mcp-async-pipeline-must-return-run-id-not-block
When wrapping long-running pipelines (23+ stages) as MCP servers, the run tool must return a run_id immediately and not block. Status must be polled via a separate status tool and artifacts fetched via get_output. Blocking a single MCP tool call on a multi-stage pipeline causes tool timeouts and breaks ORACLE Hermes orchestration. This is the correct design pattern for any async MCP pipeline wrapper.