feat: improve R1 format setting labels for reasoning models #10971
Conversation
…ng models Updated the "Enable R1 model parameters" setting label and description to be more descriptive and inclusive of other reasoning/thinking models. The new labeling explicitly mentions: - DeepSeek R1 - Kimi-K2 - QWQ - NVIDIA NIM and other OpenAI-compatible providers This makes it clearer to users that the setting should be enabled when using any reasoning/thinking model through OpenAI-compatible APIs, not just R1. Addresses Issue #10969 (user suggestion from @bozoweed)
Reviewed the new R1 format auto-detection changes. The implementation correctly expands support to reasoning/thinking models (DeepSeek R1, Kimi-K2-thinking, QWQ) with case-insensitive matching and well-documented code. No issues found. Mention @roomote in a comment to request specific changes to this pull request or fix all unresolved issues. |
|
@roomote why did only change label ? we get a real issue about R1 usage take a look
|
🔧 Roo Code Fixer✅ Fix CompletedProblem: Using reasoning/thinking models like Solution: Expanded the R1 format auto-detection logic in
Changes:
Commit: Tests: All 48 OpenAI provider tests and 20 R1 format tests pass ✓ |
- Auto-detect models with "-thinking" suffix (e.g., kimi-k2-thinking) - Auto-detect DeepSeek R1 models (deepseek-r1, deepseek/deepseek-r1) - Auto-detect QWQ models - Renamed deepseekReasoner to useR1Format for clarity - Users can still manually enable R1 format via the toggle This improves support for reasoning/thinking models like Kimi-K2 that require R1 format for proper response handling.
Summary
This PR attempts to address Issue #10969 (comment from @bozoweed suggesting a more generic approach).
Changes
Updated the "R1 format" setting labels and descriptions in the OpenAI Compatible provider to be more descriptive and inclusive of other reasoning/thinking models. The new labeling explicitly mentions:
Before
After
Files Changed
Why This Approach
Rather than creating a dedicated NVIDIA provider (as originally suggested), this follows the user suggestion to make the existing OpenAI Compatible provider more flexible by improving the labeling of the existing R1 format toggle. This approach:
Feedback and guidance are welcome!