Wikisage

A Karpathy-style persistent LLM wiki. Use when: (1) user says '加进wiki/ingest/摄入', (2) user says '查wiki/wiki里有没有', (3) user says '整理wiki/lint', (4) answering...

Security Scan

The skill is largely consistent with a local LLM-managed markdown wiki, but there are mismatches and optional components (AWS/embedding/Secrets Manager) and environment expectations that are not reflected in the registry metadata — review before installing or granting credentials.

Purpose & Capability

The name/description (a persistent LLM wiki) matches the code and instructions: reading/writing markdown under a WIKI_ROOT, ingest/query/lint flows, and scripts for dedup/ lint/embedding exist. However, the skill includes an optional embed.py that uses AWS (Bedrock + Secrets Manager + OpenSearch) and mentions additional env vars (AWS_REGION, WIKI_EMBED_SECRET, WIKI_WORKSPACE) even though the registry metadata lists no required env vars — this is an incoherence to surface.

Instruction Scope

SKILL.md instructs the agent to operate on files under $WIKI_ROOT via an Obsidian MCP server (mcporter) and to fall back to local read/write/edit and grep if MCP is unavailable. The flows (ingest/query/lint) and files referenced are within the wiki domain. The LLM is directed to ask user consent for changes during Layer 2 lint — no hidden global file reads are mandated. The scripts do fetch URLs when deduping and run local file I/O as expected for the task.

Install Mechanism

There is no install spec (instruction-only), which is low risk. The repository includes Python scripts (no automatic pip installs), and embed.py documents optional pip packages for embedding. No remote download/extract install behavior is present. The only higher-risk install-like action is optional instructions to run pip/npn (examples) for embedding/MCP server, but these are not automatic.

Credentials

The manifest declares no required env vars or credentials, but SKILL.md and README rely on several environment variables (WIKI_ROOT, WIKI_SKILL_DIR, MCPORTER_CONFIG) and the optional embed pipeline requires AWS credentials and a Secret (WIKI_EMBED_SECRET) in Secrets Manager (contains OpenSearch username/password). The presence of embed.py means if enabled the skill will read secrets and call AWS services (Bedrock, Secrets Manager) — this is sensitive and not reflected in the published requirements. The discrepancy between 'no required env vars' and the documented/env-driven behavior is concerning and should be clarified.

Persistence & Privilege

always:false and normal autonomous invocation settings. The skill writes to a dedicated wiki directory ($WIKI_ROOT) and maintains its own cache/log files (.ingest-cache.json, log.md, .lint-history). It does not request system-wide or other-skills' credentials. The README strongly recommends using an MCP server to sandbox writes; without MCP, writes happen via fallback which is expected but less sandboxed — this is an operational consideration rather than a permissions escalation in the manifest.

What to consider before installing

What to check before installing/use: - Confirm WIKI_ROOT and WIKI_SKILL_DIR defaults point to a directory you control; the skill will read and write files there (index.md, pages/, log.md, .ingest-cache.json, .lint-history/). - If you enable embeddings (embed.py), be aware it uses boto3 to call Bedrock and Secrets Manager and expects a Secrets Manager secret (WIKI_EMBED_SECRET) containing OpenSearch credentials — only provide these if you trust the skill and understand the network activity to AWS. - The registry metadata lists no required env vars, but SKILL.md/README rely on env vars (MCPORTER_CONFIG, WIKI_ROOT, etc.). That mismatch should be resolved: confirm required env vars and how they are provided to the agent. - The skill strongly recommends an Obsidian filesystem MCP (mcporter) to sandbox reads/writes. If you do not run MCP, the skill will fall back to direct file reads/writes and exec grep — review that fallback behavior and ensure the agent is allowed only to touch the intended wiki directory. - Review the example cron/webhook lines in README: these are user-level examples that can post lint summaries to webhooks; the skill itself does not send network webhooks unless you wire them in your scheduler or run embed.py. - If you plan to enable embedding/semantic search, audit embed.py and only grant the minimum AWS IAM permissions needed (SecretsManager:GetSecretValue, Bedrock invoke, and OpenSearch access) and ensure the secret contains expected fields. - If you need more assurance, ask the publisher to update the registry metadata to list the environment variables and optional permissions (AWS) explicitly, and to state if any network calls are made by default. If you cannot validate the maintainer, run the skill in an isolated environment or with MCP sandboxing enabled.

Wikisage Skill

基于 Karpathy llm-wiki 模式的持久化 Wiki。 LLM 负责写和维护所有内容，用户负责来源、探索方向和提问。 纯本地 markdown 文件，用 index.md 导航，无需向量数据库。

📍 路径约定（环境变量驱动）

本 skill 所有路径都基于环境变量，无硬编码：

变量	默认值	作用
WIKI_ROOT	$HOME/.openclaw/workspace/wiki	Wiki markdown 根目录
MCPORTER_CONFIG	$HOME/.openclaw/workspace/config/mcporter.json	mcporter 配置文件（可选）
WIKI_SKILL_DIR	$HOME/.openclaw/workspace/skills/wikisage	Skill 自身目录（脚本位置）

首次部署时，在 shell/agent 环境里 export 一下这三个变量即可（或用默认值）。 下文示例用 $WIKI_ROOT 这种写法代替绝对路径。

🛠 执行通道：Obsidian MCP（首选，强烈推荐）

本 skill 围绕 Obsidian filesystem MCP server 设计。 没装 MCP 也能跑（走 read/write/edit fallback），但装了会更稳：allowed-dir 边界兜底、错误更规范、LLM 不会意外写到 wiki 外面。

