setup-routines

Setup Routines — one Q&A pass, a repo's custom routines fully reconciled

Any team member runs /agentloop:setup-routines inside a repo that ships a routine catalog (.claude/routines/), answers a few all-defaulted questions, and gets:

• Cloud routines (via the RemoteTrigger tool): create-or-update by canonical name, per-runner stagger, per-trigger git sources, prompt rendered from the repo's templates
• Durable local cron jobs (system crontab + dedicated worktree/clone, optional)
• Reconciliation of what already exists: canonical-name matches are updated in place; unknown routines are listed for the human (the API has no delete)

★ Interactive skill — never run unattended. Denied by a deny-interactive-unattended hook = you are in an unattended environment → stop and say "/agentloop:setup-routines needs a human present". Do NOT silently apply defaults: registering schedules is a durable outward action.

Scope split: /agentloop:fleet-setup owns the generic per-repo issue-sweep/pr-sweep loop (driver-based, multi-repo). This skill owns the routines a repo declares in its own catalog — producers, QA sweeps, nightly deploys, GC jobs, anything repo-shaped. Never schedule issue/pr-sweep from here (it would double-fire against the fleet and the templates have long since diverged).

The catalog contract (what the consuming repo provides)

Everything repo-specific lives in the repo, not in this skill:

.claude/routines/
  catalog.json        # the machine-readable routine list (schema below)
  templates/*.md      # one prompt template per routine, {{RUNNER}} etc. placeholders

catalog.json schema (all routine fields except name/template/cron optional):

{
  "repo": "Owner/repo",                     // canonical slug, used for trigger sources
  "envSetupScript": ["line1", "line2"],     // the environment's setup-script lines (printed, see Step 2c)
  "identity": { "gitUserName": true },      // Step 2 renders `git config --global user.name <runner>` advice into report
  "presenceRegister": "CMD {{RUNNER}} {{ROUTINES}} {{MODEL}} {{LOCAL}}",  // optional post-reconcile hook, run from repo root
  "routines": [
    {
      "name": "issue-graph producer hourly",  // canonical trigger name — the reconcile key
      "template": "templates/producer.md",
      "cron": "{{MIN}} * * * *",              // {{MIN}} = (minBase + runner offset) % 60
      "minBase": 45,
      "allowedTools": ["Bash", "Read", "Glob", "Grep"],
      "model": "claude-sonnet-5",             // default; Step 1 lets the user override
      "extraSources": ["Owner/other-repo"],   // additional git sources → per-trigger scope (see doctrine below)
      "clearMcpConnections": true,            // script-only routines don't need inherited MCP connectors
      "singleton": true,                      // team should run exactly one — affects Step 1 default
      "singletonProbe": "some read-only shell command",  // optional; output shown to inform the default
      "defaultOn": false,                     // Step 1 pre-selection when the runner has no existing trigger
      "vars": { "ENV": ["test", "staging"] }, // extra template placeholders; user picks/notes a value in Step 1
      "localOnly": false,                     // true = never a cloud trigger (crontab only), e.g. GC jobs
      "local": {                              // optional local-crontab form of this routine
        "command": "bun scripts/foo.ts",      // run inside the dedicated checkout; pure-script routines skip `claude -p`
        "lockName": "foo.lock"
      },
      "envPrereqs": ["Network Access: *.example.dev", "env var FOO_BAR (optional)"],
      "note": "one per team; shown in Step 1 option description"
    }
  ]
}