codexclaw

codexclaw

Development discipline for OpenAI Codex

Get StartedHow It Works

PABCDSkillsSubagents

cxc — PABCD loop

$ codex plugin add codexclaw@codexclawAdded plugin codexclaw$ cxc orchestrate P –attest ‘{“from”:“IDLE”,“to”:“P”,…}’orchestrate P: IDLE -> P$ cxc orchestrate statusIPABCD: P (PLAN) · interview:false audit:false check:false$ cxc map src/ranked 42 files · tree-sitter + PageRank$ cxc loop validate –slug releaseOK — complete + all met criteria carry evidence

codexclaw (cxc) is a single OpenAI Codex plugin. It reuses the Codex runtime and adds development discipline on top: a cxc-dev work classifier, an IPABCD work loop driven by hooks and file state, multi-model subagent configuration over MCP, a shipped GUI dashboard, and an optional detect-only opencodex provider bridge.

It is not a Codex fork, not a harness, and not a provider proxy. Everything ships through the standard Codex plugin surfaces: skills, hooks, an MCP server, local components, and a small cxc CLI. The only scoped server exception is the opt-in loopback messenger bridge, which relays to stock codex exec and does not orchestrate subagents.

PABCD pipeline

IInterviewsettle ambiguity before a work loop starts

PPlanwrite the concrete change plan and evidence target

AAuditadversarial review against the real codebase

BBuildscoped atomic changes, verified as you go

CCheckreal build, tests, and review gate

DDoneevidence summary, close to idle

Codex runtime
  └─ codexclaw plugin
       ├─ skills   →  cxc-dev discipline + dev-* routers + PABCD/loop/interview
       ├─ hooks    →  PABCD state, goal guards, provider detection
       ├─ MCP      →  subagent model/prompt config
       ├─ GUI      →  dashboard + opt-in messenger bridge
       └─ cxc CLI  →  enable · status · doctor · orchestrate · gui

Surface map

Surface	Status	Count / scope	Primary docs
Skills	current	25 skills: eight implicit skills (dev, search, interview, pabcd, recall, loop, dev-frontend, dev-uiux-design), dev routers, qa, repo-map, and research/search helpers.	Skills
Hooks	current	12 hook JSON files: SessionStart x2, UserPromptSubmit x1, Stop x1, PreToolUse x4, PostToolUse x2, SubagentStop x1, PostCompact x1.	Hooks
MCP	current	Subagent model and prompt config: subagents_get, subagents_set, catalog_list.	MCP Tools · Subagents
Components	current	8 component packages: config guard, ops, recall, PABCD state, provider bridge, subagent config, messenger bridge, skill search. GUI is separate.	How It Works
cxc CLI	current	enable, disable, uninstall, status, orchestrate, freeze, metric, divergence, loop, goalplan, doctor, reset, subagents, map, provider, chat, memory, skill, gui, serve, service.	Commands
GUI	current	Dashboard plus shipped opt-in loopback messenger bridge through cxc serve; not a jaw-style orchestrator.	GUI Dashboard

Quick start

# Install from the public marketplace
codex plugin marketplace add lidge-jun/codexclaw
codex plugin add codexclaw@codexclaw

# Verify the install
cxc doctor

# Start a PABCD work-phase from the terminal
cxc orchestrate P --attest '{"from":"IDLE","to":"P","did":"plan the change"}'

The project gate is green at 801 tests. Build and test commands are documented separately so the landing page stays a control surface, not a release note.

Status honesty

These docs mark features current, planned, or deferred. A planned feature has a decided design but no shipped runtime yet. The docs never describe a planned feature as if it shipped.

Where to next

InstallationInstall from the public marketplace or dogfood a source checkout.->How It WorksThe plugin boundary: skills, hooks, MCP, file state — no fork, no harness.->PABCD WorkflowRun the IPABCD loop with attested phase gates and evidence ledgers.->Skills25 skills: the cxc-dev classifier, surface routers, search, recall, QA.->SubagentsPer-role model and prompt config applied to live spawn_agent calls.->Parity RoadmapWhat is current, planned, and deferred — status-honest scope.->