arl-env Python SDK

High-level Python SDK for the agent-env Gateway API. The SDK creates sandbox sessions, runs commands, streams output, transfers files, opens interactive shells, and manages SandboxWarmPool resources through the gateway.

Release and Migration

Installation

pip install arl-env
# or
uv add arl-env

Interactive shell support needs the optional dependency:

pip install "arl-env[shell]"

Authentication

If gateway authentication is enabled, provide a bearer API key through the environment or constructor:

export ARL_API_KEY="your-api-key"

from arl import SandboxSession

session = SandboxSession(
    image="python:3.12",
    gateway_url="http://localhost:8080",
    api_key="your-api-key",
)

Basic Usage

from arl import SandboxSession

with SandboxSession(
    image="python:3.12",
    gateway_url="http://localhost:8080",
) as session:
    result = session.execute([
        {"name": "hello", "command": ["echo", "Hello, World!"]},
    ])
    print(result.results[0].output.stdout)

Commands run in the executor container, which uses the requested image. The sidecar only exposes the gRPC control plane and proxies execution to the executor-agent over a Unix socket.

Profile Semantics

profile is a pool-selection key, not a resource specification. The gateway uses it to choose which SandboxWarmPool can satisfy a session request.

The caller does not choose a Kubernetes namespace. The gateway is deployed with one namespace scope, and all session and pool operations use that scope.

For a session request, selection works as follows:

Empty profile is normalized to default.
If image is provided, the gateway looks for a scoped pool with the same image and profile. If none exists, it creates an image-backed pool for that pair.
If only profile is provided, the gateway selects an existing scoped pool with the same profile.
When several pools match, the gateway picks the one with the most available warm capacity.

The value does not create CPU, memory, GPU, or scheduling behavior by itself. Those come from how the matching pool was created. A profile named gpu only means "select pools labeled gpu"; it should point to pools that were actually created with GPU resources.

Common patterns:

# Use the default profile. The gateway creates/reuses an image-backed pool.
session = SandboxSession(image="python:3.12")

# Select a pre-created profile. The selected pool determines the image.
session = SandboxSession(profile="gpu")

# Require both the image and the profile to match.
session = SandboxSession(image="python:3.12", profile="large-memory")

Use stable short names such as default, cpu, gpu, large-memory, or a pool family name such as python-pool. Keep the same value on pool creation and session creation when the session should target that pool family.

Persistent Sessions

Use manual lifecycle management when several operations should share the same workspace. A context manager deletes the session on exit; close() only closes the local HTTP client and leaves the remote session available for reattach. Always call delete_sandbox() when the session is no longer needed.

from arl import SandboxSession

session = SandboxSession(image="python:3.12", gateway_url="http://localhost:8080")
session.create_sandbox()
session_id = session.session_id
session.execute([
    {"name": "init", "command": ["sh", "-c", "echo 0 > /workspace/count.txt"]},
])
session.close()  # detach; the remote session remains active

Attach to an existing session:

from arl import SandboxSession

session = SandboxSession.attach(session_id, gateway_url="http://localhost:8080")
try:
    result = session.execute([{"name": "read", "command": ["cat", "/workspace/count.txt"]}])
    print(result.results[0].output.stdout)
finally:
    session.delete_sandbox()
    session.close()

Streaming Output

execute() uses the gateway SSE endpoint when available. Pass on_output to receive stdout/stderr chunks while the step is still running.

def on_output(stdout: str, stderr: str) -> None:
    if stdout:
        print(stdout, end="")
    if stderr:
        print(stderr, end="")

result = session.execute(
    [{"name": "loop", "command": ["sh", "-c", "for i in 1 2 3; do echo $i; sleep 1; done"]}],
    on_output=on_output,
)

File Transfer

Paths are relative to the session workspace.

session.upload_file("input.txt", "hello\n")
data = session.download_file("input.txt")

session.upload_path("local.bin", "data/local.bin")
session.download_path("data/local.bin", "out/local.bin")

History, Restore, and Trajectory

Each executed step is recorded in session history. Snapshot IDs are step-index strings used by the gateway's replay-based restore implementation.

r1 = session.execute([{"name": "write", "command": ["sh", "-c", "echo one > /workspace/x"]}])
snapshot_id = r1.results[0].snapshot_id

session.execute([{"name": "change", "command": ["sh", "-c", "echo two > /workspace/x"]}])
session.restore(snapshot_id)

history = session.get_history()
jsonl = session.export_trajectory()

WarmPool Management

WarmPoolManager uses the gateway pool endpoints. Pool creation is an admin operation when gateway auth is enabled.

