[codex] fix responses lifecycle parity and codex session serialization (#278)

* fix responses lifecycle parity

* remove websocket test debug logging

* fix responses proxy stream typecheck

* update compact upstream response expectations

* add community and deepwiki badges

* update linuxdo banner link

* fix responses review follow-ups

* preserve whitespace in responses terminal deltas

* fix responses stream reconciliation edges

* fix responses lifecycle parity and codex session serialization

* fix review follow-up edge cases

---------

Co-authored-by: apple <apple@appledeMacBook-Pro.local>

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

@@ -0,0 +1,30 @@
+const codexHttpSessionQueues = new Map<string, Promise<void>>();
+
+export async function runCodexHttpSessionTask<T>(
+  sessionId: string,
+  task: () => Promise<T>,
+): Promise<T> {
+  const normalizedSessionId = sessionId.trim();
+  if (!normalizedSessionId) {
+    return task();
+  }
+
+  const previous = codexHttpSessionQueues.get(normalizedSessionId) || Promise.resolve();
+  const run = previous
+    .catch(() => undefined)
+    .then(() => task());
+  const tail = run.then(() => undefined, () => undefined);
+  codexHttpSessionQueues.set(normalizedSessionId, tail);
+
+  try {
+    return await run;
+  } finally {
+    if (codexHttpSessionQueues.get(normalizedSessionId) === tail) {
+      codexHttpSessionQueues.delete(normalizedSessionId);
+    }
+  }
+}
+
+export function resetCodexHttpSessionQueue(): void {
+  codexHttpSessionQueues.clear();
+}

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

@@ -40,6 +40,7 @@ import {
   unwrapGeminiCliPayload,
 } from '../../routes/proxy/geminiCliCompat.js';
 import { summarizeConversationFileInputsInOpenAiBody } from '../capabilities/conversationFileCapabilities.js';
+import { readRuntimeResponseText } from '../executors/types.js';
 import { detectDownstreamClientContext } from '../../routes/proxy/downstreamClientContext.js';
 import { canRetryProxyChannel, getProxyMaxChannelRetries } from '../../services/proxyChannelRetry.js';
 import {
@@ -394,7 +395,7 @@ export async function handleChatSurfaceRequest(
         });
         let rawText = '';
         if (!upstreamContentType.includes('text/event-stream')) {
-          const fallbackText = await upstream.text();
+          const fallbackText = await readRuntimeResponseText(upstream);
           rawText = fallbackText;
           if (looksLikeResponsesSseText(fallbackText)) {
             startSseResponse();
@@ -578,7 +579,7 @@ export async function handleChatSurfaceRequest(
         rawText = collected.rawText;
         upstreamData = collected.payload;
       } else {
-        rawText = await upstream.text();
+        rawText = await readRuntimeResponseText(upstream);
         if (looksLikeResponsesSseText(rawText)) {
           upstreamData = collectResponsesFinalPayloadFromSseText(rawText, modelName).payload;
         } else {

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

@@ -33,6 +33,8 @@ import {
   unwrapGeminiCliPayload,
 } from '../../routes/proxy/geminiCliCompat.js';
 import { isCodexResponsesSurface } from '../cliProfiles/codexProfile.js';
+import { readRuntimeResponseText } from '../executors/types.js';
+import { runCodexHttpSessionTask } from '../runtime/codexHttpSessionQueue.js';
 import {
   summarizeConversationFileInputsInOpenAiBody,
   summarizeConversationFileInputsInResponsesBody,
@@ -56,6 +58,15 @@ function isRecord(value: unknown): value is Record<string, unknown> {
   return !!value && typeof value === 'object';
 }
 
+function getCodexSessionHeaderValue(headers: Record<string, string>): string {
+  for (const [rawKey, rawValue] of Object.entries(headers)) {
+    const normalizedKey = rawKey.trim().toLowerCase();
+    if (normalizedKey === 'session_id' || normalizedKey === 'session-id') {
+      return String(rawValue || '').trim();
+    }
+  }
+  return '';
+}
 function isResponsesWebsocketTransportRequest(headers: Record<string, unknown>): boolean {
   return Object.entries(headers)
     .some(([rawKey, rawValue]) => rawKey.trim().toLowerCase() === 'x-metapi-responses-websocket-transport'
@@ -324,10 +335,23 @@ export async function handleOpenAiResponsesSurfaceRequest(
           runtime: endpointRequest.runtime,
         };
       };
-      const dispatchRequest = createSurfaceDispatchRequest({
+      const baseDispatchRequest = createSurfaceDispatchRequest({
         site: selected.site,
         accountExtraConfig: selected.account.extraConfig,
       });
+      const dispatchRequest = (
+        endpointRequest: BuiltEndpointRequest,
+        targetUrl?: string,
+      ) => {
+        if (!isCodexSite || endpointRequest.path !== '/responses') {
+          return baseDispatchRequest(endpointRequest, targetUrl);
+        }
+        const sessionId = getCodexSessionHeaderValue(endpointRequest.headers);
+        return runCodexHttpSessionTask(
+          sessionId,
+          () => baseDispatchRequest(endpointRequest, targetUrl),
+        );
+      };
       const endpointStrategy = openAiResponsesTransformer.compatibility.createEndpointStrategy({
         isStream: isStream || isCodexSite,
         requiresNativeResponsesFileUrl,
@@ -507,7 +531,7 @@ export async function handleOpenAiResponsesSurfaceRequest(
             },
           });
           if (!upstreamContentType.includes('text/event-stream')) {
-            const rawText = await upstream.text();
+            const rawText = await readRuntimeResponseText(upstream);
             if (looksLikeResponsesSseText(rawText)) {
               startSseResponse();
               const streamResult = await streamSession.run(
@@ -614,7 +638,7 @@ export async function handleOpenAiResponsesSurfaceRequest(
 
