maf-sandbox

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 MafSandboxExperimentalWarning; suppress it with warnings.filterwarnings("ignore", category=maf_sandbox.MafSandboxExperimentalWarning) 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, written for use with Microsoft Agent Framework but with no dependency on it in its protocol layer.

Quickstart

pip install maf-sandbox

from maf_sandbox import Isolation, SandboxKey, SandboxRouter, SandboxSpec, CallerContext

# Implement SandboxBackend against your own provider — or install maf-sandbox-acas for a
# ready-made Azure Container Apps Sandboxes backend — then wire it into a router. Configuring
# nothing gets the production posture (the default floor is Isolation.MICROVM); a developer
# machine opts down explicitly:
router = SandboxRouter([my_backend], min_isolation=Isolation.CONTAINER)
sandbox = await router.acquire(SandboxKey(scope="tenant-1", thread_id="t-1", agent_dir="devops"), SandboxSpec(kind="bicep", image="bicep-sandbox:0.46.1", egress_allow=("mcr.microsoft.com",), work_dir="/workspace"))

This snippet never calls ensure_can_serve (below) and is checked anyway: acquire runs the same floor, capability and egress refusals itself before it ever reaches the backend, so the only thing calling ensure_can_serve first buys you is the closed-egress-vs-allowlist-spec warning, which acquire deliberately stays silent about.

samples/01_acas_bicep is that wiring as a runnable program, including the part no snippet shows well: building the CallerContext out of callables rather than values, which is what keeps a SandboxKey a property of the host's request.

Threat model

This package draws no isolation boundary itself — it is protocol and policy over whatever a SandboxBackend implementation actually provides. Isolation is a seven-rung ladder a backend declares itself onto, weakest to strongest: none (no boundary at all — the workload runs in the host process, with the host's authority), runtime (a software boundary inside the host process, e.g. a restricted interpreter or a WASM runtime's fault isolation), os_process (a separate OS process — a kernel-enforced address space, sharing the kernel and the filesystem), container (shared-kernel namespaces and cgroups), hardened_container (syscall interception in a userspace kernel — gVisor-class), microvm (a hypervisor boundary with a minimal or absent guest OS and no ambient identity reachable from inside — the default floor), and vm (a dedicated, full VM provisioned for the workload). SandboxRouter enforces the checks below on top of that declaration; the package's job is to make an unsafe backend selection fail loudly at construction or attach, not silently at first use. Beyond backend selection this layer holds no credentials, executes nothing and reaches no network, and everything security-relevant about a specific sandbox lives in the backend that implements it. It has exactly one boundary of its own, and it is on the way out rather than in: make_file_system_sink writes guest-produced bytes under a host directory, so it resolves each destination and refuses one that leaves that directory — see Getting files back below, which is also where a host landing somewhere other than a filesystem is told it owns the same question.

The vocabulary

Isolation, weakest to strongest: none < runtime < os_process < container < hardened_container < microvm < vm. SandboxRouter's default min_isolation is microvm; an unrecognised rung refuses rather than guesses which side of the floor it falls on.

SandboxKey's scope and thread come from the host's request context through CallerContext, whose fields are callables read at call time rather than values. That is deliberate: a key a caller can supply is a key a model can supply, and that would let one conversation address another's sandbox.

SandboxSpec.egress_allow is an allowlist — everything not named is denied, so an empty tuple means no network. Stating it positively means a spec that forgets to mention egress gets the closed configuration rather than the open one.

Four axes, six checks that are not conveniences

router = SandboxRouter(backends)                                   # default floor: Isolation.MICROVM
router = SandboxRouter(backends, min_isolation=Isolation.VM)       # stricter: dedicated full-VM only
router = SandboxRouter(backends, min_isolation=Isolation.NONE)  # a developer machine, opted down

Which backend, when there is more than one. By default the router resolves one at construction — selected="docker" names it, or the first registered one wins — and every workload gets that one, so a spec it cannot serve is refused with the other registered backends untouched. SandboxRouter(backends, selection=Selection.PER_SPEC) routes instead: the first registered backend that passes every check below, decided per spec. It is opt-in because of what it can move, and the claim needs its condition stated. For a router with no selected pin it can only ever serve a spec that is refused today — routing picks the first registered backend exactly as the fixed selection resolves to it — so nothing already running moves, and what changes is that a refusal becomes a running sandbox, which on a remote backend has a price. Migrating off a pin is the case to check: selected= and PER_SPEC are refused together, so a host dropping selected="second" has routing start at the first registered backend, and a workload the second was serving moves unless backends is reordered to match. Registration order is the preference, and the route is a pure function of the spec and what the backends declare — never load, latency or cost — so one spec always routes to the same backend and the warm sandbox acquire reuses stays reachable. Per spec rather than per conversation: two kinds under one key may route to different backends by design, which is why dispose asks every registered backend rather than one.

1. The minimum-isolation floor. A backend declares its own isolation, ranked on the ladder above. The router refuses, at construction, the backend it resolves to when it sits below min_isolation — or when its declared value is not a rung this package recognises, because nothing here can tell whether an unrecognised boundary is stronger or weaker than the floor. Under Selection.PER_SPEC there is no one resolved backend, so the same refusal is judged across the whole registration: construction fails when nothing registered clears the floor, and an individual backend below it is kept and never routed to. A spec may also carry its own min_isolation; the effective floor is the stricter of the host's and the spec's — a spec may raise the floor for itself and never lower it.

It refuses rather than degrades — under Selection.FIXED, where the backend it resolved to is the only one it will ever use. Promoting to a stronger backend unasked would hide a misconfiguration, and proceeding with the weaker one would break claims the host's security posture makes about every execution surface; neither is better than an error.

Under Selection.PER_SPEC a host has asked for that promotion, so routing does pass over a below-floor backend and serve on one that clears the floor. The floor itself is never crossed — every candidate is checked against it, so nothing below it can serve — but the passed-over backend would otherwise go unmentioned, which is the misconfiguration half of the paragraph above rather than the safety half. The per-spec refusal names it only when no candidate serves at all, since a successful route discards the refusals it passed over — so the router says it once, at construction, with a logger.warning naming each registered backend below the floor. It warns rather than refuses because a registration that includes a weaker backend is the arrangement this mode exists to serve, and it does not advise unregistering it: dispose and dispose_scope reach every registered backend, so a host that changed which one serves would strand whatever the old one still holds.

2. The capability match. A backend declares declarations.capabilities (a frozenset[Capability]: EXEC, RUN_CODE, HOST_TOOLS, FILES_IN, FILES_OUT, FILES_LIST, FILES_DELETE, SNAPSHOT, ATTACHED_IDENTITY) — what it can actually do — and a spec declares requires, what its workload cannot run without. ensure_can_serve(spec) raises SandboxCapabilityNotSupported when the backend is missing something the spec requires — and where the router selects per spec, that check is also what chooses, so it raises only once every registered backend has refused, naming each. Unlike the floor, silence here is a functionality claim rather than a safety one: an unstated capabilities reads as exactly DEFAULT_CAPABILITIES = {EXEC, FILES_IN} — what this package's own Sandbox protocol already obligates, so a backend written before Capability existed does not have to start lying to keep working.

