fix: corrige autenticação e ciclo de streaming do DeepAgents

README.md

@@ -253,6 +253,22 @@ flowchart TB
 
 **Design principle:** transport, orchestration, lifecycle, and provider-specific integrations remain separate concerns. Provider independence is established through contracts and factories, while streaming and cancellation behavior are validated at their respective integration boundaries.
 
+### DeepAgents: autenticação e streaming
+
+No Mac, `API_KEY` lê o perfil OCI e a chave indicados pelo arquivo de configuração.
+Para uma sessão OCI CLI, use `SECURITY_TOKEN` e um perfil com token válido.
+O DeepAgents também aceita `INSTANCE_PRINCIPAL` e `RESOURCE_PRINCIPAL` em
+ambientes OCI compatíveis. O padrão do construtor permanece `INSTANCE_PRINCIPAL`
+quando `auth_type` é omitido. Credenciais e permissões precisam ser válidas.
+Dentro do container, monte o arquivo OCI e as chaves/tokens necessários somente
+para leitura, com caminhos acessíveis ao usuário do processo, e ajuste
+`auth_file_location`; `~` refere-se ao usuário do container, não ao do Mac.
+Não há integração MCP/RAG automática nesse bloco; a implementação usa `tools=[]`.
+`agent.options.timeout_seconds` limita a espera por cada próximo evento do
+DeepAgents (inclusive eventos internos), sem incluir o tempo em que o consumidor
+está processando um fragmento já entregue. Não é prazo total da resposta nem
+limite de inicialização/autenticação. A expiração fecha o stream e propaga erro.
+
 ## Quickstart
 
 ### Requirements

src/voice_assistant/agent/deep_agent.py

@@ -16,7 +16,10 @@ Does not manage:
 
 from __future__ import annotations
 
-from collections.abc import AsyncGenerator
+import asyncio
+import math
+from collections.abc import AsyncGenerator, Iterator
+from typing import Any
 from contextlib import aclosing
 
 from deepagents import create_deep_agent
@@ -41,6 +44,23 @@ de ferramentas que não foram executadas.
 """
 
 
+def _text_fragments(content: str | list[str | dict[str, Any]]) -> Iterator[str]:
+    """Extract text without exposing reasoning, tools or other content blocks."""
+    if isinstance(content, str):
+        if content:
+            yield content
+        return
+
+    for block in content:
+        if isinstance(block, str):
+            if block:
+                yield block
+        elif isinstance(block, dict) and block.get("type") == "text":
+            text = block.get("text")
+            if isinstance(text, str) and text:
+                yield text
+
+
 class DeepAgentsAdapter(StreamingAgentAdapter):
     def __init__(
         self,
@@ -51,6 +71,10 @@ class DeepAgentsAdapter(StreamingAgentAdapter):
         max_completion_tokens: int,
         timeout_seconds: float,
         system_prompt: str = DEFAULT_SYSTEM_PROMPT,
+        *,
+        auth_type: str = "INSTANCE_PRINCIPAL",
+        auth_profile: str = "DEFAULT",
+        auth_file_location: str = "~/.oci/config",
     ) -> None:
         if not region:
             raise ValueError("OCI region is required.")
@@ -61,8 +85,15 @@ class DeepAgentsAdapter(StreamingAgentAdapter):
         if not model_id:
             raise ValueError("OCI model_id is required.")
 
-        if timeout_seconds <= 0:
+        if (
-            raise ValueError("timeout_seconds must be positive.")
+            isinstance(timeout_seconds, bool)
+            or not isinstance(timeout_seconds, (int, float))
+            or not math.isfinite(timeout_seconds)
+            or timeout_seconds <= 0
+        ):
+            raise ValueError("timeout_seconds must be finite and positive.")
+
+        self._stream_event_timeout = float(timeout_seconds)
 
         self.model = ChatOCIGenAI(
             model_id=model_id,
@@ -71,7 +102,9 @@ class DeepAgentsAdapter(StreamingAgentAdapter):
                 f"https://inference.generativeai."
                 f"{region}.oci.oraclecloud.com"
             ),
-            auth_type="INSTANCE_PRINCIPAL",
+            auth_type=auth_type,
+            auth_profile=auth_profile,
+            auth_file_location=auth_file_location,
             is_stream=True,
             model_kwargs={
                 "temperature": temperature,
@@ -142,27 +175,30 @@ class DeepAgentsAdapter(StreamingAgentAdapter):
 
         config = self._config(session_id)
 