           let replayReader: ReturnType<typeof createSingleChunkStreamReader> | null = null;
           if (websocketTransportRequest) {
-            const rawText = await upstream.text();
+            const rawText = await readRuntimeResponseText(upstream);
             if (looksLikeResponsesSseText(rawText)) {
               try {
                 const collectedPayload = collectResponsesFinalPayloadFromSseText(rawText, modelName).payload;
@@ -745,7 +769,7 @@ export async function handleOpenAiResponsesSurfaceRequest(
           rawText = collected.rawText;
           upstreamData = collected.payload;
         } else {
-          rawText = await upstream.text();
+          rawText = await readRuntimeResponseText(upstream);
           if (looksLikeResponsesSseText(rawText)) {
             upstreamData = collectResponsesFinalPayloadFromSseText(rawText, modelName).payload;
           } else {

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

@@ -1,3 +1,4 @@
+import { zstdCompressSync } from 'node:zlib';
 import Fastify, { type FastifyInstance } from 'fastify';
 import { afterAll, beforeAll, beforeEach, describe, expect, it, vi } from 'vitest';
 
@@ -191,4 +192,166 @@ describe('testRoutes proxy tester transport', () => {
     expect(url).toMatch(/\/v1\/files$/);
     expect(requestInit.body?.constructor?.name).toBe('FormData');
   });
+
+  it('decodes zstd-compressed buffered proxy test responses', async () => {
+    const payload = JSON.stringify({
+      choices: [
+        {
+          message: {
+            content: '你好，来自压缩响应',
+          },
+        },
+      ],
+    });
+    fetchMock.mockResolvedValue(new Response(zstdCompressSync(Buffer.from(payload)), {
+      status: 200,
+      headers: {
+        'content-encoding': 'zstd',
+        'content-type': 'application/json; charset=utf-8',
+      },
+    }));
+
+    const response = await app.inject({
+      method: 'POST',
+      url: '/api/test/proxy',
+      payload: {
+        method: 'POST',
+        path: '/v1/chat/completions',
+        requestKind: 'json',
+        stream: false,
+        jobMode: false,
+        rawMode: false,
+        jsonBody: {
+          model: 'gpt-4o-mini',
+          messages: [{ role: 'user', content: 'ping' }],
+        },
+      },
+    });
+
+    expect(response.statusCode).toBe(200);
+    expect(response.json()).toEqual({
+      choices: [
+        {
+          message: {
+            content: '你好，来自压缩响应',
+          },
+        },
+      ],
+    });
+  });
+
+  it('decodes zstd-compressed non-SSE stream fallback responses', async () => {
+    const payload = JSON.stringify({
+      choices: [
+        {
+          message: {
+            content: '你好，来自流式回退',
+          },
+        },
+      ],
+    });
+    fetchMock.mockResolvedValue(new Response(zstdCompressSync(Buffer.from(payload)), {
+      status: 200,
+      headers: {
+        'content-encoding': 'zstd',
+        'content-type': 'application/json; charset=utf-8',
+      },
+    }));
+
+    const response = await app.inject({
+      method: 'POST',
+      url: '/api/test/proxy/stream',
+      payload: {
+        method: 'POST',
+        path: '/v1/chat/completions',
+        requestKind: 'json',
+        stream: true,
+        jobMode: false,
+        rawMode: false,
+        jsonBody: {
+          model: 'gpt-4o-mini',
+          messages: [{ role: 'user', content: 'ping' }],
+        },
+      },
+    });
+
+    expect(response.statusCode).toBe(200);
+    expect(response.headers['content-type']).toContain('text/event-stream');
+    expect(response.payload).toBe(`data: ${payload}\n\ndata: [DONE]\n\n`);
+  });
+
+  it('encodes multiline non-SSE stream fallback responses as valid SSE', async () => {
+    const payload = 'line one\nline two\n';
+    fetchMock.mockResolvedValue(new Response(payload, {
+      status: 200,
+      headers: {
+        'content-type': 'text/plain; charset=utf-8',
+      },
+    }));
+
+    const response = await app.inject({
+      method: 'POST',
+      url: '/api/test/proxy/stream',
+      payload: {
+        method: 'POST',
+        path: '/v1/chat/completions',
+        requestKind: 'json',
+        stream: true,
+        jobMode: false,
+        rawMode: false,
+        jsonBody: {
+          model: 'gpt-4o-mini',
+          messages: [{ role: 'user', content: 'ping' }],
+        },
+      },
+    });
+
+    expect(response.statusCode).toBe(200);
+    expect(response.headers['content-type']).toContain('text/event-stream');
+    expect(response.payload).toBe('data: line one\ndata: line two\ndata: \n\ndata: [DONE]\n\n');
+  });
+
+  it('cancels and releases SSE readers when stream forwarding fails', async () => {
+    const cancel = vi.fn().mockResolvedValue(undefined);
+    const releaseLock = vi.fn();
+    const read = vi.fn().mockRejectedValue(new Error('boom'));
+    fetchMock.mockResolvedValue({
+      ok: true,
+      status: 200,
+      headers: new Headers({
+        'content-type': 'text/event-stream; charset=utf-8',
+      }),
+      body: {
+        getReader: () => ({
+          read,
+          cancel,
+          releaseLock,
+        }),
+      },
+    });
+
+    const response = await app.inject({
+      method: 'POST',
+      url: '/api/test/proxy/stream',
+      payload: {
+        method: 'POST',
+        path: '/v1/chat/completions',
+        requestKind: 'json',
+        stream: true,
+        jobMode: false,
+        rawMode: false,
+        jsonBody: {
+          model: 'gpt-4o-mini',
+          messages: [{ role: 'user', content: 'ping' }],
+        },
+      },
+    });
+
+    expect(response.statusCode).toBe(200);
+    expect(response.headers['content-type']).toContain('text/event-stream');
+    expect(response.payload).toContain('event: error');
+    expect(response.payload).toContain('"message":"boom"');
+    expect(cancel).toHaveBeenCalledTimes(1);
+    expect(releaseLock).toHaveBeenCalledTimes(1);
+  });
 });

