Feat/docker sandbox by floatlibai · Pull Request #394 · vllm-project/vime

floatlibai · 2026-08-20T06:30:28Z

No description provided.

Signed-off-by: flb_ <floatlibai@gmail.com>

gemini-code-assist

Code Review

This pull request introduces a remote Docker sandbox (DockerSandbox) as a drop-in replacement for E2BSandbox, updates the agent sandbox logic to support it, and adds an end-to-end training shell script. The review feedback highlights several critical issues in the new sandbox implementation, including potential SSH connection leaks during startup and teardown, a bug in parsing the DOCKER_SANDBOX boolean environment variable, and API mismatches with the original sandbox interface (such as default argument values and exception handling). Additionally, the reviewer recommended making hardcoded IP addresses in the launch script configurable.

gemini-code-assist · 2026-08-20T06:32:24Z

+    async def __aenter__(self) -> "DockerSandbox":
+        loop = asyncio.get_event_loop()
+        docker_tarball_dir = os.environ.get("DOCKER_TARBALL_DIR")
+        def _start():
+            self._client = _make_client()  # 每个沙箱独立连接
+            vime_head_host = os.environ.get("VIME_HEAD_HOST", "")
+            no_proxy=os.environ.get("no_proxy", f"127.0.0.1,localhost,{vime_head_host}")
+            NO_PROXY=os.environ.get("NO_PROXY", f"127.0.0.1,localhost,{vime_head_host}")
+            http_proxy=os.environ.get("http_proxy", "")
+            https_proxy=os.environ.get("https_proxy", "")
+            return self._client.containers.run(
+                self.image,
+                command="sleep infinity",
+                detach=True,
+                mem_limit=_MEM_LIMIT,
+                nano_cpus=int(_CPUS * 1e9),
+                network_mode="bridge",
+                cap_add=["SYS_PTRACE"],
+                remove=False,
+                volumes={
+                    docker_tarball_dir: {
+                        "bind": docker_tarball_dir, 
+                        "mode": "rw"
+                    }
+                },
+                environment={
+                    "http_proxy": http_proxy,
+                    "https_proxy": https_proxy,
+                    "no_proxy": no_proxy,
+                    "NO_PROXY": NO_PROXY,
+                },
+            )
+
+        self._container = await loop.run_in_executor(None, _start)
+        await self.exec("mkdir -p /workspace", user="root", check=False)
+        return self


If containers.run raises an exception during sandbox startup, the newly created DockerClient (which establishes an SSH connection) is never closed, leading to a resource/SSH connection leak. Additionally, if DOCKER_TARBALL_DIR is not set in the environment, passing None as a key in the volumes dictionary will cause a TypeError or AttributeError in docker-py.

We should wrap the container startup in a try...except block to ensure the client is closed on failure, and safely handle the volume mounting only when docker_tarball_dir is set.

async def __aenter__(self) -> "DockerSandbox": loop = asyncio.get_event_loop() docker_tarball_dir = os.environ.get("DOCKER_TARBALL_DIR") def _start(): self._client = _make_client() # 每个沙箱独立连接 try: vime_head_host = os.environ.get("VIME_HEAD_HOST", "") no_proxy = os.environ.get("no_proxy", f"127.0.0.1,localhost,{vime_head_host}") NO_PROXY = os.environ.get("NO_PROXY", f"127.0.0.1,localhost,{vime_head_host}") http_proxy = os.environ.get("http_proxy", "") https_proxy = os.environ.get("https_proxy", "") volumes = {} if docker_tarball_dir: volumes[docker_tarball_dir] = { "bind": docker_tarball_dir, "mode": "rw" } return self._client.containers.run( self.image, command="sleep infinity", detach=True, mem_limit=_MEM_LIMIT, nano_cpus=int(_CPUS * 1e9), network_mode="bridge", cap_add=["SYS_PTRACE"], remove=False, volumes=volumes, environment={ "http_proxy": http_proxy, "https_proxy": https_proxy, "no_proxy": no_proxy, "NO_PROXY": NO_PROXY, }, ) except Exception: self._client.close() self._client = None raise self._container = await loop.run_in_executor(None, _start) await self.exec("mkdir -p /workspace", user="root", check=False) return self

gemini-code-assist · 2026-08-20T06:32:24Z

+    async def __aexit__(self, *args):
+        if self._container is None:
+            return
+        loop = asyncio.get_event_loop()
+        container = self._container
+        client = self._client
+
+        def _stop():
+            try:
+                container.remove(force=True)
+            except Exception:
+                pass
+            try:
+                client.close()
+            except Exception:
+                pass
+
+        await loop.run_in_executor(None, _stop)
+        self._container = None
+        self._client = None


If self._container is None (e.g., if container creation failed or was partially initialized), __aexit__ currently returns immediately without closing self._client. This leaks the SSH connection. We should ensure self._client is always closed if it exists, regardless of whether the container was successfully created.

