Context pruning: reduce context size using telemetry signal-to-noise data #9
Description
Description
Use telemetry data to learn which context sections are actually referenced during fixes, and prune unused sections from future captures of the same issue class.
Requirements
- Parse session transcripts for section references (element, consoleErrors, networkFailures, etc.)
- Score each section by usage frequency per issue class
- Build per-class context profiles (e.g., CSS fixes skip network/perf sections)
- Apply pruning rules in captureContext() based on issue classification
- Measure: context size should shrink while fix correctness stays constant
Target
- Signal-to-noise ratio >60%
- Context sufficiency (fixes needing extra Grep/Read) <20%
Depends On
- ADR-0006: DevTools Bridge integration #6 (DevTools Bridge integration) — Done
- Fix pattern library: store and reuse successful fix patterns #8 (Fix pattern library) — needs classification data
References
- PKM ADR-0013 Benchmarks: Self-Improvement Loop 1 (Context Pruning)