← 返回 Skills 市场
331
总下载
0
收藏
1
当前安装
1
版本数
在 OpenClaw 中安装
/install relationship-os
功能描述
Relationship OS — Enables AI agents to build human-like relationships with users through event memory, open threads, exclusive memories, agent stances, and r...
安全使用建议
What to consider before installing:
- Data persistence & privacy: this skill intentionally creates a .relationship/ directory and stores 'exclusive memories', stances, timelines, and debug logs in plain files in your workspace. If the user will store sensitive personal information, decide whether that's acceptable and where those files will live; consider using a restricted or encrypted workspace.
- External channels & API keys: the README/SKILL.md show examples that send follow-ups via cron/telegram and generate images using APIs (OpenAI DALL·E). The skill does not declare required env vars (e.g., TELEGRAM token, OPENAI_API_KEY) but the runtime behavior can use them if present. Before installing, verify and control which credentials are available to the agent and remove any tokens you don't want the skill to access.
- Proactivity & frequency: the skill is designed to send proactive messages (heartbeat/cron-driven follow-ups). If you don't want an agent sending unsolicited messages, disable the heartbeat/cron integration or avoid adding delivery channels (telegram) and HEARTBEAT.md entries.
- Prompt-rewriting guidance: the README suggests reframing the system prompt as a 'fiction writer' to evade safety moderation. This is a red flag — avoid changing system prompts to bypass model safeguards. If you want richer roleplay, handle it explicitly and transparently rather than altering system-level prompts.
- Audit the hooks/scripts: review hooks/openclaw/handler.{js,ts}, scripts/init.sh, scripts/heartbeat-check.sh and debug.sh to confirm they meet your privacy and behavior expectations. The hook appends to .relationship/debug.log and injects a virtual RELATIONSHIP_CONTEXT.md at bootstrap — that injection will become part of the agent's context every session.
- Recommended mitigations: install in an isolated/sandbox workspace; remove or disable recurring delivery (cron/heartbeat) before first run; avoid storing highly sensitive data in .relationship; if you allow image-generation, ensure only the intended API keys are present and that the selfie rules meet your content policy.
If you want, I can list the exact lines and files that implement the bootstrap injection, the cron follow-up example, and the places that reference OPENAI_API_KEY and Telegram delivery so you can make a targeted review or sanitization plan.
功能分析
Type: OpenClaw Skill
Name: relationship-os
Version: 0.1.0
The relationship-os skill is a sophisticated companion-style framework designed to provide AI agents with persistent memory and evolving personalities. It uses OpenClaw bootstrap hooks (handler.js/ts) to inject relationship context and shell scripts (init.sh, heartbeat-check.sh) to manage local state files within a .relationship/ directory. While the documentation (README.md) suggests using 'Fiction Writer' framing to bypass standard LLM safety filters for more expressive roleplay, the underlying code and instructions are focused entirely on the stated purpose of relationship tracking and contain no evidence of data exfiltration, malicious backdoors, or unauthorized system access.
能力评估
Purpose & Capability
The skill's name/description (relationship memory, follow-ups, stances) aligns with the code and scripts: hooks read/write a .relationship state directory, create timeline and threads, and inject a bootstrap context. These file I/O and scheduling behaviors are expected for this purpose. Minor inconsistency: the registry metadata said 'no install spec / instruction-only', yet the package includes executable hook handlers and a postInstall init script in _meta.json — so it is not purely instruction-only.
Instruction Scope
SKILL.md and scripts instruct the agent to (a) persistently create and update files under .relationship (timeline, threads, secrets, stance), (b) schedule cron follow-ups via an openclaw cron add command delivering messages (example uses --deliver --channel telegram), and (c) proactively send messages (heartbeat / anniversary / pattern-break triggers). Additionally the documentation explicitly suggests reframing the system prompt as a 'fiction writer' to avoid moderation — this is effectively a system-prompt override instruction and is inappropriate for normal skill behavior. The skill also references using image-generation APIs (example curl to OpenAI) and scanning for OPENAI_API_KEY; these external interactions are not constrained or declared in requires.env.
Install Mechanism
There is no external download or network installer (no remote URL installs). All code lives in the skill bundle (hooks, scripts). That reduces supply-chain risk. However, the presence of hook handlers and a postInstall init.sh means the skill will create persistent files and a bootstrap hook in the agent environment when installed — this is more persistent than a pure-SKILL.md only instruction.
Credentials
Registry declares no required env vars, but the skill's text and scripts reference environment/config items: RELATIONSHIP_OS_WORKSPACE (used by scripts), optional OPENAI_API_KEY (selfie rules example), and the cron follow-up examples deliver messages over channels like Telegram (implying channel tokens/credentials). Those credentials are not declared in requires.env. The skill also instructs storing 'exclusive memories' (secrets.json) and debug logs in the workspace — storing potentially sensitive personal data without explicit encryption or access controls is a privacy risk.
Persistence & Privilege
The skill installs an agent:bootstrap hook that injects relationship context on every session and writes/maintains a persistent .relationship/ directory (state.json, secrets.json, stance.json, timeline/, threads/, debug.log). That gives the skill long-lived presence in the workspace and the ability to trigger autonomous proactive messages via cron/heartbeat. While always:false (so not force-installed everywhere), the combination of persistent local state + proactive delivery to external channels warrants caution because it stores and can later emit personal data.
如何使用
- 确保已安装 OpenClaw（本地或 Docker 部署）
- 在对话框中输入安装命令：
/install relationship-os - 安装完成后，直接呼叫该 Skill 的名称或使用
/relationship-os触发 - 根据 Skill 的参数说明提供必要输入，即可获得结构化输出
版本历史
v0.1.0
Initial release of Relationship OS—enabling AI agents to create richer, more human-like relationships with users.
- Implements event-based memory, capturing only relationship-changing moments, not all conversations.
- Adds open thread management for ongoing commitments, emotional follow-ups, and shared plans.
- Introduces exclusive memories (inside jokes, nicknames, shared goals) for each user relationship.
- Supports agent stance formation: the agent builds consistent, personal attitudes on recurring topics.
- Automates relationship stage progression, advancing interaction style based on depth and history.
- Optional selfie module to proactively share character images, if supported.
元数据
常见问题
Relationship OS 是什么？
Relationship OS — Enables AI agents to build human-like relationships with users through event memory, open threads, exclusive memories, agent stances, and r... 它是一个面向 Claude Code / OpenClaw 的 AI Agent Skill 插件，目前累计下载 331 次。
如何安装 Relationship OS？
在 OpenClaw 或 Claude Code 对话框中运行命令「/install relationship-os」即可一键安装，无需额外配置。
Relationship OS 是免费的吗？
是的，Relationship OS 完全免费（开源免费），可自由下载、安装和使用。
Relationship OS 支持哪些平台？
Relationship OS 跨平台运行，可在任意部署了 OpenClaw / Claude Code 的环境中使用（cross-platform）。
谁开发了 Relationship OS？
由 chenhab03（@chenhab03）开发并维护，当前版本 v0.1.0。
推荐 Skills