As it seems impossible to prevent claude from receiving notifications
about subagents finishing after it has produced structured output, which
breaks the structured output as it has to be the final reply, let's stop
using subagents and background tasks completely to avoid the issue.
claude keeps failing by its subagents completing after it has already
written the review for large prs. It seems to run out of turns, tries
to get the subagents to post partial reviews but doesn't seem to stop
them.
Let's insist that it waits for background tasks to stop but let's also
increase the max turns a bit so it doesn't run out as quickly.
claude wants to use python to access the JSON context so let's allow
it. Since python3 basically allows you to reimplement every other tool,
let's just enable all tools except the web related ones but enable network
isolation so it can't try to exfiltrate anything via python.
In CI we get spurious failures about unitialized variables with gcc
versions older then (depending on the case) 12, 13, or 14. Let's only
try to do this check with newer gcc which returns more useful results.
At the same time, do compile with both gcc and clang at -O2, just
disable the warning.
The old logic seems to have been confused. We compile with -Wall, at
least in some cases, which includes -Wmaybe-unitialized. So if we
_don't_ want it, we need to explicitly disable it.
The labelling approach introduced in 6089075265
means contributors can now trigger the workflow on their own when the label
is added by a maintainer and they update the PR. Hence we need to allow all
users to access the claude code action. This is safe because we already gate
the workflow ourselves to only the contributors that we want to allow.
Additionally, the claude code job has no permissions anymore except read access
to the repository and can execute very limited tools, so this should be safe.
The name doesn't actually matter, it gets replaced with the name
of the file when not archiving. So stop passing a name and pass in
the filename as the name when downloading the artifact.
Rather than have claude fetch the context itself, let's fetch the
context for it in the setup job. This has the following advantages:
- We can reduce the permissions granted to the claude job
- claude has less opportunity to mess up trying to fetch the context
itself. Specifically, it keeps spawsning a background task to fetch
the PR branch which messes up the structured output at the end, causing
the review job to fail. By pre-fetching the context it won't have to
spawn the background task. Additionally, we limit the git commands it
can execute to local ones to ensure it doesn't try to fetch the PR branch.
Finally, we fetch the branch ourselves as pr-review so claude can look at it
to review the PR.
- If a pr is labeled with claude-review, review it immediately
- If a pr labeled with claude-review is updated, review it regardless
of the author
- If a pr is opened by a maintainer, review it and add the claude-review
label. If the claude-review label is later removed, the pr won't be
auto-reviewed anymore.
Bash(gh:api *) wasn't actually working. Turns out the colon syntax
is deprecated and unnecessary. Let's stop using it which also fixes
the bug so that gh api calls are allowed now.
Claude now identifies which existing review comment threads should be
resolved (because the issue was addressed or someone disagreed) and
returns their REST API IDs in a new `resolve` array in the structured
output. The post job uses GraphQL to map comment IDs to threads and
resolve them.
Also switches all GitHub data fetching from MCP tools to `gh api` calls,
since the MCP tool strips comment IDs during its GraphQL-to-minimal
conversion and cannot be used for thread resolution.
The thread resolution GraphQL pagination is wrapped in a try/catch so
that a failure to fetch threads degrades gracefully instead of aborting
the entire post job. Unmatched comment IDs are logged for debuggability.
Adds explicit instructions to complete all data fetching before starting
review and to cancel background tasks before returning structured output,
working around a claude-code-action issue where a late-completing
background task triggers a new conversation turn that overwrites the
structured JSON result.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Switch claude-review from reviewing the entire PR diff at once to
reviewing each commit individually via subagents. Each commit review
subagent receives the PR context, preceding commit diffs, and its own
commit diff, then returns comments tagged with the commit SHA. This
ensures review comments are attached to the correct commit via the
GitHub API rather than all pointing at HEAD.
Also add Bash(gh:*) to allowed tools so subagents can fetch per-commit
diffs via `gh api` without needing local git objects, and remove CI
analysis (needs to be delayed until CI finishes to be useful).
