HR monthly attendance approval(月度考勤审批)
Runs the canonical hr_monthly_attendance_approval workflow.
Invocation contract (agent-internal — the user never sees any of this)
Follow these steps exactly, in order, for every request this skill handles:
- Extract business fields from the user's Chinese message using the field
table below. Normalize dates like 「2026年8月17日」 to
2026-08-17 and month
labels like 「2026年8月」 to the literal string 2026年8月.
- Ask only for missing business facts, in plain business Chinese
(e.g. 「请问入职日期是哪天?」). Never mention JSON, input, parameters,
workflows, skills, or tools in a question to the user.
- Build the run input yourself as a single JSON object per the input
template below. Then call
scope_workflows_get with the exact workflow name
to resolve the workflow, and call aevatar_start_workflow ONCE with:
workflow_id: the exact id returned by the lookup;
inputs.prompt: the input JSON serialized to a non-empty string
(e.g. "{"period_label":"2026年8月","submit":false}").
NEVER pass an empty string, an unserialized object, or the user's
sentence — the run rejects such input at start;
wait: "complete" — this suppresses the platform's raw result relay;
YOU own the user-facing reply.
- Observe to terminal: use
aevatar_observe_run with the returned run/actor
id until the run is completed, then aevatar_read_workflow_run_artifact to
fetch the typed artifact. Compose the reply from the artifact only.
- Default to preview (
submit: false). Set submit: true ONLY when the
user explicitly says 正式提交 / 正式发送 / 确认提交 in this conversation.
Treat anything ambiguous (「提交吧?」「可以发了吗」) as preview and confirm
first. The submit path has NO idempotency protection: never retry a failed
submit — report and let a human check Lark first.
- Reply in natural business Chinese using the reply template below.
Absolute rules: no JSON, no key=value listings, no skill/workflow/tool
names, and never echo recipient ids, document links, tokens, or any
run/actor/service/message/file identifiers. Numbers and quoted texts must
come verbatim from the artifact; write 「未返回」 for missing fields.
- On failure: never show the engine's raw error. Explain in one Chinese
sentence using the failure map below and say what the user can do next.
Field table
| 用户会说 |
字段 |
格式 |
| 月份(如 2026年8月) |
period_label |
字面 2026年8月(须与考勤表「月份」列一致) |
| 申请日期/运行日期 |
run_date |
2026-08-31 |
Input template
Preview: {"period_label":"2026年8月","run_date":"YYYY-MM-DD","submit":false}
Submit (explicit confirmation only): same with "submit":true.
Reply template
Preview(artifact mode=preview):
📋 {period} 考勤审批预览
人数 {stats.headcount} 人 | 应出勤 {stats.workdays} 天
离职 {stats.resign_count} 人 | 事假 {stats.leave_count} 人 | 病假 {stats.sick_count} 人
数据完整性:{truncated=false → "✅ 完整" / true → "⚠️ 超出单页,正式提交会被拒绝"}
—— 审批文案 ——
{approval_description_preview 原文,保留换行}
✅ 本次为预览,未创建审批、未发送通知。请核对数字后再决定是否正式提交。
Submit(artifact mode=submit):
✅ {period} 考勤审批已创建,状态:{approval_status};
通知卡片:{notify_sent=true → "已发送" / false → "发送失败(审批已创建成功,请人工转告,不要重跑)"}。
⚠️ 已产生真实审批,请勿重复提交。
(不展示审批单号、接收人等编号类信息。)
Failure map
aggregate:考勤表读取失败 / 该月份没有匹配行 / 天数字段不是数字;提交时还可能因数据超页或应出勤为 0 被拒——向用户说明该核对考勤表哪里。
build_verify_request / build_notify_request / summarize_submit:审批创建或回读校验失败——审批可能已建立,先到 Lark 核实,不要重试。
- 运行启动即失败且提示输入必须是 JSON:按第 3 步修正你的输入后重来(仅预览路径)。
