flywheel（flywheel 搜索）

Flywheel Skill

Monitor the knowledge flywheel health.

The Flywheel Model

Sessions → Transcripts → Forge → Pool → Promote → Knowledge
     ↑                                               │
     └───────────────────────────────────────────────┘
                    Future sessions find it

Velocity = Rate of knowledge flowing through Friction = Bottlenecks slowing the flywheel

Execution Steps

Given /flywheel:

Step 1: Measure Knowledge Pools

# Count top-level artifact files (avoid counting directories)
LEARNINGS=$(find .agents/learnings -maxdepth 1 -type f 2>/dev/null | wc -l)
PATTERNS=$(find .agents/patterns -maxdepth 1 -type f 2>/dev/null | wc -l)
RESEARCH=$(find .agents/research -maxdepth 1 -type f 2>/dev/null | wc -l)
RETROS=$(find .agents/retros -maxdepth 1 -type f 2>/dev/null | wc -l)
echo "Learnings: $LEARNINGS"
echo "Patterns: $PATTERNS"
echo "Research: $RESEARCH"
echo "Retros: $RETROS"

Step 2: Check Recent Activity

# Recent learnings (last 7 days)
find .agents/learnings -maxdepth 1 -type f -mtime -7 2>/dev/null | wc -l
# Recent research
find .agents/research -maxdepth 1 -type f -mtime -7 2>/dev/null | wc -l

Step 3: Detect Staleness

# Old artifacts (> 30 days without modification)
find .agents/ -name "*.md" -mtime +30 2>/dev/null | wc -l

Step 3.5: Check Cache Health

if command -v ao &>/dev/null; then
  # Get citation report (cache metrics)
  CITE_REPORT=$(ao metrics cite-report --json --days 30 2>/dev/null)
  if [ -n "$CITE_REPORT" ]; then
    HIT_RATE=$(echo "$CITE_REPORT" | jq -r '.hit_rate // "unknown"')
    UNCITED=$(echo "$CITE_REPORT" | jq -r '(.uncited_learnings // []) | length')
    STALE_90D=$(echo "$CITE_REPORT" | jq -r '.staleness["90d"] // 0')
    echo "Cache hit rate: $HIT_RATE"
    echo "Uncited learnings: $UNCITED"
    echo "Stale (90d uncited): $STALE_90D"
  fi
else
  # ao-free fallback: compute approximate metrics from files
  echo "Cache health (ao-free fallback):"

  # Learnings modified in last 30 days (active pool)
  ACTIVE_30D=$(find .agents/learnings/ -name "*.md" -mtime -30 2>/dev/null | wc -l | tr -d ' ')
  echo "Active learnings (30d): $ACTIVE_30D"

  # Forge candidates awaiting promotion
  FORGE_PENDING=$(ls .agents/forge/*.md 2>/dev/null | wc -l | tr -d ' ')
  echo "Forge candidates pending: $FORGE_PENDING"

  # Citation tracking (if citations.jsonl exists)
  if [ -f .agents/ao/citations.jsonl ]; then
    CITATION_COUNT=$(wc -l < .agents/ao/citations.jsonl | tr -d ' ')
    UNIQUE_CITED=$(grep -o '"artifact_path":"[^"]*"' .agents/ao/citations.jsonl 2>/dev/null | sort -u | wc -l | tr -d ' ')
    echo "Total citations: $CITATION_COUNT"
    echo "Unique learnings cited: $UNIQUE_CITED"
  else
    echo "No citation data (citations.jsonl not found)"
  fi

  # Session outcomes (if outcomes.jsonl exists)
  if [ -f .agents/ao/outcomes.jsonl ]; then
    OUTCOME_COUNT=$(wc -l < .agents/ao/outcomes.jsonl | tr -d ' ')
    echo "Session outcomes recorded: $OUTCOME_COUNT"
  fi
fi

Step 4: Check ao CLI Status

if command -v ao &>/dev/null; then
  ao metrics flywheel status 2>/dev/null || echo "ao metrics flywheel status unavailable"
  ao status 2>/dev/null || echo "ao status unavailable"
  ao maturity --scan 2>/dev/null || echo "ao maturity unavailable"
  ao anti-patterns 2>/dev/null || echo "ao anti-patterns unavailable"
  ao badge 2>/dev/null || echo "ao badge unavailable"

  # Knowledge maintenance
  ao dedup --merge 2>/dev/null || true
  ao contradict 2>/dev/null || true
  ao constraint review 2>/dev/null || true
  ao curate status 2>/dev/null || true
  ao metrics health 2>/dev/null || true
  ao metrics cite-report --days 30 2>/dev/null || true

  # Active pruning: archive stale, evict low-utility, and curate noisy uncited learnings
  ao maturity --expire --archive 2>/dev/null || true
  ao maturity --evict --archive 2>/dev/null || true
  ao maturity --curate --archive 2>/dev/null || true

  # Retrieval quality: use the representative live corpus when it exists
  if [ -d cli/cmd/ao/testdata/retrieval-bench-live ]; then
    ao retrieval-bench --live --corpus cli/cmd/ao/testdata/retrieval-bench-live --json 2>/dev/null || true
  fi
else
  echo "ao CLI not available — using file-based metrics"

