This is the single top-level location for browser-test evidence.

playwright/
├── artifacts/   # transient screenshots, HTML, traces, logs, and buffers
└── snapshots/   # reviewed visual baselines used by screenshot assertions

artifacts/ is ignored by Git and is uploaded by CI when the browser suite fails. snapshots/ is intentionally not ignored: visual baselines are test inputs and must be reviewed alongside the test that owns them.

The current pytest-Playwright suite produces diagnostic screenshots and other failure evidence. It does not yet contain Playwright visual-regression assertions (to_have_screenshot); adding one should place its approved baseline under snapshots/.

Use the long-lived launcher when an interactive browser driver needs to exercise CRUD workflows outside pytest:

uv run --all-packages python playwright/serve_hermetic.py

The launcher prints one JSON connection record, then keeps the server alive:

{"base_url": "http://127.0.0.1:18612", "fake_agent": true, "pid": 12345}

It reuses the pytest TestWorld boundary, creates a temporary HOME/config/ session database/plans directory, and uses the deterministic fake agent by default. No LAN authentication is enabled. Stop it with Ctrl-C; its temporary world and child server are cleaned up together.

For a different shell, connect with the installed headless driver:

agent-browser --session studyloop-crud open http://127.0.0.1:18612
agent-browser --session studyloop-crud snapshot -i

--real-agent is available only as an explicit opt-out and should not be used for deterministic CRUD assertions.