René's URL Explorer Experiment

{"@context":"https://schema.org","@type":"DiscussionForumPosting","headline":"BrokenResourceError race condition in stdio_client cleanup when context exits quickly","articleBody":"# GitHub Issue: MCP SDK stdio_client Race Condition\n\n## Repository\n`modelcontextprotocol/python-sdk`\n\n## Title\n`BrokenResourceError` race condition in `stdio_client` cleanup when context exits quickly\n\n## Body\n\n### Description\n\nWhen the `stdio_client` async context manager exits quickly (before the subprocess has finished outputting data), a race condition occurs between the `stdout_reader` task and the cleanup code in the `finally` block, resulting in `BrokenResourceError`.\n\n### Reproduction\n\nThis occurs in scenarios where:\n1. A subprocess is spawned via `stdio_client`\n2. The subprocess starts outputting data (e.g., initialization messages)\n3. The calling code exits the context quickly (e.g., user disconnects)\n4. The `finally` block closes `read_stream_writer` while `stdout_reader` is mid-send\n\n### Error\n\n```\nExceptionGroup: unhandled errors in a TaskGroup (1 sub-exception)\n  +-+---------------- 1 ----------------\n    | Traceback (most recent call last):\n    |   File \"mcp/client/stdio/__init__.py\", line 162, in stdout_reader\n    |     await read_stream_writer.send(session_message)\n    |   File \"anyio/streams/memory.py\", line 242, in send\n    |     self.send_nowait(item)\n    |   File \"anyio/streams/memory.py\", line 213, in send_nowait\n    |     raise BrokenResourceError\n    | anyio.BrokenResourceError\n    +------------------------------------\n```\n\n### Root Cause\n\nIn `mcp/client/stdio/__init__.py`, the `stdio_client` function:\n\n```python\nasync with (\n    anyio.create_task_group() as tg,\n    process,\n):\n    tg.start_soon(stdout_reader)\n    tg.start_soon(stdin_writer)\n    try:\n        yield read_stream, write_stream\n    finally:\n        # ... process cleanup ...\n        await read_stream_writer.aclose()  # \u003c-- Closes stream\n        await write_stream_reader.aclose()\n```\n\nThe `finally` block closes `read_stream_writer` before the TaskGroup has cancelled its tasks. If `stdout_reader` is in the middle of `await read_stream_writer.send(...)`, it receives `BrokenResourceError`.\n\n### Suggested Fix\n\nCancel the TaskGroup's scope before closing the streams:\n\n```python\nfinally:\n    # Cancel background tasks before closing streams\n    tg.cancel_scope.cancel()\n    \n    # ... existing cleanup ...\n    await read_stream_writer.aclose()\n    await write_stream_reader.aclose()\n```\n\nAlternatively, wrap the stream operations in `stdout_reader` with `BrokenResourceError` handling:\n\n```python\nasync def stdout_reader():\n    try:\n        async with read_stream_writer:\n            # ... existing code ...\n    except anyio.BrokenResourceError:\n        # Context is closing, exit gracefully\n        pass\n```\n\n### Environment\n\n- Python: 3.12.1\n- mcp: (version from pip)\n- anyio: (version from pip)\n- OS: Linux\n\n### Workaround\n\nWe're currently serializing MCP lifecycle operations with an asyncio.Lock to prevent overlapping enter/exit operations, which avoids triggering the race condition.\n","author":{"url":"https://github.com/newsbubbles","@type":"Person","name":"newsbubbles"},"datePublished":"2026-01-26T14:36:56.000Z","interactionStatistic":{"@type":"InteractionCounter","interactionType":"https://schema.org/CommentAction","userInteractionCount":4},"url":"https://github.com/1960/python-sdk/issues/1960"}