Agent Skills: AI-Powered Mutation Testing

AI-Powered Mutation Testing

Target: $ARGUMENTS

If no argument provided, default to files changed in the current branch (via git diff).

This skill implements the mutation testing workflow: generate targeted mutants, filter unproductive ones, run tests to find survivors, and harden the test suite by strengthening or creating tests that kill surviving mutants.

Step 1: Gather Context

1a. Identify Target Files

Determine what to mutate based on $ARGUMENTS:

1. File path — Mutate the specified source file
2. Directory — Mutate source files in the directory (exclude test files)
3. No argument — Use git diff to find changed source files:

   git diff --name-only $(git merge-base HEAD main)...HEAD -- '*.ts' '*.tsx' ':!*.spec.*' ':!*.test.*'
   

If no target files found, notify the user and stop.

1b. Gather Supporting Context

For each target source file, collect:

1. Source code — Read the full file contents
2. Existing tests — Find corresponding test files (*.spec.ts, *.test.ts) via naming convention or import analysis
3. Test coverage — Run tests with coverage for the target file to identify uncovered lines:

   bun run test -- --coverage --collectCoverageFrom='<target-file>' --silent 2>&1 | tail -20
   

4. Recent git history — Check for recent defects or frequent changes:

   git log --oneline -10 -- <target-file>
   

5. Risk factors — Assess which risk factors apply to each file:

| Risk Factor | Indicators | |---|---| | Data security / compliance | Handles PII, auth, authorization, encryption, tokens | | Integrations | External API calls, HTTP clients, webhooks, message queues | | Code vs data model | Database queries, ORM entities, schema validation | | Historic defects | Frequent bug-fix commits, complex conditional logic | | Change impact | Exported utility functions, shared modules, base classes | | Performance | Loops over large datasets, recursive calls, caching logic |

1c. Determine Test Runner

Detect the project's test framework from package.json scripts and devDependencies:

• Jest: bun run test
• Vitest: bun run test
• Other: adapt accordingly

Step 2: Generate Mutants

2a. Create Experimental Branch

git stash --include-untracked || true
git checkout -b mutation-testing/$(date +%Y-%m-%d-%H%M%S)
git stash pop || true

2b. Generate Risk-Factor-Guided Mutants

For each target source file, generate 3-5 mutants that target the identified risk factors. Apply these mutation operators:

| Operator Type | Examples | |---|---| | Decision mutations | Change > to >=, && to \|\|, === to !== | | Value mutations | Change constants, swap string literals, alter numeric values | | Statement mutations | Remove guard clauses, delete error handling, skip validation | | Integration mutations | Remove timeout handling, skip error code checks, bypass retry logic | | Security mutations | Remove auth checks, bypass input validation, expose sensitive data in logs |

For each mutant, produce:

1. Mutant ID — Sequential identifier (M001, M002, etc.)
2. File and line — Exact location of the change
3. Mutation description — What was changed and why
4. Risk factor — Which risk factor this targets
5. The code change — A minimal, atomic edit to introduce the defect

Mutant quality criteria — Each mutant must be:

• Non-trivial — Not just whitespace, comments, or formatting
• Buildable — The project must still compile/typecheck
• Realistic — Simulates a plausible programming error
• Targeted — Addresses a specific risk factor
• Atomic — Exactly one logical change per mutant

2c. Apply and Validate Each Mutant

For each generated mutant, one at a time:

1. Apply the mutation — Edit the source file to introduce the defect
2. Build check — Run bun run typecheck to verify the project still compiles
   • If build fails: discard the mutant, revert the change, log failure reason, continue to next
3. Commit the mutant — git commit -am "mutant: M00X - <description>"

Step 3: Filter Mutants

3a. Run Tests Against Each Mutant

Process mutants in reverse order (LIFO — latest commit first):

1. Run relevant tests against the mutant:

   bun run test -- <test-file-path> --silent 2>&1
   

2. Evaluate result:
   • Test fails (mutant killed) → Revert the mutant commit. Log as killed. Proceed to next mutant.
   • Test passes (mutant survived) → The mutant reveals a test gap. Keep it for Step 4.

3b. Equivalence Detection for Surviving Mutants

For each surviving mutant, verify it is not semantically equivalent to the original:

1. AST-level comparison — Compare the diff. If the change is purely syntactic (reordering independent statements, renaming to equivalent aliases), discard it.
2. Behavioral analysis — Reason about whether any input could distinguish the mutant from the original:
   • If no distinguishing input exists → Discard as equivalent
   • If a distinguishing input exists → Keep as a unique actionable mutant
   • If uncertain → Keep and flag for human review

3c. Record Mutant Metadata

For each mutant, maintain a record:

{
  "mutant_id": "M001",
  "file": "<source-file>",
  "line": 42,
  "risk_factor": "Data security / compliance",
  "description": "Removed authentication check before data access",
  "status": "survived|killed|equivalent|discarded",
  "commit_hash": "<hash>",
  "tests_run": ["test-file.spec.ts"],
  "equivalence_result": "not_equivalent|equivalent|uncertain"
}

