feat: Add Deepseek v3.2 interleaved thinking support #9925
+2,249
−27
Conversation
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
added 7 commits
December 8, 2025 19:07
Implement interleaved thinking mode support for DeepSeek V3.2 with generic capability-based architecture for extensibility to other models. Key changes: - Core thinking mode: parameter handling, temperature override, streaming - Multi-turn conversations: conditional reasoning clearing, turn detection - Task persistence: reasoning_content save/restore - Generic refactoring: capability flags replace model-specific checks Makes it easy to add interleaved thinking for other models (Kimi K2, Minimax M2) via capability flags. Note: Checkpoint commit. Backward compatibility verification and documentation pending. No manual testing performed yet.
…r OpenAI compatible providers - Set supportsInterleavedThinking default to false in openAiModelInfoSaneDefaults - Remove enabledR1Format fallback, rely solely on modelInfo.supportsInterleavedThinking flag - Add UI checkbox in OpenAICompatible settings to configure supportsInterleavedThinking for custom models
- Fix turn detection to preserve reasoning_content when user messages contain tool_result blocks (still in tool call sequence) - Fix turn detection when assistant stops sending tool calls after receiving tool results - Convert tool_use blocks from assistant messages to OpenAI tool_calls format - Convert tool_result blocks from user messages to tool role messages - Preserve tool_calls when merging consecutive assistant messages These changes ensure proper handling of tool calls and tool results in the interleaved thinking message conversion pipeline, maintaining reasoning_content preservation during tool call sequences while correctly clearing it between turns.
… history for DeepSeek models
Update three tests to expect `true` (clear reasoning_content) when assistant stops making tool calls after receiving results. These were outdated after commit 477d917 changed the implementation to detect tool sequence completion.
Contributor
All issues resolved. The implementation is clean and ready for merge.
Mention @roomote in a comment to request specific changes to this pull request or fix all unresolved issues. |
![roomote[bot]](https://avatars.githubusercontent.com/in/1546624?s=60&v=4){kind=small}
src/api/providers/openai.ts
Outdated
Comment on lines
219
to
221
| let finalReasoningContent = "" | ||
| let finalContent = "" | ||
| let finalToolCalls: any[] = [] |
Contributor
There was a problem hiding this comment.
These two variables (finalReasoningContent and finalContent) are declared but never used. The comment on line 266 mentions "Track tool calls for debug logging" but only finalToolCalls is actually used (to accumulate tool call data). Consider removing these unused variables to reduce dead code.
Suggested change
| let finalReasoningContent = "" | |
| let finalContent = "" | |
| let finalToolCalls: any[] = [] | |
| let finalToolCalls: any[] = [] |
Fix it with Roo Code or mention @roomote and request a fix.
hannesrudolph
added
the
Issue/PR - Triage
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 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.
Related GitHub Issue
Closes: # 9779
Description
reasoning_contenthandling in streaming and non-streaming modesTest Procedure
deepseek-chatanddeepseek-reasonermodelPre-Submission Checklist
Screenshots / Videos
Added "Interleaved Thinking" toggle to OpenAI provider
Documentation Updates
Does this PR necessitate updates to user-facing documentation?
Additional Notes
1. Streaming Content Behavior
Streaming content sometimes appears in chunks rather than smoothly. It's unclear whether this is due to my implementation or Deepseek's API behavior. This doesn't affect functionality but may impact user experience.
2. Reasoning Content Continuity Verification
While the implementation follows Deepseek's API requirements for passing
reasoning_contentback to the API, I haven't been able to definitively verify that Deepseek is receiving it correctly. However, the API would return 400 errors if not handled properly, and I haven't observed any such errors, suggesting the implementation is working correctly.3. Limited Testing Scope
This has only been tested with Deepseek models (
deepseek-chatanddeepseek-reasoner). No regression testing has been performed with other OpenAI-compatible providers.Get in Touch
DM skelectric on Discord for questions