name: Claude Code Review # Runs on PRs INTO this repo. We use pull_request_target (not pull_request) so # that PRs from a fork can access CLAUDE_CODE_OAUTH_TOKEN — GitHub withholds # secrets from `pull_request` runs triggered by forks, which is why the plain # `pull_request` version never worked for fork PRs. # # SECURITY: pull_request_target runs in the BASE repo with secrets and a # write-capable token. The job is gated to PRs from the trusted `jnasbyupgrade` # fork only — an arbitrary external fork can never trigger this secret-bearing # job. The workflow file always comes from the base branch (master), so a PR # cannot modify the reviewer that runs on it. We never check out the fork's PR # head ourselves here: anthropics/claude-code-action's own internal checkout # logic (setupBranch() in src/github/operations/branch.ts) already fetches the # PR branch via `git fetch origin pull//head`, which requires `origin` to be # the BASE repo -- checking out the fork directly instead (as a prior version # of this file did) points `origin` at the fork, which has no such ref, and # breaks that fetch with "couldn't find remote ref pull//head". on: pull_request_target: # labeled: lets adding the claude-debug label (see the "Check for # claude-debug label" step below) kick off a fresh run by itself, with no # push/re-run needed. Scoped in the job's `if:` below to only actually # proceed when the label added IS claude-debug -- otherwise every # unrelated label added to a PR would trigger another paid review. types: [opened, synchronize, reopened, ready_for_review, labeled] concurrency: # Concurrency cancellation resolves when a run is admitted, before the # job's `if:` is evaluated -- a job's `if:` can only no-op itself, it can't # un-cancel whatever the run already displaced. So only a labeled event # whose label is NOT claude-debug gets its own per-label group here, # keeping it from ever colliding with (and cancelling) the real review's # group. labeled+claude-debug deliberately keeps the plain group, since # it's meant to supersede an in-progress review. group: claude-review-${{ github.event.pull_request.number }}${{ (github.event.action == 'labeled' && github.event.label.name != 'claude-debug') && format('-{0}', github.event.label.name) || '' }} cancel-in-progress: true jobs: claude-review: # Trusted fork only, and skip drafts (don't spend API/CI on unfinished PRs). # To add more trusted owners, extend the head-owner check. # # SECURITY-CRITICAL: this owner check is what makes it safe to run this # pull_request_target job -- which holds base-repo secrets/token -- on # every fork PR unattended. Do not remove or loosen this condition (e.g. # drop the owner check, or allow non-owner forks) without re-evaluating # whether this job should keep running on arbitrary forks. if: >- github.event.pull_request.draft == false && github.event.pull_request.head.repo.owner.login == 'jnasbyupgrade' && (github.event.action != 'labeled' || github.event.label.name == 'claude-debug') runs-on: ubuntu-latest timeout-minutes: 60 permissions: contents: read pull-requests: write # post the review comments checks: read # read sibling check-runs for the cost gate # GitHub has no narrower "cache write" scope -- actions: write is the # only permission that lets a step save an Actions cache entry (it also # grants cancelling/deleting workflow runs and managing artifacts, which # this job doesn't use). Without it, claude-code-review's own setup step # can never write a cache, only ever miss. Granted here on top of the # existing trusted-fork-owner gate below, not instead of it. actions: write steps: # DEBUG MODE: add the "claude-debug" label to a PR to (a) skip the cost # gate below entirely -- a debug session shouldn't wait 5-20+ min per # iteration on sibling CI -- and (b) get show_full_output: true on the # Run Claude Code Review step, dumping the full raw Claude Code JSON # transcript (including tool results -- see that input's own WARNING # below) to the job log. This is how you'd catch something like a # silently-swallowed `--comment` flag (see that step's other comment). # Queried live via `gh pr view`, not the static event payload, so # adding the label and clicking "Re-run jobs" on an existing run picks # it up without needing a new push. - name: Check for claude-debug label id: debug env: GH_TOKEN: ${{ secrets.GITHUB_TOKEN }} REPO: ${{ github.repository }} PR: ${{ github.event.pull_request.number }} run: | enabled=$(gh pr view "$PR" --repo "$REPO" --json labels \ --jq 'any(.labels[]; .name == "claude-debug")' 2>/dev/null) || enabled=false echo "enabled=$enabled" >> "$GITHUB_OUTPUT" echo "claude-debug label present: $enabled" # COST GATE: the paid Claude review is the last thing to run. Wait for the # PR head's OTHER check-runs to finish and only proceed if they are clean. # If any sibling check failed we skip the review to