letta-nightly 0.6.28.dev20250220163833__tar.gz → 0.6.29.dev20250221104040__tar.gz

{letta_nightly-0.6.28.dev20250220163833 → letta_nightly-0.6.29.dev20250221104040}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: letta-nightly
-Version: 0.6.28.dev20250220163833
+Version: 0.6.29.dev20250221104040
 Summary: Create LLM agents with long-term memory and custom tools
 License: Apache License
 Author: Letta Team

{letta_nightly-0.6.28.dev20250220163833 → letta_nightly-0.6.29.dev20250221104040}/letta/__init__.py

@@ -1,4 +1,4 @@
-__version__ = "0.6.28"
+__version__ = "0.6.29"
 
 # import clients
 from letta.client.client import LocalClient, RESTClient, create_client

{letta_nightly-0.6.28.dev20250220163833 → letta_nightly-0.6.29.dev20250221104040}/letta/agent.py

@@ -322,6 +322,7 @@ class Agent(BaseAgent):
         max_delay: float = 10.0,  # max delay between retries
         step_count: Optional[int] = None,
         last_function_failed: bool = False,
+        put_inner_thoughts_first: bool = True,
     ) -> ChatCompletionResponse:
         """Get response from LLM API with robust retry mechanism."""
         log_telemetry(self.logger, "_get_ai_reply start")
@@ -367,6 +368,7 @@ class Agent(BaseAgent):
                     force_tool_call=force_tool_call,
                     stream=stream,
                     stream_interface=self.interface,
+                    put_inner_thoughts_first=put_inner_thoughts_first,
                 )
                 log_telemetry(self.logger, "_get_ai_reply create finish")
 
@@ -648,6 +650,7 @@ class Agent(BaseAgent):
         # additional args
         chaining: bool = True,
         max_chaining_steps: Optional[int] = None,
+        put_inner_thoughts_first: bool = True,
         **kwargs,
     ) -> LettaUsageStatistics:
         """Run Agent.step in a loop, handling chaining via heartbeat requests and function failures"""
@@ -662,6 +665,7 @@ class Agent(BaseAgent):
             kwargs["last_function_failed"] = function_failed
             step_response = self.inner_step(
                 messages=next_input_message,
+                put_inner_thoughts_first=put_inner_thoughts_first,
                 **kwargs,
             )
 
@@ -743,9 +747,9 @@ class Agent(BaseAgent):
         metadata: Optional[dict] = None,
         summarize_attempt_count: int = 0,
         last_function_failed: bool = False,
+        put_inner_thoughts_first: bool = True,
     ) -> AgentStepResponse:
         """Runs a single step in the agent loop (generates at most one LLM call)"""
-
         try:
 
             # Extract job_id from metadata if present
@@ -778,6 +782,7 @@ class Agent(BaseAgent):
                 stream=stream,
                 step_count=step_count,
                 last_function_failed=last_function_failed,
+                put_inner_thoughts_first=put_inner_thoughts_first,
             )
             if not response:
                 # EDGE CASE: Function call failed AND there's no tools left for agent to call -> return early

{letta_nightly-0.6.28.dev20250220163833 → letta_nightly-0.6.29.dev20250221104040}/letta/llm_api/helpers.py

@@ -202,21 +202,29 @@ def add_inner_thoughts_to_functions(
     inner_thoughts_key: str,
     inner_thoughts_description: str,
     inner_thoughts_required: bool = True,
+    put_inner_thoughts_first: bool = True,
 ) -> List[dict]:
     """Add an inner_thoughts kwarg to every function in the provided list, ensuring it's the first parameter"""
     new_functions = []
     for function_object in functions:
         new_function_object = copy.deepcopy(function_object)
-
-        # Create a new OrderedDict with inner_thoughts as the first item
         new_properties = OrderedDict()
-        new_properties[inner_thoughts_key] = {
-            "type": "string",
-            "description": inner_thoughts_description,
-        }
 
-        # Add the rest of the properties
-        new_properties.update(function_object["parameters"]["properties"])
+        # For chat completions, we want inner thoughts to come later
+        if put_inner_thoughts_first:
+            # Create with inner_thoughts as the first item
+            new_properties[inner_thoughts_key] = {
+                "type": "string",
+                "description": inner_thoughts_description,
+            }
+            # Add the rest of the properties
+            new_properties.update(function_object["parameters"]["properties"])
+        else:
+            new_properties.update(function_object["parameters"]["properties"])
+            new_properties[inner_thoughts_key] = {
+                "type": "string",
+                "description": inner_thoughts_description,
+            }
 
