feat(responses): pass conversation context and stream thinks in real time

- delegate_to_* now receives a trimmed conversation history (last ~6
  turns, 500 chars/turn) as context on both live direct-delegation
  paths (streaming and steward non-streaming), via new
  build_delegation_context helper
- _stream_direct_delegation restructured as an async generator: the
  butler 'start' think message streams BEFORE the expert runs and the
  success/error message right after it finishes, instead of all
  messages arriving after the research completed

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

src/agents/delegation.py

@@ -128,6 +128,50 @@ def _detect_action_type(expert: str, task: str) -> ActionType:
     return ActionType.RETRIEVE
 
 
+def build_delegation_context(
+    conversation_history: list[dict] | None,
+    max_turns: int = 6,
+    max_chars_per_turn: int = 500,
+) -> str:
+    """
+    Format the most recent conversation turns as delegation context.
+
+    Experts accept a context string but the live paths never passed the
+    in-scope conversation history; this trims it to the last few turns
+    so follow-up questions ("and what about X?") keep their referent.
+
+    Args:
+        conversation_history: Prior messages as {"role", "content"} dicts
+        max_turns: How many trailing turns to include
+        max_chars_per_turn: Truncation limit per turn
+
+    Returns:
+        str: Newline-joined "role: content" lines ("" when no history)
+    """
+    if not conversation_history:
+        return ""
+
+    lines = []
+    for msg in conversation_history[-max_turns:]:
+        if not isinstance(msg, dict):
+            continue
+        role = msg.get("role", "user")
+        content = msg.get("content", "")
+        if isinstance(content, list):
+            # Tolerate structured content parts
+            content = " ".join(
+                part.get("text", "") if isinstance(part, dict) else str(part)
+                for part in content
+            )
+        content = str(content).strip()
+        if content:
+            lines.append(f"{role}: {content[:max_chars_per_turn]}")
+
+    if not lines:
+        return ""
+    return "Recent conversation:\n" + "\n".join(lines)
+
+
 def get_think_message(expert: str, task: str, phase: str) -> str:
     """
     Get the appropriate think message for an expert delegation.