feat: two-phase execution, think slugs, query enrichment (v1.6.0)

Two-Phase Tatlock Execution:
- orchestrate_tool_calls() for Phase 1 coordination
- synthesize_from_results() for Phase 2 butler-toned synthesis
- Guarantees butler personality in all responses

Automatic Think Slugs:
- Deterministic butler-perspective messages during expert delegation
- ActionType enum: RETRIEVE, RESEARCH, CREATE, CONTROL, RECORD
- HOUSEHOLD_THINK_MESSAGES mapping for all experts
- Streaming delegation wrappers with automatic think messages

Steward Query Enrichment:
- Auto-fill user context (location, timezone) when not specified
- _build_enriched_query() with regex word boundary matching
- enriched_query field in StewardRecommendation schema

Documentation:
- ORCHESTRATION_SCENARIOS.md rewritten with Mermaid diagrams
- New Housekeeper and Biographer scenarios
- TESTING_IMPROVEMENTS.md for future LLM testing patterns

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

src/agents/delegation.py

@@ -9,13 +9,136 @@ This implements the agent-as-tool pattern recommended by PydanticAI:
 agents call other agents via tool wrappers, keeping each agent focused.
 """
 from dataclasses import dataclass, field
-from typing import Callable, Optional, Any
+from enum import Enum
+from typing import AsyncGenerator, Callable, Optional, Any
 
 from src.core.logging_config import get_logger
 
 logger = get_logger(__name__)
 
 
+# =============================================================================
+# Action Types for Think Slug Selection
+# =============================================================================
+
+class ActionType(Enum):
+    """
+    Categories of actions for selecting appropriate think messages.
+
+    Each expert has different action types that warrant different
+    butler-perspective messages to the user.
+    """
+    RETRIEVE = "retrieve"     # Looking up existing information
+    RESEARCH = "research"     # Conducting new research (web search, etc.)
+    CREATE = "create"         # Creating new content (pages, notes)
+    CONTROL = "control"       # Controlling devices/automations
+    RECORD = "record"         # Recording memories/notes
+
+
+# =============================================================================
+# Household Think Messages (Butler's Perspective)
+# =============================================================================
+
+HOUSEHOLD_THINK_MESSAGES: dict[str, dict[ActionType, dict[str, str]]] = {
+    "librarian": {
+        ActionType.RETRIEVE: {
+            "start": "<think>Allow me to consult the archives, sir.</think>",
+            "success": "<think>The Librarian has compiled the relevant findings.</think>",
+            "error": "<think>I'm afraid the archives proved difficult to access.</think>",
+        },
+        ActionType.RESEARCH: {
+            "start": "<think>I've dispatched the Librarian to conduct some fresh research.</think>",
+            "success": "<think>The Librarian has returned with findings, sir.</think>",
+            "error": "<think>The research proved inconclusive, I'm afraid.</think>",
+        },
+        ActionType.CREATE: {
+            "start": "<think>I'm having the Librarian prepare a new entry.</think>",
+            "success": "<think>The new material has been properly catalogued, sir.</think>",
+            "error": "<think>I'm afraid there was difficulty filing the entry.</think>",
+        },
+    },
+    "biographer": {
+        ActionType.RETRIEVE: {
+            "start": "<think>Let me consult the household records.</think>",
+            "success": "<think>The Biographer has located the relevant information, sir.</think>",
+            "error": "<think>I'm unable to locate those particular records.</think>",
+        },
+        ActionType.RECORD: {
+            "start": "<think>I've asked the Biographer to take note of this, sir.</think>",
+            "success": "<think>The household records have been updated accordingly.</think>",
+            "error": "<think>I'm afraid there was difficulty recording the entry.</think>",
+        },
+    },
+    "housekeeper": {
+        ActionType.RETRIEVE: {
+            "start": "<think>Allow me to inquire with the household staff.</think>",
+            "success": "<think>The staff reports the current status, sir.</think>",
+            "error": "<think>The household staff is momentarily unavailable, I'm afraid.</think>",
+        },
+        ActionType.CONTROL: {
+            "start": "<think>I'm instructing the household staff now, sir.</think>",
+            "success": "<think>The household has been configured as requested.</think>",
+            "error": "<think>I'm afraid the staff reports an issue with that request.</think>",
+        },
+    },
+}
+
+
+def _detect_action_type(expert: str, task: str) -> ActionType:
+    """
+    Detect action type from expert name and task description.
+
+    Used to select appropriate butler-perspective think messages.
+
+    Args:
+        expert: Name of the expert (librarian, biographer, housekeeper)
+        task: Task description
+
+    Returns:
+        ActionType: Detected action type for message selection
+    """
+    task_lower = task.lower()
+
+    if expert == "librarian":
+        if any(w in task_lower for w in ["search", "find", "look up", "research"]):
+            if any(w in task_lower for w in ["web", "online", "internet"]):
+                return ActionType.RESEARCH
+            return ActionType.RETRIEVE
+        if any(w in task_lower for w in ["create", "write", "add", "make", "new"]):
+            return ActionType.CREATE
+        return ActionType.RETRIEVE
+
+    elif expert == "biographer":
+        if any(w in task_lower for w in ["remember", "note", "record", "save", "store"]):
+            return ActionType.RECORD
+        return ActionType.RETRIEVE
+
+    elif expert == "housekeeper":
+        if any(w in task_lower for w in ["turn", "set", "activate", "enable", "disable", "toggle"]):
+            return ActionType.CONTROL
+        return ActionType.RETRIEVE
+
+    return ActionType.RETRIEVE
+
+
+def get_think_message(expert: str, task: str, phase: str) -> str:
+    """
+    Get the appropriate think message for an expert delegation.
+
+    Args:
+        expert: Name of the expert
+        task: Task description (used to detect action type)
+        phase: One of "start", "success", "error"
+
+    Returns:
+        str: Butler-perspective think message
+    """
+    action_type = _detect_action_type(expert, task)
+    expert_messages = HOUSEHOLD_THINK_MESSAGES.get(expert, {})
+    action_messages = expert_messages.get(action_type, expert_messages.get(ActionType.RETRIEVE, {}))
+    return action_messages.get(phase, f"<think>Consulting {expert}...</think>")
+
+
 @dataclass
 class DelegationTask:
     """
@@ -301,6 +424,103 @@ async def delegate_to_housekeeper(
         )
 
