The shared daemon is reached over native HTTP, not a stdio bridge

Status: accepted

Hosts that support Streamable HTTP talk to the shared velesdb-memory daemon directly. The pinned mcp-remote stdio bridge remains only for hosts whose config accepts stdio alone, and those must be restarted after a long idle period.

Why. The server expires an MCP session after 60 minutes of inactivity, and SSE pings do not renew it. A request carrying an expired id is refused immediately — measured at 0.048 s — but mcp-remote@0.1.38 neither completes the call nor re-initialises, so the host waits for its own timeout instead. A fresh native session runs initialize, list and recall in 0.1–0.2 s.

Evidence. PR #1807, which shipped the native wiring and the session guards.