🏠 Index | ⬅️ Prev (Ch.12) | 🌐 中文版
Ch.13 Frontier Watch: The 2026 Codex Ecosystem Overhaul
🎯 The Real Problem: Rapid tool churn, deprecated commands (Codex desktop merging into ChatGPT code mode), model evolutions (GPT-5.6), and unexpected CLI breaking changes.
💡 Tangible Output & Takeaway: 2026 migration command battle map (winget/brew setups, CLI 0.14x configurations, Plugins ecosystem integration), and compatibility test scripts.
⚡ Viral Screenshot Quote: "Tools retool every quarter, but mental models remain your moat. Zero fluff—only an actionable battle map of what deprecated and what to adopt."
The first twelve chapters built a version-independent orchestration methodology. But methodology must land on real tooling. In 2026, the Codex ecosystem went through four structural shifts: the desktop merger, the model transition, the maturing plugin economy, and security becoming its own product line. This chapter breaks down each shift with concrete migration commands and configurations.
13.1 The Desktop Merger: Codex App Folds into the ChatGPT Client
In July 2026, the standalone Codex desktop app was retired. All of its capabilities moved into the ChatGPT desktop client as a dedicated "Codex Code Mode," available to Free, Plus, and Enterprise users.
Migration steps:
# macOS: download ChatGPT.dmg and drag it into Applications
# Windows: install via winget
winget install OpenAI.ChatGPTAfter signing in, click the Codex tab in the left sidebar to enter Code Mode. Two caveats:
- Account vs. API Key: Signing in with a ChatGPT account unlocks the full feature set (cloud tasks, cross-device sync, Computer Use); an API key grants only basic local coding.
- Standalone components are unaffected: the Codex CLI, the VS Code extension, and Codex Cloud continue to evolve independently. The multi-surface matrix from Ch.02 still holds — the "desktop app" cell is simply replaced by the ChatGPT client.
New capabilities at a glance:
- Upgraded built-in browser: search browsing history from the address bar; the Chrome extension can reference open tabs.
- Multi-repo Review: multi-folder projects show changed lines across all repositories in one review view — no more tab-hopping between diffs.
- Sites: create, deploy, and manage hosted web projects directly in the client, with Annotations for in-place "point and edit" workflows.
13.2 The Model Transition: GPT-5.6 Terra and Luna Take Over
As of August 31, 2026, GPT-5.4 and GPT-5.4 mini are retired from Codex for ChatGPT-authenticated sessions. The official replacements:
| Retired model | Successor | Role |
|---|---|---|
gpt-5.4 | gpt-5.6-terra (GPT-5.6 Terra) | Primary deep-reasoning model |
gpt-5.4-mini | gpt-5.6-luna (GPT-5.6 Luna) | Lightweight fast model (Guardian approvals, etc.) |
Migration checklist:
# 1. Audit your configs for hard-coded legacy models
grep -rn "gpt-5.4" ~/.codex/config.toml .
# 2. Explicitly upgrade the default model in config.toml# ~/.codex/config.toml
model = "gpt-5.6-terra"
# Lightweight background duties (e.g. auto-review) can be pinned separately
[profiles.guardian]
model = "gpt-5.6-luna"⚠️ Note: API-key-authenticated sessions may continue calling the GPT-5.4 family, but ChatGPT sign-in sessions must switch. Audit every Automation, workspace default, and custom agent.
13.3 CLI 0.14x: The Breaking Changes You Must Know
The Codex CLI is now in the 0.14x era (latest stable: 0.147.0 as of August 2026). These changes directly affect every command example earlier in this book:
npm install -g @openai/codex@latest1. --full-auto Is Officially Removed
# ❌ Old syntax (now errors out)
codex exec --full-auto "fix all lint errors"
# ✅ New syntax: express autonomy via sandbox mode
codex exec --sandbox workspace-write "fix all lint errors"2. The Hooks Engine Graduates to Stable
The interception layer from Ch.05 is now a first-class citizen. Configure SessionStart / Stop hooks inline in config.toml; hooks can observe MCP tool calls, apply_patch, and long-running Bash sessions:
# ~/.codex/config.toml
[hooks]
session_start = "bash scripts/bootstrap-env.sh"
stop = "npm run lint --silent"3. Agent Plugins and Marketplaces
The plugin system is now production-grade, with local, personal, workspace, and remote catalogs — including third-party marketplaces for Amazon Bedrock and Claude Code:
# Unified plugin management entry point
codex plugin list
codex plugin marketplace add https://plugins.example.com/catalog.json
codex plugin install security-workbenchMention a plugin in chat with @plugin to auto-inject its context (MCP / App / Skill).
