Issue
When processing large diffs (many files or large changes), the token counting and truncation logic may not properly account for all providers' token limits, leading to API errors.
Current Behavior
src/gac/diff_scoring.py implements smart_truncate_diff but uses a fixed model (anthropic:claude-3-haiku-latest) for token counting- Different providers have vastly different context windows:
- OpenAI GPT-4o: 128k tokens
- Anthropic Claude: 200k tokens
- Some local models (Ollama): 4k-8k tokens
- Gemini: 1M+ tokens
- The
max_tokens parameter in CLI defaults to 4000 but isn't validated against the selected model's limits
Risk
- Users with local models (Ollama, LM Studio) will hit token limit errors on moderately large diffs
- No clear error message when token limit is exceeded - just generic API errors
- Grouped commits (
-g flag) multiply token usage but scaling is arbitrary (multiplier = min(5, 2 + (num_files // 10)))
Issue
When processing large diffs (many files or large changes), the token counting and truncation logic may not properly account for all providers' token limits, leading to API errors.
Current Behavior
src/gac/diff_scoring.pyimplementssmart_truncate_diffbut uses a fixed model (anthropic:claude-3-haiku-latest) for token countingmax_tokensparameter in CLI defaults to 4000 but isn't validated against the selected model's limitsRisk
-gflag) multiply token usage but scaling is arbitrary (multiplier = min(5, 2 + (num_files // 10)))