fix(server): prevent Model Context Protocol response timeouts

What changed

Pinned the emcp dependency to the upstream commit that keeps request responses on the original transport request, and added a controller regression test for sessions with stale event-stream registrations.

Why

Tuist’s hosted Model Context Protocol endpoint could intermittently leave agent clients waiting for a timeout even after the server handled the tool call. The symptom was a slow first call followed by a fast retry.

Root cause

The released emcp package could route a request response to a stored Server-Sent Events process identifier. If that identifier pointed to a closed process, the original request received an empty accepted response and the client kept waiting until its own timeout fired.

Approach

The dependency is temporarily pinned to the upstream fix commit until a released package includes the change. The upstream fix is tracked in PJUllrich/emcp#5. The new regression test initializes a session, registers a stale event-stream process, and verifies that a normal tools/list request still receives the response inline.

Impact

Agent clients should no longer see intermittent tool-call timeouts caused by stale event-stream registrations. The dependency pin should be reverted back to a released package once the upstream fix is published.

Validation

• mix format mix.exs test/tuist_web/controllers/mcp_controller_test.exs
• mix test test/tuist_web/controllers/mcp_controller_test.exs passed with 9 tests and 0 failures.

No GitHub comments yet.