skip_environment_setup still uses /sandbox/.sandbox-venv/bin/python

[CoreJa]

Describe the bug
Adding flag skip_environment_setup under k8s(haven't tried other backend yet), container will use /sandbox/.sandbox-venv/bin/python as its default python interpreter, which obviously does not exists.

To Reproduce
Steps to reproduce the behavior:
Run the following:

# Define custom pod manifest
pod_manifest = {
    "apiVersion": "v1",
    "kind": "Pod",
    "metadata": {
        "name": "any_pod",
    },
    "spec": {
        "containers": [
            {
                "name": "sandbox-container",
                "image": "ghcr.io/vndee/sandbox-python-311-bullseye",
                "tty": True,
                "resources": {
                    "limits": {
                        "memory": "512Mi",
                        "cpu": "500m"
                    },
                    "requests": {
                        "memory": "256Mi",
                        "cpu": "250m"
                    }
                }
            }
        ],
    },
}

session = SandboxSession(
    client=k8s_client,
    backend=SandboxBackend.KUBERNETES,
    lang=SupportedLanguage.PYTHON,
    pod_manifest=pod_manifest,
    skip_environment_setup=True,
    verbose=True,
)

session.open()

session.run("print('Hello from LLM Sandbox!')")

Expected behavior
It should come with a stdout with the following string.

Screenshots

Outcome:

2025-11-30 15:03:18 - llm_sandbox.core.session_base - INFO - Copying /var/folders/c4/9xg90tkd4sd6xdp3cpwfl68w0000gp/T/tmpj2j4571_.py to /sandbox/79ab6a7de6c34c2399b27a57bf656534.py
2025-11-30 15:03:22 - llm_sandbox.core.session_base - INFO - Executing command: id -u
2025-11-30 15:03:23 - llm_sandbox.core.session_base - INFO - STDOUT: 0

2025-11-30 15:03:23 - llm_sandbox.core.session_base - INFO - Executing command: /sandbox/.sandbox-venv/bin/python /sandbox/79ab6a7de6c34c2399b27a57bf656534.py
2025-11-30 15:03:24 - llm_sandbox.core.session_base - ERROR - STDERR: /bin/sh: 1: /sandbox/.sandbox-venv/bin/python: not found

Additional context

As for session.run_command, it is well executed:

session.execute_command("python --version")

2025-11-30 15:07:03 - llm_sandbox.core.session_base - INFO - Executing command: python --version
2025-11-30 15:07:04 - llm_sandbox.core.session_base - INFO - STDOUT: Python 3.11.13

Is there a parameter or anything like it that I can pass to specify the python it should be using? Any help is much appreciated!

[github-actions]

The GitHub issue describes a bug that occurs when using the skip_environment_setup flag in a Kubernetes-based setup within the sandbox. With this flag enabled, the system incorrectly tries to use /sandbox/.sandbox-venv/bin/python as the default Python interpreter, even though this path does not exist. The issue arises when trying to run a custom pod manifest with a Python container, resulting in errors when the system attempts to execute commands using the missing interpreter. The user provides detailed reproduction steps, including the Python code to create and open a sandbox session, and shared screenshots of the erroneous output, where the execution fails with a "not found" error. The expected behavior is for the command to complete successfully with the correct Python output, but the bug prevents this from happening.

[dosubot]

Hi @CoreJa! I'm Dosu and I’m helping the llm-sandbox team.

This happens because the code always sets the Python interpreter path to /sandbox/.sandbox-venv/bin/python (derived from the workdir), even when skip_environment_setup=True—so the virtual environment is never created, but the code still tries to use it as the interpreter. There’s no fallback to a system Python like /usr/bin/python or a way to specify a custom interpreter path when skipping environment setup, so you get a "not found" error if the venv doesn’t exist in the image [source].

As a workaround, you can either build your image so that it provides a Python interpreter at /sandbox/.sandbox-venv/bin/python, or patch the code to allow specifying a custom interpreter path when skipping environment setup. There’s currently no built-in config option for this [source]. If you want to avoid modifying the image, you could also try setting a different workdir and ensuring the expected interpreter exists at {workdir}/.sandbox-venv/bin/python, but the limitation remains until the code supports custom interpreter paths for this mode.

To reply, just mention @dosu.

---

Docs are dead. Just use Dosu.

[CoreJa]

I found RuntimeContext with pip_executable_path that might help. I'll definitely look into it.

Haven't read the whole source code, but it seems to me that

self._python_env_dir = (workdir_path / PYTHON_VENV_DIR_NAME).as_posix()

this line 83 under session_base.py is somehow fixed there, and I didn't see any configuration setup I can modified it there. I'll be using it as my image venv as a workaround for now.

[vndee]

Hey @CoreJa, could you check if version 0.3.29 has resolved this issue?