| name | ai-regression-testing |
| description | Use this skill for regression testing in AI-assisted development, where the same model writes AND reviews code and so carries identical blind spots into both steps. It covers sandbox/mock-mode API testing without a database, an automated bug-check workflow, and the specific AI-introduced regression patterns to test for. Activate after an agent modifies API routes or backend logic, after fixing a bug, or when running a bug-check. |
AI Regression Testing
Testing patterns for AI-assisted development. When the same model writes code and then reviews its own work, it carries the same assumptions into both steps โ only automated tests catch the resulting blind spots.
When to activate
- An agent (re-forge, Claude Code, Cursor, Codex) modified API routes or backend logic
- A bug was found and fixed โ prevent its re-introduction
- The project has a sandbox/mock mode usable for DB-free testing
- Running a
/bug-check style review after changes
- Multiple code paths exist (sandbox vs production, feature flags)
The core problem
The failure loop: AI writes fix โ AI reviews fix โ AI says "looks correct" โ bug still exists. The most common AI-introduced regression is sandbox/production path inconsistency โ a field added to one path but not the other, repeated across multiple "fixes" because self-review shares the blind spot. A test catches it on the first run.
Sandbox-mode API testing
Most AI-friendly projects have a sandbox/mock mode โ the key to fast, DB-free API tests.
- Force sandbox mode in test setup: set the sandbox env flag and blank out real DB/service URLs so no database is needed.
- Use a request helper to construct route requests (method, body, headers, a sandbox user id header) and a response parser returning
{ status, json }.
- Assert the response contract: define a
REQUIRED_FIELDS list and assert every field is present, and assert specific previously-buggy fields are not undefined.
- Test sandbox/production parity: assert the sandbox path returns the same shape as production for the same endpoint.
What to do
- Run automated checks first (mandatory). Run the test suite and the build/type-check before any AI review. A failing test or type error is the highest-priority bug โ found mechanically, no AI judgment needed.
- Then do focused AI review, with known blind spots in mind: sandbox/production consistency, response shape matching frontend expectations, SELECT/query completeness, error handling with rollback, optimistic-update races.
- For each bug fixed, write a regression test named after the bug (e.g. "BUG-R1 regression"). Write tests for bugs that were found โ not for code that already works.
Common AI regression patterns
| Pattern | Test strategy | Priority |
|---|
| Sandbox/production mismatch | Assert same response shape in sandbox mode | High |
| SELECT/query column omission | Assert all required fields present in response | High |
| Error state leakage (stale data) | Assert related state is cleared on error | Medium |
| Missing rollback on optimistic update | Assert state restored on API failure | Medium |
| Type cast masking null/undefined | Assert field is not undefined | Medium |
Strategy: test where bugs were found
Don't chase 100% coverage. Test the exact endpoints/areas where bugs appeared. AI tends to make the same category of mistake repeatedly and bugs cluster in complex areas (auth, multi-path logic, state). Once tested, that exact regression cannot recur, and the test suite grows organically with fixes.
DO / DON'T
DO: write the test immediately when a bug is found (ideally before fixing); test response shape, not implementation; run tests as step 1 of every bug-check; keep tests fast (< 1s with sandbox mode); name tests after the bug they prevent.
DON'T: write tests for code that never had a bug; trust AI self-review as a substitute for automated tests; skip the sandbox path because "it's just mock data"; reach for integration tests when a unit test suffices; chase a coverage percentage instead of regression prevention.
