[MVP] First automated backend release: plan, dependencies and hosted drill #555

First automated MVP — master plan

This is the current release checklist for one real Terminal-Bench-style task-authoring project:

approved guide + locked policies -> task import/screen/release -> claim/start -> ZIP PRE feedback or verified immutable Submission -> automatic POST -> governed correction or shared FinalAcceptance -> submitter ContributionRecord and applicable conditional awards.

Use an explicitly approved human_review_required=false project. Preserve default-true guards, but live human review is outside this first milestone. This is narrower than the roadmap's full human-review v0.1 release. No task-claim lease/skip, frontend, external payments, generic agent judging (#492), expanded evaluator CI (#496), reputation or unrelated cleanup is required here.

Delivered foundations — do not rebuild or wait for them

• Public guide setup/approval/activation and bounded local connected journey ([PILOT] Connect public guide setup to a governed contribution outcome #552/Activate supported automatic acceptance and prove public contribution journey #550).
• Task import/screen/release, optional task-local criteria ([PILOT-02] Import, screen and release task batches atomically #489/fix(tasks): preserve optional criteria through screening #574).
• Public ZIP intake/Submission and contribution exports (Expose admission-backed public Submission intake #549/[PILOT-10] Stable contribution export pagination #497).
• Checker-driven revisions, immediate exhaustion, timeout and reclaim ([PILOT-08] Revision caps, guarded rebases, false-policy activation #495/[PILOT-08] Close exhausted assignments and conserve paused revision time #573). PILOT-08 is merged and is not a waiting dependency.

These are implementation/local-proof facts, not complete current-head hosted certification. The historical connected drill used recorded inference, locally signed Flow-compatible identities, MinIO and development checker containers.

Merged implementation — do not rebuild

• [MVP] Expose authorized checker registration and acceptance lifecycle controls #580 — supported checker registration/readback and lifecycle controls, delivered by Expose authenticated checker registration and readback #589; issue closed.
• [MVP] Return actionable intake feedback for unreadable ZIP uploads #584 — malformed-ZIP actionable feedback, delivered by fix(artifacts): return feedback for unreadable ZIPs #591; issue closed.
• [PILOT] Confirm manager agreement on the prepared Terminal-Bench corpus #576 technical assets — original guide, five task briefs/examples and calibrated corpus (Prepare the real terminal benchmark pilot corpus #587).
• [PILOT-06] Finish real-sandbox negative and resource proofs for Terminal-Bench checks #493 implementation — project PRE and offline build/reference/no-op/stability POST, development success/interruption proof (feat(checkers): add terminal benchmark project checks #594).
• [MVP] Complete uncertain-delivery inspection and recovery #581 known-terminal recovery — public authorized discovery/retry and exact replay (Expose supported Operator checker recovery #588).
• [MVP] Finish AWS S3 activation repairs and live custody proof #582 source foundation — guarded AWS provider and immutable activation proof custody (Activate AWS S3 artifact storage behind live proof #592); remaining source findings are recorded in [MVP] Finish AWS S3 activation repairs and live custody proof #582.
• [MVP] Deploy and verify the hosted API, workers and gVisor runtime #583 release preparation — pinned images/configuration, startup checks and image-build/inspection proof (Add bounded hosted release configuration #590).
• Per-finding evidence and attempt outcomes ([PILOT-READ-1] Stored attempt outcomes and per-finding checker evidence #575), contributor reads ([PILOT-READ-2] Contributor work reads and task-list filters #577), and Project Manager/Finance reads ([PILOT-READ-3] Project Manager and Finance reads C1–C9 and D1 #595).

Remaining checklist — existing scope only

• [PILOT] Confirm manager agreement on the prepared Terminal-Bench corpus #576 — explicit manager agreement with the proposed guide/mechanical scope and offline material/profile; technical corpus preparation is complete.
• [PILOT-06] Finish real-sandbox negative and resource proofs for Terminal-Bench checks #493 — real sandbox negative-container and resource-enforcement cases still absent from retained runtime evidence.
• [MVP] Complete uncertain-delivery inspection and recovery #581 — supported uncertain-delivery inspection and source-specific request/completion/acceptance-ack-loss reconciliation, then interruption proof.
• [MVP] Finish AWS S3 activation repairs and live custody proof #582 — three merged-source activation corrections, authorized live AWS access/custody and retention/restore proof.
• [MVP] Deploy and verify the hosted API, workers and gVisor runtime #583 — actual authorized deployment, identity/runtime configuration, measured isolation/limits and operational recovery/backup proof.
• [PILOT-12] CLI submit, replay status, and contributor findings #499 — contributor submission/history revision-summary display. Main still prints unavailable in the current public API in those two commands; task-context support does not finish this presentation. REST rehearsal can proceed independently.
• Run and record the local real-project rehearsal below, using the merged corpus and public operations.
• Repeat on the intended hosted stack and satisfy the release criteria below.

