Skip to content

armgit / letta

Public Imported from GitHub
0

Platform for stateful agents: AI with advanced memory that can learn and self-improve over time.

feat: support approval requests for parallel tool calls (#5538)

cthomas69b15d606c12
diff --git a/letta/agents/letta_agent.py b/letta/agents/letta_agent.py
index 1da8bf3f8..81f38488b 100644
--- a/letta/agents/letta_agent.py
+++ b/letta/agents/letta_agent.py
@@ -1,3 +1,4 @@
+import json
import uuid
from collections.abc import AsyncGenerator
from datetime import datetime
@@ -18,7 +19,7 @@ from letta.agents.helpers import (
_safe_load_tool_call_str,
generate_step_id,
)
-from letta.constants import DEFAULT_MAX_STEPS, NON_USER_MSG_PREFIX
+from letta.constants import DEFAULT_MAX_STEPS, NON_USER_MSG_PREFIX, REQUEST_HEARTBEAT_PARAM
from letta.errors import ContextWindowExceededError
from letta.helpers import ToolRulesSolver
from letta.helpers.datetime_helpers import AsyncTimer, get_utc_time, get_utc_timestamp_ns, ns_to_ms
@@ -41,7 +42,7 @@ from letta.schemas.letta_response import LettaResponse
from letta.schemas.letta_stop_reason import LettaStopReason, StopReasonType
from letta.schemas.llm_config import LLMConfig
from letta.schemas.message import Message, MessageCreateBase
-from letta.schemas.openai.chat_completion_response import ToolCall, UsageStatistics
+from letta.schemas.openai.chat_completion_response import FunctionCall, ToolCall, UsageStatistics
from letta.schemas.provider_trace import ProviderTraceCreate
from letta.schemas.step import StepProgression
from letta.schemas.step_metrics import StepMetrics
@@ -1698,14 +1699,11 @@ class LettaAgent(BaseAgent):
request_heartbeat=request_heartbeat,
)
if not is_approval and tool_rules_solver.is_requires_approval_tool(tool_call_name):
+ tool_args[REQUEST_HEARTBEAT_PARAM] = request_heartbeat
approval_message = create_approval_request_message_from_llm_response(
agent_id=agent_state.id,
model=agent_state.llm_config.model,
- function_name=tool_call_name,
- function_arguments=tool_args,
- tool_call_id=tool_call_id,
- actor=self.actor,
- continue_stepping=request_heartbeat,
+ tool_calls=[ToolCall(id=tool_call_id, function=FunctionCall(name=tool_call_name, arguments=json.dumps(tool_args)))],
reasoning_content=reasoning_content,
pre_computed_assistant_message_id=pre_computed_assistant_message_id,
step_id=step_id,
diff --git a/letta/agents/letta_agent_v2.py b/letta/agents/letta_agent_v2.py
index c216e3721..eb5d15db2 100644
--- a/letta/agents/letta_agent_v2.py
+++ b/letta/agents/letta_agent_v2.py
@@ -1,4 +1,4 @@
-import asyncio
+import json
import uuid
from datetime import datetime
from typing import AsyncGenerator, Optional, Tuple
@@ -19,7 +19,7 @@ from letta.agents.helpers import (
_safe_load_tool_call_str,
generate_step_id,
)
-from letta.constants import DEFAULT_MAX_STEPS, NON_USER_MSG_PREFIX
+from letta.constants import DEFAULT_MAX_STEPS, NON_USER_MSG_PREFIX, REQUEST_HEARTBEAT_PARAM
from letta.errors import ContextWindowExceededError, LLMError
from letta.helpers import ToolRulesSolver
from letta.helpers.datetime_helpers import get_utc_time, get_utc_timestamp_ns, ns_to_ms
@@ -37,7 +37,7 @@ from letta.schemas.letta_message_content import OmittedReasoningContent, Reasoni
from letta.schemas.letta_response import LettaResponse
from letta.schemas.letta_stop_reason import LettaStopReason, StopReasonType
from letta.schemas.message import Message, MessageCreate, MessageUpdate
-from letta.schemas.openai.chat_completion_response import ToolCall, UsageStatistics
+from letta.schemas.openai.chat_completion_response import FunctionCall, ToolCall, UsageStatistics
from letta.schemas.step import Step, StepProgression
from letta.schemas.step_metrics import StepMetrics
from letta.schemas.tool_execution_result import ToolExecutionResult
@@ -933,14 +933,11 @@ class LettaAgentV2(BaseAgentV2):
)
if not is_approval and tool_rules_solver.is_requires_approval_tool(tool_call_name):
+ tool_args[REQUEST_HEARTBEAT_PARAM] = request_heartbeat
approval_message = create_approval_request_message_from_llm_response(
agent_id=agent_state.id,
model=agent_state.llm_config.model,
- function_name=tool_call_name,
- function_arguments=tool_args,