Suggested change

async def __aexit__(self, *args):

if self._container is None:

return

loop = asyncio.get_event_loop()

container = self._container

client = self._client

def _stop():

try:

container.remove(force=True)

except Exception:

pass

try:

client.close()

except Exception:

pass

await loop.run_in_executor(None, _stop)

self._container = None

self._client = None

async def __aexit__(self, *args):

loop = asyncio.get_event_loop()

container = self._container

client = self._client

def _stop():

if container is not None:

try:

container.remove(force=True)

except Exception:

pass

if client is not None:

try:

client.close()

except Exception:

pass

await loop.run_in_executor(None, _stop)

self._container = None

self._client = None

gemini-code-assist · 2026-08-20T06:32:24Z

+    async def exec(
+        self,
+        cmd: str,
+        *,
+        user: str = "root",
+        check: bool = True,
+        timeout: int = _DEFAULT_TIMEOUT,
+        env: dict[str, str] | None = None,
+    ) -> tuple[int, str, str]:


In the Sandbox protocol and the E2BSandbox implementation, check defaults to False. However, in DockerSandbox.exec, it currently defaults to True. This discrepancy breaks the "drop-in replacement" contract and can cause unexpected crashes when switching between sandbox backends. Let's change the default to False to match the protocol.

Suggested change

async def exec(

self,

cmd: str,

*,

user: str = "root",

check: bool = True,

timeout: int = _DEFAULT_TIMEOUT,

env: dict[str, str] | None = None,

) -> tuple[int, str, str]:

async def exec(

self,

cmd: str,

*,

user: str = "root",

check: bool = False,

timeout: int = _DEFAULT_TIMEOUT,

env: dict[str, str] | None = None,

) -> tuple[int, str, str]:

gemini-code-assist · 2026-08-20T06:32:24Z

+    async def read_file(self, sandbox_path: str, user: str = "root") -> str:
+        loop = asyncio.get_event_loop()
+
+        def _get() -> str:
+            bits, _ = self._container.get_archive(sandbox_path)
+            buf = io.BytesIO(b"".join(bits))
+            with tarfile.open(fileobj=buf) as tf:
+                member = tf.getmembers()[0]
+                f = tf.extractfile(member)
+                return f.read().decode("utf-8", errors="replace") if f else ""
+
+        return await loop.run_in_executor(None, _get)


In E2BSandbox.read_file, any exceptions (such as file not found) are caught and an empty string "" is returned. In DockerSandbox.read_file, if the file does not exist, get_archive will raise a docker.errors.NotFound exception, which propagates and crashes the caller. To ensure it acts as a true drop-in replacement, we should catch exceptions and return "".

Suggested change

async def read_file(self, sandbox_path: str, user: str = "root") -> str:

loop = asyncio.get_event_loop()

def _get() -> str:

bits, _ = self._container.get_archive(sandbox_path)

buf = io.BytesIO(b"".join(bits))

with tarfile.open(fileobj=buf) as tf:

member = tf.getmembers()[0]

f = tf.extractfile(member)

return f.read().decode("utf-8", errors="replace") if f else ""

return await loop.run_in_executor(None, _get)

async def read_file(self, sandbox_path: str, user: str = "root") -> str:

loop = asyncio.get_event_loop()

def _get() -> str:

try:

bits, _ = self._container.get_archive(sandbox_path)

buf = io.BytesIO(b"".join(bits))

with tarfile.open(fileobj=buf) as tf:

members = tf.getmembers()

if not members:

return ""

f = tf.extractfile(members[0])

return f.read().decode("utf-8", errors="replace") if f else ""

except Exception:

return ""

return await loop.run_in_executor(None, _get)

gemini-code-assist · 2026-08-20T06:32:24Z

+if os.environ.get("DOCKER_SANDBOX", True):
+    from .docker_sandbox import DockerSandbox as E2BSandbox, Sandbox
+else:
+    from vime.agent.sandbox import E2BSandbox, Sandbox


Using os.environ.get("DOCKER_SANDBOX", True) will return a truthy string like "0" or "False" if the environment variable is set to disable the docker sandbox. In Python, any non-empty string is truthy, meaning the else block will never be reached. We should parse the environment variable as a boolean properly.

