0e2edfdebf
Quick Win 1: Close Self-Report Feedback Loop [9/10 impact] - Added self-report-fixer.js module with automatic fix classification - Pattern-based detection for high-confidence fixes (e.g., prompt rubrics) - Deduplication and severity-based categorization of reports - Designed for extension into triage-self-feedback pipeline Quick Win 2: Activate Continuous Model Learning [8/10 impact] - Added model-learner.js with ModelPerformanceTracker class - Per-task-type tracking: success rate, latency, cost, token efficiency - Auto-demotion for models failing >50% on specific task types - A/B testing infrastructure for hypothesis testing on low-risk tasks - Failure analysis with pattern detection (e.g., timeouts, quality issues) - Storage: .sf/model-performance.json, .sf/model-failure-log.jsonl Quick Win 3: Automate Knowledge Injection [7/10 impact] - Added knowledge-injector.js with semantic similarity scoring - Integrated into auto-prompts.js for execute-task prompts - queryKnowledge already exists in context-store.js (60% done) - Enhanced with: semantic matching, confidence filtering, contradiction detection - Tracks knowledge usage for feedback loop Integration: - Modified auto-prompts.js to inject knowledge via knowledgeInjection variable - Added getKnowledgeInjection helper for graceful degradation - All new modules pass build check and are in dist/ Status: Core infrastructure in place; ready for integration into dispatch loop. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
2 lines
236 B
Text
2 lines
236 B
Text
<!-- sf-doc: version=2.75.3 template=docs/references/uv-llms.txt state=pending hash=sha256:e8a998667c0f830a15b68e207f6b69e6377dd7e82728833f842678f72864e9b6 -->
|
|
Reference slot for uv/Python tooling guidance intended for LLM consumption.
|