Autonomous Loops — Patterns for Agent Automation

⚠️ Security & Consent — READ FIRST

This skill teaches unattended repository automation. It can create branches, open pull requests, retry CI failures, and merge code without a human in the loop. That power is dangerous in the wrong hands and on the wrong target.

Before using any pattern in this skill, you MUST:

1. Get explicit, informed consent from the repository owner. "I want to automate PRs" is not consent for "merge my production code at 3 AM without review."
2. Default to dry-run. All patterns below start in --disable-commits / no-write mode. Only flip the switch when you understand what each flag does.
3. Use protected branches + least-privileged credentials. The GitHub token this loop uses must not have admin/org-owner scopes. A scoped fine-grained PAT scoped to one repo is the minimum.
4. Set hard limits. Every loop must have --max-runs, --max-cost, and --max-duration. Without these, a runaway loop will burn budget or destroy state.
5. Never loop on production. Run on a fork, a scratch repo, or a clearly-labeled branch. The main branch should be merged by a human.
6. Audit your inputs. Specs, directory listings, and captured diffs may contain secrets, customer data, internal paths, or proprietary code. Never pipe those into an external agent command without scrubbing them first.

Patterns in this skill that touch the network, filesystem, or git remotes will refuse to run unless the corresponding safety flags are explicitly set. If a pattern silently merges or pushes, that is a bug — file an issue.

See references/security-checklist.md for the full audit before first use.