Codex CLI Agentic Engineering from First Principles (Daniel Vaughan)（Z-Library）

【One-Line Pitch】 A practical field guide to working with OpenAI's Codex CLI as a real engineering tool rather than a novelty: it builds mental models for prompting, configuration, safety, and orchestration, then scales them to teams. Best for developers and engineering leads who already write code and want to multiply output without losing control of the repo. 【Book Arc】 - **Opening (~0%–13%)**: Frames the shift from autocomplete to agentic coding, situates Codex CLI in the 2026 tool landscape, and gets you installed, authenticated, and running a first session across its interfaces. - **Early (~13%–25%)**: Builds the working fundamentals — effective prompting, AGENTS.md as durable project context, approval modes and trust boundaries, and how to diagnose agent failures from session transcripts. - **Early–Middle (~25%–38%)**: Extends the agent's reach through model/reasoning-effort selection and MCP servers, including connecting common services, building simple servers, and enterprise authentication and governance. - **Middle (~38%–46%)**: Covers lifecycle control and scale — hooks that intercept agent events, the skills ecosystem and plugin packaging, and context-window management with compaction and memory. - **Middle–Late (~46%–54%)**: Moves into orchestration — sub-agents, parallel execution, multi-agent patterns, and Git worktrees for isolated execution. - **Late–Ending (~54% onward)**: The excerpts thin out here; the blurb indicates CI/CD integration, security hardening, and enterprise deployment, but the excerpts do not cover those chapters in detail. 【Book Arc】 (reading-order map, continued as bullets) - **Opening (~0%–13%)**: what Codex CLI is, why the terminal matters, and first-run setup. - **Early (~13%–25%)**: prompting, AGENTS.md, approvals/sandboxing, debugging. - **Middle (~25%–46%)**: MCP, hooks, skills, context management. - **Late (~46%–54%)**: sub-agents, orchestration patterns, worktrees. - **Ending (~54%+)**: CI/CD, security, enterprise rollout (per blurb; excerpts do not cover). 【Key Takeaways】 - **Agentic coding is a different discipline from autocomplete** (Opening): the book opens by arguing that the real shift is task delegation and supervision, not faster typing — which is why prompting, trust, and context get whole chapters. - **AGENTS.md is the backbone of reliable behavior** (Early): durable project context (commits, testing, style, file maps) belongs in configuration, not repeated prompts; the book treats common failure modes and monorepo patterns as first-class. - **Trust is a dial, not a switch** (Early): four approval modes plus filesystem/network sandboxing let you match autonomy to risk, with team-level strategies and hook-based programmatic approval for stricter control. - **Failures are diagnosable** (Early): session transcripts, approval mode as a diagnostic, context-overflow symptoms, and structured logging turn "the agent went weird" into a checklist you can run. - **Context is the scarce resource** (Middle): quadratic growth, compaction, thread resume/fork, sub-agent delegation, and memory pipelines are framed as engineering problems with architectures and failure taxonomies. - **Extensibility has layers** (Middle): MCP servers, hooks, skills, and plugins form a customization stack — start with read-only integrations, then add write-back only behind explicit safety boundaries. - **Parallelism must be earned** (Late): sub-agents and multi-agent patterns pay off for decomposable work, but the book is explicit about when not to parallelize and which orchestration anti-patterns to avoid. - **Isolation prevents collisions** (Late): the "one agent, one worktree" principle keeps concurrent agents from stepping on each other's working tree. 【Reading Tips】 - Read Chapters 1–2 and 7–9 closely before touching anything else; prompting, AGENTS.md, and approval modes are the load-bearing skills everything later assumes. - Skim the competing-tools chapter (Ch. 5) unless you're choosing a stack; its value is the decision framework, not the product tour. - Treat MCP, hooks, and skills as a menu: pick the one integration your workflow actually needs and build it end-to-end rather than reading all three superficially. - Expect version churn — the book cites specific releases and PR numbers, so verify flags against your installed version; the author flags version currency himself. - Do the exercises in the context and orchestration chapters; those concepts (compaction, fan-out, worktrees) only click once you've hit a real failure. 【Coverage Limits】 The excerpts are dominated by the table of contents and early-to-middle chapters; CI/CD, security hardening, and enterprise deployment are named in the blurb but not detailed here, so this guide's late-book picture is thinner than the front half.