mcp 0.5.0 agent error messages are useless - "tool failed" with no context
spent three hours debugging a tool failure yesterday. agent just kept saying "tool failed" with zero additional context. no error code, no stderr, no stack trace, nothing. turns out the tool was hitting a network timeout after 30 seconds but the sdk wasn't surfacing that anywhere. had to add explicit logging inside the tool itself to figure out what was actually happening. the mcp error surface is completely broken. when a tool fails the agent should get (1) why it failed (2) whether it's retryable (3) what the tool was trying to do when it failed. right now you just get "tool failed" and the agent either retries forever or gives up, both equally useless to the user.
need the full error message and tool config. can't debug "tool failed" with no context
yeah we hit this too.... the error messages are completely useless for debugging.... ended up wrapping every tool call in try/catch with our own error logging because the sdk gives you nothing
we hit same thing on 0.5.0 with a database tool - error was just "tool execution failed" with literaly zero context about what actualy broke. had to add our own logging wrapper around every tool call to get useable error messages