Not a sales pitch. If you search for OpenAI Codex App or CLI tutorials on Dev.to or Reddit, you'll find almost nothing besides basic "look, it generated a generic HTML template" posts. There is a massive gap between running simple chat queries in a browser and orchestrating a local autonomous agent to handle complex, multi-file codebases while maintaining a clean Git history.

To bridge this gap, we spent weeks documenting every single configuration setting, CLI flag, repository guideline, and case study into a comprehensive developer manual: The Codex Orange Book.

To prove this isn't another generic, AI-generated guide, here is a sneak peek of the actual configuration blueprints, repository manifests, and local model routing integrations we use.

1. The Sandbox & CLI Setup

By default, letting an autonomous agent (running ChatGPT 5.5 under the hood) run terminal commands and modify local files on your development machine is a massive risk. If you don't bound its scope, it can run commands that alter files outside your workspace.

We recommend running the CLI with strict sandboxing and manual command approvals:

codex --sandbox workspace-write --ask-for-approval on-request

• workspace-write restricts file modifications strictly to the current directory.
• on-request forces the agent to prompt you before executing any bash script or compiler command. You remain the final gatekeeper.

2. The AGENTS.md & .coderules.json Manifest

ChatGPT 5.5 has a tendency to over-engineer simple tasks. If you ask for a basic UI tweak, it wants to write abstract classes, custom interfaces, or install random npm packages.

We solve this by placing a localized manifest named AGENTS.md in our repository root. Codex reads this on startup to align with the specific "house rules" of your repository.

Here is the exact template we use for production Vite projects:

# Project Manifest & Agent Rules

## Coding Principles
- If 200 lines of code can be written in 50, rewrite it.
- Keep components focused and reusable. Use Vanilla CSS/custom properties.
- Do not create abstractions, interfaces, or factory patterns for one-off code.
- Read related code before modifying — understand context first.
- Every line changed must trace directly to the user's request.

## Stack & Scripts
- Stack: HTML5 + JavaScript (ES6+) + CSS3
- Build tool: Vite (run `npm run dev` to start, `npm run build` to verify compile)
- Formatting: Prettier/ESLint (run `npm run lint` before committing)

## Off-Limits
- Do not modify configuration pipelines in `.github/` unless explicitly asked.
- Never hardcode API keys or credentials. Use `process.env`.
- Do not touch styling design tokens in `index.css` without approval.

To enforce this programmatically in our commercial pipelines, we pair the manifest with a machine-readable .coderules.json config file:

{
  "rules_manifest": "AGENTS.md",
  "concurrency_limit": 1,
  "enforce_standards": true,
  "verification_pipeline": {
    "on_edit": "npm run lint",
    "on_complete": "npm run build"
  }
}

3. Local Model Routing via CC Switch (DeepSeek Integration)

For developers looking to cut down API token costs, the manual includes a step-by-step setup in Appendix A using CC Switch (a third-party proxy tool) to route Codex queries to the DeepSeek engine.

Instead of default OpenAI models, you can proxy commands locally through specific ports, dropping token expenses significantly while keeping code reasoning incredibly sharp.

4. Real-World Case Studies: From Scratch to Staging

We don't teach coding in a vacuum. The guide is built around 5 practical case studies:

• Case Studies 1–3: Scaffolding and building a complete, responsive e-commerce web app (Pet Treats store) using Vite, structuring category layouts, adding product cards, and setting up secure checkout validation flows.
• Case Study 4: Deploying PPT presentation Skills to generate investor pitch decks directly from raw Markdown.
• Case Study 5: Integrating motion graphic automation plugins to render promotional video clips.

Every step in these case studies is backed by detailed, step-by-step screenshots (over 100+ annotated diagrams in the full guide) tracking terminal commands, browser previews, editor diffs, and Git commits, making it incredibly friendly for beginners and solo developers alike.

Get the Field Manual

We've packaged this entire manual, complete with all config templates, CLI cheat sheets, local model routing setups, and case study workspace assets into a comprehensive handbook: The Codex Orange Book Field Manual.

If you want a highly visual, battle-tested guide to master ChatGPT 5.5 / 5.4 agent workflows:

Get the Field Manual on Gumroad: [https://jyhcraft4.gumroad.com/l/gcqko]

How are you guys structuring your local agent setups? Let's discuss in the comments.