Narzędzia powłoki

Pakiet Python w wersji beta agent-framework-tools zapewnia narzędzia do wykonywania powłoki i rozpoznawania środowiska za pośrednictwem agent_framework.tools przestrzeni nazw.

Narzędzie Użyj go, gdy
LocalShellTool Polecenia są uznawane za zaufane lub zatwierdzone pojedynczo i powinny działać w środowisku hosta, w którym działa proces agenta.
DockerShellTool Polecenia powłoki generowane przez model wymagają izolacji kontenera OCI.
ShellEnvironmentProvider Model potrzebuje typu aktywnej powłoki, systemu operacyjnego, katalogu roboczego oraz wersji zainstalowanych narzędzi CLI.
ShellPolicy Potrzebujesz filtra wstępnego opartego na liście dozwolonych lub liście blokowanych przed zatwierdzeniem lub wykonaniem.

Warning

Wykonywanie poleceń powłoki może modyfikować pliki, uruchamiać procesy, uzyskiwać dostęp do poświadczeń i komunikować się z systemami zewnętrznymi. Użyj poziomu wykonywania z najmniejszymi uprawnieniami, który umożliwia wykonanie tego zadania.

Instalowanie pakietu

dotnet add package Microsoft.Agents.AI.Tools.Shell --prerelease

Używanie lokalnej powłoki i świadomości środowiska

LocalShellExecutor obsługuje tryby bezstanowe i trwałe. ShellEnvironmentProvider bada aktywne środowisko i dodaje do kontekstu agenta wiarygodne wskazówki dotyczące powłoki.

using Azure.AI.Projects;
using Azure.Identity;
using Microsoft.Agents.AI;
using Microsoft.Agents.AI.Tools.Shell;
using Microsoft.Extensions.AI;

var endpoint = Environment.GetEnvironmentVariable("FOUNDRY_PROJECT_ENDPOINT") ?? throw new InvalidOperationException("FOUNDRY_PROJECT_ENDPOINT is not set.");
var deploymentName = Environment.GetEnvironmentVariable("FOUNDRY_MODEL") ?? "gpt-5.4-mini";

// WARNING: DefaultAzureCredential is convenient for development but requires careful consideration in production.
// In production, consider using a specific credential (e.g., ManagedIdentityCredential) to avoid
// latency issues, unintended credential probing, and potential security risks from fallback mechanisms.
var aiProjectClient = new AIProjectClient(new Uri(endpoint), new DefaultAzureCredential());

const string Instructions = """
    You are an agent with a single tool: run_shell. Use it to satisfy the
    user's request. Do not describe what you would do — actually run the
    commands. Reply with the final answer derived from real output.
    """;

// --------------------------------------------------------------------
// 1. Stateless mode — each call gets a fresh shell.
// --------------------------------------------------------------------
Console.WriteLine("### Stateless mode\n");
await using (var statelessShell = new LocalShellExecutor(new() { Mode = ShellMode.Stateless, AcknowledgeUnsafe = true }))
{
    var envProvider = new ShellEnvironmentProvider(statelessShell);
    var statelessAgent = aiProjectClient.AsAIAgent(new ChatClientAgentOptions
    {
        ChatOptions = new()
        {
            ModelId = deploymentName,
            Instructions = Instructions,
            Tools = [statelessShell.AsAIFunction(requireApproval: false)],
        },
        AIContextProviders = [envProvider],
    });
// --------------------------------------------------------------------
// 2. Persistent mode — one shell, reused across calls. State carries.
// --------------------------------------------------------------------
Console.WriteLine("\n### Persistent mode\n");
await using (var persistentShell = new LocalShellExecutor(new() { Mode = ShellMode.Persistent, AcknowledgeUnsafe = true }))
{
    var envProvider = new ShellEnvironmentProvider(persistentShell);
    var persistentAgent = aiProjectClient.AsAIAgent(new ChatClientAgentOptions
    {
        ChatOptions = new()
        {
            ModelId = deploymentName,
            Instructions = Instructions,
            Tools = [persistentShell.AsAIFunction(requireApproval: false)],
        },
        AIContextProviders = [envProvider],
    });

    var persistentSession = await persistentAgent.CreateSessionAsync();

