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.

fix: llm error interrupting stream for agent loop (#5696)

cthomas06d2cde43d7c
diff --git a/letta/agents/letta_agent_v2.py b/letta/agents/letta_agent_v2.py
index e42438db7..d36e2fa71 100644
--- a/letta/agents/letta_agent_v2.py
+++ b/letta/agents/letta_agent_v2.py
@@ -1164,25 +1164,29 @@ class LettaAgentV2(BaseAgentV2):
# TODO: This can be broken by bad configs, e.g. lower bound too high, initial messages too fat, etc.
# TODO: `force` and `clear` seem to no longer be used, we should remove
if not skip_summarization:
- if force or (total_tokens and total_tokens > self.agent_state.llm_config.context_window):
- self.logger.warning(
- f"Total tokens {total_tokens} exceeds configured max tokens {self.agent_state.llm_config.context_window}, forcefully clearing message history."
- )
- new_in_context_messages, updated = await self.summarizer.summarize(
- in_context_messages=in_context_messages,
- new_letta_messages=new_letta_messages,
- force=True,
- clear=True,
- )
- else:
- # NOTE (Sarah): Seems like this is doing nothing?
- self.logger.info(
- f"Total tokens {total_tokens} does not exceed configured max tokens {self.agent_state.llm_config.context_window}, passing summarizing w/o force."
- )
- new_in_context_messages, updated = await self.summarizer.summarize(
- in_context_messages=in_context_messages,
- new_letta_messages=new_letta_messages,
- )
+ try:
+ if force or (total_tokens and total_tokens > self.agent_state.llm_config.context_window):
+ self.logger.warning(
+ f"Total tokens {total_tokens} exceeds configured max tokens {self.agent_state.llm_config.context_window}, forcefully clearing message history."
+ )
+ new_in_context_messages, updated = await self.summarizer.summarize(
+ in_context_messages=in_context_messages,
+ new_letta_messages=new_letta_messages,
+ force=True,
+ clear=True,
+ )
+ else:
+ # NOTE (Sarah): Seems like this is doing nothing?
+ self.logger.info(
+ f"Total tokens {total_tokens} does not exceed configured max tokens {self.agent_state.llm_config.context_window}, passing summarizing w/o force."
+ )
+ new_in_context_messages, updated = await self.summarizer.summarize(
+ in_context_messages=in_context_messages,
+ new_letta_messages=new_letta_messages,
+ )
+ except Exception as e:
+ self.logger.error(f"Failed to summarize conversation history: {e}")
+ new_in_context_messages = in_context_messages + new_letta_messages
else:
new_in_context_messages = in_context_messages + new_letta_messages
diff --git a/letta/agents/letta_agent_v3.py b/letta/agents/letta_agent_v3.py
index 7dc284021..a7aee2130 100644
--- a/letta/agents/letta_agent_v3.py
+++ b/letta/agents/letta_agent_v3.py
@@ -224,10 +224,10 @@ class LettaAgentV3(LettaAgentV2):
force=False,
)
- except:
- if self.stop_reason and not first_chunk:
- yield f"data: {self.stop_reason.model_dump_json()}\n\n"
- raise
+ except Exception as e:
+ self.logger.error("Error during agent stream", e)
+ if first_chunk:
+ raise # only raise if first chunk has not been streamed yet
if run_id:
letta_messages = Message.to_letta_messages_from_list(
@@ -424,6 +424,7 @@ class LettaAgentV3(LettaAgentV2):
force=True,
)
else:
+ self.stop_reason = LettaStopReason(stop_reason=StopReasonType.llm_api_error.value)
raise e
step_progression, step_metrics = self._step_checkpoint_llm_request_finish(
@@ -505,8 +506,6 @@ class LettaAgentV3(LettaAgentV2):
# TODO should we be logging this even if persisted_messages is empty? Technically, there still was an LLM call
step_progression, step_metrics = await self._step_checkpoint_finish(step_metrics, agent_step_span, logged_step)
except Exception as e:
- import traceback
-
self.logger.warning(f"Error during step processing: {e}")
self.job_update_metadata = {"error": str(e)}