Hardening a manual Claude Code review workflow in GitHub Actions

Ahnii!

goformx uses a mixed-provider workflow: Codex writes some pull requests, Claude writes others, and each gets reviewed by the other provider on request. The Claude side runs on the official Claude Code Action, triggered by a maintainer commenting @claude review on a PR. That sounds simple to wire up. It wasn’t safe to wire up as-is, because the action’s default “tag mode” is built for implementation, not read-only review — three gaps in the default, five fixes to close them.

Why the Default Example Wasn’t Enough

The action ships an official progress-tracked review example that auto-triggers on every PR event and posts a tracked comment. It’s a good starting point, but three things about tag mode don’t fit a review-only job:

• Tag mode is implementation-capable by default. It’s designed for “@claude fix this,” not “@claude look but don’t touch.”
• GitHub Actions expression matching is case-insensitive. An if: condition comparing a comment body to @claude review will also match @Claude Review, @CLAUDE REVIEW, and anything else GitHub considers equal — not the exact trigger you intended.
• Tool permissions in the SDK accumulate, they don’t replace. Passing a restricted tool list doesn’t override the mode’s defaults; both apply, so an allow-list alone can’t get you to read-only.

The workflow adapts the official example at a pinned action revision, keeps the manual-only trigger and subscription auth, and layers on the following fixes.

Fix 1: Enforce the Exact Trigger

The outer if: on the job is a first, cheap filter, but because Actions expressions are case-insensitive it can’t be the whole check. A trusted actions/github-script preflight step re-validates everything in JavaScript before any model call happens:

const comment = context.payload.comment;
if (!context.payload.issue?.pull_request || comment?.body !== '@claude review' ||
    comment.user?.login !== 'jonesrussell' || comment.user?.type !== 'User' ||
    context.actor !== 'jonesrussell') {
  core.setFailed('An exact human maintainer review request is required.');
  return;
}

This rejects extra text, trailing newlines, wrong casing, bot comments, and anyone who isn’t the maintainer — all before the job spends a single turn talking to Claude. It also checks that the PR is still open, and records head_sha/base_sha for later verification.

Fix 2: Bound the Diff, Fail Closed

Tag mode’s native changed-file context is a file list, not a diff — Claude would have to reconstruct the actual changes from scratch. Instead, a setup step downloads the real diff to runner-temp storage with a hard size cap:

const {data: diff} = await github.rest.pulls.get({
  ...context.repo, pull_number: context.issue.number, mediaType: {format: 'diff'}
});
if (typeof diff !== 'string' || !diff.trim() || Buffer.byteLength(diff, 'utf8') > 1048576) {
  throw new Error('Review diff is missing, invalid or exceeds the 1 MiB review bound.');
}
require('node:fs').writeFileSync(diffPath, diff, {mode: 0o600, flag: 'wx'});

Missing, empty, or oversized diffs fail the job before Claude ever runs. The cap is 1 MiB; there’s no silent fallback to a partial or truncated review.

Fix 3: Deny Tools Instead of Just Allowing Them

The SDK’s argument parser accumulates allowedTools rather than replacing the mode defaults, so the workflow can’t rely on an allow-list alone. It sets the permission mode explicitly, restricts built-in tools to Read, Glob, Grep, and layers an explicit deny list on top:

claude_args: >-
  --max-turns 16
  --permission-mode default
  --setting-sources user
  --strict-mcp-config
  --tools "Read,Glob,Grep"
  --disallowedTools "Bash,Edit,Write,NotebookEdit,Agent,Task,mcp__github_ci__*,mcp__github_file_ops__*"

Some of those denials cover capabilities that aren’t even enabled at the pinned action revision. That’s deliberate — defense against future defaults changing underneath the pin.

Fix 4: Keep Git Metadata Off Limits

Tag mode checks out the PR head and stores its short-lived job token in .git/config. A review agent with unrestricted Read could read that token straight out of the checkout. The workflow denies it explicitly in the action’s settings:

{"disableAllHooks":true,"permissions":{"deny":["Read(./.git)","Read(./.git/**)"]}}

Denying reads to .git also blocks Grep/Glob from searching that directory, and hooks are disabled entirely so only user settings and the action’s own MCP configuration load.

Fix 5: Catch Stale or Unverified Reviews

A successful job run isn’t proof the review still applies. If the PR gets pushed to while Claude is reviewing it, the findings could describe a revision that no longer exists. A final step re-checks the PR’s current head_sha/base_sha against what the preflight recorded, and confirms the local checkout matches too:

const checkedHead = require('node:child_process').execFileSync(
  'git', ['rev-parse', 'HEAD'], {encoding: 'utf8'}
).trim();
if (pr.state === 'open' && pr.head.sha === process.env.REVIEW_HEAD &&
    pr.base.sha === process.env.REVIEW_BASE && checkedHead === process.env.REVIEW_HEAD) return;

If any of those checks fail, the workflow posts a comment marking the review STALE or UNVERIFIED with a link to the run, and fails the job. That comment matters as much as the review itself — without it, a stale finding just sits there looking authoritative.

What Changed, at a Glance

Testing Without Spending a Turn

The workflow’s own tests don’t call the model at all. A separate CI job runs mocked-response tests against the preflight and revision-check logic, plus actionlint against both workflow files:

node --test .github/scripts/manual-review.test.cjs
go run github.com/rhysd/actionlint/cmd/actionlint@v1.7.12 -shellcheck= -pyflakes= \
  .github/workflows/manual-claude-review.yml .github/workflows/manual-review-tests.yml

That proves the guardrails and workflow syntax are correct. It doesn’t prove the provider’s runtime tool enforcement — a live @claude review request against an authorized PR is still the only way to confirm end to end that the final comment gets posted and failures are reported correctly.

Budget Stays Bounded Too

The job itself is capped three ways: 16 turns, an 8-minute limit on the review step, and a 20-minute ceiling on the whole job. Auth comes from a CLAUDE_CODE_OAUTH_TOKEN repository secret tied to a Claude Max subscription. There’s no API-key fallback, so a misconfigured secret fails the preflight instead of silently billing somewhere else.

None of this is exotic. It’s the same instinct as any other CI hardening: don’t trust a string comparison you haven’t tested for case sensitivity, don’t hand out more filesystem access than the job needs, and don’t let a slow job’s output outlive the code it described.

Baamaapii