4. Subagents and Multi-Agent Orchestration
The CLI natively spawns subagents with independent context windows, slicing large efforts into parallel pipelines:
> Spawn two subagents: one adds integration tests for src/api,
another refactors state management in src/hooks in parallel.
Aggregate the diffs for my review at the end.Combined with git worktree isolation, multiple agents work in parallel without polluting each other — the official realization of Ch.04's goal-driven philosophy.
5. Guardian Auto-Approval and --approve-for-me
Human confirmation is no longer the only path for high-risk operations. The new --approve-for-me flag routes approval requests to a Guardian subagent (powered by GPT-5.6 Luna): in-policy actions pass automatically; violations are blocked with reasons:
codex --approve-for-me "upgrade dependencies and make tests pass"This complements Ch.08's mobile approval gateway: Guardian handles the low-risk queue; only genuinely dangerous operations reach your phone.
6. The MCP 2026-07-28 Protocol
Paginated discovery, multi-round requests, and non-blocking server startup are now supported. Legacy .mcp.json configs remain compatible, but new MCP servers should target the 2026-07-28 spec.
13.4 The Skills Economy: Compounding Repeated Workflows into Assets
If AGENTS.md is a project's "constitutional law," Skills are the "standard operating procedures" for specific task types: PR review, Feishu doc grooming, PPT generation, CI repair, security scanning. Any workflow repeated three times or more deserves to become a Skill.
# Standard Skill layout (under ~/.codex/skills/ or inside a plugin package)
skills/
└── pr-review/
└── SKILL.md # a procedural manual with YAML frontmatterSkills ship with plugins to marketplaces for uniform team installation. This delivers on Ch.05's promise of cross-project reuse: the protocol constrains behavior; the Skill freezes the procedure.
13.5 Codex Security and Daybreak: Security Becomes a Product Line
In August 2026, OpenAI introduced the two-tier Daybreak access system for defenders:
- Daybreak Blue: general-purpose models (GPT-5.6 Sol) for vulnerability discovery, secure code review, detection engineering, incident response, malware analysis, and patch validation — start here for most defensive work.
- Daybreak Red: purpose-trained models (GPT-5.6 Cyber) for explicitly authorized vulnerability reproduction, exploit validation, penetration testing, and red teaming; requires separate approval and provisioning.
Together with the Codex Security plugin and CLI, security workflows are fully productized:
# Run a deep scan in the workbench
codex plugin install codex-security
# Or run bulk scans in CI
codex-security scan --deep --report sarif > results.sarifOperational red lines (continuous with Ch.05's sandbox boundaries):
- Work in an isolated environment with an explicitly defined engagement scope.
- Use least-privilege permission profiles.
- Configure Auto-review for actions that cross the sandbox boundary, so Guardian reviews them first.
13.6 Migration Checklist (TL;DR)
# ① Desktop: retire the old Codex App, install the ChatGPT client
winget install OpenAI.ChatGPT # Windows
# macOS: download ChatGPT.dmg
# ② Upgrade the CLI to 0.147.0+
npm install -g @openai/codex@latest
# ③ Switch models to the GPT-5.6 family (before August 31)
grep -rn "gpt-5.4" ~/.codex/ . # full audit
# ④ Replace the removed --full-auto flag
codex exec --sandbox workspace-write "<task>"
# ⑤ Enable Guardian auto-approval
codex --approve-for-me "<task>"
# ⑥ Install plugin marketplaces and the security plugin
codex plugin marketplace add <catalog-url>
codex plugin install codex-security13.7 The Constants That Survive Every Version
The tooling lesson is singular: every version number, CLI flag, and model codename hard-coded into scripts is technical debt. Centralize them in config.toml and AGENTS.md, so migration cost collapses to "edit one line."
And the three principles that outlive every release cycle are exactly what this book keeps hammering: drive by goals, not steps; set boundaries before granting trust; divide labor between human and AI — never let the AI lead the human. Tools change blood; your mindset compounds.