Informazioni di riferimento sulle API dell'agente a esecuzione prolungata (anteprima)

from datetime import timedelta
from azure.ai.agentserver.core.tasks import task, multi_turn_task, TaskContext, RetryPolicy

@task(name="summarize", timeout=timedelta(minutes=10), retry=RetryPolicy())
async def summarize(ctx: TaskContext[str]) -> str:
    ...

@multi_turn_task(name="chat", steerable=True)
async def chat(ctx: TaskContext[dict]) -> dict:
    ...

using Azure.AI.AgentServer.Core.Tasks;

// Register the task engine, then declare tasks.
builder.Services.AddResilientTasks(credential);

builder.Services.AddResilientTaskBuilder(tasks =>
{
    tasks.AddTask<string, string>("summarize", async (TaskContext<string> ctx, CancellationToken ct) =>
    {
        // ...
        return result;
    });

    tasks.AddMultiTurnTask<ChatInput, ChatOutput>("chat", async (ctx, ct) =>
    {
        // ...
        return output;
    }, steerable: true);
});

# One-shot: run returns the output directly.
result = await summarize.run(input="...")

# Multi-turn: same task_id resumes the chain; if_last_input_id enforces ordering.
r = await chat.run(task_id="conv-7", input={"msg": "hi"}, if_last_input_id=prev_id)

# Fire-and-track: start returns a TaskRun handle.
run = await chat.start(task_id="conv-7", input={"msg": "hi"})

// Resolve the task and run it.
var run = await task.StartAsync(input, new RunOptions
{
    TaskId = "conv-7",
    IfLastInputId = prevId,
});
var output = await run.GetResultAsync();

RunOptions porta TaskId, InputIde IfLastInputId (precondizione di ordinamento).

from azure.ai.agentserver.core.tasks import RetryPolicy

RetryPolicy(
    initial_delay=timedelta(seconds=1),   # base delay
    backoff_coefficient=2.0,              # multiplier per attempt (>= 1.0)
    max_delay=timedelta(seconds=60),      # cap
    max_attempts=3,                       # total tries, including the first
)
RetryPolicy.no_retry()                    # single attempt

Formula ritardata: min(initial_delay * backoff_coefficient ** attempt, max_delay).

TaskRetryPolicy.ExponentialBackoff(maxAttempts: 3, initialDelay: TimeSpan.FromSeconds(1),
    backoffCoefficient: 2.0, maxDelay: TimeSpan.FromSeconds(60), jitter: true);
TaskRetryPolicy.FixedDelay(maxAttempts, delay, jitter);
TaskRetryPolicy.LinearBackoff(maxAttempts, initialDelay, increment, maxDelay, jitter);
TaskRetryPolicy.NoRetry;

Proprietà: InitialDelay, BackoffCoefficient, MaxDelayMaxAttempts, , Jitter, RetryOn.

from azure.ai.agentserver.core.tasks import set_resilient_tasks_enabled
set_resilient_tasks_enabled(True)   # call at import time, before host startup

AddResilientTasks(credential) registra lo scanner di ripristino durante l'avvio dell'host; non è necessaria alcuna opzione separata.

from azure.ai.agentserver.responses import ResponsesAgentServerHost, ResponsesServerOptions

app = ResponsesAgentServerHost(
    options=ResponsesServerOptions(
        resilient_background=True,       # re-invoke stored background responses after a crash
        steerable_conversations=True,    # queue a new turn instead of 409 conversation_locked
        default_fetch_history_count=100,
        sse_keep_alive_interval_seconds=None,
        shutdown_grace_period_seconds=10,
    ),
)

var options = new ResponsesServerOptions
{
    ResilientBackground = true,
    SteerableConversations = true,
    DefaultFetchHistoryCount = 100,
};

from azure.ai.agentserver.core.streaming import streams

streams.use_in_memory_live()                                        # no replay, no restart survival
streams.use_in_memory_replay(cursor_fn=lambda e: e["n"], ttl_seconds=600)
streams.use_file_backed_replay(storage_dir=Path("/streams"),
                               cursor_fn=lambda e: e["n"])

stream = await streams.get_or_create(invocation_id)
await stream.emit({"n": 0, "delta": "..."})
async for event in stream.subscribe(after=last_cursor):
    ...
await stream.close()

Operazioni del Registro di sistema: await streams.get_or_create(id), await streams.delete(id). Passare cursor_fn per abilitare subscribe(after=...) e last_cursor().

using Azure.AI.AgentServer.Core.Streaming;

builder.Services.AddEventStreams(options =>
{
    // Configure the backing (live / replay / file-backed) via EventStreamOptions.
});

// Resolve the registry and obtain a per-turn stream.
EventStream stream = await registry.GetOrCreateAsync(invocationId);

Backup: BroadcastEventStream (in tempo reale), ReplayEventStream (riproduzione in memoria), FileBackedReplayEventStream (persistente). Registro di sistema: EventStreamRegistry / InMemoryEventStreamRegistry.