 
+# =============================================================================
+# Streaming Delegation Wrappers (with Think Messages)
+# =============================================================================
+
+async def stream_delegate_to_librarian(
+    task: str,
+    context: str = "",
+) -> AsyncGenerator[str, None]:
+    """
+    Stream delegation to Librarian with automatic think messages.
+
+    Yields butler-perspective think messages before and after the delegation,
+    allowing the UI to show progress to the user.
+
+    Args:
+        task: Task description
+        context: Additional context
+
+    Yields:
+        str: Think messages and final result marker
+    """
+    # Yield start message (deterministic)
+    yield get_think_message("librarian", task, "start") + "\n"
+
+    # Execute delegation
+    result = await delegate_to_librarian(task, context)
+
+    # Yield completion message (deterministic)
+    if result.success:
+        yield get_think_message("librarian", task, "success") + "\n"
+    else:
+        yield get_think_message("librarian", task, "error") + "\n"
+
+    # Yield result marker for extraction
+    yield f"__DELEGATION_RESULT__:librarian:{result.output}"
+
+
+async def stream_delegate_to_biographer(
+    task: str,
+    context: str = "",
+) -> AsyncGenerator[str, None]:
+    """
+    Stream delegation to Biographer with automatic think messages.
+
+    Args:
+        task: Task description
+        context: Additional context
+
+    Yields:
+        str: Think messages and final result marker
+    """
+    yield get_think_message("biographer", task, "start") + "\n"
+
+    result = await delegate_to_biographer(task, context)
+
+    if result.success:
+        yield get_think_message("biographer", task, "success") + "\n"
+    else:
+        yield get_think_message("biographer", task, "error") + "\n"
+
+    yield f"__DELEGATION_RESULT__:biographer:{result.output}"
+
+
+async def stream_delegate_to_housekeeper(
+    task: str,
+    context: str = "",
+) -> AsyncGenerator[str, None]:
+    """
+    Stream delegation to Housekeeper with automatic think messages.
+
+    Args:
+        task: Task description
+        context: Additional context
+
+    Yields:
+        str: Think messages and final result marker
+    """
+    yield get_think_message("housekeeper", task, "start") + "\n"
+
+    result = await delegate_to_housekeeper(task, context)
+
+    if result.success:
+        yield get_think_message("housekeeper", task, "success") + "\n"
+    else:
+        yield get_think_message("housekeeper", task, "error") + "\n"
+
+    yield f"__DELEGATION_RESULT__:housekeeper:{result.output}"
+
+
+# Mapping of streaming delegation wrappers
+STREAMING_DELEGATION_WRAPPERS = {
+    "librarian": stream_delegate_to_librarian,
+    "biographer": stream_delegate_to_biographer,
+    "housekeeper": stream_delegate_to_housekeeper,
+}
+
+
 # Future expert delegation wrappers will be added here:
 # - delegate_to_developer(task, context) -> DelegationResult
 # - delegate_to_secretary(task, context) -> DelegationResult

src/agents/steward/schemas.py

@@ -60,6 +60,10 @@ class StewardRecommendation(BaseModel):
         default_factory=dict,
         description="Pre-fetched user context from memory (profile, preferences)"
     )
+    enriched_query: str = Field(
+        default="",
+        description="User query with auto-filled context (location, timezone) when not specified"
+    )
 
     def format_for_butler(self) -> str:
         """

src/agents/steward/service.py

@@ -149,6 +149,68 @@ def _extract_missing_capabilities(text: str) -> Optional[str]:
     return None
 
