| name | sumi-run |
| description | Runs one sumi task by ID. Invoked by scheduled triggers (Routines, Claude Desktop scheduled tasks, Codex Automations) or by `/sumi-run <task-id>` on demand. Also triggers when the user says "run sumi <id>", "run my weekly sumi", "run the team update", "generate my daily standup", "run the exec brief now", "draft this week's customer digest now", or any variant referencing a specific sumi-managed task. The task ID is passed as an argument; the skill looks it up in ~/.sumi/config.json, pulls from configured sources for the cadence-appropriate window, consolidates using the configured template and notes preset, and delivers the draft to the configured destination. If config is missing or the task doesn't exist, defer to the `sumi` manager skill. |
Sumi — Run
Slash command: /sumi-run <task-id>. Also invoked by scheduled triggers with the task ID as argument.
Execute one configured summary task. Designed to run unattended on a schedule — assume no human is watching. Fail loud (visible error at the configured destination) rather than silently.
Full procedure lives at references/run-procedure.md. Read that file for the complete step-by-step — the run logic lives there so this skill body stays scannable.
Entrypoint
The task ID comes in as an argument ($ARGUMENTS when invoked by a scheduled trigger, or from the user's natural-language request).
- Read
~/.sumi/config.json.
- Locate
tasks[] entry where id === <task-id>. If not found: write a "task not found" error to the default destination (Gmail draft-to-self if Gmail is connected, else local file at ~/.sumi/drafts/_errors/YYYY-MM-DDTHH-MM.md) and stop. Do not silently fail.
- If the task has
paused: true: stop without error.
- Execute the pull → consolidate → deliver procedure described in
references/run-procedure.md.
Disambiguating when the user is vague
If the user says "run my summary" with no task ID:
- If exactly one task exists, use it without asking.
- If zero tasks exist, defer to
sumi to set one up.
- If multiple tasks exist, list them with display names + cadence labels and ask which one.
Dry-run behavior
If invoked via "run the weekly summary" / "run summary now" / "dry run the team update", behave identically to a scheduled invocation — post to the real destination. Users treat the destination as a staging area; there's no benefit to redirecting to chat.
Context hygiene
When drafting, use any relevant context the user has shared in this conversation or prior conversations (codenames, team names, stakeholders, projects, company terminology). Do not ask the user to restate these at run time — they should never see a prompt from a scheduled run.
