🤖 Codex Guard

Docs site: https://akimiya-z.github.io/codex-guard

An automatic quality gate for AI-generated pull requests. Stop TODO leftovers,
leaked secrets, sloppy commits and red CI from reaching main — with zero review
bandwidth spent on the obvious stuff.

Works with OpenAI Codex (cloud & CLI), Claude Code, Copilot, and any
other agent that opens PRs against your repo.

🐕 Dogfooding: this repo gates its own AI-generated PRs. See a real failing
example on PR #6 and the
workflow behind it in .github/workflows/codex-guard.yml.

Why

AI coding agents are great at writing code and terrible at cleaning up after
themselves. In practice, agent-written PRs tend to arrive with the same handful
of problems:

• // TODO: handle this comments that were never meant to stay
• Hardcoded API keys and connection strings copied from a chat transcript
• Commit trails like WIP, fix stuff, more changes — squashed versions of
  a messy session
• A green-looking PR that actually has failing CI on the head commit

You shouldn't need a human reviewer to catch those every single time. Codex
Guard checks the boring, deterministic things automatically, and only on PRs
that look AI-generated — so human attention goes where it matters.

How it works

Codex Guard runs on pull_request, figures out whether the PR looks
agent-generated (via label, branch prefix, or title), and then:

Each finding is posted as a GitHub check-run annotation at the exact file and
line, plus a human-readable summary comment on the PR.

Honest content-scan coverage

GitHub sometimes omits the textual patch for binary or very large files. Codex
Guard now distinguishes no findings from not scanned: every report shows
the number of eligible changed files whose patches were actually inspected. A
missing patch produces a neutral, non-blocking coverage warning with the
affected paths; the JSON and Action outputs carry the same information. The
warning also appears when GitHub's pull-request files API reaches its documented
3,000-file limit.

The output below comes from a real run on this repo
(PR #6) — a deliberate test
PR from a codex/ branch that left a TODO, credential-shaped fixtures and two
sloppy commits. Secret-shaped values are redacted here just as they are in
current reports:

## 🤖 Codex Guard

❌ **Checks failed — review the findings before merging.**

| Check | Result |
| --- | --- |
| TODO / FIXME scan | ⚠️ 2 |
| Secret scan | ⚠️ 3 |
| Commit hygiene | ⚠️ 2 |
| CI status | ✅ |

**Unfinished work**
- `scripts/sync.js:4` — `FIXME`: const aws = 'AKIA...MPLE'; // FIXME: move this to a secret store
- `scripts/sync.js:2` — `TODO`: // TODO: wire up real retry with exponential backoff.
**Potential leaked secrets**
- `scripts/sync.js:4` — AWS Access Key ID `AKIA...MPLE`
- `scripts/sync.js:5` — Connection string `post...prod`
- `scripts/sync.js:7` — OpenAI API Key `sk-p...6789`
**Commit hygiene**
- `7c84ae1` — _WIP stuff_ (by Akimiya-z)
- `1affcf8` — _tmp_ (by Akimiya-z)

> Detected as an AI-generated PR (branch prefix "codex/").

Quick start

From the root of your Git repository:

npx --yes codex-guard init
git add .github/workflows/codex-guard.yml
git commit -m "ci: add Codex Guard"

The installer starts in observe mode: findings are annotated, but they do
not fail the workflow while you tune the policy. Three setup presets keep the
rollout explicit:

Prefer to add it by hand? The generated workflow is:

# Generated by codex-guard init
name: Codex Guard
on:
  pull_request:

permissions:
  contents: read
  statuses: read
  pull-requests: write
  checks: write

jobs:
  codex-guard:
    runs-on: ubuntu-latest
    steps:
      - uses: Akimiya-z/codex-guard@v1
        with:
          preset: 'observe'

That's it. Codex Guard now reports on matching PRs without blocking them.

Upgrading an existing workflow? Add statuses: read to its permissions
block. checks: write already includes read access for check runs. Without
status access, Codex Guard reports incomplete CI visibility and returns a
neutral result instead of claiming every check is green.

One-click from the GitHub Actions Marketplace.

Turn on enforcement

After a few representative PRs, choose how strongly to enforce:

- uses: Akimiya-z/codex-guard@v1
  with:
    preset: 'balanced' # or 'strict'

balanced warns on unfinished markers while blocking secrets, commit hygiene,
and red CI. strict blocks every default finding. For a custom mix, use
fail-on and the individual inputs. Then require the status check under
Settings → Branches → Require status checks → Codex Guard. A blocking
finding will prevent the PR from merging until it is resolved (or the PR is
marked with an ignore label — see "Opting out").

Detecting agent PRs

By default Codex Guard only gates PRs it believes were written by an agent,
so human-authored PRs are never slowed down:

• Label matches one of codex-generated, agentic, ai-generated
• Branch starts with codex/, copilot/, claude-auto, gh-codex/
• Title contains Generated by Codex, Generated by Claude, Generated by Copilot

All of these are configurable — or set gate-agents-only: false to gate every PR.

