codex-bot-andresmgsl/ceremony
1
0
Fork 0
forked from heavy-duty/ceremony
Code Pull requests Activity

docs(builder): WIP — slim BUILDER.md to the rules, bare local cites

Browse source
This commit is contained in:
cndgrr 2026-08-04 13:11:32 +00:00
parent 2e11855839
commit 6ebdd2efd9
2 changed files with 243 additions and 342 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

580
BUILDER.md
View file

@ -7,226 +7,169 @@ triage bug, and the move is to say so on the issue, not to guess.
## Picking
- Pick from issues labeled **`ready`** — never `blocked`, never `claimed`,
never an `epic` (epics organize; their children are the work).
- Respect dependency order: inside an epic, take the earliest unblocked
unclaimed child. Between epics and strays, prefer the issue that unblocks
the most other work.
- In a repository that adopts version epics, read [RELEASES.md](RELEASES.md)
before choosing among release-window members.
- **Your own red head outranks a new claim.** A failing check at the head
of a PR you authored is picked up **before claiming another issue** —
repairing your own red PR comes ahead of new work, which is why the
engine's duty order evaluates ci-red between resume and build (crew#17:
ceremony#163 sat with full-panel approvals at its head, mergeable, and
stranded on an HTTP 429 in a job that never ran the PR's code, because no
wake covered a red head that owed no round and had no conflict). Red and
green here are the ruled terms of the review round below: a cancelled or
stale check is not a green head; a skipped or neutral one is. The
recovery path (crew#17): inspect the check at the head and record the
failing check and its failure class; rerun a clearly retryable
infrastructure failure without changing code; when the failure belongs to
the branch, return to the normal fix-round and worklog discipline; leave
visible evidence when a rerun cannot be started or the cause is
uncertain; never repeatedly rerun a deterministic branch failure without
a corrective commit; and proceed to handoff once the check is green and
current-head approvals stand. A PR of yours with a red head is **not
parked** — the next move is yours, whatever the round's verdict state
says (shape 2 below carves this out explicitly). How the engine detects a red
head — its ledger, its quiet rules, the rollup's node shapes — is crew's
to describe, not this file's.
never an `epic` (epics organize; their children are the work). Inside an
epic take the earliest unblocked unclaimed child; between epics and strays
prefer the issue that unblocks the most other work. Where a repository
adopts version epics, [RELEASES.md](RELEASES.md) governs the choice among
release-window members.
- **Your own red head outranks a new claim.** A failing check at the head of
a PR you authored is picked up before claiming another issue; a red head
that owes no round and holds no conflict is otherwise nobody's next move,
and the PR strands mergeable (#163). The recovery path: record the failing
check and its failure class; rerun a clearly retryable infrastructure
failure without changing code; when the failure belongs to the branch,
return to the normal fix-round and worklog discipline; leave visible
evidence when a rerun cannot be started or the cause is uncertain; never
repeatedly rerun a deterministic branch failure without a corrective
commit; and hand off once the check is green and current-head approvals
stand. Such a PR is **not parked** — the next move is yours whatever the
round's verdict state says. Red and green are the review round's ruled
terms below; how the engine detects a red head is crew's to describe, not
this file's.
- **One build at a time.** You hold at most one issue on which you are
writing or revising a deliverable — finish or release that work before
starting new work. The rule counts build work in flight, not claims: a
claim does not consume the slot while it is **parked**, meaning the next
move belongs to someone else. Exactly five shapes qualify:
**parked** claim, one whose next move belongs to someone else, does not
consume the slot. Exactly five shapes park:
1. the issue carries `needs-ruling`, its escalation names a decider, and
its `Blocked:` line stops the remaining work;
2. the deliverable is in a review round where every outstanding verdict
belongs to someone else — either the round is awaiting its first
verdicts, or it was answered whole and the owed re-requests posted —
by head, not by verdict: every panelist after a push, the
non-approvers alone at an unchanged head (the review round, steps
1–2). This is the *live* round; shape 4 is
the *passed* one — they are sequential and do not overlap. A red
check at the current head takes the deliverable **out of this
shape**: mid-round CI going red is exactly the state that reads as
"waiting on the panel" and is not — the next move is yours (the
red-head rule above), and reading it as parked is what strands the
PR;
2. the deliverable is in a **live** review round, every outstanding
verdict someone else's — awaiting its first verdicts, or answered whole
with the owed re-requests posted, by head and not by verdict (steps 1–2
below). A red check at the current head takes it **out of this shape**:
that state reads as waiting on the panel and is not, the next move is
yours, and reading it as parked strands the PR;
3. every remaining acceptance criterion is operator-owned, stated as such
by triage on the issue;
4. the deliverable is **handed off** — the round passed, no `blocker:*`
stands, and you set `state:needs-human` per Handoff (below). The
remaining move is the human's merge.
