plat 414

← Pulse platform issue index

PLAT-414 — Stop browser finds nothing; a crashed tab stays stuck; tall screenshots kill Chrome

Coordination	Value
State	fixed on main; proved on a candidate build running beside Excellence's live service; not deployed
Date	2026-10-04
Owner	browser
Related	PLAT-401 (launcher), PLAT-374

Found

• Stop browser. The browser panel (WorkflowLiveBrowser) has no stop button: its controls are Start browser, Take control / Give back, Start/Stop recording and Reconnect. The only stop affordance is the runtime-health control, which calls the workspace GET /api/browser/processes and POST /api/browser/cleanup. Both found browsers with ps aux | grep chromium and pkill -f chromium, so the managed Chrome (/opt/google/chrome/chrome) was never listed or stopped, while the list also showed other accounts' processes. The owner's error text "could not stop session ... start again" is not in the code; the chat Stop button (SessionStopButton, "Could not stop the session. Please try again.") stops a chat session, not a browser. Reproduced on Excellence with the matrix: the session's Chrome was not in the list and cleanup killed 0 of 1.
• Stuck tab. After chrome://crash, open timed out ("Operation timed out") until close plus open. The tool already recovered "CDP response channel closed" but not timeouts or crashed tabs in headless mode.
• Tall screenshot. screenshot --full of a 600000 px page killed Chrome ("CDP response channel closed").
• Side findings. A "Failed to save screenshot ... No such file or directory" error was treated as a dead session and the healthy browser was killed and relaunched. An empty <tmp>/.agent-browser/o/p<owner> folder and <session>.config/.target files were left per closed session.

Done

• workspace/handlers/browser_processes.go: list and cleanup match Chromium, Chrome and the headless shell, only for the service account; all kills the listed pids instead of pkill -f chromium.
• agent_go/pkg/browser/executor.go: headless timeout/crashed-tab recovery (close once, retry once, BROWSER_STUCK otherwise; no retry for side-effecting commands), SCREENSHOT_TOO_TALL above 16000 px, "Failed to save" is not a dead session, .config/.target removed with the session files.
• workspace/browserconfig.RemoveEmptySessionSocketDirs, called after a managed session is closed (executor and workspace close).
• Tests: executor_stuck_test.go, browser_close_test.go, browser_processes_chrome_test.go; live check executor_live_recovery_test.go (skipped without AW_LIVE_BROWSER_WORKSPACE_URL).
• Proof on Excellence (candidate workspace build on port 24901 beside the live service, deleted afterwards): the matrix passes 55/55 including three new stop rows (process list shows the session's Chrome, cleanup kills it and the session relaunches, close leaves no socket folder); the live executor test opens a page, opens chrome://crash, and the next open logs "Stuck session ... closing it and retrying once" and succeeds; a 600000 px full-page screenshot returns SCREENSHOT_TOO_TALL and the browser stays usable.

Left

• Deploy workspace and agent to Excellence, then run verify-browser-matrix.py against the live service.
• screenshot without a path writes under $HOME/.agent-browser/tmp/screenshots, which the sandbox does not grant; callers must pass a path (the tool brokers it through managed staging). A bare screenshot returns the save error.
• The matrix talks to /api/execute directly, so its crashed-tab row still needs close plus open; the recovery lives in the agent_browser tool and is covered by the unit and live executor tests.
• /tmp/.agent-browser is shared by every account on a host and holds old .config/.target files of closed sessions; only sessions closed after this fix are cleaned.