src/server/routes/api/test.ts

@@ -2,6 +2,7 @@ import { FastifyInstance, FastifyReply, FastifyRequest } from 'fastify';
 import { randomUUID } from 'node:crypto';
 import { fetch, File as UndiciFile, FormData as UndiciFormData } from 'undici';
 import { config } from '../../config.js';
+import { readRuntimeResponseText } from '../../proxy-core/executors/types.js';
 
 type UndiciRequestInit = Parameters<typeof fetch>[1];
 
@@ -521,7 +522,7 @@ async function fetchProxyBuffered(
   });
 
   const contentType = upstream.headers.get('content-type') || '';
-  const text = await upstream.text();
+  const text = await readRuntimeResponseText(upstream);
 
   if (!upstream.ok) {
     throw new UpstreamProxyError(upstream.status, normalizeErrorPayload(text));
@@ -656,14 +657,14 @@ async function sendStreamingEnvelope(
   }
 
   if (!upstream.ok) {
-    const text = await upstream.text();
+    const text = await readRuntimeResponseText(upstream);
     cleanupClientListeners();
     return reply.code(upstream.status).send(normalizeErrorPayload(text));
   }
 
   const contentType = upstream.headers.get('content-type') || '';
-  const reader = upstream.body?.getReader();
-  if (!reader) {
+  const reader = contentType.includes('text/event-stream') ? upstream.body?.getReader() : null;
+  if (contentType.includes('text/event-stream') && !reader) {
     cleanupClientListeners();
     return reply.code(502).send({
       error: {
@@ -682,25 +683,30 @@ async function sendStreamingEnvelope(
 
   try {
     if (contentType.includes('text/event-stream')) {
-      while (true) {
-        const { done, value } = await reader.read();
-        if (done) break;
-        if (value) {
-          reply.raw.write(Buffer.from(value));
+      const streamReader = reader!;
+      try {
+        while (true) {
+          const { done, value } = await streamReader.read();
+          if (done) break;
+          if (value) {
+            reply.raw.write(Buffer.from(value));
+          }
+        }
+      } finally {
+        try {
+          await streamReader.cancel();
+        } catch {
+          // no-op
+        } finally {
+          streamReader.releaseLock();
         }
       }
     } else {
-      let text = '';
-      const decoder = new TextDecoder('utf-8');
-      while (true) {
-        const { done, value } = await reader.read();
-        if (done) break;
-        if (value) {
-          text += decoder.decode(value, { stream: true });
-        }
+      const text = await readRuntimeResponseText(upstream);
+      for (const line of text.split(/\r?\n/)) {
+        reply.raw.write(`data: ${line}\n`);
       }
-      text += decoder.decode();
-      reply.raw.write(`data: ${text}\n\n`);
+      reply.raw.write('\n');
       reply.raw.write('data: [DONE]\n\n');
     }
   } catch (error) {
@@ -711,11 +717,6 @@ async function sendStreamingEnvelope(
       reply.raw.write(`event: error\ndata: ${message}\n\n`);
     }
   } finally {
-    try {
-      await reader.cancel();
-    } catch {
-      // no-op
-    }
     cleanupClientListeners();
     if (!reply.raw.writableEnded) {
       reply.raw.end();

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

@@ -177,7 +177,7 @@ describe('chat proxy codex oauth compatibility', () => {
     const forwardedBody = JSON.parse(options.body);
     expect(forwardedBody.stream).toBe(true);
     expect(forwardedBody.instructions).toBe('');
-    expect(forwardedBody.store).toBeUndefined();
+    expect(forwardedBody.store).toBe(false);
     expect(forwardedBody.parallel_tool_calls).toBeUndefined();
     expect(forwardedBody.include).toBeUndefined();
     expect(forwardedBody.max_output_tokens).toBe(256);

src/server/routes/proxy/chat.stream.test.ts

@@ -1,3 +1,4 @@
+import { zstdCompressSync } from 'node:zlib';
 import Fastify, { type FastifyInstance } from 'fastify';
 import { afterAll, beforeAll, beforeEach, describe, expect, it, vi } from 'vitest';
 import { config } from '../../config.js';
@@ -169,6 +170,79 @@ describe('chat proxy stream behavior', () => {
     expect(response.body).toContain('data: [DONE]');
   });
 
