local-paper-navigator

Local Paper Navigator

Find and read papers from the local papers library via the three native tools. No scripts, no environment variables — the library is already mounted and the tools resolve it themselves.

paper_search    L3  one line per paper — find candidates, pick paperIds
      │
paper_read      L2  full card + section outline — understand a paper (one call is enough)
      │
paper_section   L1  ONE verbatim section — the only escape hatch into full text

Whole-paper reads are blocked by design (/papers/markdown/** and /papers/raw/** reject direct reads). A read rejection is the mechanism working, not a failure — descend a level instead.

Core loop

1. paper_search(query, limit) — 1–3 queries, one concept each, dedupe by paperId.
2. paper_read(paper_id) — for each candidate that matters. 8+ char paperId prefix works.
3. paper_section(paper_id, heading|query) — only when you must quote the full text verbatim.
4. Stop. Cite with the paperId prefix (first 8+ chars).

Empty search → the tool says so honestly. Report "not in the library" rather than widening into fabrication; try one rephrased query (English ↔ Chinese — cards match Chinese titles) before concluding.

Reading discipline

Level	Tool	Payload	When
L3 Contextual	paper_search	one line/paper	scanning, building a shortlist
L2 Analytical	paper_read	card + outline, ≤8K	understanding a paper's method + findings
L1 Technical	paper_section	one section, ≤16K	reimplementing, verbatim quotation

Defaults: L3 before L2 before L1. Escalate only when the current level cannot answer the question. Most questions end at L2.

Five Red Lines (always)

1. Track history. Don't re-run a query you already ran. Empty result → change angle, not synonyms.
2. Search a gap, not a vibe. Every query maps to one missing piece of information. No stacked-keyword bags.
3. One query = one concept. Split comparisons (A vs B), multi-property asks, and multi-year spans into separate calls.
4. Never hallucinate. Every fact (title, source, year, content) comes from a tool result. Direct reads being rejected is the design — never claim content you couldn't reach.
5. Quote-or-zero. When you claim a paper meets a criterion, quote a ≤80-char span from its card or a paper_section result. No quote → that criterion scores 0.