Found by the first post-merge sweep after the 0.6.0 merge — #198's own
acceptance probe — not by review. Three PRs in one run:
labels: #208: could not read the head commit's date:
forge_api: HTTP 404 from 'GET repos/heavy-duty/ceremony/commits/f3a1336…'
— blocker:unrequested not judged this pass
Measured against this instance:
forgejo repos/{o}/{r}/commits/{sha} -> 404
forgejo repos/{o}/{r}/git/commits/{sha} -> 200, date under `.created`
github repos/{o}/{r}/commits/{sha} -> 200, date nested
A fourth asymmetry, alongside the three lib/forge-forgejo.sh's header already
records. #198 ported this call site onto the shim with GitHub's path unchanged
— correct against GitHub, and the block it lives in (#236 D2) arrived WITH the
merge, so nothing here had ever executed it.
So it becomes a verb rather than a path at the call site: the caller wants one
timestamp and should not have to know either shape.
Cost while it stood was bounded and loud rather than silent — guarded_read
refused and the sweep said so — but blocker:unrequested could never be judged
on this forge.
The tests pin each backend's PATH and FIELD, because a stubbed forge_api cannot
catch a wrong path; that is exactly how this shipped and why it took a live
sweep to find. Swapping the paths reds the forgejo pair; swapping the fields
reds the github one.
test/run.sh 28/28 under jq 1.7 and jq 1.6; forge-backends 124/124; shellcheck
0.10.0 and actionlint clean.
Refs #209
@codex-reviewer-andresmgsl caught a test that describes evidence it does not
collect — mine, and it is the class this PR is about.
Two cases were titled "naming the verb, path and status" and asserted only the
substring "500". The PUT boundary happened to satisfy the contract because
forgejo_write already passes "PUT $endpoint" to forgejo_http_ok. The GET
boundary did not: the diagnostic was `HTTP 500 from 'repos/o/r/issues/5'`, with
no verb at all — so a caller could not tell a failed READ from a failed WRITE
of the same path, and #192's acceptance criterion asks for exactly that
distinction.
Reads now pass "GET $endpoint" on both non-paginated and paginated paths, and
the two tests assert the complete expected diagnostic as one substring rather
than a status code that any failure would contain. Reverting the verb reds the
GET case.
Also, per the same review: the failed GET is asserted to write nothing, and the
failed PUT to have attempted exactly one write.
forge-backends 117/117 (was 115), test/run.sh 22/22 under jq 1.7 and jq 1.6,
shellcheck 0.10.0 and actionlint clean.
Refs #192
@codex-reviewer-andresmgsl's four gaps, all real, all taken.
1. PRESERVED IDS COME FROM THE ISSUE. The removal path read only .labels[].name
and then re-resolved every preserved label through the repository-wide list
— so preservation depended on a paginated read with nothing to do with this
issue, and an incomplete one would drop a bystander. It now keeps
name<TAB>id from the issue payload, subtracts removals by name, and resolves
ONLY added names. Fixture: a bystander on the issue with id 14 that is
absent from the repo-list fixture entirely must still survive the PUT.
2. AN ABSENT REMOVAL WRITES NOTHING. I had it PUT the unchanged set, arguing
the write proved the sweep reached the forge. The GET already proves that,
and replacing a set with itself opens ceremony#128's window for no state
change — most calls here are exactly this case, since the reconcilers call
--remove-label unconditionally. Short-circuits when the wanted set equals
the current one. This was the policy-shaped choice flagged for @andres; the
reviewer's reasoning is better than mine was.
3. EVERY LABEL MUTATION REACHES THE TALLY. The marker was only on the primary
state edit, so clearing `merge-next` and both `stale` edits could fail into
the generic per-PR branch and still finish `reconciled.` and exit 0. All
four sites go through one `label_write` helper, so a future call site cannot
reopen it by forgetting to mark itself. Probe: a failed NON-primary write
(unstale on a blocked PR) must fail the sweep.
4. NO SUCCESS TOKEN IN A FAILURE TAIL. "NOT reconciled." still contains
"reconciled.", which a log-tail consumer greps for. The line is now
"sweep incomplete", and the test asserts the whole output is free of the
token rather than only of the success prefix.
Also added the two fault boundaries the acceptance plan named and the fixtures
never proved: a failed current-label GET and a failed replacement PUT, each
non-zero with the backend's verb/path/status diagnostic.
Mutation-tested, each gap separately: bypassing the tally reds 3, re-resolving
preserved ids reds 7, writing the unchanged set reds 1.
forge-backends 115/115 (was 110), labels-reconcile 175/175 (was 172),
test/run.sh 22/22 under jq 1.7 and jq 1.6, shellcheck 0.10.0 and actionlint
clean.
Refs #192
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
Both findings are @codex's on !193 (#1583), and both are real.
The asset name travels as a QUERY VALUE, and the artifact-hook contract
permits any file the consumer drops in RELEASE_ASSETS_DIR. Raw
interpolation meant `release asset.tgz` made curl reject the URL outright
(exit 3), and '&', '#', '+', '%' silently changed the name or the query's
shape. `gh release create` handled all of those, so a 1:1 port had to.
