[Bug]: CLI v1.14.0 fails with Anthropic models via LiteLLM proxy - temperature and top_p cannot both be specified

[jpshackelford]

Is there an existing issue for the same bug?

• I have checked the existing issues.

Describe the bug and reproduction steps

CLI v1.14.0 fails immediately when using Anthropic models via LiteLLM proxy. The error indicates both temperature and top_p are being sent, which Anthropic does not allow.

Steps to reproduce:

1. Install CLI v1.14.0: uvx openhands@1.14.0
2. Configure model via LiteLLM proxy (e.g., litellm_proxy/prod/claude-opus-4-5-20251101)
3. Start a conversation and send any message
4. Error occurs immediately

Error message:

Conversation Error
Code: LLMBadRequestError

Detail:
litellm.BadRequestError: AnthropicException - {
  "type": "error",
  "error": {
    "type": "invalid_request_error",
    "message": "`temperature` and `top_p` cannot both be specified for this model.
                Please use only one."
  }
}

Workaround: Delete ~/.openhands/agent_settings.json and re-run openhands login to regenerate with current SDK defaults.

Root Cause

Two contributing factors have been identified:

1. Stale agent_settings.json: Users who configured CLI with older SDK versions (pre-v1.12.0) have temperature: 0.0 and top_p: 1.0 persisted in their settings. These were SDK defaults that were serialized when saving the agent config.

2. LiteLLM model recognition: For proxy models like litellm_proxy/prod/claude-opus-4-5-20251101, the SDK strips litellm_proxy/ but LiteLLM doesn't recognize prod/claude-opus-4-5-20251101 as a valid model path. This causes supports_reasoning_effort to return False, so temp/top_p aren't stripped from the request.

See upstream issue for full investigation: OpenHands/software-agent-sdk#2686

OpenHands Installation

CLI (via uvx)

OpenHands Version

1.14.0 (regression from 1.13.x)

Model Name

litellm_proxy/prod/claude-opus-4-5-20251101 (Anthropic Claude Opus 4.5 via LiteLLM proxy)

Operating System

Linux

---

This issue was created by an AI assistant (OpenHands) on behalf of the user.

[jpshackelford]

Reproduction Confirmed

Tested SDK versions directly to isolate the regression:

from openhands.sdk.llm.options.chat_options import select_chat_options

class MockLLM:
    model = "litellm_proxy/prod/claude-opus-4-5-20251101"
    temperature = 0.7
    top_p = 0.95
    # ... other fields

result = select_chat_options(MockLLM(), {}, has_tools=True)

Results:

SDK Version	Output
v1.11.5	{max_completion_tokens: 4096} ✅
v1.14.0	{max_completion_tokens: 4096} ✅
v1.15.0	{top_p: 0.95, temperature: 0.7, max_completion_tokens: 4096} ❌
v1.16.1	{top_p: 0.95, temperature: 0.7, max_completion_tokens: 4096} ❌

The regression was introduced in SDK v1.15.0 by PR #2277.

Upstream issue filed: OpenHands/software-agent-sdk#2686

[jpshackelford]

@malhotra5 How do you have your model name set? Am I using an odd model name form? Mine was set in cli via:

$ openhands login

[jpshackelford]

Summary of Investigation

The original diagnosis (PR #2277 being the cause) was incorrect. Two factors are involved:

1. Stale settings: ~/.openhands/agent_settings.json contains temperature: 0.0 and top_p: 1.0 from pre-v1.12.0 SDK defaults
2. Model recognition: LiteLLM doesn't recognize prod/claude-opus-4-5-20251101 after stripping litellm_proxy/, so supports_reasoning_effort is False and temp/top_p aren't stripped

Workaround: Delete ~/.openhands/agent_settings.json and re-run openhands login.

See OpenHands/software-agent-sdk#2686 for full analysis and fix options.

---

This comment was created by an AI assistant (OpenHands) on behalf of the user.

[jpshackelford]

CLI Fix Options

Required:

1. Upgrade to fixed SDK version when it lands (tracks [Bug]: LiteLLM proxy models send both temperature and top_p to Anthropic (regression in v1.15.0) software-agent-sdk#2686)

Optional (nice-to-have):
2. Add startup cleanup to remove stale temperature: 0.0 / top_p: 1.0 from agent_settings.json

Once SDK is fixed, CLI could do:

from openhands.sdk.llm.options.chat_options import supports_reasoning_effort

if supports_reasoning_effort(settings.llm.model):
    if settings.llm.temperature is not None or settings.llm.top_p is not None:
        settings.llm.temperature = None
        settings.llm.top_p = None
        save_agent_settings(settings)

This cleanup doesn't provide protection — the SDK handles stripping at request time. It only keeps the settings file tidy to reduce user confusion if they inspect it.

Open question: Is #2 worth doing?

---

This comment was created by an AI assistant (OpenHands) on behalf of the user.

[jpshackelford]

@malhotra5 @enyst I am happy to open a PR for part 2 above if you think it's worth doing. What do you think?

[enyst]

I think changing agent_settings.json means overwriting user choices there. I mean, we did write these two by default, yes, but on the other hand people can always overwrite them and maybe some have.

Although... Maybe with a notification, we can do it. That is, at the next update, maybe we can show the user at first run of that update, a small notification box that temperature and top_p have been cleared out and they can set them in agent_settings.json... WDYT?