageQueue = messageQueue
-      const parsed = parseJsonObject(raw);
+      .catch(() => undefined)
-      if (!parsed) {
+      .then(async () => {
-        writeResponsesWebsocketError(socket, 400, 'Invalid websocket JSON payload');
-        return;
-      }
-
-      const requestModel = asTrimmedString(parsed.model) || asTrimmedString(lastRequest?.model);
-      if (requestModel && !await isModelAllowedByPolicyOrAllowedRoutes(requestModel, authContext.policy)) {
-        writeResponsesWebsocketError(socket, 403, 'model is not allowed for this downstream key');
-        return;
-      }
-      const supportsIncrementalInput = selectedChannelSupportsIncrementalInput(selectedChannel, requestModel)
-        || await supportsResponsesWebsocketIncrementalInput(parsed, lastRequest, authContext);
-      const shouldHandleLocalPrewarm = shouldHandleResponsesWebsocketPrewarmLocally(
-        parsed,
-        lastRequest,
-        supportsIncrementalInput,
-      );
-      const normalized = normalizeResponsesWebsocketRequest(
-        parsed,
-        lastRequest,
-        lastResponseOutput,
-        supportsIncrementalInput,
-      );
-      if (!normalized.ok) {
-        writeResponsesWebsocketError(socket, normalized.status, normalized.message);
-        return;
-      }
-
-      if (authContext.source === 'managed' && authContext.key?.id) {
-        await consumeManagedKeyRequest(authContext.key.id);
-      }
-
-      lastRequest = normalized.nextRequestSnapshot;
-      if (shouldHandleLocalPrewarm) {
-        lastResponseOutput = [];
-        for (const payload of synthesizePrewarmResponsePayloads(normalized.request)) {
-          socket.send(JSON.stringify(payload));
-        }
-        return;
-      }
-
-      if (!shouldReuseSelectedChannel(selectedChannel, requestModel)) {
-        selectedChannel = requestModel
-          ? await tokenRouter.selectChannel(requestModel, authContext.policy)
-          : null;
-      }
-
-      const codexWebsocketChannel = selectedChannelSupportsCodexWebsocketTransport(selectedChannel, requestModel)
-        ? selectedChannel
-        : null;
-
-      if (codexWebsocketChannel) {
-        const downstreamHeaders: Record<string, unknown> = {
-          ...(request.headers as Record<string, unknown>),
-          [RESPONSES_WEBSOCKET_TRANSPORT_HEADER]: '1',
-          ...(supportsIncrementalInput ? { [RESPONSES_WEBSOCKET_MODE_HEADER]: 'incremental' } : {}),
-        };
-        const providerHeaders = buildOauthProviderHeaders({
-          extraConfig: typeof codexWebsocketChannel.account.extraConfig === 'string'
-            ? codexWebsocketChannel.account.extraConfig
-            : null,
-          downstreamHeaders,
-        });
-        const prepared = buildUpstreamEndpointRequest({
-          endpoint: 'responses',
-          modelName: asTrimmedString(codexWebsocketChannel.actualModel) || requestModel,
-          stream: true,
-          tokenValue: codexWebsocketChannel.tokenValue,
-          sitePlatform: codexWebsocketChannel.site.platform,
-          siteUrl: codexWebsocketChannel.site.url,
-          openaiBody: normalized.request,
-          downstreamFormat: 'responses',
-          responsesOriginalBody: normalized.request,
-          downstreamHeaders,
-          providerHeaders,
-          codexExplicitSessionId: deriveCodexExplicitSessionId(normalized.request, websocketSessionId),
-        });
-        const requestUrl = `${codexWebsocketChannel.site.url.replace(/\/+$/, '')}${prepared.path}`;
-
         try {
-          const runtimeResult = await codexWebsocketRuntime.sendRequest({
+          const parsed = parseJsonObject(raw);
-            sessionId: websocketSessionId,
+          if (!parsed) {
-            requestUrl,
+            writeResponsesWebsocketError(socket, 400, 'Invalid websocket JSON payload');
-            headers: prepared.headers,
+            return;
-            body: prepared.body,
-          });
-          lastResponseOutput = collectResponsesOutput(runtimeResult.events);
-          for (const payload of runtimeResult.events) {
-            socket.send(JSON.stringify(payload));
           }
-        } catch (error) {
+
-          const runtimeError = error instanceof CodexWebsocketRuntimeError
+          const requestModel = asTrimmedString(parsed.model) || asTrimmedString(lastRequest?.model);
-            ? error
+          if (requestModel && !await isModelAllowedByPolicyOrAllowedRoutes(requestModel, authContext.policy)) {
-            : new CodexWebsocketRuntimeError('upstream websocket request failed');
+            writeResponsesWebsocketError(socket, 403, 'model is not allowed for this downstream key');
-          if (runtimeError.status && runtimeError.events.length === 0) {
+            return;
-            const fallbackOutput = await forwardResponsesRequestViaHttp({
+          }
-              app,
+          const supportsIncrementalInput = selectedChannelSupportsIncrementalInput(selectedChannel, requestModel)
-              socket,
+            || await supportsResponsesWebsocketIncrementalInput(parsed, lastRequest, authContext);
-              request,
+          const shouldHandleLocalPrewarm = shouldHandleResponsesWebsocketPrewarmLocally(
-              payload: normalized.request,
+            parsed,
-              preserveIncrementalMode: false,
+            lastRequest,
-            });
+            supportsIncrementalInput,
-            if (fallbackOutput) {
+          );
-              lastResponseOutput = fallbackOutput;
+          const normalized = normalizeResponsesWebsocketRequest(
+            parsed,
+            lastRequest,
+            lastResponseOutput,
+            supportsIncrementalInput,
+          );
+          if (!normalized.ok) {
+            writeResponsesWebsocketError(socket, normalized.status, normalized.message);
+            return;
+          }
+
+          if (authContext.source === 'managed' && authContext.key?.id) {
+            await consumeManagedKeyRequest(authContext.key.id);
+          }
+
+          lastRequest = normalized.nextRequestSnapshot;
+          if (shouldHandleLocalPrewarm) {
+            lastResponseOutput = [];
+            for (const payload of synthesizePrewarmResponsePayloads(normalized.request)) {
+              socket.send(JSON.stringify(payload));
             }
             return;
           }
-          lastResponseOutput = collectResponsesOutput(runtimeError.events);
-          for (const payload of runtimeError.events) {
-            socket.send(JSON.stringify(payload));
-          }
-          writeResponsesWebsocketError(
-            socket,
-            runtimeError.status || 408,
-            runtimeError.message,
-            runtimeError.payload,
-          );
-        }
-        return;
-      }
 
