本教程的代码不是 24 个孤立 demo，而是一条围绕桌面 Agent Harness 展开的渐进学习路径：每章继承上一章的稳定边界，只新增一个主要机制。tests/test_project_structure.py 会检查每个 code.py 都声明 PROGRESSION 元数据。
- 每章
code.py必须定义PROGRESSION。 chapter必须等于目录名。- 除 s01 外，每章
builds_on必须指向前一章。 adds和preserves必须非空。- s24 必须把前面机制收束为一个完整 harness。
| 章节 | 继承 | 本章新增 | 保留不变 |
|---|---|---|---|
s01_agent_loop |
起点 | minimal agent loop single bash tool tool_use/tool_result feedback |
interactive CLI |
s02_tool_dispatch |
s01_agent_loop | tool dispatch map read/write/edit/glob tools workspace path guard |
same agent loop shape |
s03_deferred_loading |
s02_tool_dispatch | compact deferred tool directory deterministic tool discovery session-scoped schema loading |
single-source tool registry dispatch boundary |
s04_permission_hooks |
s03_deferred_loading | allow/ask/deny decisions workspace path scope separate user approval auditable execution outcomes |
multi-tool execution boundary hook lifecycle |
s05_electron_shell |
s04_permission_hooks | main/renderer/preload split IPC bridge process isolation |
agent request boundary |
s06_sidecar_server |
s05_electron_shell | sidecar control plane JSON-RPC routing ring buffer logs |
desktop process boundary |
s07_session_management |
s06_sidecar_server | logical session and runtime separation create/resume/close lifecycle ACP-like HTTP boundary |
sidecar-managed runtime |
s08_model_routing |
s07_session_management | lite/default/craft routing cost tracking agent-to-model mapping |
session runtime context |
s09_jsonl_transcript |
s08_model_routing | sequenced transcript evidence derived replay state partial-tail recovery |
model turn event shape |
s10_workspace_memory |
s09_jsonl_transcript | workspace-scoped fact log policy-driven memory distillation atomic curated memory view |
append-only evidence and restart recovery |
s11_user_memory |
s10_workspace_memory | user-level memory preference dedupe identity prompt blocks |
workspace memory layer |
s12_cloud_memory |
s11_user_memory | remote profile injection history recall tool memory selector |
three-layer memory model |
s13_output_externalization |
s12_cloud_memory | large output threshold tool-results swap files page-fault reads |
context budget mindset |
s14_context_compact |
s13_output_externalization | token pressure detection structured compaction summary preservation |
externalized output pointers |
s15_prompt_assembly |
s14_context_compact | runtime prompt segments budgeted context blocks assembly order |
memory and compaction inputs |
s16_skills_system |
s15_prompt_assembly | SKILL.md discovery frontmatter parsing on-demand skill loading |
prompt assembly pipeline |
s17_mcp_connectors |
s16_skills_system | connector config trust workflow MCP tool namespace |
lazy capability loading |
s18_experts_system |
s17_mcp_connectors | expert packages expert prompt injection session-level expert state |
external capability model |
s19_visualizer |
s18_experts_system | visualizer protocol SVG/HTML widget generation theme-aware output |
specialized output routing |
s20_result_presentation |
s19_visualizer | present_files flow artifact cards deliverable prioritization |
visual output artifacts |
s21_sqlite_database |
s20_result_presentation | SQLite WAL database session metadata usage tracking |
deliverable and session persistence |
s22_automation_scheduler |
s21_sqlite_database | RRULE scheduling automation run history runtime state table |
SQLite persistence layer |
s23_audit_sandbox |
s22_automation_scheduler | hash-chain audit log command safety classifier sandbox policy |
scheduled autonomous execution boundary |
s24_comprehensive |
s23_audit_sandbox | integrated mini harness end-to-end agent pipeline all-layer wiring |
all previous chapter mechanisms |
先看 PROGRESSION["adds"]，再搜索源码里的 NEW in sXX、FROM sXX、LAYER 注释。这样读者能区分：哪些是上一章留下来的 harness 骨架，哪些是本章新加的机制。