fix: support reasoning_details format for Gemini 3 models

[daniel-lxs]

Summary

Fixes support for Gemini 3 models on OpenRouter by implementing proper handling of the reasoning_details array format.

Problem

Gemini 3 models were causing 400 errors because they use the new reasoning_details array format instead of the legacy reasoning string format. The reasoning details need to be preserved and sent back to the API in multi-turn conversations (especially for tool calling workflows).

Solution

Implemented full support for the reasoning_details format following OpenRouter's documentation:

1. Created OPEN_ROUTER_REASONING_DETAILS_MODELS set to track affected models
2. Accumulate full reasoning_details array during streaming (not just text extraction)
3. Store reasoning_details on assistant messages via ApiMessage type
4. Set preserveReasoning: true for models in the set
5. Preserve reasoning_details when converting messages and sending back to API

Changes

• packages/types/src/providers/openrouter.ts: Added model constant
• src/api/providers/openrouter.ts: Added accumulator and getter method
• src/api/providers/fetchers/openrouter.ts: Set preserveReasoning flag
• src/api/transform/openai-format.ts: Preserve reasoning_details in conversion
• src/core/task-persistence/apiMessages.ts: Added type support
• src/core/task/Task.ts: Store and send back reasoning_details

Testing

✅ All OpenRouter provider tests pass (12/12)
✅ All reasoning preservation tests pass (6/6)
✅ All Task core tests pass (34 passed, 4 skipped)
✅ All transform tests pass (4/4)
✅ All provider tests pass (102/102)

Related Issues

• Upstream issue: Gemini 3.0 Pro Preview returns 400 INVALID_ARGUMENT error when used via OpenRouter or Cline account cline/cline#7551
• Follows OpenRouter documentation on reasoning tokens

Notes

• Uses the preserveReasoning flag mechanism (no needless conditionals)
• Only affects models explicitly listed in the set
• Maintains backward compatibility with legacy format
• Easy to add more models as needed

---

Important

Adds support for reasoning_details format in Gemini 3 models, ensuring compatibility with OpenRouter and preserving details in multi-turn conversations.

• Behavior:
  • Adds support for reasoning_details format in Gemini 3 models, preserving details in multi-turn conversations.
  • Introduces OPEN_ROUTER_REASONING_DETAILS_MODELS set to track models using reasoning_details.
  • Accumulates reasoning_details during streaming in openrouter.ts.
  • Preserves reasoning_details in openai-format.ts and Task.ts.
• Files:
  • openrouter.ts: Adds accumulator for reasoning_details, resets accumulator per request, and processes reasoning details for Gemini models.
  • openai-format.ts: Preserves reasoning_details when converting messages.
  • Task.ts: Stores and sends back reasoning_details in addToApiConversationHistory().
• Testing:
  • All relevant tests pass, including OpenRouter provider tests, reasoning preservation tests, Task core tests, and transform tests.

^{This description was created by for 6b741a2. You can customize this summary. It will automatically update as commits are pushed.}

[roomote]

Rooviewer   See task on Roo Cloud

Re-review completed successfully. No new issues found.

✅ All checks passed - The changes correctly implement universal reasoning_details handling by removing the model allowlist and processing the chunks dynamically when present in the stream.

Previous reviews

• 7939d06: Review #1
• fe1a9f5: Review #2

_{Mention @roomote in a comment to request specific changes to this pull request or fix all unresolved issues.}