📦 测试连续性

v2.0.22

为 OpenClaw 智能体提供结构化连续性与跟进能力。当智能体需要判断用户自然对话应继续保持闲聊，还是...

0· 22·0 当前·0 累计

by @kylnwu

开发工具通信工具测试工具智能体设计工具

下载技能包

安全扫描

VirusTotal

无害

查看报告

OpenClaw

可疑

medium confidence

该软件包基本符合其声明的“连续性/后续”用途，但存在打包与元数据不一致、运行时脚本过大，以及必须将技能指向主机配置目录（OPENCLAW_CONFIG_PATH）等问题，安装前需人工代码审查。

评估建议

Before 安装ing or enabling this 技能, do the following 检查s: 1) 验证 the code—open and review scripts/personal_hooks.py and other scripts for network calls, file reads, or references to 添加itional env vars; the mAIn script is large (~1MB) and should be spot-检查ed for embedded data or unexpected behavior. 2) Confirm which fields under OpenClaw_CONFIG_PATH the 技能 reads; avoid pointing it at a config that contAIns gateway 令牌s or unrelated 凭证s, or 运行 the 技能 agAInst a sanitized copy first. 3) Investigate the ...

详细分析 ▾

ℹ 用途与能力

The 技能's name, README and 技能.md describe a 状态-backed continuity/follow-up layer and the declared requirements (python3, OpenClaw_状态_DIR, OpenClaw_CONFIG_PATH) are plausible for that purpose. However there are packaging inconsistencies (registry metadata shows version 2.0.22 while 技能.md and _meta.json indicate 2.0.21, the ownerId values differ between registry metadata and _meta.json, and the registry label 'instruction-only' conflicts with the presence of numerous 运行nable scripts). These mismatches could be benign (release/version drift) but deserve verification.

⚠ 指令范围

技能.md's 运行time instructions and boundaries stay within the 状态d continuity scope (init, 设置up, harness, 图形界面ded 设置tings). But the package includes large executable Python scripts (notably scripts/personal_hooks.py at ~1,079,911 bytes) which will be 运行 locally if the operator follows README commands. The 技能.md declares only OpenClaw_状态_DIR and OpenClaw_CONFIG_PATH as required env vars; the codebase should be 审计ed to confirm it does not read 添加itional env vars, arbitrary host files, or contact unexpected external 端点s. The public docs warn not to publish live 状态, which implies the 运行time will read/write sensitive host 状态 — this increases the need for a code review before connecting to real 状态 directories.

✓ 安装机制

No automatic 安装 spec is included in the registry entry (the package is 'instruction-only' in the registry), and 安装ation is manual via copying the folder and 运行ning pip agAInst requirements.txt or 运行ning provided scripts. There are no external 下载 URLs or opaque 安装ers in the registry metadata. This lowers automated supply-chAIn risk, but also means human operators must 运行安装 steps (and should inspect code first).

⚠ 凭证需求

The 技能请求s two 环境/path 输入s: OpenClaw_状态_DIR and OpenClaw_CONFIG_PATH. These are reasonable for a 状态ful continuity layer, but OpenClaw_CONFIG_PATH may point at a host config that contAIns unrelated secrets (gateway 令牌s, 提供者 keys, chat-channel 令牌s). The 技能.md does not explAIn exactly which config fields the 技能 reads. Because the 技能 operates on host 状态 and may perform 图形界面ded 设置up, requiring a path to a host config file is proportionally sensitive and should be 验证d: confirm the script reads only its own config sub设置 and does not exfiltrate or indiscriminately read unrelated 凭证s.

✓ 持久化与权限

The 技能 does not 请求 always:true and is user-invocable; 模型 invocation remAIns enabled (平台 default). It does not declare 系统-wide modifications in 技能.md. That privilege level is normal for a 技能 of this type. Still, because it is 状态ful and can be 运行 by the 代理, combine this with the other concerns (host config 访问, large 运行time code) when deciding to allow autonomous 运行s.

