DottaandPaperclip
baf651c8d0
Reviewed and verified at f1c48c924d.
Co-Authored-By: Paperclip <noreply@paperclip.ing>
3.0 KiB
Run dispatch without idle retry scans
The run-dispatch module owns retry admission, its durable scheduled_retry_at
deadline, and atomic promotion. The heartbeat queue owns concurrency, current
execution authority, and dispatch. The shared coordinator owns notification
coalescing, nonoverlap, transaction uncertainty, drain/resume and shutdown.
The heartbeat facade and index.ts only connect these owners.
A scheduled retry is a future obligation. Startup reconstructs its next stored
deadline through the run-dispatch adapter. Writes that create, move, cancel or
invalidate retries notify the retry worker in their transaction. Promotion uses
the existing gate and compare-and-set, and its committed write wakes ordinary
queued dispatch. A queued run behind occupied capacity waits for the owning
run's lifecycle transition. Configuration and authority changes also wake it.
Retry-schedule notifications are separate from ordinary queue eligibility.
notifyRunDispatch wakes only queued dispatch; ordinary issue, agent, budget
and company edits cannot repeatedly scan the global retry schedule.
notifyScheduledRetryDispatch wakes retry reconciliation as well as dispatch
for actual retry creation, rescheduling and cancellation. Its producers include
bounded and operator-requested retries, recovery retries, native capacity and
safe-replacement successors, and the existing retry-cancellation transition
boundary. Shared runtime changes do not impose a spacing floor on a saved deadline.
An idle queue has no recurring read or timer-driven dispatch attempt. Notifications batch for one second and carry the changed company to queued dispatch. Known retry deadlines still preempt batching. Cancellation-owned retries stay with their cancellation worker and do not retain a promotion deadline.
The supported writer contract is service writes through the shared database connection scope, including caller-owned and nested transactions. Existing explicit domain notification boundaries cover lifecycle, issue review state, agent configuration/status, company lifecycle and scheduling settings. Direct SQL writers and independent replicas must provide an explicit ingress signal or restart reconstruction; a process-local signal does not observe their commits. This change adds no permanent database listener or write interception.
The remaining periodic recovery chain still owns deferred comment receipts, execution-wait cancellation recovery, orphan/resource recovery, stranded issue assignments, dependency wake repair, task watchdogs, silent active runs and stale locks. Those duties retain their backstop until their separate producer/deadline migrations are qualified. This slice removes only the periodic scheduled-retry promotion call. Ordinary queued dispatch can still be reached by that backstop; the existing atomic claim and per-agent start lock prevent duplicate execution.
These workers use internal process deadlines. They do not register Cloud wake alarms and do not relax the durable idle-sleep safety proof.