Read Paper

Read Paper

Turn paper-like input into an auditable artifact before analysis. Never treat an abstract, landing page, or repository index as a paper's full text.

Routing contract

Read <plugin-root>/references/agent-routing.md, resolving <plugin-root> by going up two directories from this file. The main thread owns user Q&A and any archive writes. Delegated roles return analysis only.

1. Acquire and validate

Read references/paper-acquisition.md. Create a temporary paper packet plus a page/section-preserving text artifact, then run:

python3 "{skill_root}/scripts/validate-paper-packet.py" packet.json \
  --mode deep-dive --require-anchor equation:7

Use --mode hub-triage for a hub; omit --require-anchor when none was requested.

Choose the route from the packet, not from the URL suffix:

• single_paper + full_text → deep-dive.
• single_paper + partial/abstract_only → stop the deep-dive. State the coverage limit; offer an explicitly scoped abstract-level reading or request the PDF/full text. Do not update the landscape as if the paper was read.
• literature_hub + hub_index → hub-triage. A repository, Awesome list, proceedings page, or reading list is not a single paper.

If validation fails, report the failed invariant and repair acquisition before analysis. Never silently replace an arXiv PDF with its /abs/ page. Translate each user-requested equation, figure, table, page, or section into a repeatable --require-anchor TYPE:ID; missing anchors limit or block the answer.

2. Gather context by path

Pass only relevant existing paths, such as project-skill/SKILL.md, docs/papers/landscape.md, exp/summary.md, and a directly related experiment README/result. Do not inline whole files or preload every experiment.

3. Delegate analysis

Use <plugin-root>/agents/domain-expert.md through the portable routing contract. For a paper, request Mode 4; for a hub, request Mode 5 seed-hub triage. The task must contain:

Paper packet path: {packet_path}
Research context paths:
- {relevant_path_or_none}
User focus: {question_or_none}
Read the artifacts directly. Return analysis to the main thread.

Do not paste the fetched paper into the delegation prompt.

4. Continue and archive

Present the result with its coverage/provenance and invite follow-up questions. Answer in the main thread, preserving unresolved questions. For hub triage, begin with input_kind=literature_hub so the boundary is visible.

When “save/archive/store/存档/保存” was requested, including in the original task, continue without asking again. Honor an explicit wiki destination using that wiki's source/analysis rules; do not also create an unrequested project archive. Preserve the validated packet and exact locators in either destination. For a project archive:

1. Freeze packet and exact text under docs/papers/artifacts/{short-name}/, rewrite only its relative text_path, then revalidate the archived packet. Copy an accessible source PDF when privacy/licensing permit and record its hash; otherwise retain the locator and explain why raw bytes are absent.
2. Write {short-name}-deep-dive.md with links/hashes, Evidence Ledger, Methodology, assumptions, Bridge Analysis, and Open Questions. A hub instead gets {short-name}-hub-triage.md; linked papers remain unread.
3. Update the landscape only with artifact-supported claims; report all paths.