root/fusion
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

fix(engine): keep settle-window ledger updates synchronous on task moves

Browse Source 
FN-8656 moved the recentEngineTodoRequeues / dispatch-oscillation reset behind
await resolveTaskParkedColumns, racing sync delete/update handlers. A hold
requeue then delete in the same tick could re-set the settle guard after the
clear and leave the card undispatchable. Consume the sync parked lanes in the
emitter prologue again.
This commit is contained in:
gsxdsm
2026-08-02 17:33:11 -07:00
parent f21c07929a
commit a84e16118a
1 changed files with 36 additions and 20 deletions
Download Patch File Download Diff File Expand all files Collapse all files

56
packages/engine/src/scheduler.ts
View File

@@ -1103,6 +1103,42 @@ export class Scheduler {
}
}
/*
FNXC:SchedulerDispatchOscillation 2026-08-02-17:20:
The settle-window ledger (`recentEngineTodoRequeues`) and the dispatch-oscillation reset are
CONSUMED SYNCHRONOUSLY, so they must stay in the emitter tick with the rest of the synchronous
prologue — reading the sync, emitter-lane-aware `parked`, never the async `resolvedParked`.
FN-8656 moved these two arms behind `await resolveTaskParkedColumns` while converting lane
lookups; that broke the ordering against the synchronous `task:deleted`/`task:updated`
handlers, which also mutate `recentEngineTodoRequeues`/oscillation state. Concretely: a card
requeued to hold then deleted in the same tick had its settle-window SET run a microtask AFTER
the delete's synchronous CLEAR, so the guard survived deletion and the card never re-dispatched
(todo-inprogress-flapping.test.ts "clears the settle-window guard when the task is deleted" and
"resets dispatch oscillation state on forward transition to in-review"). This obeys FN-8656's
own stated rule: convert to the async resolver ONLY where the answer is consumed asynchronously.
Production emitters carry lanes, so `parked` resolves renamed hold/wip/review columns here too;
the lane-less SQLite polling emitters retain the legacy fallback exactly as before.
*/
if (from === parked.wip && to === parked.hold) {
if (source === "engine") {
this.recentEngineTodoRequeues.set(task.id, task.columnMovedAt ?? new Date().toISOString());
} else {
this.recentEngineTodoRequeues.delete(task.id);
}
} else if (to === parked.review || parked.terminal.has(to)) {
this.recentEngineTodoRequeues.delete(task.id);
if (task.dispatchStormCount != null || task.lastDispatchAt != null || task.executeRequeueLoopCount != null || task.executeRequeueLoopSignature != null) {
void this.store.updateTask(task.id, {
dispatchStormCount: null,
lastDispatchAt: null,
executeRequeueLoopCount: null,
executeRequeueLoopSignature: null,
}).catch((error) => {
schedulerLog.warn(`Failed to reset dispatch oscillation state for ${task.id} on move to ${to}: ${error instanceof Error ? error.message : String(error)}`);
});
}
}
const resolvedParked = mergeParkedColumns(await resolveTaskParkedColumns(this.store, task.id), lanes);
// FN-3895/FN-3924: complement periodic stale-blockedBy self-healing with immediate
@@ -1181,26 +1217,6 @@ export class Scheduler {
}
}
if (from === resolvedParked.wip && to === resolvedParked.hold) {
if (source === "engine") {
this.recentEngineTodoRequeues.set(task.id, task.columnMovedAt ?? new Date().toISOString());
} else {
this.recentEngineTodoRequeues.delete(task.id);
}
} else if (to === resolvedParked.review || resolvedParked.terminal.has(to)) {
this.recentEngineTodoRequeues.delete(task.id);
if (task.dispatchStormCount != null || task.lastDispatchAt != null || task.executeRequeueLoopCount != null || task.executeRequeueLoopSignature != null) {
void this.store.updateTask(task.id, {
dispatchStormCount: null,
lastDispatchAt: null,
executeRequeueLoopCount: null,
executeRequeueLoopSignature: null,
}).catch((error) => {
schedulerLog.warn(`Failed to reset dispatch oscillation state for ${task.id} on move to ${to}: ${error instanceof Error ? error.message : String(error)}`);
});
}
}
// Event-driven scheduling: when a task moves to "done" (completion) or "todo" (retry/manual move),
// trigger scheduling immediately so waiting tasks can start without waiting
// for the next poll interval (up to 15 seconds).