← All playbooks
InfrastructureWe run this
Update your agent framework without breaking production
Updates are merged and tested in a separate copy first. Production only moves when you say go.
Why we run it
Our Hermes install carries local patches. A scheduled job prepares each upstream update in a separate worktree, runs the tests, and asks before touching the live install.
Steps
- Keep your local changes on their own branch, not as loose edits in the live folder.
- Every two hours, fetch upstream and merge it into a separate worktree.
- Run the tests that cover your local changes. Report pass, skip and fail counts.
- Check whether dependencies or the desktop app changed, so you know what else needs rebuilding.
- Don't apply while long jobs are running. Wait for them to finish.
- Apply only after a human says go, then restart services and verify they came back.
The promptIn the kit
Every 2 hours: 1. Fetch upstream. If there's nothing new, stop silently. 2. Merge upstream into the separate update worktree (never the live checkout).
Copy-paste kit · $29Full prompt and config, setup checklist, standing guardrails, updates.
Also in The Full Bench.
Guardrails
- The live install is never modified by the scheduled job.
- No restarts while long-running jobs are active.
- A saved config doesn't prove running services picked it up. Verify after restart.
What broke
- An upstream test failed on its own, unrelated to our patches. Reporting exact test names made it easy to tell 'their bug' from 'our bug'.
- Changing a config file didn't change the services already running. They kept the old value until restarted, so 'saved' and 'live' are now checked separately.
Our stack
Hermes cron job, git worktrees, the project's own test suite.
Want it running on your accounts?We install and adapt it for you, from $500.
Have us install it