fix: verdict contract, head-bound approvals, serialized reconcile — and a testable state machine
Round-1 blockers, all three reviewers concurring:
- COMMENTED agreement now counts: agreement_signal recognizes the live bots'
durable markers (Verdict: Approve / I agree with everything / leading ✅) —
the gate to needs-human can actually close. Formal verdicts remain the
contract (CONTRIBUTING), this is the documented transitional workaround.
- Every counting verdict is bound to the head SHA; a stale approval parks the
PR in addressing (agent owes re-request) instead of promoting unreviewed
code. CHANGES_REQUESTED blocks at any head, per GitHub's own semantic.
- reconcile serializes under ONE job-level concurrency group; scope stays
per-PR. No more cron-vs-event race on the request-the-human-once guard.
- Sweep resilience: per-PR subshell (one failure logs and continues), label
edits warn instead of wedging; the self-heal claim now matches reality
(dispatch-only bootstrap).
- The state machine is extracted pure (globals in, state out) and sourceable:
test/labels-reconcile.sh proves 14 fixture transitions — comment-only
agreement, stale approval, comment-without-verdict, human precedence and
human-block — wired into CI.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-18 19:31:00 +00:00
|
|
|
#!/usr/bin/env bash
|
|
|
|
|
set -euo pipefail
|
|
|
|
|
|
2026-07-18 19:51:52 +00:00
|
|
|
# Fixture tests for the labels-reconcile state machine: a comment is a
|
|
|
|
|
# non-verdict whatever its body says (the AUTHOR escalates by requesting the
|
|
|
|
|
# human), a stale approval does not promote unreviewed code, and an explicit
|
|
|
|
|
# human request outranks everything.
|
fix: verdict contract, head-bound approvals, serialized reconcile — and a testable state machine
Round-1 blockers, all three reviewers concurring:
- COMMENTED agreement now counts: agreement_signal recognizes the live bots'
durable markers (Verdict: Approve / I agree with everything / leading ✅) —
the gate to needs-human can actually close. Formal verdicts remain the
contract (CONTRIBUTING), this is the documented transitional workaround.
- Every counting verdict is bound to the head SHA; a stale approval parks the
PR in addressing (agent owes re-request) instead of promoting unreviewed
code. CHANGES_REQUESTED blocks at any head, per GitHub's own semantic.
- reconcile serializes under ONE job-level concurrency group; scope stays
per-PR. No more cron-vs-event race on the request-the-human-once guard.
- Sweep resilience: per-PR subshell (one failure logs and continues), label
edits warn instead of wedging; the self-heal claim now matches reality
(dispatch-only bootstrap).
- The state machine is extracted pure (globals in, state out) and sourceable:
test/labels-reconcile.sh proves 14 fixture transitions — comment-only
agreement, stale approval, comment-without-verdict, human precedence and
human-block — wired into CI.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-18 19:31:00 +00:00
|
|
|
# Dependency-free beyond jq; no network, no daemon — pure decide_state.
|
|
|
|
|
|
|
|
|
|
cd "$(dirname "$0")/.."
|
|
|
|
|
# shellcheck source=.github/scripts/labels-reconcile.sh
|
|
|
|
|
. .github/scripts/labels-reconcile.sh
|
|
|
|
|
|
|
|
|
|
# The DRAFT/HEAD_SHA/REQUESTED/REVIEWS_JSON assignments below are the state
|
|
|
|
|
# machine's inputs, consumed inside the sourced decide_state — not unused.
|
|
|
|
|
# shellcheck disable=SC2034
|
|
|
|
|
BOT1="${BOTS[0]}" BOT2="${BOTS[1]}" BOT3="${BOTS[2]}"
|
|
|
|
|
pass=0 fail=0
|
|
|
|
|
|
|
|
|
|
expect() { # $1 = description, $2 = want, $3 = got
|
|
|
|
|
if [ "$2" = "$3" ]; then
|
|
|
|
|
pass=$((pass + 1))
|
|
|
|
|
else
|
|
|
|
|
fail=$((fail + 1))
|
|
|
|
|
printf 'FAIL: %s — want %s, got %s\n' "$1" "$2" "$3"
|
|
|
|
|
fi
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
rev() { # $1=login $2=state $3=commit $4=body $5=submitted_at → one review object
|
|
|
|
|
jq -n --arg u "$1" --arg s "$2" --arg c "$3" --arg b "$4" --arg t "$5" \
|
|
|
|
|
'{user: {login: $u}, state: $s, commit_id: $c, body: $b, submitted_at: $t}'
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
reviews() { jq -s '.' <<<"$*"; } # collect review objects into an array
|
|
|
|
|
|
|
|
|
|
# -- drafts are building, whoever is requested --------------------------------
|
|
|
|
|
DRAFT=true HEAD_SHA=head1 REQUESTED="" REVIEWS_JSON='[]'
|
|
|
|
|
expect "draft PR is building" state:building "$(decide_state)"
|
|
|
|
|
|
|
|
|
|
# -- fresh ready PR with bots requested ---------------------------------------
|
|
|
|
|
DRAFT=false REQUESTED="$BOT1
|
|
|
|
|
$BOT2
|
|
|
|
|
$BOT3" REVIEWS_JSON='[]'
|
|
|
|
|
expect "requested bots mean bots-reviewing" state:bots-reviewing "$(decide_state)"
|
|
|
|
|
|
|
|
|
|
# -- a bot that never reviewed keeps the round open ---------------------------
|
|
|
|
|
REQUESTED="" REVIEWS_JSON="$(reviews \
|
|
|
|
|
"$(rev "$BOT1" APPROVED head1 "" t1)" \
|
|
|
|
|
"$(rev "$BOT2" APPROVED head1 "" t2)")"
|
|
|
|
|
expect "missing bot review means bots-reviewing" state:bots-reviewing "$(decide_state)"
|
|
|
|
|
|
2026-07-18 19:51:52 +00:00
|
|
|
# -- a comment is a non-verdict, agreement body or not: the author escalates --
|
fix: verdict contract, head-bound approvals, serialized reconcile — and a testable state machine
Round-1 blockers, all three reviewers concurring:
- COMMENTED agreement now counts: agreement_signal recognizes the live bots'
durable markers (Verdict: Approve / I agree with everything / leading ✅) —
the gate to needs-human can actually close. Formal verdicts remain the
contract (CONTRIBUTING), this is the documented transitional workaround.
