Loop ops

Loop Ops

There are two halves to loop engineering. loop-engineering writes the prompt a single run executes. loop-ops (this skill) wraps that prompt into a recurring, often unattended operation and makes it safe to leave running.

That power cuts both ways: a recurring loop amplifies judgment — good and bad. So this skill is mostly about restraint: start small, gate the risky stuff, cap the cost, and keep a human able to read what shipped.

Your job when this skill triggers: produce an operating plan for the loop — pattern, cadence, autonomy level, budget, safety gates, the schedule command, and the state/run-log files. It's the plan, not the running system: scaffold it and hand it over; only actually schedule/launch it if the user explicitly asks.

Respond and write the plan in the user's language; reasoning in English.

The building blocks of a recurring loop

A loop that runs unattended is assembled from six primitives (Cobus Greyling's "five building blocks + memory"):

1. Automation / scheduling — what fires it, and how often (cron, /loop 1d …, GitHub Actions, a watcher).
2. Worktrees / isolation — a safe, parallel place to work so a bad run can't corrupt the main branch or other loops.
3. Skills — the persistent project knowledge each run loads.
4. Plugins & connectors (MCP) — how it reaches your real tools (git, tickets, CI), and at what scope.
5. Sub-agents — the maker/checker split: an implementer does the work, a separate verifier checks it.
6. Memory / state — the durable spine (a STATE.md + a run-log) that lives outside any single run, so each run orients and the operator can audit.

For team/shared agents, bind tools, connector credentials, memory, state, and events to one tenant/channel/principal authority context. Channel visibility does not grant connector-data access, and another channel never inherits memory or a side-effect token.

The one rule that matters most: phased rollout (L1 → L2 → L3)

Never let a new loop act unattended on day one. Earn it in stages:

• L1 — report-only. It observes and proposes; it changes nothing. You read its output and judge whether you'd have done the same. Every loop starts here.
• L2 — assisted. It may make narrow, reversible changes behind safety gates (e.g. patch-only dependency bumps, draft PRs), still with a human gate on anything risky.
• L3 — unattended. It acts on its own for the allowlisted, low-risk slice — only after L1/L2 have shown it behaves.

State the level in the plan, and default new loops to L1. Graduating a level is a deliberate decision, not a default.

Workflow for an operating plan

1. Pick a pattern. Map the goal to a known recurring pattern (PR babysitter, daily triage, CI sweeper, dependency sweeper, changelog drafter, post-merge cleanup, issue triage). See references/recurring-patterns.md — each gives a cadence, a starting autonomy level, a cost band, and its specific safety gates.
2. Set the autonomy level — start L1 unless there's a reason not to.
3. Set the budget controller. Use an external scheduler/quota ledger to atomically reserve per-run/per-day capacity before dispatch and expensive increments, plus a persistent max-attempts cap. Token/$ values are hard caps only when the runtime meter can stop execution; otherwise report them as observed metrics.
4. Define the safety gates. A denylist (paths/actions it must never touch), a human gate for risky/irreversible/ambiguous actions, auto-merge rules, and MCP scopes. For L2/L3, map each rule to environment enforcement: credentials, sandbox/network policy, branch protection, scheduler pause, and active cancellation. See references/operating-safety.md.
5. Choose cadence + scheduler only after detecting what the user's environment actually supports. Never invent a /loop command or scheduler syntax; provide a platform-specific command only when verified, otherwise give a portable schedule specification. Host-specific background sessions are adapters, not CORE. On verified Claude Code W29+, /fork is a separately managed background session and /subtask stays in-session; never emit either from stale assumptions.
6. Scaffold control state and events. Use machine-readable control-state.json for budget, gate, cancellation, and claims; append structured events to run-events.jsonl; derive the glanceable Markdown status rather than treating prose as the source of truth. Templates and schemas live in assets/.
7. Get the per-run prompt from loop-engineering (the thing that actually runs each tick), and make sure its stop conditions, human gate, and budget match this plan.
8. Run the readiness check in references/operating-safety.md before anything goes past L1.

Output contract

Deliver a single operating plan using assets/ops-plan-template.md, filled in: pattern, cadence + schedule command, autonomy level (with why), per-run and per-day budget, safety gates (denylist / human gate / auto-merge / MCP scopes), the control-state/event-log scaffolding, enforcement owners, and a pointer to the per-run loop prompt. Add a 2–4 line note: which pattern, the key risks, and what must be true before promoting L1 → L2 → L3.

The quality bar

• Autonomy starts at L1 (report-only) unless the user explicitly overrides.
• There's an externally enforceable cost ceiling (per run and per day) and a kill switch that prevents launches and cancels active work.
• Risky / irreversible / ambiguous actions hit a durable, action-bound human gate with full context; the scheduler stays paused until a matching decision.
• A denylist names what the loop must never touch; MCP/tool scopes are minimal.
• Shared/team operation has an explicit authority context and a seeded cross-namespace memory/credential denial test.
• State + run-log are externalized so the operator can audit what shipped.
• If several loops touch the same repo, multi-loop coordination is addressed.
• A human can read the run-log in 30 seconds and know the loop's health.

Reference files

• references/recurring-patterns.md — the 7 named recurring patterns + their gates.
• references/operating-safety.md — rollout levels, human gate, cost/budget, denylist, auto-merge, MCP scopes, multi-loop coordination, failure modes, kill switch, and the operational-readiness checklist.
• assets/ops-plan-template.md — the fill-in operating plan (the deliverable).
• assets/control-state.example.json, assets/run-event.schema.json — control state and append-only audit contract.
• assets/STATE-template.md, assets/run-log-template.md — human-readable views, not enforcement state.
• scripts/quota_ledger.py — cross-platform, single-host atomic reference controller; distributed schedulers must provide equivalent reservation semantics.

Method note

The operating plan applies the suite's core principles: start with the simplest safe deployment, keep tools narrowly scoped, externalize state, verify outcomes, and increase autonomy only after observed evidence. For per-run prompt discipline, see loop-engineering.