The morning KPI digest that writes itself

Trigger: cron, weekdays 07:00 · Pattern: fetch → narrate → deliver · Sample: examples/use-cases/exec-digest.toml · Status: runs today (intel-remote,tools-http-tls,trigger-cron)

The problem

Every leadership team has a morning ritual: someone opens four dashboards, squints, and types "quick numbers update 🧵" into Slack. The numbers were always available; the narrative — what moved, why it matters, what to watch — is the part that costs a person's morning. And the days it gets skipped are, by Murphy, the days something moved.

What the agent does

At 07:00 on weekdays, with no inbound network surface at all:

1. GET the KPI snapshot from your metrics API.
2. One bounded LLM step turns the JSON into a three-section markdown brief: What moved (numbers and direction), Why it matters (one insight each), Watch today (max three items).
3. Post it to the leadership Slack channel and file a dated copy on disk — the searchable archive of every morning's story.

The model sees data and writes prose. The schedule, the data source, the destination, and the spend are all declared in the workflow, out of its reach.

[[triggers]]
type = "cron"
schedule = "0 7 * * 1-5"
start_node = "morning"

The quiet advantages of the boring architecture

• It's a daemon with one job. No inbound HTTP routes exist in this workflow — the only way in is the clock. Attack surface: zero listening sockets beyond /healthz.
• The cost is a budget, not a surprise. max_llm_tokens = 25000 per run, every run. A digest costs what it costs, predictably — multiply by 21 workdays and that's the monthly line item, known in advance. The conformance suite's cost forecasting does that arithmetic for you.
• A bad morning fails loudly, not weirdly. Metrics API down → declared fail node with a clear reason in the audit log, and no digest — instead of a hallucinated digest about data that didn't arrive. The when = "error" edge from the fetch is the difference.

Variations that stay in bounds

• Pull from several sources and let a parallel node fan the fetches out concurrently (see content-localization for the pattern).
• Weekly board pack: same graph, schedule = "0 16 * * 5", a longer prompt, and the file becomes the draft your CFO edits instead of writes.
• Anomaly-only mode: add a diff_compute gate like the churn monitor's so quiet days post nothing — alert fatigue is a design choice, not a fate.

Honest limits

Slack's incoming webhooks happily accept the brief; richer Block Kit formatting means shaping the POST body precisely, which template_render handles. What the runtime won't do is be the dashboard — it writes the morning's story; your metrics stack keeps the charts.