fix: add configurable timeout_keep_alive to prevent ReadError on keep-alive connection reuse

[nihongye]

Problem

The SandboxManager client uses a long-lived httpx.AsyncClient with connection pooling to access the sandbox manager server.

uvicorn's default timeout_keep_alive is 5 seconds, while httpx's default keepalive_expiry is also 5 seconds. This creates a race condition:

1. Client sends request, server responds, connection goes idle
2. After ~5s, server sends FIN to close the idle connection
3. Client (unaware of FIN) tries to reuse the connection for a new request
4. Server responds with RST since the connection is closing
5. Client receives httpx.ReadError instead of a response

Timing pattern observed:
T+0.000s: Request/Response completes, connection idle T+5.000s: Server sends FIN (timeout_keep_alive=5s default) T+5.001s: Client sends new request on same connection T+5.002s: Server sends RST → httpx.ReadError

Solution

Set timeout_keep_alive to 120 seconds (configurable) so the server keeps idle connections alive longer than clients expect. This ensures the server never closes a connection that the client might still try to reuse.

Key principle: Server timeout_keep_alive > Client keepalive_expiry

Changes

• config.py: Add TIMEOUT_KEEP_ALIVE: int = 120 to Settings (configurable via env var)
• sandbox manager server app.py: Pass timeout_keep_alive=settings.TIMEOUT_KEEP_ALIVE to uvicorn.run()
• agent_app.py: Add timeout_keep_alive=120 parameter to AgentApp.run() for clients using connection pools

[cla-assistant]

All committers have signed the CLA.