    // State carries across calls in persistent mode: cd into temp, then
    // verify the next call sees the new CWD.
    Console.WriteLine(await persistentAgent.RunAsync("Change directory into the system temp folder, then print the current working directory.", persistentSession));
    Console.WriteLine();
    Console.WriteLine(await persistentAgent.RunAsync("In a NEW shell call, print the current working directory again. Tell me whether it still matches the temp folder.", persistentSession));
    Console.WriteLine();

    // Same idea with an exported variable: set in one call, read in the next.
    Console.WriteLine(await persistentAgent.RunAsync("Set the environment variable DEMO_TOKEN to the value 'hello-world'.", persistentSession));
    Console.WriteLine();
    Console.WriteLine(await persistentAgent.RunAsync("Print the current value of DEMO_TOKEN. Tell me exactly what value the shell reports.", persistentSession));
    Console.WriteLine();

    PrintSnapshot(envProvider.CurrentSnapshot!);
}

ShellPolicy jest również dostępny do wstępnego filtrowania poleceń. Dedykowany przykład DockerShellExecutor, który można uruchomić, nie jest obecnie opublikowany.

Instalowanie pakietu

pip install agent-framework-tools --pre

Pakiet jest instalowany psutil w celu zakończenia podrzędnych drzew procesów, gdy upłynął limit czasu wykonywania.

Użyj LocalShellTool

LocalShellTool uruchamia polecenia bezpośrednio na hoście. Domyślnie używa trwałej powłoki, 30-sekundowego limitu czasu, obcięcia danych wyjściowych do 64 KiB, ograniczenia do katalogu roboczego oraz wymagania zatwierdzenia każdego polecenia.

import asyncio
from typing import Any

from agent_framework import Agent, Message
from agent_framework.openai import OpenAIChatClient
from agent_framework.tools import LocalShellTool
from dotenv import load_dotenv

# Load environment variables from .env file
load_dotenv()
async def main() -> None:
    print("=== OpenAI Agent with LocalShellTool Example ===")
    print("NOTE: Commands will execute on your local machine.\n")

    client = OpenAIChatClient(model="gpt-5.4-nano")

    async with LocalShellTool() as shell:
        agent = Agent(
            client=client,
            instructions="You are a helpful assistant that can run shell commands to help the user.",
            tools=[client.get_shell_tool(func=shell.as_function())],
        )

        query = "Use the shell tool to execute `python --version` and show only the command output."
        print(f"User: {query}")
        result = await run_with_approvals(query, agent)
        if isinstance(result, str):
            print(f"Agent: {result}\n")
            return
        if result.text:
            print(f"Agent: {result.text}\n")
        else:
            printed = False
            for message in result.messages:
                for content in message.contents:
                    if content.type == "function_result" and content.result:
                        print(f"Agent (tool output): {content.result}\n")
                        printed = True
            if not printed:
                print("Agent: (no text output returned)\n")


async def run_with_approvals(query: str, agent: Agent) -> Any:
    """Run the agent and handle shell approvals outside tool execution."""
    current_input: str | list[Any] = query

    while True:
        result = await agent.run(current_input)
        if not result.user_input_requests:
            return result

        next_input: list[Any] = [query]
        rejected = False
        for user_input_needed in result.user_input_requests:
            if user_input_needed.function_call is None:
                continue
            print(
                f"\nShell request: {user_input_needed.function_call.name}"
                f"\nArguments: {user_input_needed.function_call.arguments}"
            )
            user_approval = await asyncio.to_thread(input, "\nApprove shell command? (y/n): ")
            approved = user_approval.strip().lower() == "y"
            next_input.append(Message("assistant", [user_input_needed]))
            next_input.append(Message("user", [user_input_needed.to_function_approval_response(approved)]))
            if not approved:
                rejected = True
                break
        if rejected:
            print("\nShell command rejected. Stopping without additional approval prompts.")
            return "Shell command execution was rejected by user."
        current_input = next_input


if __name__ == "__main__":
    asyncio.run(main())

Użyj polecenia mode="stateless", gdy każde wywołanie powinno zostać uruchomione w nowym procesie. Użyj zmiennej środowiskowej AGENT_FRAMEWORK_SHELL lub argumentu konstruktora shell , aby zastąpić rozpoznaną powłokę.

