The merge base is the commit the fork's author branched from, so a
branch cut from an old commit selected an old detector, and a fix to
the detector never reached the PRs an attacker controls. The workflow
itself comes from the default branch, so the detector now does too.
The head is also checked out with its links intact. Flattening them
hid a CRD behind a file link from the detector while Helm still
shipped it, and made the verdict differ from the one the same-repo
gate reaches.
Assisted-by: LLM
Signed-off-by: Myasnikov Daniil <myasnikovdaniil2001@gmail.com>
A commit status belongs to the commit. When two open fork PRs carry the
same head, a success already standing on it may have been earned by one
of them on another base, or may rest on an approval that was withdrawn
on a PR whose run waits for a maintainer and so never reaches resolve.
The sweep skipped such a head and left that success in place, which let
a duplicate fork PR keep a withdrawn approval from taking effect.
The sweep now writes the same failure resolve writes, once per head,
when a success set by a workflow stands on a head that only fork PRs
carry. It still never creates a status, and a head that an open
same-repo PR carries stays with api-gate-report.
Assisted-by: LLM
Signed-off-by: Myasnikov Daniil <myasnikovdaniil2001@gmail.com>
api-gate exits 2 to report a sizeable change, and the runner runs a step
without an explicit shell as bash -e. The fork detect step switched
errexit off only after reading the exit code, so the first failing
command ended the step with 2 before sizeable=true was recorded. Every
fork PR with a sizeable API change then got "gate evaluation did not
complete", which the sweep deliberately ignores, so an owner's approval
could not turn it green by either path. The same-repo step already
switches errexit off before the call.
Assisted-by: LLM
Signed-off-by: Myasnikov Daniil <myasnikovdaniil2001@gmail.com>
A commit status belongs to the commit, so two open PRs that carry the
same head share one "API Review Gate" status. The fork workflow
attributed that status to a fork PR by head SHA alone. A fork PR opened
at a commit of this repository could therefore overwrite the verdict of
the same-repo PR that owns it: resolve wrote a failure after
api-gate-report had posted success, and the sweep applied the fork PR's
reviews to a status the other PR had earned. A fork can reference any
commit of the repository's network, so anyone able to open a PR could
hold another PR's gate red.
The fork workflow now leaves a head that an open same-repo PR carries to
api-gate-report, and the sweep skips any head carried by more than one
open PR. Two fork PRs at one head still fail closed in resolve, since
their bases, and so their verdicts, can differ.
The header of the same-repo workflow said skipped and cancelled runs
post nothing. That stopped being true once a review posts pending ahead
of its evaluation, so it now says they post no verdict.
Assisted-by: LLM
Signed-off-by: Myasnikov Daniil <myasnikovdaniil2001@gmail.com>
On a fork PR an API owner's approval could never turn the status green.
For a pull_request_review run on a fork PR, workflow_run.head_repository
names the base repository and pull_requests is empty, so the fork
publisher took every review-triggered run for a same-repo one and
skipped it, while the same-repo reporter skipped it for being a fork. A
sizeable fork PR stayed on the failure its push run had posted.
Fixing the guard alone would leave most forks stuck. The repository
makes every run from an external contributor wait for a maintainer's
approval, and a run in action_required never fires workflow_run, so
their review-triggered runs never reach the publisher. And a fork's run
executes the fork's own copy of the gate workflow, so the conclusion the
publisher relayed was whatever the fork made it.
The privileged workflow therefore reaches the verdict itself, from
inputs a fork cannot forge, and a scheduled sweep covers the runs that
never execute. The sweep only moves a status the full evaluation already
set after finding a sizeable change, so it cannot create a pass.
The same-repo gate now posts pending as soon as an actionable review
arrives: a dismissal can turn a green head red, and the old success
would stay mergeable for the minute the evaluation takes. Both paths
also ignore COMMENTED reviews when taking each reviewer's latest, since
an owner who approved and then commented had withdrawn their own
approval as far as the gate was concerned, which GitHub does not do.
Assisted-by: LLM
Signed-off-by: Myasnikov Daniil <myasnikovdaniil2001@gmail.com>
The gate's merge context was its job name, which inherits two GitHub
behaviours that together make it both unenforceable and self-poisoning.
A required *job* that skips counts as PASSED by branch protection, so
any path that skips the evaluating job hands out a free pass — the hole
that let fork PRs merge with e2e never run (#3257). And check runs are
not deduplicated per name: every workflow run gets its own check suite,
so `GET /commits/{sha}/check-runs?filter=latest` returns the same set as
`filter=all`, and the rollup is worst-of. Two runs on one head SHA
therefore leave two verdicts standing for one context name, and a
cancelled run poisons the context until the next push. Commit statuses
are deduplicated per context, newest wins.
Rename the evaluating job off the gate context and publish "API Review
Gate" as a commit status: `api-gate-report` for same-repo PRs, and a
privileged workflow_run for forks, whose GITHUB_TOKEN is read-only on
both the pull_request and the pull_request_review trigger and so can
never post a status itself. Skipped and cancelled runs post nothing, so
declining to post can only ever preserve a verdict some real run
reached — it can never invent one.
That invariant is what makes it safe to stop re-evaluating on the plain
`commented` reviews the review bots post: such a review cannot move a
verdict that counts only APPROVED reviews, and the run that skips now
leaves the previous status untouched instead of overwriting it.
Also split the concurrency group so a run that computes no verdict
cannot cancel one that does. A bot commenting during the ~30s
evaluation used to cancel it and then contribute nothing, leaving a
cancelled run as that SHA's only api-gate verdict (41d1153c, 4b1966c5).
For the gate to actually block a merge, main branch protection must
require the "API Review Gate" status context; until then it is
advisory, as the job name was.
Assisted-by: LLM
Signed-off-by: Myasnikov Daniil <myasnikovdaniil2001@gmail.com>