feat(twister): durable recurring tasks — scheduleRecurring + migrate all connector chains by KrisBraun · Pull Request #215 · plotday/plot

KrisBraun · 2026-06-22T16:24:47Z

Problem

Self-rescheduling connector task chains (watch/subscription renewals, polling loops, self-heal) could die permanently and silently. The continuation of a periodic chain lived only in (a) a transient RUN_QUEUE message and (b) the connector's own "schedule the next occurrence" call — so any single broken hand-off killed the whole chain: a queue retry-exhaustion drop during a DB storm, a twist suspension, a deploy/eviction between fire and reschedule, or a callback that threw before rescheduling. A Gmail connection's hourly self-heal stopped ~June 4 and never restarted; sync flat-lined 18+ days with no error anywhere (a vanished chain is silent).

Fix (this PR — the SDK + connector side)

A new durable scheduleRecurring(key, callback, { intervalMs, firstRunAt? }) primitive whose cadence is owned by the platform (the CallbacksState Durable Object re-arms the next occurrence at fire time, independent of the queue hand-off). A dropped/suspended/crashed run now loses at most one beat, never the chain. intervalMs is a safety ceiling (max gap); firstRunAt sets a precise earlier fire (e.g. renew at expiry−24h) — the callback no longer needs to reschedule itself.

This PR contains:

twister:Tasks.scheduleRecurring abstract + Twist/Tool base-class convenience delegators; changeset (@plotday/twister minor). scheduleTask JSDoc narrowed to one-shot keyed.
All 13 connectors migrated off the fragile patterns:
- Bare-runTask + manual-token self-heal/renewal (the incident form): gmail, outlook-mail — token bookkeeping removed, upgrade() re-asserts on deploy.
- Keyed scheduleTask renewals/polls/daily-syncs: google-calendar, google-drive, google-chat, outlook-calendar, airtable, jira, apple-calendar, asana, google-tasks, slack, ms-teams.
SDK docs (connectors/AGENTS.md, RUNTIME.md, TOOLS_GUIDE.md) now recommend scheduleRecurring for recurring chains.

Companion PR

The runtime side — the DO recurring_interval_ms column + ceiling clamp, the alarm() advance-instead-of-delete, the api scheduleRecurring impl, a RUN_QUEUE dead-letter queue, and a backstop reconcile sweep — lands in a follow-up PR on plotday/core after this merges (it bumps the submodule pointer to this branch's merged commit).

Testing

Every connector: pnpm build + pnpm exec tsc --noEmit clean.
pnpm validate-changesets passes.
Reviewed end-to-end (whole-branch review); two cadence-ceiling bugs (airtable/asana polling too frequently) caught and fixed in 265f1aa.

🤖 Generated with Claude Code

Also adds scheduleRecurring() convenience helper to Twist and Tool base classes (mirroring scheduleTask/cancelScheduledTask), so connectors and tools can call this.scheduleRecurring() directly without going through this.tools.tasks.

…duleRecurring

…urring

…ring

…tale comments

…TOOLS_GUIDE + RUNTIME

…l cadence The migration set intervalMs (the safety ceiling) below the connector's actual POLL_INTERVAL_MS (airtable 6h, asana 5min), so the create() clamp min(firstRunAt, now+intervalMs) pinned the effective cadence to 60s — polling 360x/5x too often. Set the ceiling to POLL_INTERVAL_MS so the intended cadence is honored.

KrisBraun · 2026-06-22T16:24:47Z

Problem

Self-rescheduling connector task chains (watch/subscription renewals, polling loops, self-heal) could die permanently and silently. The continuation of a periodic chain lived only in (a) a transient RUN_QUEUE message and (b) the connector's own "schedule the next occurrence" call — so any single broken hand-off killed the whole chain: a queue retry-exhaustion drop during a DB storm, a twist suspension, a deploy/eviction between fire and reschedule, or a callback that threw before rescheduling. A Gmail connection's hourly self-heal stopped ~June 4 and never restarted; sync flat-lined 18+ days with no error anywhere (a vanished chain is silent).

