B1: deliverable_keys answered a multiset, so a `+` title whose segments
normalize to one key made collision_flags find the issue adjacent to
itself and chain it to its own number — the comment asked an issue to
declare `Blocked by` itself, and two such carriers corrupted the chain
between them. The keys are deduped where the set property belongs.
B2: window_in_scope excluded only blocked/epic/post-merge, admitting
`needs-triage` and label-less issues, so the sweep could add
`needs-triage` to an issue and then tell it about a mint-time membership
call in the same pass. Both flags now call one unblocked_claimable
predicate — #293 D2 corrected gives one gloss on `unblocked` and D3b
says D3 uses it.
The window log line says "unblocked", not "ready": D3b corrected exactly
that wording, and the flag fires on `claimed` too.
Both checks ride the existing sweep walk and write nothing but comments
(#293 D1). The board read now answers the whole payload, because both
decisions are over the WHOLE board — every open issue's labels and title,
and every open release issue's body — and a second pagination for the same
rows would be a second board free to disagree with this one mid-sweep.
Fixtures still owed.
Claiming a needs-ruling issue dated it through the assigned timeline
event, silencing the 7-day escalation nudge at exactly the moment a
builder started working through it. The ruling block now reads
last_issue_comment_activity (D1); the reclaim clock keeps the assignment
(D2) because there the assignment IS the claim; post-merge hands its
evidence read to the ruling block instead of reading again (D6, D7); the
claimed branch reads both clocks at its top, before anything it posts.
LABELS.md and lib/ruling.sh now say what each surface's clock reads
(D4, D5). The #257-era order compositions move to the comments read —
the timeline is no longer an input the issue clocks take, and its
unreadability no longer holds unrelated writes hostage; that narrowing
is pinned rather than implied.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
"no activity for N days" was true of the old clock and is now imprecise:
an assignment no longer counts, so an item assigned yesterday would read
a nudge claiming nine days of nothing. It says "no comment" instead —
the fact the sweep actually read.
Refs #254
The nudge rode `last_issue_activity`, which counts `assigned` timeline
events because assignment is the claim the 48-hour reclaim protects.
`post-merge` has no claim: an assignee there is the invalid composition
the flag beside it reports, so counting the assignment let a broken board
buy the item another 7 days of silence — this issue's failure direction
taken backwards.
One computation, two clocks over it: `issue_activity_at` is the body,
`last_issue_activity` keeps the reclaim and ruling clocks byte-identical,
and `last_issue_comment_activity` is the evidence clock. Both clocks are
read before this branch posts anything, the ruling one included — read
after, it would date the issue by the evidence nudge's own comment and
silence the ruling nudge, which is the self-silencing the branch already
guarded against in the other direction.
Refs #254
WIP checkpoint: the nudge itself, tests still owed. Reuses
ruling_nudge_decision so the 7-day rule keeps one spelling, addresses the
triage actor (post-merge is triage's completion queue), and carries no
idempotency marker — the comment is itself activity, so it self-rate-limits.
Refs #254
The echo body wrote `no longer blocked by #9` unbackticked, twice in one
sentence. GitHub linkifies both, so every echo posted a pair of "mentioned in"
events onto this repo's issue #9 — an issue with nothing to do with the one
being echoed. On a 15-minute cron that is one pair per `blocked` issue on the
board, and the first sweep after merge would have written the whole board's
worth at once.
The file's own convention two branches down already avoids this: the
`blocked-unparseable` comment code-spans its `Blocked by #N` for exactly this
reason. The prose is unchanged, the rendered set is unchanged, and the marker
digests the parsed set rather than the comment body, so no marker moves.
ensure_comment's any-occurrence grep answers "have I ever said this", which
is right for a flag like blocked-unparseable and wrong for a value that
changes. A -> B -> A found A's own first echo and stayed silent, leaving the
thread's newest echo asserting B while the sweep gated on A: a stale parse
presented as the current one, and the third edit did change the parsed set,
so the criterion says it speaks.
blocked_parse_echo_needed compares this parse's marker against the LAST
blockers-parsed-* marker on the thread. The read stays inside guarded_read /
skip_issue, so an unreadable history still fails closed (#247 D1) rather than
answering "nothing echoed yet" and re-posting. ensure_comment is untouched
for every other caller.
Refs #252
The marker claimed to be scoped to the parsed set's value and was scoped
to a lossy rendering of it: `tr -c '[:alnum:]' '-'` maps `acme/widgets#9`
and `acme-widgets#9` — both parses this reconciler accepts — onto one
marker, so a declaration edited between them found the old echo and said
nothing. Silence in exactly the case the echo exists to speak about.
The identity is now a digest of the exact rendered set. The readable slug
stays in front of it and decides nothing. Distinguishing `/` would have
closed the reported pair and left the class: `-`, `_` and `.` are all
legal in a qualifier and all collapse the same way, so all four are
pinned, and the sweep probe observes the second echo actually landing.
