Mine Session History

Extract actionable learnings from past Claude Code conversations and synthesize them into project documentation. Turns the goldmine of corrections, mandates, and discoveries across all sessions into persistent knowledge.

$ARGUMENTS

Phase 1: Extract Raw Corrections

Run the extraction script to scan all session transcripts for user corrections:

node scripts/extract-learnings.js \
  --project=. \
  --output=/tmp/learnings-raw.md \
  --since-days=90 \
  --max-pairs=500

Adjust --since-days and --max-pairs based on user arguments.

Read the output file to understand what was extracted.

Phase 2: Audit Existing Documentation

Before proposing any changes, read ALL existing documentation to understand what's already captured — CLAUDE.md, docs/*.md, .claude/rules/. Build a mental map of what's already documented.

Phase 3: Synthesize & Deduplicate

Work through the extracted corrections by topic:

Quality filter — only propose learnings that:

• Appear multiple times across sessions (pattern, not one-off)
• OR are clearly universal (e.g., "never use DateTime.UtcNow")
• Would prevent a real mistake if documented
• Are concise enough for a one-liner or short paragraph

Route by category:

Phase 4: Present Proposals

Present ALL proposed changes in a single summary before writing anything:

## History Mining Results

### Statistics
- Sessions scanned: X
- Corrections found: Y
- Already documented: Z
- New learnings: N

### Proposed Updates

#### 1. docs/FRONTEND.md
- **Add to Reactivity section**: "Never unwrap .value in addToast() calls — refs stay reactive"

#### 2. Skipped (already documented)
- "Use npm run scripts" — in CLAUDE.md line 12

Phase 5: Apply with Approval

Use AskUserQuestion to get approval before writing. Options: apply all, let me pick, save proposals to file, or skip.

After applying, show a summary table with file, change description, and line delta.

Gotchas

• Never add session-specific context — "We fixed bug #42" is not a learning
• Prefer updating over adding — extend existing bullets, don't create new sections
• One-liners only for CLAUDE.md — it has a 300-line budget
• Route domain knowledge to docs/*.md — that's where detailed patterns belong
• Don't duplicate linter rules — if the toolchain enforces it, don't document it
• Multiple occurrences = high confidence — a pattern seen across 3+ sessions is almost certainly worth documenting
• Single occurrence = verify first — read the surrounding context before proposing