5. the claim is **held by directive** — triage or the operator has told
you to stop, the direction names what the hold waits on, and that thing
is not yours to move. This is not "waiting for a good moment": somebody
else has decided the work must not proceed, and only they end it.
And it ends the same way it started: **on the labels.** When the queue
labels and any prose — an issue body header, a triage comment, an
operator's comment — disagree about whether a hold stands, the most
recent queue-label event by the hold's owner governs, and the prose is
stale until someone corrects it. So before standing down *or* standing
up on a hold, read the issue's **label events**
stands, `state:needs-human` is set per Handoff (below), and the merge
is the human's. Shapes 2 and 4 are sequential and never overlap;
5. the claim is **held by directive** — triage or the operator stopped the
work, the direction names what the hold waits on, and only they end it;
this is never "waiting for a good moment". A hold ends the way it
started, **on the labels**: where the queue labels and any prose
disagree about whether it stands, the most recent queue-label event by
the hold's owner governs and the prose is stale until corrected, an
operator being free to lift by label alone (#149, #151). So before
standing down *or* standing up, read the issue's **label events**
(`gh api /repos/{owner}/{repo}/issues/{n}/timeline`), not only its
comments: an operator may lift by label alone, and on 2026-07-24 did,
twice, on [#149](https://github.com/heavy-duty/ceremony/issues/149)
and [#151](https://github.com/heavy-duty/ceremony/issues/151). Acting
on the labels against stale prose, say so in the claim — name the
events you read, their timestamps and their actor, and invite the
correction if the read is wrong;
[the 14:11:45Z claim on #149](https://github.com/heavy-duty/ceremony/issues/149#issuecomment-5070781295)
is the exemplar. Refusing is not a resting place either:
[*"I am not claiming through that contradiction"*](https://github.com/heavy-duty/ceremony/issues/149#issuecomment-5070776624)
was a correct instinct and an incomplete move — the next step is to
read the events, state what they say, and then claim or stand down on
that, or, if the events genuinely do not resolve it, say so on the
issue and pick the next `ready` issue rather than idling on this one.
Not parked — these are what the rule defends against: waiting on
yourself, waiting on CI (a red head is your own work, above; a pending
one resolves without you), or waiting for a good moment. An issue you have
simply stopped working on is not parked either — that is abandonment,
and its move is unchanged: unassign and restore `ready` (Claiming,
below).
The 2026-07-23 board is why the rule counts work and not claims: one
builder correctly held
[#15](https://github.com/heavy-duty/ceremony/issues/15) (`offsite`,
round answered whole, one verdict outstanding) and
[#16](https://github.com/heavy-duty/ceremony/issues/16) (`needs-ruling`
hard block, triage said hold) parked beside the one active build,
[#73](https://github.com/heavy-duty/ceremony/issues/73).
comments; acting on the labels against stale prose, say so in the claim
— name the events, their timestamps and their actor, and invite the
correction. Refusing to claim through the contradiction is not a
resting place either: where the events genuinely do not resolve it, say
so on the issue and pick the next `ready` issue.
Not parked: waiting on yourself, waiting on CI (a red head is your own
work; a pending one resolves without you), or waiting for a good moment.
An issue you have simply stopped working on is abandoned, not parked —
unassign and restore `ready` (Claiming, below). The rule counts work and
not claims because parked claims are legitimately held beside the one
active build (#15, #16, #73).
## Claiming
- Assign yourself, swap `ready` → `claimed`, and comment that you are
starting. The claim is a promise of a draft PR soon — a claim with no PR
and no activity is what the staleness sweep reclaims unless `offsite`
records that its PR lives in another repository.
- **A park is declared, never inferred.** When your claim enters a parked
shape (Picking, above), say so in a comment on that issue, naming what it
waits on and who owns the next move. No new label: the comment is
activity, so it feeds the same reclaim clock the `needs-ruling`
([#52](https://github.com/heavy-duty/ceremony/issues/52)) and `offsite`
([#68](https://github.com/heavy-duty/ceremony/issues/68)) exemptions
already guard — a parked claim nobody can name is an abandoned one.
Shape 4 alone is exempt from the separate comment: the factual handoff
comment plus the `state:needs-human` write *is* its declaration — both
halves are already there, what the claim waits on (the merge) and who
owns the next move (the human), and both are visible to any scan as a
`labeled` event with the comment beside it. No second comment is owed on
the issue. Every other shape still declares as above.