3. The egress rule, unchanged in substance. egress_allow was a contract nothing checked, so a backend that reads it and one that ignores it have the same type, the same methods and the same passing tests — each one declares an Egress level instead: allowlist (deny by default, allow the named hosts), closed (all or nothing), or unrestricted (cannot confine egress at all). ensure_can_serve(spec) refuses the last one. Here silence is not read charitably: an undeclared egress is treated as unrestricted and refused, because a backend written before the property existed cannot have been enforcing an allowlist it never read.

Missing in either direction is refused, and the symmetry is the rule rather than an omission. Confining less than the spec asks silently widens what the workload was designed to reach; confining more hands it a posture it was not built for, and a workload that fails at whatever it could not fetch fails somewhere no reader of the spec would look. A backend serves the mode it declares or turns the workload away.

4. The guest-shape match. A backend declares declarations.os_families (a frozenset[OsFamily]) — the guest shapes it hands out, posix or windows — and a spec declares requires_os_family, the shape its commands and scripts are written for. ensure_can_serve(spec) raises SandboxOsFamilyNotSupported on a mismatch, so a POSIX workload meets a Windows guest at attach rather than at its first command. The axis is path grammar and argv quoting, and nothing else: a spec asking for posix and getting it can still meet an image with no shell, because what is installed in a guest is a property of the image, and one backend may be handed many. docs/sandbox/guest-platform-and-commands.md settles where that separate question is answered. Silence here is neither of the readings above — an unstated os_families is the absence of an answer, read as frozenset(), which refuses a spec that asks and leaves every spec that does not exactly as it was. A backend with no guest in the operating-system sense, such as one serving a language runtime, has nothing to declare and declares nothing.

5. The transfer-ceiling match. A spec carries TransferLimits per direction — max_bytes_per_file, max_total_bytes, max_files — and a backend may declare its own ceilings as limits. A spec asking above them raises SandboxTransferLimitsNotPermitted rather than being clamped: a workload served a smaller cap than it declared fails part-way through a collection, and a partial artifact set is worse than none because the model cannot tell what it did not get. Silence follows the safety rule, not the capability one — an undeclared ceiling is the default ceiling, and a bigger ask is refused.

6. The isolation scope. How much of a conversation one sandbox serves. A spec declares isolation_scope — conversation, the default, one sandbox reused across the conversation's calls; or call, one created for the call and deleted when it returns — and a backend declares declarations.isolation_scopes, the scopes it can serve. SandboxRouter(min_isolation_scope=...) is the host's floor on the same axis, and the effective scope is the stricter of the two, exactly as with min_isolation. ensure_can_serve(spec) raises SandboxScopeNotEnforced otherwise, because a backend cannot answer a per-call workload by sharing: every call would succeed with the separation absent. Silence here is the one declaration read as a claim rather than as the absence of one — an unstated isolation_scopes means {conversation}, the get-or-create every backend already did.

call costs a cold start per call and buys what cleanup cannot: two calls of one conversation never meet in one filesystem, so a reclaim that failed, a program that would not stop, and anything a call left unnamed all stay where no later call can address them. A backend declares it once it folds SandboxKey.call_id into whatever names a sandbox — a container name, a label set, its own registry — and maf_sandbox.conformance.assert_call_scope_conformance is what holds it to that. None of it makes a removal an erasure: a snapshotted disk image can keep blocks an unlink released, which is a property of the backend's storage and not something the protocol states.

Note that the checks answer to different owners. How strong the boundary must be here, and how little of a conversation one sandbox may serve, are the host's policy, read from min_isolation and min_isolation_scope — and a spec may raise either floor for itself, never lower it. What a sandbox may reach, and what it must be able to do, are properties of the workload, stated in its spec. Keeping the axes apart is deliberate: merging isolation into a "required capabilities" list would let a workload ask for a weaker boundary than the deployment mandates.

ensure_can_serve is also the whole of a wiring test, in your own repository, against your own backend choice:

router.ensure_can_serve(bicep_sandbox_spec())

Getting files back — the declaration, and where it lands

A workload's only return channel used to be ExecResult.stdout, which is right for a diagnostic and wrong for a rendered image. Capability.FILES_OUT is the pull surface, and it is narrow in two deliberate ways: this library never discovers what a workload produced, and it never decides where the bytes go.

Declare it. A DeclaredOutput names one artifact as a literal path relative to work_dir, in SandboxSpec.declared_outputs. Literal rather than a glob: resolving a pattern means enumerating a directory, which is the primitive Capability.FILES_LIST exists to gate, so a kind that cannot name its outputs in advance requires that capability and a backend serving only FILES_OUT refuses it. media_type is declared rather than sniffed, because sniffing lets guest-produced content decide how the host handles it. required=False is how a workload says an absence is normal — a renderer exiting non-zero produces no file, and the model needs that diagnostic rather than a transfer error stacked on top of it. name is the spelling the artifact lands under and defaults to path; the two come apart as soon as a kind writes into a per-call directory, which warm sandbox reuse forces on any kind whose outputs would otherwise persist into the next round.

disposition keeps the two flows apart because they answer to different legs of a host's policy: LAND goes to the sink and the question is confidentiality, while CONSUME is parsed by the kind that asked for it and the question is integrity. A CONSUME output is still counted against every cap — files_out bounds the collection the spec declared, not the subset of it that lands.

Receive it. await collect_outputs(sandbox, spec, sink=...) returns LandedArtifacts in declaration order. The order of its phases is part of the contract rather than an implementation detail: everything the declaration alone decides — a sink for anything that lands, a valid name for every output, no two landing names that collide — is settled before the sandbox is touched, then every declared output is stat-ed and capped, then the landing ones are read, and only then is anything delivered. Delivery is a push nothing can take back, so a refusal arriving after the first deliver could not leave the host as it found it.

spec.files_out is a TransferLimits and all three of its fields are load-bearing: a byte ceiling alone does not bound a collection, since ten thousand files one byte under the per-file cap cost exactly what the cap was written to prevent. What comes back when a collection does not fit is specific rather than generic — SandboxTransferCapExceeded names both the cap and the file that breached it, SandboxOutputMissing names a required output that was not there, SandboxOutputSizeUnknown is a backend that could not say how large something was, and SandboxArtifactNameCollision is two landing names that are one file at the destination: identical, or differing only by case or by Unicode form.

Land it. An OutputSink wraps a single async def deliver(artifact) -> LandedArtifact. This library holds no opinion about where an artifact goes — a directory, a blob container, a file store — which is what keeps that flow visible to the host's own information-flow policy instead of buried in a dependency. LandedArtifact.display is the one line the model is allowed to see; handle is the host's own reference, and nothing renders it into the transcript.

validate_artifact_name is lexical, so a sink still has to confine its own destination. It refuses .., absolute paths, backslashes and empty segments, so the name cannot traverse — which is not the same as safe, because it says nothing about what is already sitting at the path that name resolves to. A symlink in the output directory carries the write straight out of it: the same failure class as #142, on the host side of the boundary.

make_file_system_sink(root) is that check, packaged. It resolves each destination, refuses anything leaving root with SandboxLandingNotConfined, refuses one that is already there with SandboxLandingExists, creates the parents a nested name needs, and writes. That second refusal is the default because the name check in collect_outputs is per collection — a name is not fresh just because this call is, and a root more than one conversation lands in is a channel between them. It is an exclusive create rather than a look, so nothing takes the destination in between. A workload landing one stable name wants existing="replace", since its own previous call is then the commonest thing in the way; and either way the refusal arrives per artifact, during delivery, so a collection whose third name is occupied leaves the first two landed. Reach for it rather than writing the four lines yourself — two samples here wrote them by hand and only one got it right. Pass display when the kind introduces its artifacts in its own words. It stays a check rather than a guarantee, and that is a property of the filesystem rather than of the helper: resolving and writing are two calls, so a host landing genuinely hostile output wants no-follow primitives underneath. What it closes is the standing case — something already in the way when the run started.