Opting out of a specific PR

Add a label named codex-guard-ignore (configurable) to a PR and Codex Guard
will pass it without running checks. Useful when a human has already reviewed
and accepted the changes.

Inputs

CI results are never guessed

Codex Guard reads both commit statuses and check runs, follows every results
page up to GitHub's 3,000-result safety cap, and treats every completed
non-success conclusion as a failure. Pending checks produce a neutral result,
not a premature green check. If one GitHub CI API is unavailable, the report is
neutral and explicitly says visibility is incomplete; if both are unavailable,
the CI check fails closed. Previous Codex Guard check runs are ignored so a
rerun cannot inherit its own old failure.

Configuration file

Every input above can be overridden per-repo with a .github/codex-guard.yml
file on the default branch — so agents can't just loosen the policy in their PR.
Unknown keys are ignored (typo tolerant).

# .github/codex-guard.yml
preset: balanced
gate-agents-only: true
agent-labels:
  - codex-generated
  - agentic
todo-patterns:
  - TODO
  - FIXME
  - XXX
secret-exclude-paths:
  - README.md
  - docs/
comment-mode: replace
request-changes: true

A copy-paste template lives at examples/codex-guard.yml.
Policy resolution is deterministic: the workflow preset supplies a baseline,
the repository preset replaces that baseline, and individual repository keys
win last. This keeps existing workflows compatible because an empty workflow
preset retains the legacy defaults.

Diagnose an installation

Run the read-only local doctor after installing or changing policy:

npx --yes codex-guard doctor
npx --yes codex-guard doctor --json  # machine-readable output

It checks the pull-request trigger, Action step, effective workflow/job
permissions, setup preset, policy YAML, unknown keys, boolean values,
comment-mode, and fail-on. It exits 1 for configuration errors and 0
when only warnings or passes remain. This is a local-file diagnostic; GitHub
branch protection and organization policy still live in repository settings.

Local dry-run (CLI)

The CLI installs the workflow and previews the same checks locally before CI —
no token or waiting:

# install the GitHub workflow (safe observe mode by default)
npx --yes codex-guard init

# scan a diff from npm — no install needed (v1.7.0+)
npx --yes codex-guard --diff <(git diff origin/main)

# from this repo's source
git diff origin/main > /tmp/patch.diff
node src/cli.js --diff /tmp/patch.diff

# or straight against a ref (bare `--git` scans tracked + untracked changes)
npx --yes codex-guard --git --commits
node src/cli.js --git origin/main --commits

PowerShell does not support Bash process substitution (<(...)). Save the
diff to a temporary UTF-8 file instead:

$patch = New-TemporaryFile
try {
  git diff origin/main | Set-Content -LiteralPath $patch -Encoding utf8
  npx --yes codex-guard --diff $patch
} finally {
  Remove-Item -LiteralPath $patch
}

Exit codes: 0 = no blocking findings, 1 = blocking findings, 2 = usage
error. --json prints the raw report (same shape as the findings-json
output) for scripts. In --git mode the CLI automatically loads
.github/codex-guard.yml, so local content checks use the same preset,
patterns, exclusions, and blocking rules as the Action. Use --preset or
individual flags for an explicit local override, --config <path> for another
policy file, or --no-config to bypass repository policy. Explicit CLI flags
win over the loaded file. Commit hygiene is checked only when --commits is
passed. Git-ignored files stay ignored; ordinary untracked files are treated as
all-added patches, so a new file is scanned before its first git add. Binary
or untracked files larger than 8 MiB are listed explicitly as unscanned instead
of being silently treated as clean.

Agent skill (pre-submit self-check)

Same checks, one level further up: packed as a skill, so Codex or Claude
Code run them themselves before opening a PR — the CI gate then never sees a
dirty diff that the agent didn't already spot locally. This repo ships the
skill in skills/codex-guard/SKILL.md; install it with:

bash skills/install.sh              # installs for Codex and Claude Code
# or copy skills/codex-guard/ into ~/.codex/skills/ or ~/.claude/skills/

The skill tells the agent to run node src/cli.js --git --commits before
submitting, fix TODO/secret/commit findings, and only open the PR once the
checks pass (or the exception is documented in the description).

Codex plugin

The same skill is packaged as an official Codex plugin (skill-only form,
schema follows openai/plugins): plugins/codex-guard/.codex-plugin/plugin.json
plus a marketplace manifest at .agents/plugins/marketplace.json. Codex users
can add this repository as a plugin marketplace from the Codex app or CLI —
install steps are in OpenAI's plugin docs:
https://developers.openai.com/codex/plugins/build.

plugins/codex-guard/.codex-plugin/plugin.json
plugins/codex-guard/skills/codex-guard/SKILL.md
.agents/plugins/marketplace.json

Once installed, Codex runs the pre-submit checks itself before opening or
updating a PR — same behavior as the standalone skill, one packaging step
closer to discoverable on GitHub (see the codex-plugin topic).