- tool_call_id=tool_call_id,
- actor=self.actor,
- continue_stepping=request_heartbeat,
+ tool_calls=[ToolCall(id=tool_call_id, function=FunctionCall(name=tool_call_name, arguments=json.dumps(tool_args)))],
reasoning_content=reasoning_content,
pre_computed_assistant_message_id=pre_computed_assistant_message_id,
step_id=step_id,
diff --git a/letta/agents/letta_agent_v3.py b/letta/agents/letta_agent_v3.py
index 2bc72b4dc..fa0931292 100644
--- a/letta/agents/letta_agent_v3.py
+++ b/letta/agents/letta_agent_v3.py
@@ -1,4 +1,5 @@
import asyncio
+import json
import uuid
from typing import AsyncGenerator, Optional
@@ -31,7 +32,7 @@ from letta.schemas.letta_message_content import OmittedReasoningContent, Reasoni
from letta.schemas.letta_response import LettaResponse
from letta.schemas.letta_stop_reason import LettaStopReason, StopReasonType
from letta.schemas.message import Message, MessageCreate, ToolReturn
-from letta.schemas.openai.chat_completion_response import ToolCall, UsageStatistics
+from letta.schemas.openai.chat_completion_response import FunctionCall, ToolCall, UsageStatistics
from letta.schemas.step import StepProgression
from letta.schemas.step_metrics import StepMetrics
from letta.schemas.tool_execution_result import ToolExecutionResult
@@ -676,6 +677,38 @@ class LettaAgentV3(LettaAgentV2):
)
return persisted_messages, continue_stepping, stop_reason
+ # 2. Check whether tool call requires approval
+ tool_names = [tc.function.name for tc in tool_calls]
+ requires_approval = (
+ not is_approval_response
+ and tool_names
+ and any(tool_rules_solver.is_requires_approval_tool(tool_name) for tool_name in tool_names)
+ )
+ if requires_approval:
+ approval_message = create_approval_request_message_from_llm_response(
+ agent_id=agent_state.id,
+ model=agent_state.llm_config.model,
+ tool_calls=tool_calls,
+ reasoning_content=content,
+ pre_computed_assistant_message_id=pre_computed_assistant_message_id,
+ step_id=step_id,
+ run_id=run_id,
+ )
+ messages_to_persist = (initial_messages or []) + [approval_message]
+
+ for message in messages_to_persist:
+ if message.run_id is None:
+ message.run_id = run_id
+
+ persisted_messages = await self.message_manager.create_many_messages_async(
+ messages_to_persist,
+ actor=self.actor,
+ run_id=run_id,
+ project_id=agent_state.project_id,
+ template_id=agent_state.template_id,
+ )
+ return persisted_messages, False, LettaStopReason(stop_reason=StopReasonType.requires_approval.value)
+
if tool_returns:
assert len(tool_returns) == 1, "Only one tool return is supported"
tool_return = tool_returns[0]
@@ -752,44 +785,7 @@ class LettaAgentV3(LettaAgentV2):
# 4. Unified tool execution path (works for both single and multiple tools)
- # 4a. Check for single tool approval case (special handling required)
- if len(tool_calls) == 1 and not is_approval_response:
- single_tool = tool_calls[0]
- tool_name = single_tool.function.name
- tool_args = _safe_load_tool_call_str(single_tool.function.arguments)
- tool_args.pop(REQUEST_HEARTBEAT_PARAM, None)
- tool_args.pop(INNER_THOUGHTS_KWARG, None)
-
- if tool_rules_solver.is_requires_approval_tool(tool_name):
- tool_call_id = single_tool.id or f"call_{uuid.uuid4().hex[:8]}"
- approval_message = create_approval_request_message_from_llm_response(
- agent_id=agent_state.id,
- model=agent_state.llm_config.model,
- function_name=tool_name,
- function_arguments=tool_args,
- tool_call_id=tool_call_id,
- actor=self.actor,
- continue_stepping=True,
- reasoning_content=content,
- pre_computed_assistant_message_id=pre_computed_assistant_message_id,
- step_id=step_id,
- run_id=run_id,
- append_request_heartbeat=False,
- )
- messages_to_persist = (initial_messages or []) + [approval_message]
-
- for message in messages_to_persist:
- if message.run_id is None:
- message.run_id = run_id
-
- persisted_messages = await self.message_manager.create_many_messages_async(
- messages_to_persist,
- actor=self.actor,
- run_id=run_id,
- project_id=agent_state.project_id,
- template_id=agent_state.template_id,
- )
- return persisted_messages, False, LettaStopReason(stop_reason=StopReasonType.requires_approval.value)
+ # 4a. TODO