- Every counting verdict is bound to the head SHA; a stale approval parks the
PR in addressing (agent owes re-request) instead of promoting unreviewed
code. CHANGES_REQUESTED blocks at any head, per GitHub's own semantic.
- reconcile serializes under ONE job-level concurrency group; scope stays
per-PR. No more cron-vs-event race on the request-the-human-once guard.
- Sweep resilience: per-PR subshell (one failure logs and continues), label
edits warn instead of wedging; the self-heal claim now matches reality
(dispatch-only bootstrap).
- The state machine is extracted pure (globals in, state out) and sourceable:
test/labels-reconcile.sh proves 14 fixture transitions — comment-only
agreement, stale approval, comment-without-verdict, human precedence and
human-block — wired into CI.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-18 19:31:00 +00:00
|
|
|
REVIEWS_JSON="$(reviews \
|
|
|
|
|
"$(rev "$BOT1" COMMENTED head1 "✅ **Reviewed — I agree with everything.**" t1)" \
|
2026-07-18 19:51:52 +00:00
|
|
|
"$(rev "$BOT2" APPROVED head1 "" t2)" \
|
|
|
|
|
"$(rev "$BOT3" APPROVED head1 "" t3)")"
|
|
|
|
|
expect "comment-only agreement still parks on the author" state:addressing "$(decide_state)"
|
|
|
|
|
# ...and the author's escalation — requesting the human — flips it
|
|
|
|
|
REQUESTED="$HUMAN"
|
|
|
|
|
expect "author escalation flips to needs-human" state:needs-human "$(decide_state)"
|
|
|
|
|
REQUESTED=""
|
|
|
|
|
|
|
|
|
|
# -- three formal approvals need no author judgment ---------------------------
|
|
|
|
|
REVIEWS_JSON="$(reviews \
|
|
|
|
|
"$(rev "$BOT1" APPROVED head1 "" t1)" \
|
|
|
|
|
"$(rev "$BOT2" APPROVED head1 "" t2)" \
|
|
|
|
|
"$(rev "$BOT3" APPROVED head1 "" t3)")"
|
|
|
|
|
expect "three formal approvals reach needs-human" state:needs-human "$(decide_state)"
|
fix: verdict contract, head-bound approvals, serialized reconcile — and a testable state machine
Round-1 blockers, all three reviewers concurring:
- COMMENTED agreement now counts: agreement_signal recognizes the live bots'
durable markers (Verdict: Approve / I agree with everything / leading ✅) —
the gate to needs-human can actually close. Formal verdicts remain the
contract (CONTRIBUTING), this is the documented transitional workaround.
- Every counting verdict is bound to the head SHA; a stale approval parks the
PR in addressing (agent owes re-request) instead of promoting unreviewed
code. CHANGES_REQUESTED blocks at any head, per GitHub's own semantic.
- reconcile serializes under ONE job-level concurrency group; scope stays
per-PR. No more cron-vs-event race on the request-the-human-once guard.
- Sweep resilience: per-PR subshell (one failure logs and continues), label
edits warn instead of wedging; the self-heal claim now matches reality
(dispatch-only bootstrap).
- The state machine is extracted pure (globals in, state out) and sourceable:
test/labels-reconcile.sh proves 14 fixture transitions — comment-only
agreement, stale approval, comment-without-verdict, human precedence and
human-block — wired into CI.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-18 19:31:00 +00:00
|
|
|
|
|
|
|
|
# -- a comment WITHOUT a verdict parks the PR on the agent --------------------
|
|
|
|
|
REVIEWS_JSON="$(reviews \
|
|
|
|
|
"$(rev "$BOT1" COMMENTED head1 "🔧 Reviewed — I agree with most; feedback below." t1)" \
|
|
|
|
|
"$(rev "$BOT2" APPROVED head1 "" t2)" \
|
|
|
|
|
"$(rev "$BOT3" APPROVED head1 "" t3)")"
|
|
|
|
|
expect "comment without verdict is addressing" state:addressing "$(decide_state)"
|
|
|
|
|
|
|
|
|
|
# -- changes requested blocks, at any head ------------------------------------
|
|
|
|
|
REVIEWS_JSON="$(reviews \
|
|
|
|
|
"$(rev "$BOT1" CHANGES_REQUESTED old1 "blockers below" t1)" \
|
|
|
|
|
"$(rev "$BOT2" APPROVED head1 "" t2)" \
|
|
|
|
|
"$(rev "$BOT3" APPROVED head1 "" t3)")"
|
|
|
|
|
expect "changes-requested blocks even from an old head" state:addressing "$(decide_state)"
|
|
|
|
|
|
|
|
|
|
# -- a stale approval must not promote unreviewed code ------------------------
|
|
|
|
|
REVIEWS_JSON="$(reviews \
|
|
|
|
|
"$(rev "$BOT1" APPROVED old1 "" t1)" \
|
|
|
|
|
"$(rev "$BOT2" APPROVED head1 "" t2)" \
|
|
|
|
|
"$(rev "$BOT3" APPROVED head1 "" t3)")"
|
|
|
|
|
expect "stale approval is addressing (agent owes re-request)" state:addressing "$(decide_state)"
|
|
|
|
|
|
|
|
|
|
# -- a re-requested bot reopens the round even with an old approval on file ---
|
|
|
|
|
REQUESTED="$BOT1"
|
|
|
|
|
expect "re-requested bot means bots-reviewing" state:bots-reviewing "$(decide_state)"
|
|
|
|
|
REQUESTED=""
|
|
|
|
|
|
|
|
|
|
# -- only the LATEST review per bot counts ------------------------------------
|
|
|
|
|
REVIEWS_JSON="$(reviews \
|
|
|
|
|
"$(rev "$BOT1" CHANGES_REQUESTED head1 "blockers" t1)" \
|
|
|
|
|
"$(rev "$BOT1" APPROVED head1 "" t2)" \
|
|
|
|
|
"$(rev "$BOT2" APPROVED head1 "" t3)" \
|
2026-07-18 19:51:52 +00:00
|
|
|
"$(rev "$BOT3" APPROVED head1 "" t4)")"
|
fix: verdict contract, head-bound approvals, serialized reconcile — and a testable state machine
Round-1 blockers, all three reviewers concurring:
- COMMENTED agreement now counts: agreement_signal recognizes the live bots'
durable markers (Verdict: Approve / I agree with everything / leading ✅) —
the gate to needs-human can actually close. Formal verdicts remain the
contract (CONTRIBUTING), this is the documented transitional workaround.