         # Cast OrderedDict back to a regular dict
         new_function_object["parameters"]["properties"] = dict(new_properties)
@@ -225,9 +233,11 @@ def add_inner_thoughts_to_functions(
         if inner_thoughts_required:
             required_params = new_function_object["parameters"].get("required", [])
             if inner_thoughts_key not in required_params:
-                required_params.insert(0, inner_thoughts_key)
+                if put_inner_thoughts_first:
+                    required_params.insert(0, inner_thoughts_key)
+                else:
+                    required_params.append(inner_thoughts_key)
                 new_function_object["parameters"]["required"] = required_params
-
         new_functions.append(new_function_object)
 
     return new_functions

{letta_nightly-0.6.28.dev20250220163833 → letta_nightly-0.6.29.dev20250221104040}/letta/llm_api/llm_api_tools.py

@@ -140,6 +140,7 @@ def create(
     stream: bool = False,
     stream_interface: Optional[Union[AgentRefreshStreamingInterface, AgentChunkStreamingInterface]] = None,
     model_settings: Optional[dict] = None,  # TODO: eventually pass from server
+    put_inner_thoughts_first: bool = True,
 ) -> ChatCompletionResponse:
     """Return response to chat completion with backoff"""
     from letta.utils import printd
@@ -185,7 +186,9 @@ def create(
             else:
                 function_call = "required"
 
-        data = build_openai_chat_completions_request(llm_config, messages, user_id, functions, function_call, use_tool_naming)
+        data = build_openai_chat_completions_request(
+            llm_config, messages, user_id, functions, function_call, use_tool_naming, put_inner_thoughts_first=put_inner_thoughts_first
+        )
         if stream:  # Client requested token streaming
             data.stream = True
             assert isinstance(stream_interface, AgentChunkStreamingInterface) or isinstance(

{letta_nightly-0.6.28.dev20250220163833 → letta_nightly-0.6.29.dev20250221104040}/letta/llm_api/openai.py

@@ -94,6 +94,7 @@ def build_openai_chat_completions_request(
     functions: Optional[list],
     function_call: Optional[str],
     use_tool_naming: bool,
+    put_inner_thoughts_first: bool = True,
 ) -> ChatCompletionRequest:
     if functions and llm_config.put_inner_thoughts_in_kwargs:
         # Special case for LM Studio backend since it needs extra guidance to force out the thoughts first
@@ -105,6 +106,7 @@ def build_openai_chat_completions_request(
             functions=functions,
             inner_thoughts_key=INNER_THOUGHTS_KWARG,
             inner_thoughts_description=inner_thoughts_desc,
+            put_inner_thoughts_first=put_inner_thoughts_first,
         )
 
     openai_message_list = [
@@ -390,7 +392,7 @@ def openai_chat_completions_process_stream(
     chat_completion_response.usage.completion_tokens = n_chunks
     chat_completion_response.usage.total_tokens = prompt_tokens + n_chunks
 
-    assert len(chat_completion_response.choices) > 0, chat_completion_response
+    assert len(chat_completion_response.choices) > 0, f"No response from provider {chat_completion_response}"
 
     # printd(chat_completion_response)
     return chat_completion_response

{letta_nightly-0.6.28.dev20250220163833 → letta_nightly-0.6.29.dev20250221104040}/letta/orm/__init__.py

@@ -4,6 +4,7 @@ from letta.orm.base import Base
 from letta.orm.block import Block
 from letta.orm.blocks_agents import BlocksAgents
 from letta.orm.file import FileMetadata
+from letta.orm.identities_agents import IdentitiesAgents
 from letta.orm.identity import Identity
 from letta.orm.job import Job
 from letta.orm.job_messages import JobMessage

{letta_nightly-0.6.28.dev20250220163833 → letta_nightly-0.6.29.dev20250221104040}/letta/orm/agent.py

@@ -1,7 +1,7 @@
 import uuid
 from typing import TYPE_CHECKING, List, Optional
 
-from sqlalchemy import JSON, Boolean, ForeignKey, Index, String
+from sqlalchemy import JSON, Boolean, Index, String
 from sqlalchemy.orm import Mapped, mapped_column, relationship
 
 from letta.orm.block import Block
@@ -61,14 +61,6 @@ class Agent(SqlalchemyBase, OrganizationMixin):
     template_id: Mapped[Optional[str]] = mapped_column(String, nullable=True, doc="The id of the template the agent belongs to.")
     base_template_id: Mapped[Optional[str]] = mapped_column(String, nullable=True, doc="The base template id of the agent.")
 
