maf-sandbox-docker

Experimental. This package is early-stage (pre-1.0, Development Status :: 4 - Beta) — its API may change or be removed in a future release without notice. Importing it emits a one-time MafSandboxDockerExperimentalWarning; suppress it with warnings.filterwarnings("ignore", category=maf_sandbox_docker.MafSandboxDockerExperimentalWarning) once you've read the notice.

This package is not affiliated with, endorsed by, or a product of Docker Inc. or Microsoft — it is a third-party sandbox backend for Microsoft Agent Framework.

app  ->  maf_sandbox  ->  maf_sandbox_docker  ->  the container

The sandbox backend for everyone wslc leaves out: plain Docker containers, driven through the docker command-line client, on any machine with a Docker-compatible engine — macOS, Linux, Windows with WSL 2, and every GitHub Actions ubuntu-latest runner. No subscription, no login, and no dependency but maf-sandbox itself. A workload written against the protocol runs here unchanged, which is what makes it a workload rather than an integration.

Quickstart

pip install maf-sandbox-docker

from maf_sandbox import Isolation, SandboxRouter
from maf_sandbox_docker import DockerSandboxBackend, DockerSandboxConfig

router = SandboxRouter([DockerSandboxBackend(DockerSandboxConfig())], min_isolation=Isolation.CONTAINER)

samples/06_docker_codeact runs those two lines end to end: an agent that executes model-written Python in a container and reads the result back out. Its siblings 03_acas_codeact and 04_wslc_codeact are the same program on a microVM-isolated Azure backend and on wslc, and the diff between any two of them is two imports and one constructor.

Requirements

A Docker-compatible engine, reachable through the docker client. Docker Desktop (macOS, Linux, Windows with WSL 2) and Docker Engine (Linux, rootful or rootless) are what this backend supports. The client's own configuration — DOCKER_HOST, the active context, TLS settings — is inherited, because every call is a subprocess that inherits this process's environment; point DockerSandboxConfig.docker_path at a different client binary to use another one. Colima, OrbStack, Rancher Desktop and Podman expose Docker-compatible sockets and may work through the same client (Podman's default outbound network is called podman, so set outbound_network="podman" in allowlist mode), but they are not officially supported and nothing here is verified against them.

Docker Engine 28.0.0 or newer, to serve Egress.ALLOWLIST with hosts on the list. Such a sandbox is attached to an internal network whose bridge holds no host address. That is the bridge driver's gateway_mode_ipv4=isolated and its IPv6 twin gateway_mode_ipv6=isolated, both of which arrived in that engine. Both are set, because a daemon with IPv6 enabled would otherwise keep the v6 half addressed, and the route back with it. An older daemon rejects the options, and the acquire fails saying so rather than serving the workload on an addressed bridge — a bridge address is a route to the host that the allowlist does not cover. The floor is only for that case, because it is only there that a network is built at all: Egress.CLOSED, and an ALLOWLIST spec whose egress_allow is empty, both run on --network none and on any supported engine.

Every call spawns the docker client, so the host's event loop has to be one that can start subprocesses — asyncio's default Proactor loop on Windows does, and a host that installs WindowsSelectorEventLoopPolicy has to undo that first, or every acquire fails with a message saying so.

Hosts this backend does not serve: Windows without WSL (Docker Desktop's Hyper-V backend is documented by Docker but not its default, needs Pro or Enterprise, and is not verified here; Windows Home has no route at all), GitHub Actions' windows-latest (Windows containers only) and macos-latest (no Docker, no nested virtualization). For WSL-less Windows the eventual answer is a separate backend over Docker's "Docker Sandboxes" micro-VM product.

What this backend declares

Isolation.CONTAINER. A container shares the host kernel, below SandboxRouter's default min_isolation=Isolation.MICROVM floor — construct the router with min_isolation=Isolation.CONTAINER and it admits this backend; leave the floor at its default and construction raises SandboxBackendNotPermitted. A Docker Desktop or Colima VM does not lift the rung: one shared VM kernel serves every container, the same shape wslc's WSL 2 utility VM has, and the ladder classifies that at container. The declaration is a constant — no configuration raises it, because a security level the backend cannot verify must not become one the router repeats.

declarations.egress_modes = {closed} by default, {allowlist, closed} with a proxy configured. With no proxy configured every container is created --network none: a network namespace with only loopback, enforced by whichever kernel runs the container. That serves a workload declaring Egress.CLOSED, and refuses one declaring Egress.ALLOWLIST — the router never substitutes a mode, so denying everything is no longer offered as a stricter stand-in for a host list. Never UNRESTRICTED: a container backend always cuts or proxies, so it cannot serve a workload that asked to run open.

