vLLM's Responses API was registering one MCP cleanup callback per session and re-closing already-closed sessions

GitHub·medium signalXBlueskyLinkedInCopy link

PR #56988 fixes a cleanup-ordering bug in vllm/entrypoints/openai/responses/context.py: with multiple MCP tool sessions, both Responses contexts registered the same cleanup callback once per session and each callback visited every called tool, sending duplicate cleanup requests to open sessions and calling into already-closed ones. The fix registers the callback once, after sessions are opened. Tests cover ParsableContext and HarmonyContext including unused tools, repeated initialization and partial initialization failure; the PR notes it does not address the separate cross-task cancel-scope error discussed in #25697.