Skip to content

fix(server): pressing Stop now clears orphaned delegated follow-ups #21607

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Closed
sybeeria wants to merge 1 commit into from
Closed

fix(server): pressing Stop now clears orphaned delegated follow-ups #21607

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
243 changes: 136 additions & 107 deletions apps/server/src/orchestration-v2/Orchestrator.ts
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -9594,6 +9594,121 @@ const makeOrchestrator = Effect.fn("orchestrationV2.Orchestrator.layer")(functio
return undefined;
});

const cancelIdleOrphans = (input: {
readonly command: Extract<
OrchestrationV2ServerCommand,
{ readonly type: "run.interrupt" | "thread.stop" }
>;
readonly events: Ref.Ref<Array<OrchestrationV2DomainEvent>>;
readonly projection: Pick<OrchestrationV2ThreadProjection, "subagents" | "turnItems">;
}) =>
Effect.gen(function* () {
const { command, events, projection } = input;
const pending = yield* Ref.get(events);
const children = new Map<
ThreadId,
ProjectionRecords<"runs" | "providerThreads" | "runtimeRequests" | "turnItems"> | null
>();
const idleChildren = new Map<ThreadId, boolean>();
let nodes: ReadonlyArray<OrchestrationV2ExecutionNode> | undefined;
for (const recorded of projection.subagents) {
const updated = pending.findLast(
(event) => event.type === "subagent.updated" && event.payload.id === recorded.id,
);
const task = updated?.type === "subagent.updated" ? updated.payload : recorded;
if (
task.origin !== "app_owned" ||
task.childThreadId === null ||
!isOrchestrationV2WorkActive(task.status)
)
continue;
const childThreadId = task.childThreadId;
const child = children.has(childThreadId)
? (children.get(childThreadId) ?? null)
: yield* projectionStore
.getThreadRecords(
childThreadId,
["runs", "providerThreads", "runtimeRequests", "turnItems"],
{
turnItemTypes: ["command_execution", "dynamic_tool", "subagent"],
turnItemStatuses: ["pending", "running", "waiting"],
},
)
.pipe(
Effect.catchTags({
ProjectionStoreThreadNotFoundError: () => Effect.succeed(null),
}),
Effect.mapError(
(cause) => new OrchestratorProjectionError({ threadId: childThreadId, cause }),
),
);
children.set(childThreadId, child);
if (child === null) continue;
if (
child.thread.lineage.parentThreadId !== command.threadId ||
child.thread.lineage.relationshipToParent !== "subagent" ||
(child.thread.forkedFrom?.type === "node" &&
child.thread.forkedFrom.nodeId === task.id) ||
child.runs.some((run) => run.delegatedTaskId === task.id)
)
continue;
const idle =
idleChildren.get(task.childThreadId) ??
(!child.runs.some((run) => run.status === "queued" || isBlockingRun(run)) &&
!child.providerThreads.some(
(thread) =>
thread.status === "active" || (thread.pendingBackgroundTasks?.length ?? 0) > 0,
) &&
!child.runtimeRequests.some((request) => request.status === "pending") &&
child.turnItems.length === 0 &&
Option.isNone(yield* Effect.findFirst(child.runs, awaitsRestartContinuation)));
idleChildren.set(task.childThreadId, idle);
if (!idle) continue;
// Stop can cancel an idle orphan without guessing which completed run owned its result.
const now = yield* DateTime.now;
const emitEvent = emit(events, command);
yield* emitEvent({
type: "subagent.updated",
threadId: command.threadId,
nodeId: task.id,
occurredAt: now,
payload: {
...task,
status: "cancelled",
completedAt: now,
updatedAt: now,
completionDelivery: { state: "disposed", observedByRunId: null },
},
});
nodes ??= (yield* projectionStore.getThreadRecords(command.threadId, ["nodes"])).nodes;
const node = nodes.find((candidate) => candidate.id === task.id);
if (node !== undefined)
yield* emitEvent({
type: "node.updated",
threadId: command.threadId,
nodeId: node.id,
occurredAt: now,
payload: { ...node, status: "cancelled", completedAt: now },
});
for (const item of projection.turnItems) {
if (item.type !== "subagent" || item.subagentId !== task.id) continue;
yield* emitEvent({
type: "turn-item.updated",
threadId: command.threadId,
nodeId: task.id,
occurredAt: now,
payload: { ...item, status: "cancelled", completedAt: now, updatedAt: now },
});
}
}
}).pipe(
Effect.mapError((cause) =>
Schema.is(OrchestratorProjectionError)(cause)
? cause
: new OrchestratorProjectionError({ threadId: input.command.threadId, cause }),
),
);

