openclaw-fleet-status

Check fleet operational status — tasks (pending / in-progress / completed today), agents (registered, running, idle, stuck), board state (Plane / Mission Control), recent activity, blockers. Read-only

SKILL.md

GitHub

openclaw-fleet-status

The read-only operational checkup skill for an OpenClaw fleet deployment. Produces a single concise report that answers: how many tasks are in flight, which agents are healthy, what's blocked, what's been done today. Distinct from openclaw-health (deeper system audit) and openclaw-fleet-status MCP tool (programmatic — this is the human-facing variant).

Trigger phrases (when to load this skill)

Load when the conversation matches any of:

  • Direct verb: operator says "fleet status", "what's the fleet doing", "are agents running", "anything blocked", "show me the board", "how busy are we".
  • Morning/EOD check-in: operator opens a session and asks for an orientation snapshot before deciding what to work on.
  • Pre-deploy gate: about to ship a change that affects fleet behavior; want to know current state before touching anything.

Do NOT load when:

  • Deeper diagnostic needed (system health, hooks, integrations) — load openclaw-health.
  • Operator wants to add a new agent — load openclaw-add-agent.
  • Operator wants Mission Control re-connected — load openclaw-configure-mc.
  • The status request is for a specific task ID — use the task-tracking skill (aicp-ops-tasks) or read the task file directly.

Operations

This skill has 2 named operations. Execute in order. The first is the bulk of the work; the second is presentation.

Operation 1: Gather signals from each fleet surface

Trigger: skill loaded.

Process:

  1. Fleet daemon / Make targets — from the fleet project root (openfleet/ or operator-named):
    make status   # if defined; otherwise:
    openclaw agents list           # registered agents + their workspaces
    openclaw agents status --all   # running / idle / stuck per agent
    
    Capture: agent count, per-agent state, last heartbeat timestamp.
  2. Mission Control board — read board file or hit MC API:
    cat openfleet/board.json 2>/dev/null    # if file-based
    curl -s http://localhost:<mc-port>/api/board   # if MC API exposed
    
    Capture: tasks per column (todo / in-progress / done-today), oldest in-progress, recent activity timestamps.
  3. Plane integration — if Plane is wired (per docs/architecture/fleet-integration.md):
    curl -s -H "Authorization: Bearer $PLANE_API_KEY" \
      https://plane.<host>/api/v1/workspaces/<ws>/issues/?state=in_progress
    
    Capture: linked Plane issues open under each agent, age of the oldest.
  4. AICP backend health — fleet agents call AICP for inference; if AICP is down, fleet is stuck:
    .venv/bin/aicp --check 2>&1 | head -20
    
    Capture: backend availability summary (the failover-chain line is the most informative).
  5. Recent work — past 24h activity:
    find openfleet/agents/*/logs -name "*.log" -mtime -1 | head
    
    Capture: log freshness per agent. Stale logs (>2h on a "running" agent) is a stuck-agent signal.
  6. Identify blockers: any task in_progress for >2h with no log activity, any agent registered but not running, any Plane issue waiting on operator review.

Quality bar (Operation 1 done when):

  • All 6 signal sources queried (or explicitly skipped with reason: "Plane not wired", etc.).
  • Per-agent state captured (registered / running / idle / stuck).
  • Board state captured (count per column).
  • AICP backend health captured.
  • Blocker list assembled with evidence for each (timestamp, task ID, log path).

Operation 2: Format the status report

Trigger: Operation 1 signals gathered.

Process:

Produce a single concise report in this shape:

FLEET STATUS — <timestamp>

Agents (<N> registered):
  ● <name>       running   last activity 12 min ago    on T<NNN>
  ● <name>       idle      last activity  2 hr ago     queue empty
  ⚠ <name>       stuck     last activity  4 hr ago     on T<NNN> — no log progress

Board (Mission Control / Plane):
  todo:        <N> tasks  (oldest <age>)
  in-progress: <N> tasks  (oldest <age>)
  done-today:  <N> tasks

AICP backend:
  failover chain: local → k2_6_openrouter → openrouter → claude
  health:        <X/Y backends OK>

Recent (24h): <N> tasks completed across the fleet.