Co-developed-by: Claude Opus 4.6 <noreply@anthropic.com>
Address feedback from facebook/bpfilter#472:
- Fix setFailed error message counting file-level comments (without
line numbers) that are intentionally skipped, use inlineComments.length
instead of comments.length
- Fix double severity prefix in inline comments: the prompt told Claude
to prefix body with **must-fix**/etc but the post job also prepended
"Claude: ", producing "Claude: **must-fix**: ...". Now the prompt says
not to prefix and the post job adds "Claude **severity**: " using the
structured severity field
- Move error tracking instructions to a top-level section after all phases
so they apply to all runs, not just the first run
- Clarify that line is optional: use "should be" instead of "must be"
and document that omitting line still surfaces the comment in the
tracking comment summary
- Distinguish cancelled vs failed in tracking comment message
- Add side: "RIGHT" and subject_type: "line" to createReviewComment
per GitHub API recommendations
- Downgrade partial inline comment posting failures to warnings; only
fail the job when no comments at all could be posted
Co-developed-by: Claude Opus 4.6 <noreply@anthropic.com>
- Use github.paginate() for listComments to handle PRs with 100+ comments
- Make line optional in review schema to allow file-level comments
- Skip createReviewComment for comments without a line number
- Fix failed count to exclude skipped file-level comments
- Pass review result via env var instead of expression injection
- Use core.warning() instead of console.log() for JSON parse failures
- Fix MARKER insertion for single-line summaries that have no newline
- Require "@claude review" instead of just "@claude" to trigger
Co-developed-by: Claude <claude@anthropic.com>
This workflow runs on any comment to a github PR. 99% of the time the
workflow will be skipped yet it will still cancel any previous ongoing
workflows. Let's not cancel in progress workflow but instead queue the
workflow so we don't cancel in progress reviews any time a comment is
posted on a PR that is being reviewed.
The workflow is split into two jobs for least-privilege:
1. 'review' job — runs Claude with read-only permissions (contents: read,
id-token: write for AWS OIDC, actions: read). Claude produces a structured
JSON review via --json-schema with a 'comments' array and a 'summary'
string. Its tools are restricted to read-only operations (Read, LS, Grep,
Glob, Task, and various Bash prefixes for common read-only commands).
Claude also has access to CI MCP tools to analyze failed workflow runs.
2. 'post' job — only has pull-requests: write. Reads the structured JSON
output from the review job and posts inline comments individually (so
re-runs only add new comments). Maintains a tracking comment with a
<!-- claude-pr-review --> marker that is created on first run and updated
in-place on subsequent runs, preserving existing item order, wording,
and checkbox state. Posts a notification comment when the tracking
comment is updated or left unchanged.
Comment deduplication is handled by Claude in the prompt rather than in
the posting script, allowing for better semantic understanding of whether
two comments address the same issue.
The PR number is resolved via github.event.pull_request.number with a
fallback to github.event.issue.number for issue_comment events where
github.event.pull_request is not populated. The concurrency group uses
the same fallback.
Co-developed-by: Claude <claude@anthropic.com>
This seems to be what all the tools are standardizing on, except
claude (https://github.com/anthropics/claude-code/issues/6235) so
add a symlink from CLAUDE.md to AGENTS.md for now until they support
it as well.
I also had claude extend the instructions a bit.
Co-developed-by: Claude <claude@anthropic.com>
This will allow maintainers to mention claude in comments on issues
and prs to do stuff like review something or try to reproduce a bug
or other stuff. Let's give it a try and see whether we like it or
not.
Set `persist-credentials: false` for actions/checkout.
By default, using `actions/checkout` causes a credential to be persisted on
disk. Subsequent steps may accidentally publicly persist the credential, e.g.
by including it in a publicly accessible artifact via actions/upload-artifact.
However, even without this, persisting the credential on disk is non-ideal
unless actually needed.
Link: https://docs.zizmor.sh/audits/#artipacked
Daan De Meyer
Zbigniew Jędrzejewski-Szmek
dependabot[bot]
Jörg Behrmann
Dmitry V. Levin