fix: remove duplicate assistant message when content and tool_calls both present

In CrewOrchestrator.run(), when the LLM returns both content and tool_calls,
the code was appending two consecutive assistant messages to self._messages:
1. {"role": "assistant", "content": response.content}
2. {"role": "assistant", "content": ..., "tool_calls": [...]}

Back-to-back model-role messages cause Gemini (and technically any
OpenAI-compatible API) to reject the conversation on the next turn.
The fix is to drop the redundant bare-content append — the content is
already included in the combined message that follows.

Also update get_available_models() to list Gemini models with the
required gemini/ provider prefix so the helper reflects correct usage.

Fixes #8

pentestagent/agents/crew/orchestrator.py

@@ -184,9 +184,6 @@ class CrewOrchestrator:
                     # If there are tool calls, the content is "thinking" (reasoning before action)
                     if response.content:
                         yield {"phase": "thinking", "content": response.content}
-                        self._messages.append(
-                            {"role": "assistant", "content": response.content}
-                        )
 
                     def get_tc_name(tc):
                         if hasattr(tc, "function"):

pentestagent/llm/llm.py

@@ -363,8 +363,8 @@ class LLM:
             "claude-sonnet-4-20250514",
             "claude-opus-4-20250514",
             # Google
-            "gemini-2.5-pro",
-            "gemini-2.5-flash",
+            "gemini/gemini-2.5-pro",
+            "gemini/gemini-2.5-flash",
             # Others via LiteLLM
             "ollama/llama3",
             "ollama/mixtral",