Keyed singleton scheduled tasks (scheduleTask) + migrate connectors off leak-prone renewal pattern

[KrisBraun]

Why

A connector's self-renewing scheduled task (watch/webhook renewal, polling, daily digest) was hand-managed with runTask({ runAt }) + a stored token + cancelTask() before re-scheduling. That pattern is easy to get wrong: a renewal callback that re-schedules itself, plus any extra entry into the setup path (an onChannelEnabled re-dispatch, a re-init, a race), leaks parallel self-perpetuating task chains that accumulate without bound and eventually trip the runtime execution quota.

This is exactly what auto-suspended the Google Drive connector (≈350 → 3351 worker invocations/day, near-zero DB writes — a steady 24/7 background climb from accumulated renewal chains). google-calendar had the identical latent bug, and several pollers (airtable, google-tasks) had no token/cancel at all, so they could fork unboundedly with no way to stop them.

The leaky pattern was even documented as the recommended approach in connectors/AGENTS.md, so connectors copied it faithfully.

What

New SDK primitive (@plotday/twister) — makes the leak impossible by construction:

• this.scheduleTask(key, callback, { runAt }) — schedules a singleton task per key; re-scheduling under the same key atomically cancels-and-replaces any pending task. At most one live scheduled task per key.
• this.cancelScheduledTask(key) — tears it down (no-op if none / already ran).

Exposed on the Tasks tool interface and as Tool/Twist helpers. Changeset included (minor). The runtime side (the task_key column + atomic replace-on-create in CallbacksState, and the Tasks-tool wiring) ships in the matching core PR, which also bumps the submodule pointer to this branch.

Docs — fixed the connectors/AGENTS.md watch-renewal snippet (it was teaching the leaky pattern), added a checklist item + pitfall, and updated twister/docs/RUNTIME.md (periodic-work example) and TOOLS_GUIDE.md (scheduling section) to steer recurring tasks to scheduleTask.

Connector migration — moved every connector with a self-renewing/recurring scheduled task onto the keyed primitive:

Class	Connectors
Watch/subscription renewal	google-drive, google-calendar, google-chat, ms-teams, jira, outlook-calendar
Periodic poll	airtable (+ webhook renewal), google-tasks, asana, apple-calendar
Daily digest	slack (members + custom-emoji), ms-teams, google-chat

Each chain is keyed per resource (watch-renewal:<id>, poll:<id>, members-sync:<channelId>, …) so re-scheduling replaces rather than forks, and disable paths cancel by key. Pollers that previously had no teardown now cancel on disable. google-drive/google-calendar additionally stop the prior provider watch before re-creating it (so a redundant setup can't orphan a still-firing watch). Added a google-drive regression test.

Left as-is (not affected): gmail and outlook-mail already cancel-before-reschedule (no leak) and keep their self-heal logic; github's runTask override is a pass-through; linkedin's retry terminates.

Backwards compatibility

scheduleTask/taskKey are purely additive. Any pre-deploy in-flight (un-keyed) renewal task fires once more after deploy, then the chain becomes keyed — no lingering leak. Suspended connectors un-suspend on their next deploy.

Verification

• All 13 touched connectors build + lint clean; no leftover *_renewal_task/poll_task/cancelTask references.
• twister builds; changeset validates.
• Runtime behavior (atomic replace / independent keys / cancel) is covered by an integration test in the core PR (5 tests), plus a google-drive connector regression test here.

🤖 Generated with Claude Code