Practical patterns, starters & CLI tools for loop engineering with AI coding agents. Design systems that prompt and orchestrate agents (inspired by Addy Osmani and Boris Cherny). Includes loop-audit, loop-init, loop-cost.
CI Sweeper — Hermes Agent (hermes cron)
This example maps the CI Sweeper pattern to Hermes's native scheduler. Week one is deliberately report-only: each run classifies failing checks, updates a state file, and exits without changing code or GitHub state.
Keep delivery local while you tune the loop. After its reports are consistently useful, the same cron job can deliver summaries to a connected channel.
Setup
Copy the shared state and ci-triage skill, then confirm Hermes can discover the
skill:
cp starters/ci-sweeper/ci-sweeper-state.md.example ci-sweeper-state.md
mkdir -p ~/.hermes/skills/ci-triage
cp starters/ci-sweeper/.claude/skills/ci-triage/SKILL.md \
~/.hermes/skills/ci-triage/SKILL.md
hermes skills list | grep ci-triage
The skill format is shared across tools. Use .hermes/skills/ci-triage/ instead
when the skill should be available only inside this project.
Week one — report only
Start with a 15-minute cadence on active repositories. The prompt has an early exit so green CI remains a cheap no-op.
hermes cron create "*/15 * * * *" \
--name "CI sweeper" \
--deliver local \
--skill ci-triage \
--workdir "$PWD" \
"Run the ci-triage skill. Read ci-sweeper-state.md, then inspect the latest CI checks on the default branch and active PRs. If every watched check is green, update Last run and stop. For each failure, record the check URL, job or test, first failing commit SHA, classification (regression, flake, infrastructure, config, or unknown), confidence, evidence, proposed human action, and attempt count. Move entries between Active Failures, Watch, and Resolved (last 7d). Week one is report-only: do not edit source or workflow files; do not create a worktree, branch, commit, or PR; do not retry CI; do not post GitHub comments or change checks; do not merge or enable auto-merge. Escalate ambiguous, security, release, secret, permission, and infrastructure failures. End with the top three human actions."
--workdir "$PWD"pins the repository and injects itsAGENTS.md,CLAUDE.md, or.cursorrulesinto the fresh cron session.--deliver localwrites to~/.hermes/cron/output/; it does not add routine reports to a human chat.--skill ci-triageattaches the shared classifier without granting a fixer permission to act.
Run one tick manually before relying on the schedule:
hermes cron list
hermes cron run <job-id>
hermes cron status
Pause the job while changing its scope: hermes cron pause <job-id>. Resume only
after reviewing the prompt and the latest state file.
State file
Keep enough evidence to distinguish a new regression from a known flake or infrastructure incident:
# CI Sweeper State
Last run: 2026-08-10 09:30 UTC
Mode: report-only
## Active Failures
### main @ abc1234 — unit-tests / parses empty input
- Check: https://github.com/example/project/actions/runs/123
- Classification: regression (medium confidence)
- Evidence: first fails at abc1234; reproduced twice without a retry
- Proposed action: human confirms scope and validation command
- Attempts: 0 / 3
## Watch (flakes / infra)
### release/2.x @ def5678 — integration / registry login
- Classification: infrastructure (high confidence)
- Evidence: registry returned 503 across unrelated jobs
- Proposed action: escalate; no retry or code change
## Resolved (last 7d)
- main @ 789abcd — lint failure fixed by PR #456
Do not put credentials or unredacted secrets from CI logs into this file.
Optional channel delivery
Keep local as the trust-building default. Once reports have been reviewed over
several runs, edit the job to use --deliver origin, a configured home channel
such as slack or feishu, or a specific platform:chat_id:thread_id target.
Set the channel's allowFrom and mention rules first; delivery changes where the
summary appears, not what the loop is allowed to do.
A CI webhook can replace polling when your Hermes gateway is already configured for trusted event delivery. Keep the same report-only prompt and the same state transition rules.
Week two — isolated fixes with an independent verifier
Enable fixes only after a human approves the exact failing check, allowed paths, and validation command. Each attempt follows this boundary:
- Create an isolated attempt with
npx @cobusgreyling/loop-worktree create --run-id <run-id> --pattern ci-sweeper. - Run
minimal-fixinside the printed worktree; never edit the primary checkout. - Run the approved test command in that worktree.
- Ask a separate verifier to review the diff and test output, either with a
review-only
delegate_taskcall or a chained Hermes cron job using--context-from. - Present an approved patch or draft PR to a human. Never merge or enable auto-merge from the loop.
- After three attempts on the same root cause, stop and escalate with the state history instead of retrying again.
The verifier must be independent of the implementer and must reject unrelated changes, denylisted paths, or a fix that only hides the symptom. Security, authentication, payments, secrets, permissions, release, and infrastructure failures always remain human-owned.