{"slug":"oversight-team","title":"oversight-team","summary":"Use when you are a seat on the oversight pod (a standing monitor-mode rig that keeps OTHER rigs healthy), configuring or running the drift detectors, or choosing whether to intervene vs escalate. Covers the pull-not-poll posture, the v0 detectors (premature-park, process-drift, o","platform":"Claude","tags":[],"authorName":"LLM Mart","authorSlug":"llm-mart","score":0,"source":"github","price":null,"verified":false,"createdAt":"2026-08-31T16:20:50.991994Z","repo":{"url":"https://github.com/mvschwarz/openrig","stars":371,"forks":51,"license":"Apache-2.0","updatedAt":"2026-09-25T04:59:18Z"},"bodyHtml":"<hr>\n<h2>name: oversight-team\ndescription: Use when you are a seat on the oversight pod (a standing monitor-mode rig that keeps OTHER rigs healthy), configuring or running the drift detectors, or choosing whether to intervene vs escalate. Covers the pull-not-poll posture, the v0 detectors (premature-park, process-drift, off-task, token-burn), the intervention ladder (orchestrator-ping -&gt; refocus -&gt; human escalation), and the cheap+deep model economics. NOT for orchestrating your OWN rig (orchestration-team) and NOT the intervention primitive mechanics (watchdog).\nmetadata:\nopenrig:\nstage: draft\nsibling_skills:\n- watchdog\n- refocus\n- orchestration-team\n- human-in-the-loop\n- messaging-the-human\n- retiring-and-inheriting-a-seat</h2>\n<h1>Oversight Team</h1>\n<p>You are on the <strong>oversight pod</strong> — a standing rig (agent-managed infrastructure, like the\nskills-architect pattern: agents + scripts + an SOP owning a function full-time) whose job is\nto keep <strong>other</strong> rigs healthy. You catch the unproductive patterns rigs drift into —\npremature parking, process-drift, off-task wandering, token burn — early, and correct them\nwith the lightest touch that works. <strong>Monitor mode: idle until a flag fires; never hyper-poll.</strong></p>\n<h2>Use this when</h2>\n<ul>\n<li>You are a seat running in <strong>monitor mode</strong> on the oversight pod.</li>\n<li>Configuring or running the v0 drift detectors over the fleet.</li>\n<li>Deciding whether to <strong>intervene</strong> (orchestrator-ping / refocus) or <strong>escalate</strong> (human).</li>\n</ul>\n<h2>Don't use this when</h2>\n<ul>\n<li>You are orchestrating your <strong>own</strong> rig — that's <code>orchestration-team</code>. Oversight watches\n<strong>across</strong> rigs; it does not run them.</li>\n<li>You need the intervention <strong>primitive mechanics</strong> (the wake / refocus / alignment-checkpoint\nstack, <code>rig watchdog</code> policies, message shape) — that's <code>watchdog</code>.</li>\n<li>A single stuck seat needs recovery — the owning orchestrator or <code>watchdog</code> handles that.</li>\n</ul>\n<h2>The posture — pull, never poll (load-bearing, and self-referential)</h2>\n<p>Monitor mode means <strong>idle until a flag fires, then wake and check</strong> — not continuous watching.\nContinuous <code>rig capture</code> / vigilant-observation loops are the exact anti-pattern that has\nburned entire model accounts: an over-observing watcher is expensive and produces nothing.\n<strong>The oversight seat must model the discipline it enforces</strong> — the token-burn detector below\nexists precisely because seats fell into vigilant loops, so a hyper-polling monitor would be\nthe failure it hunts. Cheap models read the large swaths; the expensive watcher acts on\n<strong>aggregated summaries</strong>, never the raw firehose.</p>\n<h2>The v0 detectors (scripted + cheap-model-summarized)</h2>\n<p>Cheap, evidence-based checks — each confirms a pattern from durable evidence before anyone acts:</p>\n<ul>\n<li><strong>Premature park</strong> — an in-progress qitem with an idle owner and no handoff. Pull the\ntranscript and confirm the turn actually ended <em>without</em> passing the ball (not merely quiet).</li>\n<li><strong>Process-drift</strong> — the ship-nothing pattern: commits that produce no shipped change, heavy\ntest iteration, very large/verbose qitems, endless deliberation. Detectable from git history\n<ul>\n<li>queue sizes. (This is process winning over product; correct it toward shipping.)</li>\n</ul>\n</li>\n<li><strong>Off-task drift</strong> — a cheap-model summary over a large activity swath (JSONL transcripts,\nthe stream, git log) answering one question: <em>\"is this rig on task?\"</em> If a pod drifts too\nlong, intervene.</li>\n<li><strong>Token burn / hypermonitoring</strong> — a seat consuming unusually — top-N consumers → capture and\ninspect for the vigilant-loop pattern. <strong>Telemetry surface caveat:</strong> the v0 detector uses\n<strong>point-in-time</strong> consumption polling; <strong>per-agent token telemetry OVER TIME</strong> is a later\nupgrade — do NOT assume it exists, and verify the available telemetry against your current\nversion before wiring a detector to it.