Skip to main content
skill CX Codex

codex-host — Codex skill: what it does and how to install

Runs the full great_cto pipeline through a controller that owns writes and approval gates. Use it when you need the whole workflow, not a single skill.

Install into Claude Code · Skills CLI (Vercel) All methods and platforms ↓
npx skills add avelikiy/great_cto --skill codex-host -a claude-code
Open on GitHub

What codex-host does

The skill hands a great_cto run to a controller. It reads shared/pipeline.toml, tracks run state and limits writes to paths explicitly allowed at startup. Before starting, doctor checks readiness and list shows the project’s runs.

Start with an absolute project path and operator-owned policy files stored outside that project. The controller may pause for human approval before continuing or releasing artifacts; pass back its issued token only after approval. A blocked doctor state means the run must not start, and exit code 2 is not a successful result.

The local and GitHub Release adapters publish immutable artifacts, but do not activate them in production. A separate adapter and approval contract are needed to activate production traffic. That boundary matters if your goal is deployment rather than a release artifact.

What it helps with

  • run the full great_cto pipeline for a project
  • check controller readiness and inspect project runs
  • restrict controller writes to approved paths
  • approve a gated continuation or artifact release
  • resume a run after reviewing its recorded reason

Example prompts for the agent

  • Run the full great_cto pipeline for this project and allow changes in src, tests and docs.
  • Check great_cto readiness and list runs for /absolute/project.
  • Show me the release approval summary and wait for my explicit approval.

How to install codex-host

1Skills CLI (Vercel)

npx skills add avelikiy/great_cto --skill codex-host -a claude-code

2Manually: put the files in place

~/.claude/skills/codex-host/SKILL.md

Copy the whole skill folder; for one project use .claude/skills/codex-host/

Step by step, with commands for each platform: How to install skills in Claude Code: five ways

Questions about codex-host

How do I install codex-host in Codex?

Run npx skills add avelikiy/great_cto --skill codex-host -a codex. The skill file is placed at .agents/skills/codex-host/SKILL.md.

What does codex-host do?

It runs the full great_cto workflow through a controller that manages run state, permitted writes and approval gates.

Does codex-host add great-cto to PATH?

No. It invokes the pinned npm package with npx --yes [email protected]; installing a Codex plugin does not put the great-cto binary on PATH.

Source text

SKILL.md · MIT
---
name: codex-host
description: Run the great_cto controlled Codex lifecycle with controller-owned writes, verifier evidence, human gates and optional artifact release.
when_to_use: Use when the user asks Codex to run or inspect a full great_cto pipeline rather than only use an individual skill.
effort: high
allowed-tools: Read, Bash
---

# Controlled Codex host

Use the version-pinned npm entrypoint below. A Codex plugin install loads skills
and MCP configuration, but it does not add the `great-cto` npm binary to `PATH`.
Do not imitate Claude Code hooks or manually chain roles: the controller reads
`shared/pipeline.toml`, persists the cursor and owns every write and gate
transition.

## Preflight

```sh
npx --yes [email protected] codex-host doctor
npx --yes [email protected] codex-host list --dir /absolute/project
```

`doctor.state=blocked` means do not start. Docker may be unavailable when no
checks/release policy is requested; GitHub CLI may be unavailable when the local
adapter is used. The run store must not be group/world accessible.
For mixed-host runs, also require `doctor.checks.claude.state=available`.

## Start

Policies must be operator-owned absolute files outside the worker project.
Allowed paths are explicit controller write boundaries.

```sh
npx --yes [email protected] codex-host start \
  --dir /absolute/project \
  --prompt "the requested outcome" \
  --allow src,tests,docs \
  --checks-policy /absolute/operator/checks.json \
  --release-policy /absolute/operator/release.json
```

Read the returned `status`, `pending` and `release` object. Never treat exit 2 as
success: it means a gate, manual action or blocked evidence requires attention.

Since 3.47.0, independent
graph roles can run on both hosts with
`--routes qa-engineer=claude-code,security-officer=codex` on `start`. The
controller dispatches only a symmetric join pair concurrently, checks both
proposals for overlap, serializes writes and runs the existing verifier and
human gates. The route map is fixed for the run. An interrupted wave must be
inspected; never start replacement workers against an uncertain wave.

## Gates and recovery

Show the operator the exact gate/release summary and wait for explicit approval.
Then pass back the controller-issued token; never synthesize or persist one in a
project file.

```sh
npx --yes [email protected] codex-host approve RUN_UUID --token GATE_TOKEN
npx --yes [email protected] codex-host approve-release RUN_UUID --token RELEASE_TOKEN
npx --yes [email protected] codex-host resume RUN_UUID
```

Use `recover` only after inspecting the recorded reason. Recovery reuses the
same candidate and refuses unknown partial writes. `cancel` revokes unexecuted
approval but does not delete external audit evidence.

## Release boundary

The local and GitHub Release adapters publish immutable artifacts and declare
`activation: none`. They do not deploy traffic. Local rollback means selecting
a previous directory; GitHub rollback is a newly gated superseding release.
Production activation needs a separate adapter and approval contract.

Detailed policy schemas and limitations are in `docs/HOST-CODEX.md`.

© avelikiy · MIT · avelikiy/great_cto

Similar skills

All skills for Agent setup: memory, context, orchestration →

From the same set34

decision-eval Runs a scoring agent on architectural alternatives in ADRs and outputs ranked evaluations to docs/decisions/. Helpful when choosing between competing designs. CX Codex +1 ★ 107repo stars crystallize crystallize mines session logs and lessons for recurring patterns, then drafts reusable skills. Run it when you want to turn project knowledge into agent instructions. CX Codex +1 ★ 107repo stars cost-model Pins down one format for the cost section in plans and architecture docs. Useful when a report claims savings and someone will ask how you got the number. CX Codex +1 ★ 107repo stars committed-aesthetic Replaces vague design catalogues with testable rules for coding agents. Helps build distinct interfaces from scratch without falling into generic AI layouts. CX Codex +1 ★ 107repo stars brainstorming Takes a raw product idea through three steps: divergent bets, a two-round argument between four personas on four models, and a synthesised call. CX Codex +1 ★ 107repo stars archetype-review-base Standardizes domain review reports in Claude Code. It creates TM-{slug}.md threat models, checks source quotes, and enforces repro steps for critical findings. CX Codex +1 ★ 107repo stars anydesign Analyzes images, websites, and Figma files, then creates design.md with design tokens, components, and reconstruction notes. For a single element, it produces element.md. CX Codex +1 ★ 107repo stars anti-patterns A reference list of known SDLC anti-patterns that great_cto agents cite when they push back on an architecture, a plan or a diff. CX Codex +1 ★ 107repo stars

Project sponsors

Companies that keep this analytics open
Gold sponsor

Impressions, clicks and code copies — open statistics. We keep the numbers public for the same reason NeBlask itself is public. Become a sponsor