Both 0.2.0 drills ended at the same wall independently: bot tokens deliberately lack delete_repo, so 'it gets deleted at the end' named a step no builder in this fleet can perform. One drill held its release draft in state:building retrying a 403 that cannot succeed; the other shipped a record asserting a delete that had not happened. Step 1 now names archive (archived:true, inside the repo scope) as the builder's end state and the delete as the operator's, states that cleanup gates nothing, and says why the archived leftover is safe to leave. The record now states the disposal its author actually observed. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
4 KiB
Drills
What a drill means in this repo: an end-to-end rehearsal of both doors of
the release workflow on a disposable repo. The contract suite proves every
decision offline — facts → decide → notes against fixtures, the merge door's
step sequence replayed in release-exercise.yml — but the doors themselves
only ever run live: gating on a real push event, the tag create, the
publish, the -dev re-arm (release.yml's "what is honestly untested"). The
drill is where they run live before a version rests on them.
The rehearsal
-
Create a scratch private repo. It is disposable by design — but the disposal is split, because the builder cannot perform the delete: at the end the builder archives it (
PATCH /repos/{owner}/{repo}witharchived: true, inside thereposcope every fleet identity holds), and deleting it is the operator's step —delete_repois deliberately absent from bot tokens, fleet doctrine and not a misconfiguration, so no builder that will ever run a drill can do it. Do not retry the delete and do not wait on it: both 0.2.0 drills ended at that wall independently (#135) — one builder held its release draft instate:buildingre-trying a 403 that cannot succeed, the other wrote a record asserting a delete that had not happened. Cleanup gates nothing — not ready-for-review, not the review panel, not the merge. The archived leftover is safe to leave: private, no consumers, and outsideheavy-duty/ceremony's ref namespace — the namespace the "never a branch named like the tag" rule below protects. -
Install the docs/CONSUMERS.md caller stubs, pinned to a fork ref carrying the release candidate tree. The candidate's
CEREMONY_SELF_REFis by construction the tag this release has not created yet, so the consumer path cannot resolve directly from the candidate. Rewrite that pin to a canonical candidate SHA in every carrier on the fork ref, and record the fork ref and rewritten pin in the drill record.Never create a branch named like the tag on
heavy-duty/ceremonyto paper over this deadlock: it would shadow the tag for every consumer until someone remembers to delete it. The 0.1.0 drill (#11) is the worked example of this standing fork-ref shape. -
Give it a fixture
VERSION/CHANGELOG.md/changelog.d//drills/in the armed state (X.Y.Z-dev, the fragments directory with itsREADME.mdmarker plus at least one fragment for the ceremony to consume). -
Exercise both doors, one probe at a time:
- a merge-door ceremony publishes exactly one release and re-arms main
to
-dev; - a mislabeled ordinary PR is a green NOTICE no-op;
- a bare-version PR without the
releaselabel refuses; - a re-run of the completed ceremony refuses;
- a tag-door release from a manual tag;
- a mismatched tag refuses.
Every refusal must refuse creating nothing — a probe that leaves a tag or a release behind on a refusal path is a failed probe.
- a merge-door ceremony publishes exactly one release and re-arms main
to
The record
One file per version, drills/X.Y.Z.md — the shape the siblings use: what
was run, where, the result of each probe, failures written down plainly. The
record is the evidence; the scratch repo is the evidence's scaffolding. The
record names the scratch repo by full owner/name and states its disposal
state as its author observed it when the record was written — archived
and pending the operator's delete, or deleted only if the author genuinely
performed the delete. Never a disposal the author did not observe: the
record is the only thing that survives the drill, and 0.2.0's record shipped
its first draft asserting a cleanup that had not happened (#135) — false
evidence in the one file whose job is to be evidence.
actions/drill-recorded refuses any bare-version tree whose record is
missing or blank. A waived drill is still a record: the file says WAIVED and
why — a maintainer's call, visible and reviewable in the release PR's diff,
never a silent skip.