-        async for message, metadata in self.agent.astream(
+        events = self.agent.astream(
-            {
+            {"messages": [HumanMessage(content=text)]},
-                "messages": [
-                    HumanMessage(content=text)
-                ]
-            },
             config=config,
             stream_mode="messages",
-        ):
+        )
-            # Ignore tools and subagent outputs.
+        async with aclosing(events):
+            while True:
+                try:
+                    async with asyncio.timeout(self._stream_event_timeout):
+                        message, metadata = await anext(events)
+                except StopAsyncIteration:
+                    break
+                except TimeoutError as exc:
+                    raise TimeoutError(
+                        "Deep Agent timed out waiting for the next stream event."
+                    ) from exc
+
                 if metadata.get("langgraph_node") != "model":
                     continue
-
-            # Ignore tool-call fragments.
                 if getattr(message, "tool_call_chunks", None):
                     continue
 
-            content = message.content
+                for fragment in _text_fragments(message.content):
-
+                    yield fragment
-            if isinstance(content, str) and content:
-                yield content
 
 
     async def aclose(self) -> None:

tests/test_deep_agent.py

@@ -64,6 +64,7 @@ def test_deep_agent_stream_can_be_cancelled():
     async def scenario():
         adapter = object.__new__(DeepAgentsAdapter)
         adapter._closed = False
+        adapter._stream_event_timeout = 1.0
 
         started = asyncio.Event()
         cancelled = asyncio.Event()
@@ -341,3 +342,53 @@ def test_deep_agent_checkpoint_survives_cancellation():
     started = None
     asyncio.run(scenario())
 
+@pytest.mark.parametrize(
+    ("options", "expected_auth"),
+    [
+        ({}, ("INSTANCE_PRINCIPAL", "DEFAULT", "~/.oci/config")),
+        (
+            {"auth_type": "API_KEY", "auth_profile": "LOCAL", "auth_file_location": "/config/oci"},
+            ("API_KEY", "LOCAL", "/config/oci"),
+        ),
+        (
+            {"auth_type": "SECURITY_TOKEN", "auth_profile": "SESSION"},
+            ("SECURITY_TOKEN", "SESSION", "~/.oci/config"),
+        ),
+        ({"auth_type": "RESOURCE_PRINCIPAL"}, ("RESOURCE_PRINCIPAL", "DEFAULT", "~/.oci/config")),
+    ],
+)
+def test_deep_agent_factory_configures_authentication(monkeypatch, options, expected_auth):
+    import voice_assistant.agent.deep_agent as module
+    from voice_assistant.config.loader import load_provider_catalog
+    from voice_assistant.config.models import AppConfig
+    from voice_assistant.providers import create_agent_adapter
+
+    model_factory = MagicMock()
+    agent_factory = MagicMock()
+    monkeypatch.setattr(module, "ChatOCIGenAI", model_factory)
+    monkeypatch.setattr(module, "create_deep_agent", agent_factory)
+    config = AppConfig.model_validate({
+        "stt": {"provider": "nvidia"},
+        "tts": {"provider": "voxcpm"},
+        "agent": {
+            "provider": "deepagents",
+            "options": {
+                "region": "sa-vinhedo-1",
+                "compartment_id": "test-compartment",
+                "model_id": "test-model",
+                "temperature": 0.3,
+                "max_completion_tokens": 256,
+                "timeout_seconds": 60.0,
+                **options,
+            },
+        },
+    })
+
+    adapter = create_agent_adapter(config=config, catalog=load_provider_catalog())
+
+    model_factory.assert_called_once()
+    kwargs = model_factory.call_args.kwargs
+    assert (kwargs["auth_type"], kwargs["auth_profile"], kwargs["auth_file_location"]) == expected_auth
+    assert adapter.model is model_factory.return_value
+    assert agent_factory.call_args.kwargs["model"] is adapter.model
+