 
+def _build_enriched_query(user_request: str, memory_context: dict[str, Any]) -> str:
+    """
+    Build an enriched query by appending user context when not specified.
+
+    When the user asks location-dependent questions (weather, nearby, etc.)
+    without specifying a location, this appends their known location.
+    Similarly for timezone-dependent queries.
+
+    Args:
+        user_request: The user's original request
+        memory_context: Pre-fetched memory context with profile/preferences
+
+    Returns:
+        str: Query with context appended, or original query if no enrichment needed
+
+    Example:
+        >>> query = _build_enriched_query(
+        ...     "What's the weather?",
+        ...     {"profile": {"location": "Amsterdam", "timezone": "Europe/Amsterdam"}}
+        ... )
+        >>> query
+        "What's the weather?\n\n[User Context: location=Amsterdam, timezone=Europe/Amsterdam]"
+    """
+    if not memory_context:
+        return user_request
+
+    request_lower = user_request.lower()
+    profile = memory_context.get("profile", {})
+    preferences = memory_context.get("preferences", {})
+
+    context_parts = []
+
+    # Check if location is needed and not specified
+    location_keywords = ["weather", "temperature", "forecast", "nearby", "local", "here"]
+    # Use word boundary pattern to avoid false positives like "at" in "what"
+    location_prepositions = [r'\bin\b', r'\bat\b', r'\bnear\b', r'\baround\b', r'\bfor\b']
+    location_specified = any(re.search(p, request_lower) for p in location_prepositions)
+
+    if any(word in request_lower for word in location_keywords):
+        if not location_specified and profile.get("location"):
+            context_parts.append(f"location={profile['location']}")
+
+    # Check if timezone is needed and not specified
+    time_keywords = ["time", "schedule", "meeting", "appointment", "when", "today", "tomorrow"]
+    timezone_specified = any(word in request_lower for word in ["timezone", "tz", "utc", "gmt"])
+
+    if any(word in request_lower for word in time_keywords):
+        if not timezone_specified and profile.get("timezone"):
+            context_parts.append(f"timezone={profile['timezone']}")
+
+    # Add preferences if relevant
+    if preferences.get("temperature_unit") and "weather" in request_lower:
+        context_parts.append(f"temperature_unit={preferences['temperature_unit']}")
+
+    # Build enriched query
+    if context_parts:
+        context_str = ", ".join(context_parts)
+        return f"{user_request}\n\n[User Context: {context_str}]"
+
+    return user_request
+
+
 async def _prefetch_memory_context(user_request: str) -> dict[str, Any]:
     """
     Pre-fetch user context that might be needed for this request.
@@ -277,6 +339,9 @@ async def analyze_request(
             context = _extract_conversation_context(analysis_text, conversation_history)
             missing = _extract_missing_capabilities(analysis_text)
 
+            # Build enriched query with auto-filled context
+            enriched_query = _build_enriched_query(user_request, memory_context)
+
             recommendation = StewardRecommendation(
                 recommended_capabilities=capabilities,
                 reasoning=analysis_text,
@@ -284,6 +349,7 @@ async def analyze_request(
                 conversation_context=context,
                 missing_capabilities=missing,
                 memory_context=memory_context,
+                enriched_query=enriched_query,
             )
 
             # Update log context with results

src/agents/tatlock.py

@@ -630,6 +630,241 @@ class TatlockAgent(AgentInterface):
 
         logger.info("tatlock_scoped_run_complete")
 
