Correctness: No child process death detection for MCP servers #80

New issue

Open

opened 2026-05-11 13:48:50 +00:00 by thabeta · 0 comments

thabeta commented

2026-05-11 13:48:50 +00:00

Owner

Severity: High

Location

crates/hero_aibroker_lib/src/mcp.rs — McpServer

Finding

MCP servers are spawned as child processes but there is no mechanism to detect if they crash or exit:

No SIGCHLD handler or child.try_wait() polling
McpManager has no health check loop for server liveness
Dead MCP servers remain in the tools_cache indefinitely
Tool discovery returns tools from dead servers

Impact

Tools from crashed MCP servers appear available but fail on invocation
No automatic restart of crashed MCP servers
Silent degradation of tool availability

Recommendation

Spawn a background task per MCP server that wait()s on the child
Implement periodic health checks (ping via JSON-RPC)
Auto-restart with backoff on unexpected exit
Remove tools from cache when server dies

## Severity: High ## Location `crates/hero_aibroker_lib/src/mcp.rs` — `McpServer` ## Finding MCP servers are spawned as child processes but there is no mechanism to detect if they crash or exit: - No `SIGCHLD` handler or `child.try_wait()` polling - `McpManager` has no health check loop for server liveness - Dead MCP servers remain in the `tools_cache` indefinitely - Tool discovery returns tools from dead servers ## Impact - Tools from crashed MCP servers appear available but fail on invocation - No automatic restart of crashed MCP servers - Silent degradation of tool availability ## Recommendation - Spawn a background task per MCP server that `wait()`s on the child - Implement periodic health checks (ping via JSON-RPC) - Auto-restart with backoff on unexpected exit - Remove tools from cache when server dies