Platform for stateful agents: AI with advanced memory that can learn and self-improve over time.
feat: prevent summarization mid approval request (#5172)
cthomasec2375f332af
diff --git a/letta/agents/letta_agent_v2.py b/letta/agents/letta_agent_v2.pyindex 3c672b27d..df44d13d3 100644--- a/letta/agents/letta_agent_v2.py+++ b/letta/agents/letta_agent_v2.py@@ -694,7 +694,7 @@ class LettaAgentV2(BaseAgentV2):curr_dynamic_section = extract_dynamic_section(curr_system_message_text)- # refresh files+ # refresh filesagent_state = await self.agent_manager.refresh_file_blocks(agent_state=agent_state, actor=self.actor)# generate just the memory string with current state for comparison@@ -1152,6 +1152,11 @@ class LettaAgentV2(BaseAgentV2):total_tokens: int | None = None,force: bool = False,) -> list[Message]:+ # always skip summarization if last message is an approval request message+ latest_messages = in_context_messages + new_letta_messages+ if latest_messages[-1].role == "approval" and len(latest_messages[-1].tool_calls) > 0:+ return in_context_messages+# If total tokens is reached, we truncate down# 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