pre-land-review

name	pre-land-review
description	The review gate that must pass before any commit lands. Called automatically by the /commit skill — do NOT invoke directly when the user says "commit", "land", "ship it", etc. (use /commit instead). Invoke directly only when you want to run the review without committing, or when another skill references it. This is a MANDATORY gate — do not create a git commit without completing this checklist.
user-invocable	true
argument-hint	[git-ref range, default HEAD~1]

Pre-Land Review Checklist

This skill is the landing gate for vibeProj. Every commit must pass through it. The checklist has two tiers: deterministic checks (enforced by the pre-commit hook) and AI-powered analysis (run as review agents).

Tier 1: Deterministic Checks

Run each command. ALL must pass before committing.

uv run ruff check src/ tests/
uv run ruff format --check src/ tests/
uv run python scripts/check_projections.py --all

If any fail, fix the issues before proceeding. The pre-commit hook will also enforce these, but catching them here avoids a failed commit attempt.

Tier 2: AI-Powered Review Agents

The pre-commit hook CANNOT do this — it requires AI judgment. Reviews run as dedicated agents so they analyze the code with fresh eyes, without being biased by the context of having written the code.

Step 1: Gather context

1. Run git diff --cached --name-only (or git diff HEAD~1 --name-only) and save the file list.
2. Run git diff --cached (or git diff HEAD~1) and save the full diff.
3. Categorize the changed files:
   • kernel/GPU code (fused_kernels.py, helmert.py, _ds_device_fns.py, runtime.py, CUDA source strings): needs GPU code review + performance + zero-copy analysis
   • pipeline/transform (pipeline.py, transformer.py): needs performance + zero-copy analysis
   • projection xp (projections/*.py): needs maintainability
   • CRS/ellipsoid (crs.py, ellipsoid.py): needs maintainability
   • tests only: skip AI analysis (deterministic checks suffice)

Step 2: Launch review agents

Based on the categories, spawn the applicable agents in parallel using the Agent tool. Each agent gets the file list and full diff injected into its prompt. Launch all applicable agents in a SINGLE message.

Category	Agent	When
GPU Code Review	gpu-code-reviewer	kernel/GPU/NVRTC code changed
Zero-Copy	zero-copy-reviewer	pipeline/transform/kernel code changed
Performance	performance-reviewer	kernel/pipeline/transform code changed
Maintainability	maintainability-reviewer	any non-test source code changed

Prompt template for each agent:

Review the following changes for vibeProj.

## Changed Files
{file_list}

## Full Diff
{diff}

The agents already know their review procedure and severity rules from their agent definitions. You only need to provide the diff context.

Step 3: Collect and report

Wait for all agents to complete, then compile results.

Report Format

## Pre-Land Review

### Changed Files
[list with categories]

### Deterministic Checks
[PASS/FAIL for each]

### Agent Reviews

#### GPU Code Review: [CLEAN / BLOCKING ISSUES]
[findings or "N/A — no GPU code touched"]

#### Zero-Copy Analysis: [CLEAN / LEAKY / BROKEN]
[findings or "N/A"]

#### Performance Analysis: [PASS / FAIL]
[findings or "N/A"]

#### Maintainability: [CONSISTENT / GAPS / BROKEN]
[findings or "N/A"]

### Overall Verdict
[LAND / FIX REQUIRED]

Note: LAND requires zero BLOCKING findings across all agents.

Severity Rules

All review agents follow these severity rules (defined in their agent definitions):

• BLOCKING — Must fix before landing. Default for all findings.
• NIT — Only for pure style preferences with zero functional or performance impact.

"Existing codebase does it too" is NEVER a valid NIT justification. If new code builds on a broken upstream pattern, that is BLOCKING — fix the upstream too.

Rules

• ALL deterministic checks must pass.
• ANY BLOCKING agent finding means FIX REQUIRED.
• Test-only changes need only deterministic checks (skip agents).
• Be concise — this is a gate, not a code review.

After Review

If the verdict is LAND, write the content-addressable review marker:

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

The marker is single-use — delete it after a successful commit.

Then proceed with the commit.