-      const forwardedOutput = await forwardResponsesRequestViaHttp({
+          if (!shouldReuseSelectedChannel(selectedChannel, requestModel)) {
-        app,
+            selectedChannel = requestModel
-        socket,
+              ? await tokenRouter.selectChannel(requestModel, authContext.policy)
-        request,
+              : null;
-        payload: normalized.request,
+          }
-        preserveIncrementalMode: supportsIncrementalInput,
+
+          const codexWebsocketChannel = selectedChannelSupportsCodexWebsocketTransport(selectedChannel, requestModel)
+            ? selectedChannel
+            : null;
+
+          if (codexWebsocketChannel) {
+            const downstreamHeaders: Record<string, unknown> = {
+              ...(request.headers as Record<string, unknown>),
+              [RESPONSES_WEBSOCKET_TRANSPORT_HEADER]: '1',
+              ...(supportsIncrementalInput ? { [RESPONSES_WEBSOCKET_MODE_HEADER]: 'incremental' } : {}),
+            };
+            const providerHeaders = buildOauthProviderHeaders({
+              extraConfig: typeof codexWebsocketChannel.account.extraConfig === 'string'
+                ? codexWebsocketChannel.account.extraConfig
+                : null,
+              downstreamHeaders,
+            });
+            const prepared = buildUpstreamEndpointRequest({
+              endpoint: 'responses',
+              modelName: asTrimmedString(codexWebsocketChannel.actualModel) || requestModel,
+              stream: true,
+              tokenValue: codexWebsocketChannel.tokenValue,
+              sitePlatform: codexWebsocketChannel.site.platform,
+              siteUrl: codexWebsocketChannel.site.url,
+              openaiBody: normalized.request,
+              downstreamFormat: 'responses',
+              responsesOriginalBody: normalized.request,
+              downstreamHeaders,
+              providerHeaders,
+              codexExplicitSessionId: deriveCodexExplicitSessionId(normalized.request, websocketSessionId),
+            });
+            const requestUrl = `${codexWebsocketChannel.site.url.replace(/\/+$/, '')}${prepared.path}`;
+
+            try {
+              const runtimeResult = await codexWebsocketRuntime.sendRequest({
+                sessionId: websocketSessionId,
+                requestUrl,
+                headers: prepared.headers,
+                body: prepared.body,
+              });
+              lastResponseOutput = collectResponsesOutput(runtimeResult.events);
+              for (const payload of runtimeResult.events) {
+                socket.send(JSON.stringify(payload));
+              }
+            } catch (error) {
+              const runtimeError = error instanceof CodexWebsocketRuntimeError
+                ? error
+                : new CodexWebsocketRuntimeError('upstream websocket request failed');
+              if (runtimeError.status && runtimeError.events.length === 0) {
+                const fallbackOutput = await forwardResponsesRequestViaHttp({
+                  app,
+                  socket,
+                  request,
+                  payload: normalized.request,
+                  preserveIncrementalMode: false,
+                  authToken: authContext.token,
+                });
+                if (fallbackOutput) {
+                  lastResponseOutput = fallbackOutput;
+                }
+                return;
+              }
+              lastResponseOutput = collectResponsesOutput(runtimeError.events);
+              for (const payload of runtimeError.events) {
+                socket.send(JSON.stringify(payload));
+              }
+              writeResponsesWebsocketError(
+                socket,
+                runtimeError.status || 408,
+                runtimeError.message,
+                runtimeError.payload,
+              );
+            }
+            return;
+          }
+
+          const forwardedOutput = await forwardResponsesRequestViaHttp({
+            app,
+            socket,
+            request,
+            payload: normalized.request,
+            preserveIncrementalMode: supportsIncrementalInput,
+            authToken: authContext.token,
+          });
+          if (forwardedOutput) {
+            lastResponseOutput = forwardedOutput;
+          }
+        } catch {
+          writeResponsesWebsocketError(socket, 500, 'internal websocket proxy error');
+        }
       });
-      if (forwardedOutput) {
-        lastResponseOutput = forwardedOutput;
-      }
-    } catch {
-      writeResponsesWebsocketError(socket, 500, 'internal websocket proxy error');
-    }
   });
 }
 