- Every counting verdict is bound to the head SHA; a stale approval parks the
PR in addressing (agent owes re-request) instead of promoting unreviewed
code. CHANGES_REQUESTED blocks at any head, per GitHub's own semantic.
- reconcile serializes under ONE job-level concurrency group; scope stays
per-PR. No more cron-vs-event race on the request-the-human-once guard.
- Sweep resilience: per-PR subshell (one failure logs and continues), label
edits warn instead of wedging; the self-heal claim now matches reality
(dispatch-only bootstrap).
- The state machine is extracted pure (globals in, state out) and sourceable:
test/labels-reconcile.sh proves 14 fixture transitions — comment-only
agreement, stale approval, comment-without-verdict, human precedence and
human-block — wired into CI.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-18 19:31:00 +00:00
|
|
|
expect "later approval supersedes earlier block" state:needs-human "$(decide_state)"
|
|
|
|
|
|
|
|
|
|
# -- an explicit human request outranks the bot rounds ------------------------
|
|
|
|
|
REQUESTED="$HUMAN" REVIEWS_JSON="$(reviews \
|
|
|
|
|
"$(rev "$BOT1" COMMENTED head1 "feedback, no verdict" t1)")"
|
|
|
|
|
expect "human requested outranks bots" state:needs-human "$(decide_state)"
|
|
|
|
|
REQUESTED=""
|
|
|
|
|
|
|
|
|
|
# -- human CHANGES_REQUESTED puts the ball back on the agent ------------------
|
|
|
|
|
REVIEWS_JSON="$(reviews \
|
|
|
|
|
"$(rev "$BOT1" APPROVED head1 "" t1)" \
|
|
|
|
|
"$(rev "$BOT2" APPROVED head1 "" t2)" \
|
|
|
|
|
"$(rev "$BOT3" APPROVED head1 "" t3)" \
|
|
|
|
|
"$(rev "$HUMAN" CHANGES_REQUESTED head1 "not yet" t4)")"
|
|
|
|
|
expect "human block with bots approving is addressing" state:addressing "$(decide_state)"
|
|
|
|
|
# ...and re-requesting the human hands it back to them
|
|
|
|
|
REQUESTED="$HUMAN"
|
|
|
|
|
expect "re-requested human is needs-human again" state:needs-human "$(decide_state)"
|
|
|
|
|
REQUESTED=""
|
|
|
|
|
|
2026-07-18 20:16:55 +00:00
|
|
|
# -- an old human comment must not wedge the handoff (codex, #85 round 3) -----
|
|
|
|
|
REVIEWS_JSON="$(reviews \
|
|
|
|
|
"$(rev "$HUMAN" COMMENTED old1 "early thoughts" t0)" \
|
|
|
|
|
"$(rev "$BOT1" APPROVED head1 "" t1)" \
|
|
|
|
|
"$(rev "$BOT2" APPROVED head1 "" t2)" \
|
|
|
|
|
"$(rev "$BOT3" APPROVED head1 "" t3)")"
|
|
|
|
|
expect "old human comment + three approvals is needs-human" state:needs-human "$(decide_state)"
|
|
|
|
|
expect "old human comment still needs a fresh request" needed "$(human_request_needed && echo needed || echo not-needed)"
|
|
|
|
|
# ...a stale human APPROVAL likewise needs a re-request for the new head
|
|
|
|
|
REVIEWS_JSON="$(reviews \
|
|
|
|
|
"$(rev "$HUMAN" APPROVED old1 "" t0)" \
|
|
|
|
|
"$(rev "$BOT1" APPROVED head1 "" t1)" \
|
|
|
|
|
"$(rev "$BOT2" APPROVED head1 "" t2)" \
|
|
|
|
|
"$(rev "$BOT3" APPROVED head1 "" t3)")"
|
|
|
|
|
expect "stale human approval needs a fresh request" needed "$(human_request_needed && echo needed || echo not-needed)"
|
|
|
|
|
# ...a HEAD-CURRENT human approval needs nothing more
|
|
|
|
|
REVIEWS_JSON="$(reviews \
|
|
|
|
|
"$(rev "$HUMAN" APPROVED head1 "" t0)" \
|
|
|
|
|
"$(rev "$BOT1" APPROVED head1 "" t1)" \
|
|
|
|
|
"$(rev "$BOT2" APPROVED head1 "" t2)" \
|
|
|
|
|
"$(rev "$BOT3" APPROVED head1 "" t3)")"
|
|
|
|
|
expect "head-current human approval needs no request" not-needed "$(human_request_needed && echo needed || echo not-needed)"
|
|
|
|
|
# ...and a live request suppresses re-requesting
|
|
|
|
|
REQUESTED="$HUMAN"
|
|
|
|
|
expect "live human request suppresses re-request" not-needed "$(human_request_needed && echo needed || echo not-needed)"
|
|
|
|
|
REQUESTED=""
|
|
|
|
|
|
fix(labels): state:needs-human means a human could merge it right now
decide_state() derived state from three inputs -- draft flag, requested
reviewers, submitted reviews -- and read NOTHING about mergeability or checks.
Combined with the `if requested "$HUMAN"` short-circuit at the top of its
precedence, the label was sticky: once the maintainer was requested, the PR
read state:needs-human through conflicts, through red CI, through a
force-push that staled every approval. Nothing demoted it.
