Repository navigation
Description
What would you like to see?
Tests and reusable-workflow changes should be exercised against a real disposable image before they are distributed to consumer qualification gates. Lint, feature dry-run and helper unit tests do not prove GNOME/AT-SPI startup, host-vs-container probes, bootc reboots or result propagation work.
Verified starting point
Baseline main 6d36c72f52ec891ed66a8d947073854338581e38: .github/workflows/pr-validate.yml and unit-tests.yml trigger on pull_request/merge_group/main pushes but do not boot a candidate VM; e2e.yml is workflow_call only. update-v1-tag.yml advances managed v1 after main merges. The live main rules inspected during this investigation required Lint & syntax, Behave dry-run and pytest. That proves the observed required GHA checks, not the absence of all external lab signals.
PR validation workflow and managed tag publisher.
Scope / safety
Audit existing ghost-lab/commit-status and GHA runtime validation first; reuse an actually working signal instead of creating competing gates or inferring health from merged fixes. Execute the PR's real tests and workflow revision using correct test_repository/test_ref and immutable candidate identity, not defaults that silently select main/v1. Cover changed GUI/SSH/lifecycle paths where applicable. Untrusted PR code must not receive publishing credentials or execute through a privileged pull_request_target checkout. Separate read-only VM validation from trusted publication.
Acceptance
- A reviewed change-path policy identifies the minimal representative VM jobs required for tests/shared, suite steps/hooks, runner, composite action and reusable-workflow changes; irrelevant docs-only changes need not boot.
- Actual runs record the PR/merge revision and image digest and demonstrate that modified workflow/test code, not released main/v1 code, executed.
- Setup failure, absent/zero required coverage or a deliberately broken assertion prevents qualification/distribution; an observed good candidate succeeds. Reuse the parent's fail-closed verdict contract.
- Existing external statuses are audited end to end and reconciled with docs; no claim that ghost-lab works/is dead based only on prose or an action check-run API.
- Demonstrate the path on a real PR and any required merge-group/publisher transition, with retained VM evidence. The outcome must be available before a failing test/workflow change reaches consumers.
- Full affected workflow/helper test modules and actual VM paths pass; caller refs and matching docs/skills are updated consistently.
Human decision boundary
Required status checks, branch rules, v1 publication policy and CI credential/interface changes need explicit maintainer approval. The user requested tracking toward blocking tests, not unilateral ruleset/permission changes. Persistent lab orchestration stays in projectbluefin/lab. Parent: #1021.
Activity
Danathar commentedon Oct 28, 2026
Triage: accepted. Maintainer-authored design issue, part of the #1021 rollout plan.
hivecommons-hive commentedon Oct 28, 2026
PR #1042 covers this issue: #1042
The PR body says this is partial/non-closing work (
Refs/Part ofor leave-open wording), so Hive will not add a closing keyword. GitHub will not show its issue-list PR icon without a closing keyword; the stage comment is the link.feat(ci): add change-path policy for candidate-VM validation
3 remaining items
Danathar commentedon Oct 28, 2026
/hive decision Option A: a post-merge gate that holds v1. On a push to main, run scripts/vm_validation_plan.py on the merged diff, call ./.github/workflows/e2e.yml with test_ref set to the merged SHA and a digest-pinned image, and move v1 only when the #1022 fail-closed verdict (scripts/assert_passed.py) passes. Record the SHA and image digest in the job summary. Two conditions: (1) keep a manual workflow_dispatch path to move v1 anyway, for emergencies; (2) when a failed verdict holds v1 back, surface it visibly (a failed run plus an issue or a comment that names the held SHA), so a stuck tag is never silent. No skip-screenshot-publish input on e2e.yml for now; only Option B would need it, and #1037/#1050 already edit e2e.yml.
Reason: consumers only use @v1, so gating the tag protects every image repo from a broken change, and only merged, trusted code runs on the VM, with no write tokens for PR code and no ruleset or permission changes. Option B would put a hand-approved, up-to-2-hour VM run in front of every fork PR, and most testsuite PRs come from first-time fork contributors. B can still be added later if A proves too loose.
hivecommons-hive commentedon Oct 28, 2026
Approved by @Danathar. Decision: Option A: a post-merge gate that holds v1. On a push to main, run scripts/vm_validation_plan.py on the merged diff, call ./.github/workflows/e2e.yml with test_ref set to the merged SHA and a digest-pinned image, and move v1 only when the #1022 fail-closed verdict (scripts/assert_passed.py) passes. Record the SHA and image digest in the job summary. Two conditions: (1) keep a manual workflow_dispatch path to move v1 anyway, for emergencies; (2) when a failed verdict holds v1 back, surface it visibly (a failed run plus an issue or a comment that names the held SHA), so a stuck tag is never silent. No skip-screenshot-publish input on e2e.yml for now; only Option B would need it, and #1037/#1050 already edit e2e.yml.
approved-directionis applied and Hive's ownstage/045-needs-decisionremoved. Anyholdlabel is left exactly as it was.hivecommons-hive commentedon Oct 28, 2026
Summary: I checked whether this is finished. It is not yet: the shared test tag that every image project uses is still moved forward on every change, with no test-machine run in front of it. The approved plan (option A) is still to be built, and the remaining change sits in a part of the repository that this agent is not allowed to change.
Verified on
main@f077951d:scripts/vm_validation_plan.py(feat(ci): list which real test-machine runs a change needs before it ships #1042) and the fail-closed verdict inscripts/assert_passed.py(fix(e2e): mandatory-coverage verdict blocks GNOME and KDE false green #1045) are both onmain..github/workflows/update-v1-tag.ymlstill force-pushesv1toHEADon every push tomain. It does not run the plan, calle2e.yml, check the verdict, offer the emergencyworkflow_dispatch, or report a held SHA. Those are the parts of the approved Option A that are left.All of the remaining work is under
.github/workflows/. This lane's App token is minted without the Workflows permission, so I cannot push it. It needs a human or an ISSUES_PRS_MERGE agent to land. I have not opened a PR. I addedhive/verified-open.— hive: agent=architect
🐝 Hive Agent:
architect| Instance:hosted-projectbluefin-knuckle-gjvq| SHA:9170788feat(v1): move v1 only after the merged tests boot (Option A) (#1055)
hivecommons-hive commentedon Oct 28, 2026
PR #1060 covers this issue: #1060
The PR body says this is partial/non-closing work (
Refs/Part ofor leave-open wording), so Hive will not add a closing keyword. GitHub will not show its issue-list PR icon without a closing keyword; the stage comment is the link.hivecommons-hive commentedon Oct 28, 2026
Fixed by #1055 (merged dadee6e). Closing. Reply /reopen if the problem persists.
fix(v1): hold v1 only for failures the tests at v1 do not have (#1060)