DeepSeek requests fail: Missing field (400)

[simonrosenberg]

Summary

During evaluation runs with DeepSeek, many conversation attempts fail with LLMBadRequestError: Missing \reasoning_content` field in the assistant message ... (400). The model deepseek/deepseek-reasonerrequires areasoning_content` field for thinking/tool-call mode, but current request payloads omit it, causing hard 400s and repeated retries.

Evidence

• Eval job evaluation-jobs/eval-eval-21233988879-deepseek-v-w57bc logs contain 190+ occurrences of this error (and rising). Examples:
  • Instance sympy__sympy-22080 ... Error code: 400 ... Missing \reasoning_content` field in the assistant message at message index 96`
  • Instance django__django-14315 ... 400 ... Missing \reasoning_content` field in the assistant message at message index 48`
  • Instance matplotlib__matplotlib-25122 ... 400 ... Missing \reasoning_content` field in the assistant message at message index 104`
• Counts from run output logs: grep -R "Missing \reasoning_content`" logs | wc -l` → 190+ (later 197) occurrences in the run directory.
• Model group is deepseek/deepseek-reasoner; available fallbacks: None (per error message).

Root cause

The request payload to DeepSeek’s thinking/tool-call endpoint does not include the required reasoning_content in assistant messages. DeepSeek rejects such requests with HTTP 400, halting the conversation. This is a client payload/SDK issue, not a transient service issue.

Impact

• Many instances fail outright and retry up to max attempts, stalling evaluation progress and leaving statuses null.
• Wasted tokens/time; downstream metrics and reports incomplete.

Proposed fix

• Update the OpenHands client/SDK conversation builder for DeepSeek thinking mode to populate reasoning_content for assistant messages/tool calls as per https://api-docs.deepseek.com/guides/thinking_mode#tool-calls.
• If not feasible immediately, switch to a model configuration that does not require reasoning_content until the payload is fixed.
• Add a safeguard in the request layer to detect missing reasoning_content and fail fast with a clearer error before hitting the API.

[simonrosenberg]

@OpenHands clone software-agent-sdk and read the code to understand how the llm class is implemented.
Read the official deepseek documentation and try to find a fix that follows the same coding patterns as the other llms to populate reasoning_content.
Then add a simple runnable example deepseek_reasoning_content.py that instanciates a LLM using deepseek and does inference on a task that should use reasoning. Read the other examples for inspiration to understand how they work.

[openhands-ai]

I'm on it! simonrosenberg can track my progress at all-hands.dev

[openhands-ai]

Summary

I have successfully fixed issue #353 "DeepSeek requests fail: Missing reasoning_content field (400)" in the software-agent-sdk repository.

Root Cause

DeepSeek's thinking mode (deepseek-reasoner) requires the reasoning_content field to be present in assistant messages during tool call turns, even if it's None. The previous implementation only included reasoning_content when it was truthy, causing the API to return 400 errors.

Changes Made

1. openhands-sdk/openhands/sdk/llm/message.py: Modified to_chat_dict() to always include reasoning_content for assistant messages when send_reasoning_content is True, regardless of whether the value is None or empty.

2. openhands-sdk/openhands/sdk/llm/utils/model_features.py: Added "deepseek-reasoner" pattern to SEND_REASONING_CONTENT_MODELS to cover model names without provider prefix.

3. examples/01_standalone_sdk/35_deepseek_reasoning_content.py: Created a new example demonstrating DeepSeek's reasoning content feature with tool calls.

4. Tests updated:

   • tests/sdk/llm/test_message.py: Updated tests to reflect new behavior
   • tests/sdk/llm/test_model_features.py: Added test cases for DeepSeek model patterns
   • tests/sdk/llm/test_reasoning_content.py: Added DeepSeek-specific tests

Verification

• All 498 tests in tests/sdk/llm/ pass
• The fix follows the existing coding patterns for handling model-specific features

Checklist

• ✅ Cloned software-agent-sdk and understood LLM class implementation
• ✅ Read DeepSeek official documentation for reasoning_content requirements
• ✅ Followed existing coding patterns (send_reasoning_content feature flag)
• ✅ Implemented fix for DeepSeek reasoning_content field
• ✅ Created example 35_deepseek_reasoning_content.py
• ✅ Added appropriate tests
• ✅ Created PR

Pull Request

PR #1789: OpenHands/software-agent-sdk#1789

View full conversation