Two defects, one cause, and the second is why the first survived a week.
THE WRITE. Removal was a per-label `DELETE .../labels/{id}` loop. On this
instance that call returns HTTP 500 for every removal under the token the
sweep actually holds — measured inside Actions, probe run 701, where the same
`PUT .../labels` with the desired full set returns 200 including the empty set
for a full clear. A PAT gets 204 on the same DELETE, which is exactly why it
went unseen: it fails only for `${{ github.token }}`.
Net effect before this: on Forgejo the state machine could only ever ADD
labels. Every `state:*` transition needing the previous state cleared and every
`blocker:*` that should lift was inert. Both PRs open right now carry stale
`blocker:*` labels that are false and that nothing can remove.
So the removal path is read-current, compute-wanted, one PUT — the same shape
the assignee branch beside it already used. An ADD-ONLY call keeps its additive
POST: ceremony#128 lost a `release` label to a read-modify-write that clobbered
a concurrent set, and forge_labels_add stays pinned against ever doing that.
The window is accepted here and only here, where the caller asked to REMOVE
and no additive verb can say that. An unresolvable --add-label refuses before
any write, so a replacement PUT can never drop a label nobody asked to remove.
THE REPORTING. `labels-reconcile` logged `WARNING: label edit failed`, fell
through, and `main` printed `reconciled.` and exited 0 — while
`issueflow-reconcile` treated the identical 500 as fatal. One cause, two
contradictory policies, and the wrong one hid the write defect.
A failed write is fatal now, and the tally reaches main's exit code. That
second half is load-bearing: making reconcile_pr fatal alone is not enough,
because the loop swallows a per-PR non-zero into a log line and finishes. The
per-PR tolerance is right and stays — one bad PR must not blind the board — but
it now applies to READS. A sweep that could not write exits non-zero and never
prints `reconciled.`
The diagnostic says what was attempted and that it did not happen. The old text
blamed a missing label and told the operator to bootstrap, when the label was
present and the call returned 500 — #101's rule is report, do not diagnose.
Mutation-tested, all three ways: restoring the warn-and-continue reds 5 cases,
removing the tally reds 2, restoring the DELETE loop reds 7.
test/run.sh 22 files 0 failed under jq 1.7 and jq 1.6; shellcheck 0.10.0 and
actionlint clean.
Refs #192
@codex-reviewer-andresmgsl #4780, concurred by @grok-reviewer-andresmgsl
#4785. All four real.
1. Three issueflow call sites still named per_page=100. The backend
sanitized it so it worked, but the frozen term and the changelog both say
no call site names a page size — and a contract that holds only because
something downstream cleans up is not the contract. Endpoints now carry
their logical query alone.
2. The suite's summary and `[ "$fail" -eq 0 ]` gate sat in the MIDDLE of
test/labels-reconcile.test.sh, and the eight outstanding_requests expects
were appended after them. Proven before fixing: a deliberately broken
term-4 assertion printed FAIL, was excluded from the totals, and the
suite still exited 0. Those assertions were decorative. The gate moves to
the true end, with a note that nothing goes below it; the reported count
goes 157 -> 164, which is the eight that were never being counted.
3. forge_labels_add and forge_request_reviewer arrived with the port and had
no boundary pins. Both backends now have them, and the labels_add cases
pin the property ceremony#128 turns on: an additive POST, never a PUT of
the whole set, exactly one write so nothing is read-modify-written.
Mutation-verified — making it RMW/PUT, or routing github through
`issue edit --add-label`, each red their own cases.
4. The historical comment said the old gathers were `forge_api graphql`. My
own mechanical port rewrote it; before #188 they were `gh api graphql`
and the abstraction did not exist.
Refs #188
Term 4. GitHub clears requested_reviewers when a verdict lands, so the field
answers "who still owes a verdict" by itself. Forgejo never clears it —
measured: rig!140 listed all three panelists with all three verdicts in, and
rig!146 still lists three while MERGED, so the field is stale even on a
closed PR.
Read raw on Forgejo that is not a cosmetic over-count. `requested` drives
three decisions, and a permanently-true field pins a PR at
state:bots-reviewing for life and stops blocker:unrequested from ever being
true: the sweep believes a round is live forever and no staleness can
correct it.
So the requested set is intersected with who has NOT submitted a verdict for
the current head, derived from /pulls/{n}/reviews — the read that is true on
both forges. On GitHub the filter removes nothing, because the field is
already accurate; term 5 holds by construction rather than by care.
A STALE approval — an approval of an older head — still owes a verdict. That
is the case that matters: treating it as answered would let a stale round
read as complete, which is the shape #136 exists to prevent.
Mutation-verified both ways: reading the field raw again reds three cases,
and treating STALE as answered reds two.
Also documents @grok-reviewer-andresmgsl's ask (#4763): every panel= account
must be able to read the repo, or the forge refuses the review request —
422 naming the account on Forgejo. A real failure mode for private
consumers, and it fails loudly rather than sweeping blind.
Refs #188
Term 1 completed. All 52 runtime gh call sites in the three reconcilers and
lib/ruling.sh now go through forge_* verbs; the three remaining matches in
labels-reconcile are prose in comments. lib/facts.sh is deliberately
untouched — it is the release door, and the ruling keeps release.yml out of
this issue.
