Omniweft · Engineering note
Omniweft keeps retry behind the authority boundary
Merged worker lifecycle and restart-expiry handling preserve terminal proposals and the exact prepared request. Uncertain delivery can recover a receipt; expired retention requires reconciliation.
Source: https://louijiecompo.com/updates/omniweft-worker-recovery/
| Phase | Status | World revision |
|---|---|---|
| timeout | timed_out | 0 |
| cancelled | cancelled | 0 |
| crashed | crashed | 0 |
| superseded | superseded | 0 |
| replacement | committed | 1 |
| restarted | committed | 2 |
Recorded statuses from the public archive. Prevented proposals leave the world at revision zero.
PR #21 integrates PR-015 worker lifecycle and cancellation. The parent keeps mutation authority while bounded workers produce proposals. Timeout, cancellation, crash and supersession before submission prevent mutation; late results are drained without becoming new proposals.
PR #22 preserves timed_out / DEADLINE_EXPIRED when restart first observes an expired ready proposal. Explicit retry retains the original prepared transaction and can recover its receipt without a duplicate edit. Expired retention requires reconciliation, not a replacement request key.
The README now records PR-015 as merged; yesterday's mismatch is already corrected upstream. This contract assumes the parent and volatile policy host survive. Separate native persistence does not establish supervisor or retry-host crash durability, OS sandboxing or model inference. The table below reads retained CPU records, not a new engine run.
Follow the evidence
Read source evidence ↗ (opens in a new tab)
Related reading
- Retrying AI Workers Without Duplicating Actions
- From Python Agent to Vulkan Frame: Designing a Bounded Control Path
Follow writing and engineering notes
Add this feed to your RSS reader to follow new articles and engineering notes. No signup or tracking on this site.
Follow via RSS