Platform for stateful agents: AI with advanced memory that can learn and self-improve over time.
fix: disable reasoning toggle when agent type is letta agent v1 [PRO-1346] (#4946)
Charles Packer81bf13271288
diff --git a/letta/llm_api/anthropic_client.py b/letta/llm_api/anthropic_client.pyindex d01d8da79..6e737aa7a 100644--- a/letta/llm_api/anthropic_client.py+++ b/letta/llm_api/anthropic_client.py@@ -186,7 +186,7 @@ class AnthropicClient(LLMClientBase):) -> dict:# TODO: This needs to get cleaned up. The logic here is pretty confusing.# TODO: I really want to get rid of prefixing, it's a recipe for disaster code maintenance wise- prefix_fill = True+ prefix_fill = True if agent_type != AgentType.letta_v1_agent else Falseif not self.use_tool_naming:raise NotImplementedError("Only tool calling supported on Anthropic API requests")diff --git a/letta/schemas/llm_config.py b/letta/schemas/llm_config.pyindex 65f7a462c..12269bd17 100644--- a/letta/schemas/llm_config.py+++ b/letta/schemas/llm_config.py@@ -1,4 +1,4 @@-from typing import Literal, Optional+from typing import TYPE_CHECKING, Literal, Optionalfrom pydantic import BaseModel, ConfigDict, Field, model_validator@@ -259,7 +259,62 @@ class LLMConfig(BaseModel):return config.model_endpoint_type == "openai" and config.model.startswith("gpt-5")@classmethod- def apply_reasoning_setting_to_config(cls, config: "LLMConfig", reasoning: bool):+ def apply_reasoning_setting_to_config(cls, config: "LLMConfig", reasoning: bool, agent_type: Optional["AgentType"] = None):+ """+ Normalize reasoning-related flags on the config based on the requested+ "reasoning" setting, model capabilities, and optionally the agent type.++ For AgentType.letta_v1_agent, we enforce stricter semantics:+ - OpenAI native reasoning (o1/o3/o4/gpt-5): force enabled (non-togglable)+ - Anthropic (claude 3.7 / 4): toggle honored (default on elsewhere)+ - Google Gemini (2.5 family): force disabled until native reasoning supported+ - All others: disabled (no simulated reasoning via kwargs)+ """+ # Late import to avoid circular import with Agent schema using LLMConfig+ if TYPE_CHECKING:+ from letta.schemas.enums import AgentType # pragma: no cover+ else:+ try:+ from letta.schemas.enums import AgentType # type: ignore+ except Exception:+ AgentType = None # type: ignore++ # V1 agent policy: do not allow simulated reasoning for non-native models+ if agent_type is not None and AgentType is not None and agent_type == AgentType.letta_v1_agent: # type: ignore+ # OpenAI native reasoning models: always on+ if cls.is_openai_reasoning_model(config):+ config.put_inner_thoughts_in_kwargs = False+ config.enable_reasoner = True+ if config.reasoning_effort is None:+ if config.model.startswith("gpt-5"):+ config.reasoning_effort = "minimal"+ else:+ config.reasoning_effort = "medium"+ if config.model.startswith("gpt-5") and config.verbosity is None:+ config.verbosity = "medium"+ return config++ # Anthropic 3.7/4: toggle honored+ if cls.is_anthropic_reasoning_model(config):+ config.enable_reasoner = bool(reasoning)+ config.put_inner_thoughts_in_kwargs = False+ if config.enable_reasoner and config.max_reasoning_tokens == 0:+ config.max_reasoning_tokens = 1024+ return config++ # Gemini (Vertex or Google AI): force disabled for v1 until native support+ if cls.is_google_vertex_reasoning_model(config) or cls.is_google_ai_reasoning_model(config):+ config.put_inner_thoughts_in_kwargs = False+ config.enable_reasoner = False+ config.max_reasoning_tokens = 0+ return config++ # Everything else: disabled (no inner_thoughts-in-kwargs simulation)+ config.put_inner_thoughts_in_kwargs = False+ config.enable_reasoner = False+ config.max_reasoning_tokens = 0+ return config+if not reasoning:if cls.is_openai_reasoning_model(config):logger.warning("Reasoning cannot be disabled for OpenAI o1/o3/gpt-5 models")diff --git a/letta/services/agent_manager.py b/letta/services/agent_manager.pyindex ead5199e9..f852e5f6c 100644--- a/letta/services/agent_manager.py+++ b/letta/services/agent_manager.py@@ -315,7 +315,20 @@ class AgentManager:raise ValueError("llm_config and embedding_config are required")if agent_create.reasoning is not None:- agent_create.llm_config = LLMConfig.apply_reasoning_setting_to_config(agent_create.llm_config, agent_create.reasoning)+ agent_create.llm_config = LLMConfig.apply_reasoning_setting_to_config(+ agent_create.llm_config,+ agent_create.reasoning,+ agent_create.agent_type,+ )+ # For v1 agents, enforce sane defaults even when reasoning is omitted+ elif agent_create.agent_type == AgentType.letta_v1_agent:+ # Default togglable models (Anthropic 3.7/4) to enabled; others disabled+ default_reasoning = LLMConfig.is_anthropic_reasoning_model(agent_create.llm_config)+ agent_create.llm_config = LLMConfig.apply_reasoning_setting_to_config(+ agent_create.llm_config,+ default_reasoning,+ agent_create.agent_type,+ )# blocksblock_ids = list(agent_create.block_ids or [])@@ -642,7 +655,11 @@ class AgentManager:if agent_update.reasoning is not None:llm_config = agent_update.llm_config or agent.llm_config- agent_update.llm_config = LLMConfig.apply_reasoning_setting_to_config(llm_config, agent_update.reasoning)+ agent_update.llm_config = LLMConfig.apply_reasoning_setting_to_config(+ llm_config,+ agent_update.reasoning,+ agent.agent_type,+ )scalar_updates = {"name": agent_update.name,diff --git a/tests/test_providers.py b/tests/test_providers.pyindex b8ac3d015..a990ff6b9 100644--- a/tests/test_providers.py+++ b/tests/test_providers.py@@ -2,6 +2,7 @@ from typing import Literal, Optionalimport pytest+from letta.schemas.enums import AgentTypefrom letta.schemas.llm_config import LLMConfigfrom letta.schemas.providers import (AnthropicProvider,@@ -356,7 +357,7 @@ def test_reasoning_toggle_by_provider(handle=handle,context_window=1024,)- new_config = LLMConfig.apply_reasoning_setting_to_config(config, reasoning=reasoning)+ new_config = LLMConfig.apply_reasoning_setting_to_config(config, reasoning=reasoning, agent_type=AgentType.memgpt_v2_agent)assert new_config.enable_reasoner == expected_enable_reasonerassert new_config.put_inner_thoughts_in_kwargs == expected_put_inner_thoughts_in_kwargs