Fix (this PR — the SDK + connector side)

A new durable scheduleRecurring(key, callback, { intervalMs, firstRunAt? }) primitive whose cadence is owned by the platform (the CallbacksState Durable Object re-arms the next occurrence at fire time, independent of the queue hand-off). A dropped/suspended/crashed run now loses at most one beat, never the chain. intervalMs is a safety ceiling (max gap); firstRunAt sets a precise earlier fire (e.g. renew at expiry−24h) — the callback no longer needs to reschedule itself.

This PR contains:

twister:Tasks.scheduleRecurring abstract + Twist/Tool base-class convenience delegators; changeset (@plotday/twister minor). scheduleTask JSDoc narrowed to one-shot keyed.
All 13 connectors migrated off the fragile patterns:
- Bare-runTask + manual-token self-heal/renewal (the incident form): gmail, outlook-mail — token bookkeeping removed, upgrade() re-asserts on deploy.
- Keyed scheduleTask renewals/polls/daily-syncs: google-calendar, google-drive, google-chat, outlook-calendar, airtable, jira, apple-calendar, asana, google-tasks, slack, ms-teams.
SDK docs (connectors/AGENTS.md, RUNTIME.md, TOOLS_GUIDE.md) now recommend scheduleRecurring for recurring chains.

Companion PR

The runtime side — the DO recurring_interval_ms column + ceiling clamp, the alarm() advance-instead-of-delete, the api scheduleRecurring impl, a RUN_QUEUE dead-letter queue, and a backstop reconcile sweep — lands in a follow-up PR on plotday/core after this merges (it bumps the submodule pointer to this branch's merged commit).

Testing

Every connector: pnpm build + pnpm exec tsc --noEmit clean.
pnpm validate-changesets passes.
Reviewed end-to-end (whole-branch review); two cadence-ceiling bugs (airtable/asana polling too frequently) caught and fixed in 265f1aa.

🤖 Generated with Claude Code

Also adds scheduleRecurring() convenience helper to Twist and Tool base classes (mirroring scheduleTask/cancelScheduledTask), so connectors and tools can call this.scheduleRecurring() directly without going through this.tools.tasks.

…duleRecurring

…urring

…ring

…tale comments

…TOOLS_GUIDE + RUNTIME

…l cadence The migration set intervalMs (the safety ceiling) below the connector's actual POLL_INTERVAL_MS (airtable 6h, asana 5min), so the create() clamp min(firstRunAt, now+intervalMs) pinned the effective cadence to 60s — polling 360x/5x too often. Set the ceiling to POLL_INTERVAL_MS so the intended cadence is honored.

KrisBraun · 2026-06-22T16:24:47Z

Problem

Self-rescheduling connector task chains (watch/subscription renewals, polling loops, self-heal) could die permanently and silently. The continuation of a periodic chain lived only in (a) a transient RUN_QUEUE message and (b) the connector's own "schedule the next occurrence" call — so any single broken hand-off killed the whole chain: a queue retry-exhaustion drop during a DB storm, a twist suspension, a deploy/eviction between fire and reschedule, or a callback that threw before rescheduling. A Gmail connection's hourly self-heal stopped ~June 4 and never restarted; sync flat-lined 18+ days with no error anywhere (a vanished chain is silent).

Fix (this PR — the SDK + connector side)

A new durable scheduleRecurring(key, callback, { intervalMs, firstRunAt? }) primitive whose cadence is owned by the platform (the CallbacksState Durable Object re-arms the next occurrence at fire time, independent of the queue hand-off). A dropped/suspended/crashed run now loses at most one beat, never the chain. intervalMs is a safety ceiling (max gap); firstRunAt sets a precise earlier fire (e.g. renew at expiry−24h) — the callback no longer needs to reschedule itself.