Declared once, the declaration **stands** until the park's facts change:
a resumption that finds nothing changed posts nothing — the standing
declaration is the record, and silence while parked is compliant, not
abandonment-shaped. Re-declaring on every resume is the flood
[rig#145](https://github.com/heavy-duty/rig/pull/145) drowned in — 38
near-identical audits in one night, each saying nothing changed
([#177](https://github.com/heavy-duty/ceremony/discussions/177)). What
re-opens the duty to comment is the facts changing — the named wait
resolves or changes hands, the parked shape changes, or the claim
unparks — and each owes one new comment. The one place silence has a
cost: a parked claim with **no open PR** still feeds the 48-hour
reclaim clock, so there the builder refreshes the declaration before
the window closes. That refresh is the only repeat a park ever owes,
and its cadence is the reclaim window's, not any duty loop's. None of
this loosens the abandonment rule below: a claim that was never parked
and has simply stopped moving is abandoned, not silent.
- **Pick up `attention` before anything else.** On your claim, first post a
short pickup comment and remove `attention`; the removal is the ack. A
demand on a parked claim is usually its unpark, so take the slot back under
the existing rule below rather than leaving the demand parked. A demand
that *is* the park is different: the pickup comment is the declaration,
so one comment does both jobs, and the demand does not take the slot back.
- **A directed hold keeps its bookkeeping visible.** The PR carries `blocked`
with a comment naming what it waits on; the issue stays `claimed` and
carries `attention` until the builder acknowledges it. Nobody unassigns
the issue, and the 48-hour reclaim does not fire because the claim has an
open PR. Unparking follows the existing rule below.
- **Unparking is a claim like any other.** When the wait ends, the parked
issue is work again and takes the slot. If you are already active
elsewhere, finish or release that work first, and say which you did on
both issues — the slot is still one. Nothing counts claims per builder
and no reconciler path enforces any of this: `claim_decision()` sees one
issue at a time by construction, and no such machinery should be built
expecting it to have been specified here. The discipline is the
declaration, not a counter.
- **Abandoning is fine; ghosting is not.** If you stop, say where you got to,
push the branch if it holds anything useful, unassign, and restore
`ready`.
starting. The claim promises a draft PR soon: a claim with no PR and no
activity is what the staleness sweep reclaims, unless `offsite` records
that its PR lives in another repository.
- **A park is declared, never inferred.** Comment on the issue naming what
the claim waits on and who owns the next move; no new label, because the
comment is the activity that feeds the same reclaim clock the
`needs-ruling` (#52) and `offsite` (#68) exemptions already guard. Shape 4
alone owes no separate comment — the factual handoff comment plus the
`state:needs-human` write already name the wait (the merge) and its owner
(the human), both visible to any scan.
- **A declaration stands until the park's facts change.** A resumption that
finds nothing changed posts nothing, because re-declaring on every resume
floods the record with audits each saying nothing changed (#177); silence
while parked is compliant, not abandonment-shaped. One new comment is owed
each time the facts change — the named wait resolves or changes hands, the
parked shape changes, or the claim unparks. The one place silence costs: a
parked claim with **no open PR** still feeds the 48-hour reclaim clock, so
refresh the declaration before that window closes; that refresh is the
only repeat a park ever owes, at the reclaim window's cadence and not any
duty loop's.
- **Pick up `attention` before anything else.** Post a short pickup comment
and remove `attention`; the removal is the ack. A demand on a parked claim
is usually its unpark, so take the slot back rather than leaving the
demand parked — unless the demand *is* the park, where the pickup comment
doubles as the declaration and the slot stays free.
- **A directed hold keeps its bookkeeping visible.** The PR carries
`blocked` with a comment naming what it waits on; the issue stays
`claimed` and carries `attention` until the builder acknowledges it.
Nobody unassigns the issue, and the 48-hour reclaim does not fire because
the claim has an open PR.
- **Unparking is a claim like any other.** The parked issue is work again
and takes the slot; if you are already active elsewhere, finish or release
that work first and say which you did on both issues. Nothing counts
claims per builder and no reconciler path enforces any of this — the
discipline is the declaration, not a counter, and no such machinery should
be built expecting it to have been specified here.
- **Abandoning is fine; ghosting is not.** Say where you got to, push the
branch if it holds anything useful, unassign, and restore `ready`.
## Building
- Branch per issue; open the PR **as a draft early**, `Closes #N` in the
body. `Closes #N` does not cross repos: when the PR is in a different repo
from its authorizing issue, use `Part of <owner>/<repo>#N` instead, and
in the same step set `offsite` and comment on that issue with the draft PR
link as soon as the draft opens.
