fix: improve GLM-4.6 thinking token support for better compatibility

[roomote]

Description

This PR improves GLM-4.6 thinking token support to ensure compatibility with various OpenAI-compatible endpoints, particularly addressing the issue reported with ik_llama.cpp.

Problem

As reported in #8547 and by @ChicoPinto70, the previous implementation in PR #8548 did not work with all OpenAI-compatible endpoints, specifically ik_llama.cpp. The issue was that:

• The thinking: { type: "enabled" } parameter was always added for GLM-4.6 models
• Some endpoints (like ik_llama.cpp) may not recognize this parameter and could reject requests or fail to process them correctly

Solution

This implementation takes a more conservative and compatible approach:

Key Changes:

1. Optional thinking parameter: The thinking parameter is now disabled by default and only added when explicitly enabled via configuration (openAiEnableThinkingParameter: true)

2. Multiple parsing strategies: The implementation now handles thinking tokens through three different methods to ensure maximum compatibility:

   • XML tags (<think>...</think>) parsing using XmlMatcher
   • Direct reasoning_content field in the response delta (as used by some implementations)
   • The optional thinking parameter for endpoints that support it

3. Comprehensive testing: Added extensive tests covering all scenarios including:

   • Default behavior (no thinking parameter)
   • Explicit enabling of thinking parameter
   • XML tag parsing for GLM-4.6
   • reasoning_content field handling
   • Mixed format support

Benefits

• ik_llama.cpp compatibility: Works with local ik_llama.cpp setups by not sending unsupported parameters by default
• Flexibility: Can be configured to work with different endpoint implementations
• Backward compatibility: Does not break existing setups
• Future-proof: Can handle thinking tokens regardless of how the endpoint provides them

Testing

All tests pass ✅:

• Unit tests: 13 passed
• Type checking: Successful
• Linting: No issues

Related Issues

Fixes #8547
Addresses feedback from PR #8548

For @kavehsfv

This implementation ensures GLM-4.6 thinking tokens work across different OpenAI-compatible endpoints. Once merged, it will be included in the next release cycle. The fix is backward-compatible and should work with your setup.

cc: @ChicoPinto70 - This should now work with your ik_llama.cpp setup. The thinking parameter is disabled by default for maximum compatibility.

---

Important

Improves GLM-4.6 thinking token support by making the thinking parameter optional and adding multiple parsing strategies for better compatibility.

• Behavior:
  • thinking parameter is now optional for GLM-4.6 models, enabled via openAiEnableThinkingParameter.
  • Supports multiple parsing strategies for thinking tokens: XML tags, reasoning_content, and optional thinking parameter.
• Testing:
  • Added tests in base-openai-compatible-provider.spec.ts for default behavior, explicit enabling, XML tag parsing, reasoning_content handling, and mixed formats.
• Functions:
  • Updated createStream() and createMessage() in base-openai-compatible-provider.ts to handle new logic for thinking tokens.
  • Added isGLM46Model() and shouldAddThinkingParameter() to determine model type and parameter inclusion.

^{This description was created by for d9e20b2. You can customize this summary. It will automatically update as commits are pushed.}

[roomote]

The third condition lowerModel === "glm-4.6" is redundant since it's already covered by the first condition lowerModel.includes("glm-4.6"). The includes check will return true whenever the exact match check would return true.

Suggested change

	return lowerModel.includes("glm-4.6") || lowerModel.includes("glm-4-6") || lowerModel === "glm-4.6"
	return lowerModel.includes("glm-4.6") || lowerModel.includes("glm-4-6")