Skip to content

Modal Mission author Activity proof — 2026-07-26

Status: A4 passed against the supported runtime path and real Modal sandboxes.

Result

Archetype source ae4eae3d ran one disposable coding-agent Mission through the Activity-backed Modal author path. The public fixture was everettVT/archetype-modal-proof-20260726-a4-1, a disposable proof repository that has since been deleted; the identities below remain the recorded evidence.

The base branch remained at 9a94713a93e3f4f49cb5129259faa87dc1a11296 with proof.txt equal to unproved. The author changed only that file, both exact-revision validators passed, and the proof branch published commit 67ed4f0cbc20e1e024de8ad049f804e458fa4ea6 with proof.txt equal to modal activity proved.

Durable ordering

Evidence Value
World 019f9fc7-a7da-71d0-82b1-fe6a3f48390b
Run 019f9fc7-a7da-71d0-82b1-fe7be9782c96
Dispatch source tick 1
Activity observation/settlement tick 2
Final world head 7
Activity ID 70a37d8f972f2494837f9dba8364cbb418b203558cbe0166f69ae925bb544f2d
Activity result digest 695a5e337e8b83d0c65df453265801e7559a0c238973138e1d76a7054c408eb9
Complete fact-bundle digest 74eeeeb68bd74de069fd4d25aa4f49081f48e90bd1f503a3add0a226df259568
Attempt / fence 1 / 1

The Activity catalog retained the immutable tick-1 admission, exact provider operation, bounded result reference, and tick-2 settlement. The tick-2 CompleteAuthorActivityObservation bound one sandbox, one author execution, two validations, one commit, one candidate, eight provenance relations, and zero friction facts to the same result digest. The task became accepted only after a separate critic approved the exact base-to-head diff.

Provider evidence

Evidence Value
Modal workspace / Environment vangelis-tech / main
Modal App ap-13Icl2bVNWkY4672goRJfg (archetype-agent-missions-a4-proof)
Author sandbox sb-BBHOAIlQCNBeqIPLXXN7c8
Critic sandbox sb-YucMchzdILbWDmXMe6S9MN
Provider operation missions.author:38326552c51ddeaee066345dac3c2bcd00d897f874b4271ce70906ef9f6c80a9
First-result Dict arc-author-results-v1-askxhkteilmybcgk55f7hj2uscamhehz
Candidate digest 8963804e3450f0452e82248a95079836b55cdf4c3fdc9c190a006df9872281a3
Diff digest 79ddc92ae2d8dec5ed220217a1e36aaf73d1c9383da620c1d1374af462ae6b21
Validator bundle digest 06f1d7a490ea9b1473068ca9a2e4e099105dfc09b3ec9ad672fc3027463e8efa

The author and critic sandbox identities differ. The critic receipt reviewed the exact candidate head, diff digest, validator bundle digest, and policy digest and concluded approved with zero findings.

Cold-process recovery

After the Mission and its runtime closed, a separate Python process reconstructed the Modal adapter from the durable Activity request and provider operation identity. reconcile() read the exact first-result Dict entry and returned the original author sandbox, final revision, and result digest. It did not call the author harness or start another sandbox.

The branch was then built as a wheel and installed into an empty virtual environment. From /tmp, with no source checkout on the import path, that installed artifact reconstructed the same request and adapter and recovered the same provider operation, result digest, and final Git revision.

Fault found by the proof

The first attempt stopped before provider execution because the receipt-pinned Mission reader returned append-only history rather than the state at the receipt's latest tick. Validators and relations therefore appeared twice at tick 1. A3 now applies the lazy latest projection before materialization and has a two-tick regression. The failed attempt left the committed tick pending, failed shutdown closed, and created no Git proof branch or provider result.