Observed twice in one afternoon, in two different shapes. Three PRs sat at
state:needs-human while CONFLICTING for hours -- the board inviting a merge
GitHub had already disabled. And #119, after a rebase, read MERGEABLE, four
green checks, state:needs-human, with ZERO reviews bound to its head: every
visible signal saying "merge me" over a tree no reviewer had seen. That second
shape is the dangerous one, because unlike a conflict nothing on the page
contradicts it.
The rule the label now keeps: state:needs-human means a human could merge this
RIGHT NOW, so anything making that false outranks the request that put it
there.
CONFLICTING or failing checks -> state:needs-rebase (new; the agent's to fix)
approvals staled by a push -> state:addressing (nobody reviewed this tree)
An UNFINISHED round still yields to an explicit human request -- a maintainer
pulling a PR to themselves early is deliberate, and MISSING (nobody has
reviewed yet) is a different fact from STALE (everyone reviewed something
else). That distinction is why the two are handled in different arms rather
than collapsed.
UNKNOWN mergeability is deliberately NOT treated as unmergeable: GitHub
reports it for about a minute after every merge while it recomputes, and
flapping every open PR through needs-rebase on each merge would be worse than
the bug being fixed. A failed read of either fact degrades to the same "do not
know" value, for the same reason -- an API hiccup must not relabel the board.
Also adds merge-next, because a correct needs-human still does not say WHICH
PR to merge first, and order matters when they conflict through CHANGELOG.md.
Queue order is intent, so the reconciler never sets it; it only CLEARS it the
moment the PR stops being mergeable-by-a-human -- precisely the staleness that
made needs-human untrustworthy.
Both live shapes are pinned in test/labels-reconcile.sh (19 -> 29 fixtures),
including that UNKNOWN does not trigger needs-rebase and that a draft outranks
a conflict. Proven non-vacuous: dropping the mergeability arm fails 4
assertions, dropping the STALE precedence fails 2, restoring returns 29/0.
Closes #136
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
2026-07-20 15:18:49 +00:00
|
|
|
# ---------------------------------------------------------------------------
|
|
|
|
|
# #136: state:needs-human must mean "a human could merge this RIGHT NOW".
|
|
|
|
|
# Both cases below were observed live in this repo on 2026-07-20, and both
|
|
|
|
|
# showed state:needs-human while being unmergeable in different ways.
|
|
|
|
|
# ---------------------------------------------------------------------------
|
|
|
|
|
ALL_APPROVE="$(reviews \
|
|
|
|
|
"$(rev "$BOT1" APPROVED head1 "" t1)" \
|
|
|
|
|
"$(rev "$BOT2" APPROVED head1 "" t2)" \
|
|
|
|
|
"$(rev "$BOT3" APPROVED head1 "" t3)")"
|
|
|
|
|
|
|
|
|
|
# -- flavour 1: not mergeable. The merge button is disabled, yet the board
|
|
|
|
|
# said "your turn" on #119/#120/#127 for hours.
|
|
|
|
|
DRAFT=false HEAD_SHA=head1 REQUESTED="" REVIEWS_JSON="$ALL_APPROVE" MERGEABLE=CONFLICTING CHECKS=SUCCESS
|
|
|
|
|
expect "a CONFLICTING PR is needs-rebase, not needs-human" state:needs-rebase "$(decide_state)"
|
|
|
|
|
REQUESTED="$HUMAN"
|
|
|
|
|
expect "...even with the human explicitly requested" state:needs-rebase "$(decide_state)"
|
|
|
|
|
|
|
|
|
|
# -- red CI is the same claim: not something a human should merge.
|
|
|
|
|
REQUESTED="" MERGEABLE=MERGEABLE CHECKS=FAILURE
|
|
|
|
|
expect "a red PR is needs-rebase" state:needs-rebase "$(decide_state)"
|
|
|
|
|
REQUESTED="$HUMAN"
|
|
|
|
|
expect "...and a human request does not override red CI" state:needs-rebase "$(decide_state)"
|
|
|
|
|
|
|
|
|
|
# -- UNKNOWN is NOT unmergeable. GitHub reports it for ~a minute after every
|
|
|
|
|
# merge while it recomputes; treating it as broken would flap every open PR
|
|
|
|
|
# into needs-rebase on each merge — worse than the bug being fixed.
|
|
|
|
|
REQUESTED="" MERGEABLE=UNKNOWN CHECKS=PENDING
|
|
|
|
|
expect "UNKNOWN mergeability does not trigger needs-rebase" state:needs-human "$(decide_state)"
|
|
|
|
|
|
|
|
|
|
# -- flavour 2 (the dangerous one): mergeable, green, human requested, and
|
|
|
|
|
# NOBODY has reviewed this head. Observed on #119 after a rebase: every
|
|
|
|
|
# signal read "merge me" and nothing on the page contradicted it.
|
|
|
|
|
MERGEABLE=MERGEABLE CHECKS=SUCCESS REQUESTED="$HUMAN"
|
|
|
|
|
REVIEWS_JSON="$(reviews \
|
|
|
|
|
"$(rev "$BOT1" APPROVED oldhead "" t1)" \
|
|
|
|
|
"$(rev "$BOT2" APPROVED oldhead "" t2)" \
|
|
|
|
|
"$(rev "$BOT3" APPROVED oldhead "" t3)")"
|
|
|
|
|
expect "stale approvals outrank the human request (nobody reviewed this tree)" state:addressing "$(decide_state)"
|
|
|
|
|
|
fix(labels): an unrecognised check outcome blocks, and a staled round outranks an unfinished one
Round 2. Two blockers, both real, both closing the same hole this PR
exists to close — a `state:needs-human` that invites a human to merge a
tree that will not merge.
The check-rollup classifier enumerated the outcomes that block and let
the rest fall into `else "SUCCESS"`, so ERROR, CANCELLED and STALE all
read as green. Inverted: it now lists the outcomes that DON'T block —
SUCCESS, NEUTRAL, SKIPPED, plus the pending set — and treats everything
else as blocking. The direction is the point. The rollup mixes two
closed enums (CheckRun.conclusion, StatusContext.state) and an outcome
the list forgets is one we cannot certify as mergeable; the costs are
not symmetric, since a false FAILURE parks the PR on the agent who
looks, while a false SUCCESS is #136 exactly.
