fast_llm: Fix UnboundLocalError, add docstring

Previously, fast_llm was trying to access 'response' in the finally
block, which would fail with UnboundLocalError if the chat call raised
an exception (because of missing API key, etc.)

UnboundLocalError: cannot access local variable 'response' where it is
not associated with a value

Moved the return statement into the try block while keeping state
restoration in finally, ensuring proper error propagation while
maintaining conversation state.

This way, if there's an API key issue or any other problem, any errors
from `llm.interpreter.chat()` propagate up and  can be handled at the
appropriate level in the call stack.

Added a detailed docstring to clarify the purpose, behavior, and usage
of the `fast_llm` function.

Author: endolith

interpreter/core/computer/ai/ai.py

@@ -78,16 +78,36 @@ def chunk_responses(responses, tokens, llm):
 
 
 def fast_llm(llm, system_message, user_message):
+    """
+    Creates a temporary chat context to process a single query, then restores the original chat state.
+
+    This is used for auxiliary queries (like summarization) that shouldn't affect the main conversation.
+    Particularly important for local LLMs where creating new instances is expensive.
+
+    Args:
+        llm: The LLM instance to use (typically computer.interpreter.llm)
+        system_message: The system prompt for this specific query
+        user_message: The user message/content to process
+
+    Returns:
+        str: The LLM's response content
+
+    Note:
+        This function temporarily replaces the LLM's conversation state (messages and system prompt),
+        runs the query, then restores the original state. This allows us to run one-off queries
+        without disrupting the main conversation context.
+    """
     old_messages = llm.interpreter.messages
     old_system_message = llm.interpreter.system_message
     try:
         llm.interpreter.system_message = system_message
         llm.interpreter.messages = []
         response = llm.interpreter.chat(user_message)
+        return response[-1].get("content")
     finally:
+        # Always restore the old state (before returning)
         llm.interpreter.messages = old_messages
         llm.interpreter.system_message = old_system_message
-        return response[-1].get("content")
 
 
 def query_map_chunks(chunks, llm, query):