A sink landing somewhere that is not a filesystem — a blob container, a UI panel — writes its own deliver and owns the equivalent question for that destination.

make_file_store_sink(store, *, provenance=None) is the packaged one for an agent_framework AgentFileStore, and it lands <call_id>/<name> rather than <name>: the model reads its own output back with a file-read tool instead of being told by the workload which names landed, and one call's file can never answer for the next call's. The folder is the host-minted call id, which collect_outputs(call_id=...) supplies — required rather than optional, because the sink declares OutputSink.per_call. A destination that already exists is refused with SandboxLandingExists rather than replaced, every landing is recorded into provenance before the bytes are written, and an artifact whose bytes are not UTF-8 is refused with SandboxLandingNotText rather than mangled into a store that holds text. Point it at a store the model can read and not write — never the one the agent's file_access_write writes to. sandbox_outputs_read_tools(store) is the other half — <prefix>_ls and <prefix>_read over that store and nothing else, read-only by construction rather than by a flag. It exists because FileAccessProvider names its tools from fixed constants, so a second one of those is a name collision rather than a second store. docs/sandbox/hosts.md carries the wiring and the trade: those two tools carry no label of their own, so a host that withholds a workload's guest output and then wires them has moved that output onto a path it classifies rather than kept it away from the model.

from pathlib import Path

from maf_sandbox import (
    Capability, DeclaredOutput, SandboxSpec, TransferLimits, collect_outputs,
    make_file_system_sink,
)

spec = SandboxSpec(
    kind="diagram",
    image="diagram-sandbox:1",
    egress_allow=(),
    work_dir="/workspace",
    requires=frozenset({Capability.EXEC, Capability.FILES_IN, Capability.FILES_OUT}),
    declared_outputs=(DeclaredOutput(path="diagram.png", media_type="image/png", required=False),),
    files_out=TransferLimits(max_bytes_per_file=8 * 1024 * 1024, max_total_bytes=16 * 1024 * 1024, max_files=4),
)

landed = await collect_outputs(sandbox, spec, sink=make_file_system_sink(Path("out")))

Reclaim the sandboxes when the conversation ends. router.scope(scope, thread_id) is an async context manager that calls dispose_scope however the block ends, and cannot mask an application error on its way out — dispose_scope already swallows and logs each backend's failure. Its own reason is why this is packaged rather than left to every host to remember: a sandbox nobody reclaims is a sandbox somebody pays for.

async with router.scope(scope, thread_id) as reclaimed:
    ...                                    # attach tools, run the turn
print(f"Disposed {reclaimed.disposed} sandbox(es).")   # the count arrives after the block

A workload whose artifact names are not knowable when its tool is built passes the same DeclaredOutput type to collect_outputs(outputs=...) instead. That is refused unless the spec sets outputs_named_at_call_time: without the flag, the tool was attached with no sink required of it and no outbound cap agreed, and collecting there would land artifacts behind both checks.

samples/08_docker_codeact_files is all of the above as a runnable program, against a real engine.

Host tools — the contract, and the backends that serve it

Capability.HOST_TOOLS is the one capability where trust crosses outward: a called function body runs in the host process, with the host's privileges, driven by model-written code, and each host-tool call bypasses whatever middleware the host runs. maf-sandbox-docker and maf-sandbox-acas declare it; maf-sandbox-wslc does not. The safety contract shipped first, before anything could use it, and it is what a host configures either way: HostToolRegistry starts empty (nothing is callable until a developer registers it, and registering emits a one-time, suppressible MafSandboxHostToolsWarning); @sandbox_tool(source=..., sink=..., identity=...) makes the developer answer every information-flow leg with no defaults (None is an answer — "not that role"); a require_declared gate refuses unstamped functions at registration, which is the only place the declaration is ever read — register captures it, HostToolRegistry.aggregate() seals the registry as it derives policy from it, and a stamp swapped or removed afterwards reaches nothing; allowed_identities (default frozenset({Identity.APP})) refuses at registration a tool exercising a broader authority — an Identity.USER tool, or an unstamped one read as APP — so user authority is opt-in (frozenset({Identity.APP, Identity.USER})), a tool declaring identity=None is always allowed, and denied_identities on the router stays the attach-time backstop; each run is bounded by a host-tool-call cap (DEFAULT_MAX_HOST_TOOL_CALLS_PER_RUN, refusals included) and by response size caps that reuse TransferLimits; arguments are validated host-side at the registry's one door, never in a guest shim; and a host whose posture wants a hard stop rather than awareness passes denied_capabilities={Capability.HOST_TOOLS} or denied_identities={Identity.USER} to its router.

One sentence to read before registering anything, because a declaration reads like a control and is not one: Identity.APP is not the safe option, only the declared one. It is the application's full authority, and the only real bounds on it are the emptiness of the registry and the host-tool-call cap — least privilege for host-tool calls comes from what a host registers, never from what it declares. Identity.USER is served only where a host mints it: give the registry mint_user_identity, an async callback returning that run's authority, and it reaches the tool body as user_identity. Without one, such a tool registers and its call is refused. Registering one raises the whole surface to approval-gated either way.

Reaching the host from inside — host_tool_calls_over_exec

The contract says what may be called; it does not say how a host-tool call reaches a host whose guest speaks an exit code, stdout, and a stat-and-read pull surface. host_tool_calls_over_exec is that channel, built from those primitives and nothing else, and it is a helper a kind composes rather than anything the protocol requires. A kind writes the program and the generated shim (host_tool_shim) into a fresh per-run directory (guest_run_layout); host_tool_calls_over_exec writes the launcher itself, starts it detached and then polls for request files, resolves each one through HostToolRun.call — the same one door, with the same gates, cap and ceilings — and writes the answer back. It needs EXEC, FILES_IN and FILES_OUT, and deliberately not FILES_LIST. One run is two directories, and that is what keeps a guest-supplied name away from the machinery serving its own call: WORK_DIRECTORY is the program's working directory and the only one a kind puts model-named files in, while the shim, the launcher, the output, the exit marker and the calls directory sit in a sibling nothing a model names can reach. The program itself lives in the second one, beside the shim, because sys.path[0] follows the script rather than the working directory — run from the work directory it would put a guest file named maf_host_tools.py ahead of the real module, which is exactly the substitution a list of reserved names is hardest to get right about.