Important

LocalShellTool nie jest piaskownicą. Zatwierdzenie jest podstawową granicą zabezpieczeń. Wyłączenie zatwierdzania wymaga acknowledge_unsafe=True.

Ogranicz polecenia za pomocą polecenia ShellPolicy

ShellPolicy stosuje listy zezwoleń i blokad oparte na wyrażeniach regularnych przed wykonaniem. Reguły odmowy mają pierwszeństwo.

import asyncio

from agent_framework import Agent
from agent_framework.openai import OpenAIChatClient
from agent_framework.tools import LocalShellTool, ShellPolicy
from dotenv import load_dotenv
load_dotenv()
async def main() -> None:
    client = OpenAIChatClient(model="gpt-5.4-nano")

    shell = LocalShellTool(
        mode="stateless",
        approval_mode="never_require",
        acknowledge_unsafe=True,
        policy=ShellPolicy(
            allowlist=[
                r"^ls(\s|$)",
                r"^pwd$",
                r"^cat\s[^|;&]+$",
                r"^git\s+(status|log|diff)(\s|$)",
                r"^python\s+--version$",
            ],
        ),
        timeout=10,
    )

    agent = Agent(
        client=client,
        instructions=(
            "You can run a narrow set of read-only shell commands (ls, pwd, cat, "
            "git status/log/diff, python --version). Anything else will be rejected."
        ),
        tools=[client.get_shell_tool(func=shell.as_function())],
    )

    query = "Summarise the current directory and print the Python version."
    print(f"User: {query}")
    result = await agent.run(query)
    print(f"Agent: {result.text}")

Warning

Polityka poleceń jest wstępnym filtrem użyteczności, a nie granicą bezpieczeństwa. Składnia powłoki, aliasy, zmienne, interpretery i zakodowane ładunki mogą omijać proste mechanizmy dopasowywania wzorców.

Dodaj ShellEnvironmentProvider

ShellEnvironmentProvider sprawdza rodzaj powłoki, jej wersję, system operacyjny, katalog roboczy i wybrane wersje CLI, a następnie dołącza te informacje przed uruchomieniem agenta. Domyślna lista sond to git, node, python i docker.

import asyncio

from agent_framework import Agent
from agent_framework.openai import OpenAIChatClient
from agent_framework.tools import (
    LocalShellTool,
    ShellEnvironmentProvider,
    ShellEnvironmentProviderOptions,
)
from dotenv import load_dotenv
load_dotenv()
def _print_snapshot(label: str, provider: ShellEnvironmentProvider) -> None:
    snapshot = provider.current_snapshot
    if snapshot is None:
        print(f"[{label}] no snapshot captured")
        return
    print(f"\n[{label}] snapshot:")
    print(f"  family            = {snapshot.family.value}")
    print(f"  os                = {snapshot.os_description}")
    print(f"  shell_version     = {snapshot.shell_version}")
    print(f"  working_directory = {snapshot.working_directory}")
    for tool, version in snapshot.tool_versions.items():
        print(f"  {tool:<17} = {version}")


async def _ask(agent: Agent, query: str) -> None:
    print(f"\nUser: {query}")
    result = await agent.run(query)
    if result.text:
        print(f"Agent: {result.text}")


async def main() -> None:
    client = OpenAIChatClient(model="gpt-5.4-nano")
    options = ShellEnvironmentProviderOptions(
        probe_tools=("git", "python", "uv", "node"),
    )

