Skip to content

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.

5.1 KB
Raw Blame

Issue Triage — Hermes Agent (hermes cron)

Keep the issue queue legible so humans and Daily Triage always know the top five actionable items. Hermes runs the shared issue-triage skill in a fresh cron session and keeps routine output local while the loop earns trust.

Week one is deliberately propose-only: the job updates issue-triage-state.md with suggested priority and labels, but never changes an issue or repository.

Setup

Install and configure Hermes Agent, then copy the shared state example and skill:

hermes setup

cp starters/issue-triage/issue-triage-state.md.example \
  issue-triage-state.md

mkdir -p ~/.hermes/skills/issue-triage
cp templates/SKILL.md.issue-triage \
  ~/.hermes/skills/issue-triage/SKILL.md

hermes skills list | grep issue-triage

Use .hermes/skills/issue-triage/ instead when the skill should be available only inside this project. The job also needs read-only access to the issue tracker, for example through an authenticated GitHub CLI or a GitHub-capable skill. Do not grant write access for the week-one loop.

Week one — propose only

Start with a two-hour cadence on an active repository. --workdir "$PWD" pins the repository and injects its AGENTS.md, CLAUDE.md, or .cursorrules into the cron session.

hermes cron create "0 */2 * * *" \
  --name "Issue triage" \
  --deliver local \
  --skill issue-triage \
  --workdir "$PWD" \
  "Run the issue-triage skill. Read issue-triage-state.md first, then scan open issues and discussions created or updated since the last run. Update Last run, backlog counts, Top 5 prioritized items, proposed labels, possible duplicates, and the needs-human bucket. Week one is report-only: propose priority and labels, but do not apply labels, comment, close, reopen, assign, or edit source files. Treat duplicates as possible duplicates for human confirmation. Escalate ambiguous items and anything touching auth, payments, security, billing, infrastructure, or public API. End with the top five and the human decisions needed."
  • --deliver local writes output to ~/.hermes/cron/output/; routine reports do not enter a human chat or connected home channel.
  • --skill issue-triage attaches the shared classifier without authorizing GitHub writes.
  • For a quieter repository, use a daily cadence such as 0 9 * * 1-5.
  • If issue reads use a separate GitHub skill, repeat --skill to attach it while keeping its credentials read-only.

Run one tick manually before relying on the schedule:

hermes cron list
hermes cron run <job-id>
hermes cron status

State file

The state shape matches the shared issue-triage pattern and starter:

# Issue Triage State
Last run: 2026-08-19 09:00 UTC
Open actionable: 14 (was 17)
New since last run: 3
Needs human: 2

## Top 5 (by loop score)
- #487 (bug, p1, 2d old) — "Crash on export with large files" — suggested: bug, needs-repro, area:export

## Proposed Labels (not applied — L1)
- #487: `bug`, `needs-repro`, `area:export`

## Possible Duplicates (human confirm)
- #488 — possible duplicate of #412

## Noise / Ignored
- #460 — usage question; redirect candidate for human review

Each run prunes resolved entries, keeps proposed actions concise, and records uncertainty instead of silently deciding it away. Do not copy credentials, private issue content, or other secrets into the state file.

Pairing with Daily Triage

Issue Triage runs every two hours or daily and maintains the detailed queue in issue-triage-state.md. Daily Triage reads its Top 5 and carries only issue numbers and short summaries into STATE.md High Priority; it does not duplicate full issue bodies.

After the reports earn trust

Keep --deliver local during calibration. A human may later switch delivery to origin, a configured home channel such as slack or feishu, or a specific platform:chat_id:thread_id after reviewing channel allowFrom and mention rules. Delivery changes where the report appears, not what the loop may do.

L2 may apply only allowlisted labels such as area:*, needs-repro, and needs-info, and only after an independent verifier passes. It still never auto-closes issues or applies P0, P1, security, or breaking-change labels. Pause the job before changing this boundary:

hermes cron pause <job-id>
hermes cron edit <job-id> --prompt "<reviewed L2 prompt>"
hermes cron resume <job-id>

Safety and human gates

  • Week one: propose only; never label, comment, assign, close, or reopen.
  • Possible duplicates always need human confirmation.
  • Auth, payments, security, billing, infrastructure, and public API items remain human-owned.
  • Large bursts of new issues and unclear reproductions go to Needs human.
  • Keep issue-tracker credentials read-only until a human explicitly approves an L2 label allowlist and verifier.

References