Stabilize Ward for safe autonomous burndown and verified releases

coilyco-ops commented

2026-07-28 02:25:02 +00:00

Member

Goal

Bring Ward from source-green but operationally yellow to an end-to-end dependable control plane. Ward must admit only trusted automation inputs, publish verifiable release artifacts, report its queue truthfully, and carry one real pull request through observable CI plus exact-commit QA.

This issue is intentionally the overview goal. The goal run should keep this issue as the coordination record, create or link one child issue per discrete code change, sequence those children by dependency, and post the final evidence here.

Why now

The 2026-07-27 snapshot shows several individually tractable gaps that combine into one unsafe operating state:

Ward main and the v0.823.0 promote, release, tap, and Scoop tasks report success.
#1571 documents a P0 actor-admission gap. Ward currently loses or does not enforce authorship across triage, prompt admission, reservations, WARD-* outcomes, QA, and merge evaluation. The immediate fleet containment is coilyco-flight-deck/agentic-os#766. Supporting broker provenance work is tracked at coilyco-flight-deck/cli-guard#261.
The stable v0.823.0 release reports platform assets of only 279 to 285 bytes and a 268-byte checksum file, while the same commit's staged draft contains roughly 13 to 15 MB binaries and a 512-byte checksum file. The strongest current explanation is that release promotion copied Forgejo asset metadata documents instead of downloading the asset bytes. The release is at https://forgejo.coilysiren.me/coilyco-flight-deck/ward/releases/tag/v0.823.0.
The only open pull request, #1525, has a failing Forgejo Actions result, no available job log through the observed API path, and no QA verdict for its current commit.
ward agent dispatch-health reports four queued items, one failed item, and no running work. That count overstates actionable work because #1443 has later completion evidence after a false salvage reopen, and #1080 appears satisfied by the landed Ward operator split.
Ward ships an opt-in QA role and exact-commit QA merge evidence, but one remaining review path still describes brokered QA as pending. The wording no longer matches the shipped boundary.

Ordered workstreams

1. Contain untrusted automation input

The goal run must land the default-deny containment from coilyco-flight-deck/agentic-os#766 before broad autonomous burndown resumes on externally writable trackers, unless the full Ward admission gate lands first.
Engineering must implement #1571 so only configured automation actors can author machine-state records.
Ward must keep unapproved external issue and comment content out of execution prompts.
Ward must bind collaborator approval to an immutable content snapshot and invalidate that approval after untrusted mutation.
Ward must fail closed when actor provenance is missing or ambiguous.

2. Repair and prove release artifact promotion

The goal run must create or link a focused Ward child issue for the release defect before implementation.
Engineering must make release promotion download the actual staged bytes, verify them against SHA256SUMS, upload them to the stable release, then verify the stable bytes again.
The release workflow must reject implausibly small platform binaries and metadata-shaped payloads.
The mirror workflow must prove that Forgejo and GitHub publish identical intended bytes when the mirror is enabled.
Verification must install from each supported public release channel and confirm the reported Ward version plus the expected sidecar behavior.

3. Exercise the real CI and QA contract

The goal run must give #1525 an explicit final disposition: merged after proof, or closed with a linked replacement or superseding commit.
Engineering must obtain a green Forgejo Actions result for the pull request's current commit without using repeated speculative pushes as a live debugging loop.
QA must post a structured pass for the exact green commit before a director merges the pull request.
Engineering must replace stale pending brokered QA language with wording that describes the current opt-in, exact-commit QA gate.
An operator must restore or document the supported read-only path to the failing job evidence if Forgejo continues to return no log for a completed failure.

4. Reconcile tracker and dispatch truth

The goal run must reconcile #1443 against its completion record and current green release state.
The goal run must reconcile #1080 against the landed operator-boundary and embedded-policy commits.
The goal run must inspect stale reservations on #1530 and #1550, then clear or redispatch each one from evidence.
The goal run must give blocked records #1524, #1526, and #1557 an explicit disposition.
Ward's dispatch-health and director queue views must agree about active, failed, blocked, and completed work.

5. Publish end-to-end proof

The goal run must post one final evidence comment on this issue.
That comment must link the actor-admission or containment proof, a correct stable release, the verified checksums and asset sizes, the decisive CI run, the exact-commit QA verdict, the pull request disposition, and before-and-after queue snapshots.

