Zum Inhalt

A workflow that only the repositories run starts here only by hand

  • Status: accepted, amends 0045 and 0083 in what the blueprint itself runs
  • Date: 2026-11-17

Context

The blueprint keeps the checks of every kind of project, and a few bots that only the repositories made from it run, in its own .github/workflows/, since the template includes them from there. In the blueprint itself each of their jobs carries the line if: github.repository != 'Dennis-Otto/repo-blueprint', which the template removes. The workflows kept their triggers, though, so GitHub made a run for every push of a pull request, every merge into main and every schedule, and each of these runs skipped all its jobs.

From 2026-11-10 to 2026-11-16 that made about 3,850 runs in the blueprint that did nothing (issue 1003): CI (node), CI (php), CI (generic), CI (home-assistant), CI (container), CI (github-action) and Action test 533 or 532 times each, Newest Node 110 times, and Property tests, Upstream and Blueprint update 13 times, four of these because Bot health started Property tests again for a schedule that it took as late. The release pull request alone, with 173 heads that week, had 1,211 of them. They cost no minutes of a runner, but they filled the list of runs and the checks of every pull request with skipped entries, and each one sent its workflow_run, check_suite and check_run events to every bot that listens.

Options

  1. Move the files out of .github/workflows/, such as to .github/stacks/, where GitHub doesn't run them. Renovate's manager of GitHub Actions, CodeQL's analysis of the workflows, the OpenSSF Scorecard, actionlint and zizmor read only .github/workflows/: the pins of the checks of every kind would go stale here, and every repository would get them stale.
  2. A filter on each trigger, such as branches-ignore for every branch. GitHub has no filter on the repository, and a filter that matches nothing still leaves the schedule and makes the files hard to read.
  3. Only workflow_dispatch here, the real triggers from the template. The files stay where every tool reads them; the template puts back the triggers that the blueprint's copy leaves out.

Decision

Option 3.

  • In the blueprint, a workflow of which no job runs here, by the rule of idle() of blueprint/website.py (every job carries the guard or only waits for such jobs), has only workflow_dispatch in its block on:, with its inputs where it has some, after the comment that by_hand of includes/replace.jinja holds. GitHub makes no run of it for a pull request, a merge or a schedule, and Bot health finds no schedule of it that it could take as late. These are the checks of every kind but Python, Action test, Newest Node, Property tests, Upstream and Blueprint update.
  • The guard stays, so that a run started by hand still does nothing here, and the page Automation of the website still leaves these workflows out (0045).
  • In every repository, the template of each workflow replaces by_hand with the triggers that the workflow had, through the same replaced() that removes the guard; replaced() moves each schedule into the slot of the repository after its replacements (0083). The render stops when the blueprint's copy no longer holds by_hand. Every repository made from the blueprint gets the same files as before, byte for byte, for every kind and with the answers of every repository in tests/answers/. CI (python) is the blueprint's own CI and keeps its triggers; the checks of the other kinds get the same events.
  • The real triggers are still checked before a release. In the blueprint, zizmor and CodeQL see these workflows with workflow_dispatch alone, so every leg of the Variants runs zizmor on the workflows of its project, with the version and the configuration of the project's own Workflow security, next to the actionlint that it ran before. Its audits that depend on a trigger, such as dangerous-triggers and cache-poisoning, see the real ones there. It runs offline: the pins are the blueprint's own, whose advisories the blueprint's zizmor looks up.

What stays

  • No required check and no bot of a repository loses a trigger: test_each_kind_gets_the_triggers_that_the_blueprint_leaves_out renders every kind and compares the block on: of each of these workflows with the one it had, the checks of every kind with that of CI (python), each schedule in the slot of the project.
  • Nothing new starts by itself in the blueprint: test_a_workflow_of_which_no_job_runs_here_starts_only_by_hand fails when a workflow of which no job runs here has a trigger other than workflow_dispatch, and when a workflow of which a job runs here carries by_hand.
  • The concurrency of these workflows, with github.event.pull_request.number, is the same text, and valid for a start by hand, where the number is empty.
  • The list of workflow_run of Bot health keeps its names: the template of bot-health.yml replaces it word for word, and a workflow that never runs here never ends here either.
  • No required check of the blueprint named one of these workflows (required_checks of .github/repository.toml).

Consequences

About 3,850 runs a week less in the blueprint, and a list of the checks of each pull request without seven skipped entries. A change of the triggers of one of these workflows is made in its template, template/.github/workflows/, not in the blueprint's copy; the tests above say where. A new workflow of this kind takes by_hand and a template with its triggers, or the test of the idle workflows fails.