Pre Commit Audit

Pre-Commit Audit — File Size, Anonymity, Secrets

Three-pass safety scan before commit. Blocks pain-points: oversized commits (1.3GB parquet incident), CCS-style anonymity breaches, leaked credentials. Each pass is a hard gate; user OKs proceed.

Hard Rules

Existential — block proceed

1. Size pass blocks anything >10MB unless gitignored OR user explicitly approves. Threshold matches block-large-files.sh hook for consistency.
2. Anonymity pass runs only on paper-relevant paths: paper-*/, *.tex, *.bib, *.md inside paper directories. Scope is data-driven — we don't false-flag README authorship.
3. Secrets pass uses an entropy heuristic + known-prefix list. No regex-only matching that misses high-entropy keys. Block on confirmed secret; warn on suspicious-but-uncertain.
4. All four passes run on the same file set (default: staged for commit). Don't mix staged-only size check with all-files anonymity check.
5. The Unicode pass is the one pass that is always in scope, including infra-only commits. It is identity-blind — it reads codepoints, never names — so it cannot false-flag authorship the way the anonymity pass can.

Format — catch in review

6. Output a single consolidated table (file × pass × verdict) — not four separate reports.
7. Severity tiers: BLOCK (must fix), WARN (proceed with confirm), OK.
8. Always show the file path and line number when flagging — make it copy-paste-fixable.

When to Use

• Before any git commit on a paper / project repo
• Pre-submission checklist (research projects)
• Before pushing to anonymous-artifact repo (the highest-risk anonymity surface)
• After a sub-agent has written files (verify before committing their work)
• Wired as a session-close Phase 4 sub-step (alternate path to running the hook)

When NOT to Use

• Before commits on infrastructure-only changes (skills, hooks, rules) — anonymity scan would false-flag README author lines. Exception: Phase 4 (Unicode) stays valid there and can be run alone: pre-commit-audit --unicode-only
• Auto-generated commits (CI, dependabot) — they don't need the scan
• When the change is purely metadata (.gitignore, LICENSE, etc.)

Modes

Architecture

Phase 1 (size)        → find files >10MB; check .gitignore status
Phase 2 (anonymity)   → grep paper paths for author/affiliation/email
Phase 3 (secrets)     → entropy + prefix heuristics on all staged files
Phase 4 (unicode)     → invisible-carrier scan via check_invisible_unicode.py
Phase 5 (consolidate) → single table; user OK to proceed

Phase 1: Size

THRESHOLD_BYTES=$((10 * 1024 * 1024))

git diff --cached --name-only --diff-filter=ACM | while read f; do
    if [ -f "$f" ]; then
        size=$(stat -f%z "$f" 2>/dev/null || stat -c%s "$f" 2>/dev/null || echo 0)
        if [ "$size" -gt "$THRESHOLD_BYTES" ]; then
            ignored=$(git check-ignore "$f" 2>/dev/null && echo "ignored" || echo "tracked")
            echo "$f|$size|$ignored"
        fi
    fi
done

Verdict: BLOCK if any tracked file >10MB. WARN if gitignored but staged anyway.

Phase 2: Anonymity

Single source of truth: skills/_shared/double-blind-anonymity-checklist.md holds the authoritative pattern list (paper-side P1-P8, artifact-side A1-A6) plus the CCS 2026 #1328 incident report that motivates the checks. Both pre-commit-audit (this skill) and anonymous-artifact reference it; do not duplicate the patterns inline.

Scope for pre-commit-audit: only files staged for commit at paths matching paper-*/, **/*.tex, **/*.bib, **/*.cls, **/*.sty. The full checklist also covers structured-metadata files (pyproject.toml, package.json, Cargo.toml, CITATION.cff, etc.) — those are checked by anonymous-artifact since they only matter when pushing the artifact, not on every commit.