Once CANCELLED blocks, superseded runs must be dropped first: a re-run
does not evict the run it replaced, and this PR's own tip carries a
CANCELLED `scope` beside the SUCCESS `scope` that superseded it. Each
context now collapses to its newest entry before anything is judged,
keyed on workflow + job name because a bare job name is only unique
within its workflow. That preserves the re-run case the panel split
over while still blocking a cancelled run that is the newest word.
The classifier also moved out of main() into checks_state(). That is
why no fixture caught this: it was inline in the fetch loop, so the
fixtures could only inject CHECKS= as an already-decided string.
Second, decide_state() returned from inside the bot loop on the first
MISSING, so a STALE belonging to a later bot in BOTS was never read —
a round that was both unfinished and staled came out needs-human over
a head nobody had reviewed. The whole round is now collected before
precedence is applied to it as a unit, STALE before MISSING.
test/labels-reconcile.sh: 29 -> 44 fixtures, pinning the check-outcome
enum, the supersede rule (both orders, plus same name in another
workflow), and the mixed round at both ends of BOTS. All verified
non-vacuous against the round-1 code.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
2026-07-20 16:04:16 +00:00
|
|
|
# -- ...and a round that is BOTH unfinished and staled is still the agent's.
|
|
|
|
|
# Deciding inside the bot loop made this depend on BOTS order: the MISSING
|
|
|
|
|
# returned before any later bot's STALE was read, so the mixed round came
|
|
|
|
|
# out needs-human with nothing bound to the head. Pinned at both ends of
|
|
|
|
|
# the array, because the whole failure was one of ordering.
|
|
|
|
|
MERGEABLE=MERGEABLE CHECKS=SUCCESS REQUESTED="$HUMAN"
|
|
|
|
|
REVIEWS_JSON="$(reviews \
|
|
|
|
|
"$(rev "$BOT1" APPROVED oldhead "" t1)" \
|
|
|
|
|
"$(rev "$BOT2" APPROVED oldhead "" t2)")"
|
|
|
|
|
expect "stale approvals + a bot yet to review is addressing, not needs-human" \
|
|
|
|
|
state:addressing "$(decide_state)"
|
|
|
|
|
REVIEWS_JSON="$(reviews "$(rev "$BOT3" APPROVED oldhead "" t3)")"
|
|
|
|
|
expect "...and the same when the stale verdict is the LAST bot in BOTS" \
|
|
|
|
|
state:addressing "$(decide_state)"
|
|
|
|
|
|
fix(labels): state:needs-human means a human could merge it right now
decide_state() derived state from three inputs -- draft flag, requested
reviewers, submitted reviews -- and read NOTHING about mergeability or checks.
Combined with the `if requested "$HUMAN"` short-circuit at the top of its
precedence, the label was sticky: once the maintainer was requested, the PR
read state:needs-human through conflicts, through red CI, through a
force-push that staled every approval. Nothing demoted it.
Observed twice in one afternoon, in two different shapes. Three PRs sat at
state:needs-human while CONFLICTING for hours -- the board inviting a merge
GitHub had already disabled. And #119, after a rebase, read MERGEABLE, four
green checks, state:needs-human, with ZERO reviews bound to its head: every
visible signal saying "merge me" over a tree no reviewer had seen. That second
shape is the dangerous one, because unlike a conflict nothing on the page
contradicts it.
The rule the label now keeps: state:needs-human means a human could merge this
RIGHT NOW, so anything making that false outranks the request that put it
there.
CONFLICTING or failing checks -> state:needs-rebase (new; the agent's to fix)
approvals staled by a push -> state:addressing (nobody reviewed this tree)
An UNFINISHED round still yields to an explicit human request -- a maintainer
pulling a PR to themselves early is deliberate, and MISSING (nobody has
reviewed yet) is a different fact from STALE (everyone reviewed something
else). That distinction is why the two are handled in different arms rather
than collapsed.
UNKNOWN mergeability is deliberately NOT treated as unmergeable: GitHub
reports it for about a minute after every merge while it recomputes, and
flapping every open PR through needs-rebase on each merge would be worse than
the bug being fixed. A failed read of either fact degrades to the same "do not
know" value, for the same reason -- an API hiccup must not relabel the board.
Also adds merge-next, because a correct needs-human still does not say WHICH
PR to merge first, and order matters when they conflict through CHANGELOG.md.
Queue order is intent, so the reconciler never sets it; it only CLEARS it the
moment the PR stops being mergeable-by-a-human -- precisely the staleness that
made needs-human untrustworthy.
Both live shapes are pinned in test/labels-reconcile.sh (19 -> 29 fixtures),
including that UNKNOWN does not trigger needs-rebase and that a draft outranks
a conflict. Proven non-vacuous: dropping the mergeability arm fails 4
assertions, dropping the STALE precedence fails 2, restoring returns 29/0.
Closes #136
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
2026-07-20 15:18:49 +00:00
|
|
|
# -- but an UNFINISHED round still yields to an explicit human request: a
|
|
|
|
|
# maintainer pulling a PR to themselves early is deliberate, and was the
|
|
|
|
|
# original precedence. MISSING differs from STALE — nobody has reviewed
|
|
|
|
|
# YET, versus everyone reviewed something else.
|
|
|
|
|
REVIEWS_JSON="$(reviews "$(rev "$BOT1" APPROVED head1 "" t1)")"
|
|
|
|
|
expect "an unfinished round still yields to an explicit human request" state:needs-human "$(decide_state)"
|
|
|
|
|
REQUESTED=""
|
|
|
|
|
expect "...and without that request it is still bots-reviewing" state:bots-reviewing "$(decide_state)"
|
|
|
|
|
|
fix(labels): an unrecognised check outcome blocks, and a staled round outranks an unfinished one
Round 2. Two blockers, both real, both closing the same hole this PR
exists to close — a `state:needs-human` that invites a human to merge a
tree that will not merge.