    print("=== stateless mode ===")
    async with LocalShellTool(
        mode="stateless",
        approval_mode="never_require",
        acknowledge_unsafe=True,
    ) as shell:
        provider = ShellEnvironmentProvider(shell, options)
        agent = Agent(
            client=client,
            instructions="Use the shell tool to answer the user's question.",
            tools=[client.get_shell_tool(func=shell.as_function())],
            context_providers=[provider],
        )
        await _ask(agent, "Show me the current working directory.")
        await _ask(agent, "Now `cd ..` then show the working directory again.")
        await _ask(agent, "Show the working directory once more — did `cd` persist?")
        _print_snapshot("stateless", provider)

    print("\n=== persistent mode ===")
    async with LocalShellTool(
        mode="persistent",
        confine_workdir=False,
        approval_mode="never_require",
        acknowledge_unsafe=True,
    ) as shell:
        provider = ShellEnvironmentProvider(shell, options)
        agent = Agent(
            client=client,
            instructions="Use the shell tool to answer the user's question.",
            tools=[client.get_shell_tool(func=shell.as_function())],
            context_providers=[provider],
        )
        await _ask(agent, "Show me the current working directory.")
        await _ask(agent, "Now `cd ..` then show the working directory again.")
        await _ask(agent, "Show the working directory once more — did `cd` persist?")
        _print_snapshot("persistent", provider)

Użyj DockerShellTool

DockerShellTool wymaga Dockera lub Podmana na PATH. Ustawienia domyślne wyłączają sieć, uruchamiają kontener jako użytkownik inny niż root, używają głównego systemu plików w trybie tylko do odczytu, usuwają uprawnienia systemowe, ograniczają pamięć do 512 MiB i liczbę procesów w kontenerze do 256.

from agent_framework.tools import DockerShellTool

async with DockerShellTool(
    image="mcr.microsoft.com/azurelinux/base/core:3.0",
    approval_mode="never_require",
) as shell:
    result = await shell.run("uname -a && id")
    print(result.stdout)

Domyślnym obrazem jest mcr.microsoft.com/azurelinux/base/core:3.0. Przekaż docker_binary="podman" , aby użyć narzędzia Podman. Dedykowany przykład z możliwością DockerShellTool uruchamiania nie jest obecnie publikowany.

Wybierz warstwę wykonania

Scenario Narzędzie Granica izolacji
Zaufane polecenia programistyczne LocalShellTool Zatwierdzenie w procesie hostującym
Niezaufane polecenia powłoki DockerShellTool Kontener OCI z domyślnymi flagami izolacji
Niezaufany kod wygenerowany bez powłoki Hyperlight CodeAct mikroVM Hyperlight

Go zapewnia lokalne wykonywanie poleceń powłoki i sprawdzanie środowiska za pomocą tool/shelltool. Zobacz Używanie lokalnej powłoki.

DockerShellTool wskazówki nie są obecnie dostępne dla języka Go.

Korzystanie z narzędzi powłoki przy użyciu Harness Agent

Zwykli agenci i HarnessAgent używają tej samej dwuczęściowej konfiguracji powłoki: zarejestruj funkcję wykonawcy jako narzędzie i dodaj ShellEnvironmentProvider, gdy model powinien otrzymywać kontekst powłoki, systemu operacyjnego, katalogu roboczego i wersji CLI. HarnessAgent nie tworzy ani nie posiada wykonawcy powłoki:

using System.IO;
using Microsoft.Agents.AI;
using Microsoft.Agents.AI.Tools.Shell;
using Microsoft.Extensions.AI;

await using var shell = new LocalShellExecutor(new LocalShellExecutorOptions
{
    WorkingDirectory = Directory.GetCurrentDirectory(),
    Timeout = LocalShellExecutor.DefaultTimeout,
});

