/configure:gitattributes

Audit or create a repo's .gitattributes, applying the three high-value attributes — merge=union for append-only tables, linguist-generated for build output, and LF normalization — while never applying merge=union to a file it would corrupt. See .claude/rules/gitattributes.md for the convention.

When to Use This Skill

Context

• Existing attributes: !find . -maxdepth 1 -name .gitattributes -type f
• Shell scripts present: !find . -name '*.sh' -not -path '*/.git/*' -path '*/scripts/*'
• Changelogs: !find . -name 'CHANGELOG.md' -not -path '*/.git/*'
• Generated trees: !find . -type d -name generated -not -path '*/.git/*'
• Rendered diagrams: !find . -path '*/docs/diagrams/*' -name '*.svg'
• Lockfiles: !find . -maxdepth 3 \( -name 'uv.lock' -o -name 'bun.lock' -o -name 'package-lock.json' -o -name 'pnpm-lock.yaml' \)

Parameters

Default (no flag): propose the additions, write them, and git add the file (staged, not committed).

Execution

Execute this .gitattributes configuration workflow:

Step 1: Read existing state

If a .gitattributes exists (from Context), Read it and record which paths are already covered. Otherwise plan to create one with a header comment.

Step 2: Classify candidates into the three buckets

Build the proposed line set from the Context detections, skipping any path already covered:

1. LF normalization — if any *.sh is present, add * text=auto eol=lf and *.sh text eol=lf. (No-op on an all-LF tree; protects shell shebangs.)
2. linguist-generated (always safe — display-only, no merge effect) — add one line per present group: **/CHANGELOG.md, .release-please-manifest.json (if present), each */generated/** tree, docs/diagrams/*.svg, and each lockfile found.
3. merge=union (the risky one) — do NOT auto-apply. Identify candidate one-line-per-entry, append-only files (e.g. a Markdown table every PR appends a single row to). Run the one-line test from .claude/rules/gitattributes.md. Exclude changelogs (multi-line entries), JSON, and code.

Step 3: Decide union marks interactively

For each merge=union candidate, confirm with AskUserQuestion ("Mark <path> as merge=union? It must be append-only with one logical entry per line."). Apply only confirmed paths. Under --fix, skip union entirely (apply only the zero-risk buckets) and list the candidates as recommendations.

Step 4: Write or report

• --check-only: print the proposed additions as a diff-style block; write nothing.
• Otherwise: append the missing lines to the existing .gitattributes (or create it with a header), each non-obvious line carrying a # comment for why it qualifies. Then git add .gitattributes.

Step 5: Verify any new union mark

For each newly added merge=union line, confirm it took effect: git check-attr merge -- <path> should report merge: union. Recommend a sandbox check (git merge --no-ff of two divergent single-row appends → zero conflict markers, both rows present) before relying on it.

Step 6: Report

Print a summary: lines added per bucket, union candidates deferred, and the git diff --cached hint. If --check-only, prefix with "DRY RUN — no files modified".

Important Notes

• merge=union only for one-line-per-entry append-only files. It keeps both sides of a conflicted hunk, so on an edited line it duplicates content, and on JSON it breaks syntax. When unsure, leave it off — the resolve-additive-conflicts.py pre-pass (in repos that have it) union-merges additive conflicts generically without the mark.
• linguist-generated is always safe — it only changes GitHub's diff display and language stats, never merge behavior.
• LF normalization is a no-op on an all-LF tree but prevents a future CRLF commit from breaking a shell shebang.
• Never auto-commit; stage and let the user review with git diff --cached.

Agentic Optimizations

See Also

• .claude/rules/gitattributes.md — the convention this skill applies
• /configure:repo — full repo onboarding (offers this skill)
• scripts/resolve-additive-conflicts.py (claude-plugins) — the deterministic union pre-pass for additive conflicts