[thymikee]

Summary

A PR that only changes unit tests (*.test.ts / __tests__/ under src/ or packages/*/src/) no longer runs the device smoke lanes. iOS smoke already worked this way. Android, Linux, the macOS phase of Apple, and the web smoke steps in CI now use the same affected-check rule: a lane runs only when selectChecks picks one of its checks. Pushes to main, empty or unknown diffs, and fail-open plans (scripts/, .github/, package.json, …) still run every lane.

• scripts/device-lane-impact.ts <check-id>... holds the shared rule. apple-ci-impact.ts now uses it for iOS smoke, XCTest, and the new macOS selection.
• The macOS phase also runs when the darwin-only bridge-proof inputs change.
• Apple skips the build when neither phase is selected. The final status guard expects skipped for every unselected phase.
• Scope: the lanes also route by platform family, as iOS already did. For example, an Android-only change skips Linux, web, and macOS.
• 9 files changed.

Example: #3898, which only changed a test file, now skips Android and Linux.

Validation

At d0c85f95b7c6be5ce7f63713e95e6045e9e85949:

• pnpm typecheck, pnpm check:fallow, check:gate-manifest, and the agent-guidance, docs-prose, production-exports, affected-selector, and depgraph gates pass.
• vitest run scripts/__tests__/{device-lane,apple-ci}-impact.test.ts: 59 passed.
• check:affected --run --head failed open to the full set. vitest-related ran 17,008 tests with 11 timeouts under host load around 950. All 8 failing files pass when rerun alone (65 tests). The gates after it did not run locally, so CI is authoritative for them.
• Risk: family routing for Android, Linux, and web has not had the 90-day replay that iOS got. Pushes to main still run every lane as the backstop.

🤖 Generated with Claude Code

[github-actions]

Size Report

Metric	Base	Current	Diff
Installed (including dependencies)	5.54 MB	5.53 MB	-1.3 kB
Package (unpacked)	5.54 MB	5.53 MB	-1.3 kB
Package (download)	1.68 MB	1.68 MB	-463 B

Startup median (7 runs, lower is better):

Scenario	Base	Current	Diff
CLI --version	42.4 ms	42.6 ms	+0.1 ms
CLI --help	86.1 ms	86.1 ms	+0.0 ms

[thymikee]

This PR is ready at d0c85f9, and I found no problems that need a change. All 19 checks pass at this head, and there are no conflicts. Because the PR changes scripts/ and .github/, every selector fails open, so this run executed every lane. That means no Actions run has yet taken the skip path. The evidence for it is GitHub's documented skipped outcome for a false step if, plus the guard test that runs the shell. I did not rerun the vitest cases or the gates locally. I did not check cross-family execution, such as an Android-tagged module reached from a Linux or web session. The family rule accepts that, and pushes to main stay the backstop. Android, Linux and web routing has had no 90-day replay, as the PR body says.

Not blocking, and you can take or leave these: (1) The if: steps.impact.outputs.run != 'false' guard sits on every step after the selection (linux.yml), and device-lane-impact.test.ts enforces that for Android and Linux today. A small selection job with one job-level needs and if would remove the per-step repetition, and a skipped job still satisfies a required check. (2) selectMacosSmoke falls back to selectAppleBridgeProof (apple-ci-impact.ts), which selects any file under packages/platform-apple/src/, including unit tests, so a test-only Apple PR still runs the full macOS phase to reach a 4-file proof. It also misses the proof's own capture-kit test file, as main already did. Could that selector own exactly what the proof runs, ideally as an affected-check id that selectDeviceLane covers? (3) The CLI in device-lane-impact.ts accepts any CheckId, such as lint, and reportSelection labels every Android, Linux and web summary 'Device lane'. Could the arguments be limited to the lane checks that LEAF_LANES and BUILD_OWNERSHIP already list, with a lane label passed in? The device-lanes.ts header also still says only the ios.yml routing was replayed.

[github-actions]

PR Preview Action v1.8.1
Preview removed because the pull request was closed.
2026-10-22 11:12 UTC