bibtidy

Use this skill for any request to validate, check, clean up, or fix a BibTeX .bib file.

If the request includes a .bib path, operate on it. If the path is missing or does not exist, ask for it.

Assumptions:

• Process entries one by one.
• Only brace-style BibTeX like @article{...} is supported; reject parenthesized @article(...).
• Preserve @string, @preamble, and @comment blocks verbatim.

Resolve Tools

All bundled scripts live in tools/ next to this file. Resolve TOOLS_DIR once:

for d in \
  "${CODEX_HOME:-$HOME/.codex}/skills/bibtidy/tools" \
  "$HOME/.claude/skills/bibtidy/tools" \
  "${CLAUDE_PLUGIN_ROOT:-/dev/null}/skills/bibtidy/tools"; do
  [ -f "$d/crossref.py" ] && TOOLS_DIR="$d" && break
done

Useful commands:

• python3 $TOOLS_DIR/compare.py <file.bib> [--key KEY]
• python3 $TOOLS_DIR/crossref.py doi <DOI>
• python3 $TOOLS_DIR/crossref.py search "<title>"
• python3 $TOOLS_DIR/crossref.py bibliographic "<query>"
• python3 $TOOLS_DIR/duplicates.py <file.bib>
• python3 $TOOLS_DIR/edit.py <file.bib> <patches.json>

Patch Rules

Use edit.py for all actual .bib changes. Do not edit the file directly with agent editing tools and do not rewrite the whole file.

• fix patch: include key, action, urls, explanation, and fields; entry_type is optional.
• In fields, set a field to null to remove it; omit a field to leave it unchanged.
• not_found: comment out the original entry; do not add URL lines.
• duplicate: add % bibtidy: DUPLICATE of <other_key> — consider removing above the original entry; no URL lines.
• review: add one or more % bibtidy: <URL> lines plus % bibtidy: REVIEW, <reason> above the unchanged entry.
• Clean entries get no comments.

For fix patches, edit.py should produce:

• fully commented original entry
• one % bibtidy: <URL> line per source
• one % bibtidy: <explanation> line
• corrected entry

Use source values verbatim. If the bib entry uses and others and a verified source provides the full author list, replace it with the full list.

Workflow

Each entry has a web-search budget of 1 total, used in at most one of Wave A or Wave B.

1. Read the file and note the path.
2. Clear the platform log: > <file>.bib.cc.log in Claude Code or > <file>.bib.codex.log in Codex.
3. Back up the file: cp <file>.bib <file>.bib.orig
4. Run python3 $TOOLS_DIR/duplicates.py <file.bib> before metadata fixes.
5. Run python3 $TOOLS_DIR/compare.py <file.bib> for CrossRef candidates.
6. Wave A web search: every entry with error set or no candidates must be web-verified. Those entries have spent their budget.
7. After each wave, classify every entry as one of:
   • Clean: confirmed, no changes
   • Fix: confident correction
   • Escalate: still ambiguous and budget unused
   • Not found: no paper found after required search
   • Review: budget spent and still uncertain
8. Wave B web search: only entries marked Escalate and not yet searched. After Wave B, final outcomes are only Clean, Fix, Not found, or Review.
9. Write a real patches.json file and apply fixes with python3 $TOOLS_DIR/edit.py <file.bib> patches.json.
10. Run python3 $TOOLS_DIR/duplicates.py <file.bib> again. Resolve every unresolved same-key collisions warning and rerun until the warning is gone.
11. Manually review likely related entries after fixes. Strong clues: repeated citation keys, repeated DOIs, same normalized title with overlapping authors, obvious preprint→published pairs. If two entries should be linked but not auto-removed, apply a duplicate patch.
12. Validate format, delete the backup, and print a summary table with rows: total entries, verified, fixed, not found, needs review, exact duplicates removed, near-duplicates flagged.

For files with more than 30 entries, work in batches of about 15 and report progress. Entry count must match before and after.

Compare Carefully

compare.py returns raw CrossRef candidates plus discrepancies. Treat them as hints, not truth.

• A missing candidate field is not evidence that the bib field should be removed.
• If a verified candidate supplies a missing standard venue field such as volume, issue, or pages, add it.
• Never add a doi field when the bib entry currently lacks one.
• Treat author vs authors and journal vs booktitle as schema mismatches until verified.
• If a preprint has a verified published version, update title, venue, year, volume, number, and pages together.
• When the published title differs, replace the title verbatim; do not partially edit it.
• Use -- for BibTeX page ranges.
• Missing pages is not automatically an error for venues that do not publish page numbers.

Web Verification

Use subagents when available; otherwise do the same work sequentially. Cap at 6 subagents per wave and distribute entries evenly.

Each web-search subagent should return only JSON with:

• key
• source_urls
• fields
• notes

Rules for subagents:

• fields is either a fix-patch dict or null.
• Use null inside fields to remove a stale field.
• Verify against authoritative pages when possible: DOI page, publisher page, venue page, arXiv, OpenReview, etc.
• Check title, full author list, year, journal/booktitle, volume, number/issue, pages, and DOI.
• Put a value in fields when the verified source disagrees with the bib entry or when the bib entry is missing a standard field that the venue publishes.
• Do not add a missing doi field just because you found one; mention it in notes if useful.
• If a standard field is genuinely unavailable, say so in notes.

Preserve

• Entry order
• All unchanged fields
• Empty lines between entries
• User % comments that are not % bibtidy:
• @string, @preamble, and @comment blocks
• LaTeX macros and brace-protected capitalization in titles

If rate-limited, note it and continue with the next entry.