-    # Identity
-    identity_id: Mapped[Optional[str]] = mapped_column(
-        String, ForeignKey("identities.id", ondelete="CASCADE"), nullable=True, doc="The id of the identity the agent belongs to."
-    )
-    identifier_key: Mapped[Optional[str]] = mapped_column(
-        String, nullable=True, doc="The identifier key of the identity the agent belongs to."
-    )
-
     # Tool rules
     tool_rules: Mapped[Optional[List[ToolRule]]] = mapped_column(ToolRulesColumn, doc="the tool rules for this agent.")
 
@@ -79,7 +71,6 @@ class Agent(SqlalchemyBase, OrganizationMixin):
 
     # relationships
     organization: Mapped["Organization"] = relationship("Organization", back_populates="agents")
-    identity: Mapped["Identity"] = relationship("Identity", back_populates="agents")
     tool_exec_environment_variables: Mapped[List["AgentEnvironmentVariable"]] = relationship(
         "AgentEnvironmentVariable",
         back_populates="agent",
@@ -130,7 +121,13 @@ class Agent(SqlalchemyBase, OrganizationMixin):
         viewonly=True,  # Ensures SQLAlchemy doesn't attempt to manage this relationship
         doc="All passages derived created by this agent.",
     )
-    identity: Mapped[Optional["Identity"]] = relationship("Identity", back_populates="agents")
+    identities: Mapped[List["Identity"]] = relationship(
+        "Identity",
+        secondary="identities_agents",
+        lazy="selectin",
+        back_populates="agents",
+        passive_deletes=True,
+    )
 
     def to_pydantic(self) -> PydanticAgentState:
         """converts to the basic pydantic model counterpart"""
@@ -160,6 +157,7 @@ class Agent(SqlalchemyBase, OrganizationMixin):
             "project_id": self.project_id,
             "template_id": self.template_id,
             "base_template_id": self.base_template_id,
+            "identity_ids": [identity.id for identity in self.identities],
             "message_buffer_autoclear": self.message_buffer_autoclear,
         }
 

letta_nightly-0.6.29.dev20250221104040/letta/orm/identities_agents.py

@@ -0,0 +1,13 @@
+from sqlalchemy import ForeignKey, String
+from sqlalchemy.orm import Mapped, mapped_column
+
+from letta.orm.base import Base
+
+
+class IdentitiesAgents(Base):
+    """Identities may have one or many agents associated with them."""
+
+    __tablename__ = "identities_agents"
+
+    identity_id: Mapped[str] = mapped_column(String, ForeignKey("identities.id", ondelete="CASCADE"), primary_key=True)
+    agent_id: Mapped[str] = mapped_column(String, ForeignKey("agents.id", ondelete="CASCADE"), primary_key=True)

{letta_nightly-0.6.28.dev20250220163833 → letta_nightly-0.6.29.dev20250221104040}/letta/orm/identity.py

@@ -2,11 +2,13 @@ import uuid
 from typing import List, Optional
 
 from sqlalchemy import String, UniqueConstraint
+from sqlalchemy.dialects.postgresql import JSONB
 from sqlalchemy.orm import Mapped, mapped_column, relationship
 
 from letta.orm.mixins import OrganizationMixin
 from letta.orm.sqlalchemy_base import SqlalchemyBase
 from letta.schemas.identity import Identity as PydanticIdentity
+from letta.schemas.identity import IdentityProperty
 
 
 class Identity(SqlalchemyBase, OrganizationMixin):
@@ -14,17 +16,35 @@ class Identity(SqlalchemyBase, OrganizationMixin):
 
