fix(backend): Handle ChatCompletionMessage objects in token counting functions

The actual issue was in the _msg_tokens function in prompt.py which expected
dict objects but was receiving ChatCompletionMessage objects. This fix:
- Updates _msg_tokens to convert non-dict messages to dicts using json.to_dict
- Updates compress_prompt to handle mixed message types
- Updates estimate_token_count to accept Any types in addition to dicts

This ensures that all message objects are properly converted to dictionaries
before accessing their properties, preventing AttributeError exceptions.

Co-authored-by: Nicholas Tindle <ntindle@users.noreply.github.com>

autogpt_platform/backend/backend/blocks/smart_decision_maker.py

@@ -220,7 +220,9 @@ class SmartDecisionMakerBlock(Block):
                 return missing_input
 
             conversation_history = data.get("conversation_history", [])
-            pending_tool_calls = get_pending_tool_calls(conversation_history)
+            # Convert to dict to handle ChatCompletionMessage objects
+            conversation_history_dicts = [json.to_dict(p) for p in conversation_history if p]
+            pending_tool_calls = get_pending_tool_calls(conversation_history_dicts)
             last_tool_output = data.get("last_tool_output")
 
             # Tool call is pending, wait for the tool output to be provided.
@@ -462,7 +464,7 @@ class SmartDecisionMakerBlock(Block):
         input_data.conversation_history = input_data.conversation_history or []
         prompt = [json.to_dict(p) for p in input_data.conversation_history if p]
 
-        pending_tool_calls = get_pending_tool_calls(input_data.conversation_history)
+        pending_tool_calls = get_pending_tool_calls(prompt)
         if pending_tool_calls and input_data.last_tool_output is None:
             raise ValueError(f"Tool call requires an output for {pending_tool_calls}")
 

autogpt_platform/backend/backend/util/prompt.py

@@ -15,11 +15,15 @@ def _tok_len(text: str, enc) -> int:
     return len(enc.encode(str(text)))
 
 
-def _msg_tokens(msg: dict, enc) -> int:
+def _msg_tokens(msg: dict | Any, enc) -> int:
     """
     OpenAI counts ≈3 wrapper tokens per chat message, plus 1 if "name"
     is present, plus the tokenised content length.
     """
+    # Handle ChatCompletionMessage objects by converting to dict
+    if not isinstance(msg, dict):
+        msg = json.to_dict(msg)
+
     WRAPPER = 3 + (1 if "name" in msg else 0)
     return WRAPPER + _tok_len(msg.get("content") or "", enc)
 
@@ -46,7 +50,7 @@ def _truncate_middle_tokens(text: str, enc, max_tok: int) -> str:
 
 
 def compress_prompt(
-    messages: list[dict],
+    messages: list[dict | Any],
     target_tokens: int,
     *,
     model: str = "gpt-4o",
@@ -94,7 +98,11 @@ def compress_prompt(
     list[dict]  – A *new* messages list that abides by the rules above.
     """
     enc = encoding_for_model(model)  # best-match tokenizer
-    msgs = deepcopy(messages)  # never mutate caller
+    # Convert any ChatCompletionMessage objects to dicts first
+    messages_as_dicts = [
+        json.to_dict(m) if not isinstance(m, dict) else m for m in messages
+    ]
+    msgs = deepcopy(messages_as_dicts)  # never mutate caller
 
     def total_tokens() -> int:
         """Current size of *msgs* in tokens."""
@@ -162,7 +170,7 @@ def compress_prompt(
 
 
 def estimate_token_count(
-    messages: list[dict],
+    messages: list[dict | Any],
     *,
     model: str = "gpt-4o",
 ) -> int: