Test the real PostHog Electron app

Drive the actual running app over the Chrome DevTools Protocol with agent-browser. The dev app already launches with --remote-debugging-port=9222 (see apps/code/package.json), so an agent can connect, snapshot the UI, interact, and (only when asked) screenshot the live app.

This exercises real state: live tRPC, workspace-server, GitHub/Slack and whatever profile is signed into ~/.posthog-code. Pick the right surface:

Goal	Tool
Verify or screenshot a change in the real app, live data	this skill (agent-browser + CDP :9222)
Regression coverage in CI	Playwright E2E (apps/code/tests/e2e/)

Prerequisites

npm i -g agent-browser && agent-browser install   # once

The app must be running with remote debugging. pnpm dev (or pnpm dev:code) already passes --remote-debugging-port=9222. Preflight + connect:

pnpm app:cdp        # checks agent-browser + that the app is up on :9222, then connects

If it reports the app is not reachable, the app isn't running. Start it, then retry pnpm app:cdp.

Launching the app yourself (background / no TTY)

Do not use pnpm dev from a non-interactive shell. It builds deps then hands off to the phrocs TUI process-multiplexer, which aborts without a controlling terminal (bubbletea: could not open TTY: /dev/tty: device not configured). An agent in a background shell has no TTY, so pnpm dev cannot launch it headlessly.

Build the workspace deps (TTY-safe, the same step pnpm dev runs first), then launch only the Electron app with its stdin held open:

pnpm build:deps                        # turbo build of @posthog/code deps
tail -f /dev/null | pnpm dev:code      # run in the background; leave it running

pnpm dev:code is pnpm --filter code start, i.e. electron-vite dev --watch — just the app, no phrocs TUI; it watches and rebuilds main/preload and hot-reloads the renderer (fine for screenshotting/interacting). The CDP port (:9222) is opened by the app itself in dev (the remote-debugging-port switch in apps/code/src/main/bootstrap.ts), not by a CLI flag.

The tail -f /dev/null | prefix is a harmless guard and no longer strictly required. The old electron-forge start ran an interactive "type rs to restart" stdin reader that hit EOF in a no-stdin shell, treated it as quit, and tore the Electron window down before the CDP port ever opened. electron-vite dev has no such reader, so a backgrounded pnpm dev:code with no stdin stays up on its own; keeping the pipe does no harm.

Then wait for the port and connect (poll, don't sleep blindly):

until curl -s localhost:9222/json/version >/dev/null; do sleep 1; done
agent-browser connect 9222

After you launch it: leave it up, then idle-shutdown

Never auto-kill an instance you didn't start. If pnpm app:cdp found the app already up on :9222, it's the user's own — when you're done just agent-browser close your session and leave the app running.

When you launched it, don't tear it down the instant you finish. The app survives between turns, so leave it up and end your turn by telling the user it's still running and asking if they want anything else — a follow-up needs no relaunch.

So a forgotten app doesn't linger, arm a 10-minute idle watchdog at launch. Touch a marker file on launch and after every interaction; the watchdog tears the app down once that file sits untouched for 10 minutes (re-touching resets the clock):

touch /tmp/posthog-dev-lastuse        # arm now; re-run after each interaction

Then start the watchdog once, as a background task (touch the marker first or it fires immediately). It polls, then self-exits after it fires or once the marker is removed:

while sleep 30; do
  last=$(stat -f %m /tmp/posthog-dev-lastuse 2>/dev/null || echo 0)
  [ $(( $(date +%s) - last )) -ge 600 ] && break
done
pid=$(pgrep -f "remote-debugging-port=9222" | head -1)
[ -n "$pid" ] && kill -TERM "-$(ps -o pgid= -p "$pid" | tr -d ' ')" 2>/dev/null

To stop early (user says "done" / "shut it down"): close the session, group-kill the app, and drop the marker so the watchdog exits:

agent-browser close
pid=$(pgrep -f "remote-debugging-port=9222" | head -1)
[ -n "$pid" ] && kill -TERM "-$(ps -o pgid= -p "$pid" | tr -d ' ')" 2>/dev/null
rm -f /tmp/posthog-dev-lastuse

Group-killing the launcher (kill -TERM -PGID) takes down tail, pnpm, electron-vite and Electron together, so nothing — not even the tail -f /dev/null stdin pipe — lingers. Matching remote-debugging-port=9222 hits only your dev instance (prod has no debug port and a separate posthog-code-dev profile), so it never touches the user's app. Verify with curl -s localhost:9222/json/version (fails) and pgrep -fl posthog-code-dev (empty).

