pattern-escalation

Pattern Escalation

TL;DR: Escalate recurring patterns up 3 tiers: lessons.md → MEMORY.md → CLAUDE.md.

Level 1 — tasks/lessons.md

Add/update: problem → root cause → fix → prevention rule. If variant of existing lesson: UPDATE that entry. Do not duplicate.

Level 2 — MEMORY.md (project root)

Add one-liner to relevant section. Format: **Pattern**: Rule. Ensures every future session-start reads it.

Level 3 — CLAUDE.md (architectural patterns only)

Criteria: affects schema, pipeline design, security, or data integrity. Flag to Will before committing. Do NOT auto-commit CLAUDE.md changes.

Memory curation (every ~10 sessions or when lessons.md > 40 entries)

Classify each lesson:

• Structural → promote one-liner to MEMORY.md, remove detail from lessons.md
• Active → keep
• Obsolete → delete Goal: lessons.md stays actionable, not a graveyard.

Skill accuracy check

Whenever a skill's advice is followed and produces a wrong result: update the skill immediately. Don't wait for pattern-escalation. Stale skills give confident wrong guidance. Pipeline numbers, DB patterns, and institutional checklist items are the most likely to drift.

Proposal-aware escalation

Evolution proposals live at .claude/cortex/proposals.local.md. Each proposal has:

• id (YYYYMMDD-HHMMSS-slug), status (pending/approved/rejected/expired), domain, occurrences, severity
• type: hook-rule | skill-update | claude-md-amendment
• probation=3: new hooks run in warn-only mode for 3 sessions before blocking

Escalation path

1. 1st occurrence — Lesson captured in tasks/lessons.md
2. 2nd occurrence — Existing lesson updated with new variant
3. 3rd+ occurrence — Conversation-analyzer generates a proposal with evidence
4. User review — Proposals surfaced at session start; reviewed via /analyze-session
5. Approved — Implementer applies the change, updates proposal status=applied with commit SHA
6. Rejected — User marks status=rejected with reason

Amplification path

When a lesson has been followed consistently (5+ sessions since written, 0 recurrences in that domain):

1. Skill reference — Add as "Known pitfall" in the relevant SKILL.md
2. Hook enforcement — If detectable by static analysis of tool input (file paths, code patterns, command strings), propose a new hook matcher
3. All new hooks include probation=3 — warn-only for 3 sessions, then blocking

Review protocol (at session start)

When session-start surfaces pending proposals:

1. Read each pending proposal
2. Assess: still relevant? evidence strong? risk acceptable?
3. Mark status=approved (implement now), status=rejected (with reason), or leave pending
4. Approved proposals get implemented in the current session if time allows

Staleness rule

Proposals pending for 10+ sessions without review → mark status=expired. The pattern may resurface later if it recurs, generating a fresh proposal with updated evidence.

---

See Also

• session-end — Session end triggers pattern escalation check for recurring issues [lifecycle]
• session-start — Session start surfaces pending escalation proposals for review [lifecycle]
• systematic-debugging — Debugging discovers patterns that feed into escalation [upstream]