Refs #252
The idempotency contract is the marker's scope, so both directions are
pinned: an unchanged set must reuse its marker (or a 15-minute cron
repeats itself forever) and a changed one must not (or a misparse hides
under a marker the thread already carries). crew#308's negated clause is
replayed through the sweep, and the empty parse is echoed beside the
untouched `blocked-unparseable` flag.
Refs #252
The clause parse is exact and unforgiving, and its output was invisible:
every incident in this class was found by a human running the parser by
hand, hours or days late. The sweep now states what it read — one marker
comment per distinct parsed set, comment-only, no label writes.
Refs #252
The per-read guards closed the reported class — a failed read never reaches
a decision function — and left one layer standing. A pass could mutate and
only THEN reach a guarded read, fail it, and report the issue as skipped:
`stale` removed, or `needs-triage` minted, under a log line saying the
sweep had touched nothing. That is the same false report #247 exists to
close, told from the other end, and the panel reproduced it on four
separate compositions.
Fixed as the ordering invariant rather than per site. Inside
reconcile_issue_pass's subshell, run() and log() stage their effects, and
commit_staged_effects replays them in order once the pass has completed.
skip_issue emits its own line directly and exits, so the buffer dies with
the subshell. A skip therefore implies zero `gh issue edit`, zero
`gh issue comment`, and no log line about a mutation that never landed —
for compositions nobody has written yet, because reconcile_issue has no way
to mutate directly. Reads stay where they are: they may happen anywhere,
since nothing lands until the end.
Stated per site it would hold until the next composition. Two consequences
worth naming: reconcile_ruling is covered without touching lib/ruling.sh,
because it posts through the sourcing script's run()/log() — the PR surface
keeps its own and is unaffected; and a genuine crash mid-pass now also
lands nothing, where before it left the earlier mutations applied. D4's
handler string, D6's tail and D7's exit 0 are all unchanged, and the
healthy path is byte-identical: every staged write commits under the same
`>/dev/null` its call site already applied.
Refs #247
`gh api` prints a 5xx response body to stdout AND exits non-zero, and
GitHub's 5xx body is a JSON object. Inside the per-issue subshell that
payload passed `has("pull_request") | not`, emptied `.labels[]`, and
`queue_decision` — correct on the input it was handed — wrote
`needs-triage` onto a healthy epic. The run then logged `reconciled.`
and exited 0 (crew#329, #247).
errexit could not have caught it: a command whose status is tested by
`||` runs with errexit suppressed, and the suppression extends through
the whole subshell body, so the `|| log` handler is what disables the
errexit that would have aborted at the failed read. Removing the handler
revives errexit and loses #91's resilience, and an inline `set -e` does
not re-arm it. Explicit per-read checks are the mechanism.
Every read inside that subshell is now checked — the issue read on its
status AND on its payload shape (an HTTP 200 whose body is `null` exits
0 and empties the label set just the same), both reads in
`last_issue_activity`, and the comments read in
`issue_comment_has_marker`. On failure the issue is left exactly as it
is, the reason rides its own `#$n:` line, and the subshell exits with a
distinguished status the sweep counts, so a deliberate skip is not
reported as a crash and a genuine crash is still named byte-identically.
`read_failure_reason` moves to lib/read.sh beside a new `guarded_read`,
sourced by both reconcilers: labels-reconcile's copy was the only one,
and the issue surface needs the identical rule.
Refs #247
The predicate's fixtures cannot see the read that sets HEAD_COMMIT_AT, so a
sweep probe drives it both ways: read, and the blocker is written off a dated
head; denied, and the denial is named on its own line while the state still
converges — this read narrows one blocker, it does not skip the PR the way an
unreadable rollup does. Renaming the assignment reds the probe.
The read also moves after the mergeability/checks skip: a PR the sweep walks
away from must not pay for a call whose only consumer is a blocker that pass
will never decide.
Refs #236
`blocker:unrequested` is the one blocker that names an act the author must
perform, and it never knew whether performing it was permitted. BUILDER.md's
review round requires a green check at the head before requesting, so a
builder waiting out a pending run is complying — and the blocker fired on
compliance (crew#318 ~12:44Z, ceremony#235 12:30Z, both 2026-08-03).
Gate the branch on CHECKS ∈ SUCCESS | NONE (D1): PENDING is CI's move, which
state:addressing already says, and FAILURE belongs to blocker:ci-red rather
than to a second label on the same stall. Then require the supporting facts —
the head's own date and the round's newest submitted review — to have stood
for RECONCILE_UNREQUESTED_GRACE (default 300s, D2), measured off those
timestamps because this sweep is stateless per pass. A timestamp that cannot
be read refuses the blocker.
Refs #236
issue_probe's merged-PR argument becomes a spec list — `PR` or `PR@<iso>`
— so a probe can state merge order; the bare form keeps every existing call
site literal.
The direct-drive cases cover crew#176's shape (the lower number merged
later), agreeing orders, interleaved issues, the mergedAt tie broken by
highest PR number under both input orders, and the empty answer. The
end-to-end probe is crew#321's: a marker already standing for the
later-merged, lower-numbered PR must suppress the transition, which
selecting by number could never do.
