Workbench schedules

Make the real scheduled tasks match the desired set. Desired state lives in the workbench at .workbench/schedules.json. Regenerate it first so it reflects the current profile:

python3 "${CLAUDE_PLUGIN_ROOT}/scripts/workbench.py" schedules --root "<workbench>"

Each routine there has id, name (always [Workbench] …), cron, timezone, when (human text) and prompt. The prompt is deliberately short: it invokes the plugin skill, with the copy in .workbench/routines/ as a fallback, so upgrading the plugin upgrades the routines without touching the schedule.

1. Pick the runner

The routines need to read and write the workbench folder, so where they run decides what's guaranteed. The runner follows from two facts in .workbench/profile.json: which surface the owner has (surface: cowork or code) and where the vault lives (vault_home: local or repo). Setup records both. Don't ask the owner to pick a runner; derive it:

Store the result as runner in the profile (cowork-local, code-desktop, code-cloud).

Runner 1, Cowork local (Cowork-only users). Scheduled tasks in Cowork, against the local vault folder (or the local clone of a repo vault). They run only while the computer is on and the desktop app is open. A task that needs local files runs locally even where Cowork can run tasks in the cloud. Anthropic documents no catch-up, retry or failure notice for this case, so recovery is built into the routines: each one appends to Tracking/Signals/RUN_LEDGER.md, the briefing names any routine missing from it, the dream files every unprocessed capture whatever its date, and the recap writes a reconstructed file for a missed day.

Runner 2, Code Desktop local (Code users, local vault). Scheduled tasks in the Claude Code desktop app. Documented behavior: on wake or app start, exactly one catch-up run for the most recent missed run within 7 days; older misses are skipped and listed in the task's History. The run ledger still records every run, so the briefing can tell a catch-up from a skip.

Runner 3, Code cloud (Code users, repo vault). Cloud scheduled tasks against the private GitHub repo. A research preview. Each run starts from a fresh clone, can push only to claude/ branches, and the minimum interval is one hour. No catch-up and no automatic retry are documented. Add this to the end of each prompt:

"The workbench is the repository <owner/repo>. Before reading anything, fetch and merge every unmerged claude/workbench-* branch into your working copy, so you see the earlier runs' work; if a merge conflicts, stop, append a failed line to the run ledger naming the branches, and push nothing else. When done, commit only the files this routine owns, with message <routine-id>: YYYY-MM-DD, and push to the branch claude/workbench-<routine-id>-YYYY-MM-DD. Never force-push."

The owner's side is one ritual: the first local session of the day pulls and merges the claude/ branches before the briefing reads the vault (or the owner merges the newest one in one click on GitHub; each branch already contains the earlier ones). Because each routine commits only files it owns, merges conflict only where git can show it. If cloud tasks are unavailable or switched off for the account, fall back to Runner 2 against the local clone; never run both for the same routine.

Plan notes. On Team and Enterprise plans, Cowork sessions run in the cloud only if the organization turns on "Run Cowork in the cloud" (Organization settings > Cowork), and even then a task that requires local files runs locally. On personal plans there's no location choice. Either way, a workbench in a local folder runs locally.

What isn't a scheduler. A session-only cron (one that dies when the chat closes) is not a scheduler for this; don't use it. If the tools for the derived runner aren't in this session (for example, Cowork-only and you're in a plain chat), write the table from schedules.json into your reply and tell them how to create the tasks by hand in the Claude desktop app (the Scheduled section) with each prompt. Stop there.

Duplicates across surfaces. The same routine can exist on more than one surface (a Cowork task and a Code Desktop task, or a desktop task and a cloud task). Inventory every surface you can see in step 2, and keep exactly one per routine, on the derived runner.

2. Inventory what exists

List every scheduled task you can see. Sort each into:

• Ours, current: name starts with [Workbench] and matches a desired id/name.
• Ours, retired: name starts with [Workbench] but no desired routine has that name.
• Look-alikes: not prefixed, but clearly an older version of the same job, e.g. "daily briefing", "vault-dream", "EOD recap", "weekly plan", or any task whose prompt points at this workbench folder. Show these to the person and ask, one line each: replace with the workbench version, or leave alone? Replacing means pausing (disabling) the old task, not deleting it, so its wording survives and it can be switched back on. Delete only if the tool can't pause, and only with an explicit yes.
• Unrelated: leave alone and don't mention them.

If two tasks map to the same routine, that's a duplicate. Keep one (prefer the one already named correctly) and delete the other after saying so.

3. Reconcile

Show the plan as one small table (routine · action · when) and get a yes, then apply it:

Cron expressions are in the routine's timezone. If the scheduling tool only accepts UTC, convert it and note the conversion, since daylight saving shifts it twice a year; prefer a tool option that takes a timezone when there is one.

If the tool supports it, turn on notifications only for failures. Nobody needs six pings a day.

4. Verify and record

1. List tasks again and confirm each desired routine exists exactly once with the right schedule.
2. Offer to run the Morning Briefing once now as a test, and look at the file it wrote.
3. Append to Tracking/FRICTION_LOG.md under ## Resolved one line: today's date, "routines reconciled", and what was created, updated or deleted. That's the audit trail if a routine later goes missing.

Finish with the final table: routine, when (in their timezone), where it runs.