Getting it into your agent
One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.
npx agentmods add skills/ryanzhao1011/workframe/librariannpx skills add ryanzhao1011/workframe --skill librariangit clone --depth 1 https://github.com/ryanzhao1011/workframeWrote this? Show the measurements
A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.
[](https://agentmods.dev/skills/ryanzhao1011/workframe/librarian)<a href="https://agentmods.dev/skills/ryanzhao1011/workframe/librarian"><img src="https://agentmods.dev/badge/skills/ryanzhao1011/workframe/librarian.svg" alt="Measured on agentmods" height="20"></a>What it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5 | $0.00055 | $0.08353 |
| Opus 5 | $0.00028 | $0.04176 |
| Sonnet 5 | $0.00011 | $0.01671 |
| Haiku 4.5 | $0.00006 | $0.00835 |
Grade B, and why
librarian scanned grade B with 1 finding against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 4d ago.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Reads agent configuration directoriesmediumAgent snooping
.claude/, .codex/, .gemini/ hold keys, settings and other credentials a mod has no legitimate need for.
python "$(cat .claude/workframe-state/plugin-root.txt)/bin/workframe-recompute-skill-metrics" How it starts
The opening of the file, as written. The whole thing — 319 lines — stays where its author put it; the contents beside it link to each section on GitHub.
Librarian 记忆整理技能
6 步整理流程
第 1 步：读取状态（动态角色发现 + shared 特殊角色）
先处理 shared/（第 0 号"特殊角色"）：
Read .claude/agent-memory/shared/MEMORY.md（跨角色权威事实；不存在则跳过）
Read .claude/agent-memory/shared/notes.md（共享缓冲；不存在则跳过）
shared 层的整理规则与 role 层一致（D/U/R/A 评估、容量管理、快照保存），但写入条件更严：
- 提升到
shared/MEMORY.md必须满足"影响 ≥2 个角色" + D/U/R/A ≥2 项（主 Claude 不计入角色数，理由见 2b 三问第 3 问） - 单条压缩到 ≤200 字符。shared 是各 subagent 启动时全量注入的——每加一句，此后每次委派都要付一次 token；判据是「这句删掉，另一个域的人会不会做错事」，答否就删。展开论证留在
shared/notes.md或权威文档，MEMORY 只留结论 - shared 变更在整理日志中单列（与 role 层变更分开列，标明「本轮 shared 新增/修改 N 条」+ 逐条一行摘要），供用户事后复核。不逐条弹确认——已有「≥2 消费域 + D/U/R/A ≥2 + ≤200 字符」三道门槛，再加人工确认属于冗余打扰
- 冲突以
shared/MEMORY.md为权威，role MEMORY 中的同类条目降级为 notes
再动态发现项目中其他角色：
用 Glob 扫描 .claude/agent-memory/*/ 得到所有子目录，排除 shared
例如返回：[pm, dev, qa, prompt-eng]（core 默认）
[pm, dev, qa, prompt-eng, ceo, designer, ...]（项目级扩展）
对发现的每个角色 <role>，读取其记忆文件：
.claude/agent-memory/<role>/MEMORY.md
.claude/agent-memory/<role>/notes.md
统计每个 MEMORY.md 的当前字符数（预算：role ≤8000 字符 / shared ≤4000 字符——字符才反映注入开销，行数可被单行超长钻空）。
为什么动态发现：core plugin 提供 4 通用角色（pm/dev/qa/prompt-eng），项目可手动在 .claude/agents/ 和 .claude/agent-memory/ 下新增项目级角色（如 ceo、designer、content-operator 等）。Librarian 必须覆盖所有角色，因此不能硬编码列表。
第 2 步：评估 notes.md + 决定落点
2a. D/U/R/A 准入评估
遍历每个角色的 notes.md 中的微反思条目，跳过下列两类已结案条目：
- 标注
→ 已提升至 X的（本 SOP 自己提升过的） - 标注
→ 已被 {日期} 纠正取代的（correction-detection第 3 步 supersede 时移进来的 历史留档）——它们是被推翻的旧口径，重新提升等于把用户纠正过的错误再送回高置信区 - Durability：30 天后还重要？
- Uniqueness：目标落点中尚未记录？
- Retrievable：未来需要回忆？
- Authority：来源可靠？
满足 ≥2 项 → 标记为"待提升"。
单人/小团队项目的判定口径：用户当面拍板的偏好，Authority 已满分，不要因为"是否对所有人普适"而压低评分——这类知识库只服务当前用户，普适性不是准入条件。历史上大量高价值偏好因这条误判滞留在 notes（实测某项目 181 行 notes 里全是可执行规则，却因自设"需多次验证"门槛从未提升）。
2b. 落点分层决策
MEMORY.md 不是唯一落点。按知识性质选：
| 落点 | 判据 | 加载时机 | 权限 |
|---|---|---|---|
