Log OpenAI OAuth error responses

[k2v7n24cbf-cyber]

Log OpenAI OAuth error responses

Summary

This PR is intentionally logging-only.

• Adds warning logs for final non-2xx responses from the OpenAI OAuth / ChatGPT Codex backend fetch path.
• Logs the upstream response body, status, status text, request URL, and selected diagnostic request-id headers.
• Reads the body from Response.clone(), so the original response is still returned unchanged to the AI SDK.
• Does not change model selection.
• Does not change generateObject() usage.
• Does not add a generateText() fallback.
• Does not change title generation, memory selection, parsing, validation, retries, or provider routing behavior.

Reason / RCA

In a live OpenAcme deployment using OpenAI ChatGPT OAuth, auxiliary helper calls such as title generation and memory selection were failing with only:

Bad Request

The existing logs showed the failing requests were going through the ChatGPT OAuth / Codex backend path, but they did not preserve the upstream OpenAI error response body. Because of that, we could not determine the actual rejected field or backend validation reason.

The objective of this PR is only to expose the missing evidence. It does not attempt to fix or work around the failure.

Implementation Notes

The OpenAI OAuth fetch wrapper in packages/llm-provider/src/registry.ts now calls a logging helper after the final response is available, including after the existing 401 refresh-and-retry path.

For any final non-2xx response, it logs:

• provider: "openai"
• auth: "oauth"
• request URL
• HTTP status and status text
• selected diagnostic headers:
  • cf-ray
  • openai-processing-ms
  • x-openai-request-id
  • x-request-id
  • x-stainless-request-id
• response body, truncated at 4000 characters

It does not log authorization headers or request bodies.

Validation

• pnpm --filter @openacme/config build
• pnpm --filter @openacme/auth build
• pnpm --filter @openacme/llm-provider check-types
• pnpm --filter @openacme/llm-provider test -- subagent-model.test.ts
• pnpm --filter @openacme/llm-provider build
• Push hook passed:
  • full pnpm build
  • server e2e: 13 files / 26 tests passed