Verify observable browser outcomes

Resolve the requested workflow, intended local target and relevant application tests before choosing a script. Reuse the existing runner, fixtures and package graph. An open port or one detected service does not establish the application's identity or authorize its mutations. Use disposable test data and a known test account for state-changing flows.

Establish runtime and server readiness

Check installed package and browser versions against their current stable release metadata. Provision dependencies explicitly in an isolated workspace from a complete lockfile. Keep the original skill package unchanged. Avoid implicit installation during test execution.

Prefer the project's server lifecycle integration. If starting a server, retain its process identity, drain or redirect both output streams, bound startup, and check that the launched process remains alive. Verify an application-specific HTTP or UI condition before testing; a successful TCP connection alone is insufficient. Record ownership before stopping processes and clean up the whole owned process tree with a mechanism qualified for the operating system.

Read a helper's interface before invocation and inspect its implementation when assessing its fitness or diagnosing failure. The original bundled server helper has reproduced readiness, log-draining and child-cleanup defects; it is not a qualified general server supervisor.

Act and assert

Navigate to the explicit target and wait for the relevant rendered condition. Prefer accessible roles, labels and stable test IDs. Use retrying assertions for dynamic state. Network idle and fixed sleeps do not prove UI readiness.

Arm response, download or popup waits before the triggering action. Match the request identity and expected status, then assert the user-visible outcome. Treat navigation or authentication failures as failures. Do not blindly retry an action whose effect may already have occurred; first inspect the resulting state or use a controlled idempotent fixture.

Use isolated contexts and the project's viewport/device matrix. A screenshot is evidence to inspect, not a correctness verdict. Distinguish functional, accessibility and visual coverage. Control clocks, animations and data for image comparisons and review baseline changes independently.

Preserve evidence and cleanup

Use portable output paths and unique run directories. Register console and page-error capture before navigation; keep diagnostics within the task's data boundaries. Traces and authentication state can contain sensitive information. Close contexts and browsers in finally blocks so artifacts finish even when assertions fail. Preserve the command's actual status and failed attempts.

Report tested outcomes, failed/skipped/not-run checks, exact versions and artifact paths. Do not count swallowed errors, retries or screenshot creation as passing behavior. See qualified controls and limits.