Repository navigation
Human action 124
Canonical work item: #124.
Status: Done (2026-09-23). v0.17.0 is tagged at cf371ea and published. Gate 3: run 35775710024, 15 of 15. Gate 10: passed on lab-windows-a with latency compared. The steps below are kept as the record of what was done.
When required: Before v0.17.0 is tagged and published. The issue carries current status, decisions, and completion evidence.
Owner role: Windows qualification-machine owner (lab-windows-a).
Eight of the ten release-gate items in docs/testing.md are already clear, recorded in PR #123 and on the issue: clean checkout, the NUL/binary-attribute scan over 388 tracked and untracked files, all six maintained demos, expected-failure recovery, no surviving process, version-constant agreement, and the packed-artifact install and smoke test outside the checkout. Gate 3 also passed: a manually dispatched full qualification, run 35739589896, 15 of 15 jobs across both platforms including the Node 20 floor. Two items remain, and both need this machine.
Prerequisites: Quiet Windows host with the lab-windows-a identity, a clean checkout, no competing suites, and platform defaults with no engine or session overrides.
Decided on 2026-09-22: re-qualify and tag after #122. The release candidate is cf371ea. PR #127 folded #122's entry into ## 0.17.0, so there is no Unreleased section, and v0.17.0 ships --squash/--fixup with the fix for the rebase --abort recovery defect rather than with the defect. The rationale is recorded on the issue.
Because of this, run 35739589896 (pinned to a4685aa) does not cover the candidate. Gate 3 and gate 10 both have to be met again at cf371ea.
- Check out
cf371eaand confirm the checkout is clean and the version constants agree:node bin/vlab.js versionmust printvcs-lab 0.17.0, andpackage.json, bothpackage-lock.jsonfields, andsrc/version.jsmust match. -
Gate 10, first. Run
npm run test:benchmark -- --host lab-windows-aon the quiet machine, before the CI dispatch (a count regression found here saves a wasted matrix run). Inspect the deterministic findings as well as latency. - Confirm the selected reference is
hosts.lab-windows-aandlatencySkippedis false. A latency skip, a deterministic-only pass, or a historical OS entry does not satisfy this gate. If a count has grown, investigate the cause before deciding whether to re-record; do not relax tolerances to pass. -
Gate 3. Dispatch the full qualification on
cf371ea(.github/workflows/ci.yml, manual dispatch — a routine PR or main run does not satisfy gate 3) and confirm every job passes (15 of 15) before continuing. - Tag
cf371eaasv0.17.0with an annotated tag and push the tag. - Publish the GitHub release for
v0.17.0, which closes the remaining half of gate 8.
Verification: an annotated v0.17.0 tag on the chosen commit, pushed to origin; a published GitHub release; and a passing lab-windows-a benchmark check with latencySkipped false. Gate 3's dispatched run must name the same commit the tag does.
Failure handling and cleanup: if the benchmark reports growth, preserve the output and investigate before re-recording — a Linux run cannot substitute, because #68 scoped latency qualification to Windows only. Benchmark fixtures normally clean themselves; remove only identified disposable leftovers after evidence is retained. If the tag is pushed to the wrong commit, delete it locally and on origin before re-tagging, and say so on the issue.
Known and expected: cst metadata validate may report 2 errors about missing note attachments. This depends on what the checkout has fetched, not on the release. The two attachments are ordinary compact-landed branch commits kept alive by refs/vcs-lab/retention; a clone that has not fetched refs/vcs-lab/* cannot see them and reports the errors, and a clone that has reports 0 errors. See #120. If the count is 2, run git fetch origin 'refs/vcs-lab/*:refs/vcs-lab/*' and re-check before treating it as a finding. Gate 5 asks for no unexpected errors; this is neither unexpected nor a regression from this release.
Completion: attach the benchmark evidence and the step 0 decision to issue #124, remove human-action-required when both remaining items are satisfied, update the private board, and point the Human TODO entry at the completed outcome.
Publish only evidence suitable for the public issue and wiki. Redact private paths, repository content, and credentials from shared output.