commit

name	commit
description	PROACTIVELY USE THIS SKILL when the user says "commit", "land", "land this", "ship it", "done", "wrap up", "let's finish", or any intent to commit work. This is the ONLY user entrypoint for the commit workflow — do not invoke /pre-land-review directly for commits. Orchestrates the full landing flow: pre-land review, staging, review marker, and git commit.
user-invocable	true
argument-hint	[optional commit message override]

Commit — Full Landing Flow

You are landing work. Follow these steps exactly in order. Do not skip any step. Do not create a git commit without completing the review.

Step 1: Run /pre-land-review

Invoke the pre-land-review skill. This runs:

• All deterministic checks (ruff, zero-copy compliance)
• AI-powered sub-agent reviews (GPU code review, zero-copy enforcer, performance analysis, maintainability enforcer) as applicable

If the review finds BLOCKING issues, stop here. Fix them and re-run /commit. Do not proceed to Step 2 with blocking findings.

Step 2: Stage changes

After the review passes with verdict LAND:

1. Run git status to see what needs staging.
2. Run git diff --cached --name-only to see what is already staged.
3. Stage the appropriate files. Prefer staging specific files by name over git add -A. Never stage .env, credentials, or large binaries.
4. If the user specified which files to commit, stage only those.

Step 3: Write the review marker

The commit-msg hook (if present) requires a content-addressable review marker. Write it:

printf '{\n  "timestamp": "%s",\n  "staged_hash": "%s",\n  "files": [%s],\n  "verdict": "LAND"\n}\n' \
  "$(date -u +%Y-%m-%dT%H:%M:%SZ)" \
  "$(git diff --cached | sha256sum | cut -d' ' -f1)" \
  "$(git diff --cached --name-only | sed 's/.*/"&"/' | paste -sd,)" \
  > .claude/.review-completed

IMPORTANT: If you stage any additional files AFTER writing the marker, the hash will no longer match. Always write the marker as the LAST step before git commit.

Step 4: Create the commit

1. Analyze the staged diff to draft a concise commit message.
2. If the user provided $ARGUMENTS, use that as the commit message (or incorporate it).
3. Create the commit using a HEREDOC for proper formatting:

git commit -m "$(cat <<'EOF'
<commit message here>

Co-Authored-By: Claude <co-author tag>
EOF
)"

4. Run git status after the commit to verify success.

Step 5: Report

Tell the user the commit was created. Show the commit hash and summary.

Rules

• NEVER skip the pre-land review.
• NEVER use --no-verify or -n flags. The pre-land-gate hook blocks these.
• If the pre-commit hook fails (ruff, zero-copy, etc.), fix the issues and retry. Do NOT amend — create a new commit.
• The review marker is single-use: deleted after a successful commit.
• If the commit fails for any reason, re-run /commit from the top.

Review Enforcement — Zero Tolerance

These rules govern how you handle findings from the pre-land review. They are NON-NEGOTIABLE.

NEVER reclassify BLOCKING as NIT

If a sub-agent flags something as BLOCKING, it stays BLOCKING. You do not have authority to downgrade it. The only valid NIT is a known codebase-wide gap that requires a coordinated migration — not "it works fine" or "it's minor". When in doubt, it is BLOCKING.

NEVER excuse new debt with existing debt

"Other functions don't do this either" is not a justification. Every new line of code must meet the standard. If existing code has a shortcoming, your new code must not compound that problem by building on it. Instead:

• If the fix is small (< 5 minutes): fix the upstream function too.
• If the fix is structural: spawn a cuda-engineer agent with the context needed to fix the upstream API, then build your code against the fixed version.

The goal is to shrink the cleanup backlog, not grow it.

NEVER accept CPU work in GPU code paths

If a review finds host round-trips (D->H->D), CPU fallbacks, or numpy calls inside a GPU-dispatched function, this is ALWAYS BLOCKING. Common excuses that are NOT valid:

• "The upstream API returns host arrays" — Fix the upstream API or spawn a cuda-engineer agent to do so.
• "It's just one small transfer" — Small transfers have large latency. Every D->H->D is a pipeline stall. Fix it.
• "We'll clean it up later" — No. Later never comes.

What to do when a finding requires upstream fixes

Do NOT mark the review as LAND and leave a TODO. Instead:

1. Spawn a cuda-engineer agent (background, worktree) with:
   • The specific function/API that needs fixing
   • What it currently returns (e.g., host array)
   • What it should return (e.g., device array)
   • The downstream code that needs it
2. Wait for or incorporate the fix.
3. Re-run /commit against the complete change.