Codex Guard v1.15.1

Windows contributor experience.

Highlights:

• Documentation now includes a PowerShell local dry-run recipe, contributed by @zihanw52-design in PR #22.
• npm run check is now implemented in Node rather than a POSIX shell loop.
• CI runs the test suite on Windows Node 24 while retaining the protected Linux status checks.
• Test fixtures handle Windows filename rules and CRLF frontmatter.
• The DeepSeek Harness bundle invokes npx through Node on Windows, avoiding a command-shell boundary for a git ref.

Validation: 146 tests, Linux Node 22/24, Windows Node 24, coverage, manifest validation, Markdown lint, zero production audit findings, and a packed-package syntax-check smoke test all passed.

Codex Guard v1.15.0

Action and CLI policy parity.

Highlights:

• GitHub Action now accepts observe, balanced, and strict policy presets.
• Local Git scans automatically load .github/codex-guard.yml.
• Explicit CLI preset, config, no-config, warning, and blocking overrides are supported.
• Policy precedence is documented and backward compatible for workflows without a preset.
• Unknown fail-on checks now fail validation instead of accidentally weakening enforcement.
• Doctor validates workflow and repository preset values.
• Generated workflows, examples, docs, Action outputs, and project dogfooding use the shared model.

Validation: 144 tests, Node 22/24 CI, coverage, manifest validation, Markdown lint, zero production audit findings, and a clean tarball install/init/doctor smoke test all passed.

Codex Guard v1.14.0

Local confidence and easier rollout.

Highlights:

• New read-only codex-guard doctor command validates workflow YAML, PR triggers, Action steps, permissions, policy keys and policy values.
• New observe, balanced, and strict installer presets; --strict remains compatible.
• Local --git scans now include non-ignored untracked files before git add.
• Binary and untracked files larger than 8 MiB are reported as unscanned.
• Quoted policy booleans are coerced correctly, and todo-blocking now maps to the correct internal option.
• Marker-shaped identifiers no longer create false unfinished-work findings.
• Terminal output is sanitized and exported source context is bounded.

Validation: 136 tests, Node 22/24 CI, coverage, manifest validation, Markdown lint, package audit, and tarball install smoke tests all passed.

Codex Guard v1.13.0

Trust-boundary hardening and more truthful CI gating.

Highlights:

• Sanitizes and bounds attacker-controlled report text; sweep output reaches shell steps only through an environment variable.
• Paginates commit statuses, check runs, and report comments.
• Pending or partially visible CI can no longer appear green; both CI APIs unavailable now fail closed.
• Handles every completed GitHub check conclusion and ignores previous Codex Guard runs.
• Makes sweep-label effective and detects GitHub Copilot copilot/ branches by default.
• Pins internal GitHub Actions by full commit SHA and locks the workflow validator.

Upgrade note: existing workflows should add statuses: read to their permissions block. Generated workflows already include it.

Validation: 120 tests, Node 22/24 CI, coverage, manifest validation, Markdown lint, package audit, and package dry-run all passed.

Codex Guard v1.12.0

Honest content-scan coverage

Codex Guard now distinguishes no findings from not scanned.

• reports scanned / eligible changed-file coverage
• returns a neutral, non-blocking warning when GitHub omits a textual patch
• lists affected paths with bounded, sanitized output
• exposes coverage in Action outputs, findings-json, and sweep reports
• reports uncertainty at GitHub's 3,000-file pull-request API limit
• declares and tests the existing ci-failure-count output

This design borrows the mature-tool principle of separating visibility from enforcement: incomplete coverage is impossible to mistake for a complete clean scan, but it does not suddenly break existing users.

Full changelog: v1.11.0...v1.12.0

Codex Guard v1.11.0

One-command safe setup

New adopters can now add Codex Guard from the root of any Git repository:

npx --yes codex-guard init

The generated workflow starts in observe mode so teams can see real findings before deciding what should block. Use --strict for immediate enforcement and --force for an explicit overwrite.

This release also fixes the Action owner in all bundled workflow examples and adds a regression test to keep them installable.

Full changelog: v1.10.0...v1.11.0

v1.10.0 — Trust and packaging hardening

Codex Guard 1.10.0 focuses on trust: reports no longer re-leak the secrets they detect, observe mode is genuinely non-blocking, and the Action now ships as a reproducible bundle with clean runtime dependencies.

Highlights

• Redacts secret-shaped values from every exported source context, including findings-json and TODO report lines.
• Makes soft-fail produce a neutral check-run with honest observe-mode wording.
• Paginates commit inspection beyond 100 commits.
• Handles large local diffs safely and passes git refs without shell interpolation.
• Replaces committed node_modules with a checked-in dist/ bundle that CI verifies against source.
• Updates the GitHub Actions runtime dependencies and removes all known production audit findings.
• Adds an npm OIDC trusted-publishing workflow for future releases.
• Updates first-party workflows to actions/checkout@v7 and actions/setup-node@v7.

Install or upgrade

- uses: Akimiya-z/codex-guard@v1

Pin this exact release with Akimiya-z/codex-guard@v1.10.0, or run the CLI without installing it:

npx --yes codex-guard@1.10.0 --git --commits

The v1 major tag points to this release. npm users receive 1.10.0 from the latest dist-tag.

v1.9.0 — DeepSeek Harness plugin

One more home for the gate: inside DeepSeek Harness.\n\n- dsh/ bundle declares dsh.bundle and registers a codex_guard tool via the official @deepseek-ai/dsh-tools API; the tool runs npx --yes codex-guard --git in the current repo.\n- Tests run against the real dsh-tools packages, including an end-to-end tool call in a throwaway git repo (catches FIXME + AWS key).\n- npm ships dsh/; repo topic dsh-plugin. 81 tests.

v1.8.0 — Codex plugin + marketplace

Installable inside Codex, alongside GitHub and npm.\n\n- Skill-only plugin per the official openai/plugins schema (plugins/codex-guard/.codex-plugin/plugin.json).\n- Marketplace manifest (.agents/plugins/marketplace.json) — Codex users can add this repo as a plugin marketplace.\n- npm package now ships the plugin bits; new codex-plugin topic on GitHub.\n- 77 tests.

v1.7.1 — wider secret net

More secrets caught, same zero-cost positioning.\n\n- New patterns: npm tokens, SendGrid, Telegram bots, Azure storage connection strings, JWTs, explicit AWS secret keys.\n- Test fixtures are built at runtime — GitHub's own push-protection blocked a literal fake SendGrid key in a test file, in exactly the way Codex Guard flags on PRs. 😄\n- AI whole-PR summary gate dropped by design — per-token LLM cost contradicts free/deterministic Codex Guard.\n- 74 tests.

🛒 Now on GitHub Marketplace: https://github.com/marketplace/actions/codex-guard-pr-quality-gate