@quill-io/sensor-frontier-ai-news

A World2Agent sensor that watches a configurable list of frontier AI lab blogs and emits a news.post.published signal per new post. Built for people who need to catch model releases the day they drop without manually refreshing seven lab blogs.

Two source mechanisms in one sensor:

• RSS for labs that ship a feed (OpenAI, Google DeepMind, HuggingFace, Meta).
• Sitemap diff for labs that don't (Anthropic, Mistral, Qwen) — fetch sitemap.xml, filter by URL prefix (/news/, /blog/), enrich new URLs with HTML <title> + og:description.

Cold-start is non-noisy: on first poll, the sensor seeds every existing post as already-seen and emits only the single most-recent post per feed. You get a "what's each lab up to right now" snapshot, not the back catalog.

Default feeds

| Lab | kind | URL | path_prefix | |---|---|---|---| | OpenAI | rss | https://openai.com/blog/rss.xml | — | | Google DeepMind | rss | https://blog.google/technology/google-deepmind/rss/ | — | | HuggingFace | rss | https://huggingface.co/blog/feed.xml | — | | Meta | rss | https://about.fb.com/feed/ | — | | Anthropic | sitemap | https://www.anthropic.com/sitemap.xml | /news/ | | Mistral | sitemap | https://mistral.ai/sitemap.xml | /news/ | | Qwen | sitemap | https://qwenlm.github.io/en/sitemap.xml | /blog/ |

Add or remove labs via config. xAI is excluded from defaults because it 403s anonymous bots.

Signal: news.post.published

| Field | What's in it | |---|---| | event.summary | [<lab>] <title> · <date> — <snippet> | | source_event.data | lab, source_kind (rss|sitemap), url, title, published_at | | attachments[0] | post snippet (inline text; absent for some sitemap-only sources) | | attachments[1] | reference URI to the original post |

Example summary:

[Anthropic] Introducing Claude Opus 4.8 · 2026-05-12 — A new frontier
in agent capability and long-context reasoning, with state-of-the-art
performance on SWE-bench and computer-use benchmarks.

Configuration

| Parameter | Type | Default | What it does | |---|---|---|---| | feeds | array | 7 frontier labs | List of { lab, url, kind, path_prefix? } entries | | interval_seconds | number | 1800 | Poll interval (30 min default — blog feeds change infrequently, faster polling mostly wastes bandwidth) | | max_post_age_days | number | 7 | Defensive back-catalog filter | | fetch_concurrency | number | 2 | Per-feed HTML enrichment concurrency cap (politeness knob) |

Install

In a Claude Code session with the @world2agent/claude-code-channel plugin installed:

/world2agent:sensor-add @quill-io/sensor-frontier-ai-news

The install flow reads SETUP.md, runs a 4-question Q&A (which feeds, interval, user context, surfacing aggressiveness), writes the sensor entry to ~/.world2agent/config.json, and writes a per-user handler skill.

Design notes

• Quantitative filtering at the sensor, semantic filtering at the handler. The sensor never inspects post meaning; it just emits one signal per genuinely new URL. The handler skill classifies (release / benchmark / research / policy / etc.) and decides what to surface.
• No auth required. All sources are public.
• xAI deliberately not in defaults. They block bots; users who want xAI can add it manually but should expect it to fail until they wire a workaround.

License

MIT