Have GPT-6 Astra watch severe weather alerts

TX · 8 active alerts

Heat Advisory    Montague, Cooke, Grayson, Wise, Denton, Collin, Tarrant, Dallas …
Heat Advisory    Hunt, Delta, Hopkins
Heat Advisory    Fannin, Lamar
Heat Advisory    Alfalfa, Grant, Kay, Major, Garfield, Noble …
Heat Advisory    Sevier, Howard, Little River, Hempstead, Nevada, Miller …
Flood Warning    Hardin, TX

That is a real session, unedited: one question in, one free data call, a readable summary out. The Actionway weather alerts capability returns the same records the U.S. alert feed publishes — event, area, timing, full text — as structured JSON, so the agent groups and rewrites them instead of scraping a page.

What happened in between

OpenAI released GPT-6 Astra on September 3, 2026 around long-running agent tasks, and since Codex CLI 0.153.4 it is the bundled default model. Watching a feed for hours is that kind of task: check, compare with the last check, rewrite the page, wait, repeat. What the model needs from outside is a trustworthy feed, and the Skill installed with the Actionway CLI is how it finds one.

Before answering, the agent checked the capability index it keeps locally, saw that the weather services are free, and said so with the options. A free call needs no wallet check, so it ran immediately; a priced service would have shown its price alongside and run at that price once picked. The response is JSON, so the agent can diff it against the previous hour's copy, keep only what changed, and rebuild a status page — an HTML file, a Slack-ready text, a CSV — as ordinary local code work that costs nothing.

Turn one check into a watch

The loop is the agent's job: ask for the alerts, fetch a single alert's full text with the detail service when a line needs it, filter by event type with the alert types list, and write the page. Because every response is structured and stored, the page can cite itself — which alert, which counties, fetched when — and a reader can trust it without re-checking the source. The stock and company data services follow the same pattern when the feed you need is a market instead of a storm.

Free calls still run through the same catalog, so nothing about the setup changes when you add a priced service later: the agent lists it with its price, and picking it is the approval. Only runs the platform flags for extra confirmation pause first, and an empty wallet stops a priced run before anything is charged.

What the hourly page should contain

A status page is only useful if a reader can trust it without opening the source, so ask the agent to put four things on it: the time of the last check, the list of active alerts with event and counties, what changed since the previous check (new, expired, updated), and a link or identifier for each alert so anyone can pull the full text. Keep the previous hour's JSON beside the page; the "what changed" section is a diff of the two files, which is ordinary local code and never a paid call.

Decide up front what counts as worth a notification. Eight heat advisories on a July afternoon are expected; a new flood warning is not. Tell the agent which event types matter — the free alert types list gives it the exact names to filter on — and have it write a one-line summary only when something in that set appears or expires. Because the weather services are free, the watch can run every hour on an empty wallet; only if you later add a priced service, such as a market quote for a company in the affected area, does the wallet come into it, and that call is listed with its price before it runs.

Set it up once

npm install --global @actionway/cli@latest
actionway init
mkdir -p "$HOME/.codex/skills"
cp -R "$HOME/.actionway/skill/actionway" "$HOME/.codex/skills/actionway"

You need Node.js 20 or newer and an Actionway account; after sign-in the agent reports your wallet balance. The weather services never touch it, so a watch like this one runs on an empty wallet. The get started page lists every route per client, including the Codex plugin.