The check-rollup classifier enumerated the outcomes that block and let
the rest fall into `else "SUCCESS"`, so ERROR, CANCELLED and STALE all
read as green. Inverted: it now lists the outcomes that DON'T block —
SUCCESS, NEUTRAL, SKIPPED, plus the pending set — and treats everything
else as blocking. The direction is the point. The rollup mixes two
closed enums (CheckRun.conclusion, StatusContext.state) and an outcome
the list forgets is one we cannot certify as mergeable; the costs are
not symmetric, since a false FAILURE parks the PR on the agent who
looks, while a false SUCCESS is #136 exactly.
Once CANCELLED blocks, superseded runs must be dropped first: a re-run
does not evict the run it replaced, and this PR's own tip carries a
CANCELLED `scope` beside the SUCCESS `scope` that superseded it. Each
context now collapses to its newest entry before anything is judged,
keyed on workflow + job name because a bare job name is only unique
within its workflow. That preserves the re-run case the panel split
over while still blocking a cancelled run that is the newest word.
The classifier also moved out of main() into checks_state(). That is
why no fixture caught this: it was inline in the fetch loop, so the
fixtures could only inject CHECKS= as an already-decided string.
Second, decide_state() returned from inside the bot loop on the first
MISSING, so a STALE belonging to a later bot in BOTS was never read —
a round that was both unfinished and staled came out needs-human over
a head nobody had reviewed. The whole round is now collected before
precedence is applied to it as a unit, STALE before MISSING.
test/labels-reconcile.sh: 29 -> 44 fixtures, pinning the check-outcome
enum, the supersede rule (both orders, plus same name in another
workflow), and the mixed round at both ends of BOTS. All verified
non-vacuous against the round-1 code.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
2026-07-20 16:04:16 +00:00
|
|
|
# ---------------------------------------------------------------------------
|
|
|
|
|
# checks_state: the rollup classifier. It lived inline in main() for the first
|
|
|
|
|
# round of this PR, which is why nothing here caught it calling ERROR,
|
|
|
|
|
# CANCELLED and STALE green. Extracted so the enum can be pinned down.
|
|
|
|
|
# ---------------------------------------------------------------------------
|
|
|
|
|
rollup() { jq -n --argjson c "$1" '{statusCheckRollup: $c}'; }
|
|
|
|
|
run_() { jq -n --arg n "$1" --arg o "$2" --arg t "${3:-2026-07-20T15:00:00Z}" \
|
|
|
|
|
'{__typename:"CheckRun", workflowName:"ci", name:$n, conclusion:$o, completedAt:$t}'; }
|
|
|
|
|
ctx_() { jq -n --arg n "$1" --arg s "$2" --arg t "${3:-2026-07-20T15:00:00Z}" \
|
|
|
|
|
'{__typename:"StatusContext", context:$n, state:$s, createdAt:$t}'; }
|
|
|
|
|
|
|
|
|
|
expect "no checks at all is NONE" NONE "$(rollup '[]' | checks_state)"
|
|
|
|
|
expect "all green is SUCCESS" SUCCESS \
|
|
|
|
|
"$(rollup "[$(run_ a SUCCESS),$(run_ b SUCCESS)]" | checks_state)"
|
|
|
|
|
expect "a queued run is PENDING" PENDING \
|
|
|
|
|
"$(rollup "[$(run_ a SUCCESS),$(run_ b QUEUED)]" | checks_state)"
|
|
|
|
|
expect "a plain failure is FAILURE" FAILURE \
|
|
|
|
|
"$(rollup "[$(run_ a SUCCESS),$(run_ b FAILURE)]" | checks_state)"
|
|
|
|
|
|
|
|
|
|
# -- the round-1 gap: outcomes that are neither success nor pending, and that
|
|
|
|
|
# leave a required check unsatisfied. All three reached the old `else`.
|
|
|
|
|
expect "a commit status ERROR blocks" FAILURE \
|
|
|
|
|
"$(rollup "[$(run_ a SUCCESS),$(ctx_ lint ERROR)]" | checks_state)"
|
|
|
|
|
expect "a CANCELLED run blocks" FAILURE \
|
|
|
|
|
"$(rollup "[$(run_ a SUCCESS),$(run_ b CANCELLED)]" | checks_state)"
|
|
|
|
|
expect "a STALE run blocks" FAILURE \
|
|
|
|
|
"$(rollup "[$(run_ a SUCCESS),$(run_ b STALE)]" | checks_state)"
|
|
|
|
|
expect "an outcome the enum does not know blocks, it does not pass" FAILURE \
|
|
|
|
|
"$(rollup "[$(run_ a SUCCESS),$(run_ b SOME_FUTURE_STATE)]" | checks_state)"
|
|
|
|
|
|
|
|
|
|
# -- NEUTRAL and SKIPPED satisfy branch protection; path-filtered jobs skip
|
|
|
|
|
# constantly, and calling that red would park every PR on the agent.
|
|
|
|
|
expect "NEUTRAL and SKIPPED are not failures" SUCCESS \
|
|
|
|
|
"$(rollup "[$(run_ a SUCCESS),$(run_ b NEUTRAL),$(run_ c SKIPPED)]" | checks_state)"
