Long-running delivery

Start in the intended checkout. anvil status --json reports its resolved State; an uninitialized replacement checkout does not establish that project history is absent. Inspect existing ownership and recorded attempts before claiming:

anvil status --json
anvil packet T001 --attempt --format json
anvil evidence-preflight T001 --json

For a bundle use anvil packet B001 --attempt --bundle --format json. The coordinator owns every member mutation; see coordinating a bundle. For a named PRD, retain its identity, for example anvil prd parse --prd release and anvil packet release:T001 --attempt --format json.

Attempt reads and evidence preflight never initialize State, reap claims, write packets, renew custody, or authorize submission. Preflight reports the current claim and generation, complete buffer digest and bytes, record/skipped counts, and literal required-command coverage. It preserves failed captures and refuses overflow instead of selecting a passing suffix. Its proof ceilings are 16 items and 1 MiB. Attempt responses are bounded to 64 KiB; bundle members share one State frontier and cumulative bounds. Default anvil packet T001 still writes the ordinary packet.

If complete inspection reports overflow, observe that the repository runner has stopped before releasing the rejected generation with anvil release T001. Preserve the original buffer and historical evidence. Inspect readiness and claim a fresh eligible generation; capture its literal required checks again. Do not trim, choose passing records, raise limits or relabel old captures. Bundle custody and replacement-generation rules remain coordinator-owned.

The repository owns its verification runner, prerequisites, environment checks, timeouts and process cleanup. Verification profiles freeze metadata and declared source bytes; Anvil does not execute a runner or install dependencies. Run the repository's supported preflight and verification commands, then submit actual captured evidence using the existing workflow:

anvil evidence-preflight T001 --json
anvil submit T001 --commands "pytest -q" --files-changed src/feature.py

Replace the example command and file with the task's literal requirements and actual changes. Submission re-inspects the buffer at the native mutation boundary; an earlier passing preflight cannot waive later drift. Profile guards check the actual prepared target at claim creation and the originating claim at submission and acceptance. Legitimate commits may advance that target after creation. Three independent adversarial reviews and the required immutable approval authority remain separate gates.

For an actual runner observation, write the schema-1 timing receipt with exact claim attribution and observed UTC/monotonic measurements, then ingest it:

anvil progress T001 tests --timing-file timing.json
anvil packet T001 --attempt --format json

The receipt must be a complete regular UTF-8 JSON file of at most 16 KiB. Timing requires the exact active ordinary owner and cannot combine with --attestation-file or --bundle. It is audit data: success, failure or interruption grants no proof, renewal, acceptance or ownership. Complete monotonic execution and interrupted observations remain distinct; overlapping UTC intervals use a union rather than summed execution time. Current age needs an explicit --observation-at UTC timestamp. Unknown runner stop, reservation release and external waits stay unknown; the view supplies no forecast.

Accepted-attempt correction records exact invalidation provenance while retaining historical acceptance and evidence bytes. An old signed proof may still verify cryptographically after native acceptance has been invalidated; consult the current State for acceptance validity. Exact retries cannot reopen a newer accepted generation. Replay does not consult current profile files. A new claim generation inherits no old proof coverage. Integration, PR/merge and deployment facts require recorded delivery evidence. Installed artifact checks, native platform checks, actual project pilots and final user validation must be completed separately; source tests do not establish them.