agent-knowledge
/agent-knowledge
Agent/LLM 应用工程的本地全文知识库(harness、skills、context/memory、tools/PTC/MCP、 runtime、multi-agent、evaluation、training、security、loops,规模见 kb.py stats), 带检索、渐进阅读、联网补新与入库流程。Use when the user asks about agent engineering, harness design, AGENTS.md, skill authoring, context engineering, agent memory, tool calling/MCP, long-running agents, multi-agent systems, agent evaluation or training, or wants the latest practices beyond model knowledge, or asks to research/archive new articles on these topics. Also use when designing any agent framework, auto-research pipeline, or coding-agent workflow and you need prio
- Author
- NonTrivialLiu ↗
- Repository
- agent-knowledge ↗
- Licence
- MIT
- Last repository change
- 13.11.2026
Assessment and security
- Files in this package
- 4
- Scripts
- 0
- HTML comments
- 0
- URLs in SKILL.md
- 0
No signals detected. Repository changed on 13.11.2026, without references to retired models or APIs.
MIT, open: use and modify while retaining author attribution.
What it does and when to use it
Agent/LLM 应用工程的本地全文知识库(harness、skills、context/memory、tools/PTC/MCP、 runtime、multi-agent、evaluation、training、security、loops,规模见 kb.py stats), 带检索、渐进阅读、联网补新与入库流程。Use when the user asks about agent engineering, harness design, AGENTS.md, skill authoring, context engineering, agent memory, tool calling/MCP, long-running agents, multi-agent systems, agent evaluation or training, or wants the latest practices beyond model knowledge, or asks to research/archive new articles on these topics. Also use when designing any agent framework, auto-research pipeline, or coding-agent workflow and you need prio
Original author description
Agent/LLM 应用工程的本地全文知识库(harness、skills、context/memory、tools/PTC/MCP、 runtime、multi-agent、evaluation、training、security、loops,规模见 kb.py stats), 带检索、渐进阅读、联网补新与入库流程。Use when the user asks about agent engineering, harness design, AGENTS.md, skill authoring, context engineering, agent memory, tool calling/MCP, long-running agents, multi-agent systems, agent evaluation or training, or wants the latest practices beyond model knowledge, or asks to research/archive new articles on these topics. Also use when designing any agent framework, auto-research pipeline, or coding-agent workflow and you need prio
How it works
- 工作流
- 检索方法(先 rg,再 Read,绝不整库读)
- 渐进阅读(按深度递进,不贪快也不全读)
- ⭐ core 必读机制
- 入库(新增文章)
- 维护
SKILL.md preview
---
name: agent-knowledge
description: >-
Agent/LLM 应用工程的本地全文知识库(harness、skills、context/memory、tools/PTC/MCP、
runtime、multi-agent、evaluation、training、security、loops,规模见 kb.py stats),
带检索、渐进阅读、联网补新与入库流程。Use when the user asks about agent
engineering, harness design, AGENTS.md, skill authoring, context engineering,
agent memory, tool calling/MCP, long-running agents, multi-agent systems, agent
evaluation or training, or wants the latest practices beyond model knowledge, or
asks to research/archive new articles on these topics. Also use when designing
any agent framework, auto-research pipeline, or coding-agent workflow and you
need prior art. Not for: 大模型训练理论本身、与 agent 工程无关的 ML 问题。
---
# agent-knowledge:Agent 工程知识库
跨 Codex / Claude Code / ZCode 共用的本地知识库:全文 Markdown(Anthropic、
OpenAI、Google、DeepSeek、LangChain、Cloudflare、METR、arXiv 等),每篇带单行
YAML frontmatter(title/url/published/fetched/type/topics/summary,部分含
priority: core)。权威元数据在 frontmatter;`references/<类>/CATALOG.md` 是脚本
再生的浏览视图(⭐ core 必读置顶)。
## 工作流
1. **本地检索 + 渐进阅读**(见下两节)。
2. **联网补新**:本地结果缺失、明显过旧(published 超一季)、或用户要"最新"
时,联网搜增量(来源范围见 `references/ingest.md`)。
3. **遇到好文章就入库**:检索或联网中发现有具体机制、真实数据且库内缺失或
更优的文章——入库是默认动作,按 ingest.md 七步走完再回答用户;仅当内容是
重复转述、纯新闻稿或与库内已有篇目无增量差异时才跳过。
## 检索方法(先 rg,再 Read,绝不整库读)
frontmatter 全部单行,rg 一条命令可同时搜元数据与正文。**至少跑 2-3 组
关键词**(同义词、中→英翻译后的术语、相邻概念),合并去重得到候选集:
```bash
cd <本skill目录>
rg -i -l --no-heading "database agent" references/ -g '!CATALOG.md' # 全文
rg -i "topics:.*(skill|trigger)" references/ -g '!CATALOG.md' -l # 元数据
rg -i "summary:.*(压缩|compaction)" references/ -g '!CATALOG.md' -l # 中文摘要
