兰台簿 Erdoschn/dsh-lantai

兰台·簿 · dsh-lantai

一本账册：智能体学到的经验记一笔，记什么要审，过期要核销。

A ledger for what your agent learned: every lesson is written down, what gets
written has to pass inspection, and what goes stale gets retired.

装上只要一分钟，它能告诉你手上的技能库哪几条在误导你。

One minute to install, and it will tell you which skills in your library are
currently misleading you.

dsh plugin install dsh-lantai

然后问它一句「体检一下我的技能」：

7 条 skill，4 个主题，0 个严重、0 个中等问题、3 个提示 [low] 主题「windows」只有 2 条 → 知识缺口，也是下一步该学什么的方向

模型随手写一些，但没通过核验被打回

A model writes something offhand — and it is rejected outright

核验驳回，一字未写： [name-not-kebab] name "Git_Auth" 必须是 kebab-case
[description-vague] description 太笼统 [body-too-short] 正文少于 6 行，撑不起一份操作手册 [missing-verification] 缺少「验证」小节：没有验证方式的 skill 是谣言，不是经验 [missing-boundaries] 缺少「边界」小节：不说边界的经验会被当成万能的 [not-concrete] 正文里没有具体的命令、路径或字段名 [disable-guardrail] 不能建议关闭或绕过沙箱与审批 [push-to-default-branch] 不能建议直推默认分支；改为推送分支后开 PR

最后两条是安全问题：那条草稿在教人关掉沙箱、直推主分支。

The last two are safety findings: that draft was teaching people to switch off the
sandbox and push straight to main.

如果没有核验，它会入册，然后在往后每一个匹配的会话里生效——你不会再读它，但它一直影响智能体的判断。

Without the check it would have been filed away, and then taken effect in every
matching session from then on — you would never read it again, but it would keep steering the agent's judgement.

为什么需要核验

Why the check exists

一条技能会被注入到此后每一个匹配会话，成为智能体的默认前提。

A skill is injected into every matching session from then on, and becomes one of
the agent's standing assumptions.

所以判断标准不是「这条有没有用」，而是「它错了会怎样」。错的技能不像错的搜索结果——你不看就没影响；它会主动参与每一次判断，而且没人会再读一遍。

So the question is not "is this useful" but "what happens if it is wrong". A wrong
skill is not like a wrong search result — a search result you simply don't read. A skill takes part in every judgement, and nobody rereads it.

记住的东西如果错了、旧了、互相打架，比不记得更糟。

Something remembered that is wrong, stale, or contradicting itself is worse than
something forgotten.

三个动作，都在做减法

Three actions, all of them subtractive

体检（兰台·诊） —— 不管技能是你写的、别人给的还是从 GitHub 拉的，一次给出：矛盾、孤岛、薄覆盖、逾期、缺验。每条都带「该做什么」。只读，不改文件。

Health check (兰台·诊) — whether the skills are yours, someone else's, or pulled from GitHub, one call reports: contradictions, islands, thin coverage, overdue reviews, and missing verification. Every finding says what to do about it. Read-only; it changes no files.

图谱 —— 主题聚簇 + 可证实的实边，输出可直接嵌 README 的 Mermaid。边全从文本推出，所以导入的技能也能建图。

Graph (兰台·谱) — themes plus structural edges that can be proven from text,
emitted as JSON or as a Mermaid diagram you can drop into a README. The edges are all
inferred from the text, so imported skills get a graph too.

核验 —— 写入前先审，驳回时一字不写。规则四类：形状、实质、安全、唯一。

The check (核验) — everything is inspected before it is written; a rejection writes not a single byte. Four kinds of rule: shape, substance, safety, uniqueness.