A run that overruns is signalled, not just reported. The program is started detached, so the bound expires in the supervisor rather than inside an exec a backend could tear down along with its container — which used to mean a timed-out program kept running and the only remedy was disposing the whole sandbox, taking every other call in that conversation with it. The launcher records the program's pid and the supervisor signals it on the way out, over the same exec this transport already runs on: no protocol method, no capability beyond the ones the host-tool-call path requires. The message says the signal was sent, not that the program died — kill reports success for a signal the kernel accepts and discards, and the pid comes from a file the program can rewrite, so a guest that names pid 1 or another process outlives a call that reports and was sent SIGKILL. Where the signal could not be sent at all the message says "could not be signalled, so it may still be running", and disposal is what stops it. Making the stop something the host can rely on against hostile code is #463 — the pid and the session are still read from files inside the run. Children the program spawned are killed with it where the guest has setsid, unless they left the group. The launcher starts the program inside a session of its own — setsid runs a shell, that shell leads the session, and the program is its child — and the supervisor signals the process group the program starts in, so a program that forked a spinner does not leave it burning CPU for the rest of the conversation. A descendant that calls setpgid stays in the session, leaves that group, and survives — reach says "group" and means it. Where setsid is missing the program shares the launcher's session, only its own pid can be signalled — a group signal there would reach the whole container — and the message says so rather than implying otherwise: "and was sent SIGKILL, which reaches it alone — anything it spawned is still running". The signalled processes stay as zombies until the sandbox goes if nothing in the image reaps them, which is normal for a container whose pid 1 is not an init. The guest needs sh, nohup, printf, mv, mkdir, rm and kill, and uses setsid when it is there.

The shim is not a control. It runs where model-written code can read, edit or ignore it, and a program that writes request files itself is served identically. That is the design: every gate is host-side, and a check running in the guest would be decoration.

The transport tries not to let its own files outlive the call. It removes the ones it owns — the program, the shim, the launcher, the captured output, the exit marker, the pid, and every request and response the run exchanged with the host — on every exit path, success included, over the same exec it uses for everything else — best-effort, not a retention guarantee: a guest without rm, a removal that times out, or a non-zero exit each leave that traffic readable, logged and nothing more. What it cannot remove is WORK_DIRECTORY: artifacts live there and a kind collects them after the transport has returned, so removing it would delete the outputs of every successful run. reclaim_run(sandbox, layout) is the other half, a kind's to call in a finally once it has collected, and it takes the whole run directory. A False from it is a data-retention failure rather than a tidiness one: nothing comes back for it — the protocol's delete is capability-gated and this transport does not require it — and acquire is get-or-create, so a run directory that survives is readable by every later run in the same sandbox for the life of the conversation. A kind that takes its place in the guest from SandboxToolSession.guest_call_path() does not have to remember any of this: sandboxed_tool removes that path, and everything under it, when the call returns — after a result, a refusal and an exception alike — and hands a removal that did not happen to the host's on_reclaim_failure as a ReclaimFailure — a notification, delivered after the framework has already disposed the sandbox by default (see Upgrading to 0.23 below). A kind that composes its own path keeps reclaim_run, and keeps the finally. Disposal is no longer the host's to arrange: the framework disposes an unclean sandbox itself and refuses the key until a disposal lands, and a host loosens that only by opting down on the router with reclaim=ReclaimConfig(failed_reclaim_policy=FailedReclaimPolicy.KEEP), never per kind.

It costs round trips — several backend calls per host-tool call, plus polling, plus one on every return to reclaim, and one more to stop the program on a run that overran. It serves one outstanding call at a time. This module's own docstring counts those costs exactly, beside the code that decides them; whether the trade is worth it is a measurement rather than an assumption.

A result the model may read half of

A sandbox result is rarely uniformly derived: a compiler's diagnostics quote a template the model wrote, while the sentence naming what to do about them is a constant the package ships. Under one label a kind has to choose — claim trusted over the guest's text, or declare honestly and watch MAF's information-flow module hide the whole result behind a variable reference. So a tool body may answer with a list of items instead of one string, and MAF labels and hides each item separately: the standing guidance stays readable while everything the call produced is hidden.

from agent_framework import Content
from maf_sandbox import SourceIntegrity
from maf_sandbox.maf import labelled_result_item

return [
    labelled_result_item(RECOVERY_ROUTE, SourceIntegrity.TRUSTED),
    Content.from_text(rendered_diagnostics),
]

Label as little as you can, and never every item. A per-item label replaces the item's whole label, confidentiality included, and this package has no confidentiality value to put there — those are the host's vocabulary, carried verbatim. An item left unlabelled takes the call's own label instead, and the result's combined label is the most restrictive across every item, so one unlabelled item is what keeps the host's classification. sandboxed_tool refuses a result whose every item carries a label, because nothing in it is left to carry the call's; labelled_result_item refuses SourceIntegrity.UNTRUSTED for the same reason from the other side, since the untrusted item is the one holding what the call produced. str stays valid and stays the common case.

What may carry TRUSTED is narrow — text whose value and whose presence are independent of everything the call touched, which in practice means standing guidance emitted on every return path. A count, an exit status, a size, or a line emitted only on failure all fail that test however they are split out. docs/sandbox/information-flow.md carries the rule and the measurements behind it.

Recording what the sandbox did

This package logs, at warning, and a log line is neither structured nor keyed. A deployment asked which conversation reached that host, which host tools ran under whose authority, what crossed the boundary and with what label answers from records, so SandboxObserver is the seam that hands them over: six frozen events, in this package's own vocabulary, joined by the SandboxKey that addresses a sandbox. SandboxAcquired and SandboxDisposed always carry one; HostToolCalled, StoreFileRead and OutputsCollected type it SandboxKey | None, since each has a case with no sandbox behind it; and ToolCallEnded carries keys, a tuple of every key the call touched — acquired, refused, or only read the store under — so that each of its other events has a call to join to.

from maf_sandbox import (
    HostToolCalled,
    HostToolRegistry,
    Isolation,
    SandboxAcquired,
    SandboxObserver,
    SandboxRouter,
)
from maf_sandbox.testing import InProcessSandboxBackend


class Records(SandboxObserver):
    """Override what you want; every event the base class answers with nothing."""

    def sandbox_acquired(self, event: SandboxAcquired) -> None:
        emit(thread=event.key.thread_id, egress=str(event.spec.egress), refused=event.refusal)

    def host_tool_called(self, event: HostToolCalled) -> None:
        emit(tool=event.tool, sink=event.sink, how=event.outcome, bytes=event.response_bytes)


def emit(**attributes: object) -> None:
    """Wherever this host's records go — a queue, an exporter, a SIEM."""


records = Records()
# Both registration points, since `host_tool_called` above comes from the registry and never
# from the router. The floor is lowered only for the in-process fake, which declares
# `Isolation.NONE`; a real backend leaves the default `microvm` floor where it is.
router = SandboxRouter([InProcessSandboxBackend()], min_isolation=Isolation.NONE, observer=records)
registry = HostToolRegistry(observer=records)

SandboxAcquired and SandboxDisposed come from the router; HostToolCalled from HostToolRegistry(observer=…), which is where every other host-tool policy lives; StoreFileRead from SandboxToolSession.read_file and ToolCallEnded from the wrapper sandboxed_tool builds, both reading the router's; and OutputsCollected from collect_outputs(..., observer=session.observer, key=key), which is a function rather than a policy object and so takes both as arguments.