DeepSeek Harness plugin (dsh)

A DSH bundle lives in dsh/: it declares a dsh.bundle manifest and registers
a codex_guard tool that DeepSeek Harness agents can call to pre-flight an
agent-authored diff before it becomes a PR.

dsh/
├── package.json        # dsh.bundle manifest (name: dsh-codex-guard)
├── cordis.patch.yml    # the layer a profile applies
└── index.js            # registers the codex_guard tool (dsh-tools API)

The tool shells out to the published CLI (npx --yes codex-guard --git) in the
current working directory, so it gets the same deterministic report the CI gate
uses. Node API compatibility is tested against the real @deepseek-ai/dsh-tools
registry packages (see test/dsh.test.js), including an end-to-end run in a
throwaway git repo.

Outputs

Sweeping existing PRs

Adopting Codex Guard doesn't have to be retrospective — run a sweep to
inspect every currently-open agent-generated PR in one go:

on:
  workflow_dispatch:
# … uses: Akimiya-z/codex-guard@v1 with: { sweep: 'true' }

The report is written to the run summary and sweep-report output (per-PR
numbers in sweep-json). No per-PR comments or check-runs are posted.
Use sweep-label to limit the run to PRs carrying a particular label.
Copy-paste at examples/sweep.yml.

Examples

Gate everything, and skip secrets scanning on docs:

steps:
  - uses: Akimiya-z/codex-guard@v1
    with:
      gate-agents-only: 'false'
      secret-exclude-paths: 'README.md,docs/'

Observe first, enforce later:

steps:
  - uses: Akimiya-z/codex-guard@v1
    with:
      preset: 'observe'

Integrate with a tool that adds the agent label automatically:

If your agent has a GitHub App or a bot that labels PRs, current Codex Guard will
respect whatever label you configure — and fails open (passes) when a PR has no
signal at all.

Keep gating agent PRs from forks:

Use pull_request_target so the check runs with your repo's write token. The
action reads the fork PR's files, commits and head SHA from the event payload —
no extra config. Copy-paste at examples/forks.yml.

How it compares

Community

• Ask & discuss — questions, setups, and roadmap voting live in
  GitHub Discussions;
  bug reports and feature requests go through the
  issue templates.
• Good first issues are tagged good first issue
  and help wanted — the contribution guide is in
  CONTRIBUTING.md.
• Roadmap is feedback-driven: the fastest way to influence the next release
  is to open a feature request and vote on it in Discussions.
• Every external contribution gets called out in the release notes. Thank you —
  this project lives on its users' feedback.

Development

npm install
npm test          # node:test — no framework required
npm run check     # syntax-check every source file
npm run build     # rebuild the checked-in Action bundle

Consult CONTRIBUTING.md before opening a PR — it covers the release process
(rebuild dist/, tag v1.x). See CHANGELOG.md for
release history. Copy-paste workflows live in examples/, and docs/ explains
how to record the demo GIF and smoke-test locally.

Limitations

• Secret detection is regex-based on purpose — it catches the obvious
  mistakes cheaply, but it is not a replacement for a real secret scanner
  (use gitleaks or
  zizmor in addition).
• The TODO scan only sees lines added by this PR — it won't nag you about
  pre-existing comments.
• GitHub may omit textual patches for binary or very large files. Those files
  cannot be content-scanned. Codex Guard reports their paths and uses a neutral
  check conclusion instead of silently claiming full coverage, but a
  repository-wide secret scanner should remain the authoritative control.
• GitHub's pull-request files endpoint
  returns at most 3,000 files. Reaching that limit is reported as incomplete
  coverage because additional files may be absent.
• Commit statuses and check runs also have a conservative 3,000-result safety
  cap per API. Reaching it is reported as incomplete CI visibility rather than
  silently ignoring later results.
• The local CLI scans ordinary untracked files, but reports binary files and
  untracked files larger than 8 MiB as unscanned. Git-ignored files are omitted
  by design.
• Agent detection is intentionally heuristic. Set gate-agents-only: false
  when coverage matters more than distinguishing human and agent PRs.
• Runs on pull_request events; for fork contributions use
  pull_request_target (see examples/forks.yml) and
  supply a token with the right scope.

Roadmap

• [x] Auto-request changes instead of only failing the check (opt-in)
• [x] Replace/update the report comment in place across runs
• [x] Config file support (.github/codex-guard.yml) for per-repo policy
• [x] Agent skill (SKILL.md) for pre-submit self-checks
• [x] workflow_dispatch sweep of existing agent PRs
• [x] Published to npm (npx codex-guard)
• [x] One-command observe-mode installer (npx codex-guard init)
• [x] Observe/balanced/strict installer presets and local setup doctor
• [x] Local scanning of non-ignored, untracked files
• [x] Shared Action/CLI presets with automatic local repository policy loading

Dropped by design: AI whole-PR summary gates — per-token LLM costs contradict
the project's free, deterministic positioning. Codex Guard stays zero-cost.

License

MIT