Codex Team Mode 🤖 Agent Free

What it does

Codex Team Mode is a Codex Skill that turns a single Codex session into a small coordinated team. A primary agent handles task decomposition and final validation while four specialized sub-agent roles do the actual work with clear, non-overlapping responsibilities, so complex tasks progress in parallel instead of as one long linear thread.

What is inside

• Explorer — locates relevant files, entry points and call relationships in large codebases; explicitly scoped to file location, not general research
• Executor — implements changes once targets and file ownership are clear, operating within an assigned scope with no shared write permissions across parallel tasks
• Reviewer — independently examines completed work from a fresh context and can generate an optional Markdown audit report
• ExpertAdvisor — weighs in on complex decisions, modeling, automation, or problems the main agent has failed to resolve
• A hard rule that sub-agents cannot spawn further sub-agents, so the main agent always retains oversight and validates every deliverable before accepting it
• Diagnostic scripts (current_model.py, usage_by_model.py) for checking which model is actually driving the primary agent and tracking usage by model

Works with

Built specifically for Codex (OpenAI's coding CLI/agent platform); it is installed as a Codex skill and the role configs live under Codex's own agent directory.

How to install or connect

npx skills add oil-oil/codex-team-mode

Then copy the generated Explorer.toml, Executor.toml, Reviewer.toml and ExpertAdvisor.toml role configs into ~/.codex/agents/. Once installed, $team-mode triggers delegated task completion.

Maintenance and safety

A young but active project (145+ stars, 22 forks, MIT licensed). Its single-level delegation rule — sub-agents can't cascade into further sub-agents — is a deliberate safeguard against runaway agent spawning, and every deliverable is reviewed by the main agent before being accepted.

Who should use it

Codex users tackling multi-file or multi-step tasks who want built-in role separation (explore/execute/review/advise) without hand-writing their own prompt scaffolding for each role.