Three things to know before writing one. Every way out is recorded — a refused acquire, an exhausted host-tool cap, a collection refused part-way, a call taken by a cancel — and an acquire's, a collection's or a call's failure is recorded as the exception's class name, never its message, which can carry a backend's endpoint. HostToolCalled.refusal is the exception: it holds the sanitized sentence the guest was answered with, so treat that one as guest-influenced rather than host-only. An observer cannot fail a call: its exceptions are contained and logged, and its return value is never read. It can, however, slow one down, and it is entered from more than one thread — it runs wherever the call is served, which for a synchronous tool body is a worker thread, so hand the event to a thread-safe queue or a batching exporter and do no I/O in it. A host that registers nothing builds no event at all.

docs/sandbox/observability.md carries what each event holds, what a recorder should treat as guest-chosen, and what the seam does not yet see — the egress proxy's own ALLOW/DENY lines among them.

Upgrading to 0.27

These landed in the tree tagged maf-sandbox-v0.26.0, which never reached PyPI. That tag and its GitHub Release are immutable and will stay visible; there is no 0.26.0 to install, and the same tree ships as 0.27.0. The changelog's 0.26.0 section says why.

A backend's four optional declarations became one object. capabilities, limits, egress_modes and os_families were four attributes the router read off a backend with four getattr calls. They are four fields of one BackendDeclarations, read with one, and a backend still carrying any of the four attributes is refused when the router resolves it — at construction, with the attribute named. That refusal is deliberate: none of the four was ever a member of the SandboxBackend protocol, so isinstance holds either way and nothing in the type system marks a backend half-moved, while a stray attribute is silently replaced by that field's default. On egress_modes the default enforces nothing and refuses every workload; on limits it widens a ceiling the backend meant to be narrow.

from maf_sandbox import BackendDeclarations, Capability, Egress, Isolation

class MyBackend:
    name = "mine"
    isolation = Isolation.CONTAINER
    declarations = BackendDeclarations(
        capabilities=frozenset({Capability.EXEC, Capability.FILES_IN}),
        egress_modes=frozenset({Egress.CLOSED}),
    )

Each field's default is its own silence rule, and the four still differ — capabilities reads as DEFAULT_CAPABILITIES, limits as DEFAULT_SANDBOX_LIMITS, and egress_modes and os_families as the empty set. So a field left unstated means exactly what an absent attribute used to, and a backend that declares neither the object nor any of the four attributes it replaced reads as DEFAULT_BACKEND_DECLARATIONS. One that still carries any of the four is refused at construction, per the paragraph above — declaring no object is not a way to stay unmigrated. isolation did not move: it is a protocol member, because a backend with no rung cannot be placed against a floor.

capabilities and egress_modes are now also refused when they are not a set — the router subtracts one and tests membership in the other, and a string or a list used to raise a bare TypeError out of a host's agent factory. The members are not checked, so a backend declaring plain strings still matches: Capability and Egress are StrEnum.

maf_sandbox.testing.InProcessSandboxBackend lost its capabilities=, limits=, egress_modes= and os_families= keyword arguments, replaced by one declarations=. Override with dataclasses.replace(FAKE_BACKEND_DECLARATIONS, ...) rather than constructing a bare BackendDeclarations: the fake's default states egress_modes={ALLOWLIST, CLOSED} so a workload under test attaches, and a bare object resets it to the rule that enforces nothing.

import dataclasses

from maf_sandbox import DEFAULT_CAPABILITIES, Capability
from maf_sandbox.testing import FAKE_BACKEND_DECLARATIONS, InProcessSandboxBackend

# was: InProcessSandboxBackend(capabilities=DEFAULT_CAPABILITIES | {Capability.FILES_OUT})
# is:
InProcessSandboxBackend(
    declarations=dataclasses.replace(
        FAKE_BACKEND_DECLARATIONS, capabilities=DEFAULT_CAPABILITIES | {Capability.FILES_OUT}
    )
)

A backend says a delete failed by returning, not by raising. dispose is contractually best-effort and never raises, so the refusal 0.23 shipped — a key held closed until its disposal lands — could never fire against a compliant backend: each swallowed its delete error, said nothing, and was read as having disposed. Both disposal methods now carry the answer back:

The code is the contract; the detail is not. DisposalCode is a closed set — unreachable, timeout, refused, unlisted, unknown — and it is what a caller acts on: retry an unreachable, raise the bound on a timeout, put a refused in front of a human, since it is a missing role far more often than anything transient. detail is the backend's own sentence, for a log, never to be parsed.

async def dispose(self, key: SandboxKey) -> DisposalFailure | None:
    try:
        gone = await self._client.delete(key)
    except TransportError as exc:                     # never reached the service
        return DisposalFailure("unreachable", f"{key}: {exc}")
    return None if gone else DisposalFailure("refused", f"{key}: the service kept it")

Reach for unknown rather than guessing between the others. A code chosen to look precise is worse than one that admits the backend cannot tell, because a caller branches on it either way. Several failures fold to the most actionable code — fold_disposal_failures — keeping every detail.

A third-party backend must return the new type. dispose's reason was a str; wrap it in a DisposalFailure with the code that fits. dispose_scope changes shape too: return ScopePurge(count) where you returned count.

A caller reading the count reads .disposed. Watch for if await purger.purge_scoped_thread(...): a ScopePurge is always truthy where the count it replaced was not. router.scope(...)'s record gains undisposed beside disposed, which is additive.

None means nothing was reported, not that the delete provably happened — a backend with no way to check returns it too. The conflation is with success on purpose: the alternative refuses every key served by a backend that cannot answer. Say something whenever the delete is known not to have landed, and the router will refuse the key and quote you in SandboxUnclean.

Upgrading to 0.25

The dispatch spelling is gone. 0.24 added the host_tool_call names beside the old ones so a dependent could move in its own release; this removes what was kept.

Pin maf-sandbox<0.25 to stay, or rename: an old import is an ImportError, an old attribute an AttributeError, and an old keyword a TypeError — each naming what it wanted, none of them silent.

Upgrading to 0.23

Sandbox gains reclaim, and there is no declare-or-raise escape for it. Every other protocol method a backend cannot serve may raise NotImplementedError as long as it does not declare the matching Capability — the run_code note below is that escape, spelled out. reclaim is the first member with no capability behind it: there is no spec the router could refuse before a caller arrives, and a sandbox answering with NotImplementedError leaks a directory per call out of a finally that reports rather than raises. A third-party backend adds one method, and has to genuinely implement it:

async def reclaim(self, directory: str, *, working_directory: str, timeout: float) -> None:
    ...

Three rules a caller depends on, all in the docstring. The caller created it — under working_directory, with an unguessable name — so no filesystem path check is owed before removing it: there is no attacker-chosen component to check. That is what makes reclaim payable on a backend that must refuse remove: maf-sandbox-wslc runs the check on a model-supplied path but has it answered inside the guest being confined, which an irreversible delete may not rest on, and still serves reclaim honestly, because this one owes no confinement to begin with. A directory that is not there is success. Cleanup runs in a finally and must not report a second failure over the first. Anything else raises, so the caller can escalate.

working_directory says where the directory sits; it is not a directory to run the removal from. No backend creates a spec's work_dir, so a call that wrote nothing leaves it absent, and a removal that moved there first would fail over a directory that is already gone. The target is absolute, so cwd decides nothing.

Docker, wslc and ACAS implement it as rm -rf over their own exec — for ACAS deliberately not the data-plane delete_file its remove uses, because whether that service follows a link on delete is unverified. The in-process fake removes the entries from its store and records the call, so a test can assert the cleanup ran.

