Imported from gabrielmoreira/agent-skills-mirror (mirrors/repos/mlopscommunity@Coding-Agents-Conference-skills/skills/validation-runner/SKILL.md). Install upstream with npx skills add gabrielmoreira/agent-skills-mirror --skill validation-runner. Copyright stays with the author.

Validation Runner

Overview

The Validation Runner Skill performs comprehensive validation in a separate context from the coding agent. This is critical because if an agent makes false assumptions during coding, it will bake those assumptions into its validation. By running validation as a sub-agent with fresh context, you catch issues that the coding agent missed and prevent flawed implementations from being considered "complete."

Core principle: Fresh validation context catches assumption-based bugs. If the coder made a false assumption, it contaminates their own validation. You need independent verification.

Dependency: Claude Code CLI with test infrastructure, linting tools, and the ability to run code.

When to Use

• After each coding session: Before merging code to main
• After parallelized agent work: Before merging multiple agents' code
• Before deployment: Final validation before production
• When you're uncertain about code quality: Get an independent assessment
• For critical systems: MLOps pipelines, data processing, ML models where bugs are expensive

When NOT to Use

• During live pairing or rapid iteration (too slow)
• For trivial changes (typo fixes, comments, obvious improvements)
• When you trust the coding agent's validation completely (you shouldn't)

Common Mistakes

The Validation Process

Step 1: Gather Validation Scope

Understand what we're validating:

• What code was changed? (files, lines of changes)
• What's the scope? (new feature, bugfix, refactor?)
• What systems does it touch?
• Are there any known tricky areas?
• What's the risk level? (critical, high, medium, low?)

Step 2: Environment Setup

Set up the validation environment:

# Install dependencies (if needed)
npm install  # or pip install -r requirements.txt, etc.

# Set up test database (if needed)
# Load test fixtures
# Configure test environment

# Show what we're validating
git diff HEAD~1  # or show changed files

Step 3: Run Linting

Execute all linting checks:

JavaScript/TypeScript:

npm run lint
# or
npx eslint . --fix --report-unused-disable-directives
# or
npx biome check .

Python:

pylint src/ --fail-under=8.0
flake8 src/
black --check src/
mypy src/

Go:

go vet ./...
golint ./...
go fmt ./...

General:

# Check for trailing whitespace, CRLF endings, etc.
# Check for commented-out code
# Check for debug prints (console.log, print(), println!)
# Check for TODOs/FIXMEs

Output format:

## Linting Results
✓ PASSED: ESLint (0 violations)
✓ PASSED: Black formatting
✗ FAILED: TypeScript strict null checks
  - Line 45: Variable 'user' could be undefined
  - Line 67: Cannot assign 'string' to 'number'

Custom linting checks:
✓ No console.log() in production code
✓ No commented-out code blocks
✗ WARNING: 3 TODO comments (review if intentional)

Step 4: Run Unit Tests

Execute the test suite:

# Run all tests
npm test  # or pytest, go test, etc.

# Capture output:
# - Total tests run
# - Tests passed
# - Tests failed
# - Test coverage percentage
# - Coverage gaps

Output format:

## Unit Test Results
Total tests: 342
✓ PASSED: 340
✗ FAILED: 2

Failed tests:
  - src/models/user.test.ts: "User should not allow negative age"
  - src/api/handlers.test.ts: "GET /users/{id} should return 404 for missing user"

Coverage summary:
  Statements: 87.3% (good, was 85%)
  Branches: 72.1% (ok, watch for untested error paths)
  Functions: 91.2% (good)
  Lines: 88.4% (good)

Coverage gaps (>5% uncovered):
  - src/error_handling/retry_logic.ts (coverage: 34%)
  - src/admin/cleanup_service.ts (coverage: 21%)

If coverage decreased significantly, flag it.

Step 5: Custom Pattern Checks

Run project-specific validation scripts. Create checks for patterns relevant to your codebase.

Examples for MLOps:

# Check model artifact versioning
find models/ -name "*.pkl" -o -name "*.joblib" | \
  while read f; do
    # Ensure artifacts have metadata
    [ -f "${f}.metadata.json" ] || echo "Missing metadata: $f"
  done

# Check training data assumptions
grep -r "assume_categorical\|drop_duplicates\|fillna" src/preprocessing/ | \
  grep -v "tested" | \
  while read line; do echo "Review assumption: $line"; done

# Check that model config changes are documented
if git diff HEAD src/config.yaml | grep -q "model:"; then
  [ -s CHANGELOG.md ] || echo "WARNING: Model config changed but no CHANGELOG entry"
fi

Examples for general code:

# No hardcoded credentials
grep -r "password\|api_key\|secret" src/ | \
  grep -v "^Binary" | \
  grep -v "\.env\|\.example" | \
  grep -v "test_\|TEST_" | \
  while read line; do echo "Potential hardcoded credential: $line"; done