This PR contains:

twister:Tasks.scheduleRecurring abstract + Twist/Tool base-class convenience delegators; changeset (@plotday/twister minor). scheduleTask JSDoc narrowed to one-shot keyed.
All 13 connectors migrated off the fragile patterns:
- Bare-runTask + manual-token self-heal/renewal (the incident form): gmail, outlook-mail — token bookkeeping removed, upgrade() re-asserts on deploy.
- Keyed scheduleTask renewals/polls/daily-syncs: google-calendar, google-drive, google-chat, outlook-calendar, airtable, jira, apple-calendar, asana, google-tasks, slack, ms-teams.
SDK docs (connectors/AGENTS.md, RUNTIME.md, TOOLS_GUIDE.md) now recommend scheduleRecurring for recurring chains.

Companion PR

The runtime side — the DO recurring_interval_ms column + ceiling clamp, the alarm() advance-instead-of-delete, the api scheduleRecurring impl, a RUN_QUEUE dead-letter queue, and a backstop reconcile sweep — lands in a follow-up PR on plotday/core after this merges (it bumps the submodule pointer to this branch's merged commit).

Testing

Every connector: pnpm build + pnpm exec tsc --noEmit clean.
pnpm validate-changesets passes.
Reviewed end-to-end (whole-branch review); two cadence-ceiling bugs (airtable/asana polling too frequently) caught and fixed in 265f1aa.

🤖 Generated with Claude Code

Also adds scheduleRecurring() convenience helper to Twist and Tool base classes (mirroring scheduleTask/cancelScheduledTask), so connectors and tools can call this.scheduleRecurring() directly without going through this.tools.tasks.

…duleRecurring

…urring

…ring

…tale comments

…TOOLS_GUIDE + RUNTIME

…l cadence The migration set intervalMs (the safety ceiling) below the connector's actual POLL_INTERVAL_MS (airtable 6h, asana 5min), so the create() clamp min(firstRunAt, now+intervalMs) pinned the effective cadence to 60s — polling 360x/5x too often. Set the ceiling to POLL_INTERVAL_MS so the intended cadence is honored.

KrisBraun · 2026-06-22T16:24:47Z

Problem

Self-rescheduling connector task chains (watch/subscription renewals, polling loops, self-heal) could die permanently and silently. The continuation of a periodic chain lived only in (a) a transient RUN_QUEUE message and (b) the connector's own "schedule the next occurrence" call — so any single broken hand-off killed the whole chain: a queue retry-exhaustion drop during a DB storm, a twist suspension, a deploy/eviction between fire and reschedule, or a callback that threw before rescheduling. A Gmail connection's hourly self-heal stopped ~June 4 and never restarted; sync flat-lined 18+ days with no error anywhere (a vanished chain is silent).

Fix (this PR — the SDK + connector side)

A new durable scheduleRecurring(key, callback, { intervalMs, firstRunAt? }) primitive whose cadence is owned by the platform (the CallbacksState Durable Object re-arms the next occurrence at fire time, independent of the queue hand-off). A dropped/suspended/crashed run now loses at most one beat, never the chain. intervalMs is a safety ceiling (max gap); firstRunAt sets a precise earlier fire (e.g. renew at expiry−24h) — the callback no longer needs to reschedule itself.

This PR contains:

twister:Tasks.scheduleRecurring abstract + Twist/Tool base-class convenience delegators; changeset (@plotday/twister minor). scheduleTask JSDoc narrowed to one-shot keyed.
All 13 connectors migrated off the fragile patterns:
- Bare-runTask + manual-token self-heal/renewal (the incident form): gmail, outlook-mail — token bookkeeping removed, upgrade() re-asserts on deploy.
- Keyed scheduleTask renewals/polls/daily-syncs: google-calendar, google-drive, google-chat, outlook-calendar, airtable, jira, apple-calendar, asana, google-tasks, slack, ms-teams.
SDK docs (connectors/AGENTS.md, RUNTIME.md, TOOLS_GUIDE.md) now recommend scheduleRecurring for recurring chains.

Companion PR

The runtime side — the DO recurring_interval_ms column + ceiling clamp, the alarm() advance-instead-of-delete, the api scheduleRecurring impl, a RUN_QUEUE dead-letter queue, and a backstop reconcile sweep — lands in a follow-up PR on plotday/core after this merges (it bumps the submodule pointer to this branch's merged commit).

