codex-responses-tooling

Codex Responses Tooling

Use this skill when a task should go through the local Codex Responses API configuration instead of a hardcoded gateway. It is designed as a reusable pattern for streamed Responses calls, including image generation.

What To Read First

1. Read ~/.codex/config.toml.
2. Use [model_providers.OpenAI].base_url when present.
3. Read the local auth token from Codex auth storage.
4. Build the request against POST {base_url}/responses.

Request Shape

• Set "stream": true.
• Use the user prompt as the request input.
• Add the tool definition that matches the task.
• For image generation, the tool is:

{"type":"image_generation","model":"gpt-image-2"}

• Choose the outer model with normal judgment; gpt-5.4 is a good default when the task does not require a different model.

Streaming Rules

Treat the response as SSE.

1. Read data: events line by line.
2. Ignore non-JSON or [DONE] events.
3. Watch for response.output_item.done.
4. Inspect item.type and the tool-specific payload.
5. When item.type == "image_generation_call" and item.result exists, decode item.result from base64.
6. Write the decoded bytes to .png or .jpg.

Reuse Pattern

For other Responses-based tools, keep the same flow:

• read local config first
• reuse the configured base_url
• send a streamed Responses request
• decode the final tool result from the SSE stream

Only the tool definition and the result decoder should change.

Prompting

• Keep the task concrete: subject, composition, output format, and intended use.
• If text must appear in the output image, quote it verbatim.
• For posters or banners, specify aspect ratio and negative space.

Constraints

• Do not rely on the final JSON body for the image payload.
• Do not print or expose the auth token.

Output

Return the saved artifact path and a short note about the request used.