+  it('decodes zstd-compressed non-stream chat responses before serializing downstream JSON', async () => {
+    const payload = JSON.stringify({
+      id: 'chatcmpl-zstd',
+      object: 'chat.completion',
+      created: 1_706_000_000,
+      model: 'upstream-gpt',
+      choices: [{
+        index: 0,
+        message: { role: 'assistant', content: '你好，来自 zstd 非流式响应' },
+        finish_reason: 'stop',
+      }],
+      usage: { prompt_tokens: 11, completion_tokens: 7, total_tokens: 18 },
+    });
+    fetchMock.mockResolvedValue(new Response(zstdCompressSync(Buffer.from(payload)), {
+      status: 200,
+      headers: {
+        'content-encoding': 'zstd',
+        'content-type': 'application/json; charset=utf-8',
+      },
+    }));
+
+    const response = await app.inject({
+      method: 'POST',
+      url: '/v1/chat/completions',
+      payload: {
+        model: 'gpt-4o-mini',
+        messages: [{ role: 'user', content: 'hi' }],
+      },
+    });
+
+    expect(response.statusCode).toBe(200);
+    expect(response.json()?.choices?.[0]?.message?.content).toBe('你好，来自 zstd 非流式响应');
+  });
+
+  it('decodes zstd-compressed non-SSE streaming chat responses before SSE conversion', async () => {
+    const payload = JSON.stringify({
+      id: 'chatcmpl-zstd-stream',
+      object: 'chat.completion',
+      created: 1_706_000_000,
+      model: 'upstream-gpt',
+      choices: [{
+        index: 0,
+        message: { role: 'assistant', content: '你好，来自 zstd 流式回退' },
+        finish_reason: 'stop',
+      }],
+      usage: { prompt_tokens: 11, completion_tokens: 7, total_tokens: 18 },
+    });
+    fetchMock.mockResolvedValue(new Response(zstdCompressSync(Buffer.from(payload)), {
+      status: 200,
+      headers: {
+        'content-encoding': 'zstd',
+        'content-type': 'application/json; charset=utf-8',
+      },
+    }));
+
+    const response = await app.inject({
+      method: 'POST',
+      url: '/v1/chat/completions',
+      payload: {
+        model: 'gpt-4o-mini',
+        stream: true,
+        messages: [{ role: 'user', content: 'hi' }],
+      },
+    });
+
+    expect(response.statusCode).toBe(200);
+    expect(response.headers['content-type']).toContain('text/event-stream');
+    expect(response.body).toContain('"chat.completion.chunk"');
+    expect(response.body).toContain('你好，来自 zstd 流式回退');
+    expect(response.body).not.toContain('(�/�');
+    expect(response.body).toContain('data: [DONE]');
+  });
+
   it('returns upstream_error for empty non-stream chat responses when empty-content failure is enabled', async () => {
     config.proxyEmptyContentFailEnabled = true;
 

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

@@ -1,6 +1,7 @@
 import Fastify, { type FastifyInstance } from 'fastify';
 import { afterAll, beforeAll, beforeEach, describe, expect, it, vi } from 'vitest';
 import { config } from '../../config.js';
+import { resetCodexHttpSessionQueue } from '../../proxy-core/runtime/codexHttpSessionQueue.js';
 
 const fetchMock = vi.fn();
 const selectChannelMock = vi.fn();
@@ -91,6 +92,26 @@ vi.mock('../../db/index.js', () => ({
 describe('responses proxy codex oauth refresh', () => {
   let app: FastifyInstance;
 
+  const 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 };
+  };
+
+  const waitFor = async (predicate: () => boolean, timeoutMs = 1000) => {
+    const startedAt = Date.now();
+    while (!predicate()) {
+      if ((Date.now() - startedAt) >= timeoutMs) {
+        throw new Error('Timed out waiting for test condition');
+      }
+      await new Promise((resolve) => setTimeout(resolve, 0));
+    }
+  };
+
   const createSseResponse = (chunks: string[], status = 200) => {
     const encoder = new TextEncoder();
     return new Response(new ReadableStream<Uint8Array>({
@@ -113,6 +134,7 @@ describe('responses proxy codex oauth refresh', () => {
   });
 
   beforeEach(() => {
+    resetCodexHttpSessionQueue();
     config.proxyEmptyContentFailEnabled = false;
     fetchMock.mockReset();
     selectChannelMock.mockReset();
@@ -461,6 +483,84 @@ describe('responses proxy codex oauth refresh', () => {
     expect(options.headers.referer).toBeUndefined();
   });
 
+  it('serializes concurrent codex HTTP responses requests that share the same session id', async () => {
+    const firstUpstream = createDeferred<Response>();
+    fetchMock
+      .mockImplementationOnce(() => firstUpstream.promise)
+      .mockResolvedValueOnce(new Response(JSON.stringify({
+        id: 'resp_codex_serial_2',
+        object: 'response',
+        model: 'gpt-5.2-codex',
+        status: 'completed',
+        output_text: 'second request finished',
+        usage: { input_tokens: 4, output_tokens: 2, total_tokens: 6 },
+      }), {
+        status: 200,
+        headers: { 'content-type': 'application/json' },
+      }));
+
+    const firstResponsePromise = app.inject({
+      method: 'POST',
+      url: '/v1/responses',
+      headers: {
+        session_id: 'session-http-serial-1',
+      },
+      payload: {
+        model: 'gpt-5.2-codex',
+        input: 'first request',
+      },
+    });
+
+    await waitFor(() => fetchMock.mock.calls.length === 1);
+
+    const secondResponsePromise = app.inject({
+      method: 'POST',
+      url: '/v1/responses',
+      headers: {
+        session_id: 'session-http-serial-1',
+      },
+      payload: {
+        model: 'gpt-5.2-codex',
+        input: 'second request',
+      },
+    });
+
+    await new Promise((resolve) => setTimeout(resolve, 0));
+    await new Promise((resolve) => setTimeout(resolve, 0));
+    expect(fetchMock).toHaveBeenCalledTimes(1);
+
+    firstUpstream.resolve(new Response(JSON.stringify({
+      id: 'resp_codex_serial_1',
+      object: 'response',
+      model: 'gpt-5.2-codex',
+      status: 'completed',
+      output_text: 'first request finished',
+      usage: { input_tokens: 4, output_tokens: 2, total_tokens: 6 },
+    }), {
+      status: 200,
+      headers: { 'content-type': 'application/json' },
+    }));
+
+    const firstResponse = await firstResponsePromise;
+    expect(firstResponse.statusCode).toBe(200);
+    expect(firstResponse.json()).toMatchObject({
+      output_text: 'first request finished',
+    });
+
+    await waitFor(() => fetchMock.mock.calls.length === 2);
+
+    const secondResponse = await secondResponsePromise;
+    expect(secondResponse.statusCode).toBe(200);
+    expect(secondResponse.json()).toMatchObject({
+      output_text: 'second request finished',
+    });
+
+    const [, firstOptions] = fetchMock.mock.calls[0] as [string, any];
+    const [, secondOptions] = fetchMock.mock.calls[1] as [string, any];
+    expect(firstOptions.headers.Session_id || firstOptions.headers.session_id).toBe('session-http-serial-1');
+    expect(secondOptions.headers.Session_id || secondOptions.headers.session_id).toBe('session-http-serial-1');
+  });
+
   it('records codex usage_limit_reached reset hints on upstream 429 failures', async () => {
     fetchMock.mockResolvedValue(new Response(JSON.stringify({
       error: {
@@ -525,7 +625,7 @@ describe('responses proxy codex oauth refresh', () => {
     const forwardedBody = JSON.parse(options.body);
     expect(forwardedBody.stream).toBe(true);
     expect(forwardedBody.instructions).toBe('');
-    expect(forwardedBody.store).toBeUndefined();
+    expect(forwardedBody.store).toBe(false);
     expect(forwardedBody.max_output_tokens).toBeUndefined();
 