from arl import ResourceRequirements, WarmPoolManager

manager = WarmPoolManager(gateway_url="http://localhost:8080")
manager.create_warmpool(
    name="python-pool",
    image="python:3.12",
    profile="python-pool",
    replicas=2,
    resources=ResourceRequirements(
        requests={"cpu": "500m", "memory": "512Mi"},
        limits={"cpu": "1", "memory": "1Gi"},
    ),
)
info = manager.wait_for_ready("python-pool", min_ready=1)
print(info.ready_replicas)
manager.scale_warmpool("python-pool", replicas=3)
manager.delete_warmpool("python-pool")   # drain sessions/claims and scale to zero
# manager.destroy_warmpool("python-pool")  # physically delete the WarmPool/template

Current sandbox-backed sessions support claim-scoped config_env.vars and string-valued config_env.envVars environment injection. Pool creation still rejects config_env, and sandbox-backed sessions still reject tools provisioning and config_env ConfigMap/Secret/valueFrom provisioning requests. list_tools() and call_tool() only work when the executor image already contains /opt/arl/tools/registry.json and matching tool files.

Managed Sessions

ManagedSession creates or reuses a server-side managed pool for an image and groups sessions by experiment ID.

from arl import ManagedSession

with ManagedSession(
    image="python:3.12",
    experiment_id="exp-1",
    gateway_url="http://localhost:8080",
) as session:
    result = session.execute([
        {"name": "hello", "command": ["python", "-c", "print('ok')"]},
    ])
    print(result.results[0].output.stdout)

Clean up an experiment:

from arl import GatewayClient

client = GatewayClient(base_url="http://localhost:8080")
deleted = client.delete_experiment("exp-1")

Core Classes

SandboxSession: session lifecycle, execute, replay, restore, files, logs, history, trajectory.
ManagedSession: image + experiment session flow with server-side pool creation.
GatewayClient: low-level HTTP client for all public gateway REST endpoints.
WarmPoolManager: pool create/list/get/wait/scale/logs/drain/destroy helpers.
InteractiveShellClient: WebSocket shell client.

Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

arl_env-0.21.5.tar.gz (56.5 kB view details)

Uploaded Jul 19, 2026 Source

Built Distribution

If you're not sure about the file name format, learn more about wheel file names.

The dropdown lists show the available interpreters, ABIs, and platforms. Enable javascript to be able to filter the list of wheel files.

arl_env-0.21.5-py3-none-any.whl (54.0 kB view details)

Uploaded Jul 19, 2026 Python 3

File details

Details for the file arl_env-0.21.5.tar.gz.

File metadata

Download URL: arl_env-0.21.5.tar.gz
Upload date: Jul 19, 2026
Size: 56.5 kB
Tags: Source
Uploaded using Trusted Publishing? Yes
Uploaded via: uv/0.11.29 {"installer":{"name":"uv","version":"0.11.29","subcommand":["publish"]},"python":null,"implementation":{"name":null,"version":null},"distro":{"name":"Ubuntu","version":"24.04","id":"noble","libc":null},"system":{"name":null,"release":null},"cpu":null,"openssl_version":null,"setuptools_version":null,"rustc_version":null,"ci":true}

File hashes

Hashes for arl_env-0.21.5.tar.gz
Algorithm	Hash digest
SHA256	`c3a9defa974cb3eefb8cf271605564450511d8d41c9dd53805294dbb5c2e112d`
MD5	`37e3721bd42b0f67c920fa7b3722657a`
BLAKE2b-256	`5d176195c4ad387cb5825cb0b3ceca6f40fe663067193d7f0f22656693ef8dde`

See more details on using hashes here.

File details

Details for the file arl_env-0.21.5-py3-none-any.whl.

File metadata

Download URL: arl_env-0.21.5-py3-none-any.whl
Upload date: Jul 19, 2026
Size: 54.0 kB
Tags: Python 3
Uploaded using Trusted Publishing? Yes
Uploaded via: uv/0.11.29 {"installer":{"name":"uv","version":"0.11.29","subcommand":["publish"]},"python":null,"implementation":{"name":null,"version":null},"distro":{"name":"Ubuntu","version":"24.04","id":"noble","libc":null},"system":{"name":null,"release":null},"cpu":null,"openssl_version":null,"setuptools_version":null,"rustc_version":null,"ci":true}

File hashes

Hashes for arl_env-0.21.5-py3-none-any.whl
Algorithm	Hash digest
SHA256	`fb1972526d37b73c9bb7b03d7bb0f3ee500745584b83c4300023c09552a19a79`
MD5	`c22847c7d7fa524ad4d21332f0630235`
BLAKE2b-256	`bf305911fa755c5aa46efe48a844084a816525502927ac6bbd19b76f46e5d51b`