Capsem Inspect AI Sandbox Extension (inspect-capsem-sandbox)

inspect-capsem-sandbox registers a capsem SandboxEnvironment for Inspect AI backed by isolated Capsem micro-VMs.

Installation

inspect-capsem-sandbox depends on capsem (>=0.6.3), inspect-ai, pydantic, and pyyaml, and drives the Capsem HTTP gateway (capsem-service) through the Python capsem SDK. It requires SDK APIs newer than the released capsem 0.6.3: those already on main (EXEC_TIMEOUT_CEILING_SECS, GATEWAY_REQUEST_BUDGET_SECS, decode_exec_output, and OCI container fields on Hypervisor.create) plus #286 (streaming of large binary bodies), #287 (CREATE_READY_SECS and the Hypervisor.create request_timeout), and Hypervisor.vm(id=...) foreign VM attachment; the capsem>=0.6.3 floor in integrations/inspect-ai/pyproject.toml must be bumped to the next published capsem release once cut. Install from the GitHub repository, pinned to the same branch, tag, or commit revision <rev> so the extension and the SDK match:

pip install \
  "capsem @ git+https://github.com/google/capsem.git@<rev>#subdirectory=sdk/python" \
  "inspect-capsem-sandbox @ git+https://github.com/google/capsem.git@<rev>#subdirectory=integrations/inspect-ai"

With uv:

uv add \
  "capsem @ git+https://github.com/google/capsem.git@<rev>#subdirectory=sdk/python" \
  "inspect-capsem-sandbox @ git+https://github.com/google/capsem.git@<rev>#subdirectory=integrations/inspect-ai"

Once installed, inspect_ai discovers the capsem sandbox environment through its [project.entry-points.inspect_ai] entry point.

Quickstart

Run any Inspect evaluation in a Capsem VM sandbox:

inspect eval task.py --sandbox capsem

Or configure the sandbox in a @task definition (execution_mode="vm" runs directly in a Capsem micro-VM; execution_mode="container" runs inside an OCI workload container or nested Docker container built from an image, dockerfile, or single-service compose_file):

from inspect_ai import Task, task
from inspect_capsem import CapsemSandboxConfig


@task
def my_eval() -> Task:
    return Task(
        dataset=...,
        solver=...,
        scorer=...,
        sandbox=(
            "capsem",
            CapsemSandboxConfig(execution_mode="vm", template="code", cpu_count=4, ram_gb=8),
        ),
    )

A string sandbox config is also accepted: a Dockerfile / Containerfile path, a compose.yaml / docker-compose.yml path, or a container image reference ("python:3.12-slim").

Documentation

See web/docs/src/content/docs/usage/inspect-ai.md for:

• vm vs. container execution modes, single-service Docker Compose support, and CapsemSandboxConfig fields
• Gateway discovery (CAPSEM_GATEWAY_URL, CAPSEM_GATEWAY_TOKEN, CAPSEM_RUN_DIR, CAPSEM_HOME) and INSPECT_CAPSEM_* image-cache / staging environment variables
• Execution timeouts, output limits (OutputLimitExceededError), write_file permissions, and SIGINT / inspect sandbox cleanup capsem lifecycle handling
• In-guest Dockerfile builds, host content-addressed image caching, and MITM CA trust injection inside containers

For the underlying Python gateway client, see sdk/python/README.md.