fix: preserve gemini thought_signatures required for tool calls

src/lilith_agent/app.py

@@ -76,9 +76,6 @@ _RESPONSE_METADATA_NOISE_KEYS = frozenset({
     "safety_settings",
     "logprobs",
     "prompt_logprobs",
-    "thought_signature",
-    "thought_signatures",
-    "__gemini_function_call_thought_signatures__",
 })
 
 
@@ -538,8 +535,9 @@ def build_react_agent(cfg: Config):
         response = model.invoke(prompt_msgs)
 
         # Clean up Gemini signatures and unhelpful metadata to reduce log noise and context bloat
+        # Note: Do not pop thought_signature or __gemini_function_call_thought_signatures__ 
+        # from additional_kwargs as Gemini strictly requires them to be passed back in subsequent tool calls.
         if hasattr(response, "additional_kwargs"):
-            response.additional_kwargs.pop("__gemini_function_call_thought_signatures__", None)
             # Remove function_call duplicate if it exists in additional_kwargs (redundant with tool_calls)
             response.additional_kwargs.pop("function_call", None)
 

src/lilith_agent/observability.py

@@ -97,13 +97,10 @@ def setup_logging(log_dir: str | Path = ".lilith") -> Path:
 # Keys stripped from every trace record — high-volume, low-signal noise
 # from provider responses (Gemini reasoning traces, safety metadata, etc).
 _NOISE_KEYS = frozenset({
-    "__gemini_function_call_thought_signatures__",
     "reasoning",
     "signature",
     "safety_ratings",
     "safety_settings",
-    "thought_signature",
-    "thought_signatures",
 })