agency-health
SkillAI & modelsFleet health check across three dimensions, workstream, agent, worktree. Reports severity (healthy/attention/warning/critical) and sets exit codes for CI integration. Run this when you want to know "is the fleet OK right now?", at session start, before a release, or after any disruption.
Available today. Use it from your connected AI after setup.
No other account needed.
Connect ahel once, and every AI you use reads what you have installed.
Then ask your AI: use the agency-health skill
What this skill tells your AI
The instructions your AI receives, as published by the-agency-ai/the-agency in .claude/skills/agency-health/SKILL.md and read by ahel’s review.
Three-dimensional health check over the fleet.
What it checks
Workstream dimension
- Workstream directory presence
CLAUDE-{WORKSTREAM}.mdscoping file- Seed backlog size
- QGR recency
- General activity (last-modified across the workstream directory)
Agent dimension
- Agent registration in
.claude/agents/ - Expected worktree presence
- Unread dispatch count
- Handoff freshness
- Attribution integrity (recent commits not attributed to Test User)
Worktree dimension
- Dirty working tree (modified/staged counts)
- MERGE_HEAD / REBASE_HEAD residue (critical if present)
- Divergence from main branch
- Stash pile size
.agency-agentidentity file presence.claude/settings.jsonsync with main checkout
Severity levels
- healthy — no issues
- attention — informational, worth noticing
- warning — should be addressed
- critical — requires immediate attention
Exit codes: 0 = all healthy (or attention only), 1 = warnings, 2 = critical.
When to use
- Session start — confirm the fleet is in a known state before you start work
- Before cutting a release — catch stale/dirty/broken worktrees before they become PR problems
- After a major disruption — e.g., after a merge conflict storm or a failed update
- Periodically — catches slow-burn issues like stale handoffs, unprocessed dispatches, rotting seeds
- On the cron — machine-readable output via
--jsonmeans this can feed dashboards or alerts
Invocations
./agency/tools/agency-health # full fleet report
./agency/tools/agency-health all # same
./agency/tools/agency-health worktree # only worktree dimension
./agency/tools/agency-health agent devex # only devex agent
./agency/tools/agency-health workstream iscp # only iscp workstream
./agency/tools/agency-health --json all # JSON for downstream consumers
./agency/tools/agency-health --help # full help
Example output (human mode)
the-agency fleet health (2026-04-09 12:39)
─────────────────────────────────────────────────────────
WORKSTREAMS
● agency healthy active, artifacts present
● iscp healthy active, artifacts present
● mdpal attention 12 seeds, no activity 14d
AGENTS
● captain healthy registered, handoff fresh
⚠ devex warning 3 Test User commits in last 20
BATS pollution — run test_helper.bash audit
● iscp healthy registered, handoff fresh
WORKTREES
● (main) healthy clean on main
● devex attention 2 modified, 5 behind main
⚠ iscp warning 12 modified, 95 behind main, settings drift
─────────────────────────────────────────────────────────
OVERALL: 2 warnings, 0 critical
Not in v1
--repairmode — auto-fix known issues. Deferred to v2. For now, the tool reports; the operator fixes.- Cross-workstream dependency checks — e.g., "workstream iscp depends on workstream agency, and agency is unhealthy." Deferred to v2.
- Historical trending — how has fleet health changed over time? Would require persistent storage. Deferred.
Related
- Flag #51 — the original audit tool request that seeded this
- Flag #64 — diagnostic tooling workstream
- Companion tool:
agency verify(structural validation — this is behavioral/state validation)
OFFENDERS WILL BE FED TO THE — CUTE — ATTACK KITTENS!
Signals
- GitHub stars
- 55
- Forks
- 8
- Last commit
- Sep 2026
Advanced
- Catalog kind
- skill
- Gateway key
agency-health- Source
- github.com/the-agency-ai/the-agency