Skip to main content
reopt Handbook
reopt Handbook
Advanced Codex Usage

Foundation

Setup and LoginConfigMemories

Safety and Operations

ApprovalsSecurity

Workflow

Slash CommandsPrompts and SkillsExec and AutomationThread LifecycleCloud TasksMulti-Agent Workflows

Extensions and IDE

Codex App and IDEMCP IntegrationModels and ReasoningGovernance and Rollout

Verification

VerificationUpdates
Handbook›Advanced Codex Usage
한국어English

Advanced Codex Usage

An advanced operations and extension guide for teams standardizing Codex usage.

Recently Updated Chapters

  • Exec and Automation2026-09-12

    Use Codex exec for non-interactive scripts, CI checks, and structured automation.

  • MCP Integration2026-09-12

    Connect Codex to MCP servers with explicit tool, data, and trust boundaries.

  • Setup and Login2026-09-12

    Standardize Codex installation, updates, identity, and local developer setup.

  • Thread Lifecycle2026-09-12

    Manage Codex sessions across start, resume, fork, compact, archive, delete, and handoff.

  • Models and Reasoning2026-09-05

    Choose GPT-6 Astra, GPT-5.6 Sol, Terra, Luna, and reasoning levels by task risk, cost, and latency.

September 12, 2026 release refresh

Updated for Codex CLI 0.154.0 (September 9). npm latest=0.154.0 and alpha=0.155.0-alpha.3.10 were checked on September 12, 2026 KST.

  • Sessions: experimental --worktree and /worktree create isolated checkouts for new or forked sessions, with browse/resume support. Inline questions accept choices or free text while work continues.
  • Extensions and approvals: existing sessions refresh installed plugin tools, skills, and hooks. Failed MCP OAuth refreshes surface sign-in without replaying rejected calls. Resume/fork preserve permissions, and new user instructions invalidate stale automatic approvals.
  • Compatibility: codex mcp-server and the standalone codex-mcp-server were removed. External MCP connections through codex mcp remain supported. Evaluate migration to experimental codex app-server; the official documentation does not support production workloads on that command.
  • Editing: /copy preserves formatting and supports status output; Vim gains R replace mode. Windows gains a shared background server and daemon management.

Sources: official changelog, npm dist-tags.

Previous verification snapshot (September 5, 2026)

Codex is useful as an individual coding agent, but its real leverage appears when teams standardize configuration, safety rules, automation, context, and review practices.

This handbook is for senior developers and engineering leads who already use Codex and want to make it reliable in shared repositories and team workflows.

Core View

Codex quality depends on the environment around it: repository instructions, permissions, context packs, verification commands, and review routines.

September 2026 Baseline

Checked against Codex CLI 0.153.4, npm dist-tags, official release notes, and current Codex docs on 2026-09-05 KST. npm latest is 0.153.4 and npm alpha is 0.154.0-alpha.3.

  • GPT-6 Astra (released 2026-09-03): gpt-6-astra is now Codex's top model. It requires CLI 0.153.0 or later and, from 0.153.4, is the bundled default when no model is explicitly configured. It has a 1,050,000-token context window, 128K max output, API pricing of $10 input / $1 cached input / $50 output per 1M, and an April 30, 2026 knowledge cutoff. Teams that pin model in config.toml are unaffected by the default change, so adopt it deliberately.

  • Astra security classification: OpenAI classifies Astra as its first model to reach the Critical cybersecurity capability level under the Preparedness Framework. Narrow /permissions, sandboxing, and auto-approval scope before granting it broad automation.

  • CLI 0.153.0–0.153.4: Vim undo/redo, remote marketplace management from the plugin CLI, tui.auto_recap = false, TUI history showing complete patches and individual completed commands, and earlier allowance warnings for Plus and Team. Guardian review history now survives compaction, restarts, and forks, and Full Access skips Guardian reviews for confirmation-only actions. tui.disable_paste_burst and the disabled-by-default features.context_management.experimental_mode also land here.

  • CLI 0.152.0: the planning tool is disabled by default (enable with tools.update_plan.enabled = true). MCP server names accept :, @, /, and .; individual MCP tools accept output_token_limit; app-server clients can set thread/shellCommand timeouts beyond one hour; and rate-limit banners gain usage, credit, and plan actions.

  • CLI 0.151.0: a configurable discovery grace period for optional MCP servers, extensions that can inspect or replace MCP tool results before the model sees them, per-repository plugin catalog configuration, and nested subagent tokens counted toward root goal budgets.

