Bug: MCP TaskGroup Crashes Gateway on Single Server Failure

Severity: HIGH
File: mcp_tool.py:3116
Affected versions: v0.14.0

Problem

register_mcp_servers() uses asyncio.TaskGroup to initialize MCP servers in parallel. If any single MCP server fails to connect, the ExceptionGroup cancels all other servers and crashes the entire gateway process.

This is particularly impactful because MCP servers are third-party dependencies with their own Python environments — they frequently fail due to missing packages, version conflicts, or network issues.

Observed Impact

Three MCP servers (akshare, vibe-trading, tradingagents) fail on startup due to Python dependency issues. With TaskGroup, this crashes the gateway. The current try/except Exception in _discover_and_register_server partially mitigates, but TaskGroup's ExceptionGroup can escape if the inner catch doesn't cover BaseException.

The gateway ends up in a crash loop: systemd Restart=on-failure restarts it, MCP servers fail again, gateway crashes again. One profile was observed restarting 385 times/day.

Suggested Fix

Replace TaskGroup with asyncio.gather(return_exceptions=True) and widen inner catch from except Exception to except BaseException in _discover_and_register_server:

# Before (crashes gateway on any single MCP failure):
async with asyncio.TaskGroup() as tg:
    for server in servers:
        tg.create_task(_discover_and_register_server(server))

# After (isolates failures):
results = await asyncio.gather(
    *[_discover_and_register_server(s) for s in servers],
    return_exceptions=True
)
for r in results:
    if isinstance(r, BaseException):
        logger.warning(f"MCP server failed: {r}")

And in _discover_and_register_server, change:

except Exception as e:  # before
except BaseException as e:  # after

Environment

• Hermes Agent v0.14.0
• 10 profiles, 3 failing MCP servers
• Feishu/Weixin platforms