菜单
News curator agent — bootstraps from agent-sheets/news.md directives. Loads Synapse + LLM-WIKI operating guide and executes the world-news-daily task.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
| 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 |
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
wiki/scratchpad/agent-sheets/news.md (this file gives you your task directives)wiki/scratchpad/jobs/sheet.mdwiki/scratchpad/jobs/reports/news/carryover.md — includes the Article Index (last 10 URLs ingested); do NOT re-ingest anything already indexedcurl -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"
headlines-*.md files before ingesting#news Discord channelAfter writing your carryover, load the kanban-morning-review skill:
wiki/scratchpad/jobs/reports/news/carryover.mdsearch_files or mcp_project_synapse_wiki_search for existing content on the topicmcp_project_synapse_synapse_recall for relevant episodic factshermes kanban add to create the task (the skill handles this)RSS via terminal (curl) + manual file write is the reliable primary workflow. MCP tools (wiki_fetch_url, wiki_ingest_raw) have inconsistent module paths and unreliable fetch rates for mainstream news — do not treat them as primary.
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
Parse RSS — extract <title>, <link>, <pubDate> for each <item>
Select stories — aim for 3-5 significant stories per cycle; use RSS item count as signal (empty = no recent coverage, many items = active story)
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.
Update carryover at wiki/scratchpad/jobs/reports/news/carryover.md
synapse_mcp.zettelkasten.ingestion.ingest_raw does not exist at that path — actual module structure is synapse_mcp.wiki.wiki_adapter.WikiAdapterwiki_fetch_url (defuddle-based) fails at meaningful rates for: BBC (404 on article IDs), Reuters (bot blocking), Guardian (403), Al Jazeera (slug changes)wiki/sources/articles/ is deterministic and reliablewiki_fetch_url: Only when you have a direct working URL for a specific story and RSS returned no results for that topicwiki_lint: Informational only after writes — output is noisy, do not block on itquery_knowledge / explore_connections: For cross-referencing existing wiki content, not for ingestionWhen 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}}/...` |
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.
Before finishing, update wiki/scratchpad/jobs/reports/news/carryover.md with:
The carryover is the primary mechanism to avoid duplicate work across runs. Read it at session start; update it before session end.
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.
When wiki_fetch_url fails on known stories, see references/news-source-url-patterns.md for:
Rule: Never spend more than 2 direct URL attempts on a story before switching to Google News URL resolution.
Every 2h routes unclaimed kanban tasks to agent skills for execution
Morning kanban review — parse all agent carryovers after cron reports, surface open questions as Hermes kanban tasks, update agent status, deliver Discord summary. Creates tasks in hermes kanban (not sheet.md). Proactive: each morning agent cron triggers its own carryover parse.
ArXiv research curator agent — bootstraps from agent-sheets/arxiv.md directives. Discovers, downloads, and reports on ML/AI papers from arXiv.
Ingest pipeline operator agent — bootstraps from agent-sheets/ingest.md directives. Processes raw files from raw/ inbox into structured wiki knowledge.
Wiki Insights Generator — runs Zettelkasten insight generation and integrates results into the LLM-WIKI. Runs daily at 06:00 AM.
Kanban board creation and workflow decomposition persona. Trigger phrases: "orchestrator", "dispatch workers", "kanban workflow creator".