+    async def orchestrate_tool_calls(
+        self,
+        user_message: str,
+        steward_note: str,
+        scoped_tools: list[Any],
+        message_history: list[dict],
+        tool_tracker: Any = None,
+    ) -> dict[str, Any]:
+        """
+        Phase 1: Execute tool calls and delegations, return structured results.
+
+        This is the coordination phase where Tatlock orchestrates tool calls
+        and expert delegations. The raw output is captured for Phase 2 synthesis.
+
+        Args:
+            user_message: The user's original message
+            steward_note: Note from Steward (invisible to user)
+            scoped_tools: List of tool definitions from household registry
+            message_history: Conversation history
+            tool_tracker: Optional tool call tracker for benchmarking
+
+        Returns:
+            dict with:
+                - tools_called: List of tool names that were called
+                - expert_results: Dict mapping expert names to their outputs
+                - tool_outputs: Dict mapping tool names to their outputs
+                - raw_output: The agent's raw text output
+        """
+        from pydantic_ai.models.openai import OpenAIChatModel
+        from pydantic_ai.providers.ollama import OllamaProvider
+        from pydantic_ai.settings import ModelSettings
+        from pydantic_ai.messages import (
+            ModelRequest,
+            ModelResponse,
+            UserPromptPart,
+            TextPart,
+            ToolCallPart,
+            ToolReturnPart,
+        )
+
+        logger.info(
+            "tatlock_orchestrate_tool_calls",
+            user_message_preview=user_message[:100],
+            scoped_tool_count=len(scoped_tools),
+            history_length=len(message_history),
+        )
+
+        # Create a fresh agent instance with scoped tools only
+        clean_host = self.ollama_host.rstrip('/')
+        base_url = f"{clean_host}/v1"
+
+        ollama_model = OpenAIChatModel(
+            model_name=self.model_name,
+            provider=OllamaProvider(base_url=base_url)
+        )
+
+        # Create agent with scoped tools
+        scoped_agent = Agent(
+            ollama_model,
+            system_prompt=TATLOCK_SYSTEM_PROMPT,
+            tools=scoped_tools,
+        )
+
+        # Prepend Steward's note to the request
+        enriched_message = f"{steward_note}\n\n{user_message}"
+
+        # Convert message history to PydanticAI format
+        pydantic_history = []
+        for msg in message_history:
+            role = msg.get("role")
+            content = msg.get("content", "")
+
+            if not content or not content.strip():
+                continue
+
+            if role == "user":
+                pydantic_history.append(
+                    ModelRequest(parts=[UserPromptPart(content=content)])
+                )
+            elif role == "assistant":
+                pydantic_history.append(
+                    ModelResponse(parts=[TextPart(content=content)])
+                )
+
+        # Run with scoped tools and tracker
+        result = await scoped_agent.run(
+            enriched_message,
+            message_history=pydantic_history if pydantic_history else None,
+            deps=tool_tracker,
+            model_settings=ModelSettings(extra_body={"tool_choice": "required"})
+        )
+
+        # Extract tool calls and results from the agent's messages
+        tools_called = []
+        expert_results = {}
+        tool_outputs = {}
+
+        # Parse through new messages to find tool calls and returns
+        for msg in result.new_messages():
+            if isinstance(msg, ModelResponse):
+                for part in msg.parts:
+                    if isinstance(part, ToolCallPart):
+                        tools_called.append(part.tool_name)
+            elif isinstance(msg, ModelRequest):
+                for part in msg.parts:
+                    if isinstance(part, ToolReturnPart):
+                        tool_name = part.tool_name
+                        content = part.content
+
+                        # Categorize as expert result or tool output
+                        if tool_name.startswith("delegate_to_"):
+                            expert_name = tool_name.replace("delegate_to_", "")
+                            expert_results[expert_name] = content
+                        else:
+                            tool_outputs[tool_name] = content
+
+        logger.info(
+            "tatlock_orchestration_complete",
+            tools_called=tools_called,
+            expert_count=len(expert_results),
+            tool_output_count=len(tool_outputs),
+        )
+
+        return {
+            "tools_called": tools_called,
+            "expert_results": expert_results,
+            "tool_outputs": tool_outputs,
+            "raw_output": result.output,
+        }
+
+    async def synthesize_from_results(
+        self,
+        user_message: str,
+        orchestration_results: dict[str, Any],
+        message_history: list[dict],
+    ) -> str:
+        """
+        Phase 2: Synthesize butler-toned response from gathered results.
+
+        This is the synthesis phase where Tatlock takes the coordination
+        results and produces a properly butler-toned response.
+
+        Args:
+            user_message: The user's original message
+            orchestration_results: Results from orchestrate_tool_calls()
+            message_history: Conversation history
+
+        Returns:
+            str: Butler-toned response synthesized from all results
+        """
+        from pydantic_ai.models.openai import OpenAIChatModel
+        from pydantic_ai.providers.ollama import OllamaProvider
+        from pydantic_ai.messages import ModelRequest, ModelResponse, UserPromptPart, TextPart
+
+        logger.info(
+            "tatlock_synthesize_from_results",
+            user_message_preview=user_message[:100],
+            expert_count=len(orchestration_results.get("expert_results", {})),
+            tool_count=len(orchestration_results.get("tool_outputs", {})),
+        )
+
+        # Build synthesis prompt with all available information
+        synthesis_parts = []
+        synthesis_parts.append(f"The user asked: {user_message}")
+        synthesis_parts.append("")
+
+        # Add expert findings if any
+        if orchestration_results.get("expert_results"):
+            synthesis_parts.append("Expert findings:")
+            for expert, result in orchestration_results["expert_results"].items():
+                synthesis_parts.append(f"- {expert.title()}: {result}")
+            synthesis_parts.append("")
+
+        # Add tool outputs if any
+        if orchestration_results.get("tool_outputs"):
+            synthesis_parts.append("Tool results:")
+            for tool, result in orchestration_results["tool_outputs"].items():
+                synthesis_parts.append(f"- {tool}: {result}")
+            synthesis_parts.append("")
+
+        synthesis_parts.append(
+            "Based on this information, provide a response to the user. "
+            "Maintain your butler personality - address them as 'sir', "
+            "use formal but personable language, and be helpful."
+        )
+
+        synthesis_prompt = "\n".join(synthesis_parts)
+
+        # Create synthesis agent (no tools needed)
+        clean_host = self.ollama_host.rstrip('/')
+        base_url = f"{clean_host}/v1"
+
+        ollama_model = OpenAIChatModel(
+            model_name=self.model_name,
+            provider=OllamaProvider(base_url=base_url)
+        )
+
+        # Synthesis agent uses butler prompt but no tools
+        synthesis_agent = Agent(
+            ollama_model,
+            system_prompt=TATLOCK_SYSTEM_PROMPT,
+            # No tools for synthesis phase
+        )
+
+        # Convert message history to PydanticAI format
+        pydantic_history = []
+        for msg in message_history:
+            role = msg.get("role")
+            content = msg.get("content", "")
+
+            if not content or not content.strip():
+                continue
+
+            if role == "user":
+                pydantic_history.append(
+                    ModelRequest(parts=[UserPromptPart(content=content)])
+                )
+            elif role == "assistant":
+                pydantic_history.append(
+                    ModelResponse(parts=[TextPart(content=content)])
+                )
+
+        # Run synthesis
+        result = await synthesis_agent.run(
+            synthesis_prompt,
+            message_history=pydantic_history if pydantic_history else None,
+        )
+
+        logger.info(
+            "tatlock_synthesis_complete",
+            response_preview=result.output[:100],
+        )
+
+        return result.output
+
     async def get_capabilities(self) -> dict:
         """Return current capabilities."""
         return {