所有 wiki 文件读写优先走 Obsidian filesystem MCP，而不是通用 read/write 工具。

操作	MCP 调用
读文件	mcporter call obsidian.read_text_file path=<abs path>
写/覆盖文件	mcporter call obsidian.write_file path=<abs> content=<str>
列目录	mcporter call obsidian.list_directory path=<abs>
搜文件名	mcporter call obsidian.search_files path=<abs> pattern=<glob>
改文件	mcporter call obsidian.edit_file path=<abs> edits=...
看边界	mcporter call obsidian.list_allowed_directories

所有调用都需要 --config $MCPORTER_CONFIG （mcporter 有双 config 坑：会同时读 ~/.claude.json 和项目 config，不带 --config 只会看到 claude.json 里的 server）

Fallback：MCP 不可用时（daemon 挂了、server 不 healthy），用通用 read/write/edit/exec grep 兜底，但要在回复里告诉用户"MCP 离线，走 fallback"。

全文搜索不走 MCP：MCP 的 search 只匹配文件名。找内容用：

• qmd-search（workspace 集合，BM25，快但索引可能滞后）
• exec grep -rn "关键词" $WIKI_ROOT/

---

触发条件

用户说	执行
"加进 wiki" / "ingest" / "摄入这篇"	→ ingest 流程
"查 wiki" / "wiki 里有没有" / "从 wiki 查"	→ query 流程
"整理 wiki" / "wiki 健康检查" / "lint"	→ lint 流程
涉及客户、历史决策、账号信息的技术问题	→ 先本地查 wiki，再回答
通用技术问题（无特定上下文）	→ 直接 MCP → LLM
回答完有价值的技术问题后	→ 询问"要把这些存进 wiki 吗？"

三层架构

$WIKI_ROOT/
├── raw/                  原始文档（只读，用户放入，LLM 不修改）
├── pages/                LLM 生成并维护的 markdown 文件集
│   ├── aws/              AWS 服务、架构、合规
│   ├── ai/               AI/LLM 技术
│   ├── clients/          客户信息（账号、联系人、项目）
│   ├── projects/         具体项目
│   └── ops/              运维、kubectl、DevOps
├── index.md              所有页面目录（标题 + 一行描述 + 路径），每次 ingest 后更新
├── log.md                操作日志（append-only，格式：## [YYYY-MM-DD] ingest | 标题）
└── .ingest-cache.json    SHA256 去重缓存（dedup.py 维护，不进 Obsidian vault）

只有一个 wiki 目录： $WIKI_ROOT（即 Obsidian MCP 的 allowed dir）

Query 流程

详见 scripts/query.md

核心逻辑：

1. obsidian.read_text_file 读 $WIKI_ROOT/index.md，找相关页面
2. obsidian.read_text_file 读相关页面全文，综合回答，标注来源 > 参考：[[页面名]]
3. 答案本身有价值 → 询问用户是否存回 wiki

Ingest 流程

详见 scripts/ingest.md

核心逻辑： 0. dedup.py check 去重（来源是文件/URL 时）→ DUPLICATE 就停

1. obsidian.read_text_file 读 index.md，判断是否已有相关页面
2. obsidian.write_file / obsidian.edit_file 新建 or 更新页面（一次 ingest 可能触碰 5-15 个页面）
3. obsidian.edit_file 更新 index.md
4. obsidian.edit_file 追加 log.md（## [YYYY-MM-DD] ingest | 来源标题）
5. dedup.py record 记录 SHA256 缓存（来源是文件/URL 时）

Lint 流程

详见 scripts/lint.md

检查：孤儿页面、缺失概念页、index.md 不一致、矛盾内容、过时内容 （lint.py 脚本走 Python filesystem 直接读，不经过 MCP；LLM Layer 2 整改时走 MCP）

页面模板

# 页面标题

**最后更新：** YYYY-MM-DD
**来源数量：** N
**分类：** aws/security
**置信度：** EXTRACTED  <!-- 整页默认值；段落内可局部覆盖 -->

## 概述

## 核心内容

<!-- 置信度可以在段落/句子级别用 inline tag 标注： -->
<!-- [EXTRACTED] 原文直接扒的事实 -->
<!-- [INFERRED]  基于来源推理的结论 -->
<!-- [AMBIGUOUS] 来源本身表述模糊 -->
<!-- [UNVERIFIED] AI 自己补的常识/背景，未经来源验证 -->

## 相关页面
- [[相关页面名]]

## 来源
- [[原始文档页面名]]
- [外部链接](https://...)

置信度标签规则（强制）

Tag	含义	什么时候用
EXTRACTED	从来源原文直接扒的事实	定价、API 参数、官方原话
INFERRED	基于来源推理/组合得出	"所以月成本约 $80"（来源只给了单价）
AMBIGUOUS	来源本身说得不清楚	文档自相矛盾或写得模糊
UNVERIFIED	AI 补的背景常识，没来源	写页面时为了通顺加的常识性描述

原则：

• 整页默认置信度写在 frontmatter，不要省略
• 页面内如果混合了不同置信度的内容，必须在段落开头/句尾用 inline tag 标注
• Query 时如果引用了 INFERRED / UNVERIFIED 的内容，必须在回答里明说（"这条是推断的"）

log.md 格式

每条记录格式：## [YYYY-MM-DD] {操作} | {标题}

## [2026-04-09] ingest | Karpathy llm-wiki 模式
## [2026-04-09] query | S3 Files POSIX 访问方案
## [2026-04-09] lint | 全库健康检查

可用 grep "^## \[" $WIKI_ROOT/log.md | tail -10 查最近操作。