mcp 0.5.0 tool stderr goes nowhere - agent can't see tool crashes or warnings
Spent two days debugging why our file validation tool was silently failing in production. Tool was crashing on malformed paths and writing detailed error messages to stderr ("invalid utf-8 in filename", "path traversal detected", etc). Agent never saw any of it - just hung waiting for a response that never came. Tested with a minimal repro: tool that writes "BEFORE CRASH" to stdout, crashes with sys.exit(1), writes "detailed error info" to stderr. Agent sees the stdout, then hangs indefinitely. Stderr vanishes into the void. Is there any way to surface tool stderr to the agent in 0.5.0? Our monitoring shows 340+ silent tool failures in the last week that we only caught because we added external logging. The agent should at least see that the tool failed.
Tested on 0.5.0 with a basic file reader tool - stderr literally goes to /dev/null. Agent can't see tool crashes, just hangs forever. Had to add watchdog timer in deployment wrapper to kill after 60s.
can you post the exact mcp version and the tool code you used to test this? we're evaluating 0.5.0 for production and if stderr literally goes nowhere that's a complete blocker. need repro steps