Name: Stoneforge
Author: Stoneforge

How Stoneforge orchestrates Codex CLI

OpenAI’s Codex CLI is an AI coding agent that runs in a sandboxed terminal environment with built-in safety controls. As a Codex AI coding agent, each session handles one task in isolation.

Stoneforge adds multi-agent orchestration on top of OpenAI Codex, dispatching multiple tasks to parallel Codex CLI agents and handling coordination, merging, and failure recovery.

How Stoneforge dispatches to Codex CLI

When you create tasks, the Daemon watches the queue and dispatches work to available Codex agents:

sf plan create --title "Backend Improvements"

sf task create --title "Add rate limiting to API" --plan "Backend Improvements"
sf task create --title "Optimize database queries" --plan "Backend Improvements"
sf task create --title "Add request logging middleware" --plan "Backend Improvements"

# Codex agents pick up tasks automatically
sf serve

Each Codex agent gets its own worktree, task context, and acceptance criteria.

Sandboxed execution at scale

Codex CLI’s sandboxing keeps each agent contained. Stoneforge adds git-level isolation through worktrees, so multiple Codex agents run in parallel on separate branches without interfering with each other.

Merge and review pipeline

When a Codex agent completes its task, the Steward:

Validates the branch builds and tests pass
Checks for merge conflicts with other completed work
Creates a merge request with a summary of changes
Merges automatically if checks pass, or flags for human review

Getting started with Codex

# Install Stoneforge
npm install -g @stoneforge/cli

# Initialize and configure for Codex
sf init

# Update .stoneforge/config.yaml to use codex provider
# Start orchestrating
sf serve

Codex CLI has three approval modes: suggest (approve everything), auto-edit (auto-apply file changes, approve shell commands), and full-auto (no confirmation). Set the approval field in your Stoneforge config to match your preferred level of oversight.

Learn more

Best AI Coding Agents in 2026 — see how Codex CLI compares to other AI coding agents
Run Multiple AI Coding Agents in Parallel — a practical guide to parallel agent workflows
Claude Code Orchestration — compare with another supported AI coding agent provider

Configuration

.stoneforge/config.yaml

# .stoneforge/config.yaml
orchestrator:
  agent: codex
  maxConcurrent: 4
  worktreeIsolation: true

agents:
  worker:
    provider: codex
    model: o4-mini
    approval: auto-edit
  steward:
    provider: codex
    model: o4-mini
    autoMerge: true

Codex CLI alone vs. with Stoneforge

Codex CLI alone	With Stoneforge
Single sandboxed session per task	Multiple Codex agents dispatched to parallel tasks
Manual task assignment and tracking	Automatic dispatch with priority and dependency awareness
Sandbox isolation per session	Git worktree isolation with automatic merge coordination
No built-in multi-agent coordination	Director-worker-steward role hierarchy
Manual review of each change	Steward agent runs tests and merges automatically