# OpenAI Codex: The Complete Guide (Orange Book): what the repo actually ships > The Orange Book series entry on OpenAI Codex is a book repository, not a tool. It ships PDF, EPUB and single-file HTML editions in English and Chinese, and its download is the whole product. **alchaincyf/codex-orange-book** — OpenAI Codex: The Complete Guide · 橙皮书系列 · GPT-5.5 时代的 AI 编程实战手册 - Repository: https://github.com/alchaincyf/codex-orange-book - Website: https://www.huasheng.ai/orange-books/codex - Stars: 662 · Forks: 59 - Language: HTML - License: not declared - Published: 2026-09-14 · Updated: 2026-09-14 · Language: en - Canonical page: https://hysenlabs.com/projects/alchaincyf-codex-orange-book ## A book repository, not a program The first thing to get straight is that `alchaincyf/codex-orange-book` has nothing to execute. The top level of the repository holds six book files (PDF, EPUB and single-file HTML in English and Chinese), two READMEs, two banner images, and a launch film in GIF and MP4. There is no package manifest, no build script, no source tree in a compiled language. The primary language listed is HTML, which reflects the single-file HTML editions rather than any application. So the problem it solves is not a technical one. It is the problem of a fast-moving tool with a scattered surface: OpenAI Codex exists as a CLI, a Desktop App, a Cloud service, an IDE Extension and a Chrome Extension, according to the README, plus a mobile companion shipped on May 14, 2026. Documentation for all of that lives in different places and changes monthly. This repository is one author's attempt to freeze a coherent account of it into a file you can download and read offline. Who it is for is stated plainly. The README lists engineers who want to hand implementation to an agent, Claude Code users evaluating Codex, and product managers or founders shipping MVPs alone. It also claims no prior AI programming experience is needed. That last claim is the one worth testing against your own patience, because a book that covers five product surfaces in ten chapters is not a gentle ramp. ## How the book is structured and versioned The README lays out four parts plus appendices. Foundations covers §01 to §03: the five forms of Codex, a ten-minute install, and a first project. Daily Workflow covers §04 to §06: a CLI deep dive with Auto-review and sandboxing, AGENTS.md, and the Desktop App with Computer Use, what the README calls a five-piece suite, and the Mobile Companion. Beyond the Local Machine covers §07 to §08: Cloud and the Chrome extension, then Skills, MCP, Automations and the `/goal` system, described as four extension layers. Building Real Things covers §09 to §10: idea to shipped product, and a dual-tool mental model comparing Codex with Claude Code. Appendices A to C hold a command reference, pricing, and an FAQ. Versioning is explicit and dated. v2.0.0 landed on May 14, 2026 as a major revision, and v2.0.1 on May 15, 2026 as an addendum that added a coda on Codex on mobile and shipped the English edition. The file names carry the version, so `Codex-Complete-Guide-en-v2.0.1.pdf` tells you exactly what you have. That naming discipline is a small thing that matters: when the underlying product changes, you can tell at a glance whether your copy is stale. The changelog entries are also unusually specific about what changed in Codex itself, not just in the book. v2.0.0 records GPT-5.5 as the default model, the addition of Codex for Chrome, changes to the approval model, three Pro tiers replacing two, and the persistent `/goal` system. Whether those claims are accurate is something the repository cannot prove; it is the author's account, and the README gives no citations to OpenAI's own release notes. ## Downloading the Codex Orange Book in PDF, EPUB or HTML There is no install step because there is no software. Getting started means fetching a file from the repository. The README recommends the PDF over GitHub's online preview, which it says may not render properly. If you prefer the command line, clone and pull the edition you want. This is a plain git operation against the default branch. ```bash git clone https://github.com/alchaincyf/codex-orange-book.git cd codex-orange-book ls Codex-Complete-Guide-* ``` The `ls` should list six files: English and Chinese PDF, EPUB and HTML, all at v2.0.1. The README gives the sizes, so you can sanity-check that a download completed: the English PDF is 1.7 MB, the English EPUB 5.6 MB, the English single-file HTML 232 KB. The Chinese PDF is 4.2 MB and its HTML 220 KB. If you want to read it in a browser instead, open the single-file HTML directly. It is self-contained at 232 KB, so no server is required. ```bash open Codex-Complete-Guide-en-v2.0.1.html ``` On Linux, substitute `xdg-open` for `open`. The README also points to a WorkBuddy knowledge base where the book is published one page per chapter, and suggests dropping that link into an agent as context. That is the closest thing to an interactive use of this repository, and it is worth noting that it depends on a third-party site rather than anything in the repository itself. ## The rough edges the book admits to The README has a section titled Honest About the Rough Edges, and it is the most useful part of the repository for anyone deciding whether to trust the rest. It lists five things the book calls out rather than smooths over. On benchmarks, it states that GPT-5.5 still loses to Claude Opus 4.7 on SWE-Bench Pro, 58.6% against 64.3%. On context, it states that the advertised 1M context in Codex is actually about 258K usable. On safety, it reports that Full Access mode on Windows has deleted user files, citing 370 GB, 700 GB and 240 GB reports, and says never to enable it. On stability, it names three unfixed GitHub issues in MultiAgentV2 (#16657, #17523, #14233) as of the v2 release. And it warns that Codex on mobile is the Desktop App's remote companion, not a standalone phone agent. Those are strong claims with no linked sources in the README. The benchmark numbers and the file-deletion reports are the kind of thing you would want