A sandbox the framework could not clean is disposed, by default. When a tool call's guest path could not be removed, or a program the transport stopped may have left something running — a SignalReach of "program" or "nothing" — sandboxed_tool now disposes that sandbox from the same finally, before the host is told. Better a failed run than leaked data: acquire is get-or-create, so a sandbox left warm hands the next call everything the last one could not take back. on_reclaim_failure still fires, after the disposal, and ReclaimFailure.disposal says what happened — "disposed", "failed", or "kept". A disposal that does not land makes the router refuse the key with SandboxUnclean until one does; SandboxToolSession.acquire turns that into a refusal the model reads. The opt-down is the host's, on the router, beside min_isolation — SandboxRouter(backends, reclaim=ReclaimConfig(failed_reclaim_policy=FailedReclaimPolicy.KEEP)) keeps the old behaviour — and a kind cannot lower it. A host whose callback disposed the sandbox itself can drop that code; a host that relied on a warm sandbox surviving a failed cleanup opts down explicitly. reclaim_timeout now bounds three legs — removal, disposal, report — so a failing call can cost up to three times it.

Upgrading to 0.20

Sandbox gains run_code, and Sandbox is a runtime_checkable Protocol — so a sandbox implementation that does not define it stops satisfying the protocol. isinstance(x, Sandbox) returns False and a type checker rejects it wherever a Sandbox is expected. Every backend published here answers it already, as of the previous release. A third-party backend adds one method:

async def run_code(self, code: str, *, timeout: float) -> ExecResult:
    raise NotImplementedError("this backend does not support RUN_CODE")

That is the whole migration unless you declare Capability.RUN_CODE, in which case implement it: it is the method that capability names, as exec is the method EXEC names. timeout is wall-clock from the call, so a backend that serialises calls on one sandbox spends part of it queued rather than leaving the waiting half unbounded, and a deadline that expires before the program starts raises SandboxQueuedTimeout rather than a plain TimeoutError — the caller's next move differs, retry unchanged versus make the program smaller.

The fake in maf_sandbox.testing answers it too: InProcessSandbox.run_code records each program in programs and matches outputs against the code as a substring, exactly as exec matches a command line. The two lists are separate on purpose — a test asserting a program was evaluated should not be satisfied by a shell command that happens to contain the same text.

The compatibility shim for a backend's old single egress property is gone. 0.19 read egress through a shim when egress_modes was absent; 0.20 does not, so a backend declaring only egress is now refused as undeclared — it enforces nothing the router can see. Declare egress_modes: frozenset[Egress], the set of modes it can actually enforce.

Capability.NETWORK is removed from the capability enum. No backend declared it and no spec required it; how precisely egress is confined lives in Egress.

New, and additive: assert_egress_conformance checks that the mode you declare is the mode you enforce. The two releases above made a backend say what it confines; this is the probe that holds it to the claim, in maf_sandbox.conformance beside the four suites already there. Give it a subject acquired with Egress.ALLOWLIST and that allowlist, plus one URL on the list and one off it, and it asserts the only outcome every allowlist backend shares — the allowed host answers, the denied host does not. Nothing calls it for you: a backend that declares allowlist and never runs it is exactly as it was before.

New, and additive: a workload can state the guest shape it needs. OsFamily is posix or windows; a backend declares os_families: frozenset[OsFamily] and a spec asks with requires_os_family, refused at attach with SandboxOsFamilyNotSupported on a mismatch. Nothing existing changes — a spec that asks nothing is refused by nothing, and a backend that declares nothing serves every spec that does not ask. The axis is path grammar and argv quoting and nothing else: a spec asking for posix and getting it can still meet an image with no shell, because what is installed in a guest is a property of the image, and one backend may be handed many. See docs/sandbox/guest-platform-and-commands.md.

Upgrading to 0.19

A workload now declares the egress mode it runs in, and the router refuses any backend that cannot enforce exactly that mode. SandboxSpec gains egress: Egress, defaulting to Egress.CLOSED, and a backend declares egress_modes: frozenset[Egress] — the set it can actually enforce — in place of the single egress property, which is removed. ensure_can_serve serves the workload iff spec.egress is in that set.

The tolerance is gone, and this is the change most likely to break a working deployment. Until 0.19 a backend that confined more than the spec asked was permitted with a warning — a closed backend served an allowlist spec, and the workload simply failed at whatever it could not fetch. That is now a refusal:

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

Neither direction is substituted any more: confining less silently widens what the workload reaches, and confining more hands it a posture it was not built for. If you see this, either give the backend a mode it can enforce — for maf-sandbox-docker that means configuring egress_proxy_image, without which it declares {closed} alone — or ask the kind for the mode you actually have, e.g. bicep_sandbox_spec(egress=Egress.CLOSED).

egress_allow without Egress.ALLOWLIST is refused at construction. A host list is the payload of an allowlist run, so SandboxSpec(kind=…, egress_allow=("example.invalid",)) now raises ValueError unless egress=Egress.ALLOWLIST goes with it. Kinds set both together, so this reaches you only if you build a spec by hand.

Capability.NETWORK is removed. No backend ever declared it and no spec ever required it; how precisely egress is confined lives in Egress, which is where it always was.

Every shipped backend replaced egress with egress_modes in the same release. A host that read backend.egress directly gets an AttributeError; read backend.egress_modes instead. Move core and the backends in the same step — an older backend under a 0.19 router is read through a compatibility shim and still resolves, but a 0.19 backend under an older router declares nothing the old router can see and is refused as undeclared.

Upgrading to 0.18

Sandbox.write_file takes a keyword-only working_directory, and refuses more than it used to. The path is resolved against it and then refused if it escapes, passes through a symlinked parent, lands on a symlink, or names the working directory itself — so a write that used to land can raise ValueError or NotADirectoryError, and every backend implementation has to declare the parameter. A version mismatch is invisible to an import check and surfaces at the first call: an older backend under a 0.18 caller raises TypeError: … got an unexpected keyword argument 'working_directory', and a 0.18 backend under an older caller raises missing 1 required keyword-only argument. Move core and the backends in the same step.

A tool call owns a guest path, and the framework reclaims it. New in 0.18.0, so maf-sandbox>=0.18.0 is the floor that gets it. SandboxToolSession.guest_call_path() names a place under work_dir allocated once per call, and sandboxed_tool removes it and everything under it when the call returns — after a result, a refusal and an exception alike — handing a removal that did not happen to the host's on_reclaim_failure as a ReclaimFailure. A kind that adopts it drops its own reclaim_run call and the finally around it; a kind that composes its own path keeps both, and keeps today's behaviour.

Upgrading to 0.17

A host-tool-call run's transport files are deleted now, and a kind has one call to make for the rest. host_tool_calls_over_exec removes its own directory on every exit path; the run directory — the model's shared-in files and its artifacts — is reclaim_run(sandbox, layout), which a kind calls in a finally after collecting. A kind that composes its own run directory and does not call it keeps today's behaviour for that half, so nothing breaks. If your kind reads anything out of the transport's directory after host_tool_calls_over_exec returns, it will no longer be there; nothing shipped here does.

