chore: move things to debug logging (#6610)

letta/agents/helpers.py

@@ -157,6 +157,9 @@ async def _prepare_in_context_messages_no_persist_async(
     if input_messages[0].type == "approval":
         # User is trying to send an approval response
         if current_in_context_messages and current_in_context_messages[-1].role != "approval":
+            logger.warn(
+                f"Cannot process approval response: No tool call is currently awaiting approval. Last message: {current_in_context_messages[-1]}"
+            )
             raise ValueError(
                 "Cannot process approval response: No tool call is currently awaiting approval. "
                 "Please send a regular message to interact with the agent."

letta/agents/letta_agent_v3.py

@@ -675,12 +675,18 @@ class LettaAgentV3(LettaAgentV2):
                         raise e
                     except Exception as e:
                         if isinstance(e, ContextWindowExceededError) and llm_request_attempt < summarizer_settings.max_summarizer_retries:
+                            # Retry case
+                            self.logger.info(
+                                f"Context window exceeded (error {e}), trying to compact messages attempt {llm_request_attempt + 1} of {summarizer_settings.max_summarizer_retries + 1}"
+                            )
                             # checkpoint summarized messages
                             # TODO: might want to delay this checkpoint in case of corrupated state
                             try:
                                 summary_message, messages = await self.compact(
                                     messages, trigger_threshold=self.agent_state.llm_config.context_window
                                 )
+                                self.logger.info("Summarization succeeded, continuing to retry LLM request")
+                                continue
                             except SystemPromptTokenExceededError:
                                 self.stop_reason = LettaStopReason(
                                     stop_reason=StopReasonType.context_window_overflow_in_system_prompt.value
@@ -783,6 +789,9 @@ class LettaAgentV3(LettaAgentV2):
 
             # check compaction
             if self.context_token_estimate is not None and self.context_token_estimate > self.agent_state.llm_config.context_window:
+                self.logger.info(
+                    f"Context window exceeded (current: {self.context_token_estimate}, threshold: {self.agent_state.llm_config.context_window}), trying to compact messages"
+                )
                 summary_message, messages = await self.compact(messages, trigger_threshold=self.agent_state.llm_config.context_window)
                 # TODO: persist + return the summary message
                 # TODO: convert this to a SummaryMessage

letta/monitoring/event_loop_watchdog.py

@@ -97,7 +97,7 @@ class EventLoopWatchdog:
                     pass
 
                 # ALWAYS log every check to prove watchdog is alive
-                logger.info(
+                logger.debug(
                     f"WATCHDOG_CHECK: heartbeat_age={time_since_heartbeat:.1f}s, consecutive_hangs={consecutive_hangs}, tasks={task_count}"
                 )
 

letta/server/rest_api/app.py

@@ -721,6 +721,7 @@ def start_server(
                 timeout_keep_alive=settings.uvicorn_timeout_keep_alive,
                 ssl_keyfile="certs/localhost-key.pem",
                 ssl_certfile="certs/localhost.pem",
+                access_log=False,
             )
 
     else:
@@ -759,4 +760,5 @@ def start_server(
                 workers=settings.uvicorn_workers,
                 reload=reload or settings.uvicorn_reload,
                 timeout_keep_alive=settings.uvicorn_timeout_keep_alive,
+                access_log=False,
             )

letta/server/rest_api/middleware/logging.py

@@ -109,7 +109,7 @@ class LoggingMiddleware(BaseHTTPMiddleware):
             if context:
                 update_log_context(**context)
 
-            logger.info(
+            logger.debug(
                 f"Incoming request: {request.method} {request.url.path}",
                 extra={
                     "method": request.method,

letta/services/context_window_calculator/token_counter.py

@@ -292,7 +292,7 @@ def create_token_counter(
     if use_gemini:
         client = LLMClient.create(provider_type=model_endpoint_type, actor=actor)
         token_counter = GeminiTokenCounter(client, model)
-        logger.info(
+        logger.debug(
             f"Using GeminiTokenCounter for agent_id={agent_id}, model={model}, "
             f"model_endpoint_type={model_endpoint_type}, "
             f"environment={settings.environment}"
@@ -301,14 +301,14 @@ def create_token_counter(
         anthropic_client = LLMClient.create(provider_type=ProviderType.anthropic, actor=actor)
         counter_model = model if model_endpoint_type == "anthropic" else None
         token_counter = AnthropicTokenCounter(anthropic_client, counter_model)
-        logger.info(
+        logger.debug(
             f"Using AnthropicTokenCounter for agent_id={agent_id}, model={counter_model}, "
             f"model_endpoint_type={model_endpoint_type}, "
             f"environment={settings.environment}"
         )
     else:
         token_counter = ApproxTokenCounter()
-        logger.info(
+        logger.debug(
             f"Using ApproxTokenCounter for agent_id={agent_id}, model={model}, "
             f"model_endpoint_type={model_endpoint_type}, "
             f"environment={settings.environment}"