cat references/<类>/CATALOG.md # 类目定位(单行/篇)
```
候选集先看 frontmatter 的 summary 过滤明显无关的;目标保留 **3-10 篇**进入
阅读——少于此数说明关键词太窄,换词再检索;回答横跨多主题时取上限。
候选中有 `type: analysis`(synthesis 综合文章)时优先读:它提供结论与机制
的融合分析,并相对链接到原始条目,可顺链精读。
分类:fundamentals 概念全景 · skills-design Skill 设计 · harness 工程方法/AGENTS.md ·
context-memory 上下文/记忆 · tools-execution 工具/PTC/MCP/沙盒 · runtime 长任务/运行时 ·
multi-agent 协作 · evaluation 评估/实证/成本 · training RL/环境/模型报告 ·
model-behavior 各代模型行为与提示词对冲 · reliability-security 注入/权限 ·
loops-selfimprovement Loop/自改进 · synthesis 自有综合。各类检索词明细见
`references/ingest.md` 分类表;交叉主题跨类召回靠 rg topics。
## 渐进阅读(按深度递进,不贪快也不全读)
1. **结构层**:`rg "^#{1,3} " <file>` 看标题骨架,判断该读哪几节。
2. **摘要层**:`sed -n '1,12p' <file>` 读 frontmatter summary 与开篇结论;
有 TL;DR/Summary/结论节的直接跳读首尾。CATALOG 只做类目定位,
不作为内容源。
3. **精读层**:只精读与当前任务直接相关的章节;⭐ core 文章值得整篇或读主干
章节——它们是经过筛选的概念全景与权威叙述。
各篇按需选深度:概念性问题读 core 全文即可;工程实现问题精读 2-3 篇的
相关章节 + 扫其余候选的摘要层。回答时标注来源 `title (published)`。
## ⭐ core 必读机制
每类有 ≤3 篇 `priority: core`:该主题的概念全景文、权威规范或公认经典
(审定规则见 ingest.md)。进入一个新主题时,先读该类 CATALOG.md 的 ⭐ 条目
再读检索命中的具体篇目;跨主题任务每类 core 至少读到摘要层。core 由维护者
标注,Agent 不得自封。
## 入库(新增文章)
完整 SOP 见 `references/ingest.md`(来源范围、去重、抓取、frontmatter 模板、
分类表、硬性禁令)。最短路径:
```bash
rg -i "<url关键词>" references/ # 1 去重
uv run scripts/fetch.py URL --title "T" --published D \
--out references/<类>/<slug>.md # 2 抓取
# 3 补全 frontmatter(单行!) 4 python3 scripts/kb.py lint # 5 校验
python3 scripts/kb.py catalog # 6 再生目录
```
lint 不过不算入库;禁止手改 CATALOG.md;禁止序号文件名。
## 维护
- `python3 scripts/kb.py stats` 看规模与 core 分布;`lint` 全库体检;`catalog` 再生目录。
- 无定时任务:更新由用户触发或随检索工作流顺带完成。
- 本库真实文件在 `~/.agents/skills/agent-knowledge/`;`~/.claude/skills/`、
`~/.zcode/skills/` 下是软链,不要在软链侧改结构。
Files in this package (4)
- .gitignore 1 KB
- LICENSE 1 KB
- README.md 3 KB
- SKILL.md 5 KB
Versions
The database has no other skills named agent-knowledge.
Database snapshot: 13.11.2026 · package version b14573dc8bc4 (SHA-256 digest of the entire package)
Downloads in the last 30 days: 0 · ranking
Report a problem with this skill: harmful content, incorrect descriptions or an author’s removal request.
Similar skills
- 85context-engineering-collection
A comprehensive collection of Agent Skills for context engineering, harness engineering, multi-agent architectures, and production agent systems. Use when building, optimizing, evaluating, or debugging agent systems that require effective context management and reliable operating loops. - 95ai-agent-architectures
Use this skill when the user says 'build an AI agent', 'agent architecture', 'design an autonomous agent', 'agent system prompt', 'agent with tools and memory', 'which agent framework'. Entry point for single-agent design: decomposes the request into loop, tools, memory,… - 90context-engineering
Master context engineering for AI agent systems. Use when designing agent architectures, debugging context failures, optimizing token usage, implementing memory systems, building multi-agent coordination, evaluating agent performance, or developing LLM-powered pipelines. Covers… - 100ai-agent-engineering
Use when designing or debugging an LLM agent, tool-using assistant or agent harness: "agent or workflow?", "agent picks the wrong tool", "context keeps filling up", "agent forgets what it did", "too many MCP tools", "should I add subagents", "agent drifts over hours", exploding… - 100agent-engineering-expert
Build LLM agents that use tools safely: tool design, the agent loop, memory, MCP servers, multi-agent orchestration, sandboxing and prompt-injection defence. Use when the user mentions AI agents, tool use or function calling, MCP or Model Context Protocol, autonomous workflows,… - 100building-agents
Designs and hardens LLM agents that call tools in a loop - choosing between a single call, a fixed workflow and an agent, tool design, context management, stop conditions and budgets, human approval, sandboxing, prompt-injection defenses, tracing and agent evaluation. Use when…