fix: Always include reasoning_content for DeepSeek assistant messages #1789

simonrosenberg · 2026-01-22T09:44:22Z

Summary

This PR fixes issue #353 where DeepSeek requests fail with LLMBadRequestError: Missing \reasoning_content` field in the assistant message ... (400)`.

Problem

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 API returns a 400 error if this field is missing.

According to the DeepSeek documentation:

During the process of answering question 1 (Turn 1.1 - 1.3), the model performed multiple turns of thinking + tool calls before providing the answer. During this process, the user needs to send the reasoning content (reasoning_content) back to the API to allow the model to continue reasoning.

Solution

Modified Message.to_chat_dict() to always include reasoning_content for assistant messages when send_reasoning_content is True, even if the value is None or empty string.
Added deepseek-reasoner pattern to SEND_REASONING_CONTENT_MODELS to cover model names without provider prefix (e.g., deepseek-reasoner in addition to deepseek/deepseek-reasoner).
Added example 35_deepseek_reasoning_content.py demonstrating how to use DeepSeek's reasoning mode with tool calls.
Updated tests to reflect the new behavior where reasoning_content is always included for assistant messages when the feature is enabled.

Changes

openhands-sdk/openhands/sdk/llm/message.py: Modified to_chat_dict() to always include reasoning_content for assistant messages
openhands-sdk/openhands/sdk/llm/utils/model_features.py: Added deepseek-reasoner to SEND_REASONING_CONTENT_MODELS
examples/01_standalone_sdk/35_deepseek_reasoning_content.py: New example file
tests/sdk/llm/test_message.py: Updated tests for 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

Testing

All 498 tests in tests/sdk/llm/ pass, including the new tests for DeepSeek reasoning content behavior.

Fixes #353

@simonrosenberg can click here to continue refining the PR

Agent Server images for this PR

• GHCR package: https://github.com/OpenHands/agent-sdk/pkgs/container/agent-server

Variants & Base Images

Variant	Architectures	Base Image	Docs / Tags
java	amd64, arm64	`eclipse-temurin:17-jdk`	Link
python	amd64, arm64	`nikolaik/python-nodejs:python3.12-nodejs22`	Link
golang	amd64, arm64	`golang:1.21-bookworm`	Link

Pull (multi-arch manifest)

# Each variant is a multi-arch manifest supporting both amd64 and arm64
docker pull ghcr.io/openhands/agent-server:ef3a05f-python

Run

docker run -it --rm \
  -p 8000:8000 \
  --name agent-server-ef3a05f-python \
  ghcr.io/openhands/agent-server:ef3a05f-python

All tags pushed for this build

ghcr.io/openhands/agent-server:ef3a05f-golang-amd64
ghcr.io/openhands/agent-server:ef3a05f-golang_tag_1.21-bookworm-amd64
ghcr.io/openhands/agent-server:ef3a05f-golang-arm64
ghcr.io/openhands/agent-server:ef3a05f-golang_tag_1.21-bookworm-arm64
ghcr.io/openhands/agent-server:ef3a05f-java-amd64
ghcr.io/openhands/agent-server:ef3a05f-eclipse-temurin_tag_17-jdk-amd64
ghcr.io/openhands/agent-server:ef3a05f-java-arm64
ghcr.io/openhands/agent-server:ef3a05f-eclipse-temurin_tag_17-jdk-arm64
ghcr.io/openhands/agent-server:ef3a05f-python-amd64
ghcr.io/openhands/agent-server:ef3a05f-nikolaik_s_python-nodejs_tag_python3.12-nodejs22-amd64
ghcr.io/openhands/agent-server:ef3a05f-python-arm64
ghcr.io/openhands/agent-server:ef3a05f-nikolaik_s_python-nodejs_tag_python3.12-nodejs22-arm64
ghcr.io/openhands/agent-server:ef3a05f-golang
ghcr.io/openhands/agent-server:ef3a05f-java
ghcr.io/openhands/agent-server:ef3a05f-python

About Multi-Architecture Support

Each variant tag (e.g., ef3a05f-python) is a multi-arch manifest supporting both amd64 and arm64
Docker automatically pulls the correct architecture for your platform
Individual architecture tags (e.g., ef3a05f-python-amd64) are also available if needed

DeepSeek's thinking mode requires the reasoning_content field to be present in assistant messages during tool call turns, even if it's None. The API returns a 400 error if this field is missing. Changes: - Modified Message.to_chat_dict() to always include reasoning_content for assistant messages when send_reasoning_content is True - Added 'deepseek-reasoner' pattern to SEND_REASONING_CONTENT_MODELS to cover model names without provider prefix - Added example 35_deepseek_reasoning_content.py demonstrating the feature - Updated tests to reflect the new behavior Fixes #353 Co-authored-by: openhands <openhands@all-hands.dev>

github-actions · 2026-01-22T09:47:48Z

Coverage Report •

File	Stmts	Miss	Cover	Missing
openhands-sdk/openhands/sdk/llm
message.py	280	8	97%	370, 383–384, 392, 434, 530, 659–660
openhands-sdk/openhands/sdk/llm/utils
model_features.py	46	1	97%	32
TOTAL	16113	4752	70%

all-hands-bot

The fix correctly addresses the DeepSeek reasoning_content requirement. Found a few minor issues to consider improving code clarity and maintainability.

all-hands-bot · 2026-01-22T09:48:33Z

openhands-sdk/openhands/sdk/llm/message.py

-            message_dict["reasoning_content"] = self.reasoning_content
+        # Required for models like kimi-k2-thinking and deepseek-reasoner
+        # DeepSeek requires reasoning_content to be present in assistant messages
+        # during tool call turns, even if it's None


🟠 Important: The comment mentions "during tool call turns" which could be misinterpreted to mean that the code only includes reasoning_content for messages that have tool_calls. However, the implementation (lines 299-303) includes reasoning_content for ALL assistant messages when send_reasoning_content is True, regardless of whether tool calls are present.