Print a summary table after filtering:

| Mutant | File | Risk Factor | Status |
|--------|------|-------------|--------|
| M001   | ... | ...          | survived |
| M002   | ... | ...          | killed   |

Step 4: Harden Test Suite

For each surviving, non-equivalent mutant:

4a. Determine Test Strategy

1. Existing test covers the mutated region → Strengthen the existing test
2. No test covers the mutated region → Generate a new test

4b. Strengthen Existing Test or Generate New Test

When strengthening an existing test, modify the test to:

• Add assertions that detect the behavioral difference introduced by the mutant
• Add or modify test inputs that expose the defect in the mutant code
• Preserve the test method name, signature, and overall structure
• Avoid adding trivial or unrelated checks

When generating a new test, create a test that:

• Specifically exercises the behavioral difference between original and mutant code
• Uses strong, precise assertions that detect the exact defect
• Focuses on the identified risk factor
• Includes realistic test data that exposes the mutant's flawed behavior
• Follows existing test naming conventions and structure from the codebase

4c. Validate Each Test (3-attempt limit)

For each improved or new test, validate with up to 3 attempts:

1. Revert the mutant — Switch to original code
2. Build check — Ensure the test compiles
3. Run on original code — Test must PASS on unmodified code
4. Re-apply the mutant — Switch back to mutated code
5. Run on mutant code — Test must FAIL on the mutant

If validation fails, refine the test (up to 3 total attempts). If still failing after 3 attempts, flag for manual review.

4d. Apply Decision Matrix

Classify each test result:

| Original Code | Mutant Code | Signal | Action | |---|---|---|---| | Builds + Passes | Builds + Fails | Strong detection | Keep — alert developer | | Builds + Passes | Builds + Passes | Weak assertion | Refine — strengthen assertion | | Builds + Fails | Builds + Passes | Bad test | Discard — generate new test | | Doesn't build | Any | Broken test | Discard — generate new test | | Builds + Passes | Doesn't build | Untestable mutant | Discard — notify developer | | Builds + Fails | Builds + Fails | Ambiguous signal | Discard — generate new test |

4e. Finalize Tests

For each validated test:

1. Add an inline comment above the test referencing the mutant:

   // Test hardened to kill mutant M001 (Risk Factor: Data security / compliance)
   

2. Commit the test improvement to the experimental branch

Step 5: Cleanup and Report

5a. Revert All Mutants

Remove all mutant commits from the experimental branch, keeping only test improvements:

# Revert mutant commits (identified by "mutant:" prefix in commit message)
git log --oneline | grep "^.* mutant:" | awk '{print $1}' | while read hash; do
  git revert --no-commit "$hash"
done
git commit -m "chore: revert all mutants after test hardening"

5b. Validate Clean State

1. Run full test suite on the clean code with test improvements:

   bun run test 2>&1
   

2. All tests must pass. If any fail, investigate and fix.

5c. Report Results

Print a comprehensive mutation testing report:

## Mutation Testing Report

**Target**: <files tested>
**Date**: <timestamp>
**Branch**: <experimental branch name>

### Summary
- Total mutants generated: X
- Killed by existing tests: X
- Survived (test gaps found): X
- Equivalent (discarded): X
- Build failures (discarded): X
- **Mutation Score**: X% (killed / (total - equivalent))

### Test Improvements
- Tests strengthened: X
- New tests generated: X
- Tests validated successfully: X
- Tests requiring manual review: X

### Surviving Mutants (Unresolved)
| Mutant | File | Line | Risk Factor | Description |
|--------|------|------|-------------|-------------|
| M00X   | ...  | ...  | ...         | ...         |

### Oracle Gap
- Code coverage: X%
- Mutation score: X%
- Oracle gap: X% (coverage - mutation score)

### Risk Factor Coverage
| Risk Factor | Mutants | Killed | Survived | Score |
|---|---|---|---|---|
| Data security | X | X | X | X% |
| Integrations  | X | X | X | X% |

5d. Present Options

Ask the user:

1. Merge test improvements — Cherry-pick test commits to the original branch, delete experimental branch
2. Keep experimental branch — For further review before merging
3. Discard everything — Delete the experimental branch, no changes kept

If the user chooses to merge:

git checkout <original-branch>
git cherry-pick <test-improvement-commits>
git branch -D <experimental-branch>

Never

• Modify test files to make them pass on mutants (defeats the purpose)
• Generate mutants in test files (only mutate source code)
• Skip the build check after generating a mutant
• Commit mutants to protected branches (dev, staging, main)
• Leave mutant code in the codebase after completion
• Generate more than 5 mutants per file (diminishing returns)
• Skip equivalence detection for surviving mutants
• Mark a test as valid without running it on both original and mutant code
• Use --no-verify with any git command