fix(sandbox): close AioSandbox HTTP client during provider teardown (#2872)

[18062706139fcz]

Summary

Closes #2872.

AioSandbox.__init__ allocates a host-side agent_sandbox client (which wraps an httpx.Client via SyncClientWrapper). However, AioSandboxProvider.release/destroy/shutdown only popped provider-side state and tore down the backend container — the client/transport owned by each cached AioSandbox was never explicitly closed. In long-running services that repeatedly cycle sandboxes, this accumulates unreclaimed sockets/host-side resources.

Changes

• Add AioSandbox.close() in aio_sandbox.py
  • Best-effort, idempotent close of the wrapped httpx_client (falls back to top-level client.close() if not exposed).
  • Errors are logged at WARNING and swallowed so backend/container cleanup is never blocked.
• Update aio_sandbox_provider.py
  • release() and destroy() now look up the cached AioSandbox, drop it from _sandboxes under the lock, and call sandbox.close() outside the lock before parking in the warm pool / destroying the container.
  • shutdown() inherits this behaviour because it already routes active sandboxes through destroy().

Validation

Added regression tests:

• TestClose in tests/test_aio_sandbox.py — covers happy path (closes wrapped httpx_client), idempotency, swallowed exceptions, fallback to client.close(), and missing close attribute.
• 5 new tests in tests/test_aio_sandbox_provider.py — release()/destroy()/shutdown() close the cached client, and errors during close don't break provider lifecycle.

$ pytest backend/tests/test_aio_sandbox.py backend/tests/test_aio_sandbox_provider.py \
         backend/tests/test_aio_sandbox_local_backend.py backend/tests/test_aio_sandbox_readiness.py
======================== 62 passed, 1 warning in 0.77s =========================

[CLAassistant]

All committers have signed the CLA.

[Copilot]

Pull request overview

This PR addresses a sandbox resource-leak in long-running services by ensuring the host-side HTTP client created by each AioSandbox is explicitly closed during provider teardown, preventing unreclaimed sockets/transports from accumulating over repeated sandbox lifecycle churn.

Changes:

• Added an idempotent, best-effort AioSandbox.close() to close the underlying (wrapped) httpx client.
• Updated AioSandboxProvider.release() / destroy() to close cached AioSandbox instances after removing them from provider maps (and before/alongside backend teardown).
• Added regression tests covering the new close semantics and provider lifecycle behavior (including error swallowing).

Reviewed changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated 1 comment.

File	Description
backend/packages/harness/deerflow/community/aio_sandbox/aio_sandbox.py	Introduces AioSandbox.close() to release host-side HTTP client resources.
backend/packages/harness/deerflow/community/aio_sandbox/aio_sandbox_provider.py	Ensures provider lifecycle paths close cached sandbox clients on release/destroy.
backend/tests/test_aio_sandbox.py	Adds unit tests for AioSandbox.close() behavior (happy path, idempotency, fallbacks, exceptions).
backend/tests/test_aio_sandbox_provider.py	Adds tests asserting release/destroy/shutdown close cached clients and swallow close errors.

[WillemJiang]

@18062706139fcz , here are some comments the PR.

1. Add a comment explaining why the deep attribute chain is needed (Fern-generated client lacks close())
   The deep attribute chain is fragile. The close path traverses:
   self._client._client_wrapper → SyncClientWrapper → .httpx_client → HttpClient → .httpx_client → httpx.Client
   The Ideal solution: File an upstream issue/PR to add a proper close() method to agent_sandbox.Sandbox and calling the close() method from DeerFlow.

2. Set self._client = None after closing for use-after-close safety

3. Simplify getattr(self, "_client", None) to self._client

4. Consider removing the redundant try/except in the provider or documenting it as defense-in-depth

[18062706139fcz]

@18062706139fcz , here are some comments the PR.

1. Add a comment explaining why the deep attribute chain is needed (Fern-generated client lacks close())
   The deep attribute chain is fragile. The close path traverses:
   self._client._client_wrapper → SyncClientWrapper → .httpx_client → HttpClient → .httpx_client → httpx.Client
   The Ideal solution: File an upstream issue/PR to add a proper close() method to agent_sandbox.Sandbox and calling the close() method from DeerFlow.
2. Set self._client = None after closing for use-after-close safety
3. Simplify getattr(self, "_client", None) to self._client
4. Consider removing the redundant try/except in the provider or documenting it as defense-in-depth

Thanks for the review, @WillemJiang ! All four points are addressed in the latest commit (e1ba6041)

Could you please check again and see if the PR can be merged now? Thanks!