src/core/household_registry.py

@@ -273,6 +273,65 @@ class HouseholdRegistry:
 
         return tools
 
+    def get_streaming_delegation_tools(self, names: list[str]) -> list[Any]:
+        """
+        Get streaming delegation wrapper tools for specified capabilities.
+
+        Similar to get_delegation_tools() but returns streaming wrappers
+        that yield butler-perspective think messages during execution.
+
+        These wrappers emit think slugs like:
+        - "Allow me to consult the archives, sir."
+        - "The Librarian has compiled the relevant findings."
+
+        Args:
+            names: List of member names to include
+
+        Returns:
+            List of streaming delegation wrappers and/or raw tools
+
+        Example:
+            >>> tools = registry.get_streaming_delegation_tools(["librarian"])
+            >>> async for chunk in tools[0](task="Search for Docker"):
+            ...     print(chunk)  # Yields think messages then result
+        """
+        from src.agents.delegation import STREAMING_DELEGATION_WRAPPERS
+
+        tools = []
+        for name in names:
+            member = self._members.get(name)
+            if not member:
+                logger.warning(
+                    "household_member_not_found",
+                    requested_name=name,
+                    available_names=list(self._members.keys()),
+                )
+                continue
+
+            # Check if this member has a streaming delegation wrapper
+            if name in STREAMING_DELEGATION_WRAPPERS and member.agent is not None:
+                tools.append(STREAMING_DELEGATION_WRAPPERS[name])
+                logger.debug(
+                    "streaming_delegation_wrapper_added",
+                    member=name,
+                )
+            else:
+                # No agent = direct tools (e.g., tatlock_core)
+                tools.extend(member.tools)
+                logger.debug(
+                    "raw_tools_added",
+                    member=name,
+                    tool_count=len(member.tools),
+                )
+
+        logger.info(
+            "streaming_delegation_tools_created",
+            requested_members=names,
+            total_tools=len(tools),
+        )
+
+        return tools
+
     def list_members(self) -> list[str]:
         """
         List all registered member names.

src/responses/service.py

@@ -43,7 +43,7 @@ async def _execute_single_delegation(
     Execute a single delegation to an agent.
 
