[richiemcilroy]

Loom import workflows used to poll the database every 5 to 30 seconds until their media job finished. Each poll is a workflow step plus a replay on Vercel, so a ten minute import cost about 25 invocations, and a batch of 1,200 imports cost tens of thousands.

Now each processing attempt creates a workflow hook whose token goes into the worker's webhook URL. When the worker reports complete, error or cancelled, the progress route saves the state first and then resumes the hook. The workflow then reads the status once. A typical import now uses two status reads instead of dozens.

Safety nets:

• If the signal is lost, the workflow still checks every 5 minutes, within the same 1 hour limit.
• If the signal arrives before the saved state is visible, it falls back to the old short polling.
• Tokens only resume hooks that belong to the same video. Unknown or disposed tokens are ignored.
• Each hook is disposed when its attempt ends.
• Runs already in flight stay on the deployment that started them.

Recording, instant and web uploads still use waitForVideoProcessing and are unchanged.

Follows #3118. Replaces #3119, which closed when its base branch was deleted.

Tests: video-processing-status 15, media-server-progress 59, media-processing-workflows 27 pass. The 7 share-stage-identity failures also fail on main.

Confidence Score: 5/5

The PR appears safe to merge; no actionable issues were found.

Summary

Loom imports now wait for a worker webhook instead of repeatedly checking the database.

• Loom imports wake when their media worker reports a final state.

_{Reviews (2) · Last reviewed commit: "perf: wake Loom import workflows from th..." · Reviewed by Greptile}

[richiemcilroy]

@greptileai

[capy-ai]

Capy found no issues.