     expect(response.json()).toMatchObject({
@@ -568,11 +668,11 @@ describe('responses proxy codex oauth refresh', () => {
     const secondBody = JSON.parse(secondOptions.body);
 
     expect(firstBody.instructions).toBe('');
-    expect(firstBody.store).toBeUndefined();
+    expect(firstBody.store).toBe(false);
     expect(firstBody.stream).toBe(true);
     expect(firstBody.max_output_tokens).toBeUndefined();
     expect(secondBody.instructions).toBe('');
-    expect(secondBody.store).toBeUndefined();
+    expect(secondBody.store).toBe(false);
     expect(secondBody.stream).toBe(true);
     expect(secondBody.max_output_tokens).toBeUndefined();
   });

src/server/routes/proxy/upstreamEndpoint.test.ts

@@ -628,7 +628,7 @@ describe('buildUpstreamEndpointRequest', () => {
     expect(request.body.instructions).toBe('');
     expect(request.body.prompt_cache_key).toBeUndefined();
     expect(request.body.stream).toBe(false);
-    expect(request.body.store).toBeUndefined();
+    expect(request.body.store).toBe(false);
     expect(request.body.parallel_tool_calls).toBeUndefined();
     expect(request.body.include).toBeUndefined();
     expect(request.body.max_output_tokens).toBe(4096);
@@ -762,7 +762,7 @@ describe('buildUpstreamEndpointRequest', () => {
     expect(request.body.prompt_cache_key).toBeUndefined();
     expect(request.body.instructions).toBe('');
     expect(request.body.stream).toBe(false);
-    expect(request.body.store).toBe(true);
+    expect(request.body.store).toBe(false);
     expect(request.body.parallel_tool_calls).toBe(false);
     expect(request.body.include).toEqual(['reasoning.encrypted_content', 'mcp_approval_request.details']);
     expect(request.body.previous_response_id).toBe('resp_prev_123');
@@ -848,7 +848,7 @@ describe('buildUpstreamEndpointRequest', () => {
     expect(httpRequest.headers['x-codex-beta-features']).toBeUndefined();
   });
 
-  it('applies configured payload rules before preparing codex responses requests', () => {
+  it('applies configured payload rules before preparing codex responses requests while forcing store false', () => {
     (config as any).payloadRules = {
       default: [
         {
@@ -864,6 +864,7 @@ describe('buildUpstreamEndpointRequest', () => {
           models: [{ name: 'gpt-5.4', protocol: 'codex' }],
           params: {
             'text.verbosity': 'low',
+            store: true,
           },
         },
       ],
@@ -899,6 +900,7 @@ describe('buildUpstreamEndpointRequest', () => {
     expect(request.body.reasoning).toEqual({ effort: 'high' });
     expect(request.body.text).toEqual({ verbosity: 'low' });
     expect(request.body.safety_identifier).toBeUndefined();
+    expect(request.body.store).toBe(false);
   });
 
   it('builds gemini-cli native requests with project envelope and bearer headers', () => {

src/server/routes/proxy/upstreamEndpoint.ts

@@ -413,7 +413,11 @@ function ensureCodexResponsesStoreFalse(
   body: Record<string, unknown>,
   sitePlatform: string,
 ): Record<string, unknown> {
-  return body;
+  if (sitePlatform !== 'codex') return body;
+  return {
+    ...body,
+    store: false,
+  };
 }
 
 function convertCodexSystemRoleToDeveloper(input: unknown): unknown {
@@ -1256,7 +1260,10 @@ export function buildUpstreamEndpointRequest(input: {
       ),
       sitePlatform,
     );
-    const configuredResponsesBody = applyConfiguredPayloadRules(body);
+    const configuredResponsesBody = ensureCodexResponsesStoreFalse(
+      applyConfiguredPayloadRules(body),
+      sitePlatform,
+    );
 