const dispatchThreadStop = (
command: Extract<OrchestrationV2ServerCommand, { readonly type: "thread.stop" }>,
events: Ref.Ref<Array<OrchestrationV2DomainEvent>>,
Expand All @@ -9611,111 +9726,6 @@ const makeOrchestrator = Effect.fn("orchestrationV2.Orchestrator.layer")(functio
),
);
if (projection.thread.deletedAt !== null) return undefined;
const cancelIdleOrphans = Effect.gen(function* () {
const pending = yield* Ref.get(events);
const children = new Map<
ThreadId,
ProjectionRecords<"runs" | "providerThreads" | "runtimeRequests" | "turnItems"> | null
>();
const idleChildren = new Map<ThreadId, boolean>();
let nodes: ReadonlyArray<OrchestrationV2ExecutionNode> | undefined;
for (const recorded of projection.subagents) {
const updated = pending.findLast(
(event) => event.type === "subagent.updated" && event.payload.id === recorded.id,
);
const task = updated?.type === "subagent.updated" ? updated.payload : recorded;
if (
task.origin !== "app_owned" ||
task.childThreadId === null ||
!isOrchestrationV2WorkActive(task.status)
)
continue;
const childThreadId = task.childThreadId;
const child = children.has(childThreadId)
? (children.get(childThreadId) ?? null)
: yield* projectionStore
.getThreadRecords(
childThreadId,
["runs", "providerThreads", "runtimeRequests", "turnItems"],
{
turnItemTypes: ["command_execution", "dynamic_tool", "subagent"],
turnItemStatuses: ["pending", "running", "waiting"],
},
)
.pipe(
Effect.catchTags({
ProjectionStoreThreadNotFoundError: () => Effect.succeed(null),
}),
Effect.mapError(
(cause) => new OrchestratorProjectionError({ threadId: childThreadId, cause }),
),
);
children.set(childThreadId, child);
if (child === null) continue;
if (
child.thread.lineage.parentThreadId !== command.threadId ||
child.thread.lineage.relationshipToParent !== "subagent" ||
(child.thread.forkedFrom?.type === "node" &&
child.thread.forkedFrom.nodeId === task.id) ||
child.runs.some((run) => run.delegatedTaskId === task.id)
)
continue;
const idle =
idleChildren.get(task.childThreadId) ??
(!child.runs.some((run) => run.status === "queued" || isBlockingRun(run)) &&
!child.providerThreads.some(
(thread) =>
thread.status === "active" || (thread.pendingBackgroundTasks?.length ?? 0) > 0,
) &&
!child.runtimeRequests.some((request) => request.status === "pending") &&
child.turnItems.length === 0 &&
Option.isNone(yield* Effect.findFirst(child.runs, awaitsRestartContinuation)));
idleChildren.set(task.childThreadId, idle);
if (!idle) continue;
// Stop can cancel an idle orphan without guessing which completed run owned its result.
const now = yield* DateTime.now;
const emitEvent = emit(events, command);
yield* emitEvent({
type: "subagent.updated",
threadId: command.threadId,
nodeId: task.id,
occurredAt: now,
payload: {
...task,
status: "cancelled",
completedAt: now,
updatedAt: now,
completionDelivery: { state: "disposed", observedByRunId: null },
},
});
nodes ??= (yield* projectionStore.getThreadRecords(command.threadId, ["nodes"])).nodes;
const node = nodes.find((candidate) => candidate.id === task.id);
if (node !== undefined)
yield* emitEvent({
type: "node.updated",
threadId: command.threadId,
nodeId: node.id,
occurredAt: now,
payload: { ...node, status: "cancelled", completedAt: now },
});
for (const item of projection.turnItems) {
if (item.type !== "subagent" || item.subagentId !== task.id) continue;
yield* emitEvent({
type: "turn-item.updated",
threadId: command.threadId,
nodeId: task.id,
occurredAt: now,
payload: { ...item, status: "cancelled", completedAt: now, updatedAt: now },
});
}
}
}).pipe(
Effect.mapError((cause) =>
Schema.is(OrchestratorProjectionError)(cause)
? cause
: new OrchestratorProjectionError({ threadId: command.threadId, cause }),
),
);
const latestRun = projection.runs.findLast((run) => run.status !== "queued");
const target =
projection.runs.findLast(
Expand Down Expand Up @@ -9754,7 +9764,7 @@ const makeOrchestrator = Effect.fn("orchestrationV2.Orchestrator.layer")(functio
const enqueued = yield* Ref.get(interruptEffects);
yield* Ref.update(events, (existing) => [...existing, ...written]);
yield* Ref.update(effects, (existing) => [...existing, ...enqueued]);
yield* cancelIdleOrphans;
yield* cancelIdleOrphans({ command, events, projection });
return interrupted.value;
}
yield* Effect.logWarning("thread.stop could not interrupt the running turn", {
Expand All @@ -9775,7 +9785,7 @@ const makeOrchestrator = Effect.fn("orchestrationV2.Orchestrator.layer")(functio
yield* markStoppedRun({ command, events, thread: projection.thread, run, now });
}
yield* holdStoppedThread({ command, events, projection, cohortRunIds: [], now });
yield* cancelIdleOrphans;
yield* cancelIdleOrphans({ command, events, projection });
return undefined;
});

Expand Down Expand Up @@ -11100,6 +11110,25 @@ const makeOrchestrator = Effect.fn("orchestrationV2.Orchestrator.layer")(functio
case "run.interrupt":
cancelUnsettledEffects = yield* dispatchRunInterrupt(command, events, effects);
if (command.holdQueue === true) {
// Stop must still commit when a child can't be read, so a failed repair's
// partial events are dropped.
const repairEvents = yield* Ref.make(yield* Ref.get(events));
const repaired = yield* Effect.exit(
loadProjectionForCommand(command, ["subagents", "turnItems"], {
turnItemTypes: ["subagent"],
turnItemStatuses: ["pending", "running", "waiting"],
}).pipe(
Effect.flatMap((projection) =>
cancelIdleOrphans({ command, events: repairEvents, projection }),
),
),
);
if (Exit.isSuccess(repaired)) yield* Ref.set(events, yield* Ref.get(repairEvents));
else
yield* Effect.logWarning("Stop could not cancel orphaned delegated follow-ups", {
threadId: command.threadId,
cause: repaired.cause,
});
yield* Ref.update(effects, (existing) => [
...existing,
{
Expand Down
Loading
Loading