bibcite-cli 0.5.0

Resolve papers (arXiv id / DOI / title) to canonical, normalized BibTeX for agents and humans

bibcite

Resolve papers (arXiv id / DOI / title) to canonical, normalized BibTeX, and manage .bib files so agents never hand-edit them.

The publication-matching cascade is ported from PaperMemory's bibMatcher: DBLP → Semantic Scholar → Google Scholar → CrossRef → Unpaywall. A match must have an identical normalized title, a plausible year, and a non-preprint venue.

Venue names are canonicalized against the @string table vendored in src/bibcite/data/strings.bib (journals / conferences / workshops), including year-aware rules (NIPS before 2018 vs NeurIPS, WACV before 2017).

Entry types are strict: conference/workshop papers become @inproceedings + booktitle, journal papers @article + journal, and unpublished arXiv preprints @misc + howpublished = {arXiv preprint arXiv:ID}. Types coming from authoritative source BibTeX (DBLP) are preserved.

After every write, the file is formatted with bibtex-tidy using the canonical flags in bibfile.TIDY_ARGS (requires bibtex-tidy on PATH or npx).

Install

# from a local checkout (development)
uv tool install --editable .

# from git, no checkout needed
uv tool install git+https://github.com/leo1oel/bibcite

# once published to PyPI (package name bibcite-cli, command name bibcite)
uv tool install bibcite-cli   # or: uvx --from bibcite-cli bibcite ...

# plus, once (required for the tidy step):
npm install -g bibtex-tidy

To use your own venue table instead of the vendored one, set BIBCITE_STRINGS=/path/to/strings.bib or place it at ~/.config/bibcite/strings.bib.

Usage

# Preview the BibTeX for a paper (nothing written)
bibcite get 1706.03762
bibcite get "Attention is all you need"
bibcite get 10.1109/CVPR52688.2022.01167

# Resolve and write into a .bib file, dedupe, then bibtex-tidy; prints the final key
bibcite add refs.bib 2103.14030 --json

# Add a raw BibTeX entry you already have (venue still canonicalized, file still tidied)
bibcite add refs.bib --bibtex "$(pbpaste)"

# Batch add (one query per line; shares rate-limit state, tidies once)
bibcite add refs.bib --from ids.txt

# Overwrite a bad existing entry (keeps its key), or delete one
bibcite add refs.bib <query> --replace
bibcite remove refs.bib <key>

# One-shot cleanup: upgrade preprints → tidy → lint
bibcite fix refs.bib

# Upgrade every arXiv entry in a file to its published version (bibMatcher, CLI-style)
bibcite upgrade refs.bib --dry-run

# Just format, or just lint (check is read-only)
bibcite tidy refs.bib
bibcite check refs.bib

add/upgrade/check/fix/remove print a machine-readable JSON result on stdout (action, key, venue, source, ...); all diagnostics go to stderr. add is idempotent: an existing entry returns action: exists with its key, and an existing arXiv entry matched to a published version is upgraded in place, keeping its citation key. Exit codes: 0 success, 2 paper not found (ask for a better identifier), 3 sources/tool failure (retry later). Successful matches are cached at ~/.cache/bibcite/published.json (published papers only — preprint status is never cached); bypass with --no-cache or BIBCITE_NO_CACHE=1. Entries marked pubstate = {preprint} are treated as confirmed preprint-only and muted from check/upgrade.

For agents

Never edit .bib files by hand. Call bibcite add <file> <query> --json and use the returned key in \cite{...}.