fix: inject thought signatures for Gemini 3 when switching models via LiteLLM

[daniel-lxs]

Problem

When users switch mid-task from Claude (or other models) to Gemini 3/2.5 via LiteLLM, they encounter "Corrupted thought signature" errors. This happens because Gemini 3 validates thought signatures for tool/function calling steps, and conversation history from other models lacks these signatures.

Solution

Implemented thought signature injection using provider_specific_fields.thought_signature on tool calls, following LiteLLM's official documentation:

• Thought signatures are stored in provider_specific_fields.thought_signature of tool calls
• Dummy signature is base64("skip_thought_signature_validator") - this bypasses validation
• Only the first tool call needs the signature for parallel function calls

Changes

• Added isGeminiModel() method to detect Gemini 3.x and 2.5.x models (including provider-prefixed variants)
• Added injectThoughtSignatureForGemini() method to inject dummy signatures
• Updated createMessage() to apply injection when targeting Gemini with native tool protocol
• Added 8 new tests covering model detection, injection behavior, and integration

Test Results

All tests pass (5246 passed).

Fixes: COM-489

---

Important

Injects thought signatures for Gemini models in LiteLLM to prevent errors when switching models, with new detection and injection methods and comprehensive tests.

• Behavior:
  • Adds isGeminiModel() to detect Gemini 3.x and 2.5.x models in lite-llm.ts.
  • Adds injectThoughtSignatureForGemini() to inject dummy thought signatures for Gemini models.
  • Updates createMessage() to apply thought signature injection for Gemini models with native tool protocol.
• Tests:
  • Adds tests in lite-llm.spec.ts for Gemini model detection and thought signature injection.
  • Tests cover model detection, injection behavior, and integration with createMessage().
• Misc:
  • Updates createMessage() to handle tool protocol resolution and Gemini-specific processing.

^{This description was created by for d705365. You can customize this summary. It will automatically update as commits are pushed.}

[roomote]

Rooviewer   See task on Roo Cloud

Reviewed the thought signature injection implementation for Gemini models. The core logic is sound and test coverage is comprehensive. Found one minor documentation issue.

• Fix misleading comment at line 203: says "reasoning.encrypted block" but should say "provider_specific_fields.thought_signature"

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

[roomote]

The comment mentions "inject fake reasoning.encrypted block" but the code actually injects provider_specific_fields.thought_signature. This inconsistency could confuse future developers trying to understand or modify this code.

Suggested change

	// For Gemini models with native protocol: inject fake reasoning.encrypted block for tool calls
	// This is required when switching from other models to Gemini to satisfy API validation.
	// For Gemini models with native protocol: inject thought signatures via provider_specific_fields
	// This is required when switching from other models to Gemini to satisfy API validation.

_{Fix it with Roo Code or mention @roomote and request a fix.}