paper-intake

Paper Intake

Overview

Move papers from discovery to raw evidence without passing unresolved names to capture. The visible path is chat-only discovery -> raw/evidence/ -> wiki/papers/. For read/index work, consult docs/qmd-cli.md and runtime qmd --help.

Available Tools

Tool	Purpose
discover_papers	Search, score, deduplicate, and return transient candidates for chat presentation
ingest_and_read	Capture one or many resolved arXiv URLs, IDs, or arXiv DOI values
upsert_wiki_page	Create canonical paper pages after reading and distilling
check_concept_alias	Resolve uncertain concept names before creating core concept pages

Routing Decision

Execute this decision tree before any tool call:

1. User asked for discovery without asking to ingest specific papers: call discover_papers(query=...), present returned candidates in chat, apply the agent-side presentation guidance in references/post-discovery-rescoring.md, then stop.
2. User selects a candidate from the current conversation: resolve that candidate to an arXiv URL, arXiv ID, or arXiv DOI value, then call ingest_and_read(input_value=...).
3. User provided arXiv URL(s), arXiv ID(s), or arXiv DOI value(s): call ingest_and_read(input_value=...) directly. Batch multiple resolved identifiers into one call when possible.
4. User provided title(s), acronym(s), alias(es), project name(s), or mixed natural paper references: resolve each reference to an arXiv URL, ID, or arXiv DOI value first, then call ingest_and_read with only the resolved identities. Do not pass unresolved names to MCP.
5. User asked whether something is already in the vault or wants to read local evidence: use QMD CLI directly, usually qmd query or qmd get.
6. User wants a canonical paper page after capture and reading: hand off to the paper-distillation skill.

Core Rules

• Discovery is chat-only. Treat discover_papers results as transient candidates. Present them in chat and stop unless the user also gave a specific resolved identity to capture.
• Ingest captures only resolved identities. ingest_and_read accepts arXiv URLs, arXiv IDs, and arXiv DOI values. Natural references are agent-resolved before the MCP call. Do not edit raw/evidence/ by hand. See references/ingest-rules.md for the resolved/unresolved reporting contract and capture repair behavior.
• Canonical pages are a separate skill. Distillation from raw/evidence/ to wiki/papers/ is handled by the paper-distillation skill. The canonical flow requires 1–5 concepts in key_concepts_topk, at least one wikilinked in the body, and check_concept_alias before linking uncertain concept surfaces.
• Conversation insights use the tool path. If a paper-intake conversation surfaces a durable insight about why a paper matters, save the distilled takeaway to insights/conversations/ through the Python write path. Do not save verbatim transcripts or edit insight files by hand.

Post-Discovery Re-scoring

After discover_papers returns, the agent should review returned low-score candidates against the user's research_profile.direction before presenting results. Promote semantically relevant candidates in the chat summary with a one-line reason. See references/post-discovery-rescoring.md for the full playbook.

QMD CLI Guidance

• Use QMD CLI directly for qmd query, qmd get, qmd status, qmd update, qmd embed -f, and collection/context health.
• Check docs/qmd-cli.md first for the Paper Distill mapping.
• If command details are uncertain, follow runtime qmd --help.

Output Expectations

• After discovery, summarize the strongest candidates, include resolved arXiv identities or arXiv DOI values when available, and stop for user selection.
• After ingest, report exact identifiers processed and unresolved references separately.
• If ingest cannot proceed, explain the blocking state in one sentence before suggesting the next step.

Skill / Tool Contract

This skill is responsible for intent routing, resolved-identity discipline, summarizing outcomes, and suggesting the next useful action.

Python tools are responsible for deterministic I/O, path safety, capture, and structured errors.