English | 中文
Online reading address: https://caldalis.github.io/awesome-coding-agents
A source-level study of how coding agents perceive, decide, act, and verify.
This repository is a technical documentation project about modern coding-agent systems. It studies two local source trees:
codex/: OpenAI Codex CLI source used as the Codex reference.claw-code/: a public Rust implementation used here as a Claude-Code-like comparison reference.
The goal is not to document product usage. The goal is to explain the runtime architecture behind coding agents: how they build context, call models, expose tools, classify permissions, edit files, verify work, compact history, and delegate tasks.
English chapters live in docs/. The Chinese version lives in docs/zh/. The
Astro site in web/ renders both languages.
The chapters are organized from the outer shape of a coding agent toward its inner execution mechanics:
| # | Chapter | Focus |
|---|---|---|
| 01 | Overview: What Makes a Coding Agent | Agent boundaries, runtime loops, and the core design space |
| 02 | Runtime Architecture | CLI, runtime, app server, configuration, plugins, MCP, skills, and UI layers |
| 03 | Agent Loop and Turn Execution | How a turn moves from user input to model stream, tool calls, observations, and follow-up turns |
| 04 | Tool System and Orchestration | Built-in tools, plugin tools, MCP tools, skills, routing, authorization, and execution |
| 05 | Code Search and Discovery | File discovery, text search, evidence ranking, and investigation narrowing |
| 06 | File Editing and Patch Application | Patch application, exact replacement, whole-file writes, and edit safety |
| 07 | Sandboxing and Process Security | Filesystem boundaries, network boundaries, process isolation, and defense in depth |
| 08 | Permissions and Approval Flow | Approval policies, command classification, denials, prompts, and user control |
| 09 | Context, History, and Compaction | Conversation history, token pressure, compaction, persistence, and resume behavior |
| 10 | Prompt Construction and Project Memory | Base instructions, project memory, dynamic context, and cache-aware prompt assembly |
| 11 | Model Clients, Streaming, and Caching | Provider adapters, streaming events, retries, usage tracking, and prompt caching |
| 12 | Sub-Agents and Delegation | Agent spawning, role scope, context inheritance, messaging, waiting, and cleanup |
Two longer companion essays are included:
.
├── docs/ # English source chapters
├── docs/zh/ # Chinese source chapters
├── web/ # Astro + React site wrapper
└── README_zh.md # Chinese project README
Run the site from web/:
npm ci
npm run devBuild the static site:
npm run buildPreview the production build:
npm run previewThe site is generated from the Markdown files in docs/ and docs/zh/. Chapter
metadata and route mapping are maintained in web/src/content/chapters.ts.
Claw Code is used as a public Claude-style reference implementation. It should not be described as Anthropic's closed-source Claude Code product.
Useful contributions include:
- correcting source-level analysis,
- improving diagrams or explanations,
- expanding the Chinese translation,
- adding focused deep dives,
- updating the docs for newer source versions.
Before submitting changes to the site, run:
cd web
npm run buildMIT