Automated Test-Fix Loop

Automated Test-Fix Loop Pattern

Overview

This skill provides a generalized pattern for continuous automated testing and fixing using Claude Code. The system creates a feedback loop: Test → Fail → Issue → Fix → Test that runs until all tests pass.

Use this pattern when you want:

• Automated bug fixing during development
• Regression testing after major changes
• CI/CD integration with self-healing tests
• Overnight fuzzing with automatic fixes

Architecture

Test Runner (run-tests.sh)
  ├─> Runs test suite with timeout
  ├─> Captures output and environment context
  └─> On failure: Creates GitHub issue
         ↓
    GitHub Issues (with full context)
         ↓
Issue Fixer (fix-issue.sh)
  ├─> Fetches issue from GitHub
  ├─> Invokes Claude Code with detailed prompt
  └─> Claude investigates, fixes, tests, closes issue
         ↓
Auto-Fix Loop (auto-fix-loop.sh)
  └─> Orchestrates: Test → Fix → Repeat until green

Core Components

1. Test Runner Script Template

#!/bin/bash
set -euo pipefail

# Configuration
TIMEOUT_SECONDS=10
OUTPUT_FILE="/tmp/test-output.txt"
COMMIT=$(git rev-parse HEAD 2>/dev/null || echo "unknown")
BRANCH=$(git branch --show-current 2>/dev/null || echo "unknown")
DATE=$(date -u +"%Y-%m-%d %H:%M:%S UTC")
OS_INFO=$(uname -a)

# Run tests with timeout
timeout $TIMEOUT_SECONDS <YOUR_TEST_COMMAND> 2>&1 | tee "$OUTPUT_FILE"
EXIT_CODE=${PIPESTATUS[0]}

# Annotate timeout
if [ $EXIT_CODE -eq 124 ]; then
    EXIT_CODE="124 (timeout after ${TIMEOUT_SECONDS}s)"
fi

# On failure: Create GitHub issue
if [ $EXIT_CODE -ne 0 ]; then
    # Strip ANSI color codes
    CLEAN_OUTPUT=$(sed 's/\x1b\[[0-9;]*m//g' "$OUTPUT_FILE")

    # Create issue
    gh issue create \
        --title "Test Failure: <TEST_NAME>" \
        --body "$(cat <<EOF
## Test Failure Report

**Exit Code:** $EXIT_CODE
**Date:** $DATE
**Commit:** $COMMIT
**Branch:** $BRANCH
**OS:** $OS_INFO

### Test Output
\`\`\`
$CLEAN_OUTPUT
\`\`\`

### Diagnostic Information
<Add service status, connectivity checks, etc.>

### Files Involved
- Test: <path/to/test>
- Implementation: <path/to/impl>

### Expected Behavior
<What should happen>

### Actual Behavior
<What actually happened>
EOF
)"

    exit 1
fi

echo "All tests passed!"
exit 0

Key Customization Points:

• <YOUR_TEST_COMMAND>: Language-specific test runner
  • Python: pytest tests/
  • Node.js: npm test
  • Rust: cargo test
  • Go: go test ./...
• TIMEOUT_SECONDS: Appropriate for your test suite
• Diagnostic checks: Database connections, API availability, etc.

2. Issue Fixer Script Template

#!/bin/bash
set -euo pipefail

# Parse arguments
ISSUE_NUM="$1"
MODEL="${2:-haiku}"
PERMISSION_MODE="interactive"

# Parse options
while [[ $# -gt 0 ]]; do
    case $1 in
        --model)
            MODEL="$2"
            shift 2
            ;;
        --auto-edit)
            PERMISSION_MODE="acceptEdits"
            shift
            ;;
        --no-prompts)
            PERMISSION_MODE="dangerouslySkip"
            shift
            ;;
        *)
            shift
            ;;
    esac
done

# Fetch issue
ISSUE_TITLE=$(gh issue view "$ISSUE_NUM" --json title --jq '.title')
ISSUE_BODY=$(gh issue view "$ISSUE_NUM" --json body --jq '.body')

# Create prompt with guidance
PROMPT=$(cat <<EOF
GitHub Issue #$ISSUE_NUM: $ISSUE_TITLE

$ISSUE_BODY

IMPORTANT: Project-Specific Guidance
- <What code to modify vs what code to leave alone>
- <Where to look for the bug>
- <Common pitfalls to avoid>