     Args:
-        agent_name: Name of agent (biographer, librarian)
+        agent_name: Name of agent (biographer, librarian, housekeeper)
         task: Task description
         tracker: Tool call tracker
 
@@ -67,6 +67,13 @@ async def _execute_single_delegation(
         await tracker.track_call("delegate_to_librarian", duration)
         return (agent_name, result.output)
 
+    elif agent_name == "housekeeper":
+        from src.agents.delegation import delegate_to_housekeeper
+        result = await delegate_to_housekeeper(task=task)
+        duration = time.time() - start_time
+        await tracker.track_call("delegate_to_housekeeper", duration)
+        return (agent_name, result.output)
+
     else:
         return (agent_name, f"Unknown agent: {agent_name}")
 
@@ -244,6 +251,68 @@ async def _direct_delegation(
     return "\n\n".join(results) if results else "I apologize, sir. No delegation results available."
 
 
+async def _direct_delegation_with_results(
+    user_message: str,
+    recommendation: "StewardRecommendation",
+    tracker: "ToolCallTracker",
+    conversation_id: str,
+) -> dict:
+    """
+    Directly delegate to expert agents and return structured results.
+
+    This is the Phase 1 variant of direct delegation that returns results
+    in the same format as TatlockAgent.orchestrate_tool_calls() for
+    consistent Phase 2 synthesis.
+
+    Args:
+        user_message: User's request
+        recommendation: Steward's recommendation
+        tracker: Tool call tracker
+        conversation_id: Conversation ID
+
+    Returns:
+        dict: Orchestration results with expert_results, tool_outputs, etc.
+    """
+    logger.info(
+        "direct_delegation_with_results",
+        agents=recommendation.recommended_capabilities,
+        conversation_id=conversation_id,
+    )
+
+    expert_results = {}
+    tools_called = []
+
+    for agent in recommendation.recommended_capabilities:
+        try:
+            agent_name, result = await _execute_single_delegation(
+                agent, user_message, tracker
+            )
+            expert_results[agent_name] = result
+            tools_called.append(f"delegate_to_{agent_name}")
+
+            logger.info(
+                "direct_delegation_result",
+                agent=agent_name,
+                result_preview=result[:100] if result else "empty",
+                conversation_id=conversation_id,
+            )
+        except Exception as e:
+            logger.error(
+                "direct_delegation_failed",
+                agent=agent,
+                error=str(e),
+                conversation_id=conversation_id,
+            )
+            expert_results[agent] = f"Error: {e}"
+
+    return {
+        "tools_called": tools_called,
+        "expert_results": expert_results,
+        "tool_outputs": {},  # No tool outputs for direct delegation
+        "raw_output": "",  # No raw output for direct delegation
+    }
+
+
 # Global conversation history tracker
 # In production, this would be backed by a database or Redis
 _conversation_history = ConversationHistory(max_turns=20)
@@ -379,12 +448,13 @@ async def create_response(request: ResponseRequest) -> Response:
 
 async def create_response_with_steward(request: ResponseRequest) -> Response:
     """
-    Create response using Steward preprocessing (Phase 2 flow).
+    Create response using Steward preprocessing and two-phase Tatlock execution.
 
-    This is the two-tier architecture where:
+    This is the two-tier architecture with two-phase synthesis:
     1. Steward analyzes the request and recommends capabilities
-    2. Tatlock runs with scoped tools based on recommendations
-    3. Tool usage is tracked for benchmarking
+    2. Phase 1: Tatlock orchestrates tool calls and expert delegations
+    3. Phase 2: Tatlock synthesizes butler-toned response from results
+    4. Tool usage is tracked for benchmarking
 
     Args:
         request: Response request
@@ -420,37 +490,39 @@ async def create_response_with_steward(request: ResponseRequest) -> Response:
         conversation_id=conversation_id,
     )
 
-    # Phase 1: Steward preprocessing
+    # Steward preprocessing
     enriched = await preprocess_request(
         user_message,
         conversation_history=conversation_history,
         conversation_id=conversation_id,
     )
 
-    # Phase 2: Initialize tool tracker
+    # Initialize tool tracker
     tracker = ToolCallTracker(
         recommended_capabilities=enriched.recommendation.recommended_capabilities,
         conversation_id=conversation_id,
     )
 
-    # Phase 3: Check if direct delegation is recommended
-    # If Steward recommends ONLY delegation agents (biographer/librarian),
-    # skip Tatlock and delegate directly
+    # Check if direct delegation is recommended
+    # If Steward recommends ONLY delegation agents (biographer/librarian/housekeeper),
+    # we still use two-phase but delegate directly in Phase 1
+    delegation_agents = {"biographer", "librarian", "housekeeper"}
     delegation_only = all(
-        cap in ("biographer", "librarian")
+        cap in delegation_agents
         for cap in enriched.recommendation.recommended_capabilities
     ) and enriched.recommendation.recommended_capabilities
 
