hermes - 💡(How to fix) Fix [Bug]: ERROR tools.mcp_tool: MCP tool playwright/browser_run_code call failed: Session terminated - From Gateway sessions [1 comments, 1 participants]

Bug Description

From a Hermes gateway client: Remote Playwright MCP calls results failed while same from Hermes Cli runs without issues.

Steps to Reproduce

From (Weixin) gateway client: Test remote playwright mcp using de.trip.com. Hermes agent responses:

• I'll test the MCP Playwright browser by navigating to de.trip.com and taking a screenshot to verify it's working.
• Let me get a snapshot of the page content:
• Session terminated. Let me reconnect and navigate again:
• The MCP Playwright server appears to be terminated/unavailable. Let me check if it can be restarted by looking at the server status:
• The first navigation actually worked (loaded de.trip.com successfully), but then the session terminated. Let me check the MCP configuration:
• I can see the MCP Playwright server is configured at http://mcp-playwright.mcp-playwright.svc.cluster.local:8931/mcp. The first navigation actually worked - the page loaded successfully:
• URL: https://de.trip.com/
• Title: "Buche günstige Flüge, Hotels, Mietwagen, Zugtickets und mehr | Trip.com"
• Console: 19 errors, 2 warnings However, the session terminated shortly after. etc etc...

Expected Behavior

Hermes Cli as well as Gateway client expect a successfully response back.

Actual Behavior

==> /opt/data/logs/gateway.log <== ERROR tools.mcp_tool: MCP tool playwright/browser_run_code call failed: Session terminated

Affected Component

Gateway (Telegram/Discord/Slack/WhatsApp)

Messaging Platform (if gateway-related)

No response

Debug Report

==> /opt/data/logs/gateway.log <==
ERROR tools.mcp_tool: MCP tool playwright/browser_run_code call failed: Session terminated

Operating System

Docker

Python Version

No response

Hermes Version

v0.12.0

Additional Logs / Traceback (optional)

Root Cause Analysis (optional)

Root Cause Analysis

This is a confirmed upstream issue with @playwright/mcp: microsoft/playwright-mcp#1045

The Problem

When using MCP Playwright via remote HTTP (like your setup), each tool call creates a new session, causing the browser state to be lost between calls:

Call 1: browser_navigate → works, creates page
Call 2: browser_snapshot  → "Session terminated" (new session, no page)

The MCP server uses session IDs to track browser contexts. When the client makes subsequent calls, the MCP server may assign a new session ID, creating a fresh browser context with no pages open.

Why CLI Works

In CLI mode:

1. You run hermes → starts Python process
2. First tool call → connects to MCP server, creates session A
3. Second tool call → reuses the same session A
4. State persists because the connection never drops

In Gateway mode:

1. Message arrives → Gateway spawns handler
2. Tool call → connects to MCP server, creates session A
3. Response sent → handler ends
4. Next message → new handler, new connection, session B created
5. Session A is orphaned/garbage-collected by Playwright MCP

Proposed Fix (optional)

Add "session terminated" to retry markers tools/mcp_tool.py line 1664:

_SESSION_EXPIRED_MARKERS: tuple = (
    "invalid or expired session",
    "expired session",
    "session expired",
    "session not found",
    "unknown session",
    "session terminated",  # ← ADD THIS
)

Are you willing to submit a PR for this?

• I'd like to fix this myself and submit a PR