Refine temperature/top_p handling for reasoning models

[mayeco]

Update handling of temperature and top_p for OpenAI models.

This fix issue #2278

Checklist

• If the PR is changing/adding functionality, are there tests to reflect this?
• If there is an example, have you run the example to make sure that it works?
• If there are instructions on how to run the code, have you followed the instructions and made sure that it works?
• If the feature is significant enough to require documentation, is there a PR open on the OpenHands/docs repository with the same branch name?
• Is the github CI passing?

[Copilot]

Pull request overview

Updates chat-completion option normalization to avoid stripping sampling params (temperature, top_p) for non-OpenAI “reasoning effort” models (notably Gemini), while still handling OpenAI reasoning-model quirks.

Changes:

• Only remove temperature/top_p for OpenAI reasoning models (o1*, o3*) instead of all models that support reasoning_effort.
• Add model-name normalization logic in select_chat_options to gate this behavior.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

Pull request overview

Copilot reviewed 1 out of 1 changed files in this pull request and generated 2 comments.

Comments suppressed due to low confidence (1)

openhands-sdk/openhands/sdk/llm/options/chat_options.py:59

• The Gemini 2.5-pro substring check is case-sensitive and bypasses the model_lower normalization introduced above. For consistency with other model checks in this function (and with model_features.model_matches, which is case-insensitive), use model_lower (or another normalized form) so variants like provider-prefixed or differently-cased model IDs don’t skip this defaulting behavior.

        # Gemini 2.5-pro default to low if not set
        if "gemini-2.5-pro" in llm.model:
            if llm.reasoning_effort in {None, "none"}:
                out["reasoning_effort"] = "low"

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[enyst]

Thank you for the contribution! This bit of code might be unnecessary now, since we just merged #1989

[mayeco]

yes I understand @enyst but #1989 but still not fix my issue.

If I have a reasoning model like gemini, then set the temperature to 0.0. then will be 1 (the default)

https://github.com/OpenHands/software-agent-sdk/blob/main/openhands-sdk/openhands/sdk/llm/options/chat_options.py#L43

{
  "generationConfig": {
    "maxOutputTokens": 65536,
    "temperature": 1,  // ← Incorrect: should be 0.0
    "thinkingConfig": {
      "includeThoughts": true,
      "thinkingLevel": "HIGH"
    }
  }
}

[enyst]

Oh, I agree. I just don't fully understand, why keep the code for o1 and o3?

[mayeco]

yes. maybe is not only o1 / o3 we need to make a list of models that the reasoning fail when temperature is set.
that is better @enyst ?

[enyst]

Sounds good!

[all-hands-bot]

[Automatic Post]: It has been a while since there was any activity on this PR. @mayeco, are you still working on it? If so, please go ahead, if not then please request review, close it, or request that someone else follow up.

[mayeco]

@enyst This approach is currently specific to the Gemini model. If we find that other models support reasoning_effort+temperature+top_p, we can add them to the exception list later.

For now, we are 100% sure that Gemini supports it.

https://docs.cloud.google.com/vertex-ai/generative-ai/docs/learn/prompts/adjust-parameter-values

[Copilot]

Pull request overview

Copilot reviewed 1 out of 1 changed files in this pull request and generated no new comments.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[enyst]

Thank you!