Triage closes the authorizing issue by hand when its acceptance criteria
are met; at that handoff the builder reports whether the cross-repo PR
merged or closed and clears `offsite` in the same comment. The cross-repo
merge never closes the authorizing issue. This codifies the linkage
builders already used on rig#112 and ceremony #13/#16 rather than adding a
new review obligation.
`Closes #N` also does not survive a post-merge criterion: when the issue's
body states that an acceptance criterion can only be checked after the
merge — a live proof of a workflow trigger, a released-artifact check,
anything whose subject does not exist until the change is on the base
branch — the same-repo PR uses `Refs #N` instead, and triage closes the
issue by hand on the evidence, exactly as it does for cross-repo work. The
body. Drafts are invisible to the reviewer panel on purpose: the draft
phase is yours.
- **`Closes #N` does not cross repos.** A PR in a different repo from its
authorizing issue says `Part of <owner>/<repo>#N`, sets `offsite`, and
comments the draft PR link on that issue in the same step; triage closes
that issue by hand once its acceptance criteria are met, and at that
handoff the builder reports whether the PR merged or closed and clears
`offsite` in the same comment. The cross-repo merge never closes the
authorizing issue (#13, #16).
- **`Closes #N` does not survive a post-merge criterion.** Where the issue's
body states that a criterion can only be checked after the merge — a live
proof of a workflow trigger, a released-artifact check, anything whose
subject does not exist until the change is on the base branch — the
same-repo PR says `Refs #N` and triage closes by hand on the evidence. The
merge releases the claim: the issue moves to `post-merge`, the builder
walks away, and triage owns verification and closure. If evidence later
requires corrective build work, triage returns it to `ready` or mints a
fresh `ready` issue; any builder claims from current `main`, and the
original builder has no special standing.
The issue body is what says so; you never judge which issues qualify, and
absent that instruction `Closes #N` remains the default. The exception was
bought the hard way: #143 carried `Closes #137` as doctrine then required,
and the merge closed #137 with its post-merge criterion unmet (#151).
On a `Refs #N` PR, never put a closing keyword (`close`, `closes`,
walks away, and triage owns verification and closure; corrective work is a
fresh `ready` issue any builder claims from current `main`, the original
builder holding no special standing. The issue body is what says so — you
never judge which issues qualify, and absent that instruction `Closes #N`
remains the default (#151).
- On a `Refs #N` PR, never put a closing keyword (`close`, `closes`,
`closed`, `fix`, `fixes`, `fixed`, `resolve`, `resolves`, `resolved`)
immediately before `#N` anywhere in the body — including the sentence
explaining why the PR does not close it. GitHub reads the whole body by
adjacency, not intent. Put the number first (`#N is closed by hand`) or
omit it (`triage closes the issue by hand`). A code span does not protect
the phrase: a backticked `Closes #199` still closed #199 (#200, #218).
Drafts are invisible to the reviewer panel on
purpose — the draft phase is yours.
immediately before `#N` anywhere in the body, including the sentence
explaining why the PR does not close it: GitHub reads the whole body by
adjacency, not intent, and a code span does not protect the phrase (#200,
#218). Put the number first (`#N is closed by hand`) or omit it.
- **The issue's acceptance criteria are your definition of done.** Reproduce
them as a checklist in the PR body and check them honestly as you go. If
one turns out to be wrong or unreachable, say so on the issue and get it
amended by triage — do not silently ship less than the issue says.
- Every behavior change writes one fragment, `changelog.d/<issue>.md`,
named for the authorizing issue (`<repo>-<issue>.md` when the work is
cross-repo) — the exact prose that will be published, nothing else: `- `
bullets, and in a grouped repo the `### Added` / `### Changed` /
`### Fixed` headings inside the fragment, creating a rarer kind only when
a change genuinely is one. An entry is at most 300 characters — the
fragment guard reds longer (#167) — so a genuinely long change ships
several short entries, never one long one; wrapping an entry over
continuation lines is fine and never counts against it. Every entry
**ends with its issue citation**, and the same guard reds an entry
without one: a single `(` group of `#N`, `repo#N` or `owner/repo#N`
references separated by `, `, then `)`, then the final `.` and nothing
after it — `(#262).` locally, `(#236, #250).` when one entry honestly
lands two. The citation need not name the fragment's own issue, because
the filename already carries the authorizing one (#262). Never edit
`CHANGELOG.md` for an entry — the
release PR assembles the section from the fragments (#112); the monotonic
guard still refuses anything that deletes a shipped heading.
them as a checklist in the PR body and check them honestly as you go; a
criterion that turns out wrong or unreachable goes back to triage to be
amended, never silently shipped short.
