github/AutoGPT
1
1
Fork 0
mirror of https://github.com/Significant-Gravitas/AutoGPT.git synced 2026-02-24 03:00:28 -05:00
Code Issues Packages Projects Releases Wiki Activity

fix(copilot): prevent double output from StreamFinish/mark_task_completed race (#SECRT-2021)

Browse Source 
Requested by @0ubbe

The CoPilot consistently displays every response twice. Root cause is a
race condition between StreamFinish delivery and mark_task_completed:

1. Executor publishes StreamFinish to Redis stream
2. Frontend receives it, transitions to 'ready', invalidates cache
3. mark_task_completed runs AFTER StreamFinish (separate operation)
4. Session refetch sees task still 'running' → spurious resumeStream()
5. Resume replays entire Redis stream from '0-0' → double output

Backend fix (processor.py): Intercept StreamFinish from the generator
instead of publishing it directly. mark_task_completed atomically sets
status to 'completed' THEN publishes StreamFinish, ensuring clients
never see the finish event while the task is still marked as running.

Frontend fix (useCopilotPage.ts): Only reset hasResumedRef on error
(SSE drop), not on clean stream finish. This prevents the spurious
resume even if the backend timing changes.
This commit is contained in:
Otto
2026-02-23 13:58:45 +00:00
parent a18ffd0b21
commit 8e8993be3c
2 changed files with 19 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

13
autogpt_platform/backend/backend/copilot/executor/processor.py
View File

@@ -257,10 +257,21 @@ class CoPilotProcessor:
cluster_lock.refresh()
last_refresh = current_time
# Intercept StreamFinish: don't publish it directly.
# mark_task_completed atomically sets status to "completed"
# and THEN publishes StreamFinish. Publishing StreamFinish
# before the status update causes a race where the frontend
# sees the stream as finished but the task is still "running",
# triggering a spurious resume that replays the entire stream
# (double output bug — SECRT-2021).
if isinstance(chunk, StreamFinish):
break
# Publish chunk to stream registry
await stream_registry.publish_chunk(entry.task_id, chunk)
# Mark task as completed
# Mark task as completed — this publishes StreamFinish AFTER
# atomically setting status to "completed", preventing the race.
await stream_registry.mark_task_completed(entry.task_id, status="completed")
log.info("Task completed successfully")

9
autogpt_platform/frontend/src/app/(platform)/copilot/useCopilotPage.ts
View File

@@ -208,8 +208,13 @@ export function useCopilotPage() {
queryClient.invalidateQueries({
queryKey: getGetV2GetSessionQueryKey(sessionId),
});
// Allow re-resume if the backend task is still running.
hasResumedRef.current = null;
// Only allow re-resume on error (SSE drop without clean finish).
// On clean finish (status === "ready"), the backend task is done —
// resetting the ref would allow a spurious resume if the session
// refetch races with mark_task_completed (SECRT-2021).
if (status === "error") {
hasResumedRef.current = null;
}
}
}, [status, sessionId, queryClient]);