AI request timeout too short / not configurable, and failures logged without an error message

[AhmetTumis]

AI API requests use a hardcoded 120s httpx timeout, which kills long running local/api model generations even while it still actively produces tokens.
Split the timeout (connect/write/pool short, read raised to 600s) as a quick fix, but it should be a config entry instead of hardcoded:

@staticmethod
async def do_request(url: str, headers: dict, payload: dict):
    async def _fetch():
        timeout = httpx.Timeout(connect=10, read=600, write=10, pool=10)
        async with httpx.AsyncClient(timeout=timeout) as client:
            return await client.post(url, headers=headers, json=payload)

Additionally, failures are logged with no useful message, logging.error(f"Request failed: {error}") relies on str(exception), which is empty for httpx timeout exceptions, so the terminal just shows Request failed: with nothing after it. Should log repr(error) (or the exception type name) instead so the actual failure reason is visible.

Sources: hanfor/ai_request/api_request_methods/standard_ai_api.py (timeout, line 18; error logging, line 77)

[LSchaffner]

Maybe we can just remove the timeout entirely, since the request can be cancelled through the threading interface.

[AhmetTumis]

I think removing the timeout completely might be problematic. Aside from httpx falling back to its own 5s default if we just omit the argument (so it’d need an explicit timeout=None), I’m not sure we actually want unbounded here. Cancellation covers the case where someone stops the request, but if the endpoint accepts the connection and then never answers, nothing would trip the stop flag and the request would just hang. Some LLM requests probably do need a real timeout as a backstop.