[Logan Ramos (lramos15)]

Makes pull request runs faster and frees scarce macOS and Windows runners. Three independent commits; reviewing them one at a time is easiest. The Linux cache warmer fix is separate, in #344837.

Why

Measured over about 450 successful Code OSS runs from Oct 26 to 28:

• Linux jobs set the run time. The median run takes 16.9 minutes. Linux / Remote and Linux / Electron are most often the last job to finish, because each runs two suites one after the other (about 11 minutes of tests).
• Failed runs keep using machines. Compile & Hygiene fails in about 5% of runs, at about 2 minutes in for hygiene and ESLint failures. The test jobs of those runs keep going and use about 10 macOS and 15 Windows runner hours a day.
• Windows cache misses cost 12 minutes per job. For about 22 minutes after every dependency change on main, pull request runs miss the Windows node_modules cache, and each Windows job installs the dependencies itself (about 12 minutes on the Dev Drive). The warming job installs on the OS disk, which takes 16 to 20 minutes, so it saves the cache late.

What changed

1. ci: split the slowest Linux PR test jobs (pr-linux-test.yml). Linux runners have no queue, so:

   • Electron becomes Electron (node.js) and Electron (extension host). A new INTEGRATION_TEST_SUITE env selects one, as UNIT_TEST_SUITE does for the Electron-Unit jobs.
   • Remote becomes Remote (integration) and Remote (smoke), as on Windows.
   • The Electron smoke tests use three shards instead of two, as on Windows.

   The longest Linux job drops from about 13.5 to about 9 minutes. The workflow tests and the skills that name the jobs and artifacts are updated.

2. ci: cancel the pull request run when Compile & Hygiene fails (pr.yml).

   • Compile & Hygiene reports whether one of its checks failed, as opposed to its setup, which a rerun fixes: type check, Markdown editor checks, hygiene/ESLint, compile, Codex sync, cyclic dependencies.
   • A new Cancel Run job then runs gh run cancel. It is separate so that pull request code never runs with an actions: write token.
   • Pull requests from forks get a read-only token, so their runs are not cancelled.
   • PR Gate lists the job, which its test requires, but ignores its result: Compile & Hygiene already fails the gate.

3. ci: install the Windows node_modules cache on a Dev Drive (pr-node-modules.yml). The Windows warming job moves its workspace to a Dev Drive with the same step as pr-win32-test.yml. A new test keeps the two steps the same.

Reviewer notes

• The check names Linux / Electron, Linux / Remote and Linux / Electron-Smoke (shard n/2) go away. PR Gate is the only required check, and it already covers every job of the Linux workflow.
• pr-node-modules.yml only runs on pushes to main, so this PR's checks don't exercise the Dev Drive change.
• Cancel Run only runs after a failing Compile & Hygiene, so a green run of this PR doesn't exercise it.
• This overlaps with files in ci: run macOS PR checks on three vscode-large-runners-macos-5-core runner groups #344796 (macOS runner groups), ci: cache Playwright browsers for pull request browser tests #343387 (Playwright cache) and the pnpm draft Engineering - move to pnpm12 #344791. The changes are small and should rebase easily.

How to test

• The Linux checks of this PR show the new jobs.
• Cancellation was checked on this PR with a temporary commit containing a type error (since removed). In run 40763077190:
  • Compile & Hygiene failed at 20:29:31 with Type 'string' is not assignable to type 'number'.
  • Cancel Run logged ✓ Request to cancel workflow 40763077190 submitted. 8 seconds later.
  • 35 jobs were cancelled, including every macOS and Windows test job.
  • PR Gate failed, reporting compile did not succeed (result: failure).

Locally:

• cd build && node --test lib/test/detectMissingCaches.test.ts lib/test/detectChanges.test.ts lib/test/agentHostTestSelection.test.ts: 33 passed, 0 failed. The full node --test 'lib/**/*.test.ts' also passed: 587 passed, 0 failed.
• npm run eslint on the changed test files passed, and npm run typecheck in build passed.
• actionlint 1.7.7 reports nothing new compared with main, apart from its known limits with YAML anchors and custom runner labels.

[Copilot]

🟡 Changes recommended

The cancellation probe assigns a string to a number, causing TS2322 and preventing Compile & Hygiene from passing.

1 open issue

• Remove cancellation probe causing TS2322 compile failure · New

What changed in this PR

This PR aims to shorten VS Code pull request checks and free runners sooner by increasing Linux test parallelism, cancelling runs after compilation or hygiene check failures, and accelerating Windows cache warming.

Changes:

• Split Linux Electron and Remote test jobs, and increase Electron smoke testing to three shards.
• Add an isolated cancellation job for same-repository pull requests, plus a temporary compilation-failure probe.
• Move Windows cache warming to a Dev Drive and update workflow tests and troubleshooting documentation.

File	Description
src/​vs/​base/​common/​numbers.ts	Adds a deliberate type error to test cancellation.
build/​lib/​test/​detectMissingCaches.test.ts	Checks parity between Windows Dev Drive setup steps.
build/​lib/​test/​detectChanges.test.ts	Updates expected job names and cancellation-job coverage.
build/​lib/​test/​agentHostTestSelection.test.ts	Adapts assertions to split Linux integration jobs.
.github/​workflows/​pr.yml	Reports failed checks and adds isolated run cancellation.
.github/​workflows/​pr-node-modules.yml	Moves Windows cache warming onto a Dev Drive.
.github/​workflows/​pr-linux-test.yml	Splits slow test jobs and adds a third smoke shard.
.github/​skills/​smoke-tests/​SKILL.md	Documents revised sharding and remote smoke jobs.
.github/​skills/​fix-ci-failures/​SKILL.md	Updates artifact names and download examples.

_{Review effort: Balanced}