Refiner

E2E Test Refiner

Overview

Takes a structured test case and makes it deterministic by running it repeatedly, identifying failure patterns, and refining both the test case and shared conventions. Stops when N consecutive runs pass without the runner deviating from the documented steps.

Core principle: A test isn't done until it passes N times in a row with identical execution. Every workaround discovered belongs in the test case or conventions, not in the runner's head.

When to Use

• A new test case was just authored (handoff from browser-test:author)
• An existing test is flaky or inconsistent
• User says "stabilize this test" or "make this test reliable"

Do NOT use when:

• No structured test case exists yet (use browser-test:author first)
• A stable test just needs to be executed (use browser-test:runner)

Prerequisites

• Structured test case file exists in tests/e2e/{area}/
• The Playwright MCP server must be available (configured in .mcp.json)
• Dev server must be running (or E2E_BASE_URL set to a reachable environment)

Process

digraph refiner_flow {
    rankdir=TB;
    "Read test case + conventions" -> "Execute test (fresh session)";
    "Execute test (fresh session)" -> "Pass?" [label="run"];
    "Pass?" -> "Analyze failure" [label="no"];
    "Analyze failure" -> "Update test or conventions" [label="pattern identified"];
    "Update test or conventions" -> "Execute test (fresh session)" [label="reset counter"];
    "Pass?" -> "N consecutive passes?" [label="yes"];
    "N consecutive passes?" -> "Execute test (fresh session)" [label="no, keep going"];
    "N consecutive passes?" -> "Done - test is stable" [label="yes"];
}

1. Read the test case and conventions

Read the test case file, global conventions, and area conventions. Understand every step before running.

Check the test's Environments line against E2E_ENVIRONMENT (default: local). If the current environment is not listed, refuse to execute. Use E2E_BASE_URL (default: http://localhost:5173) wherever the test references it.

Validate the three-hook structure . The test case MUST contain three top-level sections: ## Before Hook, ## Test Steps, ## After Hook. If any section is missing, stop and report — this is a hand-off bug from the author skill, not a refinement issue. Empty sections are allowed only when the markdown uses the explicit _None — {reason}._ placeholder body. Fixing missing sections is an author-skill responsibility, not a refiner-skill responsibility.

2. Execute with fresh session

Follow the Session Management procedure from global conventions:

1. Close any existing browser session (browser_close)
2. Navigate to E2E_BASE_URL - this starts a fresh browser instance
3. Verify clean session (cookie consent banner should appear on first navigation)

3. Execute the test

Follow the test case steps in hook order: ## Before Hook → ## Test Steps → ## After Hook. Do not improvise or add steps not in the document. If you must deviate, that's a signal the test case needs updating.

After Hook is mandatory and runs even if Test Steps fails. Treat the execution as a try/finally: if any Test Step fails, capture the failure evidence, then still run every step in ## After Hook so the next refinement run starts from a clean state. Record any After Hook failures separately (a teardown failure does not mask a test step failure).

4. On failure: analyze and update

When a step fails:

1. Screenshot the current state
2. Note which hook failed — Before Hook, Test Steps, or After Hook. Failures in different hooks have different fix patterns:

• Before Hook failure → setup is non-deterministic or convention drift. Fix the setup step or the referenced convention.
• Test Steps failure → standard refinement (timing, selector, missing wait).
• After Hook failure → cleanup is not defensive enough. Add a visibility guard or re-read state before mutating.

1. Identify the root cause - is it a timing issue, missing wait, unexpected element, stale selector?
2. Classify the fix:

• Hook misclassification: A step is in the wrong section (setup leaked into Test Steps, or a Test Step is restoring state and should be in After Hook). Move it to the correct section in the markdown.
• Test-specific: Update the test case file (e.g., add a conditional step for a banner)
• Area-wide pattern: Update the area conventions (e.g., document a common popup across all purchase tests)
• Global pattern: Update global conventions (e.g., a new wait strategy)

1. Apply the fix to the appropriate file
2. Reset the consecutive pass counter to 0

5. On pass: increment counter

When the test passes, increment the consecutive pass counter. Continue running until reaching N consecutive passes (default: 5).

6. Report results

When done, report:

• Total runs attempted
• Files modified (test case and/or conventions)
• Patterns discovered and where they were documented
• Final N consecutive pass confirmation

Common Refinement Patterns

| Symptom | Likely Fix | Where to Document | |---------|-----------|-------------------| | Click does nothing | Add await_element before click | Global conventions (Interaction Patterns) | | Input gets wrong value | Add click-to-focus before type | Global conventions (Interaction Patterns) | | Element obscured by overlay | Add conditional dismiss step | Area conventions or test case | | Redirect loses context | Use time-based wait + snapshot check | Global conventions (Wait Patterns) | | Stale selector after navigation | Wait for landmark on target page | Test case step | | Pre-populated data unexpected | Document in test preconditions | Test case preconditions | | Step times out at 15s/30s/60s+ | Switch to smartWaitFor(locator) from playwright-tests/utils/timeouts.ts — retry-aware cap (2/3/4 min for attempt 0/1/2+) with networkidle fail-fast. Use { timeout: currentTimeout() } for native Playwright assertions. | Compiled spec; do NOT add a new fixed timeout. | | Lazy/virtualised section never visible | Use smartScrollIntoView(locator) (re-issues scroll until target paints). | Compiled spec. | | Whole test exceeds test.setTimeout because one wait consumed the budget | Bump test.setTimeout to at least 360000 (≥1.5× retry-2 cap). Never set the test budget equal to or below currentTimeout(). | Compiled spec. |

Rules

• Never work around a problem silently. Every deviation from the written steps must be captured as a test case or conventions update.
• Reset counter on any change. If you modify the test case or conventions, the N consecutive passes must restart from 0.
• Fresh session every run. No exceptions. Session bleed between runs is the #1 source of false passes.
• Default N = 5. User can override this. For critical flows, suggest N = 10.
• Three-hook structure is load-bearing . The author skill produces tests with ## Before Hook / ## Test Steps / ## After Hook sections; the compiler relies on those boundaries to emit test.beforeEach / test.step / registerTeardown. If refinement uncovers a step in the wrong section, treat that as a legitimate fix and move it — don't accept a misclassified test as stable.
• After Hook always runs. The refiner executes Before → Test Steps → After Hook with try/finally semantics. A test isn't stable if its After Hook is fragile, even if Test Steps pass consistently — record After Hook failures and refine the cleanup steps.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: Autodesk
• Source: Autodesk/claude-browser-test-skills
• License: Apache-2.0

Install and usage instructions live in the source repository linked above.