Codex Claude Orchestrator

A Codex plugin that delegates a bounded coding task to Claude Code and brings
its result back to Codex for verification.

You give Codex the outcome. Codex decides whether delegation helps, gives Claude
one local task, and checks the changes before calling the work complete. Small
changes can stay in Codex. The plugin keeps track of which session owns the work
so two agents do not accidentally edit the same worktree.

Choose a variant

Branch	Use it when
`main`	You want Claude to work directly with repository files, without an MCP dependency.
`codeindexer`	You use CodeIndexer and want guarded semantic search and dependency discovery.

The branches are separate runtime variants. Install the one you need; do not
merge them.

This checkout is the main variant. It does not enable MCP for Claude workers.

Requirements and installation

You need Codex and Claude Code installed and authenticated, plus zsh, jq,
Git and the platform locking/process tools described in the skill. Use macOS or
Linux. Windows is not supported.

The Codex tool shell must provide an interactive PTY and CODEX_THREAD_ID.
This integration depends on that host behavior, which can change between Codex
versions.

codex plugin marketplace add coredo-eu/codex-claude-orchestrator
codex plugin add codex-claude-orchestrator@codex-claude-orchestrator

For an existing marketplace checkout, confirm it is on main before installing.

Who does what

Component	Responsibility
Codex	Understands the request, chooses the executor, makes key decisions, and verifies the result.
Claude Code	Completes one delegated local stage, then returns its changes and evidence.
Optional specialists	Help with discovery, implementation, testing, or review when useful.

The plugin keeps your selected Codex model and reasoning level. A Claude worker
starts with Sonnet 5 at high effort; its optional roles use Haiku, Sonnet, Opus,
or Fable. Native Codex fallback roles have their own model settings. No task
must use every role. See the skill for the exact routes.

Use

Start a new Codex session after installation, then ask:

Use $codex-claude-orchestrator:claude-pty-agents to fix this bug.
Give Claude one bounded task and verify its changes before finishing.

For automatic executor selection, merge the
opt-in policy
into your existing AGENTS.md. Installing the plugin alone does not change your
instructions or install the optional native agent profiles.

Boundaries

One edit owner per worktree; a worker handoff returns control to Codex.
Workers make local changes. Commits, pushes, deployments, and other external
actions stay with Codex and require the user's task to authorize them.
The plugin only controls workers registered to the current Codex thread. It
does not attach to standalone Claude sessions or change their settings.
A kill switch stops new worker interactions. Recovery checks prevent another
executor from taking over while the original worker may still be writing.

These controls are cooperative, not an operating-system sandbox. Use trusted
repositories and appropriate host permissions. Coordination runs locally;
model requests still go to OpenAI and Anthropic under your account settings.

Documentation and checks

Run the deterministic checks without a real Claude session or credentials:

./scripts/self-check.zsh

The checks use temporary state and fake worker processes. CI runs them on macOS
and Linux; real CLI compatibility still depends on your installed Codex and
Claude Code versions.

Related repositories

Codex agent configuration: portable instructions and optional native specialist profiles.
Claude agent configuration: configuration for standalone Claude Code.