  # Pool inventory
  echo "Pool depths:"
  for pool in learnings patterns forge knowledge research retros; do
    COUNT=$(ls .agents/${pool}/*.md 2>/dev/null | wc -l | tr -d ' ')
    echo "  $pool: $COUNT"
  done

  # Global patterns
  GLOBAL_COUNT=$(ls ~/.claude/patterns/*.md 2>/dev/null | wc -l | tr -d ' ')
  echo "  global patterns: $GLOBAL_COUNT"

  # Check for promotion-ready learnings (see references/promotion-tiers.md)
  echo "See: skills/flywheel/references/promotion-tiers.md for tier definitions"
fi

Step 4.5: Process Metrics (from skill telemetry)

If .agents/ao/skill-telemetry.jsonl exists, use jq to extract: invocations by skill, average cycle time per skill, gate failure rates. Include in health report (Step 6) under ## Process Metrics.

Step 5: Validate Artifact Consistency

Cross-reference validation: scan knowledge artifacts for broken internal references. Use scripts/artifact-consistency.sh (method documented in references/artifact-consistency.md). Default allowlist lives at references/artifact-consistency-allowlist.txt; use --no-allowlist for a full raw audit.

Health indicator: >90% = Healthy, 70-90% = Warning, <70% = Critical.

Step 6: Write Health Report

Write to: .agents/flywheel-status.md

# Knowledge Flywheel Health

**Date:** YYYY-MM-DD

## Pool Depths
| Pool | Count | Recent (7d) |
|------|-------|-------------|
| Learnings | <count> | <count> |
| Patterns | <count> | <count> |
| Research | <count> | <count> |
| Retros | <count> | <count> |

## Velocity (Last 7 Days)
- Sessions with extractions: <count>
- New learnings: <count>
- New patterns: <count>

## Artifact Consistency
- References scanned: <count>
- Broken references: <count>
- Consistency score: <percentage>%
- Status: <Healthy/Warning/Critical>

## Cache Health
- Hit rate: <percentage>%
- Uncited learnings: <count>
- Stale (90d uncited): <count>
- Status: <Healthy/Warning/Critical>

## Retrieval Quality
- Live corpus coverage: <percentage or unavailable>
- Live corpus learnings: <count or unavailable>
- Status: <Healthy/Warning/Critical>

## Health Status
<Healthy/Warning/Critical>

## Friction Points
- <issue 1>
- <issue 2>

## Recommendations
1. <recommendation>
2. <recommendation>

Step 7: Report to User

Tell the user:

1. Overall flywheel health
2. Knowledge pool depths
3. Recent activity
4. Any friction points
5. Recommendations

Health Indicators

Golden Signals

Four golden signals (always shown) reveal whether knowledge is truly compounding or just accumulating noise.

ao flywheel status              # table output with golden signals
ao flywheel status --json       # machine-readable

The Four Signals

Verdicts and Thresholds

Overall verdict: 3+ healthy = compounding, 3+ critical = decaying, mixed = accumulating.

Recommended Actions

Cache Eviction

Read references/cache-eviction.md for the full eviction pipeline (passive tracking → confidence decay → maturity scan → archive).

Hub Budget & Phase 4 Hardening

Phase 4 (soc-ytpq) governance for ~/.agents/learnings/ — all advisory, none block by default. Full details in references/hub-budget.md.

• Size budget: target ≤ 250 MB / ≤ 5,000 files. Restore via ao maturity --evict --target-size=250M (lowest-utility-first; respects lifecycle.IsEvictionEligible so canonical / high-confidence files are protected).
• Volume gate: ao harvest WARNs when promotions exceed --max-promotions=N (default 500; AO_MAX_PROMOTIONS=N env as fallback; ≤0 disables). WARN-only — the 2,638-promotion soc-ujls drain proves a hard gate would falsely block legitimate runs.
• Provenance: every promoted file carries source_rig: in its frontmatter (empty writers serialize as source_rig: unknown). Both harvest.Promote and pool.(*Pool).Promote content-dedup against the same hub via ~/.agents/pool/promoted-index.jsonl.
• Re-bloat triage: check SkipGlobalHub defaults to true (the agentops-b3v / soc-ujls fix), then grep -h '^source_rig:' ~/.agents/learnings/*.md | sort | uniq -c | sort -rn to identify the regressed writer. Use --target-size, never raw rm.

Key Rules

• Monitor regularly - flywheel needs attention; address bottlenecks early
• Feed the flywheel - run /retro and /post-mortem
• Prune stale knowledge - archive old artifacts

Examples

User says: /flywheel — Counts pool depths, checks recent activity, validates artifact consistency, writes health report to .agents/flywheel-status.md. Hook trigger: After /post-mortem — Compares current vs historical metrics, flags velocity drops and friction points.

Troubleshooting

Reference Documents

• references/artifact-consistency.md
• references/promotion-tiers.md
• references/hub-budget.md
• references/cache-eviction.md