learn-claude-code: Build a Claude Code-style nano agent harness from scratch

The project argues that agency comes from model training, while the product equals model plus harness; the harness consists of tools, knowledge, observation, action interfaces, and permissions. The tutorial progresses through 17 chapters from s01 to s17: s01 single loop plus Bash, s02 tool dispatch, s03 permission boundaries, s04 hooks, s05 planning, s06 sub-agents, s07 on-demand skill loading, s08 context compression, s10 task system, s11 background tasks, s13 agent teams, s14 memory, s15 integrated runtime, s16 workflow orchestration, s17 goal loop. Each chapter includes an English README, Chinese/Japanese translations, runnable code.py, and diagrams; the older 12-chapter version is kept under docs/ and agents/, plus a mapping table between old and new chapters. For developers who want to implement an agent loop themselves, the value is that it fixes the main loop and externalizes every other mechanism: adding a tool only requires registering a handler, while permissions, hooks, memory, and MCP routing never touch the core loop. Starting at s01 and adding mechanisms chapter by chapter yields a runnable minimal harness; note that the old and new chapter numbers don't correspond, so don't mix them when migrating.