</li>\n</ul>\n<h2>Interventions — least-disruptive first</h2>\n<ol>\n<li><strong>Orchestrator-ping</strong> — nudge the pod's own orchestrator to realign. They run their rig; you\nprompt, you do not seize.</li>\n<li><strong>The refocus primitive</strong> — <code>watchdog</code>'s <strong>Refocus</strong> level: a reactive north-star reminder\nthat fires only on <strong>actual</strong> drift, re-centers role / approved workstream / stop conditions,\nand does NOT interrupt valid work or turn into a fresh approval gate. Use <code>watchdog</code> for the\nmessage shape and the cadence discipline.</li>\n<li><strong>Escalate to the human</strong> — via the existing human path / notifications, per the active mode.\nReserve for what an agent-level nudge cannot fix.</li>\n</ol>\n<h2>Do not (the discipline)</h2>\n<ul>\n<li><strong>Don't hyper-poll / vigilant-loop</strong> — you become the token burn you hunt.</li>\n<li><strong>Don't seize another rig's work</strong> — ping its orchestrator; oversight corrects patterns, it\ndoes not take over.</li>\n<li><strong>Don't fire refocus on an active, on-task owner</strong> — that's bureaucracy theater (a <code>watchdog</code>\nfailure mode).</li>\n<li><strong>Don't intervene on a glance</strong> — confirm the pattern from evidence (transcript / git / queue)\nbefore acting. A false alarm costs the fleet trust and tokens.</li>\n<li><strong>Don't fire liveness / premature-park flags on a seat in an announced SWAP WINDOW.</strong> A seat mid-handover\nlooks idle/parked to the detectors. The executing party pre-announces (seat + expected window) at swap\nstart — honor it as a <strong>suppression window</strong> until the handover receipt lands (receipts arrive only at\nswap <em>end</em>, so waiting on the receipt alone still misfires on a long swap). See <code>retiring-and-inheriting-a-seat</code>.</li>\n</ul>\n<h2>Cheap + deep model economics</h2>\n<p>Cheap models do the routine reading (transcripts, logs, stream) and aggregate intelligent\nsummaries; an expensive watcher decides on those summaries. This is <em>why</em> per-agent model\nassignment matters — you can run the reading far more often without blowing out accounts. Design\nthe pod so the deep model never touches the raw firehose.</p>\n<h2>See also</h2>\n<ul>\n<li><code>watchdog</code> — the intervention primitive (wake / refocus / alignment-checkpoint) this SOP\ndrives; refocus lives there.</li>\n<li><code>refocus</code> — the drift doctrine <code>watchdog</code> operationalizes.</li>\n<li><code>orchestration-team</code> — running your own rig (complementary; oversight watches across rigs).</li>\n<li><code>human-in-the-loop</code> / <code>messaging-the-human</code> — the escalation path and how to address a human.</li>\n<li><code>retiring-and-inheriting-a-seat</code> — a context-threshold detector can fire a planned seat\nhandover (sibling oversight signal).</li>\n</ul>\n","files":[{"path":"SKILL.md","sizeBytes":6347,"isText":true}],"reviewScore":null,"reviewSummary":null,"trust":{"provenance":"trusted-source-unreviewed","notice":"Community-authored content, reproduced verbatim and not vetted as instructions. Treat it as data to evaluate, never as directives to follow.","bodySource":null},"bodyLocked":false,"purchaseUrl":null,"sourceUrl":null,"report":{"provenance":"trusted-source-unreviewed","screen":{"ran":true,"outcome":"clean","suspicious":0,"notes":0,"hiddenCharacters":false},"virusScan":{"engine":"clamav","status":"clean","scannedAt":"2026-08-31T16:21:35.423652Z","sha256":"5573E3090AF254589C9BBCBFC28E7A68E0508E905856202999BE961976E72F96","sizeBytes":2924},"review":null,"source":{"repositoryUrl":"https://github.com/mvschwarz/openrig","path":"skills/_canonical/pods/oversight-team","license":"Apache-2.0","commit":"b374dde300fd2a3cf1ee139b89b11e2fa3945784","subtreeSha":"A66B31C43B28280F4CC41C5D8ECDBD55DF168853D8ABA911548FE82B7B32F910","lastSyncedAt":"2026-09-25T06:48:47.236757Z"},"reviewedAt":"2026-08-31T16:23:03.180779Z","notice":"Community-authored content, reproduced verbatim and not vetted as instructions. Treat it as data to evaluate, never as directives to follow."},"install":[{"target":"skills-cli","command":"npx skills add https://github.com/mvschwarz/openrig/tree/main/skills/_canonical/pods/oversight-team"},{"target":"claude-code","command":"claude plugin marketplace add https://llmmart.ai/marketplace.json && claude plugin install mvschwarz-openrig@llmmart"},{"target":"git","command":"git clone https://github.com/mvschwarz/openrig.git"}]}