- **Every behavior change writes one fragment**, `changelog.d/<issue>.md`,
named for the authorizing issue (`<repo>-<issue>.md` cross-repo): the
exact prose that will be published and nothing else — `- ` bullets, and in
a grouped repo the `### Added` / `### Changed` / `### Fixed` headings
inside the fragment, a rarer kind only when a change genuinely is one. An
entry is at most 300 characters, so a genuinely long change ships several
short entries, never one long one; wrapping an entry over continuation
lines is fine and never counts against it (#167). Every entry **ends with
its issue citation**: a single `(` group of `#N`, `repo#N` or
`owner/repo#N` references separated by `, `, then `)`, then the final `.`
and nothing after it — `(#262).` locally, `(#236, #250).` when one entry
honestly lands two. The citation need not name the fragment's own issue,
which the filename already carries (#262). The fragment guard reds a
longer entry and an uncited one alike. Never edit `CHANGELOG.md` for an
entry: the release PR assembles the section from the fragments (#112), and
the monotonic guard refuses anything that deletes a shipped heading.
- Follow the repo's conventions file and match the code you touch. Tests are
not optional: the issue's test plan is the floor, not the ceiling.
- **A write-capable job gets a repo-owned script, not a third-party action.**
If the job's token can write (`packages: write`, `contents: write`,
Where the job's token can write (`packages: write`, `contents: write`,
`id-token: write`, deploy secrets), default to a script in the repo that a
test can drive; a third-party action there needs an established publisher
and a full-commit-SHA pin. Read-only jobs still SHA-pin. The full rule and
the red-flag profile a reviewer will apply are in REVIEWER.md §What you
review against, item 2 (incubator#53/#54; #216).
and a full-commit-SHA pin, and read-only jobs still SHA-pin. The full rule
and the red-flag profile a reviewer will apply are in REVIEWER.md §What
you review against, item 2 (#216).
- **Scope discipline: the PR does the issue — whole, and nothing else.**
Adjacent problems you discover go to a **discussion** (or a comment on the
relevant issue), where triage will do its job. You do not mint issues —
Adjacent problems you discover go to a **discussion**, or a comment on the
relevant issue, where triage will do its job. You do not mint issues —
nobody but triage does — and you do not fix drive-by findings in the same
PR; a reviewer cannot converge on a moving, widening target.
PR, because a reviewer cannot converge on a widening target.
## The review round
@ -234,149 +177,105 @@ triage bug, and the move is to say so on the issue, not to guess.
repo-specific facts such as the panel roster live in that repo's own
CONTRIBUTING; the shared flow lives here and is not restated there.)
1. Mark ready-for-review; request **the whole panel**. The panel is the PR
repo's `panel[<your-login>]=` line if it defines one, else its `panel=`
line; minus the author in either case (#224) — and never the roster of
the repo the issue is in. The PR repo's `.github/labels.conf` is the
machine's answer; its CONTRIBUTING roster is the human-readable answer,
and the conf governs if they disagree because that is what the state
machine reads. If the PR repo names no roster, ask triage on the
authorizing issue before marking ready-for-review; do not guess. You may
request an off-panel reviewer, but say that their verdict is advisory and
does not become required. On rig#112 this distinction mattered: requesting
codex and grok was correct for rig's panel even though ceremony's bench was
larger, and the doctrine had not said which roster governed.
**A review request requires a green check at the head.** A red check is
the author's own signal, not the panel's work: if the check is red, that
is your next task, not the panel's — fix it and push, then request. This
binds *you*, whether or not any engine enforces it. "My local suite
passed" is evidence about your machine; the check at the head is the
shared artifact the panel actually reads, and a reviewer's first act is
to read it. The one exception is a failure genuinely outside the PR — a
runner outage, a flaky dependency, a failure already present on the
default branch — and it is an exception only if the request says so
explicitly and names the evidence (e.g. "the same job fails identically
on `origin/main` at `<sha>`"). Silence about a red check is what is
prohibited; an argued exception shifts the burden to the author.
*Green* is a ruled term (operator, 2026-07-27), and it is read in two
steps, because a head carries more rollup entries than it has checks:
first pick the entry that is a check's word at this head, then
classify that entry. **A check's word at a head is its newest entry
by start time, and a `CANCELLED` entry is not that word while the
same check carries a non-cancelled entry at the same head.** The
survivor is the verdict about these bytes; the entry it displaced
reported nothing about them. Say **start** time and mean it: a
cancelled run does not stop the moment its replacement begins, so the
dead run's completion routinely postdates the live run's start, and a
reader who dates entries by completion picks the corpse. When *every*
entry a check has at the head is cancelled, nothing survives to be
its word: that check has not reported at all, and it stays not-green
by the classes below — the all-cancelled context is the case this
leaves exactly where it was. This states a collapse and not a new
class: `checks_state`'s carve-out drops a cancelled entry only where
its context keeps a non-cancelled survivor, and leaves an
all-cancelled context intact and still blocking, so doctrine and gate
partition alike on a mixed context (#139, #276). What the *machine*
drops from the rollup before it grades anything is a different
question, and crew's to describe rather than this file's.
