feat(examples): add LangGraph integration with CubeSandbox via E2B SDK

[FAUST-BENCHOU]

part of #645

Adds a minimal example under examples/langgraph-integration showing how to run a LangGraph agent (via Deep Agents) against CubeSandbox using the E2B-compatible API.

• Uses e2b + langchain-e2b as the sandbox backend; tool execution runs inside CubeSandbox MicroVMs
• Includes a sandbox-only connectivity check and a full agent demo with an OpenAI-compatible LLM (e.g. DeepSeek)
• Adds Cube envd compatibility patches (root user, stdin) and follows the same SSL setup as openai-agents-example (CUBE_SSL_CERT_FILE → SSL_CERT_FILE)
• Isolates LLM HTTPS from sandbox SSL via certifi + trust_env=False.

Test in a cloud machine
OS/Kernel: OpenCloudOS 9.4, Linux kernel 6.6.69-1.1.cubesandbox.oc9.x86_64
CPU: AMD EPYC 9K65 (reported model), 2 vCPUs (1 socket, 2 cores, 1 thread/core), KVM virtualized

My env

OPENAI_API_KEY="sk-***"
OPENAI_BASE_URL="https://api.deepseek.com/v1"

E2B_API_URL="http://127.0.0.1:3000"
E2B_API_KEY="e2b_000000"
CUBE_API_URL="http://127.0.0.1:3000"
CUBE_TEMPLATE_ID="tpl-b99d79c3f5ca422a8b3d2787"
CUBE_SSL_CERT_FILE="/etc/pki/ca-trust/source/anchors/mkcert_development_CA_60105739546307673323659108215573903552.pem"

The result is

[root@VM-0-4-opencloudos langgraph-integration]# env .venv/bin/python e2b_demo.py --model deepseek-chat
[ssl] SSL_CERT_FILE=/etc/pki/ca-trust/source/anchors/mkcert_development_CA_60105739546307673323659108215573903552.pem (for E2B data-plane HTTPS)
Backend:  E2B SDK -> CubeAPI
Model:    deepseek-chat
CubeAPI:  http://127.0.0.1:3000
Question: What OS is running? Show uname and the first 3 lines of /etc/os-release.

[sandbox] sandbox_id=a5da2232c69d4866bfe9208ff50ad4ab
[ssl] LLM client: certifi CA bundle, trust_env=False
langchain-openai injected a custom httpx transport to apply `http_socket_options`, which disables httpx's proxy auto-detection (HTTP_PROXY, HTTPS_PROXY, ALL_PROXY, http_proxy, https_proxy, all_proxy set in environment). Set `LANGCHAIN_OPENAI_TCP_KEEPALIVE=0` or pass `http_socket_options=()` to restore default proxy behavior, or supply `openai_proxy` / your own `http_client` / `http_async_client` to take full control.
[agent] running LangGraph agent (Deep Agents harness) ...

============================================================
The sandbox is running **Debian GNU/Linux 12 (bookworm)** on an **x86_64** kernel:

- **uname**: `Linux 6.6.69-opencloudos9` (PVM guest, x86_64)
- **OS**: Debian 12 (bookworm)

[done] 3785 ms

[cleanup] destroying sandbox ...

[FAUST-BENCHOU]

/cc @fslongjin @chenhengqi ptal when u are free :)

[cubesandboxbot]

Review: PR #710 — LangGraph + CubeSandbox Example

Overall this is a well-structured example with clean module separation, proper cleanup (finally + sandbox.kill()), correct SSL isolation via certifi + trust_env=False, and thorough documentation. A few issues worth addressing:

Bugs & Runtime Issues

• demo_common.py:82 — messages[-1] on an empty list raises IndexError. The if messages guard is truthy for []. Use an explicit len() check.
• env_utils.py:32-35 — If CUBE_SSL_CERT_FILE is set but the file doesn't exist, the code silently does nothing. The TLS handshake then fails with a cryptic error. Print a warning when the path is missing.

Correctness

• llm_utils.py:17-18 — Splitting on : without validation silently corrupts model names with colons (e.g., ft:gpt-4o:org:2024-01-01 → 2024-01-01). Document the expected format or validate it.
• llm_utils.py:21 — httpx.Client is created but never closed. For a one-shot demo script this is harmless, but it's a connection-pool leak if reused. Consider wrapping with a context manager.

Design

• demo_common.py:32,59 — Both run_sandbox_check() and run_agent_demo() accept an unused sandbox parameter. Either remove it or add a # keep ref for GC comment and make the intent clear.
• e2b_demo.py:21 — apply_cube_envd_patches() is called at import time, which makes the module impossible to import for testing without triggering real SDK mutations. Guarding behind if __name__ == "__main__" would improve testability.

Security

• .env.example defaults to http:// for E2B_API_URL. An attacker on the same network segment can intercept the API key and sandbox traffic. Recommend defaulting to HTTPS and documenting the TLS setup.
• env_utils.py:31 — SSL_CERT_FILE is set from user-controlled input with only an os.path.isfile() guard. Consider validating the file is a parseable PEM certificate before trusting it.

Documentation

• Both READMEs' "CubeSandbox adaptations" table lists workdir and CUBE_SSL_CERT_FILE as patches in cube_patches.py, but they're actually in demo_common.py and env_utils.py respectively. Move those rows out of the patch table or add file annotations.

Positive Highlights

• Clean separation of concerns across 5 modules
• Idempotent patching with _PATCHED guard
• Proper finally-based sandbox cleanup
• Correct SSL isolation between sandbox and LLM traffic
• Well-documented version pinning in requirements.txt

[FAUST-BENCHOU]

@kinwin-ustc Could u please help to take a review? I have tested on my OpenCloudOS 9.4

[fslongjin]

Refs #645