Claude Code 2.1.285 MCP_DISCOVERY_CACHE and Sume tools_list

Claude Code 2.1.285, released September 29, can skip rediscovering a connector's tools after a cloud session restarts. The changelog says that MCP_DISCOVERY_CACHE=1, when set in a cloud environment's variables rather than a settings file, reuses connectors' tool lists across the restart (read 2026-10-04). For Sume's hosted MCP server at https://mcp.sume.com/mcp, that is a speed gain with one thing to watch: the cached list can lag the credential you hold now.

The changelog entry is short, so this post sticks to what follows from it and from Sume's own docs.

Why could a cached list be wrong for Sume?

What tools_list returns depends on the credential. Sume's tools and gates page says an OAuth token with only mcp:read sees read-only tools, while a token with mcp:write, or an API key, sees the full set. If you restart a cloud session with a different credential, a reused list can describe a surface the new credential does not have.

How do I check after a restart?

The read-only discovery tools in the quickstart are the ground truth.

• Call mcp_health first; it reports the auth source and safety posture the server sees now.
• Call tools_list and compare against what the agent believes it has.
• If a write call returns insufficient_scope, the credential lacks mcp:write whatever the cache says.
• Pass idempotency_key on every write and paid call, so a retry after confusion cannot charge twice.

Should I turn the cache on?

For long-running cloud jobs that always use the same credential, yes: it removes a round trip at every restart. For anything that changes scopes between runs, leave it off or rerun mcp_health as the first step of each session.