|
|
|
|
|
|
|
|
|
|
# -- latest-wins. The rollup keeps superseded runs, so this PR's own tip
|
|
|
|
|
# carried a CANCELLED `scope` beside the SUCCESS `scope` that replaced it.
|
|
|
|
|
# Without collapsing, making CANCELLED block would strand it forever.
|
|
|
|
|
expect "a re-run supersedes the cancelled original" SUCCESS \
|
|
|
|
|
"$(rollup "[$(run_ scope CANCELLED 2026-07-20T15:19:39Z),\
|
|
|
|
|
$(run_ scope SUCCESS 2026-07-20T15:19:45Z)]" | checks_state)"
|
|
|
|
|
expect "...and the reverse order is not a re-run passing, it is one failing" FAILURE \
|
|
|
|
|
"$(rollup "[$(run_ scope SUCCESS 2026-07-20T15:19:39Z),\
|
|
|
|
|
$(run_ scope CANCELLED 2026-07-20T15:19:45Z)]" | checks_state)"
|
|
|
|
|
# same job name in a different workflow is a different context, not a re-run
|
|
|
|
|
expect "same name in another workflow does not supersede" FAILURE \
|
|
|
|
|
"$(rollup "[$(jq -n '{__typename:"CheckRun",workflowName:"labels",name:"scope",conclusion:"FAILURE",completedAt:"2026-07-20T15:00:00Z"}'),\
|
|
|
|
|
$(run_ scope SUCCESS 2026-07-20T15:19:45Z)]" | checks_state)"
|
|
|
|
|
|
fix(labels): date a check run by the newest stamp it carries, not by its completion
Round 3. @claude-bot-andresmgsl and @codex-bot-andresmgsl independently
caught that the supersede rule added in round 2 inverted itself on the
one shape its fixtures could not express: a re-run still in flight.
A running check does not omit `completedAt`. `gh` marshals the Go zero
time as the STRING "0001-01-01T00:00:00Z", and jq's `//` only falls
through null/false, so the sentinel won the sort key and sorted before
every real timestamp. The live re-run became the OLDEST entry in its
context, `last` discarded it, and the run it superseded was judged
instead — exactly backwards, and wrong in both directions:
green + re-run in flight -> SUCCESS (should be PENDING)
CANCELLED + re-run in flight -> FAILURE (should be PENDING)
The first is #136 restored by the very rule meant to close it: all bots
approve, mergeable, state:needs-human — over a tree whose merge button
branch protection has disabled. It was also a regression from round 1,
which caught it via `any(. == "")`. The second is the re-run flap the
supersede rule exists to prevent, narrowed rather than removed.
Fixed by taking the newest timestamp a run actually carries and
discarding BOTH spellings of absent — null and the zero sentinel —
rather than by reordering the fallbacks. An entry with no usable
timestamp now sorts LAST rather than first, so an undateable in-flight
run is never dropped in favour of a stale success. Every ambiguity
resolves toward "not settled".
The fixtures could not have caught this: the `run_()` helper sets only
`completedAt`, so every supersede fixture was a race between two
FINISHED runs. The helper now expresses an in-flight entry, and the
four new fixtures assert PENDING over both a green and a cancelled
predecessor. 44 -> 48; reverting just the dating expression fails 3 of
the 4.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
2026-07-20 16:18:22 +00:00
|
|
|
# -- a run still IN FLIGHT. `run_()` cannot express this: it always carries a
|
|
|
|
|
# real completedAt, which is exactly why the supersede rule shipped dating
|
|
|
|
|
# runs by completion and nothing caught it. Both spellings of "no
|
|
|
|
|
# completion" are pinned, because `gh` emits the zero sentinel (a string,
|
|
|
|
|
# which `//` does not fall through) while the API emits null.
|
|
|
|
|
inflight_() { jq -n --arg n "$1" --arg t "$2" --arg c "${3:-0001-01-01T00:00:00Z}" \
|
|
|
|
|
'{__typename:"CheckRun", workflowName:"ci", name:$n, status:"IN_PROGRESS",
|
|
|
|
|
conclusion:"", startedAt:$t, completedAt:(if $c == "null" then null else $c end)}'; }
|
|
|
|
|
|
|
|
|
|
expect "a re-run in flight beats the success it superseded (zero sentinel)" PENDING \
|
|
|
|
|
"$(rollup "[$(run_ build SUCCESS 2026-07-20T15:00:00Z),\
|
|
|
|
|
$(inflight_ build 2026-07-20T15:10:00Z)]" | checks_state)"
|
|
|
|
|
expect "...and the same when the absent completion is null" PENDING \
|
|
|
|
|
"$(rollup "[$(run_ build SUCCESS 2026-07-20T15:00:00Z),\
|
|
|
|
|
$(inflight_ build 2026-07-20T15:10:00Z null)]" | checks_state)"
|
|
|
|
|
expect "a replacement in flight for a CANCELLED run is pending, not failed" PENDING \
|
|
|
|
|
"$(rollup "[$(run_ build CANCELLED 2026-07-20T15:00:00Z),\
|
|
|
|
|
$(inflight_ build 2026-07-20T15:10:00Z)]" | checks_state)"
|
|
|
|
|
# an entry carrying no usable timestamp is treated as newest, not oldest —
|
|
|
|
|
# ambiguity resolves toward "not settled" rather than toward a stale success
|
|
|
|
|
expect "an undateable in-flight run is not discarded for a stale success" PENDING \
|
|
|
|
|
"$(rollup "[$(run_ build SUCCESS 2026-07-20T15:00:00Z),\
|
|
|
|
|
$(jq -n '{__typename:"CheckRun",workflowName:"ci",name:"build",conclusion:"",startedAt:null,completedAt:null}')]" \
|
|
|
|
|
| checks_state)"
|
|
|
|
|
|
fix(labels): an unrecognised check outcome blocks, and a staled round outranks an unfinished one
Round 2. Two blockers, both real, both closing the same hole this PR
exists to close — a `state:needs-human` that invites a human to merge a
tree that will not merge.
The check-rollup classifier enumerated the outcomes that block and let
the rest fall into `else "SUCCESS"`, so ERROR, CANCELLED and STALE all
read as green. Inverted: it now lists the outcomes that DON'T block —
SUCCESS, NEUTRAL, SKIPPED, plus the pending set — and treats everything
else as blocking. The direction is the point. The rollup mixes two
closed enums (CheckRun.conclusion, StatusContext.state) and an outcome
the list forgets is one we cannot certify as mergeable; the costs are
not symmetric, since a false FAILURE parks the PR on the agent who
looks, while a false SUCCESS is #136 exactly.
Once CANCELLED blocks, superseded runs must be dropped first: a re-run
does not evict the run it replaced, and this PR's own tip carries a
CANCELLED `scope` beside the SUCCESS `scope` that superseded it. Each
context now collapses to its newest entry before anything is judged,
keyed on workflow + job name because a bare job name is only unique
within its workflow. That preserves the re-run case the panel split
over while still blocking a cancelled run that is the newest word.
