Conversation Brain

@dataclass
class ConversationMemory:
    call_id: str
    turns: list[ConversationTurn]           # All conversation turns
    context: dict                            # Persistent context
    extracted_info: dict                     # Extracted entities

# conversation_brain.py:213
def add_user_transcript(self, text: str):
    """Add user transcript fragment."""
    if text:
        self._user_buffer.append(text)

# Skip greetings and simple phrases
non_actionable_phrases = [
    "hi", "hello", "hey", "thanks", "okay", "yes", "no"
]

# Quick detection of action keywords
action_keywords = [
    "open", "play", "search", "send", "check", 
    "email", "message", "youtube", "spotify"
]

prompt = f"""You are a simple intent classifier. 
Determine if the user wants you to DO something or just chatting.

Recent conversation:
{context}

User said: "{user_text}"

Is this a request to DO something?
Answer with just ONE word: YES or NO"""

cmd = [
    "clawdbot", "agent",
    "--session-id", "agent:main:main",
    "--message", processed_command,
    "--timeout", "90",
]

process = await asyncio.create_subprocess_exec(
    *cmd,
    stdout=asyncio.subprocess.PIPE,
    stderr=asyncio.subprocess.PIPE,
    env={...},
)

if response and self._on_clawdbot_response:
    # Feed response back to AI to speak it
    await self._on_clawdbot_response(response)

# From audio_bridge.py:314
prompt = f"""The system has retrieved the following information.
Please relay this information naturally and concisely.

Information to relay:
{response}"""

self._brain.set_callbacks(
    on_command=handle_command,              # Optional
    on_clawdbot_response=speak_response,    # Required
)

# 1. Buffer fragments
add_user_transcript("Check")  # Fragment 1
add_user_transcript(" my")    # Fragment 2
add_user_transcript(" emails") # Fragment 3

# 2. Flush complete turn
await flush_user_turn()
  ↓
# 3. Analyze intent
intent, command, is_actionable = await _analyze_intent(full_text)
  ↓
# 4. Route if actionable
if is_actionable:
    response = await _send_to_clawdbot_async(full_text)
    await self._on_clawdbot_response(response)

# Buffer AI's speech fragments
add_assistant_transcript("I'll")
add_assistant_transcript(" check")
add_assistant_transcript(" your emails now.")

# Flush when AI finishes speaking
await flush_assistant_turn()

ConversationBrain(
    api_key=gemini_api_key,                    # For intent analysis
    model="gemini-3-flash-preview",            # Fast classification
    telegram_chat_id=chat_id,                  # For ClawdBot agent
    call_id=call_id,                           # Session tracking
)

# Instant classification (no LLM call)
if text_lower in non_actionable_phrases:
    return "conversation", None, False

if any(kw in text_lower for kw in action_keywords):
    return "action", {...}, True

# conversation_brain.py:385
except Exception as e:
    logger.error("Error analyzing intent", error=str(e))
    # Default to actionable to avoid missing commands
    return "action", {"original_request": user_text}, True

except asyncio.TimeoutError:
    return "I'm still working on that. It's taking longer than expected."
except Exception as e:
    return f"Sorry, I encountered an error: {str(e)}"

print(f"=== BRAIN BUFFER: {len(self._user_buffer)} fragments ===")
print(f"=== BRAIN FLUSH: {num_fragments} fragments → \"{full_text}\" ===")
print(f"=== BRAIN: Actionable={is_actionable} (LLM said: {answer}) ===")
print(f"=== BRAIN: Sending to ClawdBot: \"{full_text}\" ===")

agent_system_instruction = """You are Alchemy, an AI agent with REAL capabilities:
- You CAN check and send emails
- You CAN send messages to Telegram, WhatsApp, Discord
- You CAN search the web

When asked to do something, SAY "I'll do that now" and describe what you're doing."""

# ✅ Good - clear binary choice
"Is this a request to DO something? Answer YES or NO"

# ❌ Bad - allows uncertainty
"What is the user's intent? Explain your reasoning."

def get_recent_context(self, max_turns: int = 10) -> str:
    recent = self.turns[-max_turns:]
    # ...

User: "Check my emails"
  ↓
Brain: Quick keyword match "check" + "emails"
  ↓
Classification: actionable (heuristic)
  ↓
ClawdBot: Execute GOG skill
  ↓
Response: "You have 3 unread emails: 1 from John..."
  ↓
AI speaks response to user

User: "How are you?"
  ↓
Brain: Quick match on common phrase
  ↓
Classification: conversation (heuristic)
  ↓
No ClawdBot call
  ↓
AI responds naturally from conversation

User: "What's the weather like?"
  ↓
Brain: No heuristic match
  ↓
LLM Analysis: "DO something" (get weather) → YES
  ↓
Classification: actionable
  ↓
ClawdBot: Execute search skill
  ↓
Response: "It's 72°F and sunny in San Francisco"