Then classify that entry, and classify it from its **`conclusion`**,
never its `status`: a check carrying a terminal conclusion is green or
not-green by that conclusion whatever its `status` field still
reports — the two can disagree, and on #259 a finished job's `status`
lagged its own `conclusion: success` at the head. A check with no
conclusion at all is neither class: a configured run still in progress
is not green, and waiting for it is compliance, not a stall. Picking
the newest entry never settles a live one: where the survivor is the
run still going, the head is not green and you wait on it exactly as
you would have. A **cancelled or stale** check is not a green head —
1. Mark ready-for-review; request **the whole panel**: the PR repo's
`panel[<your-login>]=` line if it defines one, else its `panel=` line,
minus the author in either case (#224) — never the roster of the repo the
issue is in. That repo's `.github/labels.conf` governs over its
CONTRIBUTING roster, being what the state machine reads; where the PR
repo names no roster, ask triage on the authorizing issue before marking
ready-for-review rather than guessing. You may request an off-panel
reviewer, saying that their verdict is advisory and does not become
required.
**A review request requires a green check at the head**, and this binds
you whether or not any engine enforces it: a red check is the author's
own signal, not the panel's work, so fix it and push, then request. The
one exception is a failure genuinely outside the PR — a runner outage, a
flaky dependency, a failure already present on the default branch — and
only if the request says so explicitly and names the evidence ("the same
job fails identically on `origin/main` at `<sha>`"); silence about a red
check is what is prohibited, while an argued exception shifts the burden
to the author.
*Green* is a ruled term (operator, 2026-07-27), read in two steps,
because a head carries more rollup entries than it has checks. **First
pick the entry that is a check's word at this head: its newest entry by
start time, a `CANCELLED` entry never being that word while the same
check carries a non-cancelled entry at the same head** — say *start* and
mean it, since a cancelled run does not stop when its replacement begins
and a reader who dates entries by completion picks the corpse. Where
*every* entry a check has at the head is cancelled, nothing survives to
be its word: that check has not reported, and it stays not-green by the
classes below. That is a collapse and not a new class — the gate's
carve-out likewise drops a cancelled entry only where its context keeps a
non-cancelled survivor, leaving an all-cancelled context blocking, so
doctrine and gate partition alike on a mixed context (#139, #276).
**Then classify that entry from its `conclusion`, never its `status`**,
which can still disagree with it (#259). An entry with no conclusion is
neither class: a configured run still in progress is not green, and
waiting for it is compliance, not a stall, so picking the newest entry
never settles a live one. **Cancelled or stale** is not a green head —
*stale* means a check belonging to a superseded head, which the
head-scoped rollup does not show anyway, so what survives there is
same-head cancellation, never a same-head node whose `status` lags its
conclusion — while a **skipped or neutral** one *is* green: those are
deliberate "passed / not applicable" conclusions, and reddening them
would red every conditional job the fleet skips on purpose. And a head
with **no checks configured** is the third ruled case, not an argued
exception: nothing is configured, so there is nothing to wait for —
the precondition is satisfied and the request goes out straight away,
no evidence or explanation owed, because the argued-exception path
above exists for a check that ran and came up red. This rules
nothing-configured, never nothing-answered-yet: a pending run has an
owner, CI, and is waited on as above. The machine partitions the same
way — `blocker:unrequested` admits the ask on `SUCCESS` and on `NONE`
alike (#236) — so doctrine and gate state one rule and each points at
the other. The costs behind the line are asymmetric: a false green
spends a three-reviewer round; a false red spends one author session.
same-head cancellation — while **skipped or neutral** *is* green, those
being deliberate "passed / not applicable" conclusions whose reddening
would red every conditional job the fleet skips on purpose. A head with
**no checks configured** is the third ruled case, not an argued
exception: nothing is configured, so there is nothing to wait for and the
request goes out straight away with no evidence owed, the
argued-exception path existing for a check that ran and came up red. That
rules nothing-configured, never nothing-answered-yet: a pending run has
an owner, CI, and is waited on as above, and the machine partitions the
same way, admitting the ask on `SUCCESS` and on `NONE` alike (#236). The
costs behind the line are asymmetric: a false green spends a
three-reviewer round; a false red spends one author session. What the
*machine* drops from the rollup before grading is crew's to describe.