  • CLI 0.150.0–0.150.1: task mentions and agent task messaging, the /copy picker, automatic task titles and /rename suggestions, clickable Markdown links, Interrupt hooks, untrusted-project instruction isolation, persistent managed deny-read policy, image-aware remote compaction, and reliability fixes.

  • CLI 0.149.0: the codex agents dashboard, codex queue, /cd//pwd//cwd, broader codex doctor diagnostics, and permission-profile restoration across resume and fork.

  • MCP transition: codex mcp-server is deprecated as of August 24; new integrations should use codex app-server while existing callers migrate.

  • CLI 0.148.0: /export writes a complete TUI conversation to Markdown, codex exec fork forks non-interactively, the resume picker archives/restores sessions, startup accepts draft prompts, and eligible workspaces see estimated credits/cost. It also adds a built-in Amazon Bedrock Runtime provider, async and MCP hook invocation, and broad resume, OAuth, rendering, and fail-closed sandbox fixes.

  • CLI 0.147.0: portable Agent Plugins and catalog search, manually ordered transcript sections, --approve-for-me, Cursor-skill import and conversation sync, and opt-in MCP 2026-07-28 support. Deprecated codex exec --full-auto is removed; use --sandbox workspace-write.

  • Model baseline: GPT-6 Astra is the top model; within GPT-5.6 use Sol for the hardest work, Terra for everyday engineering, and Luna for explicit repeated work. GPT-5.4 and GPT-5.4 mini retired from ChatGPT-sign-in Codex on 2026-08-31.

  • CLI 0.146.0: named /new and /clear threads, pinning and side-conversation switching, paginated-history forks, Agent Plugin manifests/workspace publishing, additional Bedrock and Claude Code marketplaces, remote Code Mode hosts, custom-provider web search, executor-provided skills, broader proxy handling, and live MCP/App tool refresh.

  • CLI 0.145.0: experimental paginated thread history with names, search, memories, and subagent support; broader Cursor/Claude Code /import; Bedrock login; audio; stabilized Multi-agent V2; and secure inline visualization links.

  • CLI 0.143.0–0.144.0: remote plugins are enabled by default with npm marketplace sources and local/remote versions; system proxy support covers macOS and Windows PAC/WPAD; codex remote-control pair generates manual pairing codes; MCP tool search is default and interactive/session auth is expanded; app approvals add a writes mode; reset credits show type/expiration and support explicit selection; Ultra warns about high multi-agent concurrency. The releases also harden retired-model compaction recovery, Windows writable-root deletion, terminal-control sanitization, and connector auth refresh.

Operating Model

Maturity Model

LevelStateCommon symptomPromotion condition
L1 PersonalEach developer uses Codex differentlyResults depend on individual habitsShared setup and basic rules exist
L2 StandardizedConfig and prompts are documentedSafety settings still driftApproval and sandbox defaults are governed
L3 AutomatedRepeatable tasks run through exec or agentsCI and review evidence are inconsistentVerification output is required
L4 GovernedTeam rollout is measurableChanges in Codex behavior surprise teamsChangelog, models, and policies are reviewed

Contents

Ch1. Setup and Login

Standardize installation, updates, identity, and shell integration.

Ch2. Config

Design profiles, defaults, and team policy layers.

Ch3. Memories

Manage persistent context without stale or unsafe facts.

Ch4. Approvals

Set safe autonomy boundaries with approval and sandbox rules.

Ch5. Security

Protect secrets, network access, telemetry, and auditability.

Ch6. Exec and Automation

Use non-interactive execution for scripts, CI, and repeatable work.

Ch7. Multi-Agent

Split work across bounded agents without losing review control.

Ch8. MCP

Connect tools and data sources with explicit trust boundaries.

Ch9. Governance

Roll out Codex with ownership, policy, and upgrade discipline.

Related handbooks

Claude Code Complete Guide

A practical guide to multi-session workflows, automation, and extension patterns in Claude Code.

Codex Command Master

Find Codex CLI commands by task, then follow the complete reference, availability rules, and concrete examples.

Claude Code Command Master

Find Claude Code commands by task, then follow the complete reference, availability rules, and concrete examples.

Agentic Documentation

A practical guide to documentation that AI agents can read, execute, verify, and govern.

Setup and Login

Standardize Codex installation, updates, identity, and local developer setup.

On this page

September 12, 2026 release refreshPrevious verification snapshot (September 5, 2026)Operating ModelMaturity ModelContents