We don't publish
your competitive advantage.

IFWhen using multiple ClientSession instances with separate AsyncExitStack objects, closing the first session raises RuntimeError: 'Attempted to exit a cancel scope that isn't the current task's current cancel scope'.

THENEnsure that ClientSession instances are closed in the reverse order of their creation, or use a single AsyncExitStack for all sessions, or leverage an MCP session group that manages lifecycle correctly. This avoids a conflict where anyio's TaskGroup cancel scope binds to the current task and gets overridden by a subsequent session.

IFUsing ClientSessionGroup with async with causes 'Attempted to exit cancel scope' error during session lifecycle.

THENReplace ClientSessionGroup with a custom session manager like ClientConnectionManager that manages sessions without async with. Use separate connect, initialize, list_tools, call_tool, and disconnect methods to avoid context scope coupling.

IFMCP server deployed with multiple workers (gunicorn, Kubernetes) leads to SSE session disconnection and 'Could not find session for ID' errors after initial messages.

THENSet gunicorn worker count to 1 to maintain in-memory session state in a single process. For scalable deployments, implement distributed session storage (e.g., Redis) to share state across workers, or migrate to Streamable HTTP transport for stateless operation.

IFMCP server enforces an idle session timeout (e.g., 1 minute) that terminates the session, causing clients like Claude Desktop to receive HTTP 404 'Session not found' errors on subsequent requests.

THENConfigure the MCP server’s idle timeout to a longer duration (e.g., 10 minutes) via server options. For .NET MCP SDK, set the IdleTimeout property in McpServerOptions. Alternatively, implement client-side session keep-alive or automatic reconnection to prevent session expiry. Disabling idle tracking entirely may be appropriate for interactive use cases.

IFBrowser session IDs become stale over time, causing workflows to fail during recovery.

THENUpgrade to a version where stale browser session IDs are recovered cleanly without losing context.

IFCookies from the main Chrome profile need to be refreshed for the browser automation session.

THENDelete the local .chrome-profile directory to force a fresh cookie sync. Run `rm -rf .chrome-profile` in the project root. This clears any stale cookie data and allows the next automation to pull cookies from the main Chrome profile again.

IFSSE sessions lost when deploying MCP server with multiple workers (e.g., gunicorn multi-process) or across multiple Kubernetes pods due to load balancing.

THENSet worker count to 1 to keep session state in a single process. For production scalability, implement distributed session storage like Redis to share state across workers, or deploy the MCP service independently with a single worker per pod and use sticky sessions.

IFMCP server closes session after idle timeout (e.g., 1 min); client receives 404 with -32001 'Session not found' on subsequent requests.

THENIncrease server-side idle timeout via AddMcpServer options.IdleTimeout (e.g., 30 minutes) or apply client-side patch from PR #849 to reset ping timer on keepalive/SSE comments, preventing premature client-side disconnect. If client cannot be patched, implement auto-reconnect on -32001 errors.

IFWhen creating multiple MCP ClientSession instances in the same async task and closing them in creation order, the first close raises RuntimeError about cancel scope mismatch.

THENClose MCP ClientSession instances in reverse order of creation (LIFO). Alternatively, run each client session in a separate asyncio task to avoid cancel scope conflicts. This workaround mitigates the known anyio task group behavior where a cancel scope is bound per task and is replaced by subsequent scope entries.

IFUser page refreshes endlessly every ~0.3 seconds preventing access to the LiteLLM admin UI.

THENClear browser session cookies for the LiteLLM domain or use an incognito/private browsing window as an immediate workaround. As an alternative, use the API endpoint (e.g., GET /key/list) to view and manage keys instead of the UI while the bug persists. Avoid storing per-user HTTP session state in the database that can cause reload loops.

IFAttempting to delete a session that has tasks in pending, running, or waiting_for_input state leaves orphaned workspaces or fails silently.

THENUpgrade to a version that performs safe cascade cleanup: it removes the session store and task workspaces while preserving shared uploads, and returns a 409 Conflict for sessions with active tasks.

IFMCP server closes idle session after timeout (e.g., 1 minute), causing subsequent requests to fail with 'Session not found' (HTTP 404).

THENIncrease the idle timeout on the MCP server via configuration. For .NET, set `options.IdleTimeout = TimeSpan.FromMinutes(30)` in `AddMcpServer`. Alternatively, implement SSE keepalive pings to prevent premature closure.