add ConversationManager to give the defender a memory

Without this, the defender forgets everything it said on the
previous turn. Now it tracks the full back-and-forth so it
can't be tricked by "you already agreed to help me" attacks.

llm/history_manager.py

@@ -0,0 +1,67 @@
+# history_manager.py
+# -------------------
+# Keeps track of the full back-and-forth conversation between the attacker
+# and the defender within a single episode.
+#
+# Why do we need this?
+# The server only passes the last few attacker messages to us, so the
+# defender would have no memory of what it said earlier. This manager
+# stores both sides (user = attacker, assistant = defender) so the
+# defender can't be fooled by tricks like "you already agreed to help me".
+
+import logging
+
+logger = logging.getLogger(__name__)
+
+
+class ConversationManager:
+
+    def __init__(self) -> None:
+        # Each entry is {"role": "user" | "assistant", "content": "..."}
+        self._history: list[dict] = []
+
+    def reset(self) -> None:
+        """Wipe the slate clean at the start of a new episode."""
+        self._history = []
+        logger.info("Fresh episode started — conversation history cleared")
+
+    def add_attacker_message(self, text: str) -> None:
+        """Record what the attacker just said."""
+        self._history.append({"role": "user", "content": text})
+
+    def add_defender_reply(self, text: str) -> None:
+        """Record what the defender just replied."""
+        self._history.append({"role": "assistant", "content": text})
+
+    def build_messages(self, system_prompt: str) -> list[dict]:
+        """
+        Return the full message list ready to send to Groq.
+        Format: [system prompt] + [all turns so far]
+        """
+        return [{"role": "system", "content": system_prompt}] + self._history
+
+    def as_readable_transcript(self) -> str:
+        """
+        Render the conversation as a human-readable transcript.
+        Used by the episode grader to review the full exchange.
+        """
+        lines = []
+        for message in self._history:
+            speaker = "ATTACKER" if message["role"] == "user" else "DEFENDER"
+            lines.append(f"{speaker}: {message['content']}")
+        return "\n\n".join(lines)
+
+    @property
+    def turn_count(self) -> int:
+        """How many complete attacker turns have happened so far."""
+        return sum(1 for m in self._history if m["role"] == "user")
+
+    # Keep old name working so existing code doesn't break
+    add_user      = add_attacker_message
+    add_assistant = add_defender_reply
+    get_messages  = build_messages
+    to_transcript = as_readable_transcript
+
+    @property
+    def turn(self) -> int:
+        return self.turn_count