Bibliography Filter

Produce and validate a cited-only bibliography, then replace the manuscript's canonical .bib only after explicit approval. In an Overleaf-linked project, the resolved Overleaf bibliography is canonical; a wrapper-local copy is not a second authority.

When to Use

• Before submission — strip unused references from the bibliography
• When a shared .bib (e.g., Paperpile/Paperpile export) contains hundreds of entries but the paper cites a subset
• To reduce .bib file size for Overleaf or arXiv upload
• As a cleanup step after major revision (removed sections may leave orphan citations)

Process

Step 1: Locate files

If given a directory, resolve the actual manuscript source first:

• For paper-<arm>/paper, run realpath/readlink and inspect the resolved Overleaf target before declaring a bibliography missing.
• Read \bibliography{...} or \addbibresource{...} from the canonical driver and resolve the referenced .bib relative to that source tree.
• Do not select a project-root export, backup, or stale wrapper copy merely because it has the same filename.

Then find the main .tex and .bib files:

• .tex: Glob for *.tex in the directory (and paper/ subdirectory if it exists). The main file is the one containing \begin{document}.
• .bib: Look for \bibliography{...} or \addbibresource{...} in the main .tex to find the bib filename. Fall back to globbing *.bib.

If given a specific .tex file, derive the .bib from it.

Step 2: Extract all citation keys from .tex

Scan all .tex files in the project (main + any \input{}/\include{} files) for citation commands:

\cite{key1,key2}
\citep{key1,key2}
\citet{key1}
\citealt{key1}
\citealp{key1}
\citeauthor{key1}
\citeyear{key1}
\citetext{key1}
\parencite{key1,key2}
\textcite{key1}
\autocite{key1}
\footcite{key1}
\fullcite{key1}
\nocite{key1}

Handle:

• Multiple keys in one command: \cite{key1,key2,key3} → extract all three
• Whitespace around commas: \cite{key1, key2} → trim
• Optional arguments: \cite[p.~5]{key1} or \citep[see][]{key1} → extract key1
• \nocite{*} → special case: means "include everything", so filtered.bib = full .bib (warn and stop)

Collect into a deduplicated set of cited keys.

Step 3: Parse .bib and filter

Read the .bib file. For each entry (@article{key,, @book{key,, @inproceedings{key,, etc.):

• If the entry's key is in the cited set → include in output
• If not → exclude
• Preserve @string{}, @preamble{}, and @comment{} blocks (they may be needed by included entries)

Step 4: Stage a cited-only candidate

Write a temporary candidate beside the canonical bibliography, using a collision-safe name such as .references.filtered.<timestamp>.bib. Do not create a durable second bibliography authority.

Report:

Original: N entries
Cited: M keys found in .tex
Filtered: M entries written to <temporary candidate>
Removed: N-M unused entries

If any cited keys are not found in the .bib, list them as warnings and recommend the configured bibliography validator or manual reconciliation.

Step 5: Validate before replacement

Before asking to replace anything:

1. verify every cited key is present, including crossref parents;
2. run the configured bibliography validator against the candidate;
3. compile the manuscript in a temporary build/output directory using the candidate;
4. compare bibliography entry count and rendered reference count with the expected cited set;
5. report the exact canonical target, its current hash, the candidate hash, kept keys, dropped keys, and validation result.

If any cited key is missing or validation/compilation fails, leave the canonical bibliography unchanged and stop.

Step 6: Approval-gated canonical replacement

Ask: “Replace <canonical-bib-path> with the validated cited-only candidate?” Only an explicit approval authorizes replacement.

After approval:

1. recheck the canonical file hash to detect concurrent Overleaf changes;
2. atomically replace the canonical .bib at the resolved Overleaf path;
3. compile again from the canonical source and confirm no missing citations or bibliography errors;
4. remove the temporary candidate after successful verification;
5. report the final canonical path and hash.

Do not retain both references.bib and filtered.bib in the Overleaf tree.

Dropped but strategically important comparator records belong in the literature authority, Paperpile, or a registered literature review packet—not in the live manuscript bibliography merely for safekeeping.

Edge Cases

• Multiple .bib files: preserve the driver's declared multi-file structure unless the user separately approves consolidation; stage one filtered candidate per canonical source
• \nocite{*}: Warn that all entries are cited and stop — filtering would be a no-op
• Cross-references: Some .bib entries use crossref = {parent-key}. If a cited entry cross-references another, include the parent even if it's not directly cited
• Empty result: If no citations found in .tex, warn and do not write an empty file
• Concurrent Overleaf change: if the canonical hash changes between staging and replacement, discard the replacement attempt, re-resolve, and repeat validation

---

Output Verification (Guard)

This skill writes files. Before any auto-commit, emit an outputs manifest and run the shared verifier. See skills/_shared/verify-outputs.md for the full protocol.

Required tail steps (before git commit):

1. Write manifest to <project>/.claude/state/outputs-manifest-<UTC-timestamp>.json listing every file this skill claims to have written in this invocation (paths relative to the project root).

2. Run:

   uv run python "<skills-root>/_shared/verify_outputs.py" \
       --manifest "$MANIFEST" \
       --project-root "$PROJECT_ROOT"
   

3. If the verifier exits non-zero, do not commit — surface the missing-files list to the user and stop. The verifier has already logged an error entry to ~/.local/state/ai-workflows/skill-outcomes.jsonl, which feeds the shared skill-health dashboard.

Why: closes the "hallucinated outputs" failure class (commit b2cff75, 2026-04-18).