Ca

Name: Ca
Author: arbiterForge

arbiterForge/codeArbiter

138 repo stars
Updated August 5, 2026
arbiterForge/codeArbiter

ca (codeArbiter) is a skill/plugin that acts as an orchestration layer for Claude Code, routing every intent through gated skills and reviewer agents with an append-only audit trail.

About

An orchestration layer for Claude Code that routes every intent through a slash command to a gated skill or reviewer agent, blocking commits until review gates pass. It keeps an append-only audit trail of overrides, triage, ADRs and a sprint log, and activates per-repo only after /ca:init. A developer uses it to enforce disciplined, reviewed agentic coding. It matters because it stops agents from freelancing changes.

Orchestration layer for Claude Code that gates every intent through skills and reviewer agents
Nothing commits until the review gates are green; append-only audit trail
Per-repo, explicit activation via /ca:init scaffolding .codearbiter/

Ca by the numbers

Data as of Aug 5, 2026 (Skillselion catalog sync)

At a glance

ca capabilities & compatibility

Free; a Claude Code plugin requiring only Python 3 and git.

Capabilities: agent orchestration · code review · workflow gating
Works with: github
Use cases: code review · orchestration · project management
Runs: Runs locally
Pricing: Free

/plugin marketplace add arbiterForge/codeArbiter

/plugin install ca@codearbiter

Add your badge

Show developers this skill is listed on Skillselion. Paste this into your README.

[![Listed on Skillselion](https://skillselion.com/badge/skills/arbiterForge/codeArbiter/plugins/ca.svg)](https://skillselion.com/skills/arbiterForge/codeArbiter/plugins/ca)

repo stars	★ 138
Last updated	August 5, 2026
Repository	arbiterForge/codeArbiter ↗

What it does

Govern Claude Code so every intent routes through gated skills and reviewer agents, and nothing commits until gates pass.

Who is it for?

Developers who want disciplined, gated, auditable Claude Code workflows instead of freewheeling edits.

Skip if: Users who want an unconstrained agent, or repos where per-repo setup overhead is unwelcome.

When should I use this skill?

You want to enforce review gates and an audit trail over Claude Code changes in a repo.

What you get

An orchestrated Claude Code workflow where every change is gated, reviewed and logged before commit.

Gated agentic coding workflow
Append-only audit trail (overrides.log, triage.log, ADRs, sprint log)

By the numbers

Commands include /ca:feature, /ca:fix, /ca:chore, /ca:spike, /ca:sprint, /ca:commit, /ca:audit

README.md

codeArbiter (`ca`)

An orchestration layer for Claude Code that refuses to freelance. Every intent routes through a slash command to a gated skill or reviewer agent; nothing commits until the gates are green; decisions go through SMARTS; the audit trail (overrides.log, triage.log, ADRs, sprint log) is append-only and mechanically guarded.

Full documentation, install instructions, and the command catalog live in the repository README and in COMMANDS.md (/ca:commands in-session).

The short version

Install: /plugin marketplace add arbiterForge/codeArbiter → /plugin install ca@codearbiter.
Prerequisites: Python 3 on PATH (all hooks are Python); git config user.email set (audit attribution). Optional /ca:statusline writes to your global ~/.claude/settings.json (backed up; restored on removal).
Activation is per-repo and explicit: the plugin is dormant everywhere until you run /ca:init in a repo, which scaffolds .codearbiter/ with arbiter: enabled. No .codearbiter/, no behavior: hooks exit immediately, nothing is injected, nothing is blocked.
Drive it: /ca:feature (spec-driven, with a logged small lane for small changes), /ca:fix, /ca:chore, /ca:spike, /ca:sprint (autonomous, every decision logged), /ca:commit, /ca:audit. /ca:commands lists everything.
What it writes to your repo: .codearbiter/ only. Uninstalling the plugin leaves your project state intact and yours.

Feature Forge (preview)

Some features ship in the box but aren't yet blessed: off by default, fully dormant until you opt in, labeled preview until real-world data earns them a stable promotion. In the forge now: live transcript pruning (CODEARBITER_PRUNE), which trims redundant transcript clutter to extend a session. Run it in dry mode (export CODEARBITER_PRUNE=dry) and it logs what it would prune (sizes and verdicts only, no transcript content) to ~/.codearbiter/metrics/prune-dry.jsonl. Sending that log back (open a prune-data issue) is what moves it toward on. Also in the forge: the farm (/ca:sprint --farm, needs FARM_API_KEY), a pluggable backend that runs the implementation step on cheap OpenAI-compatible workers (OpenCode Zen by default); Claude still writes the spec, the failing tests, and the plan, and every task clears the full review chain. Off by default and preview until validated. A third preview, the ca-sandbox container-isolation plugin, ships as a separate marketplace install. Full detail in the Feature Forge section of the repo README.

What's in this directory

ORCHESTRATOR.md (the persona injected in enabled repos) · commands/ · skills/ · agents/ · hooks/ (SessionStart injection + PreToolUse/PostToolUse gates) · includes/ (routing table, reference map, farm setup) · tools/ (the --farm dispatcher) · SPRINT.md (the /ca:sprint mode body).