Update to Pendar

{
  "description": "Check a whole reference list in one pass — which entries are real, which are retracted, which are ambiguous, and which cannot be confirmed. Use when the user pastes a bibliography, a reference list from a draft or a published paper, or a set of sources from an AI-generated answer and wants them checked before relying on them.",
  "included_files": [
    {
      "relative_path": "agents/openai.yaml",
      "size_in_bytes": 535
    },
    {
      "relative_path": "assets/pendar-logo.png",
      "size_in_bytes": 31724
    }
  ],
  "name": "audit-a-bibliography",
  "skill_md_contents": "---\nname: audit-a-bibliography\ndescription: Check a whole reference list in one pass — which entries are real, which are retracted, which are ambiguous, and which cannot be confirmed. Use when the user pastes a bibliography, a reference list from a draft or a published paper, or a set of sources from an AI-generated answer and wants them checked before relying on them.\nmetadata:\n  short-description: Check every reference in a list at once\n---\n\n# Audit a bibliography\n\nCheck every entry of a reference list against the published record in one call.\n`audit_bibliography` takes up to 25 citations and returns a verdict for each, in\nthe order given, plus a tally.\n\nAn entry that cannot be confirmed is reported as unconfirmed. Never substitute\nthe closest real paper, and never supply a missing title, page, year, venue, or\nDOI from memory — that is how fabricated references end up looking real.\n\n## Quick start\n\n1. Split the list into individual references, one string each. Keep each\n   reference exactly as written, in whatever style it uses (author-year, numbered\n   physics style, quoted titles). Abbreviated journal references such as\n   `Phys. Rev. Lett. 103, 086001 (2009)`, DOIs, and arXiv or PubMed IDs are all\n   supported and should not be expanded first.\n2. Set `context`:\n   - `draft` — the user's own manuscript, or sources from an AI answer.\n   - `published` — the reference list of a paper that is already published.\n3. Call `audit_bibliography` with up to 25 references. For a longer list, send\n   the next 25 in a second call.\n4. If any entries come back `not_checked`, call again with only those entries.\n\n## Reading the result\n\nEach entry in `entries` has a `verdict`:\n\n- `verified` — the paper exists and matches. `resolved_by` says how (`doi`,\n  `paper_id`, `title_search`, or `biblio` for an abbreviated reference matched on\n  its journal locator). All of these are verifications. Quote the canonical title\n  from `paper.title`. When `match_confidence` is below 0.8, show the canonical\n  title next to the reference as written so the user can confirm it is the same\n  paper. `matched_on_fields` lists what agreed with the record; `first_author`\n  there means the cited author settled a choice between papers sharing a title.\n- `retracted` — the paper exists and is retracted. Lead with that.\n- `ambiguous` — several papers match about equally. List the `candidates` and\n  ask which was meant.\n- `metadata_mismatch` — a paper matches part of the citation, but a detail\n  disagrees with the record. Two forms:\n  - with `paper` — the paper exists and was found at the cited journal, volume\n    and page, by the cited first author, but the cited title differs. `reason`\n    quotes the record's title; report the published title as the fix.\n  - without `paper` — the year, first author or journal disagrees. Show both,\n    and the `candidates` when present.\n- `not_found` — nothing in the index matches. Quote `reason`.\n- `not_checked` — not a verdict. Either the call ran out of time before reaching\n  this entry, or the paper index did not respond for it (\"The lookup failed\").\n  It was not charged; send it again in a second call.\n\n`tally` counts each verdict. `notes` carries caveats; repeat the ones that bear on\nwhat you claim.\n\n## Never\n\n- Fill in or correct a reference from memory. If an entry is `not_found`, the\n  answer is that it could not be confirmed.\n- Treat `not_found` as proof of fabrication. It is strong evidence for a\n  mainstream journal article and weak evidence for a book, chapter, thesis,\n  report, or very recent preprint. Say which case the entry looks like.\n- Report `not_checked` entries as missing.\n\n## Reporting\n\n1. A table or list in the user's order: the reference as given, the verdict, and\n   for verified entries the canonical title, year, venue, and DOI.\n2. The tally.\n3. For `draft`: say that retracted, mismatched, and unconfirmed references need\n   attention before submission. For `published`: say that unconfirmed lines could\n   not be confirmed here, which is not a finding that the published bibliography\n   is wrong.\n\n## Related workflows\n\n- To look closer at one reference, use `verify-citations`.\n- To see who has built on one of the verified papers, use\n  `trace-citation-impact`.\n"
}