knowledge-distillation

name: knowledge-distillation description: Extract and reuse knowledge from doit sessions. Automatically captures successful patterns, decisions, errors, and code patterns from completed sessions. Injects relevant past knowledge during Phase 1/2 to prevent repeating mistakes. Use when doit sessions complete (Phase 9.5.5) for extraction, or when Phase 1/2 starts for injection. Also use for /doit-learn to manually trigger knowledge extraction from historical data.

Knowledge Distillation

Capture successful patterns from doit sessions and reuse them to accelerate future development.

Overview

This module integrates with doit-skill's workflow to automatically:

1. Extract knowledge at Phase 9.5.5 (after completion summary)
2. Store in primary layers (mempalace + filesystem)
3. Inject relevant past knowledge at Phase 1/2 (before grill/planning)

Workflow Integration

Phase 9.5.5: Knowledge Extraction

Trigger: After Phase 9.5 Completion Summary, before Phase 10

Process:

1. Gather session data (git diff, worklog, spec, grill decisions)
2. Generate structured knowledge record (schema.json)
3. Present to user for confirmation/edit
4. Save to multi-layer storage
5. Report extraction status

Details: extract.md

Phase 1/2: Knowledge Injection

Trigger: Before Phase 1 grill questions and Phase 2 impact analysis

Process:

1. Generate search query from user request
2. Search multi-layer knowledge base
3. Rank by relevance + time decay
4. Inject top-3 relevant past sessions into context
5. Use injected knowledge to improve grill questions and planning

Details: inject.md

Historical Data Migration

Trigger: First run after installation, or manual /doit-learn command

Process:

1. Extract knowledge from git commit history
2. Extract from existing MemPalace drawers
3. Extract from worklog entries
4. Convert to structured format
5. Deduplicate and save

Details: migrate.md

Knowledge Record Schema

Structured JSON format for knowledge records:

{
  "id": "uuid-v4",
  "timestamp": "ISO-8601",
  "project": "project-name",
  "type": "F|B|S",
  "status": "success|failed|partial",
  "summary": "One-line description",
  "reqs": ["REQ-001: ..."],
  "decisions": [{"question": "...", "choice": "...", "rationale": "..."}],
  "errors": [{"error": "...", "root_cause": "...", "resolution": "..."}],
  "code_patterns": [{"pattern": "...", "files": ["..."], "context": "..."}],
  "files_changed": ["..."],
  "lines_added": 100,
  "lines_removed": 50,
  "confidence": 0.95,
  "tags": ["auth", "api", "database"]
}

Full schema: schema.json

Storage Layers

Configuration

Control via .doit/config.yaml:

knowledge:
  enabled: true              # Enable/disable
  auto_extract: true         # Auto-extract at Phase 9.5
  require_confirmation: true # Ask user before saving
  extract_failed: true       # Also extract failed sessions
  max_records: 1000          # Max records per project
  layers:                    # Which layers to use
    mempalace: true          # Primary
    filesystem: true         # Always

Graceful Degradation

Works with any combination of memory tools:

• Full stack: MemPalace + Filesystem → best search quality
• Partial: Any subset → still works, uses available layers
• None: Filesystem only → basic grep-based search

Manual Commands

# Extract knowledge from current session (manual trigger)
/doit-learn extract

# Inject knowledge for current request
/doit-learn inject "<request>"

# Migrate historical data
/doit-learn migrate

# View knowledge base statistics
/doit-learn stats

# Search knowledge base
/doit-learn search "<query>"

Quality Guidelines

Extract when:

• Meaningful code changes with clear rationale
• Key decisions with trade-offs documented
• Errors resolved with root cause analysis
• Code patterns created for reuse

Skip when:

• Trivial changes (typos, formatting)
• No meaningful decisions or patterns
• User explicitly requests skip
• Session failed without learning value

File Structure

learn/
├── SKILL.md          # This file — skill definition
├── extract.md        # Phase 9.5.5 extraction logic
├── inject.md         # Phase 1/2 injection logic
├── migrate.md        # Historical data migration
└── schema.json       # Knowledge record JSON schema

Integration Points

With doit-skill:

• Phase 9.5.5 hook (after completion summary)
• Phase 1 hook (before grill questions)
• Phase 2 hook (before impact analysis)

With memory tools:

• MemPalace: Primary semantic search + knowledge graph
• Tokensave: Code pattern references

With skill-creator:

• Eval framework for testing extraction/injection quality
• Iterative improvement based on user feedback