by arbiterForge
When you can't trust yourself with your code base, trust Arbiter.
# Add to your Claude Code skills
git clone https://github.com/arbiterForge/codeArbiterGuides for using ai agents skills like codeArbiter.
Last scanned: 7/25/2026
{
"issues": [],
"status": "PASSED",
"scannedAt": "2026-07-25T06:18:26.618Z",
"npmAuditRan": true,
"pipAuditRan": true,
"promptInjectionRan": true
}Shared enforcement and project-context parity across Claude Code, Codex CLI, and Pi.
Every intent routes through a gated skill or reviewer agent. Nothing commits until the gates are green. Decisions go through SMARTS. The audit trail is append-only.
Install it globally; it stays dormant until you opt a repo in.
License notice. As of v2.6.0, codeArbiter is licensed under the GNU AGPLv3, a change from its earlier MIT license. Copyright (C) 2026 SUaDtL, who reserves the right to dual-license under separate proprietary terms; commercial licenses are not offered at this time. See Dual-Licensing & Contributions.
codeArbiter is a governance layer for Claude Code, Codex CLI, and Pi. It lives in a repository of
four sibling plugins: the three governance hosts (ca, ca-codex, and ca-pi) plus the
ca-sandbox infrastructure plugin. The governance adapters are generated from one shared surface
and enforce one checked-in .codearbiter/ project store. Instead of letting the model freelance,
you drive through host-native commands. Each one routes to the process that owns the work and clears
its gates before anything ships.
The Codex path was live-verified on Codex CLI 0.144.1 with trusted SessionStart injection and a
real [H-03] PreToolUse block. See the
Claude Code + Codex evidence
for the dated proof, CI coverage, and intentional host differences.
Who it's for: teams and power users who let agents write real code and need to prove what happened.
It will not:
The gates are terse and non-negotiable. The thinking is not: it brainstorms a spec, works through a bug, and weighs a decision with you conversationally. When it enforces, it states the rule and holds the line.
The gates scale to the change: a one-line docs fix takes the small lane or /ca:chore, not the full spec-to-PR march. The weight exists because the failure mode of an eager AI assistant is plausible-but-wrong work that ships, and the gates make that hard.
codeArbiter predates OpenAI Build Week. The submission is the work added from July 13 through July
21, 2026: full Pi host support plus a corrective campaign over the existing governance core. The
new work is consolidated in PR #313. Its
verified implementation head is ec64809c1b6c32f69e3d9787ab205f7bbd563d50.
The Pi adapter adds a rich global footer, trust-aware repository activation, an extensible
allow | ask | deny execution policy, read-only plan mode, session-only background jobs, bounded
child-agent dispatch, native compaction, diagnostics, and shared-state continuity. Background jobs
never survive Pi shutdown. Parent-only UI does not leak into JSON, RPC, print, or child modes.
Codex was the primary development and orchestration environment for the submission. GPT-5.6 was used to design and implement the Pi adapter, compare its behavior with the established host, investigate failures found during live testing, review older implementation paths, and consolidate the resulting corrections. The work includes decisions and fixes across permissions, Windows process-tree cleanup, provenance handling, audit integrity, generated-source drift, CI contracts, and documentation.
The final candidate passed the six supported Pi platform cells for Pi 0.80.5 and 0.80.10 on Windows,
macOS, and Linux. It also passed Pi security analysis, CodeQL, the shared-core and generated-surface
contracts, and the repository's 46-gate final verifier. The sanitized evidence is available in
docs/reports/pi-support/promotion.md and
docs/reports/pi-support/promotion.json.
The submission supports Windows, macOS, and Linux. It requires Pi 0.80.10, Node.js 22.19 or newer, Python 3, and Git. No account or sample data is required.
Install the immutable candidate rather than a moving branch:
pi install git:github.com/arbiterForge/codeArbiter@ec64809c1b6c32f69e3d9787ab205f7bbd563d50
pi list
pi config
Open a disposable Git repository, inspect it, grant Pi project trust, and start a fresh Pi session. Then run:
/ca-init
/ca-doctor
The rich footer appears globally. Repository-specific governance appears only in an enabled, affirmatively trusted repository. The complete deterministic and trusted-live procedure is in the Pi parity-testing runbook. Judges can inspect the exact implementation, review discussion, and hosted check results in PR #313 without rebuilding the project from source.