# Database queries use parameterized queries
grep -r "SELECT.*FROM" src/ | \
  grep -E "f\"|\.format\(%s\)|'\s*\+\s*" | \
  while read line; do echo "Check for SQL injection risk: $line"; done

# No synchronous file I/O in async contexts
grep -r "fs\.readFileSync\|open.*read\(" src/ | \
  grep -v "test\|spec" | \
  while read line; do echo "Potential blocking I/O: $line"; done

Output format:

## Custom Pattern Checks
✓ PASSED: No hardcoded credentials
✓ PASSED: All SQL queries use parameterized statements
✗ FAILED: Blocking file I/O in async function
  - Line 234 of src/data_loader.py: open() blocks event loop
  - Review: Consider using aiofiles

Step 6: UI/Screenshot Validation (if applicable)

If the code includes UI changes:

# Generate screenshots of key user flows
npm run screenshot:login
npm run screenshot:dashboard
npm run screenshot:settings
# or use Playwright/Puppeteer to generate

Validate:

• Components render without errors
• Layout is correct (no overlapping, proper spacing)
• Text is readable (no cutoff, proper colors)
• Responsive design looks good (mobile, tablet, desktop views)
• Visual regressions compared to previous version

Output format:

## UI Validation
✓ Login flow: Renders correctly
✓ Dashboard: All widgets load
✗ Settings page: Form labels misaligned on mobile
  - Screenshot: validation_failures/settings_mobile.png
  - Issue: Label text overlaps input fields

Visual regressions:
  - Dashboard card spacing increased by 2px (intentional? check design)

Step 7: Run E2E Tests (if available)

If project has E2E tests:

# Run Playwright, Cypress, or other E2E framework
npx playwright test
# or
npm run e2e

# Capture:
# - Total scenarios run
# - Passed/failed
# - Any flaky tests
# - Error messages

Output format:

## E2E Test Results
Total scenarios: 24
✓ PASSED: 22
✗ FAILED: 2
⚠ FLAKY: 1 (passed on retry)

Failed scenarios:
  - "User can upload CSV file and see results" (connection timeout)
  - "Admin can deactivate user account" (permission denied, may be test data issue)

Flaky test:
  - "Model training completes in time limit" (sometimes slow CI runner)

Step 8: Analyze Relevant Logs

Check for error patterns:

# Application logs (if running)
tail -100 logs/application.log | grep -i "error\|exception\|warning"

# Test output for warnings
# Build output for deprecation warnings
# Runtime warnings from linters

Output format:

## Log Analysis
Warnings found:
  - DeprecationWarning: use_param_X is deprecated (4 occurrences)
  - TypeError in error handler (review line 342)
  - Uncaught promise rejection in data fetch (review async handling)

Recommendations:
  - Replace deprecated function calls
  - Review error handler type safety
  - Ensure all promises are caught

Step 9: Compile Validation Report

Create a comprehensive summary:

# Validation Report

**Date**: [today]
**Code scope**: [what was validated]
**Risk level**: [low/medium/high/critical]

## Summary
✓ [X] PASSED all critical checks
✗ [Y] FAILED [N] check(s), [M] warning(s)

## Results by Category

### Linting
Status: [PASS/FAIL]
- [list major issues or "no issues found"]
- Recommendations: [if any]

### Unit Tests
Status: [PASS/FAIL]
Total: 342 | Passed: 340 | Failed: 2
Coverage: 87.3% (was 85.1%) ✓ Improved
Failed tests:
- [test name]: [brief issue]
- [test name]: [brief issue]

### Custom Patterns
Status: [PASS/FAIL]
- [pattern check]: [result]
- Recommendations: [if any]

### UI Validation (if applicable)
Status: [PASS/FAIL]
- Visual regression: [status]
- Responsive design: [status]
- Issues: [list or "none"]

### E2E Tests (if applicable)
Status: [PASS/FAIL]
Total: 24 | Passed: 22 | Failed: 2
- Failed scenario: [name and reason]
- Flaky test detected: [if any]

## Issues Requiring Attention

### Critical (blocks deployment)
- [ ] [Issue 1]: [description] - FIX REQUIRED before merge
- [ ] [Issue 2]: [description] - FIX REQUIRED before merge

### High (should fix)
- [ ] [Issue 1]: [description] - Should fix before merge
- [ ] [Issue 2]: [description] - Should fix before merge

### Medium (nice to fix)
- [ ] [Issue 1]: [description] - Consider fixing
- [ ] [Issue 2]: [description] - Consider fixing

### Low (informational)
- [Note 1]: [description]
- [Note 2]: [description]

## Validation Confidence

Based on:
- Code review coverage: [good/adequate/incomplete]
- Test coverage adequacy: [good/adequate/incomplete]
- Known edge cases tested: [yes/no]
- Integration with other systems checked: [yes/no/N.A.]

