Sanitize Codex Responses message translation

[w32zhong]

Spotted this error when using Weixin with Codex OAuth and replayed chat history contains extra assistant message
fields such as reasoning_content, which the Codex Responses endpoint does not accept.

2026-05-22T16:23:31.353789809Z 2026-05-22T16:23:31.353591Z  INFO clawshell::app: Forwarding request to upstream method=GET path=/models virtual_key=vk-openai-codex                                                                           
2026-05-22T16:23:31.852544426Z 2026-05-22T16:23:31.852419Z  WARN clawshell::app: Upstream returned error oauth_provider=codex effective_path=/models status=400 Bad Request response_body={                                                   
2026-05-22T16:23:31.852576536Z   "error": {                                                                                                                                                                                                   
2026-05-22T16:23:31.852579459Z     "message": "[{'type': 'missing', 'loc': ('query', 'client_version'), 'msg': 'Field required', 'input': None}]",              
2026-05-22T16:23:31.852581951Z     "type": "invalid_request_error",                                                                                                                                                                           
2026-05-22T16:23:31.852583963Z     "param": null,                                                                                                                                                                                             
2026-05-22T16:23:31.852585910Z     "code": null                                                                                                                                                                                               
2026-05-22T16:23:31.852588046Z   }                                                                                                                                                                                                            
2026-05-22T16:23:31.852589789Z }                                                                                                                                                                                                              
2026-05-22T16:23:31.852592074Z 2026-05-22T16:23:31.852457Z  INFO clawshell::app: Request completed method=GET path=/models query= status=400 Bad Request latency_ms=498                                                                       
2026-05-22T16:23:31.900669382Z 2026-05-22T16:23:31.900559Z  INFO clawshell::app: Forwarding request to upstream method=POST path=/api/show virtual_key=vk-openai-codex                                                                        
2026-05-22T16:23:31.915812645Z 2026-05-22T16:23:31.915628Z  INFO clawshell::app: Request completed method=POST path=/api/show query= status=403 Forbidden latency_ms=15                                                                       
2026-05-22T16:23:32.722081729Z 2026-05-22T16:23:32.721815Z  INFO clawshell::app: Forwarding request to upstream method=POST path=/api/show virtual_key=vk-openai-codex                                                                        
2026-05-22T16:23:32.756354940Z 2026-05-22T16:23:32.756166Z  INFO clawshell::app: Request completed method=POST path=/api/show query= status=403 Forbidden latency_ms=34                                                                       
2026-05-22T16:23:32.777907955Z 2026-05-22T16:23:32.777738Z  INFO clawshell::app: Forwarding request to upstream method=POST path=/api/show virtual_key=vk-openai-codex                                                                        
2026-05-22T16:23:32.816114667Z 2026-05-22T16:23:32.815940Z  INFO clawshell::app: Request completed method=POST path=/api/show query= status=403 Forbidden latency_ms=38                                                                       
2026-05-22T16:23:33.026752047Z 2026-05-22T16:23:33.026625Z  INFO clawshell::app: Forwarding request to upstream method=POST path=/v1/chat/completions virtual_key=vk-openai-codex                                                             
2026-05-22T16:23:33.494637368Z 2026-05-22T16:23:33.494519Z  WARN clawshell::app: Upstream returned error oauth_provider=codex effective_path=/responses status=400 Bad Request response_body={                                                
2026-05-22T16:23:33.494667680Z   "error": {                                                                                                                                                                                                   
2026-05-22T16:23:33.494671450Z     "message": "Unknown parameter: 'input[2].reasoning_content'.",                                                                                                                                             
2026-05-22T16:23:33.494689458Z     "type": "invalid_request_error",                                                                                                                                                                           
2026-05-22T16:23:33.494692117Z     "param": "input[2].reasoning_content",                                                                                                                                                                     
2026-05-22T16:23:33.494694623Z     "code": "unknown_parameter"                                                                                                                                                                                
2026-05-22T16:23:33.494696947Z   }                                                                                                                                                                                                            
2026-05-22T16:23:33.494699117Z }

---

^{Need help on this PR? Tag @codesmith with what you need. Autofix is disabled.}

[pullfrog]

✅ No new issues found.

Reviewed changes — sanitize chat completions messages during Responses API translation by constructing a new message object with only explicitly allowed fields (role, content, phase, type), stripping unsupported fields like reasoning_content that Codex rejects.

• Build sanitized message from scratch — convert_message_to_input_items now constructs a new input_message map instead of cloning and mutating the original message, automatically dropping any fields not in the allowlist.
• Preserve phase for Codex flows — carries over the Codex-specific phase field from the source message.
• Add test for field stripping — verifies name and reasoning_content are absent while type, role, content, and phase survive translation.

  ^{｜ View workflow run ｜ Using Big Pickle (free) ｜ 𝕏}

[xichen1997]

LGTM, thanks!