…bility

Added comprehensive DI container:
- DIContainer class with lazy initialization for all components
- Protocol-based interfaces for better abstraction
- Singleton pattern for shared resources (cache, circuit breaker, etc.)
- Property-based access with setters for testing
- Global container management (get_container, set_container, reset_container)

Benefits:
- Easy mocking for unit tests (inject mock implementations)
- Clear dependency graph and reduced coupling
- Centralized lifecycle management (initialize/shutdown)
- Better testability with isolated test containers
- Backward compatible with existing code

Updated ErniPhotoAgency:
- Now accepts optional DIContainer in constructor
- Uses container for all component access
- Simplified initialization (delegates to container)
- Added shutdown() method for graceful cleanup

Updated FastAPI lifespan:
- Initialize global DI container on startup
- Proper shutdown sequence for all components
- Improved error messages for uninitialized state

Tests:
- 20 comprehensive unit tests for DI container
- Test lazy initialization, lifecycle, and mocking
- All tests passing

Resolves: openai#2 (HIGH priority)
Impact: Significantly improves code testability and maintainability

…#6)

This pull request adds a native sandbox runtime to
`openai-agents-python` by porting the relevant `universal_computer`
sandbox pieces into `src/agents/sandbox` and reshaping them around the
existing Agents SDK model instead of preserving the old UC agent stack.

The main change is a new `SandboxAgent` + `SandboxAgentRunner` flow that
fits the normal `Agent`/`Runner` paradigm while still supporting
manifests, capabilities, session state, snapshots, and sandbox-backed
tools. Along the way, this PR ports the sandbox sessions, manifests,
entries, utilities, and Docker/Modal/E2B/Unix backends, removes the
stale `universal_computer` import paths, and makes the sandbox code
Python 3.10 compatible.

This pull request also cleans up the port rather than carrying forward
legacy UC shapes wholesale. In particular, it drops the old
PTY/session-terminal surface and the legacy context-manager-oriented DX,
uses concrete sandbox types instead of a protocol-heavy type layer, and
keeps Docker as an optional extra rather than a core dependency. To make
the new flow easier to try, it adds a Docker example that streams agent
text and tool activity as it runs, plus focused sandbox tests covering
manifests, snapshots, session behavior, and the runner integration.

---------

---------

Co-authored-by: Kazuhiro Sera <seratch@openai.com>