From 077d0e2f9a68bb245ee3dae1e0415b4f2502bdc3 Mon Sep 17 00:00:00 2001
From: Kian Jones <11655409+kianjones9@users.noreply.github.com>
Date: Tue, 11 Nov 2025 18:53:02 -0800
Subject: [PATCH] feat: dump stack trace on segfault (#6121)

* dump stack trace on segfault

* log tokenizer
---
 letta/server/rest_api/app.py                  |  4 ++
 letta/services/agent_manager.py               | 10 +++++
 .../token_counter.py                          | 38 +++++++++++++++++--
 3 files changed, 49 insertions(+), 3 deletions(-)

diff --git a/letta/server/rest_api/app.py b/letta/server/rest_api/app.py
index dbeeec8b..e6fd174f 100644
--- a/letta/server/rest_api/app.py
+++ b/letta/server/rest_api/app.py
@@ -1,3 +1,4 @@
+import faulthandler
 import importlib.util
 import json
 import logging
@@ -10,6 +11,9 @@ from pathlib import Path
 from typing import Optional
 
 import uvicorn
+
+# Enable Python fault handler to get stack traces on segfaults
+faulthandler.enable()
 from fastapi import FastAPI, Request
 from fastapi.exceptions import RequestValidationError
 from fastapi.responses import JSONResponse
diff --git a/letta/services/agent_manager.py b/letta/services/agent_manager.py
index 9e8645d6..7db7e626 100644
--- a/letta/services/agent_manager.py
+++ b/letta/services/agent_manager.py
@@ -3239,8 +3239,18 @@ class AgentManager:
             model = agent_state.llm_config.model if agent_state.llm_config.model_endpoint_type == "anthropic" else None
 
             token_counter = AnthropicTokenCounter(anthropic_client, model)  # noqa
+            logger.info(
+                f"Using AnthropicTokenCounter for agent_id={agent_id}, model={model}, "
+                f"model_endpoint_type={agent_state.llm_config.model_endpoint_type}, "
+                f"environment={settings.environment}"
+            )
         else:
             token_counter = TiktokenCounter(agent_state.llm_config.model)
+            logger.info(
+                f"Using TiktokenCounter for agent_id={agent_id}, model={agent_state.llm_config.model}, "
+                f"model_endpoint_type={agent_state.llm_config.model_endpoint_type}, "
+                f"environment={settings.environment}"
+            )
 
         return await calculator.calculate_context_window(
             agent_state=agent_state,
diff --git a/letta/services/context_window_calculator/token_counter.py b/letta/services/context_window_calculator/token_counter.py
index a940cf55..12833fda 100644
--- a/letta/services/context_window_calculator/token_counter.py
+++ b/letta/services/context_window_calculator/token_counter.py
@@ -90,9 +90,24 @@ class TiktokenCounter(TokenCounter):
         ttl_s=3600,  # cache for 1 hour
     )
     async def count_text_tokens(self, text: str) -> int:
+        from letta.log import get_logger
+
+        logger = get_logger(__name__)
+
         if not text:
             return 0
-        return count_tokens(text)
+
+        text_length = len(text)
+        text_preview = text[:100] + "..." if len(text) > 100 else text
+        logger.debug(f"TiktokenCounter.count_text_tokens: model={self.model}, text_length={text_length}, preview={repr(text_preview)}")
+
+        try:
+            result = count_tokens(text)
+            logger.debug(f"TiktokenCounter.count_text_tokens: completed successfully, tokens={result}")
+            return result
+        except Exception as e:
+            logger.error(f"TiktokenCounter.count_text_tokens: FAILED with {type(e).__name__}: {e}, text_length={text_length}")
+            raise
 
     @trace_method
     @async_redis_cache(
@@ -102,11 +117,28 @@ class TiktokenCounter(TokenCounter):
         ttl_s=3600,  # cache for 1 hour
     )
     async def count_message_tokens(self, messages: List[Dict[str, Any]]) -> int:
+        from letta.log import get_logger
+
+        logger = get_logger(__name__)
+
         if not messages:
             return 0
-        from letta.local_llm.utils import num_tokens_from_messages
 
-        return num_tokens_from_messages(messages=messages, model=self.model)
+        num_messages = len(messages)
+        total_content_length = sum(len(str(m.get("content", ""))) for m in messages)
+        logger.debug(
+            f"TiktokenCounter.count_message_tokens: model={self.model}, num_messages={num_messages}, total_content_length={total_content_length}"
+        )
+
+        try:
+            from letta.local_llm.utils import num_tokens_from_messages
+
+            result = num_tokens_from_messages(messages=messages, model=self.model)
+            logger.debug(f"TiktokenCounter.count_message_tokens: completed successfully, tokens={result}")
+            return result
+        except Exception as e:
+            logger.error(f"TiktokenCounter.count_message_tokens: FAILED with {type(e).__name__}: {e}, num_messages={num_messages}")
+            raise
 
     @trace_method
     @async_redis_cache(