The CEREMONY_FORGE_CLIENT:-gh wrappers die here, in the same commit as the
sites they described, so the tree is never in a state where the declaration
lies. main() now runs forge_preflight then forge_select "".
Two sites needed judgment rather than substitution:
- labels-scope's write is forge_labels_add, a genuine additive POST on
both backends, NOT forge_issue_edit --add-label. ceremony#128 turns on
that write not being a read-modify-PUT: the labeler action computed
(labels-at-job-start union derived) and PUT the whole set, silently
dropping a label applied while the job ran. Routing it through a generic
edit verb would have quietly reopened that.
- the human-review request is forge_request_reviewer. Contrary to my
earlier reading, POST /pulls/{n}/requested_reviewers DOES exist on
Forgejo — 422 naming the reviewer's access without it, 201 with it. The
earlier 404 was a GET, which the endpoint does not serve, plus a
username that did not exist.
Test churn, all of it the term-5 boundary move:
- the suites select the github backend, so their existing gh() stubs stay
the boundary and keep intercepting;
- stubs strip the paging the shim injects, so fixtures stay keyed on the
logical endpoint (inlined in the PATH stub, which is a standalone
executable and cannot see a shell function);
- fixtures renamed off the per_page suffix for the same reason;
- recorded-mutation assertions now match the verb, not the raw gh line;
- gh() stubs carry SC2317: they are reached through the backend now, so
shellcheck can no longer see the call path.
Refs #188
The preflight half of #188, landed first so it stands alone: the forge is
decided once, before any sweep, and a client that cannot speak it exits
non-zero with a named reason.
Measured against forgejo.heavyduty.builders at 84bb1a4 — two of the three
actions reported SUCCESS having read nothing:
labels-scope exit 0 "no .github/labeler.yml" (the file is HTTP 200)
labels-reconcile exit 0 "reconciled." (zero PRs enumerated)
issueflow-reconcile exit 1 "unexpected end of JSON input"
labels-reconcile's blind-sweep warning (#96) could not fire: it counts
unreadable PRs against a list `gh pr list` never produced, and a process
substitution's failure does not trip set -e, so total stayed 0. Installing
gh makes it worse, silencing the one loud failure.
Detection is measured, not inferred from docs: a real forgejo-runner v6.3.1
job (probe task 278) shows Forgejo populating the whole GITHUB_* namespace,
so GITHUB_ACTIONS proves nothing. GITHUB_API_URL's shape, GITEA_ACTIONS and
GITHUB_SERVER_URL do. The same probe shows the runner image carries neither
gh nor stoke, which is what makes the forgejo backend REST.
Tests declare CEREMONY_FORGE at the forge boundary rather than stubbing gh
and staying silent about the forge — the boundary move term 5 asks for.
Refs #188
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>
The unit block now feeds blind_sweep_warning a sampled reason and asserts
the new lead plus two must-fail guards: the disproven 'grant checks: read
and statuses: read' diagnosis stated as fact goes red, and so does any
drift in the counted line's whole-line shape (exactly the blind PRs match,
no more, no less — a reason line that matched would double-count, a folded
reason would undercount). read_failure_reason is covered pure: D4 wording
for empty stderr, multi-line collapse to one line, 400 chars truncated to
300 plus ellipsis within the 304-byte bound, 300 passing through whole.
blind_main_probe's gh pr view stub now fails with a denial on stderr, the
way real gh fails.
Part of #101.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
CI's shellcheck gate treats info findings as red. The gh() stubs paired
with bootstrap_labels are reached only through run's "$@", which
shellcheck cannot trace (the older probes stub reconcile_pr, which calls
gh directly), so they carry reasoned SC2317 directives; the LABELS.md
doctrine parse carries SC2016 for its literal backticks. Probes now live
in named functions, matching the house *_probe() pattern.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Registry-vs-LABELS.md identity, happy path, missing label, one-name 403,
DRY_RUN narration, and the executed-subprocess dispatch (#91's lesson: a
sourced probe cannot see set -e). Cron and pull_request_target delete
nothing. Verified red without the guard: the unguarded dispatch dies on
the FIRST absent label, one delete of six attempted.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Pure decisions (shape presence, rung boundaries, Default: parse for wording
only), sweep probes for every AC path (malformed-once, conforming silence,
rungs despite activity, cron progression, missed-moment skip, re-flag
episode, unreadable comment list, malformed+rung same pass), and the
existing nudge fixtures updated to conforming escalations with pre-seeded
rung markers so each probe observes one behavior alone.
Part of #73.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Real 'gh api --jq .[].created_at' on an empty collection emits no lines;
the stub printed a literal '[]', which under byte-wise collation sorts
after ISO-8601 timestamps and poisoned the PR-surface probe's
last_activity. Route the synthesized empty array through the same jq
projection as a present fixture, and pin the test's collation so the
verdict cannot flip with the runner's ambient locale.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Issue side: invariant-1 composition, the reclaim clock stopping under a
pending ruling (with a flag-free reclaim control), the stale heal, label
churn invisible to the activity clock, the surface-level nudge reset, and
no edit anywhere naming the flag. PR side: the wired nudge riding the
stale sweep's activity computation, one nudge across two sweeps, #51's
stale skip intact.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
