The agent-facing MCP server (/mcp/agents) ran in stateful session mode,
keeping a StreamableHTTPServerTransport plus its full Server and Zod tool
schemas alive per agent connection, keyed by session id. The map was only
pruned on the SDK's onclose/onsessionclosed callbacks, which fire only on a
clean DELETE. Agents that are killed or exit without one never get reaped, so
dead sessions accumulate for the life of the daemon (observed ~77MB across 132
stale transports in a heap snapshot).
Listing and calling tools needs no cross-request state, so each request now
builds a fresh server + transport and tears both down when the response closes.
Per-request cost is ~0.6-0.9ms and ~760KB of short-lived allocation; nothing is
retained between requests. GET and DELETE return 405, which the MCP client
tolerates (it never receives a session id).