+    from src.agents.tatlock import TatlockAgent
+    tatlock = TatlockAgent()
+
     if delegation_only:
-        tatlock_response = await _direct_delegation(
+        # Direct delegation path - collect results then synthesize
+        orchestration_results = await _direct_delegation_with_results(
             user_message, enriched.recommendation, tracker, conversation_id
         )
     else:
-        # Phase 3a: Run Tatlock with scoped tools
-        from src.agents.tatlock import TatlockAgent
-        tatlock = TatlockAgent()
-
-        tatlock_response = await tatlock.run_with_scoped_tools(
+        # Phase 1: Orchestrate tool calls
+        orchestration_results = await tatlock.orchestrate_tool_calls(
             user_message=user_message,
             steward_note=enriched.steward_note,
             scoped_tools=enriched.scoped_tools,
@@ -458,14 +530,23 @@ async def create_response_with_steward(request: ResponseRequest) -> Response:
             tool_tracker=tracker,
         )
 
-        # Phase 3b: Check for text-based delegation fallback
-        # If Tatlock outputs [DELEGATE:...] instead of calling the function,
-        # we parse and execute it here
-        tatlock_response = await _handle_text_delegation(
-            tatlock_response, tracker, conversation_id
-        )
+        # Handle text-based delegation fallback if present
+        if "[DELEGATE:" in orchestration_results.get("raw_output", ""):
+            text_delegation_results = await _handle_text_delegation(
+                orchestration_results["raw_output"], tracker, conversation_id
+            )
+            # Add text delegation results to expert_results
+            if text_delegation_results != orchestration_results["raw_output"]:
+                orchestration_results["expert_results"]["text_delegation"] = text_delegation_results
 
-    # Phase 4: Finalize tool tracking
+    # Phase 2: Synthesize butler-toned response from all results
+    tatlock_response = await tatlock.synthesize_from_results(
+        user_message=user_message,
+        orchestration_results=orchestration_results,
+        message_history=conversation_history,
+    )
+
+    # Finalize tool tracking
     await tracker.finalize()
 
     # Build response output items

src/responses/streaming.py

@@ -118,11 +118,12 @@ class StreamingCoordinator:
         request: "ResponseRequest"  # type: ignore # Forward reference
     ) -> AsyncGenerator[StreamEvent, None]:
         """
-        Stream response with Steward preprocessing (Phase 2 flow).
+        Stream response with Steward preprocessing and two-phase Tatlock execution.
 
         Streams in order:
         1. Steward's analysis as reasoning summary
-        2. Tatlock's response as output text
+        2. Think slugs during expert delegation (butler-perspective messages)
+        3. Synthesized butler-toned response as output text
 
         Args:
             request: Response request
@@ -130,11 +131,17 @@ class StreamingCoordinator:
         Yields:
             StreamEvent: Stream of SSE events
         """
-        from src.responses.service import _calculate_usage, generate_id, _conversation_history
+        from src.responses.service import (
+            _calculate_usage,
+            generate_id,
+            _conversation_history,
+            _direct_delegation_with_results,
+        )
         from src.core.preprocessing import preprocess_request
         from src.core.tool_tracking import ToolCallTracker
         from src.responses.schemas import MessageOutputItem, ReasoningOutputItem, OutputTextContent
         from src.agents.tatlock import TatlockAgent
+        from src.agents.delegation import get_think_message, STREAMING_DELEGATION_WRAPPERS
         import asyncio
 
         output_items = []
@@ -152,7 +159,7 @@ class StreamingCoordinator:
 
             conversation_history = request.input[:-1] if len(request.input) > 1 else []
 
-            # Phase 1: Steward preprocessing
+            # Steward preprocessing
             enriched = await preprocess_request(
                 user_message,
                 conversation_history=conversation_history,
@@ -179,31 +186,60 @@ class StreamingCoordinator:
             )
             output_items.append(reasoning_item)
 
-            # Phase 2: Initialize tool tracker
+            # Initialize tool tracker
             tracker = ToolCallTracker(
                 recommended_capabilities=enriched.recommendation.recommended_capabilities,
                 conversation_id=conversation_id,
             )
 
-            # Phase 3: Stream Tatlock's response with scoped tools
-            tatlock = TatlockAgent()
-            tatlock_response_parts = []
+            # Check if direct delegation is recommended
+            delegation_agents = {"biographer", "librarian", "housekeeper"}
+            delegation_only = all(
+                cap in delegation_agents
+                for cap in enriched.recommendation.recommended_capabilities
+            ) and enriched.recommendation.recommended_capabilities
 