Load the canonical commands

agent-browser serves version-matched docs. Read them before driving:

agent-browser skills get electron     # Electron-over-CDP workflow (authoritative)
agent-browser skills get core         # snapshot/interact/screenshot reference

The loop

agent-browser connect 9222                      # attach (skip if you ran pnpm app:cdp)
agent-browser snapshot -i                       # interactive elements only (the app is already dark)
agent-browser click @e5                          # act on a ref from the snapshot
agent-browser snapshot -i                        # ALWAYS re-snapshot after the UI changes — this is how you verify
agent-browser close                              # done; free the session

Verifying a change is snapshot, not screenshot. The accessibility tree tells you what is on screen for almost no tokens, and it is how you confirm a test worked. Do not capture a screenshot to check your own work — only run screenshot when the user explicitly asks to see the app (see Screenshots).

Refs (@e1, @e2, …) are reassigned on every snapshot and go stale the moment the UI changes. Re-snapshot before the next ref interaction.

The renderer uses data-testid heavily, so prefer stable locators over refs when you know the target:

agent-browser find testid new-task-button click
agent-browser find role button click --name "New task"
agent-browser find text "Settings" click

Screenshots

Only screenshot when the user explicitly asks for one ("screenshot", "show me", "what does it look like"). To confirm a change worked, use agent-browser snapshot — the accessibility tree is the cheap default and is almost always enough. Auto-capturing a screenshot to "confirm" your work just burns image tokens.

agent-browser screenshot /tmp/app.png                  # viewport (absolute path = clickable)
agent-browser screenshot --full /tmp/app.png           # full page instead of viewport

Navigate to the target view first (click through the UI), then capture. agent-browser prints the saved path. Repeated captures reuse the connected session, so batches are fast.

Always let the user open the capture

When the user asked for a screenshot, they want to look at it, so close the loop every time:

1. Save to and report an absolute path (/tmp/app.png, never a bare out.png). Claude Code renders absolute paths as clickable links, so the path itself opens the PNG on click.
2. Offer to open it for them, and on a yes run open /tmp/app.png (macOS opens it in Preview). If the request was clearly "screenshot it so I can see it", just open it instead of asking.
3. The user can also open it themselves at any time with !open /tmp/app.png.

Repo specifics

• Port: 9222 (override with POSTHOG_CODE_CDP_PORT). Collides with Chrome's default debugging port; if connect attaches to the wrong target, list and pick the PostHog window: agent-browser tab then agent-browser tab --url "*".
• Multiple targets: the app has a main renderer window (page title contains "PostHog") plus possible webviews/devtools. agent-browser tab lists them; switch with agent-browser tab <index>.
• Never pass --color-scheme dark: that global flag makes agent-browser apply device emulation that forces a 1280x720 viewport and renders this Electron window blank, and it sticks in the daemon (only restarting the agent-browser daemon clears it, not close). The app is already dark, so plain agent-browser snapshot / agent-browser screenshot is what you want.
• Auth / data: you drive whatever is signed into ~/.posthog-code. If the app shows onboarding or sign-in, that is the real boot state. Do not mutate production data (don't create real tasks/PRs) while exploring.
• Boot timing: after launching the app, give it a few seconds before connecting; the renderer settles after #root > * appears and "Loading" clears.

Running alongside prod

PostHog orchestrates the agent, so the usual loop is: prod (the installed app) runs the agent, and the dev build (pnpm dev) is the system under test. They coexist by design (apps/code/src/main/bootstrap.ts): dev runs as posthog-code-dev with its own app name, userData and single-instance lock, so it never collides with prod.

• agent-browser always targets dev. Only the dev build exposes CDP on :9222; prod has no debug port, so connect 9222 can't accidentally drive prod.
• Separate auth/state. The dev instance has its own posthog-code-dev profile; it is not signed in just because prod is. Sign into the dev window once; its state persists.
• One dev instance only. Dev's single-instance lock, fixed dev callback port (8238) and :9222 mean a second pnpm dev collides and quits. Run prod + one dev.
• What reloads. Renderer/UI changes hot-reload; just re-snapshot. Main-process/Electron changes need a dev restart to take effect.

Troubleshooting

• Connection refused on :9222: the app isn't running with the debug flag. Start it (see Launching the app yourself for the headless recipe). Verify the port: lsof -i :9222 or curl -s localhost:9222/json/version.
• App launches then immediately exits; CDP never opens: the old electron-forge "quit on stdin EOF" behavior is gone under electron-vite, so the tail -f /dev/null | prefix is optional and not the fix here. This is now usually a build error, the single-instance lock (another dev instance running) or a crash — check the pnpm dev:code output. Note pnpm dev cannot run headlessly at all — its phrocs TUI needs a TTY; use pnpm dev:code.
• Snapshot is empty / wrong window: you're on the wrong target. Run agent-browser tab and switch to the "PostHog" page.
• Can't type into an input: try agent-browser keyboard type "text" (types at current focus) or agent-browser keyboard inserttext "text" to bypass key events.