DEV Community

Nova Elvaris
Nova Elvaris

Posted on

The Review Gate Pattern: Auto-Check AI Output Before It Hits Your Branch

#ai #testing #programming #productivity

You asked your AI assistant to add a caching layer. It generated 200 lines of clean code. You skimmed it, committed, pushed. The PR got merged.

Three days later: a production bug. The generated code cached error responses.

A review gate would have caught this in seconds.

The Pattern

A review gate is an automated check that runs between "AI generates code" and "code enters your branch." Think of it as a linter specifically designed for AI-generated output.

AI generates code → Review Gate → Your branch
                        ↓
                    Pass: commit
                    Fail: fix or regenerate
Enter fullscreen mode Exit fullscreen mode

Gate 1: The Diff Size Check

AI-generated code tends to be verbose. If you asked for a 20-line change and got 200 lines, something went wrong.

#!/bin/bash
# gate-diff-size.sh — Reject oversized AI changes

MAX_LINES="${1:-100}"
DIFF_LINES=$(git diff --cached --stat | tail -1 | grep -oP '\d+ insertion' | grep -oP '\d+')

if [ "${DIFF_LINES:-0}" -gt "$MAX_LINES" ]; then
  echo "❌ GATE FAIL: $DIFF_LINES insertions exceeds limit of $MAX_LINES"
  echo "   AI may have over-generated. Review before committing."
  exit 1
fi
echo "✅ Diff size OK: $DIFF_LINES insertions"
Enter fullscreen mode Exit fullscreen mode

Install as a pre-commit hook:

cp gate-diff-size.sh .git/hooks/pre-commit
chmod +x .git/hooks/pre-commit
Enter fullscreen mode Exit fullscreen mode

Gate 2: The Import Scanner

AI models love importing things you don't use — or don't have installed.

#!/bin/bash
# gate-imports.sh — Check for phantom imports in staged files

STAGED=$(git diff --cached --name-only --diff-filter=AM | grep -E '\.(ts|js|tsx|jsx)$')
FAIL=false

for file in $STAGED; do
  # Check for imports of packages not in package.json
  IMPORTS=$(grep -oP "from ['\"]([^./][^'\"]*)" "$file" | sed "s/from ['\"]//")
  for pkg in $IMPORTS; do
    BASE_PKG=$(echo "$pkg" | cut -d'/' -f1-2 | sed 's/@//' | sed 's/^/@/')
    if ! grep -q "\"$BASE_PKG\"" package.json 2>/dev/null; then
      echo "❌ Unknown import: '$pkg' in $file"
      FAIL=true
    fi
  done
done

if $FAIL; then
  echo "   Run 'npm install' or remove unused imports"
  exit 1
fi
echo "✅ All imports verified"
Enter fullscreen mode Exit fullscreen mode

Gate 3: The TODO/FIXME Detector

AI models leave breadcrumbs. Catch them before they become permanent.

#!/bin/bash
# gate-todos.sh — Block commits with AI placeholder markers

STAGED=$(git diff --cached --name-only --diff-filter=AM)
MARKERS="TODO|FIXME|HACK|XXX|PLACEHOLDER|IMPLEMENT ME|YOUR_.*_HERE"
FAIL=false

for file in $STAGED; do
  HITS=$(grep -nE "$MARKERS" "$file" 2>/dev/null)
  if [ -n "$HITS" ]; then
    echo "❌ Placeholder found in $file:"
    echo "$HITS" | head -5
    FAIL=true
  fi
done

if $FAIL; then
  echo "   AI left placeholders. Resolve before committing."
  exit 1
fi
echo "✅ No placeholders found"
Enter fullscreen mode Exit fullscreen mode

Gate 4: The Type Check Gate

If you're using TypeScript, this is free — but most people don't run tsc before committing AI code.

#!/bin/bash
# gate-typecheck.sh — Type-check staged files

STAGED_TS=$(git diff --cached --name-only --diff-filter=AM | grep -E '\.tsx?$')

if [ -n "$STAGED_TS" ]; then
  npx tsc --noEmit 2>/tmp/tsc-errors.txt
  if [ $? -ne 0 ]; then
    echo "❌ TypeScript errors in AI-generated code:"
    cat /tmp/tsc-errors.txt | head -20
    exit 1
  fi
fi
echo "✅ Type check passed"
Enter fullscreen mode Exit fullscreen mode

Combining Gates

Chain them in a single pre-commit hook:

#!/bin/bash
# .git/hooks/pre-commit

set -e

echo "🔍 Running AI review gates..."
./gates/gate-diff-size.sh 150
./gates/gate-imports.sh
./gates/gate-todos.sh
./gates/gate-typecheck.sh
echo ""
echo "✅ All gates passed — safe to commit"
Enter fullscreen mode Exit fullscreen mode

When Gates Fail

A gate failure isn't a disaster — it's the system working. When a gate fails:

  1. Read the error. It tells you exactly what's wrong.
  2. Fix or regenerate. Sometimes a one-line fix is enough. Sometimes you need to re-prompt with better constraints.
  3. Update the gate. If you keep hitting false positives, adjust the threshold. If you keep missing bugs, add a new gate.

The Numbers

Since adding review gates to my workflow:

  • Placeholder TODOs shipped to main: 0 (was ~2/week)
  • Phantom import errors in CI: 0 (was ~1/week)
  • Oversized AI commits: caught and split 4 times in the first month
  • Time cost: ~3 seconds per commit

Four bash scripts. Three seconds per commit. Zero AI-generated placeholders in production.

Your AI assistant doesn't review its own work. That's your job — but there's no reason to do it manually.

Top comments (0)