Preview limitations

Preview behavior and limitations

These limits apply to ChatGPT plan usage through Sign in with ChatGPT and https://api.openai.com/v1, both directly and through Codex app-server.

Responses API requirements

• HTTP requests: Set store: false and stream: true. Send input as an array containing the context needed for each request. Use instructions or developer messages; explicit {type: "message", role: "system"} items are rejected.
• Unsupported fields: Omit background, conversation, max_output_tokens, max_tool_calls, metadata, moderation, multi_agent, prompt, prompt_cache_retention, safety_identifier, temperature, top_logprobs, top_p, truncation, and user.
• Conversation state: Omit previous_response_id over HTTP and send the required history in input. WebSocket continuation can reference only responses from the same authenticated connection; it does not provide persistent conversation storage.
• Supported tools: Group function/custom tools in namespaces or supply them through additional_tools input items. Web search remains subject to model and account/workspace policy.
• Unsupported tools: Image generation, file search, Code Interpreter, native computer use, hosted MCP/connectors, and Responses tool_search. Client-side execution does not make tool_search supported. programmatic_tool_calling is not accepted in top-level tools on this route.
• Inputs: Text, images, and files are supported when the selected model accepts them. Audio/video input, the Files upload API, and the transcription API are not supported by this flow.

Codex app-server behavior

• Transport: The Codex app-server configuration uses stdio between your app and app-server, then HTTP/SSE to Responses. App-server sets store: false and stream: true; supports_websockets=false selects HTTP for this configuration.
• Local tools and agents: Codex can run shell and MCP tools and coordinate child agents through supported function/custom tool calls. These are separate from hosted Responses tools. Configurations that emit Responses tool_search will fail.
• Inputs and history: Use app-server’s RPC inputs; it does not accept arbitrary Responses parameters. Local thread history and thread/resume still work with store: false.
• Token renewal: Your app must refresh the OAuth token. With the env_key setup, restart app-server with the renewed ACCESS_TOKEN, then resume the saved thread.