maf-sandbox-wslc

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 MafSandboxWslcExperimentalWarning; suppress it with warnings.filterwarnings("ignore", category=maf_sandbox_wslc.MafSandboxWslcExperimentalWarning) once you've read the notice.

This package is not affiliated with, endorsed by, or a product of Microsoft — it is a third-party reference implementation of microsoft/agent-framework#7568 for Microsoft Agent Framework.

app  ->  maf_sandbox  ->  maf_sandbox_wslc  ->  the container

The developer-machine sandbox backend: a container created by wslc, the container CLI that ships with WSL, in about half a second — no subscription, no daemon, 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-wslc

from maf_sandbox import Isolation, SandboxRouter
from maf_sandbox_wslc import WslcSandboxBackend, WslcSandboxConfig

router = SandboxRouter([WslcSandboxBackend(WslcSandboxConfig())], min_isolation=Isolation.CONTAINER)

samples/02_wslc_bicep runs those two lines end to end: a one-turn agent that validates a Bicep file against the compiler and takes the container down afterwards. Its sibling samples/01_acas_bicep is the same program on a microVM-isolated Azure backend, and the diff between them is two imports and one constructor.

Requirements

Windows with WSL 2.9.3 or later. wslc is part of WSL; wsl --version reports the version and wsl --update moves it forward. There is nothing else to install. The command-line contract this backend depends on — argv passed to exec natively, cp from a tar on stdin, label filters on list, WSLC_E_* codes on stderr — was verified against wslc 2.9.4.0. Every call spawns wslc.exe, 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.

What this backend declares

Isolation.CONTAINER. A container shares the host kernel and sits next to whatever the host process holds, 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. That refusal is the feature: this is a backend for the machine you are already sitting at, and opting the floor down is the one thing that lets you use it — there is no flag left to forget. Use a microVM-isolated backend where a deployment's credentials are in the picture.

Egress.CLOSED by default, Egress.ALLOWLIST on request. With no proxy configured every container is created --network none: the CLI cannot allow one host and deny the rest, so a spec's allowlist is honoured by denying everything — confining more than a workload asked for, which the router permits with a warning precisely because the failure is loud, and a workload built for this reports the shortfall rather than passing an incomplete result off as a clean one.

Set egress_proxy_image and the declaration becomes ALLOWLIST: 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. 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 pathlib import Path
from maf_sandbox_wslc import proxy_build_context, WslcSandboxConfig

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

The backend

WslcSandboxBackend implements maf_sandbox.SandboxBackend:

The filesystem path check on a write is answered inside the guest — the file name check is host-side text arithmetic and is not, and that is the residual to know about before choosing this backend. write_file refuses a path whose parents are links, which takes classifying every component from the filesystem root down. The cp tar header settles a directory and a missing path, and streams nothing for a regular file or a link — the two kinds that rule exists to catch — so those are settled by test run in the container being confined, through core's own maf_sandbox.paths.stat_by_asking_the_guest_as_root, which spells the probe and its ordering once so that no backend in this position invents a fourth version. The probe runs as --user 0, for the reason reclaim does: the file plane writes as root, so a probe as the image's user would be blind above a directory only root can search, and a cp still lands bytes there. Root is asked for reach and never for trust, and the helper checks that reach rather than assuming it, since a uid is not a capability set. A workload running as root can replace test in its own image and be believed, so the refusal is worth what the guest is. maf-sandbox-docker answers the same question out of its engine and this one has no equivalent until it can read an entry type without asking; #495 carries that decision.

declarations.os_families is {posix}, and it is stated rather than read. A workload names the guest shape its commands and scripts are written for in SandboxSpec.requires_os_family, and the router refuses a backend whose os_families does not hold it. wslc runs Linux containers in WSL 2's utility VM and has no other guest to hand out, so there is no engine to ask the way maf-sandbox-docker asks its daemon. The declaration is what this package's argv, its rm -rf reclaim and its posixpath path arithmetic already rest on. What it changes is one direction only: an undeclared os_families is the empty set, which refuses every spec that names a family, so a posix workload this backend could always have run was turned away at attach. A windows one is still refused here, as it should be — a backend that hands out Windows guests declares them and is matched instead.

Container names are derived from the key 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.

stop is never used to tear a sandbox down. A container whose init process ignores SIGTERM takes ten seconds to stop and under a quarter of a second to remove, and there is nothing in a sandbox worth waiting for. The one place it is used is the egress proxy, and only where a host registered an observer: its record has to be closed before it is read, or a request answered between the read and the removal reaches nobody. That pays the same ten-second worst case, on the proxy alone, on an acquire that is already collecting records.

