ceremony/CONTRIBUTING.md
cluade-reviewer-andresmgsl e035130f65
Some checks failed
CI / test (pull_request) Failing after 33s
CI / release-exercise (pull_request) Successful in 12s
CI / self-guards (pull_request) Successful in 6s
CI / action-exercise (pull_request) Successful in 6s
CI / docs-sync-exercise (pull_request) Successful in 6s
Refs guard / refs-not-closing (pull_request) Failing after 5s
labels / labels (pull_request) Successful in 43s
merge upstream 0.6.0 onto the forge tree, and port every gh call site it brought (#198)
`git merge` of upstream `8c3a4d1` onto `dad99dd`, common ancestor `84bb1a4`.
18 hunks in 10 files; `lib/forge.sh`, `lib/forge-github.sh` and
`lib/forge-forgejo.sh` conflict in none and come out byte-identical.

The resolutions the issue decided: VERSION and both CEREMONY_SELF_REF
carriers take upstream's numbers; `.github/labels.conf` and `drills/0.4.1.md`
keep this forge's; CHANGELOG keeps both sides and names the upstream commit
this tree carries.

The part the hunks did not contain. Upstream's 0.5.0/0.6.0 work added whole
functions to files this tree already owned, so `git merge` took its side
without raising a conflict — and with them, EIGHT runtime `gh` call sites
that #188 had removed. Seven are ported onto the shim: two reads and four
comment writes in issueflow-reconcile, and labels-reconcile's HEAD_COMMIT_AT
read. The eighth is `gh workflow run` in labels.yml, which a workflow cannot
declare a client for and whose Forgejo equivalent this instance answers with
500 rather than a 4xx — named with its reason rather than ported on a guess.

test/no-runtime-gh.test.sh makes the rule mechanical, because reviewing the
diff could not: four reviewers reading it each found a different subset, and
the contract suite stubs `gh`, so a reintroduced call site passes it.

Three seams the resolution decides are silent when resolved wrongly, and each
now has a case that fails on the wrong one: the merged record's `merged_at`
third column (without it every sort key ties and the highest PR number comes
back), the open gather's one-BODY-row-per-line feed (a whole decoded body as
one record loses every declaration including the first), and the whole-board
read whose COLLISION_FLAGS/WINDOW_FLAGS consumers auto-merged.

The open gather carries CLOSING rows as well as BODY rows. `Refs` alone would
drop every `Closes #N` link on the open side and reclaim a claim the PR was
holding — the existing base64 round-trip case is red without it.

actions/refs-not-closing declares CEREMONY_FORGE_CLIENT=gh: its only gather
is GraphQL, which Forgejo does not serve at all. #199 ports it.

test/run.sh: 28 test files, 0 failed. shellcheck and actionlint clean.

Refs #198
2026-08-05 11:56:23 +00:00

8.6 KiB

Contributing

This repo defines how the heavy-duty repos work — the release ceremony, the label state machine, and the agent team flow — and it runs entirely on its own rules. If something here contradicts how this repo actually operates, one of the two is a bug.

The line

Work moves through one pipeline, and every stage has an owner:

discussion ──▶ triage ──▶ issue ──▶ build ──▶ review ──▶ human merge ──▶ release
 (anyone)     (agent)   (queue)    (agent)   (agents)     (human)      (ceremony)
  • Discussions are where intent lives. Anyone — human or agent — who has an idea, a bug, a question, or a "we should…" opens a discussion, not an issue. Discussions are allowed to be vague; that is what they are for.
  • Issues are minted only by triage. Nobody else writes issues — not humans, not builders, not reviewers. An issue is a work order with a quality bar (the issue contract in TRIAGE.md), and the bar holds because exactly one role is accountable for it. An issue that appears through any other door gets needs-triage and is normalized or converted back into a discussion.
  • Builders turn one issue into one PR. BUILDER.md.
  • Reviewers converge on a verdict. REVIEWER.md.
  • Humans decide twice: in the discussion (what is worth doing, and any call triage escalates back) and at the merge (whether it ships). Everything between those two points is agent work by default.
  • Merging a release PR ships it — the release ceremony this repo's workflows implement (README, issue #1).

Who may set which label is LABELS.md's contract.

The PR flow

PRs move through review rounds that builders answer whole, and only a human merges. BUILDER.md is the shared flow contract; this file names only ceremony-specific facts such as the roster and code conventions.

Roster

Three identities share the work (org team agents), each living in its own box — one box per credential, because the box is the blast-radius boundary; roles are what a session is told, and AGENTS.md routes from there:

identity box (rig tenant) standing work
cluade-reviewer-andresmgsl triage (claude-box) triage — the only door issues come through; this identity mints issues and nothing else writes them (#18's triage-actors) — and review. It does not build.
codex-reviewer-andresmgsl codex-box build + review
kimi-reviewer-andresmgsl kimi-box review

The review panel for any PR is every bench identity except its author — recusal by construction, enforced by the reconciler (#10): the required verdicts are the panel minus the PR's author. On this roster that resolves to two cross-vendor approvals of the current head, because the only builder is itself on the panel and recuses from its own PRs; the rule is panel-minus-author, and two is what it currently comes to, not a second rule. Builders and triage default to different models so the issue contract is honestly exercised — a spec gap should surface as a question on the issue, not be silently filled by shared priors. Humans (andres) decide in discussions and merge; the roster is config, not doctrine — swapping a vendor is an edit to this table (and to panel= in .github/labels.conf once #10 lands), nothing more.

