Fimeg/letta-server
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

fix: remove duplicate provider trace logging in LettaAgent (#9276)

Browse Source 
Provider traces were being created twice per step:
1. Via `request_async_with_telemetry` / `log_provider_trace_async` in LLMClient
2. Via direct `create_provider_trace_async` calls in LettaAgent

This caused duplicate records in provider_trace_metadata (Postgres) and
llm_traces (ClickHouse) for every agent step.

Removed the redundant direct calls since telemetry is now centralized
in the LLM client layer.

🤖 Generated with [Letta Code](https://letta.com)

Co-authored-by: Letta <noreply@letta.com>
This commit is contained in:
Kian Jones
2026-02-04 11:28:49 -08:00
committed by Caren Thomas
parent a2993da29a
commit cc3b0f13a6
1 changed files with 0 additions and 75 deletions
Download Patch File Download Diff File Expand all files Collapse all files

75
letta/agents/letta_agent.py
View File

@@ -49,7 +49,6 @@ from letta.schemas.openai.chat_completion_response import (
UsageStatisticsCompletionTokenDetails, UsageStatisticsCompletionTokenDetails,
UsageStatisticsPromptTokenDetails, UsageStatisticsPromptTokenDetails,
) )
from letta.schemas.provider_trace import ProviderTrace
from letta.schemas.step import StepProgression from letta.schemas.step import StepProgression
from letta.schemas.step_metrics import StepMetrics from letta.schemas.step_metrics import StepMetrics
from letta.schemas.tool_execution_result import ToolExecutionResult from letta.schemas.tool_execution_result import ToolExecutionResult
@@ -409,25 +408,6 @@ class LettaAgent(BaseAgent):
agent_step_span.add_event(name="step_ms", attributes={"duration_ms": ns_to_ms(step_ns)}) agent_step_span.add_event(name="step_ms", attributes={"duration_ms": ns_to_ms(step_ns)})
agent_step_span.end() agent_step_span.end()
# Log LLM Trace
if settings.track_provider_trace:
await self.telemetry_manager.create_provider_trace_async(
actor=self.actor,
provider_trace=ProviderTrace(
request_json=request_data,
response_json=response_data,
step_id=step_id,
agent_id=self.agent_id,
agent_tags=agent_state.tags,
run_id=self.current_run_id,
call_type=LLMCallType.agent_step,
org_id=self.actor.organization_id,
user_id=self.actor.id,
llm_config=self.agent_state.llm_config.model_dump() if self.agent_state.llm_config else None,
),
)
step_progression = StepProgression.LOGGED_TRACE
# stream step # stream step
# TODO: improve TTFT # TODO: improve TTFT
filter_user_messages = [m for m in persisted_messages if m.role != "user"] filter_user_messages = [m for m in persisted_messages if m.role != "user"]
@@ -763,25 +743,6 @@ class LettaAgent(BaseAgent):
agent_step_span.add_event(name="step_ms", attributes={"duration_ms": ns_to_ms(step_ns)}) agent_step_span.add_event(name="step_ms", attributes={"duration_ms": ns_to_ms(step_ns)})
agent_step_span.end() agent_step_span.end()
# Log LLM Trace
if settings.track_provider_trace:
await self.telemetry_manager.create_provider_trace_async(
actor=self.actor,
provider_trace=ProviderTrace(
request_json=request_data,
response_json=response_data,
step_id=step_id,
agent_id=self.agent_id,
agent_tags=agent_state.tags,
run_id=self.current_run_id,
call_type=LLMCallType.agent_step,
org_id=self.actor.organization_id,
user_id=self.actor.id,
llm_config=self.agent_state.llm_config.model_dump() if self.agent_state.llm_config else None,
),
)
step_progression = StepProgression.LOGGED_TRACE
MetricRegistry().step_execution_time_ms_histogram.record(get_utc_timestamp_ns() - step_start, get_ctx_attributes()) MetricRegistry().step_execution_time_ms_histogram.record(get_utc_timestamp_ns() - step_start, get_ctx_attributes())
step_progression = StepProgression.FINISHED step_progression = StepProgression.FINISHED
@@ -1224,42 +1185,6 @@ class LettaAgent(BaseAgent):
# TODO (cliandy): the stream POST request span has ended at this point, we should tie this to the stream # TODO (cliandy): the stream POST request span has ended at this point, we should tie this to the stream
# log_event("agent.stream.llm_response.processed") # [4^] # log_event("agent.stream.llm_response.processed") # [4^]
# Log LLM Trace
# We are piecing together the streamed response here.
# Content here does not match the actual response schema as streams come in chunks.
if settings.track_provider_trace:
await self.telemetry_manager.create_provider_trace_async(
actor=self.actor,
provider_trace=ProviderTrace(
request_json=request_data,
response_json={
"content": {
"tool_call": tool_call.model_dump_json(),
"reasoning": [content.model_dump_json() for content in reasoning_content],
},
"id": interface.message_id,
"model": interface.model,
"role": "assistant",
# "stop_reason": "",
# "stop_sequence": None,
"type": "message",
"usage": {
"input_tokens": usage.prompt_tokens,
"output_tokens": usage.completion_tokens,
},
},
step_id=step_id,
agent_id=self.agent_id,
agent_tags=agent_state.tags,
run_id=self.current_run_id,
call_type=LLMCallType.agent_step,
org_id=self.actor.organization_id,
user_id=self.actor.id,
llm_config=self.agent_state.llm_config.model_dump() if self.agent_state.llm_config else None,
),
)
step_progression = StepProgression.LOGGED_TRACE
if persisted_messages[-1].role != "approval": if persisted_messages[-1].role != "approval":
# yields tool response as this is handled from Letta and not the response from the LLM provider # yields tool response as this is handled from Letta and not the response from the LLM provider
tool_return = [msg for msg in persisted_messages if msg.role == "tool"][-1].to_letta_messages()[0] tool_return = [msg for msg in persisted_messages if msg.role == "tool"][-1].to_letta_messages()[0]