[midozy]

Problem

Using GPT-5 (or any reasoning model) with the OpenAI or Custom provider on iOS fails with an unsupported max_tokens / max_completion_tokens error.

The chat reply goes through stream(), which always sends max_tokens + temperature and has no fallback. The non-streaming send() already retries with max_completion_tokens and no temperature, and the doc comment on stream() says it falls back to send's retry, but the code never did.

Fix

In OpenAIClient.stream and CustomClient.stream, catch a 400 whose body mentions max_completion_tokens, max_tokens, temperature or unsupported (same match as send). That error arrives before any delta, so fall back to send() (which retries with the modern params) and emit the reply as one delta. Other errors are rethrown unchanged.

Trade-off: those models' replies appear all at once instead of streaming. Switching the streaming request itself to the modern params on retry would be a larger change; happy to do that if preferred.

Second commit

Raises timeoutInterval to 300s on the non-streaming OpenAI/Custom chat requests so the fallback does not time out while a reasoning model is thinking.

Testing

• xcodebuild -scheme NOOPiOS builds clean.
• Installed on a device (iPhone 16 Pro Max) and tested end to end with a live GPT-5 call: it now replies with no error. Before the second commit it failed with a network error, because GPT-5 can take over 60s to send the first byte and hit the default URLSession idle timeout; the second commit raises the timeout to 300s on the non-streaming chat requests (OpenAI and Custom).

Not touched: Android. AiCoach.kt has its own request path; I haven't checked it for the same gap.