agents2agentsai/paper-discovery

Paper Discovery

Core Behavior

1. Auto-continue. After presenting the reading view, IMMEDIATELY proceed to $paper-synthesis then $cross-paper-report. Exception: user says "just list", "don't synthesize", or "discovery only".
2. Present a landscape analysis, not a paper list. Cluster papers into approaches with tradeoffs. Never output flat lists or sections named "Summary."
3. Speed matters. Go straight to searching. Every tool call before paper_search is delay the user feels.

Hard Rules

• Do NOT open arXiv URLs — paper_search/paper_citations already return all metadata
• Do NOT call paper_get for papers already returned by search/citations — only for bare S2 IDs
• No academic citations — reference papers by title or method name

Modes

1. Citation-focused — user references a specific paper → paper_citations + 1-2 keyword searches
2. Explore — research question → faceted keyword search + citation expansion
3. Discovery — user has a research base → KB/Zotero seeds + citation/recommendation APIs

---

EXPLORE MODE

Phase 0: Check Existing Knowledge (1 call max)

Skip if KB disabled. rg "tag1\|tag2\|tag3" ${CODEX_KB_PATH}/cards/ with 2-3 tags.

Phase 1: Decompose into Facets

Parse the question into 3-5 search facets targeting different angles.

Facet design is critical — vague queries return junk. Encode the relationship, disambiguate terms, vary abstraction level, use field vocabulary.

BAD: "large language model interpretability explanations for AI models" GOOD: "LLM-generated explanations for neural network predictions", "self-rationalization large language models faithfulness"

Phase 2: Search

Citation-focused: 3 calls — paper_citations(id, limit=30) + two paper_search variants.

General explore: Keyword search (2-3 facets, paper_search(query, limit=10, sort_by=citation_count) + paper_search(query, limit=10, sort_by=year) per facet), then citation expansion on top 3-5 seeds. Budget: ≤ 20 API calls.

Phase 3: Analyze and Organize

Deduplicate (DOI → arXiv → title fuzzy). Mark known KB papers. Cluster into 3-6 approaches. Rank into reading order: Start Here, Core Methods, Cutting Edge.

Phase 4: Present

Use present_reading_view with topic-appropriate headings. Each section ≤ 40 lines. Typical sections: Landscape, Approaches, Key Insights, Reading Plan, Open Questions — adapt freely. Include key figures from papers via crop_and_store_figure when they aid understanding.

Follow-ups: use append_to_section with foldable=true.

---

DISCOVERY MODE

Uses KB/Zotero as seeds to find adjacent papers.

1. Gather seeds — list KB cards (skip if disabled), extract paper IDs. No KB: use $zotero or ask for 3-5 seed IDs.
2. Search — from top 5 seeds: paper_citations + paper_references (limit=15 each), topic searches, optional paper_recommendations. Budget: ≤ 25 calls.
3. Merge and present — dedup, filter out known papers, rank by citations + recency, present in same reading view format.

---

Post-Discovery

Skip if KB disabled. Prepend to <kb_path>/research-journal.md: date, mode, paper count, top 2-3 finds, open questions. If request reveals priorities, offer to note in research-context.md.

Pipeline

After discovery: $paper-synthesis → $cross-paper-report (default). ≤ 12 papers = standard, > 12 = large-set mode.