ceremony

One release ceremony for the whole heavy-duty family — implemented once, tested once, documented here, consumed everywhere else by reference. The approach and its constraints live in #1; this README is the operator-facing doctrine that used to live, three times over, in the consumers' CONTRIBUTINGs.

• Adopting or converting a repo → docs/CONSUMERS.md.
• Working in this repo as an agent → AGENTS.md routes you; CONTRIBUTING.md has the repo specifics.
• Operating a release, or staring at a red run on main → read on.

What a release is

A release is a PR, and merging it ships it (box#96, building on box#83; rig#47, cast#111 converged on the same doctrine). The ceremony PR — release: X.Y.Z, carrying the hand-set release label — makes three stamps:

1. The version goes bare: X.Y.Z-dev → X.Y.Z (lib/version.sh).

2. The changelog is stamped and re-armed — two edits, not one (box#108). ## Unreleased becomes ## X.Y.Z — DATE, and an empty ## Unreleased goes back on top, immediately above it:

   ## Unreleased
   
   ## 0.7.1 — 2026-07-19
   
   ### Fixed
   ...
   

   The second edit is not cosmetic and not deferrable. Between the stamp and the next re-creation of that heading, main has no ## Unreleased. A PR authored before the release wrote its entry under that heading; with the heading gone, git lands the entry under whatever now occupies the position — the section that just shipped — and it merges cleanly, no conflict, no signal. The changelog then credits a released version with a change it does not contain, and nothing but a human reading the file will ever say so (box#108; confirmed cross-repo as rig#66). The armed guard exists because of exactly this edit.

3. The drill record is present: drills/X.Y.Z.md, non-blank — the evidence the release rests on (the drill doctrine).

(This repo's own ceremony adds a fourth stamp: CEREMONY_SELF_REF — the ref consumers' runs fetch this repo at — moves to the version being released, in release.yml and every other workflow that carries it. self-ref-check.sh fails CI here, not a consumer's release, when it is stale.)

The merge is the ship decision; the tag is transcription. After the merge, release.yml asserts its way to certainty, tags the merge commit, publishes the GitHub release with the version's own changelog section as the body — the curated prose, never the generated PR list (lib/changelog.sh is the one canonical extractor) — and re-arms main by bumping to X.Y.(Z+1)-dev (release.yml). The machine does the transcription because humans err silently and machines fail loudly: everything asserts its way to certainty and fails loudly, creating nothing — a wrong release is worse than a missing one, so every failed assert leaves zero artifacts: no tag, no release, no bump.

The two doors

• The merge door — the paved road. A push to main (release.yml) runs the decide table; a merged, release-labeled PR whose version transitioned to bare is the ceremony, everything legitimate that isn't one is a green no-op, and every half-ceremony dies loudly. Use it for every normal release.

• The tag door — the fallback and the backfill. A bare X.Y.Z tag push — no v prefix, box's 0.6.0 set the scheme (release.yml) — publishes the same way. The tag is the operator's explicit act, so there is no decide and no label check; the one assert is that the tag names the tree's own version, and a mismatch refuses, creating nothing. No -dev bump either — the fallback does not rewrite main (cast's precedent). Use it when the merge path is red, for backfills, and for the first-release edge (row 4).

Tag + publish (+ the consumer's artifact hook) happen in the same job, on purpose: a GITHUB_TOKEN-created tag fires no workflows (GitHub's anti-recursion), so the merge door's tag can never re-enter the tag door and double-publish — and that job is the release's only chance to publish (release.yml, #1 constraint 2).