avoid spending money # reviewing a PR that is already known-broken. Uniform across all repos: # it discovers sibling checks dynamically (no per-repo workflow names). # - decision=run : all sibling checks completed with a good conclusion, # OR no sibling checks exist after a short grace window # (nothing to gate on), OR the poll timed out is treated # as skip (see below). # - decision=skip : at least one sibling check failed/cancelled/etc, or # we timed out waiting for still-pending checks. # We exclude this workflow's own check-run (job name `claude-review`) so the # gate never waits on or fails because of itself. - name: Wait for CI; skip the paid review if any check failed id: gate if: steps.debug.outputs.enabled != 'true' env: GH_TOKEN: ${{ secrets.GITHUB_TOKEN }} REPO: ${{ github.repository }} SHA: ${{ github.event.pull_request.head.sha }} run: | decision=skip for i in $(seq 1 72); do # ~24 min max json=$(gh api "repos/$REPO/commits/$SHA/check-runs" --paginate \ --jq '[.check_runs[] | select(.name != "claude-review")]' 2>/dev/null) || json='' [ -z "$json" ] && { sleep 20; continue; } total=$(jq 'length' <<<"$json") if [ "$total" -eq 0 ]; then [ "$i" -ge 9 ] && { decision=run; break; } # ~3 min grace: nothing to gate on sleep 20; continue fi pending=$(jq '[.[]|select(.status!="completed")]|length' <<<"$json") if [ "$pending" -eq 0 ]; then bad=$(jq '[.[]|select((.conclusion//"")|test("^(failure|cancelled|timed_out|action_required|stale)$"))]|length' <<<"$json") [ "$bad" -eq 0 ] && decision=run || decision=skip break fi sleep 20 done echo "decision=$decision" >> "$GITHUB_OUTPUT" echo "gate decision: $decision" - name: Check out base branch if: steps.debug.outputs.enabled == 'true' || steps.gate.outputs.decision == 'run' # Intentionally tracks the major-version tag (not a pinned SHA) so # upstream fixes are picked up automatically. # # No `repository:`/`ref:` here on purpose — this checks out the base # branch (master), never the fork's PR head. See the SECURITY note # above; anthropics/claude-code-action fetches the actual PR head # itself afterward via the base repo's `refs/pull//head` ref. uses: actions/checkout@v7 with: fetch-depth: 1 persist-credentials: false - name: Run Claude Code Review if: steps.debug.outputs.enabled == 'true' || steps.gate.outputs.decision == 'run' uses: anthropics/claude-code-action@v1 with: claude_code_oauth_token: ${{ secrets.CLAUDE_CODE_OAUTH_TOKEN }} # See the "Check for claude-debug label" step above -- WARNING (from # this input's own description): outputs ALL Claude messages # including tool execution results, which may contain secrets, and # these logs are publicly visible in GitHub Actions. show_full_output: ${{ steps.debug.outputs.enabled == 'true' }} # Provide github_token so the action uses it directly for GitHub API # calls instead of the OIDC->GitHub-App-token exchange, which 401s under # pull_request_target. GITHUB_TOKEN is repo/workflow-scoped (independent # of the actor's role) and has pull-requests: write here. github_token: ${{ secrets.GITHUB_TOKEN }} # A `prompt:` input puts the action in "automation mode", which by # default posts nothing until the whole run finishes -- there's no # visibility into a review that runs long. track_progress forces a # tracking PR comment with a live checklist that updates as Claude # works, so a slow run is visible instead of silent. track_progress: true # NOTE: plugin_marketplaces can't be pinned — it tracks the # marketplace repo's default branch (upstream anthropics/claude-code). plugin_marketplaces: 'https://github.com/anthropics/claude-code.git' plugins: 'code-review@claude-code-plugins' # --comment is required: without it, the code-review plugin only # prints its findings to the job log and never posts anything to # the PR (confirmed by capturing the hidden SDK transcript on a # canary PR in pgxntool-test: the review correctly found an # injected bug but ended with "No `--comment` argument was # provided, so no GitHub comments were posted"). Every review run # before this fix has been silently invisible on GitHub. prompt: '/code-review:code-review ${{ github.repository }}/pull/${{ github.event.pull_request.number }} --comment' # A direct `prompt:` (no @claude mention) runs the action in "agent # mode". In that mode, claude-code-action only installs the # github_inline_comment MCP server if it sees # mcp__github_inline_comment__create_inline_comment listed in an # --allowedTools flag inside claude_args (src/modes/agent/parse-tools.ts) -- # it does NOT look at the code-review plugin's own `allowed-tools` # frontmatter to decide that. Without this, the MCP server never # starts, the tool genuinely doesn't exist in the session, and the # plugin silently falls back to one consolidated PR comment instead # of real inline line comments. claude_args: '--allowedTools mcp__github_inline_comment__create_inline_comment'