Blockers:
  ⚠ T<NNN> — in-progress 4h on <agent>, no log activity. Likely stuck.
  ⚠ T<NNN> — Plane issue waiting on operator review since <timestamp>.
  (none) if no blockers.

Suggested next: <single-line operator action OR "all clear">

Notes on the report:

  • ≤30 lines total. Operator should grok the picture in <15s.
  • Use ●/⚠ symbols for visual scan; avoid heavy formatting.
  • Times are RELATIVE ("12 min ago"), not absolute timestamps — easier to read at a glance.
  • Suggested next action is concrete and singular ("kick T042 — stuck for 4h" not "review your blockers").

Quality bar (Operation 2 done when):

  • Report ≤30 lines.
  • Per-agent line includes state + last activity + current task (or "queue empty").
  • Board breakdown by column with oldest age.
  • AICP backend health line present.
  • Blockers section either lists each with evidence OR states "(none)".
  • Suggested-next-action is specific.

Gotchas (known failure modes — read before doing)

Gotcha 1: Counting registered agents as "running"

openclaw agents list returns 10 agents. Report says "10 agents running". But 7 of them are registered-and-stopped (the orchestrator never started them, or they crashed). Operator thinks fleet is healthy; actually 70% of capacity is dark.

The rule: state always comes from openclaw agents status, not from agents list. Distinguish "registered" from "running" in the report. Stuck agents (running but no log progress >2h) get the ⚠ marker — they look running but aren't doing work.

Gotcha 2: Stale log = OK

An agent's last log entry is 4 hours old. Report flags it as "stuck". But the agent is correctly idle — its queue is empty, nothing to log about. False alarm wastes operator attention.

The rule: cross-reference last log activity with current task assignment. Log silence + assigned-task = stuck (real flag). Log silence + no assigned task = idle (correct, no flag). The Operation 1 step 5 must capture both.

Gotcha 3: Reporting raw JSON

Pasting openclaw agents list --json output verbatim. Operator gets 200 lines of JSON, has to parse it themselves. Defeats the point of a status report.

The rule: Operation 2 produces a HUMAN-READABLE summary. Capture JSON for evidence in Operation 1; render to ≤30-line prose in Operation 2. If operator wants raw JSON, they ask for it — don't dump it preemptively.

Gotcha 4: Plane / MC URL wrong, silent failure

Operator's deployment uses Plane on a different host or MC on a different port than the skill assumed. The curl returns 404 / connection refused. Skill silently treats "no tasks" as "no tasks", reports an empty board.

The rule: every external call has explicit failure handling. If the curl fails, the report says "Plane unreachable at <url> — board state unknown" — not silent zero. Operator must see the gap, not be misled by it.

Gotcha 5: "No blockers" when AICP is degraded

Backend chain is local → k2_6_openrouter → openrouter → claude; LocalAI is down. Routes still work (failover to k2_6_openrouter), but every fleet task pays cloud-token cost it shouldn't. Report says "no blockers" because tasks are completing.

The rule: AICP backend degradation IS a blocker even if work proceeds. Report the failover state (which backend is currently the primary route) and flag if it's not the expected primary. Operator decides whether to surge.

Reference exemplars

The Extension Standards reference exemplars for skills are the second brain's model-builder and wiki-agent skills — see ~/devops-solutions-research-wiki/skills/. This skill follows their structure: trigger phrases, multiple named operations, per-operation Process + Quality bar, Gotchas with detection + rule + reasoning.

Domain context

This skill spans AICP and OpenFleet projects. AICP provides backend health visibility (aicp --check); OpenFleet runs the fleet that this skill audits. The report is the operator's morning glance + on-demand checkup tool.

Related skills

Skill When Why distinct
openclaw-health Deeper system audit (hooks, integrations, daemons) This skill is operational snapshot; openclaw-health is system audit
openclaw-add-agent Add a new agent This skill reports current state; add-agent changes it
openclaw-setup Bootstrap OpenClaw This skill reports an existing deployment; setup creates one
aicp-ops-runtime Audit AICP itself in detail This skill summarizes AICP backend in 1 line; ops-runtime drills in
pm-status-report Multi-day project status with metrics This skill is now-state snapshot; pm-status-report is reporting cadence