paper-resolver

name	paper-resolver
description	Resolve one natural-language paper request into a canonical paper list. Use when Codex needs to interpret the request, search the web, disambiguate candidate papers, and write `data/runs/<run_id>/resolved_papers.json` before acquire begins.

Use this skill to turn one natural-language request into one canonical paper list.

Interpret the request first. It may refer to:

• one paper by title or alias
• a small explicit set of papers
• a conservative collection request such as papers from one lab or company

Path rule:

• Paths under this skill such as references/... are relative to this skill directory.
• When writing files into the repo, use repo-root paths such as data/runs/<run_id>/resolved_papers.json.

Inputs:

• one natural-language request
• optional local paper lists when they help narrow candidates

Outputs:

• data/runs/<run_id>/resolved_papers.json

Output format:

• resolved_papers.json must be a top-level object with a papers list
• each item in papers must be one canonical paper record

Work:

1. interpret the natural-language request
2. search the web for candidate papers
3. for every candidate title, search for an exact or near-exact arXiv record before accepting non-arXiv links
4. disambiguate the request conservatively
5. write one canonical paper list

Each paper entry must include:

• paper_id
• title
• url
• pdf_url

When an arXiv page is available for a paper:

• use paper_id as arxiv-<arxiv_id>
• use url as https://arxiv.org/abs/<arxiv_id>
• use pdf_url as https://arxiv.org/pdf/<arxiv_id>.pdf

Complete only when:

• resolved_papers.json exists
• resolved_papers.json has a top-level papers list
• every listed paper has the required fields

Fail when:

• the request cannot be resolved reliably
• multiple plausible candidates remain and the request is still ambiguous
• the request describes an open-ended collection that cannot be bounded conservatively from search results

Guardrails:

• Resolve papers only. Do not download PDFs or enrich metadata here.
• Prefer strong identifiers over loose title matches.
• Use web search to verify candidate papers instead of guessing from memory.
• Do not stop at an official project/blog PDF link until the title has also been checked against arXiv.
• For lab or company requests, include only papers that can be tied to that entity reliably from search results.
• If the request is ambiguous, fail instead of guessing.