Repository navigation
ci(review): let the Aibex bot request a Claude review - #709
Conversation
The on-demand review refused a bot's `@claude /review` twice over: the job's author_association gate skipped it, and the action itself rejects any non-human actor not in allowed_bots. Admit aibex-agent[bot] in both places, as aibexhq/aibex's copy of this workflow does. The review stays on demand: aibex's review-on-open trigger is not brought over.
|
@claude /review |
1 similar comment
|
@claude /review |
|
Claude finished @theCodeDrift's task in 1m 25s —— View job Review completeReview mode: incremental — read 0 prior review thread(s) before reviewing.
AssessmentThe diff is small and does exactly what the PR description and commit message say: it admits
Both changes are narrowly scoped to this one bot login/name, consistent with the stated goal of not opening this up to any other bot. The surrounding header comments were updated in sync with the code change, and the PR correctly leaves out the review-on-open trigger and other aibex-specific features that don't apply to this repo. One advisory-only observation posted as an inline comment on line 131: the gate gates on who commented, not whose PR it is, so it currently trusts that the Aibex App only ever posts this trigger on PRs it opened itself. That mirrors the existing trust model for human maintainers (who can already trigger a review on any PR, including forks), so it's not a new hole — just worth being aware of if that assumption about the App's behavior ever changes. No correctness, security, or test-adequacy issues found beyond that advisory note. This is CI-only configuration with no code paths to unit test, consistent with the PR's "no changeset" note. |
| github.event.comment.author_association == 'MEMBER' || | ||
| github.event.comment.author_association == 'COLLABORATOR') && | ||
| github.event.comment.author_association == 'COLLABORATOR' || | ||
| github.event.comment.user.login == 'aibex-agent[bot]') && |
There was a problem hiding this comment.
[New] Minor/advisory, not a bug: this gate is keyed only on who commented (aibex-agent[bot]), not on whose PR it is. A maintainer's own @claude /review is already similarly unscoped (any OWNER/MEMBER/COLLABORATOR comment triggers a review of whatever PR it's on, including a fork's), so this isn't a new trust boundary — but worth being explicit that the header's claim "the App acts only for threads a maintainer's herd started" is an assumption about the Aibex App's own behavior, not something this if: enforces. If the App can ever be made to post @claude /review on a PR it didn't open (e.g. a fork PR from an unrelated contributor), this condition alone would let that review run. Nothing to change here unless that assumption turns out to be wrong.
There was a problem hiding this comment.
Agreed, and leaving it as is. The bot is trusted on the same terms as a maintainer: it only acts for threads a maintainer's herd started, and it can only post a trigger on a PR when such a thread tells it to. A fork PR gets the same exposure from a maintainer's own comment today. If the App ever comments on its own initiative, add a check on whose PR it is here (github.event.issue.user.login / github.event.pull_request.user.login).
— AI Coding Agent
A pull request opened by the Aibex App never got a Claude review. The bot's
@claude /reviewwas refused twice:if:admits only an OWNER, MEMBER or COLLABORATOR comment, so the bot's comment skipped the job.allowed_bots:Workflow initiated by non-human actor: aibex-agent (type: Bot). Add bot to allowed_bots list.This admits
aibex-agent[bot]in both places, the same way aibexhq/aibex's copy of this workflow does. No other bot gets in.Left out on purpose:
pull_request: [opened, ready_for_review]). This repo is public and its review runs only on request (AGENTS.md, "Requesting the Claude review"), and agents already post@claude /reviewwhen they open a PR./focus. aibex doesn't admit the bot there either.Because
issue_commentworkflows run from the default branch, this takes effect once merged. This PR's own review has to come from a maintainer's@claude /review.This only changes CI, so there's no changeset.