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 instructions/leviyuan/lodestar/agents-mdgit clone --depth 1 https://github.com/leviyuan/lodestarWrote 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/instructions/leviyuan/lodestar/agents-md)<a href="https://agentmods.dev/instructions/leviyuan/lodestar/agents-md"><img src="https://agentmods.dev/badge/instructions/leviyuan/lodestar/agents-md.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 | $0.02665 | $0.02665 |
| Opus 5 | $0.01333 | $0.01333 |
| Sonnet 5 | $0.00533 | $0.00533 |
| Haiku 4.5 | $0.00266 | $0.00266 |
Grade A, and why
lodestar AGENTS.md scanned grade A with 0 findings 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 today.
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.
Nothing flagged
None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.
How it starts
The opening of the file, as written. The whole thing — 60 lines — stays where its author put it; the contents beside it link to each section on GitHub.
Lodestar 项目指引
Lodestar 是一个 Bun/TypeScript daemon:它从飞书 WebSocket 接收消息,把每个群映射为一个 Session,再通过统一 AgentProcess 驱动 Codex app-server 或 Claude Agent SDK,并把 turn、工具与状态渲染到 Feishu Card Kit schema 2.0 卡片。
本仓库的维护入口和 AI 协作说明统一以 Codex/AGENTS.md 为准;这只是维护方式迁移,不是产品后端迁移。Codex、Claude、GLM、DeepSeek 与 Claude native 路径都是受支持产品能力,未经用户明确提出产品变更,不得因清理维护文件、缺少本机凭据或个人偏好而删除、弱化或伪装任一后端。
范围地图
cli.ts只处理首次配置与 PID guard,daemon.ts负责 WS、session registry、Card action 和本机通知;业务逻辑放在src/。src/是 session、双 agent 进程、token sources、飞书 API、Card Kit 队列、任务清单和 worktree 的核心实现。修改前读取src/AGENTS.md。src/cards/只维护卡片模板和渲染辅助。修改前同时读取src/AGENTS.md与src/cards/AGENTS.md。scripts/会复用生产配置并触达真实群或 daemon;运行或修改前读取scripts/AGENTS.md。docs/保存后端设计 memo;修改双后端架构前按docs/AGENTS.md选择性阅读,但最终事实仍以源码和测试为准。- 从仓库根目录启动的 agent 不会自动获得更深目录的说明;编辑子树前主动读取该路径适用的嵌套
AGENTS.md。
工作规则
- 开发运行时是 Bun,源码入口用
bun daemon.ts或bun run start;发布产物由bun run build生成可在 Node.js 18+ 运行的入口。 config.ts在 import 时同步加载配置且缺失即报错;首次安装的延迟导入只能留在cli.ts。不要吞掉配置、凭据、Codex 登录、Claude SDK native binary 或飞书 API 错误。- 配置默认在
~/.config/lodestar/config.toml,日志和 session/chat/resume/model/tasklist 等状态默认在~/.local/share/lodestar/。新增持久状态统一经src/paths.ts放到 XDG/平台数据目录,不要写回仓库。 - 凭据只存在于用户配置、登录状态或环境变量。禁止把 token、App Secret、聊天成员数据、debug context、
~/.codex/~/.claude内容或本机绝对配置复制到跟踪文件。 - API、模型目录、额度、上传或 agent 启动失败必须显式记录并向调用方显示失败;不得伪造数据、静默成功或偷偷换 provider/source。只允许针对已知瞬态条件的有限重试,最终失败仍需暴露。
- 保留用户已有改动;不要覆盖无关脏文件。依赖变更同步更新
bun.lock,不要手工改锁文件内容。 - 所有展示到工具卡片的 shell 命令第一行使用
# desc: <中文摘要>;src/cards/shell-command.ts依赖这个约定提取可读标题。 - Card action
kind、共享element_id、群命令、resume/model map 和 token source id 都是持久 wire contract。重命名时同步修改生产分发、迁移和测试,不能只改显示层。
跨模块不变式
- 一个飞书群只对应一个
Session和一个当前主 agent 进程。Codex 走codex app-server --listen stdio://JSON-RPC;Claude/GLM/DeepSeek/native 走@anthropic-ai/claude-agent-sdk的 streaming-inputquery()。AgentService可按实时身份并行启动完整 delegated Agent:它们与主进程共用全能力 coding-agent 启动面,行为范围只由主 Agent 的原始 prompt 约束,并可通过独立 capability 继续委派。不要恢复 tmux、旧 JSONL 队列或旁路进程控制。 - daemon-owned Skill 必须由
managed-skills.ts同源同步:Codex/Claude standalone 目录用于各自 user source,另生成 Claude 本地插件供排除 user settings/env 的 GLM/DeepSeek SDK 主会话显式加载。不得为了加载 Skill 把 user source 重新混入注入凭据的 TokenSource。 - Token Source 是账号、凭据、模型目录、spawn env 和额度的真相源。内置 source 包括 Codex subscription、GLM Coding Plan、DeepSeek 与 Claude native;GLM/DeepSeek 可由 config 或本机 Claude settings 探测,存在显式 Claude-side source 时 native 让位。新增 source 按 factory 注册模式扩展,禁止在 model/session 层再建固定枚举。
