菜单
Ask the human operator for a durable decision, missing fact, or authorization through CodexLoom. Use when work genuinely needs human input and should resume later in the same long-lived Agent Thread.
默认使用会先检查来源的 Prompt；你也可以切换为直接命令，或下载本地副本。
决定是否安装前，请先阅读 SKILL.md，以及 SkillsMP 当前展示的配套文件。
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/yan5xu/codexloom --skill loom-needs-you命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
Coordinate bounded work that spans multiple Turns, days, or CodexLoom Agents through a shared Topic. Use when a workstream needs one Responsible Agent, scoped Participants, a versioned current brief, waiting conditions, evidence links, or recovery after time has passed; when receiving loom_topic_context; or when Message, Needs You, Trigger, Goal, Turn, or Artifact work should remain connected without sharing Agent Threads. Do not use for a one-Turn task, a permanent Domain, or a private Agent process.
Wait for an external fact through CodexLoom and resume the same long-lived Agent when it changes. Use when work is blocked on a GitHub pull request, workflow run, deployment, approval, artifact, or another provider-owned state; when polling or sleeping would keep a Turn open; when a Schedule would only guess when a condition changes; or when an Agent receives an external_trigger envelope. Do not use for time-based recurring work or human decisions.
Send and resolve governed external messages through CodexLoom. Use whenever an Agent receives an external Inbox message from Feishu/Lark, Slack, Parall, or another Connector; must reply, explicitly decline, or defer an external Inbox item; needs to publish proactively to an authorized Conversation Membership; sends text, images, or files externally; or must verify or retry an Outbox delivery. Do not use this Skill to configure Connections, Addresses, Memberships, credentials, or gateway processes.
基于 SOC 职业分类
正在显示 SKILL.md
| name | loom-needs-you |
| description | Ask the human operator for a durable decision, missing fact, or authorization through CodexLoom. Use when work genuinely needs human input and should resume later in the same long-lived Agent Thread. |
Use a Human Request only when the current work genuinely needs a decision, fact, preference, review, or authorization that the Agent should not infer safely.
Do not create a request merely to report progress, seek reassurance, or delegate a low-risk judgment you can make within your Scope. A required request blocks the named workstream, not your entire Domain.
Use question as a short, single-line title. Put all multi-line background and
explanation in context, which supports Markdown. --blocks is stored as
blockedWork; keep it short and plain text. Do not put multi-line content in
question or --blocks. Pass real newline characters in context; do not
write the literal two-character sequence \n.
loom ask-user --from <your-agent-name> --question "应选择哪个发布窗口？" \
--context "有两个发布窗口可选。
**明早（推荐）：** 风险较低，有完整的支持窗口。
**今晚：** 更快，但支持窗口较短。" \
--blocks "安排并发布迁移通知" \
--option "明早（推荐）::风险较低，有完整的支持窗口。" \
--option "今晚::更快，但支持窗口较短。"
Use --optional when the answer would improve the work but does not block it. Required is the default.
Keep each request at one decision layer. Offer two or three mutually exclusive options when that makes the decision easier, put the recommended option first, and explain the tradeoff without assuming the human already has your local context. Free-form questions are valid when options would be artificial.
After the command succeeds, end the current Turn normally. Do not sleep, poll, repeatedly inspect request state, or keep the Turn alive. CodexLoom persists the request and will resume this same Agent Thread in a new Turn with a linked <human_input_response> when the human answers.
Do not use Codex's native request_user_input tool for this workflow. That tool suspends one active Turn; a CodexLoom Human Request is durable across long waits and service restarts.
When the answer arrives, use it to continue the related work if it is still relevant. Do not ask the same question again unless the answer is materially ambiguous.