     __tablename__ = "identities"
     __pydantic_model__ = PydanticIdentity
-    __table_args__ = (UniqueConstraint("identifier_key", "project_id", "organization_id", name="unique_identifier_pid_org_id"),)
+    __table_args__ = (
+        UniqueConstraint(
+            "identifier_key",
+            "project_id",
+            "organization_id",
+            name="unique_identifier_without_project",
+            postgresql_nulls_not_distinct=True,
+        ),
+    )
 
     id: Mapped[str] = mapped_column(String, primary_key=True, default=lambda: f"identity-{uuid.uuid4()}")
     identifier_key: Mapped[str] = mapped_column(nullable=False, doc="External, user-generated identifier key of the identity.")
     name: Mapped[str] = mapped_column(nullable=False, doc="The name of the identity.")
     identity_type: Mapped[str] = mapped_column(nullable=False, doc="The type of the identity.")
     project_id: Mapped[Optional[str]] = mapped_column(nullable=True, doc="The project id of the identity.")
+    properties: Mapped[List["IdentityProperty"]] = mapped_column(
+        JSONB, nullable=False, default=list, doc="List of properties associated with the identity"
+    )
 
     # relationships
     organization: Mapped["Organization"] = relationship("Organization", back_populates="identities")
-    agents: Mapped[List["Agent"]] = relationship("Agent", lazy="selectin", back_populates="identity")
+    agents: Mapped[List["Agent"]] = relationship(
+        "Agent", secondary="identities_agents", lazy="selectin", passive_deletes=True, back_populates="identities"
+    )
+
+    @property
+    def agent_ids(self) -> List[str]:
+        """Get just the agent IDs without loading the full agent objects"""
+        return [agent.id for agent in self.agents]
 
     def to_pydantic(self) -> PydanticIdentity:
         state = {
@@ -33,7 +53,8 @@ class Identity(SqlalchemyBase, OrganizationMixin):
             "name": self.name,
             "identity_type": self.identity_type,
             "project_id": self.project_id,
-            "agents": [agent.to_pydantic() for agent in self.agents],
+            "agent_ids": self.agent_ids,
+            "organization_id": self.organization_id,
+            "properties": self.properties,
         }
-
-        return self.__pydantic_model__(**state)
+        return PydanticIdentity(**state)

{letta_nightly-0.6.28.dev20250220163833 → letta_nightly-0.6.29.dev20250221104040}/letta/orm/sqlalchemy_base.py

@@ -68,6 +68,7 @@ class SqlalchemyBase(CommonSqlalchemyMetaMixins, Base):
         access_type: AccessType = AccessType.ORGANIZATION,
         join_model: Optional[Base] = None,
         join_conditions: Optional[Union[Tuple, List]] = None,
+        identifier_keys: Optional[List[str]] = None,
         **kwargs,
     ) -> List["SqlalchemyBase"]:
         """
@@ -143,6 +144,9 @@ class SqlalchemyBase(CommonSqlalchemyMetaMixins, Base):
                 # Group by primary key and all necessary columns to avoid JSON comparison
                 query = query.group_by(cls.id)
 
+            if identifier_keys and hasattr(cls, "identities"):
+                query = query.join(cls.identities).filter(cls.identities.property.mapper.class_.identifier_key.in_(identifier_keys))
+
             # Apply filtering logic from kwargs
             for key, value in kwargs.items():
                 if "." in key:

{letta_nightly-0.6.28.dev20250220163833 → letta_nightly-0.6.29.dev20250221104040}/letta/schemas/agent.py

@@ -83,9 +83,7 @@ class AgentState(OrmMetadataBase, validate_assignment=True):
     project_id: Optional[str] = Field(None, description="The id of the project the agent belongs to.")
     template_id: Optional[str] = Field(None, description="The id of the template the agent belongs to.")
     base_template_id: Optional[str] = Field(None, description="The base template id of the agent.")
-
-    # Identity
-    identifier_key: Optional[str] = Field(None, description="The identifier key belonging to the identity associated with this agent.")
+    identity_ids: List[str] = Field([], description="The ids of the identities associated with this agent.")
 
