Browser qa

browser-QA

Drive a real browser: start the dev server, navigate, snapshot, interact, observe, capture evidence.

Scope: $ARGUMENTS

If no scope was provided, read the recent changeset to determine what needs verification.

Tool Detection

Three viable drivers; check in order.

if [ -n "${MCP_CHROME_LOADED:-}" ] || command -v claude-mcp-chrome >/dev/null 2>&1; then
  DRIVER=mcp-chrome   # use mcp__claude-in-chrome__* tools directly
elif command -v agent-browser >/dev/null 2>&1; then
  DRIVER=agent-browser
elif command -v playwright-cli >/dev/null 2>&1; then
  DRIVER=playwright-cli
else
  echo "ESCALATE: no browser driver available. Install agent-browser (preferred), playwright-cli, or load the chrome-browser MCP." >&2
  exit 1
fi
echo "Using driver: $DRIVER"

If MCP is loaded, mcp__claude-in-chrome__* tools are the native interface and the CLI examples below translate to tool calls.

Driver-specific references (load on demand):

• agent-browser (preferred CLI): references/agent-browser/overview.md.
• playwright-cli (fallback CLI): references/playwright-cli/overview.md.
• chrome-browser MCP: tool schemas load on demand via ToolSearch.

Workflow

1. Read the changeset — git diff, git log, understand what UI surface changed.
2. Detect the dev-server start command — look for package.json scripts, Procfile, bin/rails server, vite.config.*, framework conventions. Read README and CLAUDE.md for project specifics.
3. Start the dev server in the background; trap teardown so it dies even on failure.
4. Wait for the server to be ready (poll the URL with curl, do not sleep-and-hope).
5. Drive the browser with $DRIVER: open the URL, take a snapshot, identify elements by ref, interact, re-snapshot after every navigation.
6. Verify the asserted behavior — accessibility-tree comparison, screenshot, console messages, network calls.
7. Capture evidence to .agent-history/oracle/<card-id>/<timestamp>/.
8. Kill the browser session and the dev server.
9. Report PASS / FAIL / PARTIAL with evidence.

Server Lifecycle

SERVER_LOG="/tmp/qa-server-$$.log"
PORT=5173
npm run dev >"$SERVER_LOG" 2>&1 &
SERVER_PID=$!
trap '[ -n "$SERVER_PID" ] && kill "$SERVER_PID" 2>/dev/null; [ "$DRIVER" != mcp-chrome ] && "$DRIVER" close --all 2>/dev/null' EXIT

# Poll readiness *and* watch the process — exit if it dies before responding
ready=0
for i in $(seq 1 60); do
  if ! kill -0 "$SERVER_PID" 2>/dev/null; then
    echo "dev server died before becoming ready (see $SERVER_LOG)" >&2
    exit 1
  fi
  if curl -sf "http://localhost:$PORT" >/dev/null 2>&1; then
    ready=1
    break
  fi
  sleep 0.5
done
[ "$ready" = 1 ] || { echo "dev server failed to become ready in 30s (see $SERVER_LOG)" >&2; exit 1; }

Three operational details:

• Poll the URL, don't sleep-and-hope.
• Watch the process with kill -0 so a crashed boot fails fast, not against a corpse.
• Don't trust the configured port — Vite, Next.js, and others auto-bump (5173 → 5174, 3000 → 3001) when occupied. Parse $SERVER_LOG for the bound port if the spec is port-sensitive.

Replace npm run dev and 5173 with the project's command and port (Rails: bin/rails server -p 3000; Next.js: 3000; Vite: 5173).

Browser Workflow Pattern

Both drivers share this shape; references document exact syntax.

$DRIVER open <url>            # navigate
# wait for SPA hydration:
#   agent-browser: $DRIVER wait --load networkidle
#   playwright-cli: see references/playwright-cli/overview.md
$DRIVER snapshot              # accessibility tree + element refs (agent-browser @e1, playwright-cli e1)
$DRIVER click <ref>           # interact
$DRIVER snapshot              # RE-snapshot — refs invalidate
$DRIVER screenshot <path>     # evidence
$DRIVER close

Three subtleties:

• Refs invalidate on any DOM change, not only navigation. Modals, client rerenders, infinite scroll, optimistic-update revert all invalidate. Re-snapshot after any action that may mutate the DOM.
• networkidle does not fire for websocket / SSE / long-polling apps; the wait times out. Wait on a UI condition (element appears, class changes) instead.
• Assert against the accessibility tree, not the screenshot. A 1-pixel diff is not a defect; the a11y tree is text-based and stable.

Evidence Capture for the Oracle

Save artefacts under .agent-history/oracle/<card-id>/<timestamp>/:

• before.png — screenshot before the asserted action
• after.png — screenshot after
• snapshot-before.txt — accessibility tree before
• snapshot-after.txt — accessibility tree after
• console.log — console messages captured during the flow
• diff.txt — agent-browser diff snapshot after the action; for playwright-cli, diff the captured before/after snapshots

Rules

• Detect the driver first. Do not assume which CLI is installed. The detection block belongs at the top of the script, before any tool-specific commands.
• Always start the server before driving. A browser pointed at a non-listening port produces useless errors. Start, poll for ready, then drive.
• Always trap teardown. Servers and browser sessions leak processes. Use trap … EXIT so cleanup happens on success, failure, and interrupt.
• Re-snapshot after every navigation. Element refs invalidate. Stale refs produce confusing errors that look like bugs in the page.
• Capture evidence as you go, not at the end. A test that "looked right at the time" without saved screenshots and snapshots is not evidence. The transcript directory is the deliverable.
• Name browser sessions when running concurrently. agent-browser --session qa-<timestamp>; playwright-cli -s=qa-<timestamp>. Without a session name, parallel QAs collide.

Report Format

## Browser QA Report

**Scope**: <what was verified>
**Driver**: agent-browser | playwright-cli | chrome-browser MCP
**Verdict**: PASS | FAIL | PARTIAL

### Server
- [PASS/FAIL] `<dev server command>` — <port, ready time>

### Navigation
- [PASS/FAIL] `<url>` — <observed page state>

### Interactions
- [PASS/FAIL] <action> — <observed response>
  - Evidence: `<path-to-screenshot-or-snapshot>`

### Console
- <any console errors observed during the flow>

### Issues
1. <description with reproduction steps and evidence path>

### Transcript
Path: `.agent-history/oracle/<card-id>/<timestamp>/`
Contents: <brief listing of what's in there>