Summary

• Extract OPENAI_REASONING_MODEL_PREFIXES from local variable inside generate_with_context() to a module-level constant
• Extract is_reasoning_model() as a standalone, importable function with 3-state logic
• Add is_reasoning_model: Optional[bool] = None field to LLMModelConfig

Motivation

Users of non-OpenAI providers (Gemini 2.5 thinking models, DeepSeek-R1, Qwen thinking, etc.) cannot currently configure their models as reasoning models. The hardcoded prefix list only covers OpenAI models, and expanding it with non-OpenAI prefixes is fragile:

• Gemini thinking is an optional mode — the same model name (gemini-2.5-flash) can run with or without thinking
• gemini-2.5-flash-lite shares the gemini-2.5- prefix but is NOT a thinking model
• Future model naming conventions are unpredictable across providers

Solution

A per-model config flag with auto-detection fallback:

• True → force reasoning model conventions (max_completion_tokens, no temperature/top_p)
• False → force standard conventions
• None (default) → auto-detect via OpenAI prefix list (fully backward-compatible)

llm:
  models:
    - name: "gemini-2.5-flash"
      api_base: "https://generativelanguage.googleapis.com/v1beta/openai/"
      is_reasoning_model: true     # explicit opt-in for thinking mode

    - name: "o3-mini"
      # is_reasoning_model omitted → auto-detect → true (OpenAI prefix match)

    - name: "claude-sonnet-4-5-20250929"
      api_base: "https://api.anthropic.com/v1"
      # is_reasoning_model omitted → auto-detect → false

Additional cleanup

• Removed redundant dash-prefixes (o1-, o3-, gpt-5-) already covered by base prefixes (o1, o3, gpt-5)
• Replaced full model names gpt-oss-120b/gpt-oss-20b with gpt-oss- prefix
• Rewrote test_openai_model_detection.py to import and test the real is_reasoning_model() function instead of duplicating detection logic with a non-existent api_base guard

Test plan

• 12 new tests in test_reasoning_model_detection.py (3-state logic, auto-detect, explicit override, case insensitivity, backward compat)
• Updated test_openai_model_detection.py to use real function + added gpt-oss coverage
• All 384 existing tests pass (0 failures, 0 regressions)

🤖 Generated with Claude Code