Steps:
1. Read the test output carefully
2. Examine the source code
3. Identify the root cause
4. Implement a fix
5. Test with: ./run-tests.sh
6. If tests pass, close the issue with: gh issue close $ISSUE_NUM

You MUST verify the fix by running tests before closing the issue.
EOF
)

# Build Claude Code command
CLAUDE_OPTS=(--model "$MODEL")

if [ "$PERMISSION_MODE" = "dangerouslySkip" ]; then
    CLAUDE_OPTS+=(--dangerously-skip-permissions)
elif [ "$PERMISSION_MODE" != "interactive" ]; then
    CLAUDE_OPTS+=(--permission-mode "$PERMISSION_MODE")
fi

# Invoke Claude Code
claude "${CLAUDE_OPTS[@]}" "$PROMPT"

Model Selection:

• haiku: Simple bugs, syntax errors (fast, cheap)
• sonnet: Most issues (balanced)
• opus: Complex architectural problems (slow, expensive)

Permission Modes:

• interactive (default): User approves each action
• acceptEdits: Auto-accept file edits, prompt for bash
• dangerouslySkip: No prompts (sandboxed environments only!)

3. Auto-Fix Loop Script Template

#!/bin/bash
set -euo pipefail

MAX_ITERATIONS=${1:-10}
MODEL=${2:-haiku}
TEST_MODE=${3:-}

echo "Starting auto-fix loop (max $MAX_ITERATIONS iterations, model: $MODEL)"

iteration=0
while [ $iteration -lt $MAX_ITERATIONS ]; do
    echo "========================================="
    echo "Iteration $((iteration + 1))/$MAX_ITERATIONS"
    echo "========================================="

    # Run tests
    if ./run-tests.sh $TEST_MODE; then
        echo "SUCCESS! All tests passed!"
        exit 0
    fi

    # Wait for GitHub API
    sleep 2

    # Find open issues
    OPEN_ISSUES=$(gh issue list --state open --search "Test Failure" --json number --jq '.[].number')

    if [ -z "$OPEN_ISSUES" ]; then
        echo "No open test failure issues found"
        sleep 2
        continue
    fi

    # Fix the first issue
    ISSUE_NUM=$(echo "$OPEN_ISSUES" | head -1)
    echo "Fixing issue #$ISSUE_NUM..."

    ./fix-issue.sh --model "$MODEL" --no-prompts "$ISSUE_NUM"

    iteration=$((iteration + 1))
done

echo "Max iterations reached without success"
exit 1

Sandboxing Strategies

Why Sandbox?

Using --dangerously-skip-permissions requires sandboxing because Claude Code can:

• Execute arbitrary bash commands
• Modify any files
• Make network requests

Only use --dangerously-skip-permissions in isolated environments!

Option 1: Local QEMU VM

# Start VM
qemu-system-x86_64 \
    -m 2048 \
    -hda debian-vm.qcow2 \
    -netdev user,id=net0,hostfwd=tcp::2224-:22 \
    -enable-kvm -cpu host \
    -display none -daemonize

# Copy project
rsync -avz --exclude='target/' --exclude='.git/' \
    ./ user@localhost:~/project/ -e "ssh -p 2224"

# Run in VM
ssh -p 2224 user@localhost 'cd ~/project && ./auto-fix-loop.sh'

Pros: Full isolation, easy reset, works offline Cons: Requires local resources, disk space limits

Option 2: Remote Server VM

# Copy to remote with lots of space
rsync -avz ./ user@remote:/path/to/testing/

# Start VM on remote
ssh user@remote 'qemu-system-x86_64 ... -daemonize'

# Forward VM SSH port
ssh -L 2230:localhost:3260 user@remote

# Access remote VM as if local
ssh -p 2230 user@localhost 'cd ~/project && ./auto-fix-loop.sh'

Pros: Unlimited disk space, doesn't consume local resources, can run overnight Cons: Requires network, more complex setup