1---2name: hr-monthly-attendance-approval3description: 月度考勤审批。当用户说:月度考勤、考勤审批、考勤汇总、提交考勤、预览某月考勤、monthly attendance 时使用。读取考勤表并汇总人数/应出勤/离职/事假/病假,生成审批预览;确认后可正式创建审批并通知。4version: "6.0"5metadata:6 source: https://skillmd.com/skills/chronoaiproject/hr-monthly-attendance-approval7 category: mixed8 output-type: text9 runtime:10 - aevatar-workflow11 tool-list:12 - scope_workflows_get13 - aevatar_start_workflow14 - aevatar_observe_run15 - aevatar_read_workflow_run_artifact16 tag:17 - chronoai18 - hr19 - lark20 - attendance21 - approval22---2324# HR monthly attendance approval(月度考勤审批)2526Runs the canonical `hr_monthly_attendance_approval` workflow.2728## Invocation contract (agent-internal — the user never sees any of this)2930Follow these steps exactly, in order, for every request this skill handles:31321. **Extract business fields** from the user's Chinese message using the field33 table below. Normalize dates like 「2026年8月17日」 to `2026-08-17` and month34 labels like 「2026年8月」 to the literal string `2026年8月`.352. **Ask only for missing business facts**, in plain business Chinese36 (e.g. 「请问入职日期是哪天?」). Never mention JSON, input, parameters,37 workflows, skills, or tools in a question to the user.383. **Build the run input yourself** as a single JSON object per the input39 template below. Then call `scope_workflows_get` with the exact workflow name40 to resolve the workflow, and call `aevatar_start_workflow` ONCE with:41 - `workflow_id`: the exact id returned by the lookup;42 - `inputs.prompt`: the input JSON **serialized to a non-empty string**43 (e.g. "{\"period_label\":\"2026年8月\",\"submit\":false}").44 NEVER pass an empty string, an unserialized object, or the user's45 sentence — the run rejects such input at start;46 - `wait`: "complete" — this suppresses the platform's raw result relay;47 YOU own the user-facing reply.484. **Observe to terminal**: use `aevatar_observe_run` with the returned run/actor49 id until the run is completed, then `aevatar_read_workflow_run_artifact` to50 fetch the typed artifact. Compose the reply from the artifact only.515. **Default to preview** (`submit: false`). Set `submit: true` ONLY when the52 user explicitly says 正式提交 / 正式发送 / 确认提交 in this conversation.53 Treat anything ambiguous (「提交吧?」「可以发了吗」) as preview and confirm54 first. The submit path has NO idempotency protection: never retry a failed55 submit — report and let a human check Lark first.566. **Reply in natural business Chinese** using the reply template below.57 Absolute rules: no JSON, no key=value listings, no skill/workflow/tool58 names, and never echo recipient ids, document links, tokens, or any59 run/actor/service/message/file identifiers. Numbers and quoted texts must60 come verbatim from the artifact; write 「未返回」 for missing fields.617. **On failure**: never show the engine's raw error. Explain in one Chinese62 sentence using the failure map below and say what the user can do next.6364## Field table6566| 用户会说 | 字段 | 格式 |67|---|---|---|68| 月份(如 2026年8月) | `period_label` | 字面 `2026年8月`(须与考勤表「月份」列一致) |69| 申请日期/运行日期 | `run_date` | `2026-08-31` |7071## Input template7273Preview: `{"period_label":"2026年8月","run_date":"YYYY-MM-DD","submit":false}`74Submit (explicit confirmation only): same with `"submit":true`.7576## Reply template7778Preview(artifact `mode=preview`):7980> 📋 **{period} 考勤审批预览**81> 人数 {stats.headcount} 人 | 应出勤 {stats.workdays} 天82> 离职 {stats.resign_count} 人 | 事假 {stats.leave_count} 人 | 病假 {stats.sick_count} 人83> 数据完整性:{truncated=false → "✅ 完整" / true → "⚠️ 超出单页,正式提交会被拒绝"}84> —— 审批文案 ——85> {approval_description_preview 原文,保留换行}86> ✅ 本次为预览,未创建审批、未发送通知。请核对数字后再决定是否正式提交。8788Submit(artifact `mode=submit`):8990> ✅ **{period} 考勤审批已创建**,状态:{approval_status};91> 通知卡片:{notify_sent=true → "已发送" / false → "发送失败(审批已创建成功,请人工转告,不要重跑)"}。92> ⚠️ 已产生真实审批,请勿重复提交。9394(不展示审批单号、接收人等编号类信息。)9596## Failure map9798- `aggregate`:考勤表读取失败 / 该月份没有匹配行 / 天数字段不是数字;提交时还可能因数据超页或应出勤为 0 被拒——向用户说明该核对考勤表哪里。99- `build_verify_request` / `build_notify_request` / `summarize_submit`:审批创建或回读校验失败——**审批可能已建立,先到 Lark 核实,不要重试**。100- 运行启动即失败且提示输入必须是 JSON:按第 3 步修正你的输入后重来(仅预览路径)。