From d9ace607e58caafd71224e02ed9f297b7dcacecb Mon Sep 17 00:00:00 2001
From: Tam Nhu Tran <kaitran.ntt@gmail.com>
Date: Thu, 30 Apr 2026 12:33:02 -0400
Subject: [PATCH] fix(proxy): shape direct OpenAI reasoning chat payloads

---
 src/proxy/server/messages-route.ts            | 56 ++++++++++++++--
 src/proxy/transformers/request-transformer.ts |  1 +
 .../integration/proxy/request-routing.test.ts | 66 +++++++++++++++++++
 3 files changed, 116 insertions(+), 7 deletions(-)

diff --git a/src/proxy/server/messages-route.ts b/src/proxy/server/messages-route.ts
index 965e67a3..05cf485e 100644
--- a/src/proxy/server/messages-route.ts
+++ b/src/proxy/server/messages-route.ts
@@ -2,13 +2,17 @@ import * as http from 'http';
 import type { Dispatcher } from 'undici';
 import type { OpenAICompatProfileConfig } from '../profile-router';
 import { resolveProxyRequestRoute } from '../request-router';
-import { ProxyRequestTransformer } from '../transformers/request-transformer';
+import {
+  ProxyRequestTransformer,
+  type ProxyOpenAIRequest,
+} from '../transformers/request-transformer';
 import { ProxySseStreamTransformer } from '../transformers/sse-stream-transformer';
 import { resolveOpenAIChatCompletionsUrl } from '../upstream-url';
 import { createLogger } from '../../services/logging';
 import { pipeWebResponseToNode, readJsonBody, writeJson } from './http-helpers';
 
 const REQUEST_TIMEOUT_MS = 600_000;
+const DIRECT_OPENAI_REASONING_CHAT_MODEL = /^(?:gpt-5|o[134])(?:[-.]|$)/;
 const logger = createLogger('proxy:openai-compat:messages');
 
 class ProxyInputError extends Error {
@@ -26,11 +30,46 @@ function buildUpstreamHeaders(profile: OpenAICompatProfileConfig): Record<string
   };
 }
 
+function isDirectOpenAIReasoningChatModel(
+  profile: OpenAICompatProfileConfig,
+  model: string | undefined
+): boolean {
+  return (
+    profile.provider === 'openai' &&
+    typeof model === 'string' &&
+    DIRECT_OPENAI_REASONING_CHAT_MODEL.test(model.trim().toLowerCase())
+  );
+}
+
+function shapeUpstreamChatPayload(
+  payload: ProxyOpenAIRequest,
+  profile: OpenAICompatProfileConfig
+): ProxyOpenAIRequest {
+  if (!isDirectOpenAIReasoningChatModel(profile, payload.model)) {
+    return payload;
+  }
+
+  const shaped = { ...payload };
+
+  if (shaped.max_tokens !== undefined) {
+    shaped.max_completion_tokens = shaped.max_tokens;
+    delete shaped.max_tokens;
+  }
+
+  delete shaped.metadata;
+
+  if ((shaped.tools?.length ?? 0) > 0) {
+    delete shaped.reasoning_effort;
+  }
+
+  return shaped;
+}
+
 function buildUpstreamRequest(
   profile: OpenAICompatProfileConfig,
   rawBody: unknown
 ): { body: string; route: ReturnType<typeof resolveProxyRequestRoute> } {
-  let transformed;
+  let transformed: ProxyOpenAIRequest;
   try {
     const transformer = new ProxyRequestTransformer();
     transformed = transformer.transform(rawBody);
@@ -39,11 +78,14 @@ function buildUpstreamRequest(
     throw new ProxyInputError(message);
   }
   const route = resolveProxyRequestRoute(profile, transformed);
-  const body = {
-    ...transformed,
-    model: route.model || route.profile.model,
-    stream: transformed.stream === true,
-  };
+  const body = shapeUpstreamChatPayload(
+    {
+      ...transformed,
+      model: route.model || route.profile.model,
+      stream: transformed.stream === true,
+    },
+    route.profile
+  );
   return { body: JSON.stringify(body), route };
 }
 
diff --git a/src/proxy/transformers/request-transformer.ts b/src/proxy/transformers/request-transformer.ts
index 0942244e..f1f9346f 100644
--- a/src/proxy/transformers/request-transformer.ts
+++ b/src/proxy/transformers/request-transformer.ts
@@ -127,6 +127,7 @@ export interface ProxyOpenAIRequest {
   parallel_tool_calls?: boolean;
   messages: OpenAIMessage[];
   max_tokens?: number;
+  max_completion_tokens?: number;
   temperature?: number;
   top_p?: number;
   stop?: string[];
diff --git a/tests/integration/proxy/request-routing.test.ts b/tests/integration/proxy/request-routing.test.ts
index 7e44ae13..3fe350bf 100644
--- a/tests/integration/proxy/request-routing.test.ts
+++ b/tests/integration/proxy/request-routing.test.ts
@@ -353,4 +353,70 @@ describe('openai proxy request routing', () => {
     });
     expect((bodies[0]?.body as { reasoning?: unknown } | undefined)?.reasoning).toBeUndefined();
   });
+
+  it('shapes direct OpenAI reasoning-model chat payloads after route resolution', async () => {
+    const hits: string[] = [];
+    const bodies: Array<{ label: string; body: unknown }> = [];
+    const upstreamPort = await startMockUpstream('openai', hits, bodies);
+
+    const settingsPath = writeSettings('openai', {
+      ANTHROPIC_BASE_URL: 'https://api.openai.com/v1',
+      ANTHROPIC_AUTH_TOKEN: 'openai_token',
+      ANTHROPIC_MODEL: 'gpt-5.4',
+    });
+
+    fs.writeFileSync(
+      path.join(tempDir, '.ccs', 'config.json'),
+      JSON.stringify({ profiles: { openai: settingsPath } }, null, 2),
+      'utf8'
+    );
+
+    const profile: OpenAICompatProfileConfig = {
+      profileName: 'openai',
+      settingsPath,
+      baseUrl: `http://127.0.0.1:${upstreamPort}`,
+      apiKey: 'openai_token',
+      provider: 'openai',
+      model: 'gpt-5.4',
+    };
+    proxyServer = startOpenAICompatProxyServer({
+      profile,
+      port: 0,
+      authToken: 'test-proxy-token',
+    });
+    proxyPort = await waitForServerListening(proxyServer);
+
+    const response = await requestProxy({
+      model: 'gpt-5.4',
+      thinking: { type: 'adaptive' },
+      output_config: { effort: 'max' },
+      max_tokens: 1024,
+      metadata: { trace: 'abc' },
+      tools: [{ name: 'search', description: 'Search docs', input_schema: { type: 'object' } }],
+      messages: [{ role: 'user', content: 'think with tools' }],
+    });
+
+    expect(response.status).toBe(200);
+    expect(await response.json()).toMatchObject({
+      content: [{ type: 'text', text: 'Reply from openai' }],
+    });
+    expect(hits).toEqual(['openai']);
+
+    const body = bodies[0]?.body as {
+      max_tokens?: number;
+      max_completion_tokens?: number;
+      metadata?: unknown;
+      reasoning_effort?: string;
+      tools?: unknown[];
+    };
+    expect(body).toMatchObject({
+      model: 'gpt-5.4',
+      max_completion_tokens: 1024,
+      tool_choice: 'auto',
+    });
+    expect(body.max_tokens).toBeUndefined();
+    expect(body.metadata).toBeUndefined();
+    expect(body.reasoning_effort).toBeUndefined();
+    expect(body.tools?.length).toBe(1);
+  });
 });