What happens when my PR lands on main

The merge door runs on every push to main, and the release label legitimately means two things (release ceremonies, and ordinary work on the release machinery), so the door's first act is a decision: the six-row table in lib/decide.sh (issue #8 — the comment block is the spec, and the table is contract-tested offline). Rendered for operators:

#	the tree your merge produced	the run	what it means — and your move
1	version -dev, unchanged	green NOTICE, no-op	Almost every PR — including release-flow work under the release label. Nothing to publish, nothing to do.
2	version changed, still -dev	green NOTICE, no-op	The post-release bump, or a renumber. "A dev tree is by definition not a release." Nothing to do.
3	version bare, unchanged, already released	green NOTICE, no-op	The post-release window: the ceremony landed, the -dev bump hasn't. Nothing to do.
4	version bare, unchanged, never released	red, nothing created	The label says ship but this PR did not mint the version. Mislabeled → drop the label. Meant to release → it forgot the bump; re-do the ceremony PR. A repo whose first version never carried -dev ships its first release by the tag door — the known first-release edge (cast#111; lib/decide.sh).
5	version transitioned to bare, no merged release-labeled PR behind the commit	red, nothing created	A transition nobody declared — a release is a labeled ceremony PR, not a bare push. Label a proper ceremony PR and re-do it, or publish by the tag door if the tree is genuinely right.
6	version transitioned to bare, merged release-labeled PR behind the commit	the ceremony	Tag → notes → publish → -dev re-arm. Your move afterwards: verify the release exists and main reads X.Y.(Z+1)-dev.

The green rows are the point as much as the red ones: the machinery must be safe to work on, so every legitimate non-ceremony is a green NOTICE no-op — never a red run on main per infra PR (lib/decide.sh). The label is hand-set intent and automation never guesses; the version transition is the interlock, and label-without-transition (row 4) and transition-without-label (row 5) both refuse (#1 constraint 8).

The guards

Three composite actions run in every consumer's CI (and in this repo's own). Shared shape: version-keyed where the tree's state matters, loud where it fails, and a file of its own so a test can drive it. The full war stories are in the scripts' header comments — authoritative and longer than this; what follows is the operator's cut.

changelog-armed — main never sits disarmed

The rule (actions/changelog-armed/changelog-armed.sh), keyed on the tree's version:

• -dev tree → the top section must be ## Unreleased.
• bare tree (the ceremony PR and its merge) → the top section may be ## Unreleased (re-armed) or the stamped section for exactly that version — and that version's section must exist and carry prose, because it is the one about to ship (the half-ceremony refusal, rig#67: version bumped, stamp missing — asserted through the very extractor the publisher uses, so the two cannot disagree about what a section is).

The incident: box#108 / rig#66 — the silent mislanding described above. Red means a PR entry has nowhere safe to land; the fix is to re-arm: add an empty ## Unreleased above the top stamped section.

Do not "simplify" this to "always require ## Unreleased". The unconditional form is false by construction on the ceremony PR's own tree — it makes every release unshippable — and rig#44 and cast#108 both had to revert exactly that (the script's header). The version-keyed form is what rig and cast get back by adopting this repo.

One consequence worth knowing before it happens: a ceremony PR that stamps and forgets to re-arm still passes this guard — a bare tree is allowed to be stamped. It goes red the moment the automatic -dev bump lands on main (the script). The guard does not block the release; it refuses to let main sit disarmed, which is the window a late PR falls into.

changelog-monotonic — shipped headings are append-only

The rule (actions/changelog-monotonic/changelog-monotonic.sh): the set of ## X.Y.Z headings on your branch must be a superset of the set at the merge base, and no heading may appear twice on HEAD. The rule needs no tuning because release headings are append-only by doctrine: the ceremony adds one and nothing ever legitimately removes one — so superset has no exception to carve. The ceremony's own stamp passes by construction: rewriting ## Unreleased into ## X.Y.Z — DATE adds a heading and removes none (Unreleased is not a version heading; it is changelog-armed's business).

The incidents: box#122 (caught in review of box#118) — an author adding an entry under ## Unreleased replaced the heading below it instead of inserting above it; git merges that cleanly, and the shipped section's body is silently absorbed into ## Unreleased. And box#118 itself — a bad rebase duplicated a shipped heading, which containment is blind to, which is why uniqueness-on-HEAD is a separate assert (the script).