Acceptance criteria

An authenticated external reporter cannot self-promote an issue, forge WARD-* workflow state, forge QA, or satisfy a merge gate.
Ward rejects machine-state input when actor provenance is missing or untrusted.
The fleet denies autonomous burndown for externally writable trackers until Ward enforces trusted admission.
A new stable release contains executable-size platform binaries, not metadata documents, and every published asset matches its declared checksum.
Forgejo and GitHub release assets are byte-identical when mirror publication is enabled.
Each supported installation path installs the intended release and reports the correct version.
The pull request at #1525 has an explicit final disposition, and no merge occurs without green CI plus a QA pass bound to the exact commit.
Active documentation and review output no longer claim that brokered QA is pending.
Dispatch health contains no stale or false active records and agrees with the director queue.
The tracker records at #1443 and #1080 reflect landed reality.
Ward source validation passes through the repository-declared vet, lint, test, and pre-commit verbs.
Engineering updates docs/FEATURES.md only if a child changes a public feature boundary.

Safety and sequencing

The goal run must not restart broad autonomous burndown on externally writable trackers before containment or full actor admission is effective.
Engineers and QA may inspect read-only CI, release, and runtime evidence. An operator must perform any live mutation or diagnostic action beyond observation.
Engineering must prove the release fix with local tests and API fixtures before one grounded push. The team must not use repeated release pushes to probe live CI or registry behavior.
Every child must preserve the AOSGuard operator boundary. No child should restore retired Ward operator commands.
The goal run must preserve one issue per discrete additive change and link every child back to this parent.
No participant may force-push, expose credentials, inspect raw customer payloads, or mutate live workloads as part of this goal.

Non-goals

This goal does not implement the dynamic advisor, ops, or PM role system tracked at #1328.
This goal does not perform a general v1 cleanup or product redesign.
This goal does not replace AOSGuard as the operator command surface.
This goal does not burn unrelated Ward backlog.