项目 rules（.claude/rules/local/*.md） |
跨场景稳定的纪律 / 口径，主 Claude 也需遵守 | 每次会话必载 | L2 用户审批 |
CLAUDE.md |
项目级协作规则、角色边界、目录约定 | 每次会话必载 | L2 用户审批 |
项目 skill（.claude/skills/<name>/） |
场景触发且成套的操作方法——特定场景才需要、一到场景就要整套用（SOP / 口径集 / 模板 / 清单） | 场景命中按需加载 | L2 用户审批（core plugin skill 不直接改，走 self-iteration 提案） |
<role>/MEMORY.md |
角色特有的高置信事实 | 该 subagent 被调度时 | L1 自主 |
shared/MEMORY.md |
影响 ≥2 角色的权威事实 | 各 subagent 启动时 | L1 自主（写入条件更严）；agent / 主 Claude 不直写——它们写 shared/notes.md，由本 skill 评估提升（对比上一行：role MEMORY 满足 D/U/R/A ≥2 可由角色直写，两行的「L1 自主」含义不同） |
| auto-memory（CC 官方主 Claude 记忆目录） | 主 Claude 层内容：用户偏好 / 协作习惯 / 项目状态指针——消费者是主 Claude 而非某个角色 | 主会话启动（官方注入，subagent 不注入） | 主 Claude 自维护；librarian 不写——识别到此类条目时在结果中建议归 auto-memory |
| 保留 notes | 未定型 / 场景太具体 / 待验证 | 不加载 | — |
| 移出记忆体系 | 属业务知识（竞品分析、方案论证、领域事实）→ 应进需求文档 / spec，不是记忆 | — | L2 |
What ships with it
3 files beside SKILL.md in the same directory: the scripts, references and assets a skill reads on demand. Not counted in the per-session cost; read them before you install if any of them is executable.
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- 4d ago First seen · 319 lines · 55 tokens per session scan B d78ab59b5e31
librarian is a skill published in the GitHub repository ryanzhao1011/workframe (4 stars, last pushed 17d ago), licensed MIT. It adds 55 tokens to every session and 8,353 once invoked, about $0.0003 per session on Opus 5. A static security scan graded it B with 1 finding (reads agent configuration directories). No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-31.
Other skills, from other repositories
decision-ledger
Maintain an append-only decision ledger (DECISIONS.md). Use when the user confirms a decision, reverses or changes a past decision, asks what was decided and why, or at session start to load standing decisions.
knowledge-base
Keep verified findings in memory/knowledge/ so the same research is never done twice. Use before researching any topic, when a finding passes the verify gate, or when about to answer from something learned in an earlier session.
pin
Turn a correction or working preference the user just expressed into a permanent ballast rule. Use when the user corrects your behavior, says "don't do X" / "always do Y" / "I told you before", or asks you to remember how they want things done.
brain-init
Scaffold a file-based memory system (index, decision ledger, open questions, session log, optional product truth file) so a project can serve as a durable second brain. Use when the user wants to set up memory, an external brain, or the ballast structure in a project.
recall
Sweep everything the project already holds before answering — at the first message of a session, and again whenever the subject shifts to something the current thread was not about. Use before replying, not only before researching.
project-model
When working in a codebase you've touched before, consult and update the persistent project model maintained by presence. The model lives at /.claude/presence/projects/ /model.md and contains terse notes about architecture, conventions, and verified facts. Use it BEFORE re-deriving things you've already learned, and…