Red means a shipped section was deleted (put the heading back and insert above it) or duplicated (collapse to one heading; the failure message walks through both fixes with the diff to run). This guard needs history: the consumer's checkout must use fetch-depth: 0, and in CI an unresolvable base is a hard failure, not a skip — a guard that can quietly stop guarding is the failure shape this family of checks exists to refuse (strict mode).

drill-recorded — a release carries its evidence

The rule (actions/drill-recorded/drill-recorded.sh), keyed on the tree's version: a -dev tree passes with nothing to assert (a development tree ships nothing); a bare tree — the ceremony PR and its merge — must carry drills/<version>.md with at least one non-whitespace character. One file per version, so 0.9.0.md and 0.9.0-rc1.md are simply different files and prefix confusion is unrepresentable (#1 constraint 7).

The incident: box's CONTRIBUTING said since box#96 that the release ritual must be run and recorded. No release ever did it — box#95, box#114 and box#148 all shipped as a version bump plus a changelog stamp, because the gate was a sentence in a document and the only thing standing on it was a reviewer remembering to ask. The rule moved into CI, where it fires whether or not anyone is paying attention.

Red means the release is asserting a ritual it left no evidence of. The fix is to run the drill and record it — or to waive it in writing at the same path: the guard demands a record, not a passing result (below).

The drill doctrine

Evidence, not success. The guard asserts a record exists — a failed drill honestly written down satisfies it, and so does a maintainer waiver that says plainly the drill was waived and why. What it refuses is silence: a skip must cost a deliberate, reviewable file in the diff, which is precisely what box's three silent skips never produced. CI cannot run a consumer's drill (box's wants real hardware and the better part of an hour); it can only refuse a release that never ran one.

Each repo defines what its drill means — the gate only reads the record. box asserts the isolation contract; rig asserts convergence (a machine reaches its role, idempotently); cast asserts promotion (A→B reproduces, the diff is idempotent); ceremony's own drill is a door rehearsal — both doors exercised end-to-end on a disposable repo (#11 names the six probes); incubator's is TBD in heavy-duty/incubator. Each repo states its meaning in its own drills/README.md. Three different exercises sharing a substrate is why the records are per-repo — they are not phases of one script.

Drills exercise candidate refs, not released artifacts. A ref is a static identifier that exists as soon as the release branch does, so no repo has to be released — or drilled — before another can be drilled: what looks like a box↔rig recursion at runtime dissolves into two independent tests against one fixed pair of refs. And drilling the candidate is drilling the release: a ceremony PR's diff is the stamps and nothing else, so no executable byte differs between the tree that was drilled and the tree that ships.

A cross-repo release set shares one run ID. Each repo records its own legs in its own drills/X.Y.Z.md, citing that run ID and the sibling SHAs, so the records reconcile afterwards — but the guard only ever reads the repo it runs in. If a defect shows up only in the combination: patch, re-drill, re-record. The set converges; it is not required to be right in one pass.

Troubleshooting red main

Every refusal the release flow can emit, verbatim, with cause and remedy. The catalog is generated from the sources, not paraphrased — regenerate it with:

grep -n -A2 'refuse \|>&2' lib/decide.sh lib/facts.sh .github/workflows/release.yml

$VER-style variables appear as the run interpolates them.

The decision refused (lib/decide.sh)

the version '$VER' is bare, unchanged by this PR, and never released — the label says ship but this PR did not mint the version. Refusing to guess — creating nothing. (If this PR was mislabeled, drop the label; if it was meant to release, it forgot the bump. A first release whose version never carried -dev ships by the tag door — the known first-release edge.)

Row 4 (L129–L133). The message is the remedy: drop the label, or re-do the ceremony with the bump, or take the tag door.

the version transitioned ('$BASE_VER' -> '$VER') but no merged, release-labeled PR is behind this commit — a release is a labeled ceremony PR, not a bare push — creating nothing.

Row 5 (L147–L149). Someone pushed or merged a version transition without the release label. Label a proper ceremony PR, or — if the tree is genuinely the release — publish by the tag door.

