ci: supersede #367 with combined tiered PR pipeline redesign (#369)

* ci: streamline PR pipeline gates

* test: annotate ci-test-scope stderr assertions for lint rule

---------

Co-authored-by: Colin <colin@solvely.net>
This commit is contained in:
Tom Boucher
2026-05-26 19:11:58 -04:00
committed by GitHub
parent aaa19b0609
commit 9dd2c87c6c
15 changed files with 950 additions and 192 deletions

View File

@@ -49,25 +49,44 @@ Direct harness invocation also works:
```bash
node scripts/run-tests.cjs --suite security
node scripts/run-tests.cjs --suite=security
node scripts/run-tests.cjs --files "tests/command-contract.test.cjs tests/core.test.cjs"
node scripts/run-tests.cjs --files-from .ci-selected-tests.txt
```
Unknown suites exit non-zero with the list of valid suites. Empty suites (e.g. `--suite security` before any security-tagged file exists) exit `0` with a `no tests in suite "..."` notice on stderr so CI lanes don't go red while a suite is being populated.
## CI matrix
The `Tests` workflow runs on:
The `Tests` workflow runs every PR through a scoped gate generated by
`scripts/ci-test-scope.cjs`.
| OS | Node 22 | Node 24 | Node 26 |
|---|---|---|---|
| `ubuntu-latest` | gate | gate | forward-compat (`continue-on-error`) |
| `macos-latest` | gate | gate | forward-compat (`continue-on-error`) |
| `windows-latest` | gate | gate | forward-compat (`continue-on-error`) |
| Lane | Node 22 | Node 24 |
|---|---|---|
| `ubuntu-latest` | scoped tests | unit + integration + security |
| `windows-latest` | — | scoped Windows/path/shell tests |
| `macos-latest` | full parity when required | full parity when required |
- **Node 22** is the `engines.node` floor (`>=22.0.0`) — must stay green.
- **Node 24** is the default development lane.
- **Node 26** is forward-compat. The lane reports status but does not gate the workflow — `actions/setup-node` may not yet have a stable Node 26 image at any given moment. When it stabilises, flip `continue-on-error` off.
- **Scoped tests** are selected from the changed paths, plus a small CLI/package
smoke set. They are for confidence on the affected surface, not for counting
tests.
Each matrix cell runs `test:affected` on PRs (changed tests + reverse-import dependents + a smoke fallback), then runs full named suites on push lanes. `install` and `slow` remain `main`/release push-only checks on the canonical lane to keep PR CI fast. Coverage runs in a dedicated `coverage` job on `ubuntu-latest` / Node 24 — running coverage across the full matrix would 9x the cost for no extra coverage data.
The default PR gate runs the broad `unit`, `integration`, and `security` suites
once on Ubuntu / Node 24, scoped smoke on Ubuntu / Node 22, scoped
Windows/path/shell tests on Windows / Node 24, and unit coverage once on Ubuntu /
Node 24. PRs touching workflow, package, test-runner, install, release, or
Windows-sensitive surfaces also run the full parity matrix on macOS and the
older Windows runtime, plus `install` and `slow` on the primary Ubuntu lane.
Coverage stays single-lane because multiplying coverage across OS/runtime lanes
adds cost without improving the threshold signal.
To inspect the scope locally:
```bash
npm run ci:test-scope -- --files "commands/gsd/plan-phase.md"
node scripts/ci-test-scope.cjs --base origin/next --head HEAD
```
## Best practices for forward-compat (Node 24/26)