src/server/services/modelService.ts

@@ -481,7 +481,10 @@ function buildSuccessfulRefreshResult(input: {
   };
 }
 
-export async function refreshModelsForAccount(accountId: number): Promise<ModelRefreshResult> {
+export async function refreshModelsForAccount(
+  accountId: number,
+  options?: { allowInactive?: boolean },
+): Promise<ModelRefreshResult> {
   const row = await db.select().from(schema.accounts)
     .innerJoin(schema.sites, eq(schema.accounts.siteId, schema.sites.id))
     .where(eq(schema.accounts.id, accountId))
@@ -516,7 +519,7 @@ export async function refreshModelsForAccount(accountId: number): Promise<ModelR
     return buildSkippedRefreshResult(accountId, 'site_disabled', '站点已禁用');
   }
 
-  if (account.status !== 'active') {
+  if (account.status !== 'active' && !options?.allowInactive) {
     return buildSkippedRefreshResult(accountId, 'adapter_or_status', '平台不可用或账号未激活');
   }
 

src/server/services/oauth/service.ts

@@ -242,7 +242,7 @@ async function upsertOauthAccount(input: {
       accessToken: input.exchange.accessToken,
       apiToken: null,
       checkinEnabled: false,
-      status: 'active',
+      status: 'disabled',
       oauthProvider: input.definition.metadata.provider,
       oauthAccountKey: oauth.accountKey || oauth.accountId || null,
       oauthProjectId: oauth.projectId || null,
@@ -376,7 +376,10 @@ export async function handleOauthCallback(input: {
       throw new Error('failed to persist oauth account');
     }
 
-    const refreshResult = await refreshModelsForAccount(account.id);
+    const refreshResult = await refreshModelsForAccount(
+      account.id,
+      previousAccount ? { allowInactive: true } : undefined,
+    );
     if (refreshResult.status !== 'success') {
       if (created) {
         await db.delete(schema.accounts).where(eq(schema.accounts.id, account.id)).run();
@@ -401,6 +404,13 @@ export async function handleOauthCallback(input: {
       throw new Error(errorMessage);
     }
 
+    if (previousAccount) {
+      await db.update(schema.accounts).set({
+        status: 'active',
+        updatedAt: new Date().toISOString(),
+      }).where(eq(schema.accounts.id, account.id)).run();
+    }
+
     await rebuildTokenRoutesFromAvailability();
     markOauthSessionSuccess(input.state, {
       accountId: account.id,

src/server/services/tokenRouter.siteStatus.test.ts

@@ -209,4 +209,48 @@ describe('TokenRouter site status guard', () => {
     expect(selected).not.toBeNull();
     expect(selected?.tokenValue).toBe('oauth-access-token');
   });
+
+  it('does not fall back to a stale legacy api token when an oauth-managed row has no access token', async () => {
+    const site = await db.insert(schema.sites).values({
+      name: 'codex-site',
+      url: 'https://chatgpt.com/backend-api/codex',
+      platform: 'codex',
+      status: 'active',
+    }).returning().get();
+
+    const account = await db.insert(schema.accounts).values({
+      siteId: site.id,
+      username: 'codex-user@example.com',
+      accessToken: '',
+      apiToken: 'stale-direct-token',
+      status: 'active',
+      extraConfig: JSON.stringify({
+        credentialMode: 'session',
+        oauth: {
+          provider: 'codex',
+          accountId: 'chatgpt-account-123',
+          email: 'codex-user@example.com',
+          planType: 'plus',
+        },
+      }),
+    }).returning().get();
+
+    const route = await db.insert(schema.tokenRoutes).values({
+      modelPattern: 'gpt-5.2-codex',
+      enabled: true,
+    }).returning().get();
+
+    await db.insert(schema.routeChannels).values({
+      routeId: route.id,
+      accountId: account.id,
+      tokenId: null,
+      priority: 0,
+      weight: 10,
+      enabled: true,
+      manualOverride: false,
+    }).run();
+
+    const selected = await new TokenRouter().selectChannel('gpt-5.2-codex');
+    expect(selected).toBeNull();
+  });
 });