| name | compounding |
| description | Capture post-feature learnings from Phase 7 by running 4 parallel analysis agents and writing structured learning files |
| triggers | compounding, capture learnings, phase 7 learnings, post-feature review |
Compounding runs after all links in a feature are closed (Phase 7). It spawns 4 specialist
analysis agents in parallel to extract maximum signal from the completed work — patterns,
decisions, failures, and outcome gaps — and writes durable learning files that make future
features faster.
<Use_When>
- Phase 6 review is complete and all links are marked closed
- Master Orchestrator reaches Phase 7
- User explicitly requests post-feature learnings capture
</Use_When>
<Do_Not_Use_When>
- Links are still open or in_progress
- Phase 6 review has not completed
- The feature was trivial (fewer than 3 links) — use Learner instead
</Do_Not_Use_When>
Step 1 — Verify Readiness
Before spawning agents, confirm:
- All task issues for this session are closed
- Phase 6 reviewer output exists in
.oh-my-link/history/
- A session slug is available (from session.json or plan.md frontmatter)
If not ready, output a clear message and stop.
Step 2 — Spawn 4 Parallel Analysis Agents
Spawn all 4 simultaneously. Each receives:
- The plan.md
- The Phase 6 reviewer report
- The list of closed links with their implementation notes
- The WRAP-UP.md if available
Agent 1 — Pattern Extractor
Task: Identify reusable code, architecture, and process patterns.
Output format:
## Patterns
- {pattern name}: {description} | Reuse signal: HIGH/MEDIUM/LOW
Agent 2 — Decision Analyst
Task: Evaluate key decisions — what worked, what didn't, what surprised.
Output format:
## Decisions
- {decision}: {outcome} | Verdict: GOOD/BAD/NEUTRAL | Why
Agent 3 — Failure Analyst
Task: Catalog failures, blockers, and wasted effort. Produce prevention rules.
Output format:
## Failures
- {failure}: {cause} | Prevention: {rule}
Agent 4 — Exit-State Auditor
Task: Compare planned vs actual outcomes. Identify scope drift and estimation errors.
Output format:
## Planned vs Actual
- {item}: Planned {X} | Actual {Y} | Delta: {note}
Step 3 — Synthesize Outputs
Collect all 4 agent outputs. Write to .oh-my-link/history/learnings/YYYYMMDD-{slug}.md:
# Learnings: {feature name}
Date: {YYYYMMDD}
Links: {count}
## Patterns
{from Pattern Extractor}
## Decisions
{from Decision Analyst}
## Failures and Prevention
{from Failure Analyst}
## Planned vs Actual
{from Exit-State Auditor}
## Promoted to Critical
{list any HIGH-signal patterns promoted in next step}
Step 4 — Promote Critical Findings
For any finding marked HIGH signal by any agent:
- Append a summary entry to
.oh-my-link/history/learnings/critical-patterns.md
- Format:
{date} | {slug} | {one-line summary}
- Create the file if it does not exist
Step 5 — Report
Output:
- Path to the learnings file written
- Count of patterns, decisions, failures, and deltas captured
- List of items promoted to critical-patterns.md
<Tool_Usage>
- Read: plan.md, reviewer report, WRAP-UP.md, session artifacts; verify link closure before starting
- Write: learnings file and critical-patterns.md
- Agent spawning: 4 parallel specialist subagents (Step 2)
</Tool_Usage>