[BUG] 90%+ idle CPU after first MCP tool call — anyio cancel scope busy-loop in streamable HTTP client #1226

Description

After the first request that triggers an MCP tool call, the lightspeed-stack process burns 90%+ CPU indefinitely while fully idle (no incoming requests). Restarting the container resolves it until the next MCP tool call.

Environment

• lightspeed-stack image: quay.io/lightspeed-core/lightspeed-stack:dev-20260225-f6aa46b
• llama-stack: 0.4.3
• MCP SDK (mcp): 1.26.0
• anyio: 4.12.1
• Python: 3.12.12
• Config: library mode, 1 worker, MCP server via streamable HTTP transport

Reproduction

1. Start lightspeed-stack with an MCP server configured (streamable HTTP transport)
2. Send any request that triggers MCP tool use (e.g. POST /v1/infer)
3. Wait for the response to complete
4. Observe CPU usage — the process stays at 90%+ with no further requests

Root Cause

py-spy profiling shows the main thread stuck in a tight loop:

Thread 1 (active+gil): "MainThread"
    _call_soon (asyncio/base_events.py:819)
    call_soon (asyncio/base_events.py:803)
    _deliver_cancellation (anyio/_backends/_asyncio.py:594)
    _run (asyncio/events.py:88)
    _run_once (asyncio/base_events.py:1999)
    run_forever (asyncio/base_events.py:645)

After an MCP tool call completes, the MCP streamable HTTP client's cleanup calls tg.cancel_scope.cancel(). Tasks in the scope are stuck in _must_cancel state (they're trying to write to already-closed streams), causing AnyIO's _deliver_cancellation() to re-schedule itself via call_soon() on every event loop tick without ever converging.

This is an upstream issue in the MCP Python SDK: modelcontextprotocol/python-sdk#1805. I've added confirming profiling evidence there: modelcontextprotocol/python-sdk#1805 (comment)

Evidence

podman stats with no active requests:

NAME              CPU %
lightspeed-stack  59.30%

podman top lightspeed-stack:

PID  PPID  %CPU  COMMAND
  1     0  90.1  python3.12 src/lightspeed_stack.py

3 consecutive py-spy samples (2/3 caught in _deliver_cancellation):

Sample 1: _deliver_cancellation (line 594) → call_soon
Sample 2: _deliver_cancellation (line 578)
Sample 3: select (momentarily idle between bursts)

Workaround

Restarting the lightspeed-stack container resolves the issue until the next MCP tool call.