Encoded through one boundary — jq's @uri, since jq is already a hard
dependency of this backend and a hand-rolled sed class is how the next
unescaped character gets through. Six backend cases cover it: the encoder
on a space and on the delimiters, uploads under both names, the created
release id in the path, and the multipart attachment. Mutation-checked:
dropping the encoder fails exactly the two name assertions.
docs/CONSUMERS.md's artifact-hook recovery still told operators to "run
`gh release create` by hand" and described the hook as running "before
`gh release create`" — on a Forgejo runner that is precisely the failure
this PR fixes. It now names the forge-neutral tag-door recovery first and
shows both clients for the manual path, without regressing the GitHub
guidance.
1035 assertions, 22 suites, shellcheck-all and actionlint clean.
Refs #191
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Both panel blockers on c63a550.
@kimi found the one that mattered: facts.sh got the REPO fix, release.yml's
own four call sites did not. A workflow `run:` shell carries no `set -u`, so
an unset REPO expands empty and the verb addresses `repos//…` — which 404s,
and the 404 is then read as an ANSWER. Reproduced read-only against this
instance before fixing:
forge_release_exists 0.4.1 -> "no", rc 0
forge_commit_pulls 7fc9afe4 -> "[]", rc 0 (the !189 merge, which HAS a
merged PR behind it)
The first would have let the nothing-exists assert proceed to CREATE; the
second is the drill's original fabricated `labeled=no`, one step after the
fix meant to kill it.
Fixed once rather than at four call sites, as kimi suggested: forge_select
defaults REPO from GITHUB_REPOSITORY, and forgejo_api_base — which every
verb reaches the network through — refuses an empty REPO outright. No fifth
call site can forget it.
@grok and @kimi both blocked on the same AC gap: the backend suite did not
cover the five new verbs, so the two measured asymmetries had no offline
coverage. test/forge-backends.test.sh now has 15 cases for them — singular
/pull wrapped to an array, 404 as an empty array, 500 refusing, release
present/absent/unreadable, POST /tags vs /git/refs, the publish body, and
the REPO-empty must-fail. Mutation-checked: reading the plural path fails
one case, dropping the REPO guard fails the two must-fails.
1029 assertions, 22 suites, shellcheck-all and actionlint clean.
Refs #191
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
@grok-reviewer-andresmgsl landed the term-5 pins in ff17d1e while I had the
same two asks in flight locally; theirs is on the branch and I dropped my
duplicate rather than push a competing tip. This adds only what the two
suites did not share.
Their pins are positive: the github twins DO call the right endpoints, and
forge_pr_activity emits all three timestamp sources. Mine had two negatives
they did not, and negatives are what catch the drift a 1:1-extraction path
actually suffers — a positive pin still passes if the github path GAINS
forgejo behaviour, and term 5 is a statement about what must NOT change.
- the github timeline is never reshaped. That timeline already IS the
shape ruling.sh selects on, so a projection here would be a second,
divergent normalizer maintained by nobody.
- github activity never derives inline comments from reviews. That
derivation exists on the forgejo path only because the flat endpoint
404s there; a github twin quietly adopting the workaround is the
"both backends drift together" failure term 5 forbids.
Mutation-verified: adding a --jq projection to the github timeline, and
swapping the flat PR-comments read for a reviews-derived one, each red their
own case. forge-backends 77 -> 79.
Refs #188
Codex 1566 held APPROVE: only Forgejo stubs covered forge_timeline and
forge_pr_activity. Pin the github twins as 1:1 extractions (timeline
paginate; issue comments + flat pulls comments + commits).
Cluade #4879: drop 2>/dev/null on the labels-reconcile activity call site
so a failed read still degrades last_activity but names the failure in the
job log (keep || true).
Panel-unanimous batch that was staged unpushed on 57abe15 (#4853):
- forge_timeline: project Forgejo label events into the GitHub shape
so the ruling ladder fires on this forge (measured mapping #4849)
- forge_pr_activity: stop calling /pulls/{n}/comments (404 here); use
reviews with comments_count > 0 for inline comments (#4844)
- ci.yml: install shellcheck before lint, mirroring actionlint — the
act-22.04 runner image does not ship it
Status captured before jq so an unreadable timeline cannot report empty.
@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
The coverage owed with the call-site port (@grok-reviewer-andresmgsl #4741
note 2, #4751 item 2). Live scratch-repo evidence proved these work; these
pin the request SHAPE so they keep working.
- a removal resolves name -> numeric id, and never sends the name as the
path segment (measured: DELETE .../labels/probe:one -> 422,
DELETE .../labels/149 -> 204);
- a removal of a label the repo does not have writes nothing, matching gh:
the reconcilers call --remove-label unconditionally to converge state;
- adds take names directly, one request, comma-separated values split as
gh splits them;
- an assignee removal PATCHes the SURVIVING list, because Forgejo sets
assignees rather than adding and removing them — a naive translation
would have cleared every other assignee as a side effect of removing
one, which is what the mutation test proves is caught.
Payloads are now compact JSON. They were pretty-printed, which spread a
single write across several lines — harder to read in a log, and it hid the
shape from any assertion matching a line.