Set egress_proxy_image and ALLOWLIST joins the set: each sandbox gets its own internal network and a dual-homed filtering proxy, and the spec's allowlist is enforced by topology — the container has no route out except the proxy, which opens a CONNECT tunnel only to the hosts the spec names. The HTTP_PROXY/HTTPS_PROXY variables set on the workload are how ordinary clients find the proxy, not what enforces the allowlist; the topology is. TLS is not decrypted, and the sandbox never resolves an external name itself. The proxy is shipped as source, not as an image you must trust: build it from the packaged recipe, whose only pinned dependency is its Azure Linux base.

from maf_sandbox_docker import proxy_build_context, DockerSandboxConfig

print(f"docker build -t maf-egress-proxy:local {proxy_build_context()}")  # run this once
config = DockerSandboxConfig(egress_proxy_image="maf-egress-proxy:local")

Capability.FILES_OUT, with Capability.FILES_LIST withheld. This backend reads declared outputs back out — docker cp <container>:<path> - streams a tar whose entry header carries the size, the entry type and any link target after any extended metadata, so a file is statted and read from one stream with no stat command and no shell in the image. Docker can also archive a directory, but discovering its immediate children requires transferring the whole subtree, including file contents: ten 100 MiB files cost about 1 GiB just to learn ten names. A byte ceiling would have to refuse an oversized archive, never return a partial listing. This backend declines that cost; it does not fall back to guest-controlled ls or find. A kind that cannot name its outputs in advance requires FILES_LIST and is refused here — served instead by a backend, like ACAS, that offers directory listing. The investigation records the measurements and alternatives.

Every path component is checked, not just the last one. A symlink is refused on the tar entry's type bit only when it is the entry being tarred; the engine resolves the path daemon-side, so a guest that points out at /etc gets a stat of out/hostname describing a regular file with the parent link nowhere in it. stat_file and read_file therefore stat every parent component from the filesystem root down — not from the working directory, whose own ancestors the guest can replace just as easily: with /maf-sandbox -> / unchecked, /maf-sandbox/work stats as a real directory and serves /. The check itself is maf_sandbox.paths.refuse_symlinked_ancestors, not a copy living here: this backend passes it the unconfined tar-header stat above. A link is refused as a confinement failure and any other non-directory as an ordinary ENOTDIR — the entry comes back as EntryKind.SYMLINK or EntryKind.OTHER, so a caller can tell an escape from a guest tripping over its own fifo. One residual stays open: the check and the read are separate calls and docker cp has no no-follow form, so a guest that swaps a stat-ed component for a link in between is followed. What ends that check is a component the engine said is not there, so a copy that failed for any other reason raises rather than answering "absent": the daemon names the path it could not find, and a message that names something else — a container that has gone, a socket underneath an unreachable daemon — is a failure, not an empty directory tree.

FILES_IN has a write residual too. write_file checks the path, then sends a tar to docker cp for extraction with the daemon's root authority. A guest that swaps a checked parent for a symlink between those calls can redirect placement to a directory the guest program could not write. A symlinked final component is replaced rather than followed. Stamping the entries with the guest's uid/gid controls ownership of what lands; it does not bound the authority resolving the path. The REACH write probe checks that the landed file and directory are guest-writable, so passing it is necessary but does not prove this window is bounded. The tar plane keeps file transfer usable without a shell in the image; this residual is documented under #967, and closing the window remains #456.

The pull surface supports rootfs paths only. Keep work_dir and declared outputs on the container's ordinary filesystem. docker cp does not answer from the running guest's mount namespace: a file visible to the guest under a tmpfs or another unsupported mount can be reported absent, and collect_outputs can consequently report no output. Resources under /proc, /sys and /dev are outside this contract too (Docker's copy limitations). Creation adds no volume, mount or tmpfs flags; this is not a runtime check of mounts introduced by an image or inside the guest. Adding mounted output storage requires a pull mechanism that observes it, or an explicit refusal of the affected work directory, before it can be supported.

Whether that is actually enforced is not this package's own claim either. maf_sandbox.conformance is the shared suite every backend serving FILES_OUT answers, and this is the one backend that answers it against a real engine on every pull request — a container on the runner, a hostile layout planted in it through the public surface, and the probes attacking that.

declarations.os_families = {POSIX} on a Linux daemon, and nothing on any other — asked for with create. A workload states the guest shape its commands are written for in SandboxSpec.requires_os_family, and the router refuses a backend whose os_families does not hold it. This backend reads the answer from its own daemon (docker version --format '{{.Server.Os}}') rather than taking it as configuration: a host would be restating what the engine already knows, and a value it typed could only go stale against the engine that has to back it. The read needs an await, so it lives in a factory — __init__ makes no engine calls, and a blocking read in a constructor would do subprocess I/O on your event loop against a daemon that can hang rather than refuse.