     # An advanced configuration that makes it so this agent does not remember any previous messages
     message_buffer_autoclear: bool = Field(
@@ -161,7 +159,7 @@ class CreateAgent(BaseModel, validate_assignment=True):  #
     project_id: Optional[str] = Field(None, description="The id of the project the agent belongs to.")
     template_id: Optional[str] = Field(None, description="The id of the template the agent belongs to.")
     base_template_id: Optional[str] = Field(None, description="The base template id of the agent.")
-    identifier_key: Optional[str] = Field(None, description="The identifier key belonging to the identity associated with this agent.")
+    identity_ids: Optional[List[str]] = Field(None, description="The ids of the identities associated with this agent.")
     message_buffer_autoclear: bool = Field(
         False,
         description="If set to True, the agent will not remember previous messages (though the agent will still retain state via core memory blocks and archival/recall memory). Not recommended unless you have an advanced use case.",
@@ -236,7 +234,7 @@ class UpdateAgent(BaseModel):
     project_id: Optional[str] = Field(None, description="The id of the project the agent belongs to.")
     template_id: Optional[str] = Field(None, description="The id of the template the agent belongs to.")
     base_template_id: Optional[str] = Field(None, description="The base template id of the agent.")
-    identifier_key: Optional[str] = Field(None, description="The identifier key belonging to the identity associated with this agent.")
+    identity_ids: Optional[List[str]] = Field(None, description="The ids of the identities associated with this agent.")
     message_buffer_autoclear: Optional[bool] = Field(
         None,
         description="If set to True, the agent will not remember previous messages (though the agent will still retain state via core memory blocks and archival/recall memory). Not recommended unless you have an advanced use case.",

{letta_nightly-0.6.28.dev20250220163833 → letta_nightly-0.6.29.dev20250221104040}/letta/schemas/identity.py

@@ -1,9 +1,8 @@
 from enum import Enum
-from typing import List, Optional
+from typing import List, Optional, Union
 
 from pydantic import Field
 
-from letta.schemas.agent import AgentState
 from letta.schemas.letta_base import LettaBase
 
 
@@ -17,17 +16,38 @@ class IdentityType(str, Enum):
     other = "other"
 
 
+class IdentityPropertyType(str, Enum):
+    """
+    Enum to represent the type of the identity property.
+    """
+
+    string = "string"
+    number = "number"
+    boolean = "boolean"
+    json = "json"
+
+
 class IdentityBase(LettaBase):
     __id_prefix__ = "identity"
 
 
+class IdentityProperty(LettaBase):
+    """A property of an identity"""
+
+    key: str = Field(..., description="The key of the property")
+    value: Union[str, int, float, bool, dict] = Field(..., description="The value of the property")
+    type: IdentityPropertyType = Field(..., description="The type of the property")
+
+
 class Identity(IdentityBase):
     id: str = IdentityBase.generate_id_field()
     identifier_key: str = Field(..., description="External, user-generated identifier key of the identity.")
     name: str = Field(..., description="The name of the identity.")
     identity_type: IdentityType = Field(..., description="The type of the identity.")
     project_id: Optional[str] = Field(None, description="The project id of the identity, if applicable.")
-    agents: List[AgentState] = Field(..., description="The agents associated with the identity.")
+    agent_ids: List[str] = Field(..., description="The IDs of the agents associated with the identity.")
+    organization_id: Optional[str] = Field(None, description="The organization id of the user")
+    properties: List[IdentityProperty] = Field(default_factory=list, description="List of properties associated with the identity")
 
 
 class IdentityCreate(LettaBase):
@@ -36,9 +56,12 @@ class IdentityCreate(LettaBase):
     identity_type: IdentityType = Field(..., description="The type of the identity.")
     project_id: Optional[str] = Field(None, description="The project id of the identity, if applicable.")
     agent_ids: Optional[List[str]] = Field(None, description="The agent ids that are associated with the identity.")
+    properties: Optional[List[IdentityProperty]] = Field(None, description="List of properties associated with the identity.")
 
 
 class IdentityUpdate(LettaBase):
+    identifier_key: Optional[str] = Field(None, description="External, user-generated identifier key of the identity.")
     name: Optional[str] = Field(None, description="The name of the identity.")
     identity_type: Optional[IdentityType] = Field(None, description="The type of the identity.")
     agent_ids: Optional[List[str]] = Field(None, description="The agent ids that are associated with the identity.")
+    properties: Optional[List[IdentityProperty]] = Field(None, description="List of properties associated with the identity.")