Testing

Every connector: pnpm build + pnpm exec tsc --noEmit clean.
pnpm validate-changesets passes.
Reviewed end-to-end (whole-branch review); two cadence-ceiling bugs (airtable/asana polling too frequently) caught and fixed in 265f1aa.

🤖 Generated with Claude Code

Also adds scheduleRecurring() convenience helper to Twist and Tool base classes (mirroring scheduleTask/cancelScheduledTask), so connectors and tools can call this.scheduleRecurring() directly without going through this.tools.tasks.

…duleRecurring

…urring

…ring

…tale comments

…TOOLS_GUIDE + RUNTIME

…l cadence The migration set intervalMs (the safety ceiling) below the connector's actual POLL_INTERVAL_MS (airtable 6h, asana 5min), so the create() clamp min(firstRunAt, now+intervalMs) pinned the effective cadence to 60s — polling 360x/5x too often. Set the ceiling to POLL_INTERVAL_MS so the intended cadence is honored.

KrisBraun · 2026-06-22T16:24:47Z

Problem

Self-rescheduling connector task chains (watch/subscription renewals, polling loops, self-heal) could die permanently and silently. The continuation of a periodic chain lived only in (a) a transient RUN_QUEUE message and (b) the connector's own "schedule the next occurrence" call — so any single broken hand-off killed the whole chain: a queue retry-exhaustion drop during a DB storm, a twist suspension, a deploy/eviction between fire and reschedule, or a callback that threw before rescheduling. A Gmail connection's hourly self-heal stopped ~June 4 and never restarted; sync flat-lined 18+ days with no error anywhere (a vanished chain is silent).

Fix (this PR — the SDK + connector side)

A new durable scheduleRecurring(key, callback, { intervalMs, firstRunAt? }) primitive whose cadence is owned by the platform (the CallbacksState Durable Object re-arms the next occurrence at fire time, independent of the queue hand-off). A dropped/suspended/crashed run now loses at most one beat, never the chain. intervalMs is a safety ceiling (max gap); firstRunAt sets a precise earlier fire (e.g. renew at expiry−24h) — the callback no longer needs to reschedule itself.

This PR contains:

twister:Tasks.scheduleRecurring abstract + Twist/Tool base-class convenience delegators; changeset (@plotday/twister minor). scheduleTask JSDoc narrowed to one-shot keyed.
All 13 connectors migrated off the fragile patterns:
- Bare-runTask + manual-token self-heal/renewal (the incident form): gmail, outlook-mail — token bookkeeping removed, upgrade() re-asserts on deploy.
- Keyed scheduleTask renewals/polls/daily-syncs: google-calendar, google-drive, google-chat, outlook-calendar, airtable, jira, apple-calendar, asana, google-tasks, slack, ms-teams.
SDK docs (connectors/AGENTS.md, RUNTIME.md, TOOLS_GUIDE.md) now recommend scheduleRecurring for recurring chains.

Companion PR

The runtime side — the DO recurring_interval_ms column + ceiling clamp, the alarm() advance-instead-of-delete, the api scheduleRecurring impl, a RUN_QUEUE dead-letter queue, and a backstop reconcile sweep — lands in a follow-up PR on plotday/core after this merges (it bumps the submodule pointer to this branch's merged commit).

Testing

Every connector: pnpm build + pnpm exec tsc --noEmit clean.
pnpm validate-changesets passes.
Reviewed end-to-end (whole-branch review); two cadence-ceiling bugs (airtable/asana polling too frequently) caught and fixed in 265f1aa.

🤖 Generated with Claude Code

Also adds scheduleRecurring() convenience helper to Twist and Tool base classes (mirroring scheduleTask/cancelScheduledTask), so connectors and tools can call this.scheduleRecurring() directly without going through this.tools.tasks.

