Skip to content

ci: add Claude review for pull requests #1

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
kanywst merged 1 commit into from
Oct 9, 2026
Merged

ci: add Claude review for pull requests #1

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
200 changes: 200 additions & 0 deletions .github/workflows/claude-review.yml
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,200 @@
name: Claude review

# CI (if this repository has any) answers "does it build and do the checks
# pass". This asks the questions no script here can. It reviews, it does not
# gate: nothing here is a required check.
#
# The action refuses to run when this file differs from the copy on the default
# branch, so a pull request cannot rewrite its own reviewer. The cost is that
# any change to this file, including the one that added it, goes unreviewed by
# Claude until it is merged.
#
# "It does not gate" is about findings, not about delivery. A run that reviews
# and then cannot post what it found is indistinguishable from a clean diff,
# and the check is green either way. The last step here fails when that
# happens, for the same reason the token check fails rather than reviewing
# nothing quietly.

on:
pull_request:
types: [opened, synchronize, reopened, ready_for_review]

permissions: {}

# One review per pull request. A rapid re-push cancels the run it superseded
# rather than paying for a review of a diff that no longer exists.
concurrency:
group: claude-review-${{ github.event.pull_request.number }}
cancel-in-progress: true

jobs:
review:
name: review the diff
# A pull request from a fork gets no secrets, so the job would fail on
# authentication rather than say anything useful. Drafts are skipped until
# they are marked ready, which is what the ready_for_review trigger is for.
#
# Dependabot is skipped for the same reason as a fork: a run it triggers
# reads Dependabot secrets, not Actions secrets, so the token is empty no
# matter what is configured here. A version bump is also the diff this
# review has least to say about.
if: >-
github.event.pull_request.head.repo.full_name == github.repository &&
github.event.pull_request.draft == false &&
github.event.pull_request.user.login != 'dependabot[bot]'
runs-on: ubuntu-latest
timeout-minutes: 20
permissions:
contents: read
pull-requests: write # post the summary and the inline findings
issues: read # the delivery check reads conversation comments, served from the issues endpoint
id-token: write # required by the action's default GitHub App authentication
steps:
- uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
with:
# No job in this repository pushes, so the checkout token has no
# reason to outlive the step that used it.
persist-credentials: false
fetch-depth: 0

# Without the secret the action exits successfully having reviewed
# nothing, which is worse than failing: the check goes green and the
# pull request looks reviewed. Fail here instead, where the reason is
# legible.
- name: Require the token
env:
CLAUDE_CODE_OAUTH_TOKEN: ${{ secrets.CLAUDE_CODE_OAUTH_TOKEN }}
run: |
if [ -z "$CLAUDE_CODE_OAUTH_TOKEN" ]; then
echo "::error::CLAUDE_CODE_OAUTH_TOKEN is not set. Generate one with 'claude setup-token', then: gh secret set CLAUDE_CODE_OAUTH_TOKEN"
exit 1
fi

# Anything the review posts has to be newer than this, or it is a
# leftover from an earlier push and says nothing about this diff. Two
# minutes of slack absorbs clock skew between the runner and the API.
- name: Mark the start of this review
id: started
run: echo "at=$(date -u -d '-2 minutes' +%Y-%m-%dT%H:%M:%SZ)" >> "$GITHUB_OUTPUT"

- uses: anthropics/claude-code-action@ed670b4cf9de2a5a570d130d2f6197b9e543cd64 # v1
id: claude
with:
# A subscription token, not an API key: runs against the Claude plan
# instead of API credits. Generate it with `claude setup-token`.
claude_code_oauth_token: ${{ secrets.CLAUDE_CODE_OAUTH_TOKEN }}
prompt: |
REPO: ${{ github.repository }}
PR: ${{ github.event.pull_request.number }}

This repository (0-draft/blackjack): Blackjack played by whoever turns up. Every decision is a GitHub issue and the table is the README.

Before reviewing, read whatever this repository says about itself: CLAUDE.md,
AGENTS.md, CONTRIBUTING.md, README.md (and README.ja.md if present), and the
workflows under .github/workflows. Those define what the project promises and
what CI already checks. Do not repeat what CI already does.

Look for what only a reader can catch:

1. Correctness of the change against the project's own stated rules and
invariants. A change that contradicts CLAUDE.md, CONTRIBUTING.md or the
README is a finding even when every check is green.

