changelog.yml

name: Auto Changelog

# Dispatch-only, deliberately (objectui#5409).
#
# This lane also declared `release: types: [published]` until 2026-08-23. That
# trigger was dead configuration, not a fallback: every release in this
# repository is authored by `github-actions[bot]`, created by the Changesets
# action in `changeset-release.yml` using `secrets.GITHUB_TOKEN`, and GitHub
# does not start workflow runs from events raised with that token (the
# recursive-trigger guard). Measured before removal: this workflow had 0 runs
# across the repository's whole life, against `changeset-release.yml`'s 4049
# through the identical API call — so the zero was the trigger, not the query.
#
# When to run it: at release time, as part of cutting the release. Nothing else
# invokes this workflow. The root CHANGELOG.md is a periodically hand-curated
# summary; the per-package CHANGELOG.md files that Changesets writes on the
# release commit are the source of truth for granular and current history.
on:
  workflow_dispatch:

permissions:
  contents: write
  pull-requests: read

jobs:
  changelog:
    runs-on: ubuntu-latest

    # ── NO job ceiling, and that is the DECISION, not its absence (objectui#7956) ──
    # This job runs under GitHub's 360-minute default. objectui#7956 asked for a
    # ceiling DERIVED from this job's own measured run-time distribution, in the
    # shape objectui#7270 left beside `lint.yml::lint`. It is recorded here
    # instead, because this job has no distribution to derive one from.
    #
    #   - Population: every run of this workflow, no status filter, whole
    #     retained history — `GET /actions/workflows/changelog.yml/runs` reports
    #     `total_count: 0`. Re-measured 2026-09-06T17:22Z.
    #   - Control, so that zero is a reading rather than a broken query: the
    #     identical call against `changeset-release.yml` reports a `total_count`
    #     of 5010 completed runs. The endpoint answers; this workflow has never
    #     produced a run for it to answer with.
    #   - The header above already recorded this same zero once, measured before
    #     2026-08-23 against the same endpoint while removing the dead `release`
    #     trigger. It has not moved: this workflow is dispatch-only, and it has
    #     never been dispatched.
    #
    # ⛔ Do NOT put a number here derived from another job, from the apparent
    # cost of a checkout plus git-cliff, or from what these steps "should" take.
    # A ceiling set under a job's honest slowest run converts a working job into
    # a permanently red one (objectui#7048), and this job has produced no honest
    # run to sit above; objectstack#16173 is the live counter-example, where a
    # distribution mis-estimated by ~2.6x killed a test shard while the rollup
    # read green. An invented ceiling is worse than the default it replaces,
    # because it looks derived.
    #
    # ⚠️ The accepted cost, stated plainly: a hang in a dispatched run occupies
    # a runner for six hours before reporting `cancelled`. That is the same
    # exposure objectui#7956 closed on four of its six jobs. It is accepted here
    # only because the alternative is a guess, and it is bounded in practice by
    # the fact that nothing starts this workflow except a maintainer, who is by
    # construction watching when they do.
    #
    # What would change this: one real dispatch. Measure that run the way
    # `lint.yml::lint` documents, write the derivation here, and move this job
    # out of the accept-360 table and into the `derived` table in
    # `scripts/__tests__/workflow-cache-save-bound.test.ts` in the same commit —
    # that pin holds this job's UNBOUNDED state deliberately, so adding a
    # ceiling has to be an edit someone makes on purpose rather than a silent
    # one.
    steps:
      - name: Checkout code
        uses: actions/checkout@v7
        with:
          fetch-depth: 0
          submodules: true

      # No lockfile merge driver here, deliberately (objectui#6358).
      #
      # This job configured one until 2026-08-25. A merge driver is invoked by
      # git only when git actually has to MERGE the attributed path, and this
      # job never merges: it checks out, runs git-cliff, stages exactly
      # CHANGELOG.md, commits, and pushes. There is no merge, rebase, pull,
      # cherry-pick or revert in it, and it never touches `pnpm-lock.yaml` at
      # all. The stated reason for the step -- that it "commits back to a branch
      # that may have moved" -- does not reach a driver: a push to a branch that
      # moved is REJECTED, not merged, and nothing here resolves that rejection.
      #
      # So the step was dead configuration rather than a missing-toolchain bug.
      # Note the disposition did NOT turn on the workflow being unreachable: it
      # is `workflow_dispatch`, a maintainer can start it any time, and it is
      # meant to be run at release time. It had 0 runs when this was measured,
      # but "never dispatched" is not "cannot run" -- the driver would not fire
      # even on a successful manual dispatch, which is the reason acted on.
      #
      # Restoring it needs a real merge in this job first, not the instinct that
      # every workflow touching git wants the driver. If one is ever added, add
      # `corepack enable` + `actions/setup-node` ahead of it the way
      # `changeset-release.yml` and `dependabot-auto-merge.yml` do -- the driver
      # shells out to pnpm, and this workflow pins no toolchain.

      - name: Generate changelog
        uses: orhun/git-cliff-action@v4
        id: changelog
        with:
          config: cliff.toml
          args: --verbose --latest
        env:
          OUTPUT: CHANGELOG.md

      - name: Commit changelog
        run: |
          git config --local user.email "github-actions[bot]@users.noreply.github.com"
          git config --local user.name "github-actions[bot]"
          git add CHANGELOG.md
          git commit -m "docs: update CHANGELOG.md" || echo "No changes to commit"
          git push