viant/endly/endly-go-diagnostics

Start the MCP process with endly mcp -diagnostics to enable gops. Identify that exact Endly PID with gops/process inspection; do not signal unrelated processes. Use gops stack <pid> to save goroutine stacks, then inspect the operation's executeWorkflow/executeAction stack and the goroutine it is waiting for. If the state has not progressed, capture a second stack after independent work to distinguish a stable block from normal I/O. gops stats and gops memstats can distinguish resource pressure from a mutex/channel wait. Avoid CPU/heap profiling unless needed.

Correlate the frames with MCP operation events and paused state. A dsunit.prepare channel send can mean its dataset worker is waiting on database or BigQuery I/O; inspect the worker before calling it a deadlock. A pending step/next command can mean the action is still executing: poll debug status and pausedAt before issuing continue. Never treat an RPC timeout as the operation's terminal status.

Fix the smallest demonstrated cause and add a regression test for observable behavior. Preserve the asynchronous operation/session model and external test fixtures. Use cancellation through MCP first, and ensure shutdown uses a bounded context. Retain concise stack findings and code locations without dumping secrets or large raw logs into the conversation.