from maf_sandbox_docker import DockerSandboxBackend, DockerSandboxConfig


async def wired() -> DockerSandboxBackend:
    # Asks the daemon once, and declares what it answered.
    return await DockerSandboxBackend.create(DockerSandboxConfig())


# Unchanged, and declares no family — which refuses only a spec that names one.
backend = DockerSandboxBackend(DockerSandboxConfig())

A windows daemon declares nothing, not WINDOWS. Everything this backend runs in a guest is POSIX: sh -c for a string command, rm -rf for a removal, and /-rooted path arithmetic that refuses a backslash outright. Declaring WINDOWS would be a promise no code path here backs, and would move the failure from the guest's first command to the router's certificate. A daemon that will not answer declares nothing for the same reason: silence refuses a spec that asks for a family and serves every spec that does not, which is what this backend did before it asked at all.

The declaration is a snapshot, so a cold acquire re-asks. The client resolves DOCKER_HOST and the active context on every call, so switching Docker Desktop to Windows containers moves the engine under a running backend. The router matched the old answer when your tool was attached and cannot ask again — so an acquire that is about to create or restart a container reads the daemon once more, ahead of the container and its network, and raises SandboxOsFamilyNotSupported if the answer changed. A restart counts because one that fails falls through to a create, and one that succeeds hands out a container from whichever daemon is answering now. Reusing an already-running container does not re-ask — that would cost a round trip on every tool call — and a backend built by the plain constructor never asks at all.

The backend

DockerSandboxBackend implements maf_sandbox.SandboxBackend:

Container names are derived from the key and kind rather than remembered, so acquire and dispose agree on one without a registry to keep in sync. Labels are the durable record dispose_scope selects on, and their values are digested when they are long or carry a separator — the same mapping on both sides, because transforming one and not the other makes a purge quietly select nothing.

No bind mounts, no host paths, and never the Docker socket cross into a sandbox — files go in and out only through docker cp. The hardening flags --security-opt no-new-privileges and --pids-limit go on every container; --cap-drop ALL, --memory and --cpus are opt-in through the config.

stop is never used. A container whose init process ignores SIGTERM takes ten seconds to stop and a fraction of a second to remove, and there is nothing in a sandbox worth waiting for.

Images whose user cannot be resolved

acquire raises SandboxCapabilityNotSupported when a workload requires FILES_OUT or HOST_TOOLS and the container's uid cannot be resolved. Root-owned input directories cannot promise that the guest can write outputs or host-tool markers beside them. Resolution uses Config.User, the container's /etc/passwd and /etc/group over docker cp, then id for missing values. An unset user is root; a numeric uid or any user these sources resolve remains supported.

Give the image a numeric uid:gid, readable account files, or an id it can run. Unresolved identities are retried on the next acquire, including after an unreadable Config.User; the container remains registered for retry or explicit disposal. Workloads requiring neither writing capability still run with a warning and root-owned inputs, which the guest may be unable to modify or reclaim. Commands whose result is stdout remain usable.

Proving a kind's confinement claim

Docker declares RECLAIM over its existing acquire-time reach check. Warm reuse additionally requires the kind to declare confined_to_guest_call_path=True; workloads that make no claim still use disposal. In a kind's tests, run its real call and cleanup through the shared probe on a fresh acquired sandbox:

from maf_sandbox.conformance import assert_nothing_left_behind
from maf_sandbox_docker.conformance import DockerFingerprintSubject

subject = DockerFingerprintSubject(sandbox, observer_image="trusted-python-observer:local")
results = await assert_nothing_left_behind(subject, call_and_cleanup)
assert all(result.passed for result in results)

The observer image is host-trusted tooling, available locally with Python 3.12+ as python; the helper pins its image ID. On Linux it combines rootfs changes with mounted-file contents, metadata, and process birth identities read through a separate observer container. This covers /dev/shm, which docker cp cannot observe. The observer has no network, a read-only filesystem, and only SYS_PTRACE for the kernel namespace read. It is removed on every exit path. The subject refuses dirty baselines, writable declared mounts, shared PID namespaces, privileged workloads, unreadable measurements, and exceeded limits. Non-Linux engines explicitly skip; asserting passed prevents treating a skip as proof. See the backend guide for the trust boundary and measurement limits.

The comparison measures final state. Regular-file contents use SHA-256 hashes. Verified Docker-managed mounts at /etc/hostname, /etc/hosts, and /etc/resolv.conf retain all other measured metadata but exclude ctime between observations, because Docker's archive setup changes it even on reads. Ctime still guards consistency during each read and remains compared for other mounted entries. A temporary change fully restored to the compared state leaves no residue under this measurement.

