Tracing: the default BackendSpanExporter's HTTP client is never closed #4681

Area: src/agents/tracing/processors.py:533, src/agents/tracing/provider.py:506, src/agents/tracing/setup.py
Type: bug / resource leak
Severity: low-medium

Description

BackendSpanExporter opens a pooled HTTP client in its constructor and keeps it for the
process lifetime:

# Keep a client open for connection pooling across multiple export calls
self._client = httpx2.Client(timeout=httpx2.Timeout(timeout=60, connect=5.0))

It exposes a matching close():

def close(self):
    """Close the underlying HTTP client."""
    self._client.close()

but nothing in the shutdown chain ever calls it. The chain is:

atexit → _shutdown_global_trace_provider() → DefaultTraceProvider.shutdown() →
MultiTracingProcessor.shutdown() → BatchTraceProcessor.shutdown()

BatchTraceProcessor.shutdown() sets its shutdown event, optionally calls the exporter's
private _request_shutdown(), and joins the worker thread — it never calls
self._exporter.close(). Grepping the repo, the only callers of exporter.close() are in
tests/test_trace_processor.py.

Repro

from agents.tracing.processors import BackendSpanExporter, BatchTraceProcessor

exporter = BackendSpanExporter(api_key="test_key")
processor = BatchTraceProcessor(exporter=exporter)
processor.shutdown()
print("client closed after processor shutdown:", exporter._client.is_closed)  # False

Impact

• The TLS connection(s) to api.openai.com are left open until the interpreter tears the
  socket down, producing ResourceWarning: unclosed <ssl.SSLSocket ...> under
  -W error::ResourceWarning or PYTHONDEVMODE=1 (this breaks strict CI setups that treat
  ResourceWarnings as errors).
• Applications that replace the global trace provider / processor at runtime (multi-tenant
  servers switching export API keys, test suites that build a fresh
  BatchTraceProcessor(BackendSpanExporter(...)) per case) leak one client and its
  connection pool per replacement.

Suggested fix

Have BatchTraceProcessor.shutdown() call close() on its exporter when the exporter
exposes one, mirroring the existing _request_shutdown duck-typing:

close = getattr(self._exporter, "close", None)
if callable(close):
    close()

after the final flush, so the exporter still has its client while queued spans are drained.