…duleRecurring

…urring

…ring

…tale comments

…TOOLS_GUIDE + RUNTIME

…l cadence The migration set intervalMs (the safety ceiling) below the connector's actual POLL_INTERVAL_MS (airtable 6h, asana 5min), so the create() clamp min(firstRunAt, now+intervalMs) pinned the effective cadence to 60s — polling 360x/5x too often. Set the ceiling to POLL_INTERVAL_MS so the intended cadence is honored.

KrisBraun · 2026-06-22T16:24:47Z

Problem

Self-rescheduling connector task chains (watch/subscription renewals, polling loops, self-heal) could die permanently and silently. The continuation of a periodic chain lived only in (a) a transient RUN_QUEUE message and (b) the connector's own "schedule the next occurrence" call — so any single broken hand-off killed the whole chain: a queue retry-exhaustion drop during a DB storm, a twist suspension, a deploy/eviction between fire and reschedule, or a callback that threw before rescheduling. A Gmail connection's hourly self-heal stopped ~June 4 and never restarted; sync flat-lined 18+ days with no error anywhere (a vanished chain is silent).

Fix (this PR — the SDK + connector side)

A new durable scheduleRecurring(key, callback, { intervalMs, firstRunAt? }) primitive whose cadence is owned by the platform (the CallbacksState Durable Object re-arms the next occurrence at fire time, independent of the queue hand-off). A dropped/suspended/crashed run now loses at most one beat, never the chain. intervalMs is a safety ceiling (max gap); firstRunAt sets a precise earlier fire (e.g. renew at expiry−24h) — the callback no longer needs to reschedule itself.

This PR contains:

twister:Tasks.scheduleRecurring abstract + Twist/Tool base-class convenience delegators; changeset (@plotday/twister minor). scheduleTask JSDoc narrowed to one-shot keyed.
All 13 connectors migrated off the fragile patterns:
- Bare-runTask + manual-token self-heal/renewal (the incident form): gmail, outlook-mail — token bookkeeping removed, upgrade() re-asserts on deploy.
- Keyed scheduleTask renewals/polls/daily-syncs: google-calendar, google-drive, google-chat, outlook-calendar, airtable, jira, apple-calendar, asana, google-tasks, slack, ms-teams.
SDK docs (connectors/AGENTS.md, RUNTIME.md, TOOLS_GUIDE.md) now recommend scheduleRecurring for recurring chains.

Companion PR

The runtime side — the DO recurring_interval_ms column + ceiling clamp, the alarm() advance-instead-of-delete, the api scheduleRecurring impl, a RUN_QUEUE dead-letter queue, and a backstop reconcile sweep — lands in a follow-up PR on plotday/core after this merges (it bumps the submodule pointer to this branch's merged commit).

Testing

Every connector: pnpm build + pnpm exec tsc --noEmit clean.
pnpm validate-changesets passes.
Reviewed end-to-end (whole-branch review); two cadence-ceiling bugs (airtable/asana polling too frequently) caught and fixed in 265f1aa.

🤖 Generated with Claude Code

Also adds scheduleRecurring() convenience helper to Twist and Tool base classes (mirroring scheduleTask/cancelScheduledTask), so connectors and tools can call this.scheduleRecurring() directly without going through this.tools.tasks.

…duleRecurring

…urring

…ring

…tale comments

…TOOLS_GUIDE + RUNTIME

…l cadence The migration set intervalMs (the safety ceiling) below the connector's actual POLL_INTERVAL_MS (airtable 6h, asana 5min), so the create() clamp min(firstRunAt, now+intervalMs) pinned the effective cadence to 60s — polling 360x/5x too often. Set the ceiling to POLL_INTERVAL_MS so the intended cadence is honored.

KrisBraun · 2026-06-22T16:24:47Z

Problem

Self-rescheduling connector task chains (watch/subscription renewals, polling loops, self-heal) could die permanently and silently. The continuation of a periodic chain lived only in (a) a transient RUN_QUEUE message and (b) the connector's own "schedule the next occurrence" call — so any single broken hand-off killed the whole chain: a queue retry-exhaustion drop during a DB storm, a twist suspension, a deploy/eviction between fire and reschedule, or a callback that threw before rescheduling. A Gmail connection's hourly self-heal stopped ~June 4 and never restarted; sync flat-lined 18+ days with no error anywhere (a vanished chain is silent).

