English · 中文
I read old institutions as tested coordination systems — and rebuild them as AI agent systems that are routable, reviewable, and honest about what they know.
A Tang-dynasty edict passed through three departments before it became law: one drafted, one reviewed, one executed. Persian satrapies ran on federated delegation kept honest by a roving audit. Athenian councils rotated judgment so no single hand could capture it.
These are the original multi-agent systems — battle-tested answers to the problems we hit when wiring up AI: delegating authority, gating quality, escalating failure, distributing judgment, and keeping an inspectable record.
| Discipline | Tested pattern | What I build with it |
|---|---|---|
| History | Tang three-department review · Persian satrapies · Athenian rotation | quality gates · federated delegation · distributed judgment |
| Political science | governance modes & institutional constraints | agent topologies · authority boundaries · escalation paths |
| Behavioral economics | fast/slow thinking · bounded rationality | model routing · evaluator design · failure-mode analysis |
| Knowledge management | notes · links · indexes · evidence trails | durable memory · research archives · retrieval-ready context |
| Project | What it is |
|---|---|
| agos | An honesty-first Agent OS deck: telemetry, shadow model routing, memory workbench, homelab fleet. The rule is enforced by tests — the screen never says anything the data cannot prove. Humans keep authority; models propose. |
| civagent · ⭐ 40 | The thesis in its purest form: 57 historical governance systems compiled into multi-agent orchestration patterns. v6 adds a constitutional engine (VETO / IMPEACH / EDICT), regime-as-graph typing, and a tournament pipeline (Bradley–Terry ranking + bootstrap CI, ablations) to test whether institutions actually change outcomes. |
| FuguNano · ⭐ 27 | An open, training-free reimplementation of Sakana Fugu: a control plane that makes the agents you already have run as one auditable loop — plan, dispatch, review, repair, learn — with a verifier-aware selector and a desktop Studio. |
- grimoire-skill · ⭐ 8 — parse a PDF once, forge two artifacts: typed reading notes for Obsidian and a reusable skill pack.
- oral-interview-transcription-skill · ⭐ 13 — oral-history recordings → faithful edited transcripts, with an adversarial fidelity gate; data stays local.
- skill-librarian — lint and index a large local
SKILL.mdlibrary: duplicates, dead links, name collisions, weak descriptions. - paper-reading-workflow-zh — evidence-grounded Chinese paper-reading workflow (PDF → MinerU → Obsidian).
- cn-cc · mmteam-cc — route Claude Code work to Chinese model backends; multi-model agent teams with fan-out consensus.
- ScreenBar — a SwiftUI menu-bar app that turns Macs on your tailnet into second screens.
Coordination also means contributing well inside other people's institutions. Merged since mid-June 2026:
| Repository | ⭐ | Merged | Highlights |
|---|---|---|---|
| steipete/CodexBar | 21.4k | 11 | LongCat provider (#1697) · Doubao Agent Plan usage (#2496) · Safe Storage ACL reuse (#2528) · live rollouts no longer starve cost updates (#3314) |
| jackwener/OpenCLI | 29.3k | 8 | Xiaohongshu risk-control cooldown retry (#2207) · X profile recovery (#2193) · Eastmoney column fix (#2131) |
| Hmbown/Codewhale | 41.0k | 4 | session persists before stall recovery (#3285) · proxy-aware fetch (#3577) |
| steipete/oracle | 4.0k | 4 | completion proof for short captures (#293) · hidden-window prompt submission (#302) |
| blueberrycongee/wuu | 48 | 4 | background-command ownership (#195) · stable tool prefix (#194) |
| BigPizzaV3/CodexPlusPlus | 30.9k | 2 | versioned base URL kept (#1363) · graceful inject fallback (#1360) |
| hoangsonww/Claude-Code-Agent-Monitor | 998 | 1 | non-negative subagent durations (#171) |
- 2026-09 — agos goes public: execution hardening, approvals and fleet state (#1), acceptance tree and honest session/chat/console slices (#2); CodexBar cost-update and titlebar fixes (#3314, #3315); oral-interview skill adds a Word-style review pack and podcast genre (#1).
- 2026-08 — civagent R6–R14: regime editing API + online editor, an experimental-validity layer, runtime-graph diffs, CI with teeth (#31–#34); CodexBar Alibaba Personal plan support (#3098, #3128) and z.ai 7/30-day ranges (#2524); OpenCLI Xiaohongshu cooldown (#2207).
- 2026-07 — civagent v6 hardening line: constitutional engine, Express API, tournament statistics, ablation generator, hill-climbing loop (#17–#30); FuguNano selector router, SWE-bench driver, and FuguNano Studio desktop GUI (#6–#16); four oracle browser fixes, four wuu PRs, CodexPlusPlus fixes; shipped skill-librarian and paper-reading-workflow-zh.
- 2026-06 — cn-cc
/cn:teammulti-model fan-out (#5); Codewhale TUI fixes (#3284, #3285); OpenCLI Gemini/SMZDM fixes (#1967, #1968); ScreenBar released.
BicaMind — a voice-first AI companion in an open ear-clip earphone: the same obsession with trustworthy, on-demand knowledge, leaving the terminal. Applied ML work lives in udacity-masters-ai-projects; essays connecting history, social science and agent architecture are on the blog.