-            async for chunk in tatlock.run_with_scoped_tools_stream(
+            tatlock = TatlockAgent()
+
+            if delegation_only:
+                # Direct delegation path with streaming think slugs
+                orchestration_results = await self._stream_direct_delegation(
+                    user_message=user_message,
+                    recommendation=enriched.recommendation,
+                    tracker=tracker,
+                    conversation_id=conversation_id,
+                )
+
+                # Stream think slugs that were collected during delegation
+                for think_msg in orchestration_results.get("think_messages", []):
+                    yield ReasoningSummaryDelta(delta=think_msg)
+                    await asyncio.sleep(0.05)
+
+            else:
+                # Phase 1: Orchestrate tool calls
+                orchestration_results = await tatlock.orchestrate_tool_calls(
+                    user_message=user_message,
+                    steward_note=enriched.steward_note,
+                    scoped_tools=enriched.scoped_tools,
+                    message_history=conversation_history,
+                    tool_tracker=tracker,
+                )
+
+            # Phase 2: Synthesize butler-toned response
+            tatlock_response = await tatlock.synthesize_from_results(
                 user_message=user_message,
-                steward_note=enriched.steward_note,
-                scoped_tools=enriched.scoped_tools,
+                orchestration_results=orchestration_results,
                 message_history=conversation_history,
-                tool_tracker=tracker,
-            ):
-                tatlock_response_parts.append(chunk)
-                yield OutputTextDelta(delta=chunk)
+            )
+
+            # Stream the synthesized response
+            chunk_size = 50
+            for i in range(0, len(tatlock_response), chunk_size):
+                yield OutputTextDelta(delta=tatlock_response[i:i + chunk_size])
+                await asyncio.sleep(0.02)
 
             yield OutputTextDone()
 
-            # Combine response for output item
-            tatlock_response = "".join(tatlock_response_parts)
-
             # Add Tatlock message to output items
             message_item = MessageOutputItem(
                 id=f"msg_{generate_id()}",
@@ -217,7 +253,7 @@ class StreamingCoordinator:
             )
             output_items.append(message_item)
 
-            # Phase 4: Finalize tool tracking
+            # Finalize tool tracking
             await tracker.finalize()
 
             # Calculate usage and build final response
@@ -241,6 +277,85 @@ class StreamingCoordinator:
             # Stream error event
             yield self._create_error_event(e)
 
+    async def _stream_direct_delegation(
+        self,
+        user_message: str,
+        recommendation: "StewardRecommendation",  # type: ignore
+        tracker: "ToolCallTracker",  # type: ignore
+        conversation_id: str,
+    ) -> dict:
+        """
+        Execute direct delegation with streaming think messages.
+
+        Collects think messages as delegations execute for streaming to client.
+
+        Args:
+            user_message: User's request
+            recommendation: Steward's recommendation
+            tracker: Tool call tracker
+            conversation_id: Conversation ID
+
+        Returns:
+            dict: Orchestration results with think_messages list
+        """
+        from src.agents.delegation import (
+            get_think_message,
+            delegate_to_librarian,
+            delegate_to_biographer,
+            delegate_to_housekeeper,
+        )
+        import time as time_module
+
+        expert_results = {}
+        tools_called = []
+        think_messages = []
+
+        for agent in recommendation.recommended_capabilities:
+            # Emit start think message
+            start_msg = get_think_message(agent, user_message, "start")
+            think_messages.append(start_msg + "\n")
+
+            start_time = time_module.time()
+            try:
+                # Execute delegation
+                if agent == "librarian":
+                    result = await delegate_to_librarian(task=user_message)
+                elif agent == "biographer":
+                    result = await delegate_to_biographer(task=user_message)
+                elif agent == "housekeeper":
+                    result = await delegate_to_housekeeper(task=user_message)
+                else:
+                    result = None
+
+                duration = time_module.time() - start_time
+                await tracker.track_call(f"delegate_to_{agent}", duration)
+
+                if result and result.success:
+                    expert_results[agent] = result.output
+                    tools_called.append(f"delegate_to_{agent}")
+                    # Emit success think message
+                    success_msg = get_think_message(agent, user_message, "success")
+                    think_messages.append(success_msg + "\n")
+                else:
+                    error_msg = result.error if result else "Unknown error"
+                    expert_results[agent] = f"Error: {error_msg}"
+                    # Emit error think message
+                    error_think = get_think_message(agent, user_message, "error")
+                    think_messages.append(error_think + "\n")
+
+            except Exception as e:
+                expert_results[agent] = f"Error: {e}"
+                error_think = get_think_message(agent, user_message, "error")
+                think_messages.append(error_think + "\n")
+
+        return {
+            "tools_called": tools_called,
+            "expert_results": expert_results,
+            "tool_outputs": {},
+            "raw_output": "",
+            "think_messages": think_messages,
+        }
+
     async def stream_response(
         self,
         request: "ResponseRequest"  # type: ignore # Forward reference