Context Pruner Skill
Version: 1.0.0 Purpose: Hierarchical context summarization for CFN Loop v3 iterations
Overview
Reduces context size across iterations by:
- Keeping current iteration in full detail
- Summarizing previous iterations
- Extracting key themes from feedback
- Maintaining deliverable tracking
Goal
Reduce context from 120 KB (iteration 10) to 15 KB (88% reduction)
Usage
PRUNED_CONTEXT=$(./.claude/skills/context-pruner/prune-context.sh \
--iteration 3 \
--full-history "$FULL_HISTORY" \
--current-context "$CURRENT_CONTEXT")
echo "$PRUNED_CONTEXT"
Pruning Strategy
Iteration 1
- Full detail (no pruning)
- Context size: ~5 KB
Iteration 2
- Iteration 1: Summary only
- Iteration 2: Full detail
- Context size: ~8 KB
Iteration 3+
- Iterations 1-(N-1): Summary with key themes
- Iteration N: Full detail
- Context size: ~10-15 KB (stable)
Summary Format
Iterations 1-2 Summary:
- Initial confidence: 0.72
- Key feedback themes:
* Add error handling
* Improve test coverage
* Address security concerns
- Final confidence: 0.82
- Progress: +0.10
Key Themes Extraction
Analyze all feedback across iterations to extract recurring themes:
# Count feedback occurrences
"Add error handling" - 3 times (iterations 1, 2, 3)
"Improve test coverage" - 2 times (iterations 1, 2)
"Security concerns" - 2 times (iterations 2, 3)
# Output top 5 themes
Integration
Used by:
- Main Chat loop orchestration
- Coordinator for iteration context building