     if (sitePlatform === 'codex') {
       if (providerProfile?.id !== 'codex') {

src/server/transformers/openai/responses/conversion.test.ts

@@ -437,6 +437,67 @@ describe('sanitizeResponsesBodyForProxy', () => {
       },
     ]);
   });
+
+  it('drops blank tool calls before proxying Responses requests while preserving follow-up tool outputs', () => {
+    const result = sanitizeResponsesBodyForProxy(
+      {
+        model: 'gpt-5',
+        input: [
+          {
+            type: 'function_call',
+            call_id: 'call_blank',
+            name: '   ',
+            arguments: '{"city":"Shanghai"}',
+          },
+          {
+            type: 'function_call_output',
+            call_id: 'call_blank',
+            output: '{"temp":22}',
+          },
+          {
+            type: 'message',
+            role: 'assistant',
+            content: [
+              {
+                type: 'output_text',
+                text: 'continue normally',
+              },
+            ],
+          },
+          {
+            type: 'function_call_output',
+            call_id: 'call_missing',
+            output: 'orphaned output',
+          },
+        ],
+      },
+      'gpt-5',
+      false,
+    );
+
+    expect(result.input).toEqual([
+      {
+        type: 'function_call_output',
+        call_id: 'call_blank',
+        output: '{"temp":22}',
+      },
+      {
+        type: 'message',
+        role: 'assistant',
+        content: [
+          {
+            type: 'output_text',
+            text: 'continue normally',
+          },
+        ],
+      },
+      {
+        type: 'function_call_output',
+        call_id: 'call_missing',
+        output: 'orphaned output',
+      },
+    ]);
+  });
 });
 
 describe('convertOpenAiBodyToResponsesBody', () => {
@@ -763,6 +824,45 @@ describe('convertOpenAiBodyToResponsesBody', () => {
     ]);
   });
 
+  it('drops tool outputs that do not have a matching prior tool call when converting to OpenAI-compatible messages', () => {
+    const result = convertResponsesBodyToOpenAiBody(
+      {
+        model: 'gpt-5',
+        input: [
+          {
+            type: 'function_call_output',
+            call_id: 'call_missing',
+            output: 'orphaned tool result',
+          },
+          {
+            type: 'message',
+            role: 'assistant',
+            content: [
+              {
+                type: 'output_text',
+                text: 'continue normally',
+              },
+            ],
+          },
+        ],
+      },
+      'gpt-5',
+      false,
+    );
+
+    expect(result.messages).toEqual([
+      {
+        role: 'assistant',
+        content: [
+          {
+            type: 'text',
+            text: 'continue normally',
+          },
+        ],
+      },
+    ]);
+  });
+
   it('shortens long MCP tool names consistently across tools, tool_choice and assistant tool calls', () => {
     const sharedSuffix = 'server__execute_super_long_nested_tool_name_that_needs_shortening';
     const firstName = `mcp__alpha_workspace__${sharedSuffix}`;
@@ -840,6 +940,63 @@ describe('convertOpenAiBodyToResponsesBody', () => {
       name: toolNames[0],
     });
   });
+
+  it('drops function tools with blank names when converting OpenAI-compatible input into Responses bodies', () => {
+    const result = convertOpenAiBodyToResponsesBody(
+      {
+        model: 'gpt-5',
+        messages: [{ role: 'user', content: 'hello' }],
+        tools: [
+          {
+            type: 'function',
+            function: {
+              name: '   ',
+              parameters: { type: 'object' },
+            },
+          },
+          {
+            type: 'function',
+            function: {
+              name: 'lookup_weather',
+              parameters: { type: 'object' },
+            },
+          },
+        ],
+      },
+      'gpt-5',
+      false,
+    );
+
+    expect(result.tools).toEqual([
+      {
+        type: 'function',
+        name: 'lookup_weather',
+        parameters: { type: 'object' },
+      },
+    ]);
+  });
+
+  it('drops all-invalid function tools instead of forwarding blank names into Responses bodies', () => {
+    const result = convertOpenAiBodyToResponsesBody(
+      {
+        model: 'gpt-5',
+        messages: [{ role: 'user', content: 'hello' }],
+        tools: [
+          {
+            type: 'function',
+            function: {
+              name: '   ',
+              parameters: { type: 'object' },
+            },
+          },
+        ],
+      },
+      'gpt-5',
+      false,
+    );
+
+    expect(result.tools).toEqual([]);
+  });
 });
 
 describe('convertResponsesBodyToOpenAiBody', () => {
@@ -881,6 +1038,7 @@ describe('convertResponsesBodyToOpenAiBody', () => {
             file: {
               filename: 'paper.pdf',
               file_data: Buffer.from('%PDF-demo').toString('base64'),
+              mime_type: 'application/pdf',
             },
           },
         ],
@@ -1461,6 +1619,31 @@ describe('convertResponsesBodyToOpenAiBody', () => {
     });
   });
 
+  it('drops blank native function tools and tool choices when converting to OpenAI-compatible fallback bodies', () => {
+    const result = convertResponsesBodyToOpenAiBody(
+      {
+        model: 'gpt-5',
+        input: 'hello',
+        tools: [{
+          type: 'function',
+          name: '   ',
+          parameters: {
+            type: 'object',
+          },
+        }],
+        tool_choice: {
+          type: 'tool',
+          name: '   ',
+        },
+      },
+      'gpt-5',
+      false,
+    );
+
+    expect(result.tools).toEqual([]);
+    expect(result.tool_choice).toBeUndefined();
+  });
+
   it('derives Responses reasoning config from chat-style reasoning request fields', () => {
     const result = convertOpenAiBodyToResponsesBody(
       {

src/server/transformers/openai/responses/conversion.ts

@@ -352,7 +352,7 @@ function convertOpenAiToolsToResponses(
     })
     .filter((item): item is Record<string, unknown> => !!item);
 