Option 3: Docker Container

FROM <base-image>
RUN apt-get update && apt-get install -y gh <build-tools>
COPY . /project
WORKDIR /project
CMD ["./auto-fix-loop.sh"]

docker build -t auto-tester .
docker run --rm auto-tester

Pros: Lightweight, fast startup Cons: Less isolation than VM

Usage Examples

Initial Development

# 1. Write tests for new feature (they fail)
./run-tests.sh

# 2. Run auto-fix loop (interactive)
./auto-fix-loop.sh 20 sonnet

# 3. Review changes
git diff

# 4. Commit if satisfied
git commit -am "Implement feature X"

Regression Testing

# After making changes
./run-tests.sh

# If issues created
./fix-issue.sh --auto-edit 15

# Verify
./run-tests.sh

Overnight Fuzzing (in VM)

# In sandboxed VM
nohup ./auto-fix-loop.sh 100 haiku > overnight.log 2>&1 &

# Next morning
tail overnight.log
gh issue list --state closed

Best Practices

1. Issue Quality

Include in issues:

• Test command that failed
• Exit code (with timeout annotation)
• Full test output (cleaned of ANSI codes)
• Environment: commit, branch, OS, date
• Diagnostic info: service status, connectivity
• Files involved: test code, implementation code
• Expected vs actual behavior

Bad issue: "Tests failed. Output: Error"

2. Test Timeouts

• Unit tests: 10-30 seconds
• Integration tests: 1-5 minutes
• E2E tests: 5-15 minutes

Too short = false positives. Too long = wasted time on hangs.

3. Preventing Test Modifications

Critical: Add to fix-issue.sh prompt:

IMPORTANT: The tests in tests/ are CORRECT and must NOT be modified.
Fix the implementation code in src/, not the test code.

Claude Code may try to "fix" tests to make them pass. Explicitly forbid this.

4. Iteration Limits

Set based on:

• Test suite size
• Time budget (overnight = 50+, quick check = 5-10)
• Cost concerns

Typical: 10-20 iterations

Language-Specific Adaptations

Python

# run-tests.sh
pytest tests/ --junit-xml=results.xml || EXIT_CODE=$?

Node.js

# run-tests.sh
npm test 2>&1 | tee test-output.txt || EXIT_CODE=$?

Rust

# run-tests.sh
cargo test --all 2>&1 | tee test-output.txt || EXIT_CODE=$?

Go

# run-tests.sh
go test ./... -v -timeout 30s || EXIT_CODE=$?

Troubleshooting

Claude Fixes Tests Instead of Code

Problem: Modifies test files to make tests pass

Solution: Add explicit guidance in fix-issue.sh:

IMPORTANT: Tests are CORRECT. Fix implementation code only.

Infinite Loop (Same Issue Reopens)

Problem: Fix doesn't solve the problem

Solution:

• Add more context to issues (logs, diagnostics)
• Escalate to stronger model (haiku → sonnet → opus)
• Add project-specific debugging hints

Tests Pass Locally, Fail in VM

Problem: Different environment

Solution:

• Ensure VM has all dependencies
• Check environment variables
• Verify file permissions
• Compare tool versions

Security Considerations

Safe to Auto-Fix

• ✅ Unit tests
• ✅ Integration tests (internal)
• ✅ Linting/formatting
• ✅ Type errors

Review Before Accepting

• ⚠️ Security-sensitive code (auth, crypto)
• ⚠️ Database migrations
• ⚠️ API contract changes
• ⚠️ Dependency updates

Never Auto-Fix

• ❌ Production deployments
• ❌ Credentials/secrets
• ❌ Access control
• ❌ Billing/payment code

Quick Start Checklist

• Install gh CLI: sudo apt-get install gh
• Authenticate: gh auth login
• Create run-tests.sh with your test command
• Create fix-issue.sh with project guidance
• Create auto-fix-loop.sh for orchestration
• Test manually: ./run-tests.sh (should create issue)
• Fix manually: ./fix-issue.sh 1 (check it works)
• Run loop: ./auto-fix-loop.sh 5 haiku
• For full automation: Set up sandboxed VM
• In VM: Use --no-prompts mode

References

• GitHub CLI: https://cli.github.com/
• Claude Code: https://claude.ai/claude-code
• Full pattern documentation: See AUTOMATED_TESTING_PATTERN.md in reference projects