2. **Wait for every verdict, then answer the round whole** — one reply
covering every point and stating what changed and what was verified.
That reply is the written round record: the engine mirrors it under the
PR body's **Round log**, newest last, so the builder owes the reply and
no separate body edit. At re-request time the engine takes the author's
comments posted after the newest verdict in the round and appends them
with `<!-- round:<head-sha> -->`; an existing marker makes a retry a
no-op. If the builder posted no reply, the engine records that the round
passed without one and never blocks handoff on the omission. Then push
the fixes, then re-request **by head, not by verdict**: if answering the
round pushed any commit, every
panelist's approval is now stale — an approval is of a specific tree,
and the handoff predicate counts only approvals at the current head —
so **every panelist is re-requested, the approvers included**; a
panelist left un-re-requested after a push can never approve the tree
you shipped, and the PR sits looking finished with a full set of
verdicts and nothing owed by anyone, the same silent-stall shape as
[#26](https://github.com/heavy-duty/ceremony/issues/26)/[#39](https://github.com/heavy-duty/ceremony/issues/39).
Only when the head did not move — the round was answered with argument
or evidence and nothing was pushed — do you re-request just the
non-approvers: a standing approval already covers this exact head, and
the engine absorbs a re-request at an unchanged head (the re-request
rule, [#94](https://github.com/heavy-duty/ceremony/issues/94); its
covering every point and stating what changed and what was verified. That
reply is the written round record: the engine mirrors it under the PR
body's **Round log**, newest last, appending the author's comments posted
after the round's newest verdict with `<!-- round:<head-sha> -->` (an
existing marker makes a retry a no-op), so the builder owes the reply and
no separate body edit; a round the builder left unanswered is recorded as
such and never blocks handoff.
Then push the fixes, and re-request **by head, not by verdict**. A push
makes every approval stale — an approval is of a specific tree, and the
handoff predicate counts only approvals at the current head — so **every
panelist is re-requested, the approvers included**; a panelist left
un-re-requested after a push can never approve the tree you shipped, and
the PR sits looking finished with a full set of verdicts and nothing owed
by anyone (#26, #39). Only where the head did not move — the round
answered with argument or evidence, nothing pushed — do you re-request
just the non-approvers, a standing approval already covering this exact
head and the engine absorbing a re-request at an unchanged one (#94; its
mechanism is crew's to describe). **The re-request carries the same
green-check-at-head precondition as the first request**, argued
exception included. This is where the measured cost landed: crew#40
burned two consecutive heads and four reviewer-rounds, every one
relaying a CI failure already visible in the job log (crew#45). A fix
push whose check comes up red is not ready to go back to the panel; it
is your next fix. Prefer verification over argument: when a
reviewer doubts behavior, add the test that settles it.
green-check-at-head precondition as the first request**, argued exception
included: a fix push whose check comes up red is your next fix, not the
panel's. Prefer verification over argument — when a reviewer doubts
behavior, add the test that settles it.
3. Never dismiss a review, never merge, never mark your own work as passed.
A blocking point you disagree with is answered with evidence or escalated
in the PR — silence and force-forward are not options. A panel deadlock
is one kind of human-owned decision; use the ruling ask below
([#50 D11](https://github.com/heavy-duty/ceremony/issues/50)).
in the PR; silence and force-forward are not options, and a panel
deadlock is one kind of human-owned decision (#50 D11).
**A fix round may ride a draft.** An engine may convert a PR back to draft
when a round closes; crew#139 proposes exactly that, and is still an open
proposal. What it names is the status quo without it: where an engine's own
rules make a builder's mid-round pushes *saves* rather than proposals, every
one of those saves fires CI while the PR sits ready — 41 of 106 commits
across crew's last 25 PRs, by that issue's measurement — and converting back
to draft is what would stop them. Ceremony implements no such conversion and
this passage specifies none; it is written down because a builder or a
reviewer who meets a mid-round draft has to find a state the doctrine
describes. What it means is what a draft already meant while you were
building, extended and not changed: the draft phase is yours and the panel
cannot see it (Building, above). Whose ball it is does not change either —
the round outranks the draft, so you still owe it whole, the fixes and the
reply and the flip. The label axis says the same thing in the machine's
voice rather than in this one, and [LABELS.md](LABELS.md)'s `state:building`
row is where to read it (#205).
**Ready-for-review is the act that ends the round, and it is the builder's
alone.** No engine marks a PR ready. The flip asserts that the round was
answered whole, and that assertion is the one judgement about a round its
author cannot delegate to a machine: an engine may draft a PR, which is what
crew#139 proposes engines do, but only the builder undrafts it.