-  return converted.length > 0 ? converted : rawTools;
+  return converted;
 }
 
 function convertOpenAiToolChoiceToResponses(
@@ -384,6 +384,49 @@ function normalizeResponsesBodyForCompatibility(
   };
 }
 
+const RESPONSES_TOOL_CALL_INPUT_TYPES = new Set([
+  'function_call',
+  'custom_tool_call',
+]);
+
+const RESPONSES_TOOL_CALL_OUTPUT_TYPES = new Set([
+  'function_call_output',
+  'custom_tool_call_output',
+]);
+
+function stripOrphanedResponsesToolOutputs(input: unknown): unknown {
+  if (!Array.isArray(input)) return input;
+
+  const seenToolCallIds = new Set<string>();
+  const sanitized: unknown[] = [];
+
+  for (const item of input) {
+    if (!isRecord(item)) {
+      sanitized.push(item);
+      continue;
+    }
+
+    const type = asTrimmedString(item.type).toLowerCase();
+    if (RESPONSES_TOOL_CALL_INPUT_TYPES.has(type)) {
+      const callId = asTrimmedString(item.call_id ?? item.id);
+      if (callId) seenToolCallIds.add(callId);
+      sanitized.push(item);
+      continue;
+    }
+
+    if (RESPONSES_TOOL_CALL_OUTPUT_TYPES.has(type)) {
+      const callId = asTrimmedString(item.call_id ?? item.id);
+      if (!callId || !seenToolCallIds.has(callId)) continue;
+      sanitized.push(item);
+      continue;
+    }
+
+    sanitized.push(item);
+  }
+
+  return sanitized;
+}
+
 export function normalizeResponsesMessageContent(role: string, content: unknown): Array<Record<string, unknown>> {
   return normalizeResponsesMessageContentBlocks(role, content);
 }
@@ -614,6 +657,9 @@ export function convertOpenAiBodyToResponsesBody(
 
   const responsesToolChoice = convertOpenAiToolChoiceToResponses(openaiBody.tool_choice, toolNameMap);
   if (responsesToolChoice !== undefined) body.tool_choice = responsesToolChoice;
+  if (Array.isArray(body.tools) && body.tools.length === 0) {
+    delete body.tool_choice;
+  }
 
   return normalizeResponsesBodyForCompatibility(
     normalizeResponsesRequestFieldParity(body, { verbositySource: openaiBody.verbosity }),
@@ -723,28 +769,30 @@ function toOpenAiMessageContent(content: unknown): string | Array<string | Recor
 function convertResponsesToolsToOpenAi(rawTools: unknown): unknown {
   if (!Array.isArray(rawTools)) return rawTools;
 
-  return rawTools.map((item) => {
-    if (!isRecord(item)) return item;
-    const type = asTrimmedString(item.type).toLowerCase();
+  return rawTools
+    .map((item) => {
+      if (!isRecord(item)) return item;
+      const type = asTrimmedString(item.type).toLowerCase();
 
-    if (type === 'custom' || type === 'image_generation') return item;
-    if (type !== 'function') return item;
-    if (isRecord(item.function) && asTrimmedString(item.function.name)) return item;
+      if (type === 'custom' || type === 'image_generation') return item;
+      if (type !== 'function') return item;
+      if (isRecord(item.function) && asTrimmedString(item.function.name)) return item;
 
-    const name = asTrimmedString(item.name);
-    if (!name) return item;
+      const name = asTrimmedString(item.name);
+      if (!name) return null;
 
-    const fn: Record<string, unknown> = { name };
-    const description = asTrimmedString(item.description);
-    if (description) fn.description = description;
-    if (item.parameters !== undefined) fn.parameters = item.parameters;
-    if (item.strict !== undefined) fn.strict = item.strict;
+      const fn: Record<string, unknown> = { name };
+      const description = asTrimmedString(item.description);
+      if (description) fn.description = description;
+      if (item.parameters !== undefined) fn.parameters = item.parameters;
+      if (item.strict !== undefined) fn.strict = item.strict;
 
-    return {
-      type: 'function',
-      function: fn,
-    };
-  });
+      return {
+        type: 'function',
+        function: fn,
+      };
+    })
+    .filter((item): item is Record<string, unknown> => !!item);
 }
 
 function convertResponsesToolChoiceToOpenAi(rawToolChoice: unknown): unknown {
@@ -787,11 +835,13 @@ export function convertResponsesBodyToOpenAiBody(
   stream: boolean,
   options?: { defaultEncryptedReasoningInclude?: boolean },
 ): Record<string, unknown> {
-  const normalizedBody = normalizeResponsesRequestFieldParity(body, {
-    defaultEncryptedReasoningInclude: options?.defaultEncryptedReasoningInclude,
-  });
+  const normalizedBody = normalizeResponsesBodyForCompatibility(
+    normalizeResponsesRequestFieldParity(body, {
+      defaultEncryptedReasoningInclude: options?.defaultEncryptedReasoningInclude,
+    }),
+  );
   const messages: Array<Record<string, unknown>> = [];
-  const input = normalizedBody.input;
+  const input = stripOrphanedResponsesToolOutputs(normalizedBody.input);
   let functionCallIndex = 0;
   let pendingToolCalls: OpenAiToolCall[] = [];
 
