fix: MCP CPU spike by adding timeout to session cleanup

[jwm4]

Summary

Fixes #4754

When making MCP calls through the responses API, the llama-stack server CPU usage could spike to 100% and remain there indefinitely, even after the request completes.

Root Cause

The issue occurs during MCP session cleanup in MCPSessionManager.close_all(). When tasks don't respond to cancellation, anyio's _deliver_cancellation loop can spin indefinitely, causing the CPU spike.

Solution

Added a configurable timeout (default 5 seconds) to the __aexit__ calls using anyio.fail_after(). If cleanup takes longer than the timeout, it's aborted to prevent the CPU spin.

Testing

• Verified that after the fix, CPU usage returns to idle levels after MCP requests complete
• Existing error handling catches the TimeoutError from fail_after() gracefully

[mattf]

please provide reproduction steps.

i did the following and still see 100% CPU usage -

10:53:24 in llama-stack on  fix/mcp-cpu-spike-timeout [$?] is 📦 0.4.0.dev0 …
➜ uv run llama stack run --providers agents=inline::meta-reference,inference=remote::llama-openai-compat,vector_io=inline::faiss,tool_runtime=inline::rag-runtime,files=inline::localfs
...
INFO     2026-01-28 10:53:34,588 uvicorn.error:216 uncategorized: Uvicorn running on http://['::', '0.0.0.0']:8321      
         (Press CTRL+C to quit)                                                                                         
INFO     2026-01-28 10:53:38,379 uvicorn.access:476 uncategorized: ::1:53190 - "POST /v1/responses HTTP/1.1" 200

10:53:35 in llama-stack on  fix/mcp-cpu-spike-timeout [$?] is 📦 0.4.0.dev0 …
➜ curl http://localhost:8321/v1/responses \
  -H "Content-Type: application/json" \
  -d '{
    "model": "llama-openai-compat/Llama-4-Scout-17B-16E-Instruct-FP8",
    "input": "Use the provided tool to say something.",
    "tools": [
      {
        "type": "mcp",
        "server_label": "local-mcp",
        "server_url": "http://localhost:9090"
      }
    ],
    "tool_choice": "auto"
  }'

[derekhiggins]

Also still seeing a problem
running https://github.com/derekhiggins/rhoai-auth-demo/blob/main/scripts/interactive-demo.py
python scripts/interactive-demo.py --user admin --tests mcp

proposed alternative change

[derekhiggins]

lgtm, CPU spike gone can when using MCP
thanks both.

[mergify]

This pull request has merge conflicts that must be resolved before it can be merged. @jwm4 please rebase it. https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork