@DanKornas: Want to understand Claude Code? Study the harness, not just the prompts. claude-code-from-scratch is a Python learning …

Want to understand Claude Code? Study the harness, not just the prompts. claude-code-from-scratch is a Python learning repo that reverse-engineers Claude Code-style agent architecture through 23 incremental sessions. It helps you build a practical mental model for tool-using agents by walking from a minimal perception/action loop into planning, subagents, context management, permissions, streaming, MCP, and worktree-based task isolation. Key features: • Incremental 23-session path – starts with the basic agent loop and adds one mechanism per file • Shared core foundation – http://core.py centralizes model config, tools, dispatch, permissions, and loop helpers • Agent building blocks – covers tool dispatch, TodoWrite-style planning, subagents, skills, context compaction, and task graphs • Runtime + governance examples – includes background tasks, streaming output, YAML permissions, lifecycle hooks, and session resume/fork • Advanced integration patterns – adds parallel tools, interrupts, prompt-cache tracking, MCP runtime, Redis mailboxes, and git worktree lifecycle management It’s open-source (MIT license). Link in the reply