Replace broken @claude workflow with Bedrock-based claude-review/mention (#10880)

Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
This commit is contained in:
Austin McKinley
2026-06-11 21:36:10 -07:00
committed by GitHub
parent bce0b51c2e
commit 6c04170734
4 changed files with 207 additions and 54 deletions
+63
View File
@@ -0,0 +1,63 @@
You are an experienced senior React Native engineer reviewing a pull
request in the Bluesky Social app — a cross-platform (iOS, Android, Web)
React Native + Expo application. Read the repo's CLAUDE.md before forming
an opinion; it describes the architecture, the ALF design system, and the
codebase conventions.
Your audience is other senior engineers. Write peer-to-peer, not
teacher-to-junior. Most PRs in this repo are fine; a review that says so
is a valid and common outcome.
Report a finding only if you can name a concrete scenario — specific
input, platform, navigation path, or operating condition — in which the
change causes incorrect behavior, a crash, a visual regression, a test
failure, a security issue, or a real regression visible to users. Style,
naming, and micro-optimizations are out of scope unless they introduce a
defect. Do not speculate that a change "might" break unrelated code
without pointing to the specific caller or code path. Do not repeat what
the diff does.
Where this codebase differs from a typical web app:
- Three platforms from one codebase. Web-only APIs (DOM, window),
native-only modules, and platform-specific files (.web.tsx, .ios.tsx,
.android.tsx) are common sources of single-platform breakage. When a
change touches shared code, consider all three targets.
- User-facing strings must go through Lingui (the `Trans` macro /
`useLingui`). Hardcoded English strings in UI are a finding. Do not
flag missing translations in catalog files — extraction and
compilation run in CI.
- New UI should use ALF (`#/alf`, `#/components`) rather than legacy
patterns (`#/view/com`, StyleSheet.create); flag newly written code
that adopts deprecated patterns, but don't flag pre-existing code the
PR merely touches.
- Server state lives in TanStack Query under src/state/queries. Watch
for cache-shape changes without corresponding invalidation updates,
and optimistic updates that can leave stale cache on failure.
- List rendering is performance-critical (the main feed). Changes to
feed items, FlatList usage, or anything in a hot render path deserve
scrutiny for re-render storms — unstable callback/object identities
passed to memoized children, missing memoization on expensive
computation.
- Moderation and content-filtering logic (labels, mutes, blocks,
hidden posts) is trust-and-safety-critical: a regression that shows
content that should be filtered is a blocking finding.
- Deep links, push-notification routing, and the navigation state
machine have platform-specific edge cases; changes there should name
the platforms they were verified on.
- The embed (bskyembed) and web deployment surfaces (bskyweb, link,
ogcard services in Go) ship separately from the app; changes there
have their own blast radius.
For each finding, state the scenario in one or two sentences, cite
file:line, and mark severity (blocking / non-blocking). If you are
uncertain but the potential impact is high (crash on startup, moderation
bypass, broken auth), include it and say what you are uncertain about.
Otherwise, prefer silence over guessing.
If there are no findings that meet this bar, say briefly that the PR
looks fine and note what you checked.
Post your review as a single top-level PR comment. Per-finding inline
comments are also welcome where they'd anchor a reader to the specific
lines involved.
+76
View File
@@ -0,0 +1,76 @@
name: claude-mention
# @claude mention handler on PR conversation comments, inline review
# comments, and review bodies. See the header comment in
# claude-review.yml for why this is self-contained rather than calling
# the org reusable workflows.
#
# Authorization note: this repo is public, and the load-bearing gate
# against drive-by commenters is INSIDE claude-code-action — the action
# verifies via the API that the triggering actor has write permission
# before doing anything (`allowed_non_write_users` and `allowed_bots`
# both default to deny). The `if:` below is a cheap pre-filter to avoid
# spinning up runners for the 99% of comments that don't mention
# @claude; it is not the security boundary.
on:
issue_comment:
types: [created]
pull_request_review_comment:
types: [created]
pull_request_review:
types: [submitted]
permissions:
contents: read
pull-requests: write
issues: write
actions: read
id-token: write
jobs:
mention:
