compound

name	Compound
description	Reflect on the cycle and capture high-quality lessons for future sessions

Compound Skill

Overview

Extract and store lessons learned during the cycle, and update project documentation. This is what makes the system compound -- each session leaves the next one better equipped.

CRITICAL: Store all lessons via npx ca learn -- NOT via MEMORY.md, NOT via markdown files. Lessons go to .claude/lessons/index.jsonl through the CLI. MEMORY.md is a different system and MUST NOT be used for compounding.

Methodology

1. Review what happened during this cycle (git diff, test results, plan context)
2. Detect spec drift: compare final implementation against original EARS requirements in the epic description (bd show <epic>). Note any divergences -- what changed, why, was it justified. If drift reveals a spec was wrong or incomplete, flag that for lesson extraction.
3. Spawn the analysis pipeline in an AgentTeam (TeamCreate + Task with team_name):
   • Role skills: .claude/skills/compound/agents/{context-analyzer,lesson-extractor,pattern-matcher,solution-writer,compounding}/SKILL.md
   • For large diffs, deploy MULTIPLE context-analyzers and lesson-extractors
   • Pipeline: context-analyzers -> lesson-extractors -> pattern-matcher + solution-writer -> compounding
   • Agents coordinate via SendMessage throughout the pipeline
4. Agents pass results through the pipeline via SendMessage. The lead coordinates: context-analyzer and lesson-extractor feed pattern-matcher and solution-writer, which feed compounding.
5. Apply quality filters: novelty check (>0.98 cosine similarity = skip), specificity check
6. Classify each item by type: lesson, solution, pattern, or preference
7. Classify severity: high (data loss/security/contradictions), medium (workflow/patterns), low (style/optimizations)
8. Store via npx ca learn with supersedes/related links where applicable. At minimum, capture 1 lesson per significant decision made during this cycle
9. Delegate to the compounding subagent to run synthesis: cluster accumulated lessons by similarity and write CCT patterns to .claude/lessons/cct-patterns.jsonl
10. Update outdated docs and deprecate superseded ADRs (set status to deprecated)
11. Use AskUserQuestion to confirm high-severity items with the user before storing; medium/low items are auto-stored

Docs Integration

• docs-reviewer checks if docs/ content is outdated after the cycle
• Check docs/decisions/ for ADRs contradicted by the work done
• Set ADR status to deprecated if a decision was reversed, referencing the new ADR

Literature

• Consult docs/compound/research/learning-systems/ for knowledge compounding theory, spaced repetition, and lesson extraction methodology
• Run npx ca knowledge "knowledge compounding" for indexed knowledge on learning systems
• Run npx ca search "compound" for lessons from past compounding cycles

Common Pitfalls

• Not spawning the analysis team (analyzing solo misses cross-cutting patterns)
• Capturing without checking for duplicates via npx ca search
• Skipping supersedes/related linking when an item updates prior knowledge
• Not checking if docs or ADRs need updating after the cycle
• Requiring user confirmation for every item (only high-severity needs it)
• Not classifying items by type (lesson/solution/pattern/preference)
• Capturing vague lessons ("be careful with X") -- be specific and concrete

Quality Criteria

• Analysis team was spawned and agents coordinated via pipeline
• Quality filters applied (novelty + specificity)
• Duplicates checked via npx ca search before capture
• Items classified by type (lesson/solution/pattern/preference)
• Supersedes/related links set where applicable
• Outdated docs and ADRs were updated or deprecated
• User confirmed high-severity items
• Beads checked for related issues (bd)
• Each item gives clear, concrete guidance for future sessions
• Spec drift analyzed and captured

FINAL GATE -- EPIC CLOSURE

Before closing the epic:

• Run npx ca verify-gates <epic-id> -- must return PASS for both gates
• Run pnpm test and pnpm lint -- must pass If verify-gates fails, the missing phase was SKIPPED. Go back and complete it. CRITICAL: 3/5 phases is NOT success. All 5 phases are required.