Show HN：Wienerdog——为 Claude Code/Codex 提供记忆与自我改进技能

一份真正的 CLAUDE.md / AGENTS.md——通过访谈生成，而非空白页面。你的 AI 会在每次会话中了解你是谁、你如何工作，以及你关心什么。

由两个 Agent 共享的单一记忆库——Claude Code 的 CLAUDE.md 和 Codex CLI 的 AGENTS.md 都由同一场访谈生成，并从同一个记忆库读取。你在一个工具中教给它的内容，下次打开另一个工具时依然存在——不是两个各自拥有独立记忆的 AI。记忆库本身采用 Obsidian 约定的 PARA 结构（如果你使用 Obsidian，它可与之配合；但并不要求使用 Obsidian）：你拥有的纯文本文件，可由你阅读，也可用 git 进行版本控制。它始终留在你的电脑上——Wienerdog 不会自行将其推送到任何地方；如果你想发布它，需要由你自己完成。Wienerdog 还会从它读取和写入的内容中清除看似敏感的信息（API 密钥、token），但这种清理属于尽力而为，并非保证——如果你确认真实凭证已被记录，请遵循密钥事件处理指南。

做梦（Dreaming）——夜间任务会审查当天的对话，通过质量门槛将重要信息提升为长期记忆，并把重复工作流转化为可复用技能。Wienerdog 创建的技能会持续学习：做梦流程会观察它们在实际使用中的表现，并随时间修改它们——你自己的技能以及 Wienerdog 内置技能绝不会被触碰（参见威胁模型）。每晚最多产生一次 git commit；任何内容都可以回滚。如果你的电脑在该时间关机或休眠，也无需担心——Wienerdog 会在你下次回来时自动补做。做梦流程还可以在它跨足够多会话反复观察到相同信息后，提议编辑你的身份笔记——即你是谁、你的偏好和目标——但它提议的任何内容都不会进入你的下一个会话，除非你使用 wienerdog memory approve 批准确切文本。你可以随时运行 wienerdog safety 查看每项受控能力的状态。

Google Workspace 感知能力（可选）——Gmail、Calendar 和 Drive 访问默认采用“先读取、先起草”的设计。在 Wienerdog 强制执行的路径中——一个持有凭证且仅暴露固定动词的本地能力代理——你的 AI 只能向你明确授予权限的对象、向你明确批准的收件人_发送_内容；无人值守的发送只会发往你自己的地址；授权只能由你在键盘前创建，绝不会由 AI 创建。1 你可以随时通过 /wienerdog-google-setup 连接；运行 wienerdog safety 可查看每项受控能力的状态。

适合笔记本电脑的例行流程（可选，可从目录中挑选）——完成设置后，可从菜单中选择现成的定时例行流程（晨间摘要、收件箱整理、每周回顾，……），由操作系统的原生调度器（launchd / systemd / Task Scheduler）运行。笔记本在运行时间处于合盖状态？它会补做。没有守护进程。每项例行流程均在受严格限制的 AI 副本中运行，并会在每次运行前重新检查（参见威胁模型）；你可以随时运行 wienerdog safety 查看每项受控能力的状态。

定时运行会在执行前经过验证。 每次夜间触发时，一个独立启动器会先检查应用代码是否与其记录的内容地址匹配，以及任务是否仍与其经摘要绑定的授权描述符匹配；Claude 和 Git 只会从其被固定且经过结构验证的安装位置启动。在 wienerdog sync 之外对 config.yaml 或应用目录树所做的编辑，不会改变实际运行的内容——任务会拒绝运行并发出警报（故障时默认关闭）。对于 config.yaml 编辑，修复方式是执行一次 wienerdog sync；当检查无法确认应用文件本身时，警报会刻意不建议你执行同步——同步并不是此时安全的下一步——而是要求你从可信来源重新安装。2

A proper CLAUDE.md / AGENTS.md — generated from an interview, not a blank page. Your AI knows who you are, how you work, and what you care about, in every session.

One memory vault, shared by both agents — Claude Code's CLAUDE.md and Codex CLI's AGENTS.md are both generated from the same interview and read from the same vault. Teach it something in one tool, and it's there the next time you open the other — not two AIs with two separate memories. The vault itself is Obsidian-convention PARA structure (works with Obsidian if you use it, doesn't require it): plain files you own, readable by you, versioned in git. It stays on your machine — Wienerdog never pushes it anywhere on its own; publishing it, if you ever want to, is something you do yourself. Wienerdog also scrubs secret-looking text (API keys, tokens) out of what it reads and writes, but that scrubbing is best-effort, not a guarantee — if you know a real credential got captured, follow the secret-incident guide.

Dreaming — a nightly job reviews the day's conversations, promotes what matters into long-term memory through quality gates, and turns your repeated workflows into reusable skills. Skills Wienerdog created keep learning: the dream watches how they perform in real use and revises them over time — your own skills and Wienerdog's built-ins are never touched (see the threat model). Every night is at most one git commit; anything can be reverted. If your computer is off or asleep at that time, don't worry — Wienerdog catches up automatically the next time you're back. The dream can also propose edits to your identity notes — who you are, your preferences, your goals — once it's noticed the same thing across enough sessions, but nothing it proposes reaches your next session until you approve the exact text with wienerdog memory approve. Run wienerdog safety any time to see the status of every gated capability.

Google Workspace senses (optional) — Gmail, Calendar, and Drive access that is read-first and draft-first by design. On the path Wienerdog enforces — a local capability broker that holds the credentials and exposes only fixed verbs — your AI can only send what you explicitly granted, to recipients you explicitly approved, and unattended sending goes to your own address only; grants are created by you at the keyboard, never by the AI.1 Connect it any time with /wienerdog-google-setup; run wienerdog safety to see the status of every gated capability.

Laptop-friendly routines (optional, pick from a catalog) — after setup, choose from a menu of ready-made scheduled routines (morning digest, inbox triage, weekly review, …) run by your OS's native scheduler (launchd / systemd / Task Scheduler). Laptop was closed at run time? It catches up. No daemon. Each routine runs in a locked-down copy of your AI that's checked fresh before every run (see the threat model); run wienerdog safety any time to see the status of every gated capability.

Scheduled runs are verified before they run. At every nightly fire, an independent launcher first checks that the app's code matches its recorded content address and that the job still matches its digest-bound authorization descriptor; Claude and Git are only spawned from their pinned, structurally verified install locations. Edits to config.yaml or the app tree made outside wienerdog sync don't change what runs — the job refuses with an alert instead (fail closed). For a config.yaml edit the fix is one wienerdog sync; when the check couldn't confirm the app files themselves, the alert deliberately does not tell you to sync — syncing is not the safe next step there — and asks you to reinstall from a trusted source instead.2