A hand-written, fully-commented claude code clone you can actually read

What it does

CodeAgent is a terminal AI coding assistant: you give it natural-language instructions and it reads code, edits files, runs commands, and iterates until the job is done. It supports OpenAI-compatible and Anthropic endpoints (DeepSeek by default), with a cheap auxiliary model handling compression, retrieval, and memory extraction. The pitch is explicit: this is a from-scratch, line-by-line implementation of the claude code experience — not a wrapper — with every module annotated in plain Chinese so the code reads like a tutorial.

The interesting bit

The self-learning pipeline is the unusual part: it observes conversation trajectories and distills reusable skills from them, activated by conditions like touched file paths. Combined with tiered context compression (L1 light trimming to L4 deep summarization) and cache-prefix-friendly message injection, it’s a working lab for context engineering.

Key highlights

• Tool self-registration: drop a module in tools/ and it registers; toolsets control what the model sees
• Approval panels for dangerous commands, with injection-shape detection ($(), process substitution) escalated to review
• Subagents with worktree isolation, team mode with a message bus, and a deterministic workflow engine with spend caps
• Long-term memory via MEMORY.md + JSONL, retrieved results injected temporarily without polluting history
• MCP support via project-level .mcp.json, plus secret scanning, SSRF guards, and optional bwrap sandbox on Linux

Caveats

• Windows-first: Linux runs (with sandbox support) but macOS is described as theoretically usable, not thoroughly tested
• Comments and docs are Chinese-first; the English README exists but the code commentary is the main teaching material
• At ~500 stars, it’s a young project — the author invites bug-fix PRs, which tells you something

Verdict

If you want to understand how a real coding agent works from the inside — the loop, the compression, the permission model — this is a genuinely readable specimen. If you just want a polished production assistant, the established options have more mileage.