AIAgent agent = chatClient.AsHarnessAgent(new HarnessAgentOptions
{
    AIContextProviders = [new ShellEnvironmentProvider(shell)],
    ChatOptions = new ChatOptions
    {
        Tools = [shell.AsAIFunction(requireApproval: true)],
    },
});

AsAIFunction domyślnie ma nazwę run_shell i requireApproval: true. LocalShellExecutor domyślnie używa trybu trwałego, limitu 64 KiB na każdy strumień wyjściowy i braku limitu czasu; w przykładzie jawnie użyto zalecanego 30-sekundowego LocalShellExecutor.DefaultTimeout. ShellEnvironmentProviderOptions domyślnie sonduje git, dotnet, node, python i docker, z limitem czasu wynoszącym pięć sekund dla każdej sondy.

Utwórz jedną trwałą funkcję wykonawcza dla sesji użytkownika i zlikwiduj ją po zakończeniu sesji. Nie udostępniaj tego między użytkownikami ani w różnych równoległych rozmowach, ponieważ katalog roboczy, środowisko, historia powłoki, zadania w tle i kolejka poleceń są wspólne. ShellPolicy jest tylko filtrem wstępnym; Zachowaj włączone zatwierdzenie, używaj poświadczeń o najniższych uprawnieniach i preferuj DockerShellExecutor , gdy polecenia wymagają silniejszej granicy izolacji.

Narzędzia shella są dostępne w pakiecie przedpremierowym Microsoft.Agents.AI.Tools.Shell. HarnessAgent jest dostępny na stronie Microsoft.Agents.AI.Harness.

W przypadku zwykłego agenta utwórz funkcję shella za pomocą client.get_shell_tool(func=shell.as_function()) i dodaj ShellEnvironmentProvider oddzielnie. create_harness_agent wykonuje oba kroki po przekazaniu shell_executor:

from agent_framework import create_harness_agent
from agent_framework.tools import LocalShellTool, ShellEnvironmentProviderOptions

async with LocalShellTool() as shell:
    agent = create_harness_agent(
        client=client,
        shell_executor=shell,
        shell_environment_provider_options=ShellEnvironmentProviderOptions(
            probe_tools=("git", "python"),
        ),
    )

    session = agent.create_session()
    response = await agent.run("Inspect the current repository.", session=session)

shell_executor jest opcjonalne i musi udostępniać as_function(). Fabryka dodaje narzędzie powłoki i ShellEnvironmentProvider tylko wtedy, gdy klient implementuje SupportsShellTool; w przeciwnym razie rejestruje ostrzeżenie i pomija oba te elementy. shell_environment_provider_options jest opcjonalny i jest używany tylko z shell_executor.

Domyślnie LocalShellTool używa trybu trwałego, 30-sekundowego limitu czasu, łącznego wyjścia o rozmiarze 64 KiB, ponownego osadzenia katalogu roboczego i approval_mode="always_require". Ponieważ zatwierdzanie narzędzia Harness jest domyślnie włączone, przekaż AgentSession do run. Obiekt wywołujący zarządza cyklem życia egzekutora; użyj async with lub wywołaj close(), a dla każdej sesji użytkownika utwórz jedno trwałe narzędzie. Nie udostępniaj modyfikowalnego stanu powłoki między użytkownikami ani współbieżnych konwersacji.

Powłoka hosta nie jest piaskownicą. Pozostaw włączone zatwierdzanie, używaj poświadczeń o najmniejszych uprawnieniach i używaj DockerShellTool do izolacji kontenerów. Wyłączenie akceptacji wymaga approval_mode="never_require" i acknowledge_unsafe=True; ShellPolicy samo nie jest barierą zabezpieczeń.

create_harness_agent został wydany w agent-framework-core. Integracja powłoki jest dostarczana przez pakiet wstępny agent-framework-tools i emituje element ExperimentalWarning po włączeniu.

Pakietowa wersja Go Harness nie jest obecnie dostępna. Utwórz lokalne narzędzie powłoki i dostawcę środowiska bezpośrednio na zwykłym agencie języka Go.