news-research

name

news-research

description

Use when the daemon's news-research runner instructs you to deep-research the most interesting items in a day's news master document. Selects the agent-judged top items, fetches the full article + a bounded set of linked sources + GitHub READMEs for repo mentions + prior vault context, and appends a "### 🔬 Deep dive" block under each researched item in the master doc — preserving the "## Notes" section verbatim. Idempotent — safe to re-run for the same date.

News Deep-Research

You enrich a day's news master document with deep research on the few items that most warrant it. The master doc already exists (the news-daily-master agent produced it). Your job: pick the most interesting items by judgment, research them within a bounded budget, and append findings in place.

Constraints that matter

• You own additions under existing items; you must NOT touch ## Notes or anything below it. The runner SHA-checks the ## Notes section and rejects (status failed_notes_clobbered) if a single byte changed. The user owns that section.
• Bounded depth (hard rule). Per item: the article itself, then at most 3 additional linked-source fetches total, plus a GitHub README fetch for any org/repo mention. Never git clone — fetch the README via the web only. Stop when you have enough for a tight briefing, even if under budget.
• Cap. Research at most the number of items the runner prompt states (default 3). Fewer is fine if only one or two are genuinely worth it.

Procedure

1. Read inputs

• vault_read("01-Projects/News/daily/<target_date>-master.md") — the doc you will enrich.
• vault_read("01-Projects/News/interests-profile.md") — what the user cares about.
• The runner prompt contains a ## Recent ratings block (recent ⭐/👍/👎). Treat ⭐ as the strongest "this is interesting" signal; 👎 topics are rarely worth a dive.

2. Select items (judgment)

From the master doc's items, pick the ≤N most worth a deep dive, weighing: alignment with the interests profile, ⭐/👍 rating signal, novelty, and cross-day significance. This is a judgment call — do not mechanically take the top N; if only one item deserves it, research one.

3. Research each selected item (bounded)

For each:

• WebFetch the item's external URL (full article).
• Follow at most 3 additional links per item that materially add context (primary sources, the actual paper/repo/announcement — not navigation or related-articles chrome).
• For any org/repo GitHub mention, WebFetch https://github.com/org/repo (the README renders there). No clone.
• vault_search / vault_read for prior vault context on the key entities (have we covered this company/repo/topic before?).

4. Append findings in place

Use vault_edit to insert, directly under the existing item's bullet line, a block of exactly this shape:

  ### 🔬 Deep dive

  <2–4 sentence synthesis: what's actually new, why it matters, what's the catch.>

  - **Key facts:** <specifics — numbers, versions, names>
  - **Sources:** [<title>](<url>), [<title>](<url>)
  - **Vault context:** [[01-Projects/News/entities/<Entity>|<Entity>]] — <one line on prior coverage, or "no prior coverage">

Rules:

• Insert under the item, above the next item/category heading. Never below ## Notes.
• If a ### 🔬 Deep dive block already exists under that item (re-run), replace it, don't duplicate.
• Keep it tight. This lands in the user's digest; it must be skimmable.

5. Verify before reporting

• Re-read the master doc. Confirm the ## Notes heading and everything below it is byte-for-byte unchanged from step 1.
• Confirm each researched item has exactly one ### 🔬 Deep dive block.

6. Emit the structured summary

End your response with a single fenced JSON block:

{
  "success": true,
  "items_researched": 2,
  "error": null
}

• success: true if you completed the procedure and the ## Notes verification passed; false on an unrecoverable error (master unreadable, MCP unavailable).
• items_researched: how many items you appended a deep-dive block to this run.
• error: present and non-null only when success is false.

The runner parses the LAST fenced json block, so progress JSONs earlier are fine.

Quality bar

• A deep dive must add something the one-line master summary didn't: the mechanism, the number, the caveat, the "why now."
• Prefer primary sources over commentary. If the article is itself commentary on a paper/repo, fetch the paper/repo.
• "No prior coverage" is a fine and useful vault-context answer — don't fabricate links.