View File

@@ -10,7 +10,9 @@ Targets `~DEFAULT_BRANCH` (main). Enforces:
- No deletions or force pushes
- Required linear history (no merge commits)
- All changes via pull request (0 required approvals, stale-review dismissal, thread resolution required, squash/rebase only)
- 10 required status checks: lint, Node 22/24 matrix (ubuntu/mac/windows), Changeset Required, require-issue-link, pr-template-format
- Required status checks: the aggregate `Required tests` gate plus PR policy
checks for size, branch name, changeset, docs, target branch, issue link, and
PR template format
### `release-branches`
Targets `refs/heads/release/**` and `refs/heads/hotfix/**`. Same rules as
@@ -26,7 +28,7 @@ immutable once created. Tag creation is unrestricted.
| PR | Branch | Action |
|----|--------|--------|
| PR-1 | `chore/branch-protection-specs` | Check in spec files; `enforcement: disabled` — no effect on repo |
| PR-2 | `chore/ci-skip-tests-on-docs` | Add path filter to `test.yml` + new `test-skip.yml` noop; doc-only PRs satisfy required checks in <30s |
| PR-2 | `chore/tiered-ci-gate` | Add tiered test scope detection to `test.yml`; doc-only PRs satisfy `Required tests` without macOS/Windows noop queues |
| PR-3 | `chore/branch-protection-evaluate` | Run `sync-rulesets.sh` with `ENFORCEMENT=evaluate`; 1-week dry-run via rule-suite logs |
| PR-4 | `chore/branch-protection-active` | Run `sync-rulesets.sh` with `ENFORCEMENT=active`; protection live |
@@ -63,26 +65,31 @@ triggered. Use this to validate no legitimate workflows are broken before
flipping to `active` in PR-3.
## Path filters
## Test scope detection
The `test.yml` workflow uses a `paths:` filter on its `pull_request:` trigger so
the 6-lane matrix only runs when code-touching files change. A companion workflow,
`test-skip.yml`, fires on the inverse (`paths-ignore:`) and emits instant noop
jobs with **identical job IDs and matrix dimensions**. This ensures the 7 required
status checks (`lint-tests` + 6× `test (...)`) are always satisfied — whether the
real matrix ran or the noop ran.
The `test.yml` workflow always runs, but its `changes` job classifies the PR
with `scripts/ci-test-scope.cjs` before starting expensive runners. The required
branch-protection context is the single aggregate `Required tests` job.
### Why dual workflow instead of paths-ignore alone?
Doc-only PRs run the lightweight lint and aggregate jobs only. Code-touching PRs
run the default required matrix:
GitHub required-status-checks expect a specific context string to appear as
"passed" on every PR. If `test.yml` is suppressed by `paths-ignore` on a doc-only
PR, those contexts never fire and the PR is permanently blocked. The noop workflow
produces the same context strings via matching job IDs + matrix, resolving the
deadlock.
- Ubuntu / Node 22 scoped tests
- Ubuntu / Node 24 unit, integration, and security suites
- Windows / Node 24 scoped Windows/path/shell tests
- Ubuntu / Node 24 coverage
PRs that touch workflow, package, test-runner, install, release, or Windows
sensitive surfaces also run install/slow on the primary Ubuntu lane and the full
parity matrix:
- Windows / Node 22
- macOS / Node 22
- macOS / Node 24
### Canonical code-paths list
Both `test.yml` (`paths:`) and `test-skip.yml` (`paths-ignore:`) use this list:
The `changes` job treats these paths as code-touching:
```
bin/**
@@ -90,29 +97,32 @@ get-shit-done/**
agents/**
commands/**
hooks/**
sdk/**
tests/**
scripts/**
package.json
package-lock.json
tsconfig*.json
.github/workflows/test.yml
.github/workflows/test-skip.yml
.github/workflows/**
.github/rulesets/**
```
This list also mirrors `changeset-required.yml`'s path filter. Keep all three in sync.
This list should stay aligned with changeset/docs policy where those gates care
about the same user-facing surfaces.
### Adding a new code path
When adding a directory or file that should trigger real tests:
1. Add the glob to `paths:` in `.github/workflows/test.yml`
2. Add the **same** glob to `paths-ignore:` in `.github/workflows/test-skip.yml`
3. Add the same glob to `changeset-required.yml` if changesets should be required
1. Add the path classifier to `scripts/ci-test-scope.cjs`.
2. Add targeted tests for that surface in the same classifier.
3. If it requires macOS/extra-Windows coverage, mark the classifier rule
`fullMatrix: true`.
4. Add the same glob to `changeset-required.yml` if changesets should be required
for that path
Failure to update `test-skip.yml` means doc PRs that happen to touch the new
path will deadlock (real matrix never fires, noop never fires either).
The old `test-skip.yml` inverse-path workflow was removed. Do not add required
checks for individual matrix jobs; require the aggregate `Required tests` context
instead.
## Phase-2 TODO