Apply checks P1-P8 from the shared checklist:

• P1 — title page authors blinded
• P2 — no \thanks{} / \acknowledgements / \funded / \grant revealing identity
• P3 — first-person self-reference uses third person (no "our prior work")
• P4 — self-citation bib entries blinded if author overlap
• P5 — body text doesn't name authors of self-cited works
• P6 — no identifying URLs (personal sites, lab pages, repos with handles)
• P7 — PDF metadata sanitized (pdfinfo paper.pdf)
• P8 — figures contain no identifying watermarks/captions

For each match: print path:line and the matched text (truncated). Verdict: BLOCK on any match in a path containing paper-*/. WARN on matches outside paper directories.

Exception: a frontmatter-style Made by the user in friends-repo / public-repo READMEs is allowed (intentional attribution).

Phase 3: Secrets

Patterns by category:

Skip .env.template, .env.example, *.example, comment lines starting with #.

Verdict: BLOCK on any match in a non-template file.

Phase 4: Invisible Unicode

Catches characters that are invisible on screen but change what a machine reads — zero-width carriers, bidi overrides, tag characters, and space homoglyphs. These arrive by paste (Word, PDF, web page, LLM output) and fail silently: a U+200B inside a citekey makes BibTeX drop the entry with no error, a U+00A0 makes an exact-match string edit fail on a line that looks identical.

Deterministic — delegate to the script rather than grepping by hand:

git diff --cached --name-only --diff-filter=ACM \
  | while read -r f; do [ -f "$f" ] && printf '%s\n' "$f"; done \
  | xargs -r uv run --no-sync python "<skill-dir>/scripts/check_invisible_unicode.py" --format json

Map the script's severities onto this skill's tiers:

Report path:line:col, the U+XXXX label, and the containing construct when the JSON carries a context. The context is most of the signal: the same U+00A0 is a WARN in prose and a BLOCK inside \ref{}.

Fixes are mechanical, so offer them — but never apply silently (Anti-Patterns below):

uv run --no-sync python "<skill-dir>/scripts/check_invisible_unicode.py" <path> --diff    # preview
uv run --no-sync python "<skill-dir>/scripts/check_invisible_unicode.py" <path> --apply   # write

--apply refuses Overleaf-canonical paths unless given --yes-live-surface, per reconcile-before-rewriting.md; if a flagged file is a live Overleaf surface, surface the finding and let the user fix it in the editor rather than passing that flag on their behalf.

Phase 5: Consolidate + ask

Single table:

File                                   Size   Anon   Secrets  Uni    Verdict
────────────────────────────────────────────────────────────────────────────
paper-ccs/paper/main.tex               5KB    OK     OK       OK     OK
paper-ccs/paper/references.bib         12KB   FAIL   OK       FAIL   BLOCK   (line 23: the user; line 8: U+200B in bib_entry_key)
data/raw/results.parquet               1.3GB  N/A    N/A      N/A    BLOCK   (size; not gitignored)
scripts/sync-creds.sh                  2KB    OK     FAIL     OK     BLOCK   (line 15: hf_<...>)
notes/from-word.md                     4KB    OK     OK       WARN   WARN    (4x U+00A0)

If any BLOCK rows: ask the available structured-question mechanism:

• Fix the issues and rerun — recommended
• Override and proceed — risky, requires explicit confirmation
• Cancel commit — abort

If only WARN rows: ask whether to proceed. If all OK: print "Pre-commit clean ✓" and exit silently.

Cross-References

Anti-Patterns

• Don't scan binary files for anonymity strings — false positives on PDFs, images.
• Don't rely solely on regex for secret detection — entropy heuristics catch keys with novel prefixes.
• Don't auto-fix issues — surface them for the user to handle. Auto-replacement of names risks breaking acknowledgements.
• Don't skip the .gitignore check — large gitignored files staged anyway are a signal of git add -A mistakes.
• Don't require this skill on infra-only commits. Add a path filter or let the user opt out for those — but --unicode-only remains safe and useful there.
• Don't run Phase 4's --apply as part of the audit. The audit reports; the user decides. Auto-rewriting bytes in a file the user is mid-edit on is the exact failure reconcile-before-rewriting.md exists to prevent.
• Don't flag visible non-ASCII. Accented names (López-Ibáñez), em dashes, and → are legitimate and the script deliberately ignores them; a pass that flags them will be muted within a week.