# 4b. Validate parallel tool calling constraints
if len(tool_calls) > 1:
diff --git a/letta/server/rest_api/utils.py b/letta/server/rest_api/utils.py
index ccefef891..1d60cae44 100644
--- a/letta/server/rest_api/utils.py
+++ b/letta/server/rest_api/utils.py
@@ -36,7 +36,7 @@ from letta.schemas.letta_message_content import (
TextContent,
)
from letta.schemas.llm_config import LLMConfig
-from letta.schemas.message import ApprovalCreate, ApprovalReturn, Message, MessageCreate, ToolReturn
+from letta.schemas.message import ApprovalCreate, Message, MessageCreate, ToolReturn
from letta.schemas.tool_execution_result import ToolExecutionResult
from letta.schemas.usage import LettaUsageStatistics
from letta.schemas.user import User
@@ -172,16 +172,7 @@ def create_input_messages(input_messages: List[MessageCreate], agent_id: str, ti
def create_approval_response_message_from_input(
agent_state: AgentState, input_message: ApprovalCreate, run_id: Optional[str] = None
) -> List[Message]:
- if not input_message.approvals:
- input_message.approvals = [
- ApprovalCreate(approve=input_message.approve, tool_call_id=input_message.approval_request_id, reason=input_message.reason)
- ]
- approval_messages = [
- a for a in input_message.approvals if isinstance(a, ApprovalReturn) or (isinstance(a, dict) and a.get("type") == "approval")
- ]
- first_approval = approval_messages[0] if approval_messages else None
-
- def maybe_convert_return_message(maybe_tool_return):
+ def maybe_convert_tool_return_message(maybe_tool_return: LettaToolReturn):
if isinstance(maybe_tool_return, LettaToolReturn):
packaged_function_response = package_function_response(
maybe_tool_return.status == "success", maybe_tool_return.tool_return, agent_state.timezone
@@ -200,10 +191,10 @@ def create_approval_response_message_from_input(
role=MessageRole.approval,
agent_id=agent_state.id,
model=agent_state.llm_config.model,
- approval_request_id=first_approval.tool_call_id if first_approval else input_message.approval_request_id,
- approve=first_approval.approve if first_approval else input_message.approve,
- denial_reason=first_approval.reason if first_approval else input_message.reason,
- approvals=[maybe_convert_return_message(approval) for approval in input_message.approvals],
+ approval_request_id=input_message.approval_request_id,
+ approve=input_message.approve,
+ denial_reason=input_message.reason,
+ approvals=[maybe_convert_tool_return_message(approval) for approval in input_message.approvals],
run_id=run_id,
)
]
@@ -212,29 +203,24 @@ def create_approval_response_message_from_input(
def create_approval_request_message_from_llm_response(
agent_id: str,
model: str,
- function_name: str,
- function_arguments: Dict,
- tool_call_id: str,
- actor: User,
- continue_stepping: bool = False,
+ tool_calls: List[OpenAIToolCall],
reasoning_content: Optional[List[Union[TextContent, ReasoningContent, RedactedReasoningContent, OmittedReasoningContent]]] = None,
pre_computed_assistant_message_id: Optional[str] = None,
step_id: str | None = None,
run_id: str = None,
- append_request_heartbeat: bool = True,
) -> Message:
# Construct the tool call with the assistant's message
- # Optionally set request_heartbeat in tool args (v2 behavior only)
- if append_request_heartbeat:
- function_arguments[REQUEST_HEARTBEAT_PARAM] = continue_stepping
- tool_call = OpenAIToolCall(
- id=tool_call_id,
- function=OpenAIFunction(
- name=function_name,
- arguments=json.dumps(function_arguments),
- ),
- type="function",
- )
+ oai_tool_calls = [
+ OpenAIToolCall(
+ id=tool_call.id,
+ function=OpenAIFunction(
+ name=tool_call.function.name,
+ arguments=tool_call.function.arguments,
+ ),
+ type="function",
+ )
+ for tool_call in tool_calls
+ ]
# TODO: Use ToolCallContent instead of tool_calls
# TODO: This helps preserve ordering
approval_message = Message(
@@ -242,8 +228,8 @@ def create_approval_request_message_from_llm_response(
content=reasoning_content if reasoning_content else [],
agent_id=agent_id,
model=model,
- tool_calls=[tool_call],
- tool_call_id=tool_call_id,
+ tool_calls=oai_tool_calls,
+ tool_call_id=oai_tool_calls[0].id,
created_at=get_utc_time(),
step_id=step_id,
run_id=run_id,