News Sentiment Check

Uses the check.ai: structured AI check to read recent news and classify the overall tenor. The when: arms then branch on the structured outputs the AI returned.

What it does

On each fire, the monitor's check is NOT a tool call — it's a direct AI call. The check:

1. Sends a prompt describing what to look for.
2. Declares an outputs: map naming the structured fields the AI must return (with each field's description as a hint).
3. Gets a JSON object back with those keys filled in.
4. Each key becomes a binding the when: arms can reference directly.

The arms then evaluate the structured outputs:

• Arm 1 — urgency >= 4: high-urgency catch-all. Notifies regardless of tenor.
• Arm 2 — tenor == bearish: caught only if urgency was 1-3 (arm 1 didn't match first). Notifies on bearish tenor.
• Else — quiet day or neutral/bullish tenor below the urgency threshold.

Edit before installing

In monitor.yaml:

• check.ai.prompt — rewrite for your domain ("today's crypto news", "ops channel chat from the last hour", "internal status reports"). Be specific about what you want classified and what dimensions matter.
• check.ai.outputs — declare the structured fields you want. The description text matters — it's literally the hint the AI gets for what to put in each field. Be precise.
• The when: arms reference outputs by name (e.g. tenor == bearish, urgency >= 4). Match these to whatever you declared in outputs:.

Bindings used

These come from check.ai.outputs, not from a tool. The names are whatever you declare:

• {{summary}} — paragraph string the AI wrote
• {{tenor}} — one-token classification: bullish | bearish | mixed | quiet
• {{urgency}} — integer 1-5

Add or rename fields by editing check.ai.outputs: — the bindings update automatically (they're named after the keys you declared).

Requirements

• AI assistant configured (Server Settings → AI Assistant — provider, API key, model). If no AI is configured, the check fails soft and the arms never fire.

Pattern showcase

check.ai is the structured-output AI check. Unlike an ai: arm (which returns true/false), check.ai returns a map of named fields that the arms then evaluate against. Use it when:

• The thing you're watching is fuzzy and can't be a tool call ("is the news bearish today?", "does the support inbox have an escalation?").
• You want a SINGLE AI call per fire, then cheap arm evaluation.
• You want structured downstream routing (different arms for different outputs).

Tokens: 1 LLM call per fire. Schedule accordingly.