news-agent

name	news-agent
description	News curator agent — bootstraps from agent-sheets/news.md directives. Loads Synapse + LLM-WIKI operating guide and executes the world-news-daily task.
trigger	/news-agent

News Agent

Loads: wiki/scratchpad/agent-sheets/news.md for full directives
Wiki root: /home/ty/Documents/LLM-WIKI
Operating guide: wiki/synthesis/synapse-llm-wiki-operating-guide.md

Bootstrap

1. Read the agent sheet at wiki/scratchpad/agent-sheets/news.md (this file gives you your task directives)
2. Read the jobs sheet: wiki/scratchpad/jobs/sheet.md
3. Read carryover: wiki/scratchpad/jobs/reports/news/carryover.md — includes the Article Index (last 10 URLs ingested); do NOT re-ingest anything already indexed
4. Discover via RSS only — not wiki search:

   curl -s "https://news.google.com/rss/search?q=geopolitics+may+2026&hl=en-US&gl=US&ceid=US%3Aen"
   curl -s "https://news.google.com/rss/search?q=AI+tech+policy+regulation+may+2026&hl=en-US&gl=US&ceid=US%3Aen"
   curl -s "https://news.google.com/rss/search?q=science+breakthrough+may+2026&hl=en-US&gl=US&ceid=US%3Aen"
   curl -s "https://news.google.com/rss/search?q=economy+trade+tariff+may+2026&hl=en-US&gl=US&ceid=US%3Aen"
   

5. De-duplicate against Article Index in carryover.md and recent headlines-*.md files before ingesting
6. Write 3-5 new stories per cycle (not more)
7. Update carryover.md: add new articles to Article Index, update Established/Open sections, prune index to last 10 entries
8. Report to #news Discord channel

Primary Discovery: RSS Only — NOT Wiki Search

RSS via curl is the only discovery mechanism. Do NOT use web search to scan existing wiki pages, do NOT use wiki_search as a discovery tool, do NOT use MCP tools to re-read the wiki. Re-reading wiki content as "finding news" is the primary failure mode that causes token blowups and duplicate ingestion.

curl -s "https://news.google.com/rss/search?q=geopolitics+may+2026&hl=en-US&gl=US&ceid=US%3Aen"
curl -s "https://news.google.com/rss/search?q=AI+tech+policy+regulation+may+2026&hl=en-US&gl=US&ceid=US%3Aen"
curl -s "https://news.google.com/rss/search?q=science+breakthrough+may+2026&hl=en-US&gl=US&ceid=US%3Aen"
curl -s "https://news.google.com/rss/search?q=economy+trade+tariff+may+2026&hl=en-US&gl=US&ceid=US%3Aen"

MCP tools (wiki_fetch_url, wiki_ingest_raw) are NOT discovery tools — they are ingestion tools. Use them only after selecting a story via RSS. Do not call query_knowledge, explore_connections, or wiki_search as a substitute for RSS discovery.

Article Index: Dedup Before Ingesting

Before writing any story, check the Article Index in carryover.md. If the story's URL or slug is already indexed, skip it entirely — no re-ingest, no re-write.

Target: 3-5 new stories per cycle (not more). Exceeding this means the agent is finding too much news — narrow your RSS queries if needed.

Step-by-step: RSS → Ingested Wiki Page

1. Discover via RSS:

   curl -s "https://news.google.com/rss/search?q={query}&hl=en-US&gl=US&ceid=US%3Aen"
   

   Use topic-specific queries: geopolitics+may+2026, AI+tech+policy+regulation+may+2026, science+breakthrough+may+2026, economy+trade+tariff+may+2026

2. Parse RSS — extract <title>, <link>, <pubDate> for each <item>

3. Select stories — aim for 3-5 significant stories per cycle; use RSS item count as signal (empty = no recent coverage, many items = active story)

4. Write the source page — create a markdown file locally with frontmatter and body, then copy to:

   /home/ty/Documents/LLM-WIKI/wiki/sources/articles/{slug}.md
   

   Do NOT use wiki_ingest_raw or wiki_fetch_url as primary ingestion — they fail at meaningful rates for mainstream news.