# Skip comments/reviews that don't mention @claude, non-PR issue
# comments, and events from claude[bot] itself (its review prose
# often quotes "@claude" and must not retrigger the workflow).
if: >
(
(github.event_name == 'issue_comment' &&
github.event.issue.pull_request != null &&
contains(github.event.comment.body, '@claude')) ||
(github.event_name == 'pull_request_review_comment' &&
contains(github.event.comment.body, '@claude')) ||
(github.event_name == 'pull_request_review' &&
contains(github.event.review.body, '@claude'))
) &&
github.actor != 'claude[bot]'
runs-on: ubuntu-latest
timeout-minutes: 20
concurrency:
group: claude-mention-${{ github.repository }}-${{ github.event.issue.number || github.event.pull_request.number }}
cancel-in-progress: false
steps:
- name: Checkout repository
uses: actions/checkout@v6
with:
fetch-depth: 1
- name: Configure AWS credentials (OIDC)
uses: aws-actions/configure-aws-credentials@v6
with:
role-to-assume: ${{ secrets.AWS_BEDROCK_REVIEW_ROLE_ARN }}
aws-region: us-east-2
- name: Claude
uses: anthropics/claude-code-action@v1
with:
use_bedrock: 'true'
additional_permissions: |
actions: read
track_progress: true
claude_args: |
--model global.anthropic.claude-opus-4-8
--allowedTools "mcp__github_inline_comment__create_inline_comment,mcp__github_ci__get_ci_status,mcp__github_ci__download_job_log,Bash(gh pr comment:*),Bash(gh pr diff:*),Bash(gh pr view:*)"
+68
View File
@@ -0,0 +1,68 @@
name: claude-review
# Automatic Claude review on PR creation/update, via Bedrock (OIDC, no
# long-lived tokens). Self-contained: this intentionally uses upstream
# claude-code-action defaults rather than the org reusable workflows in
# bluesky-social/.github (which a public repo cannot call, and whose
# customizations added no value over upstream).
#
# Review guidance lives in .github/claude-review-prompt.md.
on:
pull_request:
types: [opened, synchronize, reopened, ready_for_review]
permissions:
contents: read
pull-requests: write
actions: read
id-token: write
jobs:
review:
# Internal PRs only. This repo is public: fork PRs are the bulk of
# community traffic and MUST NOT trigger reviews (no Bedrock spend on
# unvetted code, and fork PRs can't mint the OIDC token anyway —
# belt-and-braces with this explicit guard). Branch PRs can only be
# created by people with write access, i.e. org members.
# Bot-authored PRs (dependabot, changesets) are also skipped.
if: >
github.event.pull_request.draft == false &&
github.event.pull_request.head.repo.full_name == github.repository &&
github.event.pull_request.user.type != 'Bot'
runs-on: ubuntu-latest
timeout-minutes: 20
concurrency:
group: claude-review-${{ github.repository }}-${{ github.event.pull_request.number }}
cancel-in-progress: true
steps:
- name: Checkout repository
uses: actions/checkout@v6
with:
fetch-depth: 1
- name: Configure AWS credentials (OIDC)
uses: aws-actions/configure-aws-credentials@v6
with:
role-to-assume: ${{ secrets.AWS_BEDROCK_REVIEW_ROLE_ARN }}
aws-region: us-east-2
- name: Claude review
uses: anthropics/claude-code-action@v1
with:
use_bedrock: 'true'
additional_permissions: |
actions: read
track_progress: true
claude_args: |
--model global.anthropic.claude-opus-4-8
--allowedTools "mcp__github_inline_comment__create_inline_comment,mcp__github_ci__get_ci_status,mcp__github_ci__download_job_log,Bash(gh pr comment:*),Bash(gh pr diff:*),Bash(gh pr view:*)"
prompt: |
REPO: ${{ github.repository }}
PR NUMBER: ${{ github.event.pull_request.number }}
Read .github/claude-review-prompt.md in the checked-out repo
and review this pull request following its guidance.
-54
View File
@@ -1,54 +0,0 @@
name: Claude Code
on:
issue_comment:
types: [created]
pull_request_review_comment:
types: [created]
issues:
types: [opened, assigned]
pull_request_review:
types: [submitted]
jobs:
claude:
if: |
(github.event_name == 'issue_comment' && contains(github.event.comment.body, '@claude')) ||
(github.event_name == 'pull_request_review_comment' && contains(github.event.comment.body, '@claude')) ||
(github.event_name == 'pull_request_review' && contains(github.event.review.body, '@claude')) ||
(github.event_name == 'issues' && (contains(github.event.issue.body, '@claude') || contains(github.event.issue.title, '@claude')))
runs-on: ubuntu-latest
permissions:
contents: read
pull-requests: read
issues: read
id-token: write
actions: read # Required for Claude to read CI results on PRs
steps:
- name: Checkout repository
uses: actions/checkout@v5
with:
fetch-depth: 1
- name: Run Claude Code
id: claude
uses: anthropics/claude-code-action@v1
with:
claude_code_oauth_token: ${{ secrets.CLAUDE_CODE_OAUTH_TOKEN }}
# This is an optional setting that allows Claude to read CI results on PRs
additional_permissions: |
actions: read
# Optional: Give a custom prompt to Claude. If this is not specified, Claude will perform the instructions specified in the comment that tagged it.
# prompt: 'Update the pull request description to include a summary of changes.'
# Optional: Add claude_args to customize behavior and configuration
# See https://github.com/anthropics/claude-code-action/blob/main/docs/usage.md
# or https://code.claude.com/docs/en/cli-reference for available options
# claude_args: '--allowed-tools Bash(gh pr:*)'
# NOTE(sfn): we can add a custom system prompt here
claude_args: |
--model claude-opus-4-8