勘程（阶梯化复核）

Review ladder (勘程)

勘程间隔由已勘次数决定。

The interval is decided by how many times the skill has been reviewed.

| 阶段 | 间隔 | 该做什么 |
| 0 | 7 天 | 轻勘：验证命令还通得过吗 |
| 1 | 14 天 | 复查 |
| 2 | 30 天 | 环境复核：版本、路径、开关变了没 |
| 3 | 60 天 | 价值复核：被更好的做法取代了吗 |
| 4 | 90 天 | 成本复核：命中率如何，还值得占上下文吗 |
| 5 | 180 天 | 终审：重写、收缩，或删除 |

| Stage | Interval | What to do |
| 0 | 7 days | Light check: does the verification command still pass? |
| 1 | 14 days | Re-check |
| 2 | 30 days | Environment: did versions, paths or flags move? |
| 3 | 60 days | Value: has a better approach replaced it? |
| 4 | 90 days | Cost: how often is it loaded, is it still worth the context? |
| 5 | 180 days | Final: rewrite, shrink, or delete |

逾期超过 180 天，诊录会建议归档——但不会自动删。

Past 180 days overdue, the health check recommends archiving — but it never deletes on its own.

归档（可撤回的删除）

Archive, not delete (归档)

skill_retire 先出示确认卡片，同意后归档，目录移到 skills.retired/<name>-<时间戳>/，留一份 RETIRED.md 记下原因与恢复方式。

skill_retire shows a confirmation card first; once you agree, the directory is
archived to skills.retired/<name>-<timestamp>/, with a RETIRED.md recording the
reason and how to restore it.

归档原因必填。 没有原因，连卡片都不会弹——你半年后看到它，得知道当初为什么。

The reason is mandatory. Without one, the card never appears — six months from now you need to know why this directory exists.

八个工具

Eight tools

| 工具 | 卡片名 | 干什么 |
| skill_health | 兰台·诊 | 诊录：矛盾、孤条、单薄、逾期、缺验 |
| skill_graph | 兰台·谱 | 主题聚簇 + 实边，输出 JSON / Mermaid |
| skill_distill | 兰台·录 | 过核验后入册 |
| skill_catalog | 兰台·索 | 列出目录、复核等级、下次复核时间 |
| skill_audit | 兰台·校 | 算到期，列出该跑的验证命令 |
| skill_attest | 兰台·勘 | 盖章，间隔自动拉长 |
| skill_retire | 兰台·削 | 确认后归档，可恢复 |
| skill_conflicts | 兰台·议 | 两两比对，只报可疑与证据 |

| Tool | Card name | What it does |
| skill_health | 兰台·诊 | Report: contradictions, islands, thin coverage, overdue, missing verification |
| skill_graph | 兰台·谱 | Themes + real edges, as JSON or Mermaid |
| skill_distill | 兰台·录 | Write only after the check passes |
| skill_catalog | 兰台·索 | List the library, review stage, next review date |
| skill_audit | 兰台·校 | Work out what is due and name the command to run |
| skill_attest | 兰台·勘 | Stamp a review; the interval grows automatically |
| skill_retire | 兰台·削 | Archive after confirmation, recoverable |
| skill_conflicts | 兰台·议 | Pairwise comparison; reports suspicions with evidence |

安全边界

Security boundary

| 兰台可控权限 | 无控制权 |
| agent 写一条没有验证方式的 skill | agent 用 shell 直接写 ~/.dsh/skills |
| 两条 skill 互相矛盾而无人察觉 | 蓄意绕过核验的 agent |
| skill 悄悄变成永远生效的默认前提 | 过期的 skill 继续被注入（只能靠勘程提醒） |

| What it does control | What it does not control |
| An agent writing a skill with no verification | An agent writing ~/.dsh/skills directly through a shell |
| Two skills contradicting each other unnoticed | An agent that deliberately routes around the check |
| 技能悄悄变成永久假设 | 过时技能仍在被注入（只有复核阶梯能提示） |

- 对比 sandbox 里的智能体权限更大——拥有写入工作区之外的权限。

安装

dsh plugin install dsh-lantai

或在配置档里加一行：

- id: lantai
name: 'dsh-lantai'
config:
skillRoot: '~/.dsh/skills'
allowReplace: false
nudge: true

和其它记忆方案的区别

| | 大多数 memory / skill 方案 | 兰台·簿 |
| 立场 | 记得越多越好 | 部分写入要求不被允许。写入的必定有效、无害。过时的应该淘汰。 |
| 写入 | 直接追加 | 过核验，不通过就驳回 |
| 过期 | 没有机制 | 勘程 |
| 删除 | 删了 | 归档 + 原因，可恢复 |
| 能否指出"这条skill现在在误导你" | 不能 | 体检的直接目的 |

License

MIT