一道唐代政令要经三省才成法：中书拟、门下审、尚书行。波斯总督制靠联邦式授权，再用巡回审计制衡。雅典议事会让判断在众人手中轮转，没有任何一只手能独揽。
这些是最早的多智能体系统——对我们接 AI 时遇到的同一批问题给出过实战验证的答案：如何授权、如何把质量关、如何升级失败、如何分散判断、如何留下可审查的记录。
| 学科 | 被验证的模式 | 我用它造什么 |
|---|---|---|
| 历史 | 唐三省审核 · 波斯总督制 · 雅典轮值 | 质量门 · 联邦式授权 · 分散判断 |
| 政治学 | 治理模式与制度约束 | 智能体拓扑 · 权限边界 · 升级路径 |
| 行为经济学 | 快慢思考 · 有限理性 | 模型路由 · 评估器设计 · 失败模式分析 |
| 知识管理 | 笔记 · 链接 · 索引 · 证据链 | 持久记忆 · 研究档案 · 可检索上下文 |
| 项目 | 是什么 |
|---|---|
| agos | 诚实优先的 Agent OS 控制台：遥测、影子模型路由、记忆工作台、homelab 机群。铁律由测试强制执行——屏幕上不出现任何数据证明不了的话。人保留决定权，模型只提议。 |
| civagent · ⭐ 40 | 主线最纯粹的化身：把 57 套历史治理制度编译成多智能体编排模式。v6 加入宪政引擎（VETO / IMPEACH / EDICT）、政体拓扑图类型化，以及锦标赛统计管线（Bradley–Terry 排名 + bootstrap 置信区间、消融实验），检验制度是否真的改变结果。 |
| FuguNano · ⭐ 27 | Sakana Fugu 的开源、免训练复现：让你手上已有的 agent 跑成一个可审计的闭环——规划、分派、审查、修复、学习——配有验证感知的选择器与桌面端 Studio。 |
- grimoire-skill · ⭐ 8 — 一次解析 PDF，炼出两件器物：落进 Obsidian 的类型化阅读笔记 + 可复用技能包。
- oral-interview-transcription-skill · ⭐ 13 — 口述史录音 → 忠实整理稿，带对抗式忠实度质量门，数据全程本地。
- skill-librarian — 给大型本地
SKILL.md技能库做体检与索引：重复、死链、重名、弱描述。 - paper-reading-workflow-zh — 证据落地的中文论文精读流程（PDF → MinerU → Obsidian）。
- cn-cc · mmteam-cc — 把 Claude Code 任务路由到国产模型后端；多模型 agent 团队与扇出共识。
- ScreenBar — SwiftUI 菜单栏应用，把 tailnet 里的 Mac 变成第二块屏幕。
协调也意味着在别人的「制度」里好好贡献。2026 年 6 月中旬以来已合并：
| 仓库 | ⭐ | 合并数 | 代表 PR |
|---|---|---|---|
| steipete/CodexBar | 2.1万 | 11 | LongCat 用量源（#1697）· 豆包 Agent Plan 用量（#2496）· 复用已授权 Safe Storage ACL（#2528）· 实时 rollout 不再饿死费用更新（#3314） |
| jackwener/OpenCLI | 2.9万 | 8 | 小红书风控冷却重试（#2207）· X 主页数据恢复（#2193）· 东方财富列名修正（#2131） |
| Hmbown/Codewhale | 4.1万 | 4 | 卡死恢复前先持久化会话（#3285）· fetch 遵守代理环境变量（#3577） |
| steipete/oracle | 4.0k | 4 | 短捕获需完成证明（#293）· 窗口隐藏时恢复提交（#302） |
| blueberrycongee/wuu | 48 | 4 | 后台命令归属记录（#195）· 稳定工具前缀（#194） |
| BigPizzaV3/CodexPlusPlus | 3.1万 | 2 | 保留带版本号的 base URL（#1363）· 注入模块缺失时优雅降级（#1360） |
| hoangsonww/Claude-Code-Agent-Monitor | 998 | 1 | 子代理平均时长不再为负（#171） |
- 2026-09 — agos 公开：执行加固、审批与机群状态（#1），验收树对齐并落地会话/聊天/控制台诚实片（#2）；CodexBar 费用更新与标题栏修复（#3314、#3315）；口述史技能新增逐段 Word 式审阅包与播客文类（#1）。
- 2026-08 — civagent R6–R14：政体编辑 API + 在线编辑器、实验效度层、运行时拓扑 diff、真正卡人的 CI（#31–#34）；CodexBar 支持阿里云个人版套餐（#3098、#3128）与 z.ai 7/30 天图表（#2524）；OpenCLI 小红书风控冷却（#2207）。
- 2026-07 — civagent v6 加固线：宪政引擎、Express API、锦标赛统计、消融生成器、爬山循环（#17–#30）；FuguNano 选择器路由、SWE-bench 驱动与 FuguNano Studio 桌面端（#6–#16）；oracle 四个浏览器修复、wuu 四个 PR、CodexPlusPlus 修复；发布 skill-librarian 与 paper-reading-workflow-zh。
- 2026-06 — cn-cc
/cn:team多模型扇出（#5）；Codewhale TUI 修复（#3284、#3285）；OpenCLI Gemini/什么值得买修复（#1967、#1968）；发布 ScreenBar。
BicaMind — 一款声音优先的 AI 伴侣，做成开放式耳夹耳机：同一种「让知识随取随用且值得信赖」的执念，走出终端。应用 ML 作品收在 udacity-masters-ai-projects；连接历史、社会科学与 agent 架构的随笔在博客。
- GitHub: @LeoLin990405
- Blog · 博客: leolin990405.github.io/blog
- LinkedIn: Zhongyue Lin
If one of my projects is useful, a star is always appreciated. · 如果哪个项目帮到了你，欢迎点个 star。