[orshih6]

OpenAIProvider decides between max_tokens and max_completion_tokens by matching the model name (gpt-5+, o1/o3/o4). Names that don't match still get max_tokens. That covers fine-tunes (ft:gpt-5...), prefixed or proxy model ids, and any future family. Newer models reject that with a 400: "Unsupported parameter: 'max_tokens' is not supported with this model. Use 'max_completion_tokens' instead." The whole turn then fails.

This change keeps the name-based choice. When the Chat Completions request fails and the error detail says max_tokens is unsupported, the provider moves the value to max_completion_tokens and resends once. Any other error, or a second failure, is thrown as before. The request code is now inside a small loop, which is why most of the diff is re-indentation.

Tests: two new cases in tests/providers/OpenAIProvider.test.ts mock fetch. One returns the unsupported-parameter 400 for a fine-tuned model id. It checks that the request is resent with max_completion_tokens and without max_tokens. That test fails without the fix. The other checks that a different 400 is not retried. The OpenAIProvider test files pass, and tsc --noEmit and eslint on the touched files are clean.

Note: the model in the report (gpt-6.1-sol) already matches the name check on current main. This fallback covers the names the check still misses.

🤖 Generated with Claude Code

Fixes #717

---

This change was prepared with an AI coding agent (Claude) and reviewed by me before submitting. I'm happy to make any changes you'd like.