manual-ui-testing · git:20260914.2afcf40 · 2026-09-14 · sha256 92b7d756fd678b75
manual-ui-testing git:20260914.2afcf40A
Immutable. This exact content never changes and is served at /api/v1/blob/92b7d756fd678b75.
No tool to install first
claude plugin marketplace add https://markdownregistry.com/mp/a/art_cyeys5fdh7eovhbb/92b7d756fd678b75a504ef07e89b09ae950fc23beaafb3647f1aaa1c9da03804.json
Then /plugin install manual-ui-testing@markdownregistry-cyeys5fdh7eovhbb in Claude Code. It installs this exact version, the one graded A here, pinned to the commit it was graded at, and leaves it there.
--- name: manual-ui-testing description: Run manual UI test cases using agent-browser against a running stack. Use when the user asks to run UI tests, test the UI, run manual tests, or verify UI behavior. metadata: internal: true user-invocable: true allowed-tools: Bash(npx agent-browser:*), Bash(agent-browser:*), Bash(just:*), Bash(doppler:*) --- # Manual UI Testing Execute the test cases in `knowledge/test-cases/ui/` against a running stack with `agent-browser`, record results, and file issues for failures. [`knowledge/test-cases/format.md`](../../../knowledge/test-cases/format.md) defines the test case format. ## Scope Each subdirectory of `knowledge/test-cases/ui/` is a category — list the directory rather than assuming a fixed set. Every case states its own preconditions (auth mode, existing data), test data, steps, and expected result; read them before running. With no scope given, run everything in dependency order: auth → org → features. ## Stack Full auth mode needs the real stack, not DEV_MODE. Start it with a unique per-worktree prefix and wait for PostgreSQL, Valkey, NATS, API, worker, UI, and Caddy to come up. Follow the canonical coding-agent startup contract in the root [`AGENTS.md`](../../../AGENTS.md), selecting the auth mode required by the test case. Verify it through the documented `/health` endpoint before testing. If a stack is already running, confirm its `PORT_PREFIX` and auth mode before using it. ## Driving the browser `agent-browser` runs headless Chromium via `npx`; the daemon persists between commands in a session. ```bash agent-browser open http://localhost:<prefix>00/<path> agent-browser wait --load networkidle agent-browser snapshot -i # → @e1 [input type="email"], @e2 [button] "Submit", … agent-browser fill @e1 "value" agent-browser click @e2 agent-browser wait --load networkidle agent-browser snapshot -i agent-browser screenshot /tmp/test_<category>_<tc>.png ``` Hints that cost time to rediscover: - Refs are invalidated by any navigation or DOM change — re-snapshot after every one. - Always `wait --load networkidle` after navigation, form submission, and before screenshots. - Chain independent commands with `&&`; do not chain when the next ref depends on reading a snapshot. - Next.js dev compilation can add 2–5s to a first page load. - Keyboard shortcuts (Ctrl+K) often do not reach headless Chromium — drive the equivalent click. - Screenshot each significant step, not just the verdict. - Element missing from a snapshot? `agent-browser scroll down` first. - Login redirect loop usually means `AUTH_MODE` does not match the test category. - Testing an **HTTPS** target (a deployed environment rather than the local stack) from a cloud agent: every page fails with `ERR_CONNECTION_RESET` until Chrome's TLS version is capped. Export the flags once, before the first `open`, so a daemon relaunch cannot drop them (EVE-807): ```bash export AGENT_BROWSER_ARGS=$'--ssl-version-max=tls1.2\n--disable-features=PostQuantumKyber' pgrep -a chrome | tr ' ' '\n' | grep ssl-version # confirm they reached Chrome ``` Chrome's TLS 1.3 ClientHello carries a post-quantum key share too large for the egress relay, which cuts the tunnel mid-handshake. `curl` is unaffected, so the proxy looks healthy by hand. A local `http://localhost:<prefix>00` stack never hits this — no TLS, no proxy. Pass these through `AGENT_BROWSER_ARGS`, not `--args`: `--args` splits on commas, so `--disable-features=A,B` becomes two argv entries and Chrome exits with `Multiple targets are not supported in headless mode`. ## Recording Write or update `.local/test-results/MANUAL_TEST_RESULTS_<YYYY-MM-DD>.md` using [`references/results-template.md`](references/results-template.md). Partial or re-test runs append to (or update) the existing file for that date. If the user asks for issues to be filed, create them via Linear MCP (EVE team, OSS project) with severity, repro steps, expected vs actual, and the test case ID (e.g. `org_creation/TC003`).