Parallel work and order

Use the merged owners and corpus. Remaining checker negative/resource proof, recovery reconciliation, and AWS repair/environment preparation can proceed independently in bounded work. The CLI owner retains #499. Public local rehearsal does not wait for cloud deployment. Final hosted proof consumes the reviewed repairs, registered exact images, approved guide/policies and authorized deployment. #492 and #496 remain held and are not prerequisites for this first milestone.

Drill sequence

1. Freeze a shareable guide, meaningful example and 5–10 real task-authoring briefs/packages with independent expected outcomes. Task-local criteria remain optional; guide-level requirements are mandatory. Do not commit private prior-project documents.
2. Rehearse locally: live guide inference -> manager inspection/correction -> separate PRE/POST approvals -> Finance policy publication/selection -> guide activation -> task import/release -> actual contributor claim/start -> real checks -> contribution evidence. Use public APIs/CLI and ordinary durable workers; bootstrap only through supported administrative operations. MinIO local evidence is not AWS proof.
3. Repeat the same corpus on the intended hosted stack with actual Flow identities, AWS S3 and gVisor. Pin source/images/policies; begin with one success and one real failure/correction, then execute the matrix.
4. Preserve failed observations. Each discovered defect gets an owner-scoped issue/repair and a focused regression that catches it; rerun that branch and the final connected journey. Never retry until green or edit product rows to pass.

Required evidence matrix

• Original guide bytes and live setup read are project-scoped; unsupported requirements cannot be silently dropped. Separate manager approvals and complete locked context are enforced before READY.
• Invalid ZIP/default/PRE failure gives actionable feedback, creates no Submission and retains the valid working assignment. Warning-only input stays nonblocking.
• Successful intake preserves server-verified exact bytes/hash/size/manifest and locked policy lineage through durable POST and acceptance.
• POST work failure exposes useful findings; acknowledged corrected submission succeeds with predecessor/history retained. Independently adjudicated bad packages fail the intended mechanical checks.
• Zero and positive revision caps, immediate exhaustion, one timeout/pause-resume case and fresh reclaim use PostgreSQL time and preserve every earlier attempt.
• Infrastructure failure spends no contributor revision; exact authorized retry recovers the original work with bounded attempts.
• Interrupt before invocation, after execution and after acceptance commit/before acknowledgement. Duplicate delivery and lost responses cannot repeat business effects. UNKNOWN is explicitly reconciled as committed, proven no effect or unresolved, never blindly retried.
• Wrong actor/project/resource, revoked authority and stale replay contexts deny without unauthorized effects or data disclosure.
• Public contribution evidence matches one FinalAcceptance, exact submitter record and frozen applicable award set; no fabricated Review/reviewer contribution or claim of payment fulfillment.
• Hosted resource/isolation, diagnostic privacy/retention, collector outage, drain/restart, supported rollback/forward recovery and backup/restore have actual retained proof.

Close/release criteria

Close only after the exact deployed candidate has required CI/reviews and real-project public-runtime evidence above, no unresolved blocker in that journey, and documented operating limits/recovery procedures. Bind evidence to source/image/corpus hashes, policy/activation IDs, task/assignment/submission/checker/outcome/contribution identities and cleanup results. GitHub-hosted CI is not hosted-product deployment proof.

This first release proves agreed mechanical task validity and the governed lifecycle, not automated difficulty, solution-leakage or general quality judging. Do not close merely because child PRs merged. Human approval, deployment authorization and PR merges remain explicit separate actions.

Current main reconciliation: 51b868071 (including merged #587, #588, #589, #590, #591, #592, #594 and #595). Before each implementation PR, reconcile current main, the smallest Commitrail record, affected canonical specifications/roadmap, risk/reviewer scope and discriminating proof. Reconcile immediate MVP vs later full-v0.1 navigation in affected changes; do not invent a documentation-only prerequisite or silently delete later requirements.