Codex, via advisor surface

## Goal Bring Ward from source-green but operationally yellow to an end-to-end dependable control plane. Ward must admit only trusted automation inputs, publish verifiable release artifacts, report its queue truthfully, and carry one real pull request through observable CI plus exact-commit QA. This issue is intentionally the overview goal. The goal run should keep this issue as the coordination record, create or link one child issue per discrete code change, sequence those children by dependency, and post the final evidence here. ## Why now The 2026-07-27 snapshot shows several individually tractable gaps that combine into one unsafe operating state: * Ward `main` and the v0.823.0 promote, release, tap, and Scoop tasks report success. * https://forgejo.coilysiren.me/coilyco-flight-deck/ward/issues/1571 documents a P0 actor-admission gap. Ward currently loses or does not enforce authorship across triage, prompt admission, reservations, `WARD-*` outcomes, QA, and merge evaluation. The immediate fleet containment is https://forgejo.coilysiren.me/coilyco-flight-deck/agentic-os/issues/766. Supporting broker provenance work is tracked at https://forgejo.coilysiren.me/coilyco-flight-deck/cli-guard/issues/261. * The stable v0.823.0 release reports platform assets of only 279 to 285 bytes and a 268-byte checksum file, while the same commit's staged draft contains roughly 13 to 15 MB binaries and a 512-byte checksum file. The strongest current explanation is that release promotion copied Forgejo asset metadata documents instead of downloading the asset bytes. The release is at https://forgejo.coilysiren.me/coilyco-flight-deck/ward/releases/tag/v0.823.0. * The only open pull request, https://forgejo.coilysiren.me/coilyco-flight-deck/ward/pulls/1525, has a failing Forgejo Actions result, no available job log through the observed API path, and no QA verdict for its current commit. * `ward agent dispatch-health` reports four queued items, one failed item, and no running work. That count overstates actionable work because https://forgejo.coilysiren.me/coilyco-flight-deck/ward/issues/1443 has later completion evidence after a false salvage reopen, and https://forgejo.coilysiren.me/coilyco-flight-deck/ward/issues/1080 appears satisfied by the landed Ward operator split. * Ward ships an opt-in QA role and exact-commit QA merge evidence, but one remaining review path still describes brokered QA as pending. The wording no longer matches the shipped boundary. ## Ordered workstreams ### 1. Contain untrusted automation input * The goal run must land the default-deny containment from https://forgejo.coilysiren.me/coilyco-flight-deck/agentic-os/issues/766 before broad autonomous burndown resumes on externally writable trackers, unless the full Ward admission gate lands first. * Engineering must implement https://forgejo.coilysiren.me/coilyco-flight-deck/ward/issues/1571 so only configured automation actors can author machine-state records. * Ward must keep unapproved external issue and comment content out of execution prompts. * Ward must bind collaborator approval to an immutable content snapshot and invalidate that approval after untrusted mutation. * Ward must fail closed when actor provenance is missing or ambiguous. ### 2. Repair and prove release artifact promotion * The goal run must create or link a focused Ward child issue for the release defect before implementation. * Engineering must make release promotion download the actual staged bytes, verify them against `SHA256SUMS`, upload them to the stable release, then verify the stable bytes again. * The release workflow must reject implausibly small platform binaries and metadata-shaped payloads. * The mirror workflow must prove that Forgejo and GitHub publish identical intended bytes when the mirror is enabled. * Verification must install from each supported public release channel and confirm the reported Ward version plus the expected sidecar behavior. ### 3. Exercise the real CI and QA contract * The goal run must give https://forgejo.coilysiren.me/coilyco-flight-deck/ward/pulls/1525 an explicit final disposition: merged after proof, or closed with a linked replacement or superseding commit. * Engineering must obtain a green Forgejo Actions result for the pull request's current commit without using repeated speculative pushes as a live debugging loop. * QA must post a structured pass for the exact green commit before a director merges the pull request. * Engineering must replace stale `pending brokered QA` language with wording that describes the current opt-in, exact-commit QA gate. * An operator must restore or document the supported read-only path to the failing job evidence if Forgejo continues to return no log for a completed failure. ### 4. Reconcile tracker and dispatch truth * The goal run must reconcile https://forgejo.coilysiren.me/coilyco-flight-deck/ward/issues/1443 against its completion record and current green release state. * The goal run must reconcile https://forgejo.coilysiren.me/coilyco-flight-deck/ward/issues/1080 against the landed operator-boundary and embedded-policy commits. * The goal run must inspect stale reservations on https://forgejo.coilysiren.me/coilyco-flight-deck/ward/issues/1530 and https://forgejo.coilysiren.me/coilyco-flight-deck/ward/issues/1550, then clear or redispatch each one from evidence. * The goal run must give blocked records https://forgejo.coilysiren.me/coilyco-flight-deck/ward/issues/1524, https://forgejo.coilysiren.me/coilyco-flight-deck/ward/issues/1526, and https://forgejo.coilysiren.me/coilyco-flight-deck/ward/issues/1557 an explicit disposition. * Ward's dispatch-health and director queue views must agree about active, failed, blocked, and completed work. ### 5. Publish end-to-end proof * The goal run must post one final evidence comment on this issue. * That comment must link the actor-admission or containment proof, a correct stable release, the verified checksums and asset sizes, the decisive CI run, the exact-commit QA verdict, the pull request disposition, and before-and-after queue snapshots. ## Acceptance criteria * An authenticated external reporter cannot self-promote an issue, forge `WARD-*` workflow state, forge QA, or satisfy a merge gate. * Ward rejects machine-state input when actor provenance is missing or untrusted. * The fleet denies autonomous burndown for externally writable trackers until Ward enforces trusted admission. * A new stable release contains executable-size platform binaries, not metadata documents, and every published asset matches its declared checksum. * Forgejo and GitHub release assets are byte-identical when mirror publication is enabled. * Each supported installation path installs the intended release and reports the correct version. * The pull request at https://forgejo.coilysiren.me/coilyco-flight-deck/ward/pulls/1525 has an explicit final disposition, and no merge occurs without green CI plus a QA pass bound to the exact commit. * Active documentation and review output no longer claim that brokered QA is pending. * Dispatch health contains no stale or false active records and agrees with the director queue. * The tracker records at https://forgejo.coilysiren.me/coilyco-flight-deck/ward/issues/1443 and https://forgejo.coilysiren.me/coilyco-flight-deck/ward/issues/1080 reflect landed reality. * Ward source validation passes through the repository-declared vet, lint, test, and pre-commit verbs. * Engineering updates `docs/FEATURES.md` only if a child changes a public feature boundary. ## Safety and sequencing * The goal run must not restart broad autonomous burndown on externally writable trackers before containment or full actor admission is effective. * Engineers and QA may inspect read-only CI, release, and runtime evidence. An operator must perform any live mutation or diagnostic action beyond observation. * Engineering must prove the release fix with local tests and API fixtures before one grounded push. The team must not use repeated release pushes to probe live CI or registry behavior. * Every child must preserve the AOSGuard operator boundary. No child should restore retired Ward operator commands. * The goal run must preserve one issue per discrete additive change and link every child back to this parent. * No participant may force-push, expose credentials, inspect raw customer payloads, or mutate live workloads as part of this goal. ## Non-goals * This goal does not implement the dynamic advisor, ops, or PM role system tracked at https://forgejo.coilysiren.me/coilyco-flight-deck/ward/issues/1328. * This goal does not perform a general v1 cleanup or product redesign. * This goal does not replace AOSGuard as the operator command surface. * This goal does not burn unrelated Ward backlog.  Codex, via advisor surface

Rows
Columns

Stabilize Ward for safe autonomous burndown and verified releases #1574