Run the live observer suite with MAF_SANDBOX_DOCKER_E2E_IMAGE naming a Linux workload image and MAF_SANDBOX_DOCKER_OBSERVER_IMAGE naming the trusted Python image. The same tests run on rootful and rootless engines. This adapter needs the core 0.37 line; the Docker release must follow that core's publication.

Explicit age-based cleanup

A killed host cannot run its cleanup. reap lets an operator find eligible Docker resources through the engine's labels and creation times, without the original conversation keys or process registry. This API does not provide automatic recovery: nothing schedules or runs it independently of the application. Cleanup ownership, lifecycle guidance and a reference operator deployment belong to #1008; it evaluates infrastructure-managed scheduling against a suite-owned runner.

import asyncio
from datetime import timedelta

from maf_sandbox_docker import DockerSandboxBackend, DockerSandboxConfig

backend = DockerSandboxBackend(DockerSandboxConfig())
result = asyncio.run(backend.reap(timedelta(hours=24)))
print(result.disposed, result.proxies_removed, result.networks_removed)
for failure in result.failures:
    print(failure)

The positive older_than duration is an explicit maximum-lifetime policy for this sweep. Age means creation time, not idleness or proof the host died: calling this API permits terminating active sandboxes older than that lifetime. An expired workload takes its proxy and network with it even if they were rebuilt more recently. Without a workload, the proxy's age decides; a network alone uses its own age. A workload exactly at the cutoff is retained; an orphan group applies the same strict comparison to the resource whose age decides. Hosts requiring long-lived active sandboxes to be protected must not use creation age as orphan detection; #1008 evaluates stronger lease or activity guarantees without requiring their implementation. scope="my-app" narrows the same operator action to that scope; omitting it covers all scopes on the configured engine. This is separate from a conversation's dispose_scope authority.

Only names from this Docker backend with all four sandbox identity labels qualify. The inventory includes stopped containers, orphaned proxies and networks whose containers are already gone. Every removal addresses the inspected resource ID, so recreating a name does not redirect it to the replacement; a network with attached containers is never forced away. After inventory, the resource whose age controls each expired group is inspected again by ID. If it has disappeared, that group is skipped so its stale age cannot expire replacement infrastructure discovered later. Inventory and revalidation failures prevent all deletion. Removal failures are returned individually, including infrastructure failures. Proxies are removed before workloads; a failed or already-absent container removal leaves the rest of that group untouched, preserving the workload age when a younger proxy needs a retry. Networks follow container removal because attached endpoints prevent their deletion. Each sweep recomputes eligibility from the resources that remain: a recently rebuilt network left alone after a failed removal may not qualify again until its own maximum lifetime elapses. The counts include only successful removals, with proxies separate from sandboxes.

The ordinary egress drain runs before removing a proxy this backend instance can attribute. A fresh operator process has no such attribution, so it cannot report those proxies' decisions; archive their logs first if you need them. This is the same attribution limit as a scope purge.

To inspect the labelled containers before maintenance, Docker also exposes their creation times directly:

docker ps -a --no-trunc --filter 'name=^maf-sandbox-docker-' --filter label=maf-sandbox.scope --filter label=maf-sandbox.thread --filter label=maf-sandbox.agent --filter label=maf-sandbox.kind --format '{{.ID}} {{.Names}} {{.CreatedAt}}'

Upgrading to 0.10

The four optional declarations moved into one BackendDeclarations. maf-sandbox 0.26 replaced capabilities, limits, egress_modes and os_families as backend attributes with one declarations object holding them as fields, and this backend follows it. A host that read them off the backend gets an AttributeError:

Nothing about what this backend declares changed — the values, and how they are derived from the config, are exactly as they were. maf-sandbox's own README carries the reasoning and what a backend author has to do.

Upgrading to 0.7

0.7.0 requires maf-sandbox 0.19, which made the egress mode a thing a workload declares and a set a backend enforces.

egress is replaced by egress_modes: frozenset[Egress]. A host that read backend.egress gets an AttributeError; read backend.egress_modes instead. Nothing in the wiring changes — the set is still derived from egress_proxy_image exactly as the single value was.

Without a proxy image this backend now refuses an allowlist workload rather than confining it further. That is the upgrade's one behavioural break, and it is most likely to reach you through a kind whose default asks for one — maf-sandbox-bicep 0.9 does:

SandboxEgressNotEnforced: sandbox backend 'docker' cannot enforce the 'allowlist'
egress the 'bicep' workload runs in (it enforces closed).

Configure egress_proxy_image if the workload is meant to reach the hosts it names, or ask the kind for Egress.CLOSED if it is meant to run offline. The old behaviour — serve it anyway, warn, and let the workload fail at the first fetch — is gone deliberately: a fetch failure deep in a tool call is a worse report than a refusal at attach.

Maintained by SOKOLAI BV.