Operator retention

reap discovers resources through WSLC, so a separate operator process can clean up after the application exits without its registry or conversation keys. It starts no timer. Save the following as an operator program, install this package in that program's environment, and run it under the Windows account that owns the WSLC engine:

import asyncio
import json
from dataclasses import asdict
from datetime import timedelta

from maf_sandbox_wslc import WslcSandboxBackend, WslcSandboxConfig

backend = WslcSandboxBackend(WslcSandboxConfig())
result = asyncio.run(backend.reap(timedelta(hours=24), scope="my-app"))
print(json.dumps(asdict(result)))
raise SystemExit(bool(result.failures))

Pause and drain acquisitions, restarts, and other resource mutations in the selected scopes before running the sweep. WSLC 2.9.3 addresses networks by name and offers no conditional identity or timestamp check on removal. The sweep rechecks identities, labels, workload presence and stop time, but these reads cannot close the interval before deletion. Workload removal uses its immutable ID without --force, so the engine refuses a workload that is running at deletion; a restart followed by another stop between inspection and deletion still requires operator coordination. Do not overlap sweeps. scope uses the same label encoding as creation; omit it only when maintenance covers all scopes on that engine.

Retention compares inspected UTC timestamps with the operator's Windows clock. Keep the Windows and WSL clocks synchronized: clock skew can advance or delay expiry by the offset. The helper does not measure or compensate for that offset.

The positive stopped_for duration applies these rules, with resources exactly at the cutoff retained:

An expired workload is removed first. Only successful removal permits deleting its proxy and network, even if they were rebuilt more recently. An expired orphan proxy similarly permits removing its network. Workload absence is checked again before each infrastructure removal. Proxies are force-removed by inspected ID; networks are never forcibly disconnected from attached containers. An eligible proxy confirmed absent by both ID and name clears its attribution and permits network cleanup without incrementing the proxy-removal count. A replacement proxy is retained, and a workload that disappears stops cleanup of its group. Only backend-shaped names with all four identity labels qualify, and grouped resources must share those labels. Missing or malformed age metadata retains the affected group with a failure; an incomplete inventory prevents all deletion. Failure counts include infrastructure, while disposed counts workloads only. Repeated sweeps tolerate resources already gone and reevaluate the resources left behind: a recently rebuilt proxy or network may need to age before a later retry qualifies it independently.

Failure codes preserve the source: unlisted covers failed queries and invalid inventory, ownership or age metadata; unreachable covers command invocation exceptions; timeout means the command's outcome is unknown; and refused means deletion was rejected and inspection confirmed the resource still exists.

The ordinary egress drain runs before removing a proxy this backend instance can attribute. A fresh operator process has no attribution for old proxy logs; archive them separately if needed. Retention does not stop running workloads and does not infer inactivity from host death. An operator needing a maximum running lifetime must supply that policy separately.

Use Windows Task Scheduler, or a Windows runner explicitly connected to the same engine, to execute the program during a maintenance window. Set its working directory, Python environment and wslc executable path explicitly, prevent overlapping runs, and monitor nonzero exits. Running as a different account or on a GitHub-hosted runner does not reach the developer's WSLC resources. A missed run extends retention. The scheduler and maintenance coordination belong to the deployment; see cleanup ownership.

The inspection contract comes from Microsoft's WSLC 2.9.3 container implementation and network implementation. Container inspection exposes timezone-qualified Created and State.FinishedAt; the reaper uses these instead of numeric listing fields. It accepts both 2.9.3 JSON arrays and the JSON-lines listings introduced by newer CLIs, including the 2.9.10 container output. Offline tests cover these shapes and retention races.

Live verification on 2026-09-08 used WSLC 2.9.4.0 and Python 3.13.12. The array listing's CreatedAt matched Unix seconds from inspected Created; StateChangedAt was also in Unix seconds but could fall in the second after inspected State.FinishedAt. WSLC 2.9.4 records state-change events separately from the inspected process exit. Inspection retained fractional seconds and a UTC suffix. Four separate-process probes passed: closed and allowlisted stopped workloads, an orphan proxy with its network, and a network alone. Running, freshly stopped and out-of-scope workloads and their infrastructure survived, and repeat sweeps removed nothing. Two further probes removed the proxy through WSLC just before its revalidation or removal, confirming that attribution clears and network cleanup continues. The WSL lifecycle clock was about five seconds ahead of Windows; the probes allow a bounded offset and wait until the inspected timestamp passes the operator's cutoff. WSLC 2.9.3 and 2.9.10 remain source-checked and covered offline, not measured live.

To repeat the live cleanup probes, set MAF_SANDBOX_WSLC_E2E_IMAGE to a runnable image and MAF_SANDBOX_WSLC_E2E_PROXY_IMAGE to a built proxy image, then run uv run pytest packages/maf-sandbox-wslc/tests/test_wslc_e2e.py -k reap -q. The separate-process probes abruptly exit the creator before another process sweeps the isolated test scope with a 30-second retention period, and include an out-of-scope control. The race probes remove a proxy during a sweep. All probes remove their resources afterward.

Maintained by SOKOLAI BV.

Upgrading to 0.13

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:

limits is not in that table because this backend never declared one — the router read its silence as DEFAULT_SANDBOX_LIMITS, and there was no backend.limits to read. backend.declarations.limits now answers with that same constant, so the ceiling is unchanged and the value is newly reachable rather than renamed.

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.