The classifier also moved out of main() into checks_state(). That is
why no fixture caught this: it was inline in the fetch loop, so the
fixtures could only inject CHECKS= as an already-decided string.
Second, decide_state() returned from inside the bot loop on the first
MISSING, so a STALE belonging to a later bot in BOTS was never read —
a round that was both unfinished and staled came out needs-human over
a head nobody had reviewed. The whole round is now collected before
precedence is applied to it as a unit, STALE before MISSING.
test/labels-reconcile.sh: 29 -> 44 fixtures, pinning the check-outcome
enum, the supersede rule (both orders, plus same name in another
workflow), and the mixed round at both ends of BOTS. All verified
non-vacuous against the round-1 code.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
2026-07-20 16:04:16 +00:00
|
|
|
# -- the classifier feeds the state machine: a cancelled required check must
|
|
|
|
|
# take the PR off the human's plate, which is the whole point of #136.
|
|
|
|
|
DRAFT=false HEAD_SHA=head1 REQUESTED="$HUMAN" REVIEWS_JSON="$ALL_APPROVE" MERGEABLE=MERGEABLE
|
|
|
|
|
CHECKS="$(rollup "[$(run_ a SUCCESS),$(run_ b CANCELLED)]" | checks_state)"
|
|
|
|
|
expect "a cancelled check reaches decide_state as needs-rebase" state:needs-rebase "$(decide_state)"
|
|
|
|
|
|
fix(labels): state:needs-human means a human could merge it right now
decide_state() derived state from three inputs -- draft flag, requested
reviewers, submitted reviews -- and read NOTHING about mergeability or checks.
Combined with the `if requested "$HUMAN"` short-circuit at the top of its
precedence, the label was sticky: once the maintainer was requested, the PR
read state:needs-human through conflicts, through red CI, through a
force-push that staled every approval. Nothing demoted it.
Observed twice in one afternoon, in two different shapes. Three PRs sat at
state:needs-human while CONFLICTING for hours -- the board inviting a merge
GitHub had already disabled. And #119, after a rebase, read MERGEABLE, four
green checks, state:needs-human, with ZERO reviews bound to its head: every
visible signal saying "merge me" over a tree no reviewer had seen. That second
shape is the dangerous one, because unlike a conflict nothing on the page
contradicts it.
The rule the label now keeps: state:needs-human means a human could merge this
RIGHT NOW, so anything making that false outranks the request that put it
there.
CONFLICTING or failing checks -> state:needs-rebase (new; the agent's to fix)
approvals staled by a push -> state:addressing (nobody reviewed this tree)
An UNFINISHED round still yields to an explicit human request -- a maintainer
pulling a PR to themselves early is deliberate, and MISSING (nobody has
reviewed yet) is a different fact from STALE (everyone reviewed something
else). That distinction is why the two are handled in different arms rather
than collapsed.
UNKNOWN mergeability is deliberately NOT treated as unmergeable: GitHub
reports it for about a minute after every merge while it recomputes, and
flapping every open PR through needs-rebase on each merge would be worse than
the bug being fixed. A failed read of either fact degrades to the same "do not
know" value, for the same reason -- an API hiccup must not relabel the board.
Also adds merge-next, because a correct needs-human still does not say WHICH
PR to merge first, and order matters when they conflict through CHANGELOG.md.
Queue order is intent, so the reconciler never sets it; it only CLEARS it the
moment the PR stops being mergeable-by-a-human -- precisely the staleness that
made needs-human untrustworthy.
Both live shapes are pinned in test/labels-reconcile.sh (19 -> 29 fixtures),
including that UNKNOWN does not trigger needs-rebase and that a draft outranks
a conflict. Proven non-vacuous: dropping the mergeability arm fails 4
assertions, dropping the STALE precedence fails 2, restoring returns 29/0.
Closes #136
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
2026-07-20 15:18:49 +00:00
|
|
|
# -- the happy path survives all of the above.
|
|
|
|
|
REVIEWS_JSON="$ALL_APPROVE" MERGEABLE=MERGEABLE CHECKS=SUCCESS REQUESTED=""
|
|
|
|
|
expect "mergeable + green + three head-current approvals is needs-human" state:needs-human "$(decide_state)"
|
|
|
|
|
# -- and a draft outranks everything, including a conflict.
|
|
|
|
|
DRAFT=true MERGEABLE=CONFLICTING
|
|
|
|
|
expect "a draft is building even when conflicted" state:building "$(decide_state)"
|
|
|
|
|
DRAFT=false MERGEABLE=MERGEABLE CHECKS=SUCCESS REQUESTED="" REVIEWS_JSON='[]'
|
|
|
|
|
|
fix: verdict contract, head-bound approvals, serialized reconcile — and a testable state machine
Round-1 blockers, all three reviewers concurring:
- COMMENTED agreement now counts: agreement_signal recognizes the live bots'
durable markers (Verdict: Approve / I agree with everything / leading ✅) —
the gate to needs-human can actually close. Formal verdicts remain the
contract (CONTRIBUTING), this is the documented transitional workaround.
- Every counting verdict is bound to the head SHA; a stale approval parks the
PR in addressing (agent owes re-request) instead of promoting unreviewed
code. CHANGES_REQUESTED blocks at any head, per GitHub's own semantic.
- reconcile serializes under ONE job-level concurrency group; scope stays
per-PR. No more cron-vs-event race on the request-the-human-once guard.
- Sweep resilience: per-PR subshell (one failure logs and continues), label
edits warn instead of wedging; the self-heal claim now matches reality
(dispatch-only bootstrap).
- The state machine is extracted pure (globals in, state out) and sourceable:
test/labels-reconcile.sh proves 14 fixture transitions — comment-only
agreement, stale approval, comment-without-verdict, human precedence and
human-block — wired into CI.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-18 19:31:00 +00:00
|
|
|
printf 'labels-reconcile tests: %d passed, %d failed\n' "$pass" "$fail"
|
|
|
|
|
[ "$fail" -eq 0 ]
|