**Where a draft suppressed the checks, green is proven at the flip and the
request still follows it.** Step 1's precondition is the whole rule and this
adds no second one — it says only *when* the head answers: marking ready is
what runs the checks the draft held back, so the order is flip, let the head
answer, then request, and the argued exception stays the only way past a red
one. Waiting there is compliance, not a stall, and the machine reads it that
way too: `blocker:unrequested` does not fire while a head's checks are pending
or red, because the one blocker that demands an act has to know when the act
is permitted (#236 — crew#318 carried it at ~12:44Z on 2026-08-03 while its
head's run was still in progress, which is the label flagging a builder for
obeying this section).
**A fix round may ride a draft**, and the draft changes nothing about who
owes what. An engine may convert a PR back to draft when a round closes, so
that mid-round saves stop firing CI on a ready PR; ceremony implements no
such conversion and this passage specifies none, but whoever meets a
mid-round draft reads it as the draft always read — the draft phase is yours
and the panel cannot see it (Building, above) — while the round outranks the
draft, so you still owe it whole, the fixes and the reply and the flip
([LABELS.md](LABELS.md)'s `state:building` row says the same in the
machine's voice, #205). **Ready-for-review is the act that ends the round,
and it is the builder's alone**: the flip asserts that the round was
answered whole, which is the one judgement about a round its author cannot
delegate to a machine, so an engine may draft a PR but only the builder
undrafts it. **Where a draft suppressed the checks, green is proven at the
flip and the request still follows it**: marking ready is what runs the
checks the draft held back, so the order is flip, let the head answer, then
request — step 1's precondition and not a second one — and the argued
exception stays the only way past a red one. Waiting there is compliance,
not a stall, and the machine reads it the same way: `blocker:unrequested`
does not fire while a head's checks are pending or red, because the one
blocker that demands an act has to know when the act is permitted (#236).
## The ruling ask
@ -384,12 +283,11 @@ Set `needs-ruling` whenever a decision belongs to a human: org policy,
published artifacts, secrets, prod, or any choice whose cost lands outside
the PR. A panel deadlock is one instance, not the definition. The builder is
the accountable flag-setter on a PR and consolidates the decision into one
comment rather than forwarding several reviewers' phrasings
([#50 D11](https://github.com/heavy-duty/ceremony/issues/50)).
comment rather than forwarding several reviewers' phrasings (#50 D11).
Keep at most these five lines above the fold and put all other analysis
inside the fold. The field labels are fixed because the ruling machinery
checks for them ([#50 D12](https://github.com/heavy-duty/ceremony/issues/50)):
checks for them (#50 D12):
```text
🧭 needs-ruling — <the decision, one line>
@ -406,11 +304,10 @@ whole problem to the human. `Blocked:` names both what stops and what
continues. Write a timed `Default:` only when you are affirmatively confident
the decision is reversible inside the PR before merge. Unsure is not a tie:
it is a hard block. Published artifacts, secrets, prod, and org policy are
hard blocks by construction ([#50 D12–D13](https://github.com/heavy-duty/ceremony/issues/50)).
hard blocks by construction (#50 D12–D13).
The ladder is anchored to the current episode's `needs-ruling` **`labeled`
event**, not its `Default:` deadline or the last activity
([#50 D13–D14](https://github.com/heavy-duty/ceremony/issues/50)):
event**, not its `Default:` deadline or the last activity (#50 D13–D14):
- **0–12h:** proceed when a still-clear, reversible default expires, and say
out loud that you did. A hard block waits.
@ -446,10 +343,9 @@ builder's behalf, in order:
the head SHA, and a pointer to the PR body's **Round log**.
The builder composes no new summary at handoff: the authored record already
lives in the Round log, mirrored mechanically from each whole-round reply as
specified above. The label write is optimistic — the reconciler validates
it, and takes it back if the PR is not actually mergeable-right-now. Then
stop: the PR is the human's. The claim is now parked as shape 4 (Picking,
above) — the handoff you just posted is its declaration, and your build slot
is free. Address what comes back (`state:addressing`) and re-hand-off the
same way.
lives in the Round log, mirrored mechanically from each whole-round reply.
The label write is optimistic — the reconciler validates it and takes it
back if the PR is not actually mergeable-right-now. Then stop: the PR is the
human's, and the claim is now parked as shape 4 (Picking, above), the
handoff comment being its declaration and your build slot free. Address what
comes back (`state:addressing`) and re-hand-off the same way.

5
changelog.d/281.md Normal file
View file

@ -0,0 +1,5 @@
### Changed
- `BUILDER.md` states its rules and cites their record bare: the incident
narratives, the links into issue comments and the cross-repo issue cites
leave the normative text, which no rule leaves with them (#281).