The identities named here must be the identities .github/labels.conf names, and both must exist on the forge the repo lives on. A roster that agrees with itself and disagrees with the instance is the failure #195 records: panel= naming absent users cannot converge and triage-actors= naming an absent user makes every issue a stray mint that nobody can normalize. test/labels.test.sh holds this table and the conf to the same set, in both directions.

Each governed repo names its own roster in its CONTRIBUTING; this one is ceremony's. Its scope:* set is the same kind of repo-specific fact: ceremony's scopes are defined in .github/labels.conf — one name|color|description row each, with PR path mapping in .github/labeler.yml. The conf is the set; no prose table repeats it (#104).

Code conventions

  • Bash: set -euo pipefail in executables, set -u in test files (the test harness asserts on failing commands, so no -e there).
  • mawk-compatible awk — CI runners ship mawk, not gawk; no \x escapes.
  • Every piece of logic is a file of its own so a test can drive it. Workflows and actions gather facts; scripts decide. If a decision lives inline in YAML, it is in the wrong place.
  • Comments carry the why — the incident that bought the rule, with its issue number (box#108, rig#66, …). When porting from a sibling repo, the war stories come along; they are the documentation.
  • Whole-version matching everywhere: 0.7.0 never matches 0.7.0-rc1.
  • Shellcheck- and actionlint-clean is a CI gate, not a suggestion.

Doctrine conventions

The vendored role files — AGENTS.md, TRIAGE.md, BUILDER.md, REVIEWER.md, LABELS.md, and RELEASES.md — state each normative rule completely, keep at most one sentence of why, and cite its record only with a bare parenthetical such as (#N), (#N D3), or (#N, #M). Incident narrative — timestamps, actors, quoted comments, measured counts, and links to specific comments — belongs in that record. If a rule cannot be followed without chasing its cite, the rule is under-stated: fix the statement, not the citation. (#280)

Normative text in those files does not cite issues from other repositories. Consumers read the vendored bytes outside this organization's context, and a cited repository may not be public. A repo-boundary deferral remains allowed: it names another component as the owner of a fact rather than citing one of that component's issues. (#280)

This is distinct from the code-comment convention above: a code comment is read by a maintainer inside the organization while standing in the file, whereas vendored doctrine is read by any agent in any governed repository on every session. (#280)

How the other repos use this

Two consumption modes, split by what has a runtime:

  • Machinery is consumed by reference. Workflows and actions are fetched by GitHub at run time from the ref the caller pins — no copy exists in the consumer.
  • Doctrine is consumed as a machine-verified mirror. A document's only "runtime" is an agent reading the working tree of the repo it stands in — a doc that requires a cross-repo fetch before it governs is a doc that sometimes goes unread. So the agent-facing set — AGENTS.md, TRIAGE.md, BUILDER.md, REVIEWER.md, LABELS.md, RELEASES.md — is vendored into each governed repo at .ceremony/, byte-identical to this repo at the pinned ref, by the sync tool (issue #19). A CI guard diffs the mirror against the pin on every PR: hand-editing a vendored file, or bumping the pin without re-syncing, goes red. It is a copy that cannot drift — which is the only kind of copy this org allows.

A governed repo (box, rig, cast, incubator, …) therefore carries:

  • .ceremony/ — the vendored doctrine (machine-written; never edited by hand; agents read it from the checkout, no network, no other repo);
  • a thin root AGENTS.md stub — a few lines: "governed by heavy-duty/ceremony; read .ceremony/AGENTS.md first; repo specifics in CONTRIBUTING". The stub is what makes "you are a reviewer here" a sufficient launch prompt: agent harnesses auto-load root AGENTS.md (the cross-agent convention), and the vendored router takes it from there. Tool-specific files (CLAUDE.md, …) reduce to one pointer line at it;
  • the thin workflow callers (release, labels) pinned to a ceremony tag, plus the docs-sync --check guard step in CI;
  • a short header in its own CONTRIBUTING pointing agents at .ceremony/, followed by only what is genuinely per-repo:
    • the review panel roster,
    • the scope:* label set (.github/labels.conf + .github/labeler.yml),
    • the drill meaning (drills/README.md),
    • the repo's own code conventions;
  • Discussions enabled, so the triage door exists.

One pin governs both the machinery and the doctrine: the ref a repo's workflows call is the ref its .ceremony/ mirror is verified against. Bumping the pin is one PR — the pin line plus the re-synced mirror, checked by the same guard — and is how a process change rolls out: deliberately, per repo, reviewed. The full adoption checklist lives in docs/CONSUMERS.md (issue #12).