- Agent 身份目录必须动态遍历 Token Source 的全部模型,使用模型目录声明的真实默认 effort,并允许主 Agent 按 run 显式选择受支持 effort;source disabled 或目录刷新失败都显示 MISS,禁止换模型/降 effort。
- 同一原始 prompt 选择多个 Agent 身份时,主 Agent 用一个 run 携带全部 identity id,由 run 内 fan-out 并发。不同 run 可并行并组成父子 DAG;每个 worker 使用独立、run-scoped、可撤销 capability,只能访问自己的 run 子树。全局并发和递归深度是透明进程治理,不能改变模型权限、注入审计语义或偷偷降级。
- Delegated Agent 必须使用完整工具面:Codex 固定
danger-full-access且不禁 apps/plugins/multi_agent/MCP/hooks;Claude 固定claude_codepreset、完整内置工具、项目 MCP 与 daemon-managed Skills。不得恢复 reviewer system prompt、read-only/safe mode、工具白名单或自动 deny。AskUserQuestion/request_user_input进入needs_input,由主 Agent 通过 answer API 回填;follow-up 必须复用 provider 原生 session。 - Agent run 每次生命周期状态变化都原子落盘;大 prompt/输出拆成独立私有正文 artifact,状态快照不得重复内嵌。委派产生的 Codex/Claude 原生 session id 另行登记并从主会话
rs/fk历史排除。父 run 取消、Session stop/kill/restart 与 daemon staged shutdown 都必须在首次 await 前封住新建入口、吊销 capability,并递归回收后代进程。 model是账号→模型→effort 三级动态面板。Codex 模型来自 app-servermodel/list,GLM/DeepSeek 来自兼容端点并允许受验证的补录,native 使用 SDK aliases;失败显示 MISS,不造假选项。- 同 provider、同 token source 的模型设置走
setModelSettings:Claude 从后续 turn 使用,Codex 持久目标需重启进程生效。跨 provider 或跨 source 会改变进程/env,只能在空闲时停止旧进程并按各 provider 独立 resume id 重新启动;活跃或排队状态按现有规则拒绝切换。 - Claude 路径保留原生 session resume/fork、
fk/bk/rs、AskUserQuestion、project profile、MCP/skills、主动/compact和 SDKtask_*后台任务;Codex 路径保留 app-server 权限、request_user_input、plan/goal、compaction、usage 与 collab 子 agent 映射。共享卡片事件不能抹平两端确有差异。 cardkit.ts独占生产卡的 sequence、队列、流式 TTL 重开、元素计数和写失败状态。session 与模板不得旁路它直调 Card Kit HTTP;专用探针只能在scripts/cardkit-probe.ts中显式操作测试卡。- Assistant 正文按完整 block 插入静态元素,不走
/content逐字输出;footer 用 element replace。公式渲染、换卡和 turn 收尾服从同一 per-card 队列,具体事务见src/AGENTS.md。 worktree.ts独占work/*分支和同级<project>[name]worktree 操作。飞书任务清单只保留基础绑定/删除,不运行自动计划、执行、审核或合并 worker。[[send: /abs/path]]由 outbound marker 流程解析并作为独立飞书文件消息发送;正文保留原标记作为可见回执,卡片模板本身不得读取本机文件。
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.
- today Changed · +2 lines · +241 tokens per session 27212a7ff56d
- 4d ago First seen · 58 lines · 2,424 tokens per session scan A af4504a5343b
lodestar AGENTS.md is an instructions file published in the GitHub repository leviyuan/lodestar (11 stars, last pushed 4d ago), licensed MIT. It adds 2,665 tokens to every session, about $0.0133 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-30.
Other instructions, from other repositories
botmux CLAUDE.md
Claude Code instructions for deepcoldy/botmux, covering botmux, 构建 & 运行, bun 开发链路, worktree 的 nodemodules:共享还是独立(改前必读) and 编译态(单文件二进制)注意.
openlark AGENTS.md
AGENTS.md instructions for foxzool/openlark, covering openlark knowledge base, structure, where to look, conventions and 模块组织.
larksnap CLAUDE.md
Instructions for AmbroseX/larksnap, covering 基本约定, plan 文件管理, 开发规范, 根据场景动态调整顺序 and 文档写入.
Claude-to-IM-skill CLAUDE.md
Instructions for op7418/Claude-to-IM-skill, covering claude.md — project guidelines for claude-to-im-skill and replying to github issues.
lark-mcp-on-agentcore AGENTS.md
AGENTS.md instructions for ddpie/lark-mcp-on-agentcore, covering agents.md, project overview, setup & commands, project structure and code style.
cc-connect-next AGENTS.md
AGENTS.md instructions for timmyagentic/cc-connect-next, covering cc-connect development guide, project overview, architecture, key design principles and core interfaces.