review

Prompt file imported from ShadyBad/margin_invest (.claude/commands/review.md). Copyright stays with the author.

/review — Weekly Phase R

You are running Phase R of the GSD Workflow Engine.

Step 1: Inventory Runs

List all runs in runs/ directory:

ls -la runs/ | sort

Count:

• Total runs
• Completed (have final/done_gate.json with "status": "DONE")
• Failed / abandoned (no done_gate or "status": "NOT_DONE")

Step 2: Failure Analysis

For every NOT_DONE run, read runs/<id>/final/done_gate.json.

Aggregate: Which gates fail most often?

Write runs/review-<timestamp>/failure_patterns.md:

# Failure Patterns

| Gate | Failure Count | Runs |
|---|---|---|
| <gate_name> | N | <run_ids> |

## Root Causes
<analysis of why these gates are failing repeatedly>

## Systemic Fixes
<what to change in CLAUDE.md, hooks, or workflow to prevent recurrence>

Step 3: Golden Cases

For every failure pattern, create a golden test case.

Write to evals/golden_cases.jsonl (append):

{"id": "gc-<timestamp>-001", "input": "<what triggered the failure>", "expected_behavior": "<what should have happened>", "gate": "<which gate failed>", "source_run": "<run_id>"}

Step 4: Security Exception Review

Scan for open security findings:

grep -r "FINDING\|TODO\|FIXME\|HACK\|security" runs/*/security/ --include="*.md" --include="*.json" 2>/dev/null | head -50

For each open finding:

• Still relevant? → Keep, schedule fix
• Mitigated? → Close with evidence
• False positive? → Document and close

Write runs/review-<timestamp>/security_exceptions.md with status of each.

Step 5: Cost Creep Audit

Count agent spawns across recent runs:

grep '"event": "agent_spawn"' runs/*/audit/events.jsonl 2>/dev/null | wc -l

Flag any run that spawned >5 agents. Assess if they were necessary.

Identify: Which phases are producing the most artifact bloat?

Write runs/review-<timestamp>/cost_analysis.md:

# Cost Analysis

## Agent Spawn Count (last N runs): <total>
## Avg agents per run: <N>
## High-spawn runs: <list>

## Artifact Size
<du -sh runs/* output>

## Recommendations
<what to cut, what to optimize>

Step 6: CLAUDE.md Health Check

Review current CLAUDE.md. Ask:

1. Are the key project commands still accurate?
2. Are there new patterns or conventions to add?
3. Is the redaction list complete?
4. Are any sections outdated?

If updates needed: edit CLAUDE.md directly and note changes.

Step 7: Cleanup

Archive or delete abandoned runs older than 30 days:

# List candidates
find runs/ -maxdepth 1 -type d -mtime +30 | head -20

Do NOT delete runs with "status": "DONE" unless explicitly requested.

Step 8: Summary Output

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
WEEKLY REVIEW COMPLETE
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
Runs Analyzed:    <N>
  DONE:           <N>
  NOT_DONE:       <N>
  Abandoned:      <N>

Golden Cases Added: <N>
Security Exceptions: <N> open / <N> closed
CLAUDE.md Updates: <N> changes

Top Failure Gate: <gate_name> (<N> failures)
Top Fix: <what to do>

Report: runs/review-<timestamp>/
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━