Platform for stateful agents: AI with advanced memory that can learn and self-improve over time.
fix: update step with resolved model info even when auto mode resolution fails partway (#10300)
cthomas775b63eca227
diff --git a/letta/agents/letta_agent_v3.py b/letta/agents/letta_agent_v3.pyindex 5e81b7bfc..8dabe5784 100644--- a/letta/agents/letta_agent_v3.py+++ b/letta/agents/letta_agent_v3.py@@ -1046,40 +1046,45 @@ class LettaAgentV3(LettaAgentV2):primary_handle = ""if is_auto_mode:- routing_client = await get_llm_routing_client()- active_llm_config, is_primary, primary_handle = await routing_client.resolve_auto_mode_config(- stored_llm_config=self.agent_state.llm_config,- actor=self.actor,- )- if not is_primary:- self.logger.info(f"[LLM ROUTER]: primary {primary_handle} rerouted, falling back to {active_llm_config.handle}")- # Content-based rerouting (e.g. images → vision-capable model)- active_llm_config = routing_client.apply_reroute_rules(- resolved_config=active_llm_config,- messages=messages,- stored_llm_config=self.agent_state.llm_config,- agent_state=self.agent_state,- )- active_llm_client = LLMClient.create(- provider_type=active_llm_config.model_endpoint_type,- put_inner_thoughts_first=True,- actor=self.actor,- )- # Update the adapter to use the resolved client and config- llm_adapter.llm_client = active_llm_client- llm_adapter.llm_config = active_llm_config- # Update persisted step with resolved model info so billing can- # identify the actual model and charge at the correct rate.- # Keep model_handle as the original auto handle so we can still- # identify that the user was on auto mode.- await self.step_manager.update_step_resolved_model_async(- actor=self.actor,- step_id=step_id,- provider_name=active_llm_config.model_endpoint_type,- provider_category=active_llm_config.provider_category or "base",- model=active_llm_config.model,- model_endpoint=active_llm_config.model_endpoint,- )+ resolved_llm_config = None+ try:+ routing_client = await get_llm_routing_client()+ active_llm_config, is_primary, primary_handle = await routing_client.resolve_auto_mode_config(+ stored_llm_config=self.agent_state.llm_config,+ actor=self.actor,+ )+ resolved_llm_config = active_llm_config+ if not is_primary:+ self.logger.info(f"[LLM ROUTER]: primary {primary_handle} rerouted, falling back to {active_llm_config.handle}")+ # Content-based rerouting (e.g. images → vision-capable model)+ active_llm_config = routing_client.apply_reroute_rules(+ resolved_config=active_llm_config,+ messages=messages,+ stored_llm_config=self.agent_state.llm_config,+ agent_state=self.agent_state,+ )+ resolved_llm_config = active_llm_config+ active_llm_client = LLMClient.create(+ provider_type=active_llm_config.model_endpoint_type,+ put_inner_thoughts_first=True,+ actor=self.actor,+ )+ # Update the adapter to use the resolved client and config+ llm_adapter.llm_client = active_llm_client+ llm_adapter.llm_config = active_llm_config+ finally:+ # Update persisted step with resolved model info so billing can+ # identify the actual model and charge at the correct rate,+ # even if resolution fails partway through.+ if resolved_llm_config is not None:+ await self.step_manager.update_step_resolved_model_async(+ actor=self.actor,+ step_id=step_id,+ provider_name=resolved_llm_config.model_endpoint_type,+ provider_category=resolved_llm_config.provider_category or "base",+ model=resolved_llm_config.model,+ model_endpoint=resolved_llm_config.model_endpoint,+ )else:active_llm_config = self.agent_state.llm_configactive_llm_client = self.llm_client