Consider clarifying this comment to be more explicit, such as: "in all assistant messages when using reasoning models, even if it's None"

all-hands-bot · 2026-01-22T09:48:33Z

openhands-sdk/openhands/sdk/llm/message.py

+        if self.send_reasoning_content:
+            if self.role == "assistant":
+                # Always include reasoning_content for assistant messages when enabled
+                # Some providers (e.g., DeepSeek) require this field to be present
+                message_dict["reasoning_content"] = self.reasoning_content


🟡 Suggestion: The nested conditionals could be simplified for better readability:

Suggested change

if self.send_reasoning_content:

if self.role == "assistant":

# Always include reasoning_content for assistant messages when enabled

# Some providers (e.g., DeepSeek) require this field to be present

message_dict["reasoning_content"] = self.reasoning_content

if self.send_reasoning_content and self.role == "assistant":

# Always include reasoning_content for assistant messages when enabled

# Some providers (e.g., DeepSeek) require this field to be present

message_dict["reasoning_content"] = self.reasoning_content

This makes the condition more explicit and easier to understand at a glance.

all-hands-bot · 2026-01-22T09:48:34Z

openhands-sdk/openhands/sdk/llm/message.py

+            if self.role == "assistant":
+                # Always include reasoning_content for assistant messages when enabled
+                # Some providers (e.g., DeepSeek) require this field to be present
+                message_dict["reasoning_content"] = self.reasoning_content


🟡 Suggestion: This PR introduces a behavior change that restricts reasoning_content to assistant messages only. Previously, reasoning_content could be included for any role if both send_reasoning_content and reasoning_content were truthy.

While this change makes sense (reasoning is produced by assistants, not users/tools/system), it would be helpful to document this behavior change in the PR description or add a note to the changelog for transparency. This helps users understand what changed and why, even if it's unlikely to affect existing code.

all-hands-bot · 2026-01-22T09:48:34Z

examples/01_standalone_sdk/35_deepseek_reasoning_content.py

+from openhands.tools.terminal import TerminalTool
+
+
+logger = get_logger(__name__)


🟢 Nit: The logger is imported and created but never used in this example. While this appears to be a pattern in other example files for consistency, consider removing it if it's not needed.

(Or keep it if it's intentional for consistency with other examples - this is a very minor point.)

openhands-ai · 2026-01-22T09:50:06Z

Looks like there are a few issues preventing this PR from being merged!

GitHub Actions are failing:
- [Optional] Docs example
- Pre-commit checks

If you'd like me to help, just leave a comment, like

@OpenHands please fix the failing actions on PR #1789 at branch `openhands/fix-deepseek-reasoning-content`

Feel free to include any additional details that might help me get this PR into a better state.

_{^{You can manage your notification settings}}

enyst

Previous thinking about this:

#1464 (review)

I’m not fully sure what we may want to do here, and it may be worth digging a bit. Maybe the other version is unreachable today? 🤔

simonrosenberg · 2026-01-22T13:13:55Z

Previous thinking about this:

#1464 (review)

I’m not fully sure what we may want to do here, and it may be worth digging a bit. Maybe the other version is unreachable today? 🤔

This is a one-shot attempt with OpenHands...I should put it as draft

openhands-ai bot mentioned this pull request Jan 22, 2026

DeepSeek requests fail: Missing field (400) OpenHands/benchmarks#353

Open

all-hands-bot reviewed Jan 22, 2026

View reviewed changes

enyst reviewed Jan 22, 2026

View reviewed changes

simonrosenberg marked this pull request as draft January 22, 2026 13:14

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

fix: Always include reasoning_content for DeepSeek assistant messages #1789

fix: Always include reasoning_content for DeepSeek assistant messages #1789

Uh oh!

simonrosenberg commented Jan 22, 2026 •

edited by github-actions bot

Loading

Uh oh!

github-actions bot commented Jan 22, 2026

Uh oh!

all-hands-bot left a comment

Uh oh!

all-hands-bot Jan 22, 2026

Uh oh!

all-hands-bot Jan 22, 2026

Uh oh!

all-hands-bot Jan 22, 2026

Uh oh!

all-hands-bot Jan 22, 2026

Uh oh!

openhands-ai bot commented Jan 22, 2026

Uh oh!

enyst left a comment

Uh oh!

simonrosenberg commented Jan 22, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

5 participants

		from openhands.tools.terminal import TerminalTool


		logger = get_logger(__name__)

fix: Always include reasoning_content for DeepSeek assistant messages #1789

Are you sure you want to change the base?

fix: Always include reasoning_content for DeepSeek assistant messages #1789

Uh oh!

Conversation

simonrosenberg commented Jan 22, 2026 • edited by github-actions bot Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Problem

Solution

Changes

Testing

Uh oh!

github-actions bot commented Jan 22, 2026

Uh oh!

all-hands-bot left a comment

Choose a reason for hiding this comment

Uh oh!

all-hands-bot Jan 22, 2026

Choose a reason for hiding this comment

Uh oh!

all-hands-bot Jan 22, 2026

Choose a reason for hiding this comment

Uh oh!

all-hands-bot Jan 22, 2026

Choose a reason for hiding this comment

Uh oh!

all-hands-bot Jan 22, 2026

Choose a reason for hiding this comment

Uh oh!

openhands-ai bot commented Jan 22, 2026

Uh oh!

enyst left a comment

Choose a reason for hiding this comment

Uh oh!

simonrosenberg commented Jan 22, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

5 participants

simonrosenberg commented Jan 22, 2026 •

edited by github-actions bot

Loading