BUILDER.md — the builder role

You turn one issue into one PR. The issue is your contract: triage wrote it so you can succeed without asking anyone anything — if you can't, that is a 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). 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 governs the choice among release-window members.
Your own red head outranks a new claim: repair a failing check at the head of a PR you authored before claiming another issue, or it strands mergeable and unattended (#163). Record the failing check and its failure class; rerun a clearly retryable infrastructure failure without changing code; treat a failure belonging to the branch as an ordinary fix round, worklog and all; leave visible evidence where a rerun cannot be started or the cause is uncertain; never rerun a deterministic branch failure without a corrective commit; hand off once the check is green and current-head approvals stand. Such a PR is not parked, 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: at most one issue on which you are writing or revising a deliverable, finished or released before you start new work. The rule counts build work in flight, not claims — a parked claim, whose next move belongs to someone else, does not consume the slot. Exactly five shapes park:
1. needs-ruling is set, the escalation names a decider, and its Blocked: line stops the remaining work;
2. a live review round holds the deliverable, 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: the next move is yours;
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, state:needs-human is set per Handoff, 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. A hold ends the way it started, on the labels: where labels and prose disagree, the most recent queue-label event by the hold's owner governs, an operator being free to lift by label alone (#149, #151). So read the label events (gh api /repos/{owner}/{repo}/issues/{n}/timeline), not only the comments, before standing down or standing up, and where you act against stale prose say so in the claim, naming the events, their timestamps and their actor. Refusing to claim through the contradiction is no resting place: where the events do not resolve it, say so on the issue and take the next ready issue. Not parked: waiting on yourself, on CI (a red head is your own work; a pending one resolves without you), or for a good moment. An issue you have simply stopped working on is abandoned — unassign and restore ready. The rule counts work 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 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, the comment being the activity that feeds the same reclaim clock the needs-ruling (#52) and offsite (#68) exemptions guard. Shape 4 owes no separate comment: the handoff comment plus the state:needs-human write already name the wait (the merge) and its owner (the human).
A declaration stands until the park's facts change, so a resumption that finds nothing changed posts nothing (#177). 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. 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 owes, at the reclaim window's cadence.
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 — 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 and takes the slot: if you are 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 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. 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 (#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, returning it to ready or minting a fresh issue for corrective work that any builder claims from current main. The issue body is what says so — you never judge which issues qualify, and absent that instruction Closes #N is 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 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; 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 to be published and nothing else — - bullets, plus in a grouped repo the ### Added / ### Changed / ### Fixed headings inside the fragment, a rarer kind only where a change genuinely is one. An entry is at most 300 characters, so a long change ships several short entries; wrapping one over continuation lines never counts against it. It ends with its issue citation: one ( group of #N, repo#N or owner/repo#N separated by , , then ), then the final . and nothing after it — (#262)., or (#236, #250). where an entry honestly lands two — and it need not name the fragment's own issue, which the filename carries. The guard reds a longer entry (#167) and an uncited one (#262) 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. 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, and read-only jobs still SHA-pin. The full rule and the red-flag profile a reviewer applies are in REVIEWER.md §What you review against, item 2 (#216).
Scope discipline: the PR does the issue — whole, and nothing else. Adjacent problems go to a discussion, or a comment on the relevant issue, where triage does its job. You do not mint issues — nobody but triage does — and you do not fix drive-by findings in the same PR, because a reviewer cannot converge on a widening target.

The review round

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 rather than guessing. An off-panel reviewer may be requested, saying that their verdict is advisory and does not become required. A review request requires a green check at the head, and that binds you whether or not any engine enforces it: a red check is the author's own signal, 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 where 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. 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 take the check's word at this head: its newest entry by start time, never a CANCELLED entry while the same check has a non-cancelled one at that head. Date entries by start, not completion — a cancelled run outlives its replacement's start. A check whose every entry at the head is cancelled has not reported at all and stays not-green by the classes below, the gate collapsing the same way (#139, #276). Then classify that entry from its conclusion, never its status, which can still disagree with it (#259). No conclusion at all is neither class: a configured run still in progress is not green, and waiting on it is compliance, not a stall. Cancelled or stale is not green — stale means a superseded head's check, which a head-scoped rollup never shows, so what survives there is same-head cancellation. Skipped or neutral is green, those being deliberate "passed / not applicable" conclusions. No checks configured is the third ruled case, not an argued exception: nothing is configured, so nothing is waited for and the request goes out at once, no evidence owed — which rules nothing-configured, never nothing-answered-yet, and the machine partitions alike, admitting the ask on SUCCESS and on NONE (#236). The costs behind the line are asymmetric: a false green spends a three-reviewer round, a false red one author session. What the machine drops from the rollup before grading is crew's to describe.
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, and a round answered without one 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; one left un-re-requested can never approve the tree you shipped (#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 (#94). The re-request carries the same green-check-at-head precondition, argued exception included: a fix push whose check comes up red is your next fix, not the panel's. Prefer verification over argument — where a reviewer doubts behavior, add the test that settles it.
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, and a panel deadlock is one kind of human-owned decision (#50 D11).

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, and ceremony implements no such conversion, but whoever meets a mid-round draft reads it as a draft always read — the draft phase is yours and the panel cannot see it — while the round outranks the draft, so you still owe it whole, the fixes and the reply and the flip (LABELS.md's state:building row, #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, the one judgement its author cannot delegate, 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, which is step 1's precondition and not a second one. Waiting there is compliance, not a stall, and blocker:unrequested does not fire while a head's checks are pending or red (#236).

The ruling ask

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).

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):

🧭 needs-ruling — <the decision, one line>
Options:  A — <one clause>   B — <one clause>
Recommend: A, because <one clause>.
Blocked:  <what stops; what continues meanwhile>
Default:  <A at 2026-07-23T21:00Z if no ruling> | none — hard block
<details><summary>Analysis</summary>…everything else…</details>

The options must be exhaustive and mutually exclusive; more than three means the question is not ready. Recommend: is mandatory — omitting it hands the 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, as published artifacts, secrets, prod and org policy are 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):

0–12h: proceed when a still-clear, reversible default expires, and say out loud that you did; a hard block waits.
at 12h: do not fire a stale default — re-read it against what has landed, and where doubt has appeared, make it a hard block.
at 24h: proceed regardless, as a PR: pick an option and state in the PR body which way you went and what doubt remains. Nothing merges by this; the human still gates the merge.
past 24h: hand the choice to triage, which picks the option, records it as a decision, and remains accountable; the operator can overturn it at merge.

A re-flag starts a fresh ladder. The ladder applies whatever Default: says, including a hard block, and an active back-and-forth still climbs it — unlike the 7-day nudge, which resets on real activity. The machine observes both clocks but never sets, clears, or decides needs-ruling.

The label stays until agreement is reached, not until the maintainer replies. The setter records the ruling, removes the label, and returns the item to its flow in the same comment (LABELS.md).

Handoff

When the round passes — every panel verdict approves the current head, and no blocker:* stands (conflicts rebased, CI green, drill recorded if this is a release PR) — the engine performs these mechanical steps on the builder's behalf, in order:

request the human's review;
set state:needs-human;
post the engine-rendered handoff comment: approvals at the current head, 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 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 parked as shape 4 (Picking, above), that handoff comment being its declaration and your build slot free. Address what comes back (state:addressing) and re-hand-off the same way.

19 KiB Raw Blame History Unescape Escape