Failing test appears → stop, diagnose, fix before continuing
Do not batch failures hoping they share a root cause
Use the iterative fix loop when a failure needs more than one pass (see iterative fix loop)

TDD Discipline

For new features: red-green-refactor, interface-first, one test at a time.

Red-green-refactor — Write failing test → implement minimal pass → refactor.
Interface-first — Confirm interface changes before implementation; design for testability.
One test at a time — Each test drives one behavior; don't batch.
Deep modules — Thin interfaces, rich internals (see coding philosophy).

Targeted re-runs drive iteration; TDD drives what to test first. See tdd-discipline.md for full workflow.

Run Strategy by Scenario

Different scenarios call for different scopes and limits. See running-strategy.md for full details.

Scenario	Scope	Workers	Max Failures	Retries
Adding new tests	Targeted file/pattern only	4	1	0
Debugging failing suite	Targeted suite	4	1	1
Post-fix verification	Targeted suite	4	2	1
Full suite / CI	Full suite	4	3	2

Adding new tests: run only the file being added. Once targeted tests pass, expand to the full project suite with --workers 4.

Debugging failures: max 1 failure before stopping, 1 retry to distinguish flaky from deterministic. Never use --max-failures 3 for targeted/debug runs — that batches failures and defeats "address failures as they come."

Iterative fix loop

When iterating on failing tests across multiple turns, use a tracked loop:

Identify the failing test(s) and generate a task ID (e.g. fix-sources-e2e)
Initialize _agent/{task-id}/tests.json with the failing test(s) as criteria
Each iteration: run targeted tests → read output → fix → update tests.json
When all tests in scope pass, expand scope and run broader suite

Use this any time you expect 2+ fix iterations so pass/fail state stays explicit between turns.

tests.json schema:

{
  "task": "fix-sources-e2e",
  "criteria": [
    {
      "file": "apps/client/src/hooks/data/auth/__tests__/requiresEmailVerification.test.ts",
      "description": "sources list invalidates after mutation",
      "status": "failing"
    }
  ]
}

status per entry: "failing" | "passing" | "skipped". Update each entry after each run. Use unique task IDs across sessions to avoid collision — prefix with a timestamp if needed (1714000000-fix-sources-e2e). Add _agent/ to .gitignore if not already present.

After review findings

When fixing council review findings on a large branch, use the review fix-loop: read prior chat synthesis, fix theme-complete batches, run the verify ladder before re-review — see review-fix-loop.md § Verify ladder. For test iteration on the same branch, _agent/{task-id}/tests.json (§ Iterative fix loop) is separate from review memory.

Flaky Tests

When debugging flakiness: use --max-failures 1 --retries 1. Do not use --max-failures 3 — it batches failures and obscures which test flaked.

Interpreting output: Playwright reports "flaky" when a test failed on first attempt but passed on retry. Use that to distinguish timing/race issues from deterministic bugs.

Mark known flaky tests so future runs surface them:

// In spec file — test-level (known flaky, fix in progress)
test('foo', { retries: 2 }, async ({ page }) => { ... });

// Or describe-level for a whole block
test.describe.configure({ retries: 1 });

Add a brief comment when using retries so it's clear the test is flaky and should be fixed, not hidden.

Commands

Canonical Nx / pytest / Playwright commands, Playwright setup, and test-layer taxonomy: docs/developer/testing.md. Apply Run Strategy when choosing scope, workers, and --max-failures.

Quick Examples

E2E

import { test } from './fixtures';

test('user can navigate home', async ({ authenticatedPage }) => {
  await authenticatedPage.goto('/workbench');
  const home = authenticatedPage.getByRole('button', { name: /home/i });
  await home.click();
  await expect(home).toHaveAttribute('aria-pressed', 'true');
});

Integration

import { renderHook, act } from '@testing-library/react';

test('hook opens pane', () => {
  const { result } = renderHook(() => useWorkbench(), { wrapper });
  act(() => { result.current.open('editor', { fileId: 'main.ts' }); });
  expect(result.current.state.layout.childNodes).toHaveLength(1);
});

Unit

test('reducer creates pane', () => {
  const { layout } = openReducer(initial, intent, env);
  expect(layout.childNodes).toHaveLength(1);
});

Reference

TDD Discipline - Red-green-refactor, interface-first, deep modules
Running Strategy - Scope, worker, and failure limit decisions
E2E Testing (Playwright) - Patterns, fixtures, selectors, wait strategies
Test-Driven Debugging - Full workflow for fixing broken features
Workbench Testing - Testing workbench-core and workbench-react
Best Practices - Guidelines for all test types

Output format

Follow references/output-schema.md. Preflight: Docker required for backend integration tests — scope to unit/file lint if unavailable.

testing

Testing

When to Use

Skill boundaries

Core Discipline