[Mikola Lysenko (mikolalysenko)]

LLM Description written by Claude Code:claude-opus-5-5

Fixes #1818

What

ci.yml gets a workflow_dispatch input os (all | linux | windows | macos, default all). A fix for a Windows-only nightly break can now be checked with

gh workflow run ci.yml --ref <branch> -f os=windows

instead of re-running the ~91 Linux and 4 macOS nightly-tier jobs that its PR run and the merge queue already cover.

Root cause

Every nightly-tier job is gated on github.event_name == 'workflow_dispatch' with no way to pick an OS, so each dispatch ran the whole nightly matrix (issue: ~97 Linux + ~41 macOS job-min per Windows-only dispatch, 5 such dispatches in the last 7 days).

How

• Each job's if: becomes (<unchanged condition>) && (<os term>). Off dispatch inputs.os is empty, so every inputs.os != '...' term is true and pull_request, merge_group, push and schedule evaluate exactly as before.
• Linux roots (node-addon, lint-ecosystems, test-release, coverage, docker-base, dispatch-tests, e2e-build, yarn-classic-matrix, yarn-berry-full, cargo-old-toolchains, hosted-e2e) skip for windows/macos; their dependents (e2e, e2e-extended, e2e-full, e2e-docker, cargo-vex-matrix*, coverage-docker/merge) skip with them.
• e2e-build-windows skips for linux/macos (e2e-windows, cargo-vex-matrix-windows follow); e2e-build-macos and yarn-berry-e2e-macos skip for linux/windows.
• Mixed matrices: test skips for linux and excludes the other OS's legs; yarn-berry-e2e skips for macos and its full-scope include narrows to the Windows row or the three Linux rows.
• clippy, release-readiness and ci-ok always run (every OS leg needs the first two). windows-compile-check never ran on dispatch and is untouched. ci-ok already counts skipped jobs as passing; required check names don't change. Depot mappings and the DISABLE_DEPOT_RUNNERS switch are untouched.

The ci-janitor routine's instructions live outside this repo; they should dispatch Windows-only fixes with -f os=windows once this lands.

• A one-OS dispatch reports its verdict as ci-ok (<os> only), not ci-ok: otherwise a Windows-only dispatch on a PR head would post a green required ci-ok while the PR's Linux jobs were red (a skipped required check would count as passing too). Every other run, os=all dispatches included, still reports ci-ok.
• OS-filtered dispatches get their own concurrency group (...-windows), so dispatching one per OS on a branch doesn't make the third replace the pending second.

CI cost

• Lean PR / merge-queue run: unchanged, ~28 Linux jobs, ~8 min (no condition changes off dispatch).
• Nightly: unchanged (99 jobs).
• workflow_dispatch -f os=windows: from ~99 jobs (~97 Linux + ~41 macOS + ~50 Windows job-min) to the Windows jobs plus clippy, release-readiness and ci-ok (~50 Windows + ~5 Linux job-min). os=all (default) keeps today's behaviour.

Validation

• python3 -m pytest scripts/tests -q: 372 passed, 6 skipped.
• New test_ci_scheduling.py::test_dispatch_os_runs_only_that_os walks the job graph and checks, for each OS value, that exactly the jobs whose matrix/runner targets that OS run (plus the always-on preflights), and that the mixed matrices narrow. It fails on main's ci.yml and passes here.
• Existing scheduling/tier tests now compare the if: without the dispatch-only OS term (job_if), so their assertions about PR/queue/push gating are unchanged.
• YAML parses; all 8 fromJSON literals in ci.yml parse as JSON.
• actionlint / zizmor are not installed in this session and installing them was not permitted; CI's own lint runs instead.
• No Rust or platform-specific code changes, so no full-scope dispatch.

/code-review high

Fixed: one-OS dispatch posting a green required ci-ok (renamed for those runs, new test test_one_os_dispatch_cannot_post_a_green_ci_ok); per-OS dispatches sharing one concurrency group (group now includes the OS, asserted in test_late_draft_run_cannot_cancel_the_ready_run); the tier test accepting either form of yarn-berry-full's gate (now requires the gated form); a figure in the input comment that mixed jobs and job-minutes.

Not changed:

• hosted_e2e=force with os=windows|macos skips hosted-e2e. It's a Linux job, and the os input's comment now says so; making force override os would add a sixth gate variant for a combination nobody uses.
• "Compute the OS selection once in release-readiness outputs": only the root job of each OS carries a term (dependents skip through needs), so it's 16 one-term edits; moving it into job outputs would make each gate depend on a script and not be readable from the if: alone.

https://claude.ai/code/session_011f9aePNkpUez9XyWCNA48w

---

Generated by Claude Code

[Mikola Lysenko (mikolalysenko)]

BugBot review

---

Generated by Claude Code

[Mikola Lysenko (mikolalysenko)]

Burn-down agent: labeled Ready for review at 39f6776f7b.

• CI: 83/83 checks on this head pass or are skipped (33 success, 50 skipped), ci-ok green.
• Bugbot: reviewed this head (check run success), no findings; no open review threads.
• Reviewer focus: the os dispatch input in .github/workflows/ci.yml. Every job's if: gains an inputs.os term that is true when inputs.os is empty, so pull_request, merge_group, push and schedule runs should behave as before. One-OS dispatches report ci-ok (<os> only) so they can't post a green required ci-ok.
• Slack: not announced this run (send tool unavailable); the next run will retry.

---

Generated by Claude Code

[Mikola Lysenko (mikolalysenko)]

Final review brief

What it does. Adds an os input (all/linux/windows/macos, default all) to ci.yml's workflow_dispatch, so a Windows-only nightly fix can be checked with -f os=windows without re-running ~95 Linux/macOS jobs. Each OS's root job gets an inputs.os != ... term. Off dispatch, inputs.os is empty, so pull_request, merge_group, push and schedule runs are unchanged.

Risk: low. CI config only, no Rust changes. The only change that touches the required check is that ci-ok's name: is now an expression. It evaluates to ci-ok whenever inputs.os is empty or all, and this PR's own pull_request run posted a green ci-ok under that name.

Look here

• ci.yml:32: the new input
• ci.yml:2905: a one-OS dispatch reports ci-ok (<os> only), so it can't post a green required ci-ok on a PR head
• ci.yml:62: per-OS concurrency group
• ci.yml:399 and ci.yml:2391: the two mixed-OS matrices narrowed by OS
• test_ci_scheduling.py:247: walks the job graph per OS value

Verified

• Read the full diff and the job graph. Every Linux root (node-addon, lint-ecosystems, test-release, coverage, docker-base, dispatch-tests, e2e-build, yarn-classic/berry, cargo-old-toolchains, hosted-e2e) is gated. Their dependents skip through needs (none use always()). e2e-build's matrix is ubuntu-only, test has no ubuntu leg, and coverage-merge skips with coverage.
• inputs is an allowed context in workflow-level concurrency and in job name.
• python3 -m pytest scripts/tests: 373 passed, 6 skipped. With main's ci.yml, the 3 new or changed scheduling tests fail, so the tests do exercise the change.
• CI: 83/83 on this head (33 success, 50 skipped). Bugbot: success, no findings. No open threads. CHANGELOG.md untouched.

Changes I made: none.

Open questions: none. Once this lands, the ci-janitor routine's instructions (kept outside the repo) should dispatch Windows-only fixes with -f os=windows.

Auto-merge is armed, so approving sends it straight to the merge queue.

---

Generated by Claude Code