fix(chat): preserve substantive answers after context handoff

loopx/capabilities/manager_context/README.md

@@ -560,7 +560,10 @@ content. Updating it in a release must not reject an otherwise identical saved
 request. Replay retains the original entry bytes, guidance, assessment and
 return history; message, brief and recipient identity conflicts still fail.
 
-The default shared Chat handoff acknowledgment is a short user-facing preview:
+The shared Chat handoff acknowledgment preserves the current audience-processed
+answer, including its Markdown, and appends receipt-backed delivery and execution
+status. It does not repeat the full brief or historical context after a substantive
+answer. A missing or blank answer uses the short user-facing preview:
 the exact recipient Agent, purpose, a bulleted result checklist (up to three
 items), execution boundaries (up to two items), and the return requirement.
 Each item has a visible text bound; omitted list items are counted. Selection

loopx/capabilities/manager_context/execution.py

@@ -6,7 +6,10 @@
 
 from pathlib import Path
 from collections.abc import Callable
-from typing import Any
+from typing import Any, TYPE_CHECKING
+
+if TYPE_CHECKING:
+    from ...chat_store import ChatSessionStore
 
 from ...agent_registry import load_goal_from_registry
 from ...collaboration_mcp import Delegations
@@ -167,11 +170,21 @@ def handoff_response(
     response: dict[str, Any],
     source_authorized: Callable[[], bool],
     execution_allowed: Callable[[], bool],
-    source_store=None,
+    source_store: "ChatSessionStore | None" = None,
 ) -> dict[str, Any]:
-    """Present context delivery and separately admitted execution on one path."""
+    """Append receipt-backed status to the existing audience-processed answer.
+
+    This presentation adapter does not reinterpret model prose as execution
+    evidence or replace the answer already shown by the Chat stream.
+    """
     from . import deliver
 
+    answer = response.get("message")
+    answer = answer if isinstance(answer, str) and answer.strip() else ""
+
+    def present(notice: str) -> str:
+        return answer + "\n\n" + notice if answer else notice
+
     try:
         if not source_authorized():
             raise ValueError("manager connection authority is no longer available")
@@ -182,8 +195,8 @@ def handoff_response(
                              execution_allowed=execution_allowed)
         return {**response, "proposals": [], "gate": None,
                 "context_handoff_receipt": receipt, "context_execution": execution,
-                "message": handoff_message(receipt, execution,
-                                           brief=response["context_handoff"].get("brief"))}
+                "message": present(handoff_message(receipt, execution,
+                    brief=None if answer else response["context_handoff"].get("brief")))}
     except (OSError, ValueError):
         return {**response, "proposals": [], "gate": None,
-                "message": "材料尚未转交：目标绑定、来源授权或持久收件回读未通过。需要修复交接链路；没有改动任务或优先级。"}
+                "message": present("材料尚未转交：目标绑定、来源授权或持久收件回读未通过。需要修复交接链路；没有改动任务或优先级。")}

tests/test_chat_project_coordination.py

@@ -71,6 +71,7 @@ def test_project_chat_nested_cli_handoff_returns_once_after_restart(project, mon
     before = registry.read_bytes(), (repo / "ACTIVE_GOAL_STATE.md").read_bytes()
     store = ChatSessionStore(root)
     controller = ChatRuntimeController(store=store, codex_bin="codex", registry_path=registry)
+    answer = "**Source comparison**\n\nThe periods differ.\n\n- Keep the correction.\n- Do not infer growth."
 
     class ModelFixture:
         upstream_thread_id = "fixture-project-model"
@@ -79,7 +80,8 @@ def close_session(self): pass
         def start_turn(self, message, sink):
             assert "OTHER_PROJECT_PRIVATE" not in message
             assert '"scope": "owner_goal"' in message and '"agent_id": "coordinator"' in message
-            return {"schema_version": "loopx_chat_agent_response_v0", "message": "Forwarding the correction", "proposals": [], "gate": None,
+            sink("answer.delta", {"text": answer})
+            return {"schema_version": "loopx_chat_agent_response_v0", "message": answer, "proposals": [], "gate": None,
                     "context_handoff": {"goal_id": "research", "agent_id": "coordinator"}}
 
