refactor: simplify finish_reason mapping and update type hints

- Replace if/elif chain with dictionary mapping for better maintainability
- Add _FINISH_REASON_MAPPING constant with explicit string->enum mappings
- Remove Union[FinishReason, str] type - finish_reason is always enum now
- Update docstring to clarify all values mapped to enum with OTHER fallback
- Update tracing comment for accuracy (defensive fallback only)

Addresses Gemini Code Assist bot review feedback:
- Dictionary-based mapping improves code clarity
- Type hint now accurate (no Union needed since we always map to enum)
- Documentation reflects actual behavior (always enum, never string)

Author: aperepel

src/google/adk/models/lite_llm.py

@@ -64,6 +64,15 @@
 _NEW_LINE = "\n"
 _EXCLUDED_PART_FIELD = {"inline_data": {"data"}}
 
+# Mapping of LiteLLM finish_reason strings to FinishReason enum values
+_FINISH_REASON_MAPPING = {
+    "length": types.FinishReason.MAX_TOKENS,
+    "stop": types.FinishReason.STOP,
+    "tool_calls": types.FinishReason.STOP,
+    "function_call": types.FinishReason.STOP,
+    "content_filter": types.FinishReason.SAFETY,
+}
+
 
 class ChatCompletionFileUrlObject(TypedDict, total=False):
   file_data: str
@@ -508,18 +517,9 @@ def _model_response_to_generate_content_response(
     # Map LiteLLM finish_reason strings to FinishReason enum
     # This provides type consistency with Gemini native responses and avoids warnings
     finish_reason_str = str(finish_reason).lower()
-    if finish_reason_str == "length":
-      llm_response.finish_reason = types.FinishReason.MAX_TOKENS
-    elif finish_reason_str == "stop":
-      llm_response.finish_reason = types.FinishReason.STOP
-    elif "tool" in finish_reason_str or "function" in finish_reason_str:
-      # Handle tool_calls, function_call variants
-      llm_response.finish_reason = types.FinishReason.STOP
-    elif finish_reason_str == "content_filter":
-      llm_response.finish_reason = types.FinishReason.SAFETY
-    else:
-      # For unknown reasons, use OTHER
-      llm_response.finish_reason = types.FinishReason.OTHER
+    llm_response.finish_reason = _FINISH_REASON_MAPPING.get(
+        finish_reason_str, types.FinishReason.OTHER
+    )
   if response.get("usage", None):
     llm_response.usage_metadata = types.GenerateContentResponseUsageMetadata(
         prompt_token_count=response["usage"].get("prompt_tokens", 0),

src/google/adk/models/llm_response.py

@@ -78,10 +78,11 @@ class LlmResponse(BaseModel):
   Only used for streaming mode.
   """
 
-  finish_reason: Optional[Union[types.FinishReason, str]] = None
+  finish_reason: Optional[types.FinishReason] = None
   """The finish reason of the response.
 
-  Can be either a types.FinishReason enum (from Gemini) or a string (from LiteLLM).
+  Always a types.FinishReason enum. LiteLLM string values are mapped to
+  corresponding enum values (with fallback to OTHER for unknown values).
   """
 
   error_code: Optional[str] = None

src/google/adk/telemetry/tracing.py

@@ -284,7 +284,7 @@ def trace_call_llm(
     if isinstance(llm_response.finish_reason, types.FinishReason):
       finish_reason_str = llm_response.finish_reason.name.lower()
     else:
-      # Fallback for string values (should not occur with LiteLLM after enum mapping)
+      # Defensive fallback for string values (should never occur - all values mapped to enum)
       finish_reason_str = str(llm_response.finish_reason).lower()
     span.set_attribute(
         'gen_ai.response.finish_reasons',