CI perf: workflow_dispatch of CI — Windows-only janitor fixes re-run the full Linux + macOS nightly matrix (~100 min/day, ~30 macOS) #1818

Measurement

Window: 7 days to 2026-10-29 08:17 UTC. There were 6 workflow_dispatch runs of CI. 5 were on ci-janitor/agent branches that fix a Windows-only nightly failure: cargo-home-windows-test, fix-ci-go-gowork-windows-tests, fix-ci-windows-yarnrc-escaped-path, windows-nightly-path-asserts and uv-settings-test-windows. The 6th was pylock-follow-uv.

A dispatch runs the whole nightly matrix. Here are the two dispatches from this morning:

In this 4-hour slice those two dispatches used 194 Linux + 83 macOS job-min. That is more than all 10 full PR CI runs together (370 Linux). Neither branch changes anything that runs on Linux or macOS, and the PR's own CI run already covers the Linux tiers.

Where the time goes

Per dispatch: ~97 Linux min in 91 jobs (the full e2e tier, the yarn/cargo-vex full matrices and coverage), ~41 macOS min (test (macos-latest, 1/2) 16.6–18.3 min each, e2e-build-macos 2.9–5.9, yarn-berry (macos) 2.5–3.0) and ~50 Windows min. Only the 4 Windows jobs (e2e-build-windows, test (windows-latest, 1/2), yarn-berry 4.12.0 (windows-latest)) test the fix.

Root cause

The only dispatch input ci.yml has is hosted_e2e. Every nightly-tier job is gated on github.event_name == 'workflow_dispatch' (for example ci.yml lines 365, 562, 825, 957, 1179, 2107, 2169, 2234, 2545), so a dispatch can't pick an OS.

Proposed fix

In .github/workflows/ci.yml:

1. Add a dispatch input os: choice [all, windows, macos, linux], default all.
2. In each job's if:, replace the bare github.event_name == 'workflow_dispatch' term with an OS-aware one. Windows jobs (test windows matrix rows, e2e-build-windows, e2e-windows, cargo-vex-matrix-windows, the windows yarn-berry row) take (github.event_name == 'workflow_dispatch' && contains(fromJSON('["all","windows"]'), inputs.os)), and likewise for macOS and Linux. Matrices that mix OSes (test, yarn-berry-e2e) filter their include on inputs.os the same way CI_SCOPE already picks the include list.
3. Have the ci-janitor routine / agent instructions dispatch with -f os=windows for Windows-only fixes.

schedule, push, pull_request and merge_group are unchanged. ci-ok already treats skipped jobs as passing, and dispatch runs aren't required checks.

Expected saving

• Per Windows-only dispatch: ~97 Linux + ~41 macOS job-min, ~95 Depot/hosted job starts and 4 macOS runner slots.
• At the 7-day rate (5 Windows-only dispatches, ~0.7/day): ~70 Linux + ~30 macOS job-min/day. On days when the janitor is iterating on a Windows nightly break (2 in 25 min this morning) it's ~190 Linux + ~80 macOS.
• Wall clock doesn't change: the Windows test shards (16–21 min) set the length of a dispatch run.

Coverage and risk

Nothing loses coverage. The nightly (schedule) still runs every job on every OS, and os: all (the default) keeps today's dispatch behavior. A janitor fix that turns out to break Linux or macOS is still caught by its PR CI run, the merge queue, and the nightly. Required check names ci-ok / clippy don't change. The risk is a typo in an if: that silently skips a job on schedule. Check with a workflow_dispatch of os: all and compare its job count (99) before merging.

Effort

S: one input and ~12 if: / matrix-include edits in one file, plus one line in the janitor instructions.

ROI

Saving ≈ 0.07k Linux ×1 + 0.03k macOS ×3 = 0.16 weighted k-min/day; confidence 0.6 (the dispatch rate swings with Windows nightly breaks); effort 1 → ROI ≈ 0.1.

Generated by Claude Code