Refs #188
a968e13 was pushed with shellcheck red. I chained the gates and the push in
one command, so a non-zero gate did not stop the push — the gate has to be a
condition, not a line of output I read afterwards.
Refs #188
@codex-reviewer-andresmgsl's three findings (#4743), all real.
1. forge_label_create is now an UPSERT, matching gh label create --force.
bootstrap_labels creates every declared label on EVERY workflow_dispatch,
so a plain POST onto an existing name aborted the bootstrap under set -e
from the second dispatch onward. Resolves name -> id and PATCHes when it
exists.
2. forge_pr_view carries createdAt/completedAt. checks_state groups repeated
contexts and selects the newest by [.startedAt, .createdAt, .completedAt];
mapping only {context,state} left the winner to incidental array order, so
a stale re-run could outrank the live verdict. The combined status carries
created_at and updated_at — measured.
3. forge_issue_edit refuses unknown flags and missing values. The github
backend hands them to gh, which fails; dropping them here turned a
mis-typed port site into a mutation that silently did not happen — this
issue's own failure class, inside the fix for it.
Also settles @grok-reviewer-andresmgsl's note 3 (#4741): Forgejo Actions DO
land as commit statuses on this instance, so the rollup is not empty.
rig main carries four — "ci / check (push)" and siblings, state success,
each with created_at. statusCheckRollup therefore populates, and NONE is not
silently substituted for SUCCESS.
Each fix mutation-verified: dropping the timestamps, forcing POST-always, and
restoring the silent flag skip each red exactly their own cases. The
newest-verdict case drives the real checks_state, not a copy.
Refs #188
@codex-reviewer-andresmgsl (#4727) and @grok-reviewer-andresmgsl (#4734):
"...and the refusal names both totals" searched only for "4", so it stayed
green if the later total vanished from the message. A case named "names
BOTH" must fail when one goes. Now asserts "4 then 9".
Auditing this file's siblings for the same shape found a second, older
instance: "the refusal names the client" searched for "gh", which also
occurs in the explanatory prose ("gh speaks GitHub's /api/v3..."), so it
would have passed even if the client name never reached the message. Now
asserts "the 'gh' client cannot speak it".
Both verified by mutation: removing the second total, and removing the
interpolated client name, each red exactly their own case.
Refs #188
@codex-reviewer-andresmgsl's three findings (#4712), each a route by which
an unprovable read could still be reported as a whole one — the guard
leaking the failure class it exists to stop.
1. x-total-count was never validated. `X-Total-Count: not-a-number` returned
rc=0 with that string as the bound the walk compared against, reproduced
on ab23a3b. Now required to be a canonical non-negative integer.
2. The total was read once. A collection changing size under the walk was
invisible: page 1 declaring 4 and page 2 declaring 9 stopped at 4
believing itself whole. Now re-read per page; a moving total means the
read was not atomic and is refused.
3. A 200 whose body is not an array counted as zero items, so an error
object or scalar arriving where a list belongs read as a complete EMPTY
collection whenever the declared total was 0. Now refused, quoting the
body. A genuinely empty array is still fine — covered.
Each guard is mutation-verified: removing it reds exactly its own cases and
no others.
Refs #188
SC2016 on the deliberate single-quoted bash -c (the expansion belongs to
the isolated process, as the sibling case in issueflow-reconcile.test.sh
already documents), and SC2317 on the curl stub, which shellcheck cannot
see is invoked indirectly by forge_api.
Found only after committing, because .github/scripts/shellcheck-all.sh
derives its lint set from `git ls-files` — an UNTRACKED file is not linted
at all. "Gates clean" measured before `git add` was measuring a set that
excluded the file just written. Verified from a clean clone at the pushed
SHA, which is what caught it.
Refs #188
Term 1's foundation. lib/forge.sh gains forge_select, which sources exactly
one of lib/forge-github.sh or lib/forge-forgejo.sh; both define the same
verbs, so no branching reaches the 61 call sites. The github backend is the
current gh invocation extracted 1:1 — term 5 is kept by making that path
boring.
The page size moves OUT of the call sites and into the backend, because it
is not portable and fails silently. Measured 2026-08-02:
?per_page=100 GitHub 100 items Forgejo 30 items (ignored)
?limit=100 GitHub 30 items Forgejo 50 items (capped)
Both answer HTTP 200 with valid JSON. Every call site here is GitHub-shaped,
so a verbatim port would have swept 30 of rig's 137 issues and printed
"reconciled." — criterion 2 failing green, the same failure class as the
blind sweep. Both page_url helpers strip a stray page-size parameter in
either dialect, so a call site cannot reintroduce it by accident.
Forgejo caps a page at 50 whatever is asked, so pagination is mandatory, not
an optimisation. The gather is then PROVEN complete against x-total-count
rather than assumed complete because a loop ended.
@kimi-reviewer-andresmgsl's hardening (#4699): a missing x-total-count is
itself a loud refusal. Header exposure is a server setting, and an assert
that cannot run must not silently pass — that is the failure class
re-entering through the guard built to stop it.
Call sites are not ported yet; that is the next commit.
Refs #188