Every host-tool-call run now issues one more exec, and a timed-out one may issue two. The extra calls are rm -rf for the cleanup, which every path pays, and kill -KILL for the stop, which only a run holding a usable pid issues — an upload that ran out never started a program, and a run whose pid is missing or unreadable has nothing to aim at. A backend double that counts exec calls, or a guest whose commands are allowlisted, will see them — the fakes in this repository's own codeact suite had to be taught to ignore them. A run's wall clock can exceed its timeout, by five separate graces. Every path pays the reclaim's _RECLAIM_GRACE (10s). A run whose last host-tool call returned after the bound pays _RESPONSE_WRITE_GRACE (2s) to record the answer, because a tool that has already acted is owed the round trip that says what it did. A run that overran pays three more _FINAL_READ_GRACE (2s each): one for the last look at the exit marker and the program's output, one for the pid lookup, and one for the signal itself — each measured fresh, because a slow guest that spends one must not leave the next with nothing. That is timeout + 18s with today's constants — the transport's own overhead, not an upper bound on the call. A host-tool call is deliberately never cancelled, so one that blocks holds the supervisor for as long as it blocks; bounding that belongs to the tool, which is the only code that knows what it is waiting on. Size an outer deadline from timeout + 18s plus whatever your slowest registered tool can take. One set any tighter loses the SandboxProgramTimeout and its output, and cancels whatever host-tool call is in flight — a tool's effect half applied with no record written, which is the trade an outer asyncio.wait_for makes on your behalf.

GuestRunLayout gained a pid field, and constructing one yourself is a TypeError until you pass it. guest_run_layout fills it in, so a kind that uses the factory — which is every kind that follows the documented path — needs no change at all. The field is where the launcher records the program's process id, which is what lets a run that overruns be stopped instead of left going.

Two more names are reserved in a run's transport directory: program_pid and program_pid.part. guest_run_layout refuses a program named for either, on the same grounds it already refuses program_exit_code.part — the launcher writes them, so a program under one of those names is written over. This is about names, not reach: a model-supplied file name cannot collide with them, because a kind writes model-named files only into work/. A program can still open anything it likes by absolute path — the shim sits where model-written code can read and edit it, and so does everything beside it.

A timed-out host-tool-call run now signals the guest program, where before it left it running. A run that reached the program gained a clause saying whether the signal was sent, so a host matching the old text no longer matches those. Only one message for a run that never got that far is unchanged, the launcher upload running out. The launcher's own exec running out with no pid gained a clause too, because the launcher backgrounds the interpreter before it writes the pid down: a call that expires between the two leaves a program running and no pid to point at, so that message now says the start could not be established rather than quietly implying none happened. If your host disposes the sandbox on every SandboxProgramTimeout to reclaim the CPU, keep doing that if you need the program actually gone. The message distinguishes a signal that was sent from one that was not, which is less than it sounds: a sent signal can be discarded, aimed at a pid the program rewrote, or leave children running, so it is not confirmation of termination and disposal is still the only thing that is. The exit marker's meaning is unchanged: the launcher waits on the program, so the code recorded is still the program's own.

Upgrading to 0.16

A host-tool run is two guest directories now, and the program's working directory is the new one. GuestRunLayout gained a work field; program, shim, launcher, output, exit_code and calls all moved from <run>/ into <run>/host_tools/. A kind that shared files into layout.directory and collected artifacts from it must use layout.work for both — this is the failure worth checking for, because nothing raises: guest_run_layout still takes the same arguments, so a kind that never named the moved paths keeps running, the program's open("input.csv") fails inside the guest, and an artifact written to the program's own working directory lands where the old collection path does not look. A run that quietly produces nothing is the symptom. Constructing GuestRunLayout yourself is the loud half — the new field makes it a TypeError.

A Python module shared into the work directory is no longer importable, and that is the second silent one. In 0.15 the program sat among the model's files, so sys.path[0] was the run directory and a kind could share helper.py beside it and have the program import helper. The program now runs from host_tools/, sys.path[0] follows it there, a working directory is never added to sys.path, and the launcher drops the inherited path entries that could put the work directory back — so the same import is a ModuleNotFoundError. If your kind shares Python rather than data, the program has to opt in:

import maf_host_tools  # first, so the real shim is in sys.modules
import os, sys
sys.path.insert(0, os.getcwd())  # now the work directory is importable
import helper

Order matters: once the work directory is on the path, a model-written file can answer any import that follows, which is what the split exists to prevent. Importing the shim first is what keeps that one safe — it is already in sys.modules and cannot be shadowed afterwards. Sharing the helper into host_tools/ instead is not an alternative; that directory is the transport's, and a name that collides with it is refused.

The split is what replaced the reserved-filename list this release was originally going to export. Two directories mean nothing a model can name reaches the transport's own files, so there is no list to keep complete, and the shim can no longer be shadowed by a guest file of the same name — sys.path[0] follows the program, which now sits beside the shim rather than among the model's files.

guest_run_layout refuses inputs it accepted in 0.15, and the first two refusals are new constraints rather than newly-enforced old ones. A run_directory containing : is rejected: the shim's directory now travels to the guest through PYTHONPATH, which separates on : and cannot quote one, so such a path would reach the interpreter as two entries — the second of them relative, resolved against the directory the guest writes into. If your run directories embed a timestamp, /runs/2026-08-17T10:30:00Z is the shape that stops working. And a program name is refused when the module it would answer to matches the shim's own module name (maf_host_tools.so and friends: the stem is reserved for the shim, because a file under it either shadows the import the program opens with or cannot run as a program), a module the generated shim imports (json, os, time), or one CPython imports at startup (encodings, site, sitecustomize, usercustomize, plus — reached through ordinary path lookup on a guest older than 3.11, whose standard library is not frozen, and refused everywhere because the guest's interpreter is not this package's to pin — abc, codecs, genericpath, io, posixpath, stat, _collections_abc, _sitebuiltins, _bootlocale) — the program shares a directory with the shim and that directory is on the path from startup, so such a name is imported instead of the module it stands for, or runs before the program does. One exact filename joins the list 0.15 already refused: program_exit_code.part, where the launcher stages the exit code before renaming it into place — that one is an old constraint newly enforced, since a program under it was truncated and renamed away by the launcher's last line in 0.15 too.

The launcher rewrites the guest's PYTHONPATH. It prepends the shim's directory and keeps an inherited entry only when it is absolute, canonical, and outside the run directory. A relative one resolves against the working directory — which the launcher has just changed to the guest's own — so an image that relies on . or a relative entry loses it here. An absolute one is dropped when it names the run directory or anything under it: nothing an image meant to name can live there, because the directory did not exist when the image was built, so an entry that does name it is either a coincidence of layout or an attempt to make the guest's own files importable at interpreter startup. /runs/current-sibling is kept when the run is /runs/current; only the tree itself goes. Every other absolute entry is passed through unchanged, including any that contain glob characters.

An entry carrying /./, /../ or // is dropped whatever it names. The comparison above is textual, so /runs/./current/work is a different string from /runs/current/work and the same directory to the interpreter. Such an entry is refused rather than normalised — an entry this cannot compare against the run tree is one it cannot vouch for. If your images export a path spelled that way, spell it canonically or it will not reach the guest.

The launcher also sets PYTHONNOUSERSITE=1, so user site-packages are off inside a run. PYTHONPATH is not the only inherited way into startup: site adds $PYTHONUSERBASE/lib/pythonX.Y/site-packages, and a sitecustomize there runs before the program exactly as one on the path would. Filtering that variable alone would leave the same hole behind HOME, which the user base falls back to, so the mechanism goes off rather than being chased through its inputs. If your image installs dependencies with pip install --user, they stop resolving inside a host-tool run — install them into the system environment instead. The failure is an ImportError naming the module, not a silent one.

