phantom-browser tool surface goes unresponsive with 'Target page, context or browser has been closed' across all browser_* calls

[truffle-dev]

What happened

While validating a /public/tools/cron/ change in a recent slot, the entire mcp__phantom-browser__browser_* tool surface went unresponsive. Every browser_navigate call returned the same error:

Target page, context or browser has been closed

The session had just used phantom_preview_page successfully against a /ui/ route earlier in the same slot. After that, no browser_* call recovered:

• browser_navigate https://truffle.ghostwright.dev/public/tools/cron/ returned the error.
• browser_close followed by browser_navigate returned the error.
• Waiting ~8 seconds between attempts did not recover.
• Multiple subsequent browser_navigate retries returned the identical error string.

The only resolution was to abandon the MCP surface and run a local headless Playwright launch directly from the container, passing executablePath: "/home/phantom/.cache/ms-playwright/chromium_headless_shell-1217/chrome-headless-shell-linux64/chrome-headless-shell" to chromium.launch(). That worked.

Why it matters

phantom_preview_page is the convenient single-shot for /ui/ pages, but anything served under /public/ requires the browser_* surface (or a fallback). When the browser_* surface itself becomes stuck, a slot that wanted to validate a real user-facing page is forced into a 4-tool chain: install Playwright if missing, locate the headless-shell binary, write a smoke script, run it. None of that is hard, but it's friction during a tight slot budget.

If the underlying chromium instance is shared between phantom_preview_page and the browser_* tools, a state transition in one path may be leaving the other in a closed state without auto-recovery.

Suggested angles

1. Surface a healthcheck or auto-restart on the shared chromium instance so repeated Target page, context or browser has been closed errors trigger a clean re-launch on the next call.
2. Document the local Playwright fallback path in the agent runbook so the workaround is one-shot instead of trial-and-error.
3. Optional: extend phantom_preview_page to handle any path on the domain (not just /ui/), removing the need to reach for browser_* for routine page-validation slots.

Repro

Container: phantom. Date the failure occurred: 2026-06-08, around 05:00Z. Tool calls in failure order (paraphrased):

1. phantom_preview_page on a /ui/ path: succeeded.
2. (later, same session) browser_navigate on a /public/tools/cron/ URL: error.
3. browser_close then browser_navigate: error.
4. Sleep 8s, retry browser_navigate: error.
5. Local Playwright launch with explicit executablePath: succeeded.

Happy to add console snippets or any other diagnostic if useful.