Overall assessment: [The code is ready for merge / Recommend fixes before merge / Code needs significant work]

## Next Steps

If PASS:
- [ ] Approve for merge
- [ ] Ready for deployment (or further testing)

If FAIL:
- [ ] Return to coding agent with specific issues to fix
- [ ] Re-run validation after fixes

Step 10: Flag for Human Review

If you encounter issues that require human judgment:

## Requires Human Review

### Design Decision Question
Issue: [description]
Question: Should we [option A] or [option B]?
Context: [relevant info]
→ Waiting for human decision

### Breaking Change Warning
Change: [description]
Impact: [what breaks]
Migration path: [is there one?]
→ Needs approval before proceeding

### Known Limitation
Situation: [description]
Why: [reason]
Workaround: [if available]
→ Document as known limitation or fix?

Special Cases

If Validation Fails Unexpectedly

• Don't blame the coding agent
• Report what failed, why, and what concrete fix would help
• Provide specific error messages and line numbers
• Consider: Is this a test infrastructure issue, or a real code issue?

If You Detect Invalid Assumptions

Example: Coding agent assumed database table X exists, but it doesn't

• Report this clearly: "Code assumes [assumption], but [reality]"
• Provide evidence: show where code makes assumption
• Suggest fix: "Either add migration to create table X, or change code to use table Y"

If Coverage Drops Significantly

• Identify what's uncovered
• Ask: Is this intentional (adding code that doesn't need testing) or a gap?
• High-risk, uncovered code (error handling, security) should be flagged

Configuration Notes

Project-Specific Setup

1. Linting tools: Specify your stack:

   JavaScript/TypeScript: ESLint + Prettier (or Biome)
   Python: pylint + flake8 + mypy + black
   Go: gofmt + vet + golangci-lint
   Custom: [add your tools]
   

2. Test command: Update validation:

   # Your project's test command
   npm test
   # or
   pytest --cov=src --cov-report=html
   # or
   cargo test

3. Custom patterns: Add checks relevant to your project:

   # MLOps: Check model artifact versions
   # Data pipeline: Validate data schema transformations
   # Security: Check for credential exposure
   # Performance: Check for N+1 queries

4. Coverage thresholds: Set acceptable coverage:

   Minimum coverage: 80%
   Critical paths: must be 95%+
   New code: should increase or maintain coverage
   

5. E2E test configuration: If using Playwright/Cypress:

   Timeout: [your setting]
   Retry failed tests: [yes/no]
   Screenshot on failure: [yes/no]
   

How to Invoke

Create a separate validation session (do NOT run this in the same session as coding):

/validation-runner

[Paste or describe the code that was changed]

Or after coding work completes:

Use the validation-runner skill to verify the code quality.
[Provide files changed or commit SHA]

Tips from Practitioners

When to run this: After the coding agent finishes, before merging. Never skip validation.

Fresh context is critical: The validator should NOT know what the coder was thinking. Fresh eyes catch assumptions.

Test flakiness: If a test fails inconsistently, that's a real problem. Flag it and investigate. Don't ignore "it sometimes passes."

Coverage gaps matter: Uncovered error handling and edge case code is high risk. Ask: should this be tested, or is it safe to leave untested?

Common validation mistakes:

• Skipping unit tests because "we did E2E testing"
• Not checking coverage reports carefully
• Ignoring deprecation warnings
• Not validating assumptions (file paths, config, external service availability)

Performance validation: If code touches data, ML models, or large datasets:

• Benchmark performance before and after
• Check for N+1 query patterns
• Verify model inference performance
• Check for memory leaks in long-running jobs

Pro tip: Save detailed validation reports. When bugs appear in production, review the validation report. You'll often find "we could have caught this."

Red flags that demand investigation:

• Test coverage decreases by >5%
• Linting violations that can't be auto-fixed
• E2E tests that are flaky
• Unit tests that timeout
• Custom pattern checks finding issues
• Log analysis showing error patterns

Key Principles

1. Fresh context catches assumption bugs. If the coder made assumptions, validation in the same context validates the wrong thing.
2. Every check serves a purpose. Linting catches style. Tests catch logic. Pattern checks catch domain-specific issues. Coverage catches untested paths.
3. Failures are data. Don't ignore or work around test failures. They tell you what's broken.
4. Assumptions are risks. Flag them and verify them. Code that assumes file paths, config existence, or service availability will break in different environments.
5. Coverage matters. Untested code is time bombs. High-risk code (error handling, security) must be tested.

Attribution

Based on techniques from the Coding Agents Lunch & Learn Series. Rob Ennals (Anthropic) and Rahul contributed the "Punch It" skill concept: "Validation must run in a fresh context. If the coding agent made assumptions, those same assumptions will be baked into the validation. You need independent verification."