Suggested change

if os.environ.get("DOCKER_SANDBOX", True):

from .docker_sandbox import DockerSandbox as E2BSandbox, Sandbox

else:

from vime.agent.sandbox import E2BSandbox, Sandbox

_use_docker = os.environ.get("DOCKER_SANDBOX", "1").lower() not in ("0", "false", "no")

if _use_docker:

from .docker_sandbox import DockerSandbox as E2BSandbox, Sandbox

else:

from vime.agent.sandbox import E2BSandbox, Sandbox

gemini-code-assist · 2026-08-20T06:32:24Z

+export MASTER_ADDR="192.168.13.190"
+export VIME_HEAD_HOST="${MASTER_ADDR}"


The IP address 192.168.13.190 is hardcoded. It is better to allow overriding it via environment variables with this IP as a default fallback, making the script more portable across different cluster setups.

Suggested change

export MASTER_ADDR="192.168.13.190"

export VIME_HEAD_HOST="${MASTER_ADDR}"

export MASTER_ADDR="${MASTER_ADDR:-192.168.13.190}"

export VIME_HEAD_HOST="${MASTER_ADDR}"

gemini-code-assist · 2026-08-20T06:32:24Z

+export SWE_SANDBOX_METADATA_FILE="${SANDBOX_METADATA_FILE}"
+
+export DOCKER_SANDBOX=1
+export DOCKER_SANDBOX_HOST="root@192.168.13.188"


The IP address 192.168.13.188 is hardcoded. It is better to allow overriding it via environment variables with this IP as a default fallback.

Suggested change

export DOCKER_SANDBOX_HOST="root@192.168.13.188"

export DOCKER_SANDBOX_HOST="${DOCKER_SANDBOX_HOST:-root@192.168.13.188}"

read-the-docs-community · 2026-08-20T06:34:08Z

Documentation build overview

📚 vime | 🛠️ Build #34149042 | 📁 Comparing 8396397 against latest (af29411)

🔍 Preview build

40 files changed · + 3 added · ± 20 modified · - 17 deleted

+ Added

± Modified

- Deleted

floatlibai added 2 commits August 11, 2026 11:02

add docker sandbox impl

5168b3c

Signed-off-by: flb_ <floatlibai@gmail.com>

fix eos token ids

8396397

Signed-off-by: flb_ <floatlibai@gmail.com>

floatlibai marked this pull request as draft August 20, 2026 06:31

gemini-code-assist Bot reviewed Aug 20, 2026

View reviewed changes

Navigation Menu

Uh oh!

Feat/docker sandbox - #394

Feat/docker sandbox#394
floatlibai wants to merge 2 commits into
vllm-project:ascendfrom
floatlibai:feat/docker_sandbox

floatlibai commented Aug 20, 2026

Uh oh!

gemini-code-assist Bot left a comment

Uh oh!

gemini-code-assist Bot Aug 20, 2026

Uh oh!

gemini-code-assist Bot Aug 20, 2026

Uh oh!

gemini-code-assist Bot Aug 20, 2026

Uh oh!

gemini-code-assist Bot Aug 20, 2026

Uh oh!

gemini-code-assist Bot Aug 20, 2026

Uh oh!

gemini-code-assist Bot Aug 20, 2026

Uh oh!

gemini-code-assist Bot Aug 20, 2026

Uh oh!

read-the-docs-community Bot commented Aug 20, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

		export MASTER_ADDR="192.168.13.190"
		export VIME_HEAD_HOST="${MASTER_ADDR}"

	export DOCKER_SANDBOX_HOST="root@192.168.13.188"
	export DOCKER_SANDBOX_HOST="${DOCKER_SANDBOX_HOST:-root@192.168.13.188}"

Uh oh!

Conversation

floatlibai commented Aug 20, 2026

Uh oh!

gemini-code-assist Bot left a comment

Choose a reason for hiding this comment

Code Review

Uh oh!

gemini-code-assist Bot Aug 20, 2026

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist Bot Aug 20, 2026

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist Bot Aug 20, 2026

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist Bot Aug 20, 2026

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist Bot Aug 20, 2026

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist Bot Aug 20, 2026

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist Bot Aug 20, 2026

Choose a reason for hiding this comment

Uh oh!

read-the-docs-community Bot commented Aug 20, 2026

Documentation build overview

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant