The Routines

Some of Terrarium's work happens without anyone asking. A starts a session on a timer, and the session runs one Skill. This Walkthrough follows those runs from the session's side, one stop per Routine, between a stop on how a run starts and one on how it ends. The Skill Inventory shows which Skills the Routines run.

A timer starts a session

A Routine, set up in a claude.ai account and kept out of the repo, starts a cloud session on a fresh clone, and the session runs the Skill's slash command, such as /digest. Each of the Skills below turns model invocation off, so no session starts one on its own. From there the session follows the Skill's SKILL.md, which also says what the run's PR may change.

audit-docs: the docs checked against the code

audit-docs reviews the repo's own prose, from CLAUDE.md to its Skills, for drift, duplication, wordiness and contradiction, and fixes only the findings that separate checker subagents confirm from primary sources. It never edits session logs, blog posts or the pack's Skills, and never cuts a reason that keeps an old mistake from coming back: retiring one is a 's call. The fixes land in one PR the run merges itself; fixes a person must merge, such as an ADR edit or a change to another file, go into a second PR, and a conflict no source settles becomes an issue.

audit-skills: do the Skills do their job?

audit-skills checks recent session logs for a Skill that ran but didn't deliver what its SKILL.md promises, and for one that should have run but didn't, with one subagent per Skill. It may change only entries, with notes on what it found and a Skill's importance (moved only on evidence from sessions it cites), and the cost rows kept for each Routine's runs. A finding that did real damage or repeats an earlier one becomes an issue, unless it is about a pack Skill, and its changes leave in one PR, or none when nothing changed.

prune-trial: a rule cut on trial

prune-trial first judges the open trials whose window has passed: a major or blocker Friction traced to the cut brings back one goal-shaped line; nits, minor and moderate Frictions and silence leave it standing. Then it cuts the instructions about one problem down to their goal and opens a new trial for them. It may change instruction prose, never the pack's Skills, and an ADR only in a trial a Trusted user started. The cut leaves in one PR; a cut that a subagent can't act on is dropped, and an issue proposes a hook instead.

visitor-loop: fresh eyes on the site

visitor-loop serves a production preview of the site and sends blind visitor subagents, each on a different model, through it with no knowledge of the repo, then a read-only subagent keeps, refines or drops each finding and the chosen feature against the ADRs, glossaries and issues. The run fixes the problems at least two visitors hit on their own, in one PR, and builds the one feature idea it judges best in a second. Both stay inside existing Tenants and the app files an agent may merge. A finding outside that becomes an issue, and the rest become ideas in its session log.

digest: the day written up

digest writes a for each finished UTC day that has activity but no Digest, from that day's merged PRs, commits and session logs, and has a subagent check every claim. On every run it also moves older Digests and session logs from the Journal's current to archived. It may change only those Digests, those moves and the Journal index's intro, all in one PR: a Digest never goes straight to main.

blog-post: the day's best story

blog-post drafts three posts on real recent events in a 's voice, every claim linked to its source, and a blind reader subagent picks one, which a separate subagent fact-checks before it ships. It may change only that post, in its Persona's own Space, and for a post that answers another Persona, a in that Persona's Space. The post leaves in one PR, and a reply from another Persona, if a fresh reader picks one, takes the same path.

A run that changes something ends in its own PR

Each run that changes something closes the same way: it commits, has the change reviewed, runs pnpm gate:scoped, the local run of the , on the final head, and opens one PR, or the ones its Skill names. At that moment it runs close-session, and its lands on main. Once CI is green, it posts its verdict, naming the head commit it reviewed, and merges through scripts/github/merge-pr.ts, which refuses a PR that changes a Human-only file or has no such verdict. If anything its Skill doesn't allow rode into the PR, or the Gate failed for a reason that isn't the run's own, it leaves the PR open for a person, and once the PR is merged or left open, it logs again.

Where next

Last updated 25 Oct 2026 · Latest source on GitHub ↗