Fix (this PR — the SDK + connector side)

A new durable scheduleRecurring(key, callback, { intervalMs, firstRunAt? }) primitive whose cadence is owned by the platform (the CallbacksState Durable Object re-arms the next occurrence at fire time, independent of the queue hand-off). A dropped/suspended/crashed run now loses at most one beat, never the chain. intervalMs is a safety ceiling (max gap); firstRunAt sets a precise earlier fire (e.g. renew at expiry−24h) — the callback no longer needs to reschedule itself.

This PR contains:

twister:Tasks.scheduleRecurring abstract + Twist/Tool base-class convenience delegators; changeset (@plotday/twister minor). scheduleTask JSDoc narrowed to one-shot keyed.
All 13 connectors migrated off the fragile patterns:
- Bare-runTask + manual-token self-heal/renewal (the incident form): gmail, outlook-mail — token bookkeeping removed, upgrade() re-asserts on deploy.
- Keyed scheduleTask renewals/polls/daily-syncs: google-calendar, google-drive, google-chat, outlook-calendar, airtable, jira, apple-calendar, asana, google-tasks, slack, ms-teams.
SDK docs (connectors/AGENTS.md, RUNTIME.md, TOOLS_GUIDE.md) now recommend scheduleRecurring for recurring chains.

Companion PR

The runtime side — the DO recurring_interval_ms column + ceiling clamp, the alarm() advance-instead-of-delete, the api scheduleRecurring impl, a RUN_QUEUE dead-letter queue, and a backstop reconcile sweep — lands in a follow-up PR on plotday/core after this merges (it bumps the submodule pointer to this branch's merged commit).

Testing

Every connector: pnpm build + pnpm exec tsc --noEmit clean.
pnpm validate-changesets passes.
Reviewed end-to-end (whole-branch review); two cadence-ceiling bugs (airtable/asana polling too frequently) caught and fixed in 265f1aa.

🤖 Generated with Claude Code

Also adds scheduleRecurring() convenience helper to Twist and Tool base classes (mirroring scheduleTask/cancelScheduledTask), so connectors and tools can call this.scheduleRecurring() directly without going through this.tools.tasks.

…duleRecurring

…urring

…ring

…tale comments

…TOOLS_GUIDE + RUNTIME

…l cadence The migration set intervalMs (the safety ceiling) below the connector's actual POLL_INTERVAL_MS (airtable 6h, asana 5min), so the create() clamp min(firstRunAt, now+intervalMs) pinned the effective cadence to 60s — polling 360x/5x too often. Set the ceiling to POLL_INTERVAL_MS so the intended cadence is honored.

KrisBraun · 2026-06-22T16:24:47Z

Problem

Self-rescheduling connector task chains (watch/subscription renewals, polling loops, self-heal) could die permanently and silently. The continuation of a periodic chain lived only in (a) a transient RUN_QUEUE message and (b) the connector's own "schedule the next occurrence" call — so any single broken hand-off killed the whole chain: a queue retry-exhaustion drop during a DB storm, a twist suspension, a deploy/eviction between fire and reschedule, or a callback that threw before rescheduling. A Gmail connection's hourly self-heal stopped ~June 4 and never restarted; sync flat-lined 18+ days with no error anywhere (a vanished chain is silent).

Fix (this PR — the SDK + connector side)

A new durable scheduleRecurring(key, callback, { intervalMs, firstRunAt? }) primitive whose cadence is owned by the platform (the CallbacksState Durable Object re-arms the next occurrence at fire time, independent of the queue hand-off). A dropped/suspended/crashed run now loses at most one beat, never the chain. intervalMs is a safety ceiling (max gap); firstRunAt sets a precise earlier fire (e.g. renew at expiry−24h) — the callback no longer needs to reschedule itself.

