Motivation and Context

In integrating latest llama-stack with MCP examples via OpenAI Responses API, I ran into a couple glitches which broke requests. I narrowed them down to GeneratorExit handling. When rebuilding llama-stack against this branch all scenarios pass.

I noticed this also fixes #1214 which was closed due to missing tests. This adds them.

The Problem

Scenario 1: GC cleanup leaks GeneratorExit

When garbage collector cleans up an orphaned generator while a background task has failed:

BaseExceptionGroup: unhandled errors in a TaskGroup (2 sub-exceptions)
├─ HTTPStatusError: Client error '401 Unauthorized'
└─ GeneratorExit  ← leaked from GC calling aclose()

Background:

• [BUG] Error during MCP session cleanup streetrace-ai/streetrace#87 - same bug

Scenario 2: User TaskGroup wraps GeneratorExit

When user code has a generator yielding inside a TaskGroup:

BaseExceptionGroup: unhandled errors in a TaskGroup (1 sub-exception)
└─ GeneratorExit  ← wrapped by user's TaskGroup during cleanup

Background:

• unusual behaviour in asyncio.TaskGroup in the presence of GeneratorExit, StopIteration, KeyboardInterrupt and SystemExit python/cpython#95571 - why GeneratorExit leaks from async generators
• GeneratorExit and asyncio.TaskGroup - missing from _is_base_error? python/cpython#135736 - why TaskGroup wraps GeneratorExit
• except* can break @trio.as_safe_channel cleanup python-trio/trio#3324 - why except* breaks generator cleanup

The Fix

Two exception handlers at the yield point handle both scenarios:

┌─────────────────────────────────────────────────────────────────┐
│                        yield streams                             │
└──────────────────────────────┬──────────────────────────────────┘
                               │
           ┌───────────────────┴───────────────────┐
           │                                       │
           ▼                                       ▼
┌─────────────────────────┐          ┌─────────────────────────────┐
│  Branch 1               │          │  Branch 2                    │
│  except GeneratorExit   │          │  except BaseExceptionGroup   │
│  → suppress             │          │  → split(GeneratorExit)      │
│                         │          │  → re-raise rest             │
└─────────────────────────┘          └─────────────────────────────┘
        │                                       │
        ▼                                       ▼
   Scenario 1:                         Scenario 2:
   GC calls aclose()                   User TaskGroup wraps
   on suspended generator              GeneratorExit

Changes

• Added GeneratorExit and BaseExceptionGroup handling in sse_client and streamable_http_client
• Added exceptiongroup>=1.0.0 as conditional dependency for Python 3.10 support
• Fixed stream cleanup: closes read_stream in finally block

How Has This Been Tested?

Parameterized tests in tests/client/test_resource_cleanup.py (each runs for both SSE and Streamable HTTP):

• test_generator_exit_on_gc_cleanup[sse/streamable] - Scenario 1: GC cleanup
• test_generator_exit_in_exception_group[sse/streamable] - Scenario 2: BaseExceptionGroup([GeneratorExit])
• test_generator_exit_mixed_group[sse/streamable] - Scenario 2: BaseExceptionGroup([GeneratorExit, ValueError])

Breaking Changes

None

Types of changes

• Bug fix (non-breaking change which fixes an issue)

Checklist

• I have read the MCP Documentation
• My code follows the repository's style guidelines
• New and existing tests pass locally
• I have added appropriate error handling