Rodrigo Rodriguez (Pragmatismo)
4ce06daf75
- Added initial 30s delay to compact prompt scheduler - Implemented async LLM summarization for conversation history - Reduced lock contention by minimizing critical sections - Added fallback to original text if summarization fails - Updated README with guidance for failed requirements - Added new `summarize` method to LLMProvider trait - Improved session manager query with proper DSL usage The changes optimize the prompt compaction process by: 1. Reducing lock contention through better resource management 2. Adding LLM-based summarization for better conversation compression 3. Making the system more resilient with proper error handling 4. Improving documentation for development practices
966 B
966 B
LLM Strategy & Workflow
Fallback Strategy (After 3 attempts / 10 minutes):
When initial attempts fail, sequentially try these LLMs:
- DeepSeek-V3-0324 (good architect, adventure, reliable, let little errors just to be fixed by gpt-*)
- DeepSeek-V3.1 (slower)
- gpt-5-chat (slower, let warnings...)
- gpt-oss-120b
- Claude (Web): Copy only the problem statement and create unit tests. Create/extend UI.
- Llama-3.3-70B-Instruct (alternative)
Development Workflow:
- One requirement at a time with sequential commits
- On error: Stop and consult Claude for guidance
- Change progression: Start with DeepSeek, conclude with gpt-oss-120b
- If a big req. fail, specify a @code file that has similar pattern or sample from official docs.
- Final validation: Use prompt "cargo check" with gpt-oss-120b
- Be humble, one requirement, one commit. But sometimes, freedom of caos is welcome - when no deadlines are set.