{letta_nightly-0.6.28.dev20250220163833 → letta_nightly-0.6.29.dev20250221104040}/letta/server/rest_api/chat_completions_interface.py

@@ -56,6 +56,7 @@ class ChatCompletionsStreamingInterface(AgentChunkStreamingInterface):
         self.current_function_name = ""
         self.current_function_arguments = []
         self.current_json_parse_result = {}
+        self._found_message_tool_kwarg = False
 
         # Internal chunk buffer and event for async notification
         self._chunks = deque()
@@ -153,12 +154,13 @@ class ChatCompletionsStreamingInterface(AgentChunkStreamingInterface):
         """No-op retained for interface compatibility."""
         return
 
-    def process_chunk(self, chunk: ChatCompletionChunkResponse, message_id: str, message_date: datetime) -> None:
+    def process_chunk(
+        self, chunk: ChatCompletionChunkResponse, message_id: str, message_date: datetime, expect_reasoning_content: bool = False
+    ) -> None:
         """
         Called externally with a ChatCompletionChunkResponse. Transforms
         it if necessary, then enqueues partial messages for streaming back.
         """
-        # print("RECEIVED CHUNK...")
         processed_chunk = self._process_chunk_to_openai_style(chunk)
         if processed_chunk is not None:
             self._push_to_buffer(processed_chunk)
@@ -197,6 +199,10 @@ class ChatCompletionsStreamingInterface(AgentChunkStreamingInterface):
         content (especially from a 'send_message' tool) is exposed as text
         deltas in 'content'. Otherwise, pass through or yield finish reasons.
         """
+        # If we've already sent the final chunk, ignore everything.
+        if self._found_message_tool_kwarg:
+            return None
+
         choice = chunk.choices[0]
         delta = choice.delta
 
@@ -219,25 +225,43 @@ class ChatCompletionsStreamingInterface(AgentChunkStreamingInterface):
                 combined_args = "".join(self.current_function_arguments)
                 parsed_args = OptimisticJSONParser().parse(combined_args)
 
-                # If the parsed result is different
-                # This is an edge case we need to consider. E.g. if the last streamed token is '}', we shouldn't stream that out
-                if parsed_args != self.current_json_parse_result:
-                    self.current_json_parse_result = parsed_args
-                    # If we can see a "message" field, return it as partial content
-                    if self.assistant_message_tool_kwarg in parsed_args and parsed_args[self.assistant_message_tool_kwarg]:
-                        return ChatCompletionChunk(
-                            id=chunk.id,
-                            object=chunk.object,
-                            created=chunk.created.timestamp(),
-                            model=chunk.model,
-                            choices=[
-                                Choice(
-                                    index=choice.index,
-                                    delta=ChoiceDelta(content=self.current_function_arguments[-1], role=self.ASSISTANT_STR),
-                                    finish_reason=None,
-                                )
-                            ],
-                        )
+                # TODO: Make this less brittle! This depends on `message` coming first!
+                # This is a heuristic we use to know if we're done with the `message` part of `send_message`
+                if len(parsed_args.keys()) > 1:
+                    self._found_message_tool_kwarg = True
+                    return ChatCompletionChunk(
+                        id=chunk.id,
+                        object=chunk.object,
+                        created=chunk.created.timestamp(),
+                        model=chunk.model,
+                        choices=[
+                            Choice(
+                                index=choice.index,
+                                delta=ChoiceDelta(),
+                                finish_reason="stop",
+                            )
+                        ],
+                    )
+                else:
+                    # If the parsed result is different
+                    # This is an edge case we need to consider. E.g. if the last streamed token is '}', we shouldn't stream that out
+                    if parsed_args != self.current_json_parse_result:
+                        self.current_json_parse_result = parsed_args
+                        # If we can see a "message" field, return it as partial content
+                        if self.assistant_message_tool_kwarg in parsed_args and parsed_args[self.assistant_message_tool_kwarg]:
+                            return ChatCompletionChunk(
+                                id=chunk.id,
+                                object=chunk.object,
+                                created=chunk.created.timestamp(),
+                                model=chunk.model,
+                                choices=[
+                                    Choice(
+                                        index=choice.index,
+                                        delta=ChoiceDelta(content=self.current_function_arguments[-1], role=self.ASSISTANT_STR),
+                                        finish_reason=None,
+                                    )
+                                ],
+                            )
 
         # If there's a finish reason, pass that along
         if choice.finish_reason is not None: