fix(formatter): preserve reasoning_content for tool-call segments in …#1323
Merged
LearningGp merged 1 commit intoMay 21, 2026
Merged
Conversation
…DeepSeek thinking mode When thinking mode is enabled and a user turn contains tool calls, all assistant messages in that turn must preserve their reasoning_content when sent back to the DeepSeek API in subsequent requests. Previously, reasoning_content was only kept for the current turn, causing HTTP 400 errors in multi-turn tool-calling scenarios. See: https://api-docs.deepseek.com/guides/thinking_mode#tool-calls Fixes agentscope-ai#1281 1.0.12-SNAPSHOT **Background:** DeepSeek's thinking mode requires `reasoning_content` to be preserved for all assistant messages within a user turn that contains tool calls. The original code only preserved it for the current turn, causing HTTP 400 errors in multi-turn tool-calling scenarios. **Changes:** - `applyDeepSeekFixes()`: Added segment-level tool call detection via `computeSegmentToolFlags()`. A segment is defined as messages between two consecutive user messages. If any message in a segment contains tool calls, all assistant messages in that segment are flagged to preserve their `reasoning_content`. - `fixMessage()`: Renamed parameter from `isCurrentTurn` to `needReasoning` to reflect that preservation is now determined by the segment analysis, not just the turn boundary. - Added `ReasoningPreservationTests` nested test class covering key scenarios: non-thinking mode fallback, multiple rounds with tool calls, mixed text/tool rounds, text-only messages in tool-call segments, and regression for text-only segments. - [x] Code has been formatted with `mvn spotless:apply` - [x] All tests are passing (`mvn test`) - [x] Javadoc comments are complete and follow project conventions - [ ] Related documentation has been updated - [x] Code is ready for review
|
|
Contributor
Author
|
您的邮件我已收到，谢谢。——收信人：吴宇杰
|
Codecov Report❌ Patch coverage is
📢 Thoughts on this report? Let us know! |
Contributor
Author
|
@LearningGp PTAL — could you help request a review from the team? All checks have passed. Thanks! |
liangxingguang
pushed a commit
to liangxingguang/agentscope-java
that referenced
this pull request
May 21, 2026
agentscope-ai#1323) When thinking mode is enabled and a user turn contains tool calls, all assistant messages in that turn must preserve their reasoning_content when sent back to the DeepSeek API in subsequent requests. Previously, reasoning_content was only kept for the current turn, causing HTTP 400 errors in multi-turn tool-calling scenarios. See: https://api-docs.deepseek.com/guides/thinking_mode#tool-calls Fixes agentscope-ai#1281 ## AgentScope-Java Version 1.0.12-SNAPSHOT ## Description **Background:** DeepSeek's thinking mode requires `reasoning_content` to be preserved for all assistant messages within a user turn that contains tool calls. The original code only preserved it for the current turn, causing HTTP 400 errors in multi-turn tool-calling scenarios. **Changes:** - `applyDeepSeekFixes()`: Added segment-level tool call detection via `computeSegmentToolFlags()`. A segment is defined as messages between two consecutive user messages. If any message in a segment contains tool calls, all assistant messages in that segment are flagged to preserve their `reasoning_content`. - `fixMessage()`: Renamed parameter from `isCurrentTurn` to `needReasoning` to reflect that preservation is now determined by the segment analysis, not just the turn boundary. - Added `ReasoningPreservationTests` nested test class covering key scenarios: non-thinking mode fallback, multiple rounds with tool calls, mixed text/tool rounds, text-only messages in tool-call segments, and regression for text-only segments. ## Checklist - [x] Code has been formatted with `mvn spotless:apply` - [x] All tests are passing (`mvn test`) - [x] Javadoc comments are complete and follow project conventions - [ ] Related documentation has been updated - [x] Code is ready for review
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
When thinking mode is enabled and a user turn contains tool calls,
all assistant messages in that turn must preserve their
reasoning_content when sent back to the DeepSeek API in subsequent
requests. Previously, reasoning_content was only kept for the
current turn, causing HTTP 400 errors in multi-turn tool-calling
scenarios.
See: https://api-docs.deepseek.com/guides/thinking_mode#tool-calls
Fixes #1281
AgentScope-Java Version
1.0.12-SNAPSHOT
Description
Background: DeepSeek's thinking mode requires
reasoning_contentto be preserved for allassistant messages within a user turn that contains tool calls. The original code only preserved it
for the current turn, causing HTTP 400 errors in multi-turn tool-calling scenarios.
Changes:
applyDeepSeekFixes(): Added segment-level tool call detection viacomputeSegmentToolFlags(). Asegment is defined as messages between two consecutive user messages. If any message in a segment
contains tool calls, all assistant messages in that segment are flagged to preserve their
reasoning_content.fixMessage(): Renamed parameter fromisCurrentTurntoneedReasoningto reflect thatpreservation is now determined by the segment analysis, not just the turn boundary.
ReasoningPreservationTestsnested test class covering key scenarios: non-thinking modefallback, multiple rounds with tool calls, mixed text/tool rounds, text-only messages in tool-call
segments, and regression for text-only segments.
Checklist
mvn spotless:applymvn test)