5. Update carryover at wiki/scratchpad/jobs/reports/news/carryover.md

Why MCP tools are not primary

• synapse_mcp.zettelkasten.ingestion.ingest_raw does not exist at that path — actual module structure is synapse_mcp.wiki.wiki_adapter.WikiAdapter
• wiki_fetch_url (defuddle-based) fails at meaningful rates for: BBC (404 on article IDs), Reuters (bot blocking), Guardian (403), Al Jazeera (slug changes)
• RSS gives you structured data with titles, dates, and links — no defuddle parsing needed
• Manual write to wiki/sources/articles/ is deterministic and reliable

When to use MCP tools (limited scope)

• wiki_fetch_url: Only when you have a direct working URL for a specific story and RSS returned no results for that topic
• wiki_lint: Informational only after writes — output is noisy, do not block on it
• query_knowledge / explore_connections: For cross-referencing existing wiki content, not for ingestion

URL Slug Patterns by Major Publisher

When defuddle returns 404 on a known article, the URL slug format may have changed. Known working patterns:

Publisher	URL Pattern	Notes
Al Jazeera	https://www.aljazeera.com/news/YYYY/M/D/{slug}	Date-based subdirectory; slug includes full headline
BBC	https://www.bbc.com/news/world-{article-id}	Numeric article ID, not slug-based
Reuters	https://www.reuters.com/{world|world/{region}}/...	Title in path, may change

When a direct URL fails: Use the browser to visit the section page, extract article links from there, then feed those links to wiki_fetch_url. The browser renders dynamic content that defuddle cannot fetch.

When browser also fails: Use Google News search for the story title, extract the canonical URL from the Google News result (which usually has a working link to the publisher), then use that URL with wiki_fetch_url.

After Writing: Update Carryover

Before finishing, update wiki/scratchpad/jobs/reports/news/carryover.md with:

• New emerging themes from this run
• Stories to keep monitoring (with brief notes on why)
• Any regions needing continued attention
• Connections to existing wiki research threads

The carryover is the primary mechanism to avoid duplicate work across runs. Read it at session start; update it before session end.

Wiki Lint: Informational Only

Running wiki_lint() after writes is good practice but the output is noisy (orphans, broken links, non-reciprocal links are endemic to this wiki). Do not let lint failures block your run. Lint is for health awareness, not gatekeeping.

Operational Notes

• Major geopolitical events, AI/tech policy, scientific breakthroughs, economic shifts = priority
• Soft news (celebrity, sports) = exclude unless it intersects your research threads
• Aim for 3-5 stories per cycle, ingested with significance not volume
• Cross-link to existing wiki threads to make news actionable knowledge

URL Fetch Failure Reference

When wiki_fetch_url fails on known stories, see references/news-source-url-patterns.md for:

• Publisher-specific URL patterns that work
• Why defuddle fails on certain sources (BBC article IDs, Reuters bot detection)
• The Google News canonical URL resolver pattern (universal fallback)
• Quick reference table for which method works per publisher

Rule: Never spend more than 2 direct URL attempts on a story before switching to Google News URL resolution.

Strict Token Budget

Target: ~15,000 input tokens per cycle (30k max).

Token-saving rules:

1. Do NOT read the Established section of carryover unless you need a specific detail. It's historical context — the Article Index and Heading are what you need.
2. Do NOT cross-reference existing wiki content during news selection. RSS gives you the story; write the summary directly. Only use wiki_search/query_knowledge if a story explicitly requires it.
3. Limit carryover reads to the Article Index and Heading sections.
4. Skip wiki lint. wiki_lint() is informational only — don't call it.
5. 3-5 stories max. Stop after 5 new stories. Do not hunt for more.
6. Do NOT re-ingest established stories. They are already tracked. Skip them entirely.

If context trends toward 30k+ tokens before writing reports: stop reading wiki, skip cross-links, write only the headlines in carryover, and flag that the cycle was context-heavy.