to verify against the original issue trackers before repeating them. The value here is directional: the author is telling you where the sharp edges are, and a book that names a losing benchmark against its own subject is at least not marketing. The real limitation of the repository is different from any of those. It is a snapshot. The last push was on 2026-08-23, and the book content is v2.0.1 from May 2026. Anything Codex changed after that is not in here, and the README does not document a process for updating the text between releases. ## Where it sits against the official Codex docs and the Claude Code book The obvious alternative is OpenAI's own Codex documentation. The difference in approach is authorship and intent. Official docs describe the current state of the product, are maintained by the vendor, and are authoritative when they conflict with anything else. This book describes one practitioner's experience of the product at a point in time, includes a comparison chapter against a competing tool, and is written to be read end to end rather than searched. If you need to know what a flag does today, the official docs win. If you want a narrative of how someone with no hand-written code shipped products with an agent, this is the shape of document that provides it. The second alternative is the sibling Orange Book on Claude Code, which the README lists first in the series. Same author, same format, same licence, different subject. The interesting overlap is §10, the dual-tool mental model, which the README describes as an honest, data-backed comparison of when to use Codex and when to use Claude Code. If you already own the Claude Code volume, §10 is the chapter that justifies looking at this one, because it is the only part written specifically for people holding both tools. A third comparison worth naming: this is the eighth book in the series, and the README lists the other seven. That matters for expectations. You are not evaluating a one-off; you are evaluating an entry in a consistent line of practical guides, which means the format, the tone and the licence are stable across the set. ## Licence, reuse and what it rules out The repository is licensed CC BY-NC-SA 4.0, stated in the README with a link to the licence deed. The README summarises it as free to share and adapt for non-commercial purposes with attribution. The GitHub metadata does not report a licence identifier, so the README is the source of record here. Three practical consequences follow, and I am describing the licence terms rather than advising on them. First, the NonCommercial clause means internal commercial training use is a question you would need to resolve with the licence text itself, not with this article. Second, ShareAlike means a translated or adapted edition carries the same licence, which constrains how you can republish it. Third, attribution is required, so redistributing the PDF inside a company wiki without the author's name attached is outside the terms as summarised. On upgrade cost, the repository is cheap to maintain and expensive to keep current. Cheap because there is no dependency graph: the artefacts are six static files, and a new version means new files with a new version string, as happened between v2.0.0 and v2.0.1. Expensive because the subject moves fast. The README itself frames v2.0.0 as a major revision after one month of Codex changes, which tells you the useful lifetime of a given edition is measured in weeks, not years. Budget for re-reading, not for patching. ## Conclusion Adopt this if you want a single downloadable document on Codex's five forms, its CLI approval model and its pricing tiers, and you accept that it is one author's field notes rather than vendor documentation. Do not adopt it if you need a tool to install, if you need a permissively licensed text you can rebrand or sell, or if you need claims traced to OpenAI's own changelog. Before you spend an afternoon on it, verify three things: that the version you download is v2.0.1, that the English edition reads as native prose rather than a literal translation, and that the rough-edges section still matches the Codex build you actually run. The licence is the boundary that decides most commercial use cases, and nothing in the repository changes that. ## FAQ ### What is the purpose of the Orange Book series? The README describes the Orange Book series as practical, hands-on guides for AI-native builders, and this Codex volume is the eighth book in it. The series includes guides on Claude Code, Harness Engineering, Agent Skills and others. ### Is the Codex Orange Book free to access? Yes. The repository ships the PDF, EPUB and single-file HTML editions directly, and the README also links to a WorkBuddy knowledge base where it is published one page per chapter. The work is licensed CC BY-NC-SA 4.0, so sharing and adaptation are permitted for non-commercial purposes with attribution. ### What is the purpose of the codex-orange-book repository? It distributes a book about OpenAI Codex. The top level holds six book files in English and Chinese plus READMEs, banners and a launch film, with no executable code. ### Which Codex Orange Book formats are available for download? PDF, EPUB and single-file HTML, in both English and Chinese, all at version v2.0.1. The README recommends the PDF because GitHub's online preview may not render properly. ### Which version of the Codex Orange Book should I download? The README presents v2.0.1, dated May 15, 2026, as the current release. The version string is part of every file name, so you can confirm which one you have. ## Sources - [alchaincyf/codex-orange-book on GitHub](https://github.com/alchaincyf/codex-orange-book) - [Issues](https://github.com/alchaincyf/codex-orange-book/issues) - [Project website](https://www.huasheng.ai/orange-books/codex) - [README](https://github.com/alchaincyf/codex-orange-book/blob/main/README.md) --- Hysen Labs editorial analysis, written from the project's own repository and release notes. Cite the canonical page: https://hysenlabs.com/projects/alchaincyf-codex-orange-book