/sheal-drift — Drift Detection

Prerequisites

command -v sheal >/dev/null 2>&1 && sheal --version 2>/dev/null || echo "sheal not installed — run: npm install -g @liwala/sheal"

Step 0: Telemetry (silent, non-blocking)

aiframework-telemetry skill_invoked skill=sheal-drift 2>/dev/null || true

Step 1: Check Drift

sheal drift --last 10 --format json --project . 2>/dev/null || echo '{"error": "sheal drift failed"}'

Step 2: Analyze Patterns

For each drifted learning:

1. Check if it appears in CLAUDE.md invariants already
2. Check if a .claude/rules/ file covers it
3. Check frequency of drift (how often the same learning is violated)

Step 3: Promotion Recommendations

Apply these rules:

• Drift on same learning 3+ times → promote to CLAUDE.md invariant
• Drift in specific file area → create .claude/rules/<domain>.md
• No drift, healthy learning → keep as-is in sheal

Step 4: Apply Promotions (with approval)

Ask: "Found N learnings with persistent drift. Promote to permanent rules?"

For each approved promotion:

• Add to CLAUDE.md Invariants section via Edit tool
• Or create new .claude/rules/ file
• Mark the sheal learning as promoted

Step 5: Report