This PR contains:

twister:Tasks.scheduleRecurring abstract + Twist/Tool base-class convenience delegators; changeset (@plotday/twister minor). scheduleTask JSDoc narrowed to one-shot keyed.
All 13 connectors migrated off the fragile patterns:
- Bare-runTask + manual-token self-heal/renewal (the incident form): gmail, outlook-mail — token bookkeeping removed, upgrade() re-asserts on deploy.
- Keyed scheduleTask renewals/polls/daily-syncs: google-calendar, google-drive, google-chat, outlook-calendar, airtable, jira, apple-calendar, asana, google-tasks, slack, ms-teams.
SDK docs (connectors/AGENTS.md, RUNTIME.md, TOOLS_GUIDE.md) now recommend scheduleRecurring for recurring chains.

Companion PR

The runtime side — the DO recurring_interval_ms column + ceiling clamp, the alarm() advance-instead-of-delete, the api scheduleRecurring impl, a RUN_QUEUE dead-letter queue, and a backstop reconcile sweep — lands in a follow-up PR on plotday/core after this merges (it bumps the submodule pointer to this branch's merged commit).

Testing

Every connector: pnpm build + pnpm exec tsc --noEmit clean.
pnpm validate-changesets passes.
Reviewed end-to-end (whole-branch review); two cadence-ceiling bugs (airtable/asana polling too frequently) caught and fixed in 265f1aa.

🤖 Generated with Claude Code

Also adds scheduleRecurring() convenience helper to Twist and Tool base classes (mirroring scheduleTask/cancelScheduledTask), so connectors and tools can call this.scheduleRecurring() directly without going through this.tools.tasks.

…duleRecurring

…urring

…ring

…tale comments

…TOOLS_GUIDE + RUNTIME

…l cadence The migration set intervalMs (the safety ceiling) below the connector's actual POLL_INTERVAL_MS (airtable 6h, asana 5min), so the create() clamp min(firstRunAt, now+intervalMs) pinned the effective cadence to 60s — polling 360x/5x too often. Set the ceiling to POLL_INTERVAL_MS so the intended cadence is honored.

KrisBraun added 18 commits June 22, 2026 10:32

feat(twister): add Tasks.scheduleRecurring durable recurring primitive
166afba

feat(google-calendar): durable watch-renewal via scheduleRecurring
e1b42e2

feat(outlook-mail): durable self-heal + subscription renewal via sche…
456a2ab
…duleRecurring

feat(outlook-mail): re-assert recurring chains on upgrade()
51a0714

feat(google-drive): durable recurring maintenance via scheduleRecurring
0b06129

feat(outlook-calendar): durable recurring maintenance via scheduleRec…
e6ebcb7
…urring

feat(jira): durable recurring maintenance via scheduleRecurring
6673091

feat(google-chat): durable recurring maintenance via scheduleRecurring
b4d53db

feat(airtable): durable recurring maintenance via scheduleRecurring
1f529e4

feat(apple-calendar): durable recurring maintenance via scheduleRecur…
ca64b98
…ring

feat(asana): durable recurring maintenance via scheduleRecurring
739f935

feat(google-tasks): durable recurring maintenance via scheduleRecurring
74b1610

feat(slack): durable recurring maintenance via scheduleRecurring
14c6337

feat(ms-teams): durable recurring maintenance via scheduleRecurring
7b6082e

docs(twister): recommend scheduleRecurring for recurring tasks; fix s…
b2a4672
…tale comments

docs(twister): point recurring-task guidance at scheduleRecurring in …
2e7de35
…TOOLS_GUIDE + RUNTIME

KrisBraun merged commit b2d19b7 into mainJun 22, 2026
1 check passed

KrisBraun deleted the durable-recurring-tasks branch June 22, 2026 16:57

plot-app-builderBot mentioned this pull request Jun 22, 2026
chore: version packages #216
Merged

KrisBraun mentioned this pull request Jun 22, 2026
fix(gmail,outlook-mail): durable recovery for stranded mailbox watches on upgrade #217
Merged