tests/test_deep_agent_stream.py

@@ -0,0 +1,146 @@
+import asyncio
+from contextlib import aclosing
+from types import SimpleNamespace
+
+import pytest
+
+from voice_assistant.agent.deep_agent import DeepAgentsAdapter
+
+
+def make_adapter(events, timeout=0.05):
+    adapter = object.__new__(DeepAgentsAdapter)
+    adapter._closed = False
+    adapter._stream_event_timeout = timeout
+    adapter.agent = SimpleNamespace(astream=lambda *args, **kwargs: events)
+    return adapter
+
+
+def event(content, node="model", tool_calls=None):
+    return SimpleNamespace(content=content, tool_call_chunks=tool_calls or []), {"langgraph_node": node}
+
+
+@pytest.mark.parametrize(
+    ("content", "expected"),
+    [
+        ("Olá ", ["Olá "]),
+        ([{"type": "text", "text": "Olá "}, {"type": "text", "text": "Lara."}], ["Olá ", "Lara."]),
+        (["Olá", " ", {"type": "text", "text": "Lara."}], ["Olá", " ", "Lara."]),
+        ([{"type": "reasoning", "text": "Interno"}, {"type": "image", "text": "Ignorar"}, {"type": "text", "text": "Resposta"}], ["Resposta"]),
+        ([{"type": "text"}, {"type": "text", "text": None}, {"type": "text", "text": ""}], []),
+    ],
+)
+def test_stream_extracts_only_text_blocks(content, expected):
+    async def scenario():
+        async def source():
+            yield event(content)
+        assert [part async for part in make_adapter(source()).stream("session", 1, "Olá")] == expected
+    asyncio.run(scenario())
+
+
+def test_stream_filters_non_model_events_and_tool_calls():
+    async def scenario():
+        async def source():
+            yield event("Resultado interno", node="tools")
+            yield event("Chamada interna", tool_calls=[{"id": "call-1"}])
+            yield event("Resposta final")
+        assert [part async for part in make_adapter(source()).stream("session", 1, "Olá")] == ["Resposta final"]
+    asyncio.run(scenario())
+
+
+@pytest.mark.parametrize("finish", ["normal", "early_close", "cancel", "error", "timeout"])
+def test_stream_closes_owned_generator(finish):
+    async def scenario():
+        closed = asyncio.Event()
+        waiting = asyncio.Event()
+
+        async def source():
+            try:
+                yield event("Olá")
+                if finish == "error":
+                    raise ValueError("provider failure")
+                if finish in ("cancel", "timeout"):
+                    waiting.set()
+                    await asyncio.Event().wait()
+            finally:
+                closed.set()
+
+        stream = make_adapter(source(), timeout=0.1).stream("session", 1, "Olá")
+        assert await anext(stream) == "Olá"
+        if finish == "normal":
+            with pytest.raises(StopAsyncIteration):
+                await anext(stream)
+        elif finish == "early_close":
+            await stream.aclose()
+        elif finish == "cancel":
+            task = asyncio.create_task(anext(stream))
+            await asyncio.wait_for(waiting.wait(), 1)
+            task.cancel()
+            with pytest.raises(asyncio.CancelledError):
+                await task
+        elif finish == "error":
+            with pytest.raises(ValueError, match="provider failure"):
+                await anext(stream)
+        else:
+            with pytest.raises(TimeoutError, match="next stream event"):
+                await anext(stream)
+        assert closed.is_set()
+    asyncio.run(scenario())
+
+
+def test_timeout_does_not_cancel_consumer_between_fragments():
+    async def scenario():
+        async def source():
+            yield event("Primeiro ")
+            yield event("segundo")
+        async with aclosing(make_adapter(source(), timeout=0.01).stream("session", 1, "Olá")) as stream:
+            assert await anext(stream) == "Primeiro "
+            await asyncio.sleep(0.04)
+            assert await anext(stream) == "segundo"
+    asyncio.run(scenario())
+
+
+def test_timeout_is_renewed_for_each_next_event():
+    async def scenario():
+        async def source():
+            for token in ("a", "b", "c", "d"):
+                await asyncio.sleep(0.02)
+                yield event(token)
+        adapter = make_adapter(source(), timeout=0.06)
+        assert "".join([part async for part in adapter.stream("session", 1, "Olá")]) == "abcd"
+    asyncio.run(scenario())
+
+
+@pytest.mark.parametrize("timeout", [0, -1, float("inf"), float("nan"), True, "60", None])
+def test_invalid_timeout_fails_before_model_construction(monkeypatch, timeout):
+    import voice_assistant.agent.deep_agent as module
+
+    def unexpected_model(**kwargs):
+        pytest.fail("Invalid configuration must not create a model")
+
+    monkeypatch.setattr(module, "ChatOCIGenAI", unexpected_model)
+    with pytest.raises(ValueError, match="finite and positive"):
+        DeepAgentsAdapter(
+            region="sa-vinhedo-1", compartment_id="test", model_id="test",
+            temperature=0.3, max_completion_tokens=256, timeout_seconds=timeout,
+        )
+
+
+def test_adapter_streams_real_deepagents_graph_with_fake_model():
+    from deepagents import create_deep_agent
+    from langchain_core.language_models import FakeListChatModel
+    from langgraph.checkpoint.memory import MemorySaver
+
+    class Model(FakeListChatModel):
+        def bind_tools(self, tools, *, tool_choice=None, **kwargs):
+            return self
+
+    async def scenario():
+        adapter = object.__new__(DeepAgentsAdapter)
+        adapter._closed = False
+        adapter._stream_event_timeout = 2.0
+        adapter.agent = create_deep_agent(
+            model=Model(responses=["Olá Lara."]), tools=[], checkpointer=MemorySaver(),
+        )
+        response = await adapter.process("session", 1, "Diga olá")
+        assert response == "Olá Lara."
+    asyncio.run(scenario())