FeishuCLassmate

项目及进度管理全流程支持，包括项目信息录入、甘特图自动更新、当日DDL主动提醒（temi线下优先，飞书兜底）、与项目进度有关的帮扶提醒。所有文档、甘特图、项目表格均存储在飞书【Lab知识库】公共协作区。使用场景：创建新项目、更新项目进度、查询项目甘特图、查看项目进度、设置项目提醒、与项目卡点帮扶相关的聊天。

让 OpenClaw 作为实验室「科研协作 Agent」：先读飞书云文档知识库中各同学项目文档再头脑风暴、 预研课题发现、研究计划与任务拆解、飞书多维表/甘特时间线推进；判断同学是否适合打扰时 **以 Temi sidecar 的 POST /photo 场景快照为准**，不用飞书用户在线/日历状态作空闲依据；长期记忆迭代。 **触发**: 「预研课题」「科研脑暴」「读知识库项目再讨论」「拆研究计划」「甘特/时间线」「谁在现场方便聊」等。

**路由入口** — 实验室参观全链路拆分为两段 skill,避免单文件过长: **准备讲解稿** → `lab-tour-prepare`; **现场 Temi 带览** → `lab-tour-run`。 触发词:「带参观」「导览」「tour」「来人了」「有访客」「参观开始」。 若用户明确「提前写稿/生成文档」→ **lab-tour-prepare**; 若「人已到现场/机器人开始讲」→ **lab-tour-run**(可附带 doc 链接)。

智能会议预约与会前提醒。使用场景：导师通过飞书发送会议需求后，自动读取参会人日程并安排共同空闲时间，创建会议并在临近会议时优先通过Temi线下寻人提醒，未找到则飞书私信提醒。

教师/管理员为来访访客「预生成」个性化导览讲解稿时使用。 读取知识库中的导览母版，自动识别访客单位信息并联网搜索公司背景，结合单位属性生成高度贴合访客兴趣的定制化导览稿，新建一篇飞书 Doc 作为当次剧本。 **触发**: 「给 XX 单位准备导览」「生成今天的参观讲解稿」「预生成导览文档」。 **个性化特性**: 自动识别用户输入中的公司名称，联网搜索公司主营业务、行业方向、技术需求、近期动态等信息，分析访客兴趣点后定制讲解大纲，重点展示与访客匹配的实验室成果。 **不要用于**: 访客已到现场的边走边讲 —— 请用 skill `lab-tour-run`。

访客到达现场后的「执行期」实验室导览:Temi 走位 + 播报 + 工位项目补充 + 问答。 **前置**:已有当次讲解 Doc(由 skill `lab-tour-prepare` 生成),或使用通用内置剧本。 **触发**: 「开始导览」「参观开始」「飞书同学 XX 到了」、用户提供 doc 链接/token。 **不要用于**: 仅在飞书里写稿不发机器人 —— 用 `lab-tour-prepare`。

Academic paper search with per-student (open_id) preference memory, Feishu identity resolution, and iterative query/refinement; see references/.

自动读取飞书LAB知识库项目文档，自主确定研究课题，联网检索前沿案例和方向，生成研究报告辅助团队。Invoke when: 每周自动触发/用户说「启动自主研究」/项目有重大进展或卡点时

Use after confirming a system is in mock/simulation mode. Stop attempting real connections and either resolve the issue or ask the user for help.

Use when launching background processes or long-running commands. Maintain a clear state map of which processes are running and their expected outcomes.

Use when starting servers, daemons, or any process expected to run indefinitely. Always use `background=true` to prevent timeout failures.

Use when repeatedly searching for the same or related strings across multiple grep/read commands. Consolidate into a single comprehensive search to avoid redundant queries and session loss.

Use when about to run Node.js/npm commands in a project directory. Verify dependencies and module availability first to avoid MODULE_NOT_FOUND failures.

Use when attempting to start a server or service that connects to an existing robot endpoint, to avoid conflicts from duplicate processes.

Use when interacting with a robot via sidecar or WebSocket. Always communicate status changes and errors to the user instead of returning NO_REPLY.

When system messages indicate fallback behavior (e.g., mock mode activation), explicitly inform the user rather than treating it as routine.

Use after receiving test output to explicitly confirm success criteria are met and communicate a clear pass/fail status to avoid ambiguity.

When asked if a capability exists (e.g., controlling a robot), verify actual tool/skill availability before responding.

Use when multiple background processes are running. Consolidate into a single workflow to avoid process sprawl and confusion about which session holds the relevant output.

When a robot command fails with a connection error, diagnose the specific issue (IP, port, protocol) before searching for alternative code paths.

Use when a robot movement command fails or the robot doesn't move. Always diagnose the root cause (connection state, WebSocket error, command parameters) before retrying.

When WebSocket or network connection to robot fails, run connectivity diagnostics before attempting the same action again.

Use when robot commands fail with identical errors across multiple attempts. Recognize when failures are persistent rather than transient network glitches.

Use when a robot connection fails and the system falls back to mock/simulation mode. Clearly communicate the connection failure to the user with actionable next steps.

Use when robot connectivity fails and diagnostic attempts do not resolve the issue. Ask the user to manually verify the robot's status rather than continuing automated diagnostics.

When encountering Feishu/Lark API permission errors, provide a clear step-by-step action plan for the admin, not just a URL.

Use when robot command retries have been exhausted. Present a clear diagnostic summary to the user and suggest next steps rather than continuing to retry.

When user provides minimal but actionable project/task requirements, use sensible defaults for missing details instead of asking excessive clarifying questions.

Use when the same connection attempt fails 3+ times consecutively, to stop and report the persistent failure to the user instead of continuing retry attempts.

When an async command completes with a failure signal (SIGKILL, SIGTERM, or falling back to mock mode), always notify the user instead of returning NO_REPLY.

Use when user messages contain dates or timestamps with vague references like '4/18' or '阶段目标' without explicit instructions. Ask clarifying questions instead of guessing intent.

Use when receiving connection errors mentioning "WebSocket", "Upgrade", or "Sec-WebSocket-Key" to understand the protocol mismatch and take appropriate action.

Use when credentials, tokens, or API keys appear in command output. Truncate or mask them before returning results to avoid leaking sensitive data.

Use when a diagnostic method succeeds where the primary method failed. Switch to the successful method as the new working baseline.

Use when making HTTP requests to an API or service. Always verify the endpoint exists by checking documentation or probing the base URL first.

Use when exploring tool or system capabilities. Always check built-in help/docs (--help, -h, README) before using exec commands to discover functionality.

Use when a shell session disappears mid-command or reports 'No active session found'. Follow systematic recovery steps before reissuing commands.

When looking up skill files for execution, use the correct workspace-relative path and verify the skill exists before reading.

Use when a robot command fails with a WebSocket 1001 error or connection-drop error. Perform connectivity diagnostics before blindly retrying the same command.

When user requests a specific robot action (move, speak, stop), immediately invoke the corresponding function without reading documentation or searching for code first.