Repository navigation
temperature=0 is always sent; current OpenAI models reject it (including the openai provider's default gpt-5.6-sol) #23
Description
Summary
LLMClient always sends temperature (agents/llm.py L769), and every agent builds it with temperature=0.0. The current OpenAI models reject that value, so with an API key every call fails with a 400, including on gpt-5.6-sol, the openai provider's own default_model (providers.py L45-L48).
Observed (2026-10-12)
The same streaming request OpenHack sends (temperature: 0.0, max_tokens, stream: true), for each model:
| model | response |
|---|---|
gpt-5.6-sol, gpt-5.6-luna |
400 Unsupported value: 'temperature' does not support 0 with this model. Only the default (1) value is supported. (param: temperature) |
gpt-6.1-sol, gpt-6-luna |
400 Unsupported parameter: 'temperature' is not supported with this model. (param: temperature) |
The same request without temperature streams normally, tool calls included.
To be precise about the setup: these calls went through our OpenAI-compatible gateway, which forwards the body to OpenAI unchanged and returned the errors above verbatim. I haven't repeated them against api.openai.com directly, but the error text and param are OpenAI's own. The ChatGPT-subscription (OAuth) path uses a different endpoint and may not be affected.
Suggested fix
There is already a pattern for this in agents/llm.py L732-L748: when the endpoint rejects prompt_cache_key, OpenHack sets _cache_key_unsupported and retries without it. The same retry for a 400 whose param is temperature (drop it and remember that per model) would cover these models and any other OpenAI-compatible endpoint with the same restriction, without keeping a list of model names.
AI-assisted: found and written up with Claude Code while testing OpenHack against an OpenAI-compatible gateway (see also #21, #22).
Activity
Lakshya-0910 commentedon Oct 14, 2026
Sent today directly to api.openai.com (no gateway): the request
LLMClient._chatbuilds forgpt-5.6-sol(main @ 4e1f532,openhack/agents/llm.py:765-781:max_tokens8192,temperature0, streamed, function tools).max_tokensfirst: "Unsupported parameter: 'max_tokens' is not supported with this model." Withouttemperature(PR Retry and omit temperature when rejected by endpoint (fixes #23) #25's retry) the answer is the same 400, and that retry never fires because the error namesmax_tokens. Same ongpt-6.1-sol.The agent loop sends tools every turn (
openhack/agents/base.py:454-458), so an OpenAI-key scan stops at its first model call, reconnaissance. I ran the requests only, not OpenHack.I'm Lakshya; I'm building a tool that checks model migrations on a team's real requests. If you'd like it run on yours, just reply here and I'll follow up privately.