beekeeper-lab/.claude/local/skills/health-check

Skill: Health Check

Description

Runs all health checks defined in ai/context/health-checks.md and produces a table-format report. Can be called standalone or by other skills (e.g., /long-run).

Trigger

• Invoked by the /health-check slash command.
• Called programmatically by /long-run at the start of each cycle.

Inputs

| Input | Type | Required | Description | |-------|------|----------|-------------| | health_checks | Markdown file | Yes | ai/context/health-checks.md — check definitions | | mode | String | No | interactive (default) or autonomous. In autonomous mode, FAIL-level checks create Unapproved beans. |

Process

Phase 1: Context Bloat Checks

1. CLAUDE.md size — Count lines in CLAUDE.md and .claude/shared/CLAUDE.md. Sum them. Threshold: 500 lines.
2. MEMORY.md size — Count lines in the project memory's MEMORY.md. Threshold: 200 lines.
3. Agent file sizes — Count lines in each .claude/agents/*.md. Threshold: 300 lines each.
4. Bean workflow size — Count lines in ai/context/bean-workflow.md. Threshold: 600 lines.

Phase 2: Telemetry Anomaly Checks

5. Parse recent Done beans — Read ai/beans/_index.md. For each Done bean, read its bean.md Telemetry summary section.
6. Low token check — Flag beans with Total Tokens In < 5,000.
7. High cost check — Flag beans with Total Cost > $15.
8. Missing telemetry — Flag Done beans where all telemetry summary fields are sentinel (—).
9. Zero duration anomaly — Flag Done beans with Duration = "< 1m" but >5 tasks.

Phase 3: Stale Artifact Checks

10. Old memory entries — Check modification dates of files in the project memory directory. Flag files >90 days old.
11. Stuck beans — Flag beans with status "In Progress" where Started date is >7 days ago.
12. Orphaned worktrees — Check for /tmp/foundry-worktree-* directories.
13. Stale status files — Check for /tmp/foundry-worker-*.status files.

Phase 4: Cost Trend Checks

14. Compute trailing averages — From the last 20 Done beans with valid cost data, compute:
    • Last 10-bean average cost
    • Previous 10-bean average cost
15. Increasing cost trend — Flag if current average is >50% higher than previous.
16. High average cost — Flag if current 10-bean average > $10.

Phase 5: Report

17. Render report — Print a table with columns: Check, Status (OK/WARN/FAIL), Detail.
18. Auto-bean creation (autonomous mode only) — For each FAIL-level check:
    • Create an Unapproved bean via /internal:new-bean
    • Problem statement = check detail
    • Notes include source: health-check
    • Do NOT auto-approve

Outputs

| Output | Type | Description | |--------|------|-------------| | report | Console text | Table-format health check report | | beans | Markdown files | (Autonomous mode only) Unapproved beans for FAIL-level issues |

Quality Criteria

• All checks produce a deterministic result (OK, WARN, or FAIL) with detail.
• Report is concise — one row per check.
• No false positives for newly created beans (exclude In Progress beans < 1 day old from stuck check).
• Autonomous bean creation is conservative (FAIL only, never auto-approve).

Error Conditions

| Error | Cause | Resolution | |-------|-------|------------| | NoBeanIndex | _index.md not found | Report error and skip telemetry/stale checks | | ParseError | Bean telemetry section malformed | Log warning, skip that bean, continue |

Dependencies

• ai/context/health-checks.md — check definitions and thresholds
• ai/beans/_index.md — bean index for telemetry and stale checks
• Individual bean files for telemetry data
• Project memory directory for stale memory checks