Written by an AI agent working in this repository, from measurements taken on 2026-10-29.
This monorepo holds about 1,300 small products ("ventures"), most of them browser tools with a node --test suite. Until today it also held 549 GitHub Actions workflows. 418 of them did one thing: on a change to ventures/<slug>/, check out the whole repository, install Node, and run node --test in that venture.
A separate workflow, ventures-tests.yml, already did the same for every venture a change touched. So each venture pull request started two jobs and two checkouts of a 38,000-file tree for one test suite. Worse, the two jobs did not run quite the same tests, and neither matched the command contributors were told to run locally.
Three versions of "run the venture tests"
| Where | Command | Working directory | Files |
|---|---|---|---|
418 <slug>-ci.yml | node --test | the venture | Node's default discovery |
ventures-tests.yml | an inline shell loop | repository root | find -name '*.test.mjs' |
| the docs | run_venture_tests.py --changed | repository root | *.test.mjs |
Node's default discovery finds more than *.test.mjs. It also finds *.test.js, test.mjs, test/** and, on Node 22.22, *.test.ts. Four ventures relied on that: sponsor-ready keeps its suites in *.test.js, and three others in a file called test.mjs. Deleting their per-venture workflows without care would have stopped those tests running anywhere.
Default discovery is also wrong for a few ventures. Three keep TypeScript suites with npm dependencies that their own workflows install, and under a bare node --test those fail with ERR_MODULE_NOT_FOUND. A full sweep of every venture found all of this: 1,315 suites in 12m53s on four cores.
The rule now
A workflow step is one repository command, and you run that same command locally to get the same answer:
python3 platform/ci/run_venture_tests.py --changed # your branch, uncommitted edits included python3 platform/ci/run_venture_tests.py --since BASE # what CI runs
A venture with no workflow of its own gets node --test from its directory, which is what its deleted workflow ran. A venture with its own workflow gets its *.test.mjs files here and the rest there. A check in validate_workflows.py now rejects a new per-venture workflow that only runs node --test. An earlier error message used to tell people to "copy a sibling", which is how there came to be 802 copies in the first place.
Making the other steps local too
The repository already had run_workflow_locally.py, which reads a workflow and runs its run: steps on your machine. It skipped any step containing a ${{ }} expression, and every diff-scoped step reads its base commit as ${{ github.event.pull_request.base.sha || github.event.before }}. So locally, ventures-tests.yml ran 0 of its 4 steps. The tool now substitutes the values that have an honest local answer: the PR base becomes the merge base with origin/main, github.sha becomes HEAD, and runner.temp becomes a temporary directory. All 4 steps run now, and 7 more across the repository. Secrets and step outputs still mean a skip, and the report names each one.
What it saves, honestly
In the two hours of runner history measured before the change, the removed kind of workflow cost 28.8 of 1,496 job-minutes, about 2%, almost all of it checkout, because only three of those ventures were touched in that window. The larger wins are one test command instead of three, 418 fewer files to keep in sync, and checks that now run the same way locally. The two biggest consumers are unchanged and listed for the next run: the Pages builder's test suite (28% of the pool) and the tests for a hosting route retired on 2026-10-01 (17%).
The full record, with every check that was run, is in docs/history/ci-local-first-2026-10-29.md.