2. Factual claims. This org publishes explainers, deep dives and reference
material. A new or edited statement about a protocol, a cloud service, a
price, a date, a version or an attribution needs to be true, not merely
plausible. Flag anything that reads as recalled rather than sourced.

3. Content parity. Where content exists in more than one language (EN/JA),
both must say the same thing. A string added in one language only, or a
translation that drifted from its source, is a finding. Japanese should be
plain technical prose, not colloquial or translationese.

4. Injection and trust boundaries. Anything driven by GitHub issues, comments
or other user input (games, guest books, ingest jobs) must not let that
input reach a shell, a GitHub expression in a run block, or rendered HTML
unescaped. Workflows should keep least-privilege permissions and pinned
actions.

5. Claims that outrun the code. A README line, doc comment or UI text that
promises behaviour the code does not have is a defect, not a nit.

Do not sign anything you post. No "Generated with" footer, no co-authored-by
line, no self-attribution of any kind, in any comment or commit. The review
stands on the findings, and a signature is noise on every one of them.

Post findings as inline comments on the lines they concern. Use one
top-level comment for the summary. Be specific and cite
`file:line`. If the diff is clean against all of the above, say so
in one line rather than manufacturing nits.

Write the summary to a file and post it with
`gh pr comment --body-file <path>`. A body built inline from a
heredoc or a command substitution does not match the allowlist and
is refused, which loses the whole review.

claude_args: |
--allowedTools "mcp__github_inline_comment__create_inline_comment,Write,Bash(gh pr comment:*),Bash(gh pr diff:*),Bash(gh pr view:*)"

# The gate below keys off the action's conclusion output. Print what it
# was, so a run that gates when it should not - or does not when it
# should - can be diagnosed from the log rather than by guessing.
- name: Report the review outcome
if: always()
env:
CONCLUSION: ${{ steps.claude.outputs.conclusion }}
EXECUTION_FILE: ${{ steps.claude.outputs.execution_file }}
run: |
echo "action conclusion: '${CONCLUSION}'"
echo "execution file: '${EXECUTION_FILE}'"

# A review that runs, is refused a tool call on the way to posting, and
# exits green leaves the pull request unreviewed behind a passing check.
# Write is allowed above so a long summary can go to a file and out
# through `gh pr comment --body-file`, which is the shape the allowlist
# matches; this catches it when something is refused anyway.
#
# Silence on its own is not the signal. A follow-up push whose review had
# nothing to add to what is already on the pull request posts nothing and
# denies nothing, and that is fine. The failure is silence *with* a
# refusal behind it. An unreadable denial count is treated as a refusal,
# because the whole point is not to call a pull request reviewed on the
# strength of a green check.
- name: Fail on a review that was not delivered
if: always() && steps.claude.outputs.conclusion == 'success'
env:
GH_TOKEN: ${{ github.token }}
REPO: ${{ github.repository }}
PR: ${{ github.event.pull_request.number }}
SINCE: ${{ steps.started.outputs.at }}
EXECUTION_FILE: ${{ steps.claude.outputs.execution_file }}
# The action's default bot_name, which this workflow does not override.
BOT: claude[bot]
run: |
set -euo pipefail
fresh() {
gh api --paginate "$1" --jq \
"[.[] | select(.user.login == \"$BOT\") | select((.$2 // \"\") >= \"$SINCE\")] | length" |
awk '{ n += $1 } END { print n + 0 }'
}
summary=$(fresh "repos/$REPO/issues/$PR/comments" updated_at)
inline=$(fresh "repos/$REPO/pulls/$PR/comments" updated_at)
reviews=$(fresh "repos/$REPO/pulls/$PR/reviews" submitted_at)
total=$((summary + inline + reviews))
echo "posted since $SINCE: $summary summary, $inline inline, $reviews review(s)"
if [ "$total" -gt 0 ]; then
exit 0
fi

denials=unknown
if [ -n "${EXECUTION_FILE:-}" ] && [ -f "$EXECUTION_FILE" ]; then
denials=$(jq -r '[.. | objects | select(.type? == "result") | .permission_denials_count?]
| map(select(. != null)) | last // 0' "$EXECUTION_FILE" 2>/dev/null || echo unknown)
fi
echo "denied tool calls: $denials"
if [ "$denials" = "0" ]; then
echo "::notice::The review posted nothing and nothing was refused, so it had nothing to add to what is already on this pull request."
exit 0
fi
echo "::error::The review ran, was refused a tool call ($denials), and posted nothing, so this pull request is unreviewed despite a green check. Widen --allowedTools above."
exit 1
Loading