VER is empty — the caller failed to establish the version at the pushed head. Refusing to decide — creating nothing. BASE_VER is empty — the caller failed to establish the version at the base. Refusing to decide — creating nothing. RELEASED='${RELEASED}' — expected yes, no, or empty. Refusing to decide — creating nothing. LABELED='${LABELED}' — expected yes, no, or empty. Refusing to decide — creating nothing. the version '$VER' is bare and unchanged, but RELEASED is empty — this state is decided by whether '$VER' is already released, and the caller did not establish that fact. Refusing to guess — creating nothing. the version transitioned ('$BASE_VER' -> '$VER') but LABELED is empty — a transition ships only behind a merged, release-labeled PR, and the caller did not establish that fact. Refusing to guess — creating nothing.

The fact-gathering guards (L92–L105, L135, L151): a missing fact must never fall through to "no". These indicate a bug upstream in lib/facts.sh or the workflow plumbing, not an operator mistake — read the run's facts: stderr line and file what you find.

The facts could not be established (lib/facts.sh, lib/version.sh)

facts: unknown VERSION_SOURCE '$VERSION_SOURCE' — expected file or package-json

L37: the caller's version-source: input is neither file nor package-json. Fix the caller.

version_read: $path: no such file version_read: $path is empty version_read: $path: no version field version_read: node is required for version-source: package-json

lib/version.sh: the tree's version source is missing, empty, or unreadable. A wrong release is worse than a missing one, so an unreadable state is never an empty print — restore the VERSION file (or package.json version field) on main.

The merge door refused (release.yml)

CHANGELOG.md has no '## $VER' section at the merge commit — the ceremony PR must stamp it; refusing to publish an empty release

L202–L205: the ceremony merged without its stamp (a state the armed guard already refuses on the PR — red main here means it was overridden). Stamp the section on main, then publish by the tag door.

tag '$VER' already exists — this release already happened, or a manual tag won the race; refusing to re-release, creating nothing. release '$VER' already exists — refusing to re-release, creating nothing.

L207–L222, the nothing-exists assert — what makes a re-run of a completed ceremony refuse instead of clobber, and what catches a manual tag racing the merge. If the release truly exists, there is nothing to do: this red is the system declining to do the thing twice. If the tag exists but the release does not (a manual tag won the race, or a failed artifact hook), recover by the tag door: delete and re-push the tag, or gh release create by hand from a fixed tree.

direct push refused (branch protection?) — opening the bump PR instead

L292–L300 — loud, but not a refusal: the post-release -dev bump could not push directly, so the run opened a release-labeled bump PR itself. Your move: merge it promptly — until it lands, main is sitting bare, where a dev install impersonates the release and the armed guard's window stays open.

The tag door refused (release.yml)

tag '$GITHUB_REF_NAME' does not match the tree's version '$ver' — creating nothing. A release is a PR, then a tag: the release PR bumps the version and stamps the changelog; the tag goes on its MERGE commit. Delete this tag and re-tag the right commit.

L333–L337. The message is the remedy.

CHANGELOG.md has no '## $VER' section — stamp the Unreleased section in the release PR before tagging; refusing to publish an empty release

L346–L349. The tagged tree was never stamped. Stamp first, then delete and re-push the tag.

Red main that is not the release workflow

Consumer CI runs its guard steps on pushes to main too (this repo's ci.yml does the same). The one guard red an operator will actually meet on main is changelog-armed after a re-arm was forgotten: the ceremony stamped without putting ## Unreleased back, the release's own -dev bump landed, and the guard now says (first line):

changelog-armed: the version is '$ver' (a development tree) but the top section of $changelog is: …

The fix is a one-line PR: add an empty ## Unreleased above the stamped section. The full message (the script) carries the same instruction.

Design lineage

The ceremony converged across box#83 → box#96, rig#32 → rig#47, and cast#96 → cast#111; this repo is those three implementations folded into one (the drift that motivated it is measured in #1). The load-bearing constraints — each bought with an incident, none of them safe to "simplify" away — are listed in #1 and carried, with their war stories, in the headers of the scripts they bind: release.yml, lib/decide.sh, lib/facts.sh, and the three guard scripts. The comments are the documentation of record; this README is their operator-facing cut.