Which AI agents does this loop work with?

Triage a repo in five minutes is written for claude-code, codex.

Is this loop safe to run unattended?

This loop is graded C (65/100) by looprepo's automatic safety evaluator, which checks for exit conditions, iteration budgets, verification steps, and scoping before publishing.

/loop maintenancemedium riskintermediatesafety C · 65 Forward Futurepre-dates current gate · under review

Triage a repo in five minutes

Triage the repo, route bounded maintenance to dedicated threads, and require proof and permission before anything lands.

prompt

→ Claude

While repository maintenance is active, wake every five minutes. Triage [repositories] and read each repository thread's latest state. Reuse one thread per repository; assign its highest-value bounded task only within granted permissions, and do not interrupt coherent active work. Require tests, live proof, autoreview, and green CI before work can land. Escalate product, access, security, or irreversible decisions. Record meaningful changes and stop when every item is landed, decision-ready, blocked, or has no work.

claude-code · codex

Use this when

Use this when Codex may coordinate maintenance across several active repositories and you want parallel work to stay steerable without duplicating or micromanaging threads.

How it runs

Define the repository scope, exclusions, and separate permissions for triage, delegation, implementation, push, CI repair, merge, and release.
Every five minutes, refresh each repository queue and read the latest state of its existing thread before choosing the highest-value eligible item.
Reuse one thread per repository, assign one bounded task, and let coherent active work continue unless it is blocked, stalled, unsafe, or off course.
Require tests, live proof, autoreview, and green CI; record the evidence, then route the next item or present the owner with one exact decision.

Done when

✓ Every repository item reaches a proven handoff or terminal state. Authorized autonomous work lands with evidence; other items are decision-ready, blocked with one exact ask, or recorded as a clean no-op.

Why it works

A five-minute heartbeat keeps the control plane current without turning polling into micromanagement. One thread per repository preserves context, while proof and authorization gates make autonomous landing auditable.

Implementation note

The source pairs Maintainer Orchestrator with github-project-triage, autoreview, and computer use for live proof. A heartbeat automates observation, not authority: triage, delegation, implementation, push, merge, and release remain separate permissions. Read current thread state before steering, and never duplicate or interrupt active work.

Source: Forward Future ↗

More maintenance loops

Weekly tech debt report

/scheduleloopreponew

Every Friday, compile a trend report of debt signals — TODO count, lint suppressions, type coverage, largest files — so the team sees drift before it compounds.

prompt

→ Claude

/schedule every Friday at 4pm, measure TODO/FIXME count, eslint-disable and ts-ignore counts, type coverage, and the five largest source files; append the numbers with week-over-week deltas to reports/tech-debt.md and call out the single worst trend in one paragraph

maintenancelow risk

Keep memory pins under control

/schedulenew

Audit and prune pinned memory contexts to stay under 7, replacing competing invariants atomically and checking load count before each pin.

prompt

→ Claude

/goal that must load every session (it is then surfaced deterministically by load pinned ). Pin sparingly — keep a context at ≤7 pinned (prune at 10; the pinned load cap of 100 is a safety net, not the budget). Decisions/patterns/status are NOT pin material. Before pinning, call load pinned to check the count; when an invariant supersedes an old one, unpin the old in the same step ( update memory(memory id=<old , delivery mode="on recall") ) so two competing invariants are never both pinned

maintenancemedium risk

source ↗

Chase a refund until it lands

/goalForward Futurenew

Open the claim, watch replies and deadlines, and keep the case moving until the money actually arrives.

prompt

→ Claude

Get my refund for [company and charge info]. Start the claim now through an approved support channel, then keep following up on replies, promises, and deadlines until the refund arrives. Keep a short case note so each follow-up has context. Stop only when the refund is received or you are genuinely blocked and need me.

maintenancemedium risk