Two static pins keep the request count honest — the sweep issues exactly two
GraphQL queries, with mergedAt selected on the merged-PR node it already
fetched.
Refs #242
post_merge_pr_for_issue answered "which merged Refs PR is this issue's
deliverable?" with sort -n | tail -n1. Merge order is not number order:
crew#176's two Refs PRs merged #184 at 19:05:16Z and #182 at 19:05:18Z.
MERGED_REF_PR_RECORDS gains mergedAt as a third column — a field on the
merged-PR node set already fetched, so no additional GraphQL request — and
the selection sorts on it, breaking ties by highest PR number so the answer
never depends on input order.
Refs #242
panel[z]]=b parsed at the round-1 head: the case pattern only proves
some ]= occurs, so the stray ] stayed inside the login and the real
author silently fell back to the base panel — the misroute D4 exists to
refuse. The login charset is now enforced with the bracket-specific
diagnostic; codex's probe and an invalid-character row are the new
must-fail fixtures.
Refs #224
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
round_outranks_draft consults the round before draft short-circuits: a
re-drafted PR carrying CHANGES_REQUESTED, an owed round-reply, or
push-staled approvals reads state:addressing; a live panel request on a
draft surfaces as state:bots-reviewing rather than being absorbed
(the must-not-paper-over combination, decided as: visible). Approvals do
not outrank draft, so a draft never reads needs-human, and a virgin
draft is byte-identical to before. LABELS.md's state:building row makes
draft evidence, not the definition.
Refs #205
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
One resolution point (panel_for_author) feeds set_required_bots; the
author's row when the conf defines one, the base panel= otherwise, minus
the author in either case. Bracket prefixes are matched quoted so the
case patterns cannot glob (D7, panela= tripwire). configured_label_rows
skips the rows so a dispatch bootstrap cannot mint a label named after
one. BUILDER.md/REVIEWER.md carry the one D9 wording; CONSUMERS.md
publishes the row as unreleased with the parse-failure warning.
Refs #224
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The shared reconcile concurrency group displaces queued sweeps as
CANCELLED, and the displaced run's successor attaches to a different PR —
so on the victim the newest self entry stayed CANCELLED, scored FAILURE,
and the sweep set blocker:ci-red off its own corpse every cadence
(crew#227). Drop rollup entries whose workflowName matches SELF_WORKFLOW
(defaulting to the ambient GITHUB_WORKFLOW — the caller's name, so no
workflow edit and no hardcoded consumer name) before the newest-per-context
collapse; an empty name filters nothing. A self-only rollup now honestly
scores NONE, and a genuine foreign failure still blocks beside a cancelled
self entry — the must-fail guard against re-opening #136.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Binding to the first marker occurrence dropped every later sentence of a
repeated declaration and let earlier prose hijack the parse — the false
ready promotion on rig#154. Each occurrence now contributes its own
clause, terminated at its own first ./; (unterminated -> end of input),
and the union feeds the unchanged classification and decision table.
Closes#184
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
During the ceremony#111 freeze, rig#126/#128 carried blocked beside
state:needs-human — the round had finished, but the hold said the merge
must not happen, and rig#126 was merged seven minutes after the
reconciler wrote the green label. decide_state() only joined the two
axes through blockers(), which emits branch facts; the hand-set blocked
label was never consulted.
blocked becomes the second exclusion on state:needs-human, exactly
parallel to needs-ruling: round says needs-human + has_label blocked ->
state:addressing. Deliberately not a blockers() emission — BLOCKERS is
machine-owned and the converge loop would strip the live hold on the
next tick, the same trap #51 names for needs-ruling.
Ruling record: discussion 122, armed default A fired 2026-07-25T09:00Z.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
checks_state discards a CANCELLED entry only when its context group holds
at least one non-cancelled sibling — before the sort, so the duplicate the
repo-global reconcile queue evicted (after it had already attached a check
to the head) cannot outvote the success that did its work (#136a17e497,
#1334002924). An all-cancelled context never reported at all and still
classifies FAILURE; {FAILURE older, CANCELLED newest} keeps its red.
The fixture that pinned the opposite rule imagined a cancelled run
replacing a success; it never saw one that replaced nothing. Rewritten
with its reason, plus the recorded a17e497 shape, the all-cancelled
groups, and the cancelled-over-FAILURE case.
Closes#139
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Bare X.Y.Z at the head where the base says otherwise, no release label,
not a draft: the sweep emits one :⚠️: per pass naming both
versions. A warning only — release is declared intent and the reconciler
never guesses intent (LABELS.md). Version read via the API, both
backends, jq not node; unreadable reads nag nobody. Plus the yq test
contract in CI and fixture tests for the guard matrix.
Part of #130.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
actions/labeler@v5 writes the whole label set (setLabels PUT) even under
sync-labels: false, so a label applied between its read and its write is
silently removed — ceremony#128 lost `release` that way. v6/v7 write the
same way. Replace the step with actions/labels-scope: same labeler.yml
mapping (changed-files/any-glob-to-any-file subset, refused loudly
otherwise), changed paths via the API, and an additive POST as the only
write.
Part of #130.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
