Observed live (show night, 2026-08-07, docs/mcp-findings.md session 4): after the MCP client sits idle for a while, the next tool call fails with -32001 (request timeout); an immediate retry succeeds. Reproduced several times during one session; not tied to a specific tool.
Suspects worth checking: the streamable-HTTP session going stale (keep-alive / SSE stream lifecycle in rmcp's LocalSessionManager), or the first call paying for something lazily re-established. A server-side keep-alive ping or a cheap session touch may fix it.
The co-DJ skill documents the workaround (retry once) for now.
Observed live (show night, 2026-08-07, docs/mcp-findings.md session 4): after the MCP client sits idle for a while, the next tool call fails with -32001 (request timeout); an immediate retry succeeds. Reproduced several times during one session; not tied to a specific tool.
Suspects worth checking: the streamable-HTTP session going stale (keep-alive / SSE stream lifecycle in rmcp's LocalSessionManager), or the first call paying for something lazily re-established. A server-side keep-alive ping or a cheap session touch may fix it.
The co-DJ skill documents the workaround (retry once) for now.