How to use the Oversight Team Monitor agent skill

What the oversight-team skill makes the agent do

The skill tells the agent to stay idle until a warning flag fires, then wake and check evidence. It does not watch continuously. The evidence can include transcripts, git history and queue sizes, and the agent uses it to confirm a pattern before acting.

When a pattern is confirmed, the skill asks for the lightest correction that works. The intervention ladder runs in order: first an orchestrator-ping to the other rig's own orchestrator, then a refocus reminder, and only then escalation to a human.

• Pull, not poll: check only when a flag fires, never in a hyper-polling loop.
• v0 detectors: premature-park, process-drift, off-task, token-burn.
• Intervention ladder: orchestrator-ping, then refocus, then human escalation.
• Cheap models summarize large logs; the expensive model decides on the summaries.

Install the skill from the ZIP

Download the ZIP for the oversight-team skill from this page and unzip it. The skill is a single file named SKILL.md. For a personal setup, place the unzipped folder in ~/.claude/skills/. For a single project, place it in that project's .claude/skills/ folder.

If you use the Claude apps instead of files on disk, you can upload the skill under Settings → Capabilities → Skills.

• Personal install: ~/.claude/skills/
• Project install: the project's .claude/skills/ folder
• Claude apps: Settings → Capabilities → Skills, then upload

Trigger the skill with natural requests

The skill loads when your request matches its description. You do not need a special command. Ask in plain language about the oversight role, the detectors, or the choice between intervening and escalating.

The description covers the oversight pod, configuring or running the drift detectors, and deciding whether to intervene or escalate. It also says not to use it for orchestrating your own rig, which belongs to the orchestration-team skill.

• "I am a seat on the oversight pod. Check whether the rig on pod two shows premature parking."
• "Configure the v0 drift detectors for the fleet and tell me which signals to use."
• "Should I ping this rig's orchestrator or escalate to the human? Here is the evidence."
• "Explain the intervention ladder for an off-task rig."

Understand the four v0 detectors

Each detector checks a pattern against durable evidence before anyone acts. Premature park means an in-progress item has an idle owner and no handoff, and the agent must confirm the turn ended without passing the work on. Process drift is commits that ship nothing, heavy test iteration, very large queue items or endless deliberation.

Off-task drift uses a cheap-model summary of activity to answer whether the rig is on task. Token burn looks at top consumers and inspects them for a vigilant-loop pattern. The source notes that v0 uses point-in-time consumption polling, and that per-agent token telemetry over time is a later upgrade that you should not assume exists.

• Premature park: idle owner, in-progress item, no handoff.
• Process drift: commits with no shipped change, heavy test churn, oversized items.
• Off-task: cheap summary answers whether the rig is on task.
• Token burn: top consumers checked for vigilant loops.

Choose the right intervention

Start with the least disruptive step. An orchestrator-ping asks the other rig's own orchestrator to realign, and you prompt rather than take control. A refocus is a north-star reminder that fires only on actual drift. It is a watchdog primitive, so use the watchdog skill for its message format and timing.

Escalate to a human only when an agent-level nudge cannot fix the problem, using the existing human path and notifications.

• Do not seize another rig's work; ping its orchestrator.
• Do not refocus an active, on-task owner.
• Confirm evidence before acting, because false alarms cost trust and tokens.
• Respect swap windows: suppress premature-park and liveness flags on a seat announced as mid-handover until the handover receipt lands.

Set up cheap and deep models

The skill describes a cost pattern. Cheap models read transcripts, logs and streams and produce aggregated summaries. An expensive watcher decides only on those summaries. This is why per-agent model assignment matters: the reading can run often without draining accounts.

Design the pod so the expensive model never reads the raw firehose of activity.