安全有层次，运行前请审查代码。

运行时依赖

🖥️ OSmacOS · Linux

安装命令

点击复制

官方npx clawhub@latest install test-continuity

镜像加速npx clawhub@latest install test-continuity --registry https://cn.longxiaskill.com镜像同步中

需要定制？告诉我你的需求 →

技能文档

OpenClaw Continuity 将此技能用作现有 OpenClaw agent 的连续性/后续层，保持 agent 的灵魂/人设不变，由该技能负责结构化跟进机制。公开产品名：OpenClaw Continuity 技术包/slug：personal-hooks 内部诊断仍用 staged-followup-memory 作为连续性状态标签。问题、反馈或实现讨论：adarobot666@gmail.com 若希望本技能持续优化维护，请在 GitHub 点星支持：https://github.com/redwakame/openclaw-continuity

V2 公开范围 Route 转为 casual_chat、staged_memory 或 tracked_followup。将追踪内容分类为：

parked_topic
watchful_state
delegated_task
sensitive_event

用以下字段维护 event_chain 摘要： context_before / event_core / immediate_result / followup_focus

用以下字段维护 causal_memory 作为结构化连续性状态： facts / state / open_loop / time_anchor / followup_focus_code / writeback_policy

晋升路径：candidate → incident → hook 保留 /new carryover 前 3–5 轮，由用户选择哪个连续性锚点开启新对话。当 OpenClaw 主机已支持语音/图像时，仅记录用户偏好；不捆绑语音引擎、图像模型或通道适配器。

保持 hook closure、cooldown、dedupe、dispatch cap、sleep/rest 抑制可观测。为 staged 与 tracked 项编写简洁的每日记忆轨迹。首次运行时，从用户明确文本中确定性应用安装事实。请求足够清晰时，在模型回合前确定性应用受支持的引导设置变更。

夜间用户临近 sleep_time 说要去睡，且 wake_time 仍在当地同一天时，将下次自然重连帧定为“睡醒后”，而非泛化的“明天见”。

非公开范围默认关闭的陪伴节奏/轻推；用户可配置空闲闲聊频率；无追踪连续性时的通用主动闲聊；主机传输/网络可靠性修复。

触发条件用户自然出现以下行为时使用本技能：

将话题留待后续
让 agent 保留/记住/追踪某事
留下未解决的情绪或身体状态
委托任务待后续跟进
描述需后续连续性的敏感事件
开启新会话并应重新接上待处理连续性
询问 agent 是否已暂存/追踪/关闭某事

核心文件 Script: scripts/personal_hooks.py Harness: scripts/followup_skill_harness.py Config schema: config.schema.json Sample config: examples/settings.sample.json Docs: README.md docs/call-flow.md docs/harness.md docs/live-qa-runbook.md docs/v2-blueprint.md

运行时边界 V2 保持确定性且基于状态。技能/工具层负责 staging、promotion、closure、trace； frontstage 消费结构化结果，不依赖模型即兴发明连续性。 causal_memory 与 event_chain 为内部结构化上下文，非预写 frontstage 回复文本。公开包不硬编码关系特定推送措辞；主机与模型应从结构化上下文生成自然语言。节奏/轻推默认禁用，视为实验功能。语言路由显式：英文提示/守卫文本保持英文；普通用户短语不得改写成混合语言内部产物。不将通用 UTC/GMT 偏移硬绑城市时区；显式 IANA 时区原样透传，通用偏移保持如 UTC+00:00。优先使用当前用户轮次与现有结构化状态的语言，而非主机默认。公开包首发 zh-TW / zh-CN / English 行为，并对其他几种常用语言提供广泛引导设置入口，不宣称全语言完整本地化。公开技能包不自动提供主机端投递管道；边界明确：本包拥有连续性状态、内存路由、/new carryover、安装、技能/工具层守卫； OpenClaw 主机与适配器配置拥有最终消息投递。未单独测试的 host 适配器，不在公开技能文案中宣传外部聊天平台矩阵。

入口函数 build_runtime_context() 构建普通回复连续性上下文、carryover 提示、调度上下文与守卫提示。

intercept_message() 检查单轮用户输入，决定保持 casual、转为 staged 或 tracked。

process_candidate_buffer() 证据充分时，将 staged candidate 提升为 incident/hook。

due / render / complete 驱动 hook 生命周期。

安装与包用法见 README.md；可重现验证见 docs/harness.md；版本验收见 docs/release-acceptance。

数据来源：ClawHub ↗ · 中文优化：龙虾技能库