The first time codeArbiter blocks you, it can feel like friction, until you see what it caught. This exchange is a real one from this project's own history: a bug fix forced through test-first, then a PR held at review until an untested seam got covered.
you /ca:fix the statusline keeps running the old version after an update
arbiter Routing to tdd (bug variant): a regression test before any fix.
→ writes a failing test, confirms it's red for the right reason
→ minimal fix → suite green → coverage + lint gates clear
you /ca:commit
arbiter commit-gate: ✓ permission ✓ branch ✓ tests ✓ secrets
✓ behavioral proof ✓ clean diff → committed.
you /ca:pr
arbiter reviewer fleet over the diff: coverage-auditor flags an untested seam.
BLOCK. Here's the gap. → (you resolve, re-run) → PR opened.
Every step is a gate you watch clear. You stay in the driver's seat; the gates keep the work honest.
codeArbiter self-hosts its plugins from this repository. Install the host adapter you use; all three
governance hosts read the same .codearbiter/ project state.
/plugin marketplace add arbiterForge/codeArbiter
/plugin install ca@codearbiter
Hooks, commands, agents, and statusline wiring load automatically; everything resolves under the /ca: namespace.
The public GitHub-slug flow is available now. It was verified against release v2.8.13 with
ca-codex 0.2.4:
codex plugin marketplace add arbiterForge/codeArbiter
codex plugin add ca-codex@codearbiter
Open /hooks, review and trust the ca-codex handlers, then start a fresh thread. Commands use the
$ca-* spelling. Run $ca-init to opt a repository in and $ca-doctor to prove the hooks are live.
For development against an unpublished checkout, use the local-clone flow:
git clone https://github.com/arbiterForge/codeArbiter
cd codeArbiter
codex plugin marketplace add .
codex plugin add ca-codex@codearbiter
Feature Forge preview.
ca-piis available for real use now, and you are welcome to install it, use it in your repositories, and report what you find. It has passed the documented automated and hosted promotion matrix, but it still needs broader real-world testing before codeArbiter claims 100% validation or stable status.
Pi distribution is Git-only. Pin the independent ca-pi release tag, then inspect the installed
source and enabled resources:
pi install git:github.com/arbiterForge/codeArbiter@ca-pi-v<version>
pi list
pi config
Pi 0.80.5 and Pi 0.80.10 are the supported hosts for this release line. ca-pi also requires Node
22.19+ and Python 3. After inspecting the project, grant Pi project trust, start a fresh session,
then run /ca-init and /ca-doctor. Generated aliases use /ca-*; /skill:ca-* is the native
fallback. The Pi parity runbook covers isolated installation, live
verification, uninstall, and the scope of the module-identity diagnostic. There is no npm release.
Pi's generated human-readable skill catalog is plugins/ca-pi/SKILLS.md, outside the
loader-scanned skills/ directory. Before the platform aggregate runs any fixture, it checks the
resolved tools workspace; a cold checkout exits with missing_prerequisite and directs you to
npm --prefix plugins/ca-pi/tools ci --ignore-scripts rather than installing dependencies itself.
Prerequisites: Python 3 on PATH. Pi installs its final TypeScript wrappers before bridge
readiness, so a missing interpreter blocks mutating calls and points to /ca-doctor; Claude Code
and Codex host shims surface an interpreter breadcrumb instead of silently disabling governance.
Also set git config user.email (overrides and ADRs are attributed to
that identity). Full version matrix: Compatibility.
The optional /ca:statusline</kbd
codeArbiter is an open-source ai agents skill for AI coding assistants such as Claude Code, Codex CLI, and ChatGPT, built by arbiterForge. When you can't trust yourself with your code base, trust Arbiter. It has 137 GitHub stars.
Yes. codeArbiter passed SkillsLLM's automated security scan — a dependency vulnerability audit plus prompt-injection heuristics — with no high-severity issues. You can read the full report in the Security Report section on this page.
Clone the repository with "git clone https://github.com/arbiterForge/codeArbiter" and add it to your Claude Code skills directory (see the Installation section above).
codeArbiter is primarily written in Python. It is open-source under arbiterForge on GitHub, so you can review or fork the full source.
Yes. SkillsLLM lists many other AI Agents skills you can browse and compare side by side. Open the AI Agents category from the badge at the top of this page, or use the Related Skills and comparison links further down to weigh codeArbiter against similar tools.
No comments yet. Be the first to share your thoughts!