@@ -928,6 +978,9 @@ export function convertResponsesBodyToOpenAiBody(
   if (normalizedBody.parallel_tool_calls !== undefined) payload.parallel_tool_calls = normalizedBody.parallel_tool_calls;
   if (normalizedBody.tools !== undefined) payload.tools = convertResponsesToolsToOpenAi(normalizedBody.tools);
   if (normalizedBody.tool_choice !== undefined) payload.tool_choice = convertResponsesToolChoiceToOpenAi(normalizedBody.tool_choice);
+  if (Array.isArray(payload.tools) && payload.tools.length === 0) {
+    delete payload.tool_choice;
+  }
   if (normalizedBody.safety_identifier !== undefined) payload.safety_identifier = normalizedBody.safety_identifier;
   if (normalizedBody.max_tool_calls !== undefined) payload.max_tool_calls = normalizedBody.max_tool_calls;
   if (normalizedBody.prompt_cache_key !== undefined) payload.prompt_cache_key = normalizedBody.prompt_cache_key;

src/server/transformers/openai/responses/normalization.ts

@@ -8,6 +8,16 @@ function asTrimmedString(value: unknown): string {
   return typeof value === 'string' ? value.trim() : '';
 }
 
+const RESPONSES_TOOL_CALL_ITEM_TYPES = new Set([
+  'function_call',
+  'custom_tool_call',
+]);
+
+const RESPONSES_TOOL_OUTPUT_ITEM_TYPES = new Set([
+  'function_call_output',
+  'custom_tool_call_output',
+]);
+
 const ALLOWED_RESPONSES_INPUT_STATUSES = new Set([
   'in_progress',
   'completed',
@@ -135,7 +145,7 @@ function normalizeResponsesContentItem(
     return fileBlock ? toResponsesInputFileBlock(fileBlock) : null;
   }
 
-  if (type === 'function_call' || type === 'function_call_output') {
+  if (RESPONSES_TOOL_CALL_ITEM_TYPES.has(type) || RESPONSES_TOOL_OUTPUT_ITEM_TYPES.has(type)) {
     return item;
   }
 
@@ -166,10 +176,71 @@ function toResponsesInputMessageFromText(text: string): Record<string, unknown>
   };
 }
 
+function normalizeResponsesToolLifecycleItem(item: Record<string, unknown>): Record<string, unknown> | null {
+  const type = asTrimmedString(item.type).toLowerCase();
+  if (!RESPONSES_TOOL_CALL_ITEM_TYPES.has(type) && !RESPONSES_TOOL_OUTPUT_ITEM_TYPES.has(type)) {
+    return withNormalizedResponsesInputStatus(item);
+  }
+
+  const callId = asTrimmedString(item.call_id ?? item.id);
+  if (!callId) return null;
+
+  const normalized: Record<string, unknown> = {
+    ...item,
+    call_id: callId,
+  };
+
+  if (typeof item.id === 'string') {
+    const id = asTrimmedString(item.id);
+    if (id) normalized.id = id;
+    else delete normalized.id;
+  }
+
+  if (RESPONSES_TOOL_CALL_ITEM_TYPES.has(type)) {
+    if (Object.prototype.hasOwnProperty.call(item, 'name')) {
+      const name = asTrimmedString(item.name);
+      if (!name) return null;
+      normalized.name = name;
+    }
+  }
+
+  return withNormalizedResponsesInputStatus(normalized);
+}
+
+function sanitizeResponsesInputToolLifecycle(items: unknown[]): unknown[] {
+  const sanitized: unknown[] = [];
+
+  for (const item of items) {
+    if (!isRecord(item)) {
+      sanitized.push(item);
+      continue;
+    }
+
+    const type = asTrimmedString(item.type).toLowerCase();
+    if (RESPONSES_TOOL_CALL_ITEM_TYPES.has(type)) {
+      const normalized = normalizeResponsesToolLifecycleItem(item);
+      if (!normalized) continue;
+      sanitized.push(normalized);
+      continue;
+    }
+
+    if (RESPONSES_TOOL_OUTPUT_ITEM_TYPES.has(type)) {
+      const normalized = normalizeResponsesToolLifecycleItem(item);
+      if (!normalized) continue;
+      sanitized.push(normalized);
+      continue;
+    }
+
+    sanitized.push(item);
+  }
+
+  return sanitized;
+}
+
 export function normalizeResponsesMessageItem(item: Record<string, unknown>): Record<string, unknown> {
   const type = asTrimmedString(item.type).toLowerCase();
-  if (type === 'function_call' || type === 'function_call_output') {
-    return withNormalizedResponsesInputStatus(item);
+  if (RESPONSES_TOOL_CALL_ITEM_TYPES.has(type) || RESPONSES_TOOL_OUTPUT_ITEM_TYPES.has(type)) {
+    return normalizeResponsesToolLifecycleItem(item) ?? item;
   }
 
   const role = asTrimmedString(item.role).toLowerCase() || 'user';
@@ -211,7 +282,7 @@ export function normalizeResponsesInputForCompatibility(input: unknown): unknown
   }
 
   if (Array.isArray(input)) {
-    return input.flatMap((item) => {
+    const normalized = input.flatMap((item) => {
       if (typeof item === 'string') {
         const normalized = item.trim();
         return normalized ? [toResponsesInputMessageFromText(normalized)] : [];
@@ -219,10 +290,11 @@ export function normalizeResponsesInputForCompatibility(input: unknown): unknown
       if (!isRecord(item)) return [item];
       return [normalizeResponsesMessageItem(item)];
     });
+    return sanitizeResponsesInputToolLifecycle(normalized);
   }
 
   if (isRecord(input)) {
-    return [normalizeResponsesMessageItem(input)];
+    return sanitizeResponsesInputToolLifecycle([normalizeResponsesMessageItem(input)]);
   }
 
   return input;