What none of this closes is a symlink from outside the run tree into it, which needs a realpath POSIX sh does not have, and PYTHONSAFEPATH does not help with any of it — sitecustomize runs before any script. If your images export PYTHONPATH or PYTHONUSERBASE at all, keep them clear of wherever your kind places run directories.

host_tool_calls_over_exec raises SandboxProgramTimeout for its own bound. A TimeoutError from it used to mean either the run running out or a backend bounding one of its own calls, and callers could not tell which. The new type — a TimeoutError subclass, so existing handlers keep working — is the first. A bare TimeoutError is the second, and says nothing about whether the program is still running — validation errors and whatever a backend raises for its own reasons come through as themselves, unchanged. It carries the program's partial output on output, and what the transport managed to do about the program on signal — sent, refused, absent, unrecorded, unknown. Branch on signal, not on the message text, which is prose and will keep moving. Only absent says nothing was started; none of the others confirms the program was stopped, so a host that needs it gone still disposes the sandbox. Raising this type yourself reports unknown unless you say otherwise — nothing claims absent but the leg that never reached a launcher.

There is a new rung, os_process, between runtime and container. A separate OS process is a real boundary — a kernel-enforced address space — and a weaker one than a container, which is a process plus namespaces and cgroups. It exists so that a backend running untrusted code in a subprocess has something honest to declare instead of understating itself as runtime or overstating itself as container. No backend in this repository provides it; this release is vocabulary.

Isolation.PROCESS is back as a name, and it means the new rung. If you upgraded through 0.14 you have already made the edit this needs: the old Isolation.PROCESS meant no boundary and is now Isolation.NONE. If you are coming from 0.13 or earlier, read the 0.14 note below first — jumping the version where the old spelling raises is the one path on which this rename is quiet.

The value is "os_process", not "process", and Isolation("process") still raises ValueError. Reusing the attribute is safe because Python resolves it where you wrote it. Reusing the string would not be: a declaration reaches this vocabulary through Isolation(raw) at run time, out of configuration nobody re-reads, so the old spelling would have come back ranked two rungs higher having claimed a boundary it never drew. It is refused instead, and it will stay refused.

Rank numbers shifted; comparisons did not. Inserting a rung renumbers everything above it — container moved from 2 to 3, and so on up. Nothing needs to change if you compare rungs with meets_floor or through ISOLATION_RANK, which is the only ordering there is. If you persisted a rank integer anywhere, it now names a different rung.

Upgrading to 0.14

Isolation.PROCESS is Isolation.NONE. The rung that provides no boundary was named for where the code runs rather than for what it protects, and read as the opposite of what it meant — "process isolation" implies a boundary, and this rung is the absence of one. One mechanical edit, in host code and in any backend you have written.

The old spelling is removed outright rather than kept as an alias, and that is the point. PROCESS is reserved for a genuine separate-OS-process rung — a kernel-enforced address space, sharing the kernel and the filesystem — which landed between runtime and container in 0.16, carrying the value "os_process". An alias would have made that reuse silent: a backend declaring "process" because it drew no boundary would come back ranked two rungs higher, having claimed one, and a host running min_isolation=Isolation.RUNTIME would begin admitting it. So in this release Isolation.PROCESS raises AttributeError, Isolation("process") raises ValueError, and a backend still declaring it is refused at construction with SandboxBackendNotPermitted. The failure is the notice.

Check your configuration, not only your code. Isolation is a StrEnum, so a floor or a declaration may reach the router as the string "process" out of a config file or an environment variable rather than as an attribute. Those fail the same way and at the same moment — but a grep for Isolation.PROCESS will not find them.

Upgrading to 0.11

0.11.0 retired the word workspace from the public vocabulary. It was carrying three unrelated things, and only one of them keeps the stem. Two edits, both mechanical.

WorkspaceContext is CallerContext, and make_workspace_context is make_caller_context. The type was never a storage concept: two of its three fields are identity, and list_files receives a store rather than holding one. Its first parameter is now list_files where it was store_walker — a positional call needs no edit, a keyword one does.

work_dir and working_directory are unchanged. They name the guest's working directory, they are the most common use of the stem by an order of magnitude, and they were never the concept being retired. If you were looking for a rename here, there isn't one.

The dependent packages moved with it: maf-sandbox-bicep and maf-sandbox-codeact take file_store where they took workspace_store, and each has its own note.

Upgrading from 0.4.x

0.5.0 replaced the deployed boolean with a declared isolation floor, and added a capability axis. Four changes need an edit; nothing else moves.

SandboxRouter(..., deployed=...) is gone — pass min_isolation instead. deployed=True becomes min_isolation=Isolation.MICROVM, which is also the default, so a deployed host can drop the argument entirely. deployed=False on a developer machine becomes the rung that host actually accepts, stated explicitly — min_isolation=Isolation.CONTAINER for a container backend, Isolation.NONE for an in-process fake. There is no longer a value meaning "anything goes": a host that wants the weakest rung names it.

DEPLOYED_ISOLATION is removed. The policy it expressed is min_isolation's default.

Isolation and Egress are StrEnums, and the ladder grew. Values are unchanged, so backend.isolation == "vm" and any stored configuration keep working. The ladder is now process < runtime < container < hardened_container < microvm < vm; a declared value outside it is refused at construction rather than silently permitted. (The bottom rung was renamed to none in 0.14 — see Upgrading to 0.14 above. This note describes the ladder as 0.5.0 shipped it.)

AcasSandboxBackend now declares microvm, not vm. ACA Sandboxes are hardware-isolated micro-VMs; vm now means a dedicated, full VM on remote infrastructure. A host that pinned min_isolation=Isolation.VM expecting ACA Sandboxes to satisfy it should use Isolation.MICROVM — the default, and the rung the micro-VM standard defines.

A backend that states no capabilities field is read as declaring DEFAULT_CAPABILITIES (exec + files_in), which is what the Sandbox protocol already obliges. Declare a wider set to serve workloads that require more.

Writing a backend

Implement name, isolation, acquire, dispose, dispose_scope, and a declarations holding a BackendDeclarations. The object is optional and every field in it has a default, but egress_modes is the one you must state — silence there is the empty set, and the router refuses every ask. One migration rule the router also enforces: do not leave any of the four declaration fields as a bare attribute on the backend — they were attributes before 0.27, the router refuses a backend that still carries one, and a stray attribute is read by nothing and silently replaced by that field's default.

The ordered path through the rest — each Sandbox method with what it owes, what to reach for, what never to do, and the probes that prove it; the bundle menu and the three things the stat it runs must be; why reclaim has no declare-or-raise escape; the acquire race, the label rules and the disposal contract; the six assert_*_conformance suites against a real instance — is docs/sandbox/backends/writing-a-backend.md, with the shipped backends' declarations beside it in docs/sandbox/backends/README.md.

Provenance

Extracted from a production agent application, where this seam was written for its first execution surface: a tool that compiles agent-authored infrastructure code in a sandbox. The minimum-isolation floor above is not a preference — it is what a security review concluded when it worked through what a shared-kernel boundary does not close for code an agent wrote.

Maintained by SOKOLAI BV.