Cronsole Claude Routines

Claude Code routines — scheduled agents in the cloud

A routine is a prompt Anthropic runs on a schedule, in an isolated checkout, ending in a pull request. It is one of two places a scheduled agent job can live, and picking the wrong one is the expensive mistake.

Choose the cloud when the work needs no local state — no installed dependencies, no .env, no local service, no machine that has to be on. Choose local when it does, or when the work must touch that machine.

1. The three constraints that shape everything

1. Nothing lives locally. There is no config file and no prompt on disk. The routine's entire definition is in the dashboard, which means the prompt is the only artifact — there is no version history behind it and nothing to diff when a run goes wrong. Keep your own copy of every prompt in the repo it acts on, or in a notes folder. Nothing else will.
2. You cannot enumerate the fleet from a fresh session. CronList is session-scoped and shows nothing from another session. Use /schedule, or RemoteTrigger with {action: "list"}. Do not read its silence as an empty fleet.
3. Deleting is dashboard-only. No API surface Cronsole has can remove a routine. The reversible move is to turn one off and leave it.

2. Build one

Keep a fleet inventory

A cloud fleet has no local record at all, so a routine you forget about keeps running, keeps opening PRs, and keeps costing money. Write down — anywhere you will actually look — each routine's name, repo, branch prefix, fire time and whether it is on. The dashboard is the source of truth and your inventory is not, which is exactly why it drifts and exactly why it is worth keeping.

3. The four rules a cloud prompt must follow

The first three are the universal unattended-agent rules. The fourth is what the cloud adds.

1. Never offer a choice. Nobody answers a question at 14:00 UTC on a Saturday. A prompt that asks "which repositories should I include?" does not fail — it stalls, and the run is over before anything happened. Give the parameter, or tell the agent to pick and say which it picked.
2. Demand explicit failure reporting. An agent that cannot finish a step narrates success instead. completed means the agent finished its turn — never that the job was done. End every prompt with: "If any step fails, say so explicitly in your final message and name the step that failed, rather than summarizing what you would have produced."
3. State the verification, and forbid a PR when it fails. Revert, leave the tree clean, say what blocked it.
4. Carry the one-open-PR rule in the prompt itself. A local runner can enforce this in code by adopting an open branch. A cloud routine branches by hand and gets no such help, so its prompt must say: check for an open PR on this prefix first; if one exists, commit onto that branch and comment on the PR rather than cutting a new one.

Why rule 4 exists: two runs a day apart both branched from the base while the first PR sat unmerged. Each read a work list missing the other's changes, picked the same item, and reimplemented it. The two PRs shared nine identical files and one had to be closed. Code fixed the local case; the cloud case is only ever fixed in prose.

4. Cronsole's side, in one paragraph

CLAUDE_CODE is a dual-mode connector, and what it can do depends on the install, not on the platform. With a readable Claude Code session on the Cronsole backend's host it uses the undocumented OAuth /v1/code/triggers API and gets a real sync plus create, setStatus, updateSchedule and run. Without one it falls back to the documented per-routine fire token: run only, and syncTasks returns the routines you declared in config — which is not a sync and is documented as not being one. So unsupportedVerbs is a getter, and a capability matrix read on one machine does not describe another. Neither mode can delete. Full detail, and what every reading on screen means, in references/cronsole-side.md.

5. Anti-patterns

• Putting a job in the cloud that needs a specific machine. No local state reaches it. If the work needs installed dependencies, a .env, a running service or a local path, it is a local job.
• Creating a routine without registering its branch prefix with whatever reviews and merges its PRs. It will open them forever and nothing will ever merge them.
• Trusting completed. Read the step list.
• Assuming CronList shows the fleet. It is session-scoped and shows nothing.
• Trying to delete a routine from Cronsole. No verb exists in either mode, by enumeration. Disconnecting the Claude source removes its tracked tasks along with the declaration — that is not a delete, and it is not reversible without re-declaring.
• Keeping the prompt only in the dashboard. It is the only artifact and it has no history.
• Copying a prompt between routines. The infrastructure is generic; the prompt is not.

Related