     monkeypatch.setattr(controller, "capabilities", lambda: [
@@ -100,6 +102,12 @@ def cli(agent, action, *args):
             message="Ask coordinator to check the corrected source with reviewer.", work_dir=repo, objective="Research delivery")
         done = controller.wait_for_turn(session_id=session["session_id"], turn_id=turn["turn_id"], timeout_sec=20)
         assert done["status"] == "completed", done
+        assert done["response"]["message"].startswith(answer + "\n\n")
+        assert done["response"]["message"].count(answer) == 1
+        streamed = [event["payload"]["text"] for event in store.events_after(session["session_id"], turn["turn_id"], None)
+                    if event["kind"] == "answer.delta"]
+        assert "".join(streamed) == answer
+        assert ChatSessionStore(root).load_turn(session["session_id"], turn["turn_id"])["response"] == done["response"]
         receipt = done["response"]["context_handoff_receipt"]
         request_id = receipt["request_id"]
         assert request_id in json.dumps(cli("coordinator", "read"))

tests/test_manager_context_execution.py

@@ -120,7 +120,7 @@ def test_model_receives_authorized_task_choices_after_manager_context_compaction
     assert "execution_binding_id" not in _turn_prompt(turn["message"], context_summary=json.dumps(unavailable))
 
 
-def test_handoff_response_preserves_receipt_and_separate_execution_status(flow):
+def test_handoff_response_preserves_answer_and_separate_execution_status(flow):
     root, registry, (_, session, turn, request, _, _), started = flow
     response = execution.handoff_response(root, registry, session=session, turn=turn,
         response={"context_handoff": request, "message": "Unverified model completion claim.",
@@ -129,12 +129,70 @@ def test_handoff_response_preserves_receipt_and_separate_execution_status(flow):
     assert response["context_handoff_receipt"]["status"] == "delivered"
     assert response["context_execution"]["status"] == "prepared"
     assert "尚未确认完成" in response["message"]
-    assert request["brief"]["purpose"] in response["message"]
-    assert "Unverified model completion claim" not in response["message"]
+    assert response["message"].startswith("Unverified model completion claim.\n\n")
+    assert request["brief"]["purpose"] not in response["message"]
+    # Preserving prose does not turn a model claim into execution evidence.
+    assert response["context_execution"]["status"] != "accepted"
     assert response["proposals"] == [] and response["gate"] is None
     assert len(started) == 1
 
 
+@pytest.mark.parametrize("message", [None, "", " \n\t"])
+def test_empty_handoff_answer_retains_compact_brief_preview(flow, message):
+    root, registry, (_, session, turn, request, _, _), _ = flow
+    response = execution.handoff_response(root, registry, session=session, turn=turn,
+        response={"context_handoff": request, "message": message},
+        source_authorized=lambda: True, execution_allowed=lambda: True)
+    assert response["message"].startswith("**已转交给 `worker`。**")
+    assert request["brief"]["purpose"] in response["message"]
+
+
+@pytest.mark.parametrize("refusal", ["revoke", "delivery", "stop"])
+def test_handoff_refusal_preserves_answer_without_inventing_execution(flow, monkeypatch, refusal):
+    root, registry, (_, session, turn, request, _, _), started = flow
+    original = normalize_agent_response({"context_handoff": request,
+        "message": f"结论：保留现有数据。\n\n- 已核对 {root / 'private' / 'source.txt'}。\n- 恢复仍须独立验收。"},
+        protected_paths=[root])
+    answer = original["message"]
+    assert str(root) not in answer
+    if refusal == "delivery":
+        from loopx.capabilities import manager_context
+        def unavailable(*args, **kwargs):
+            raise OSError("delivery unavailable")
+        monkeypatch.setattr(manager_context, "deliver", unavailable)
+    response = execution.handoff_response(root, registry, session=session, turn=turn,
+        response=original,
+        source_authorized=lambda: refusal != "revoke",
+        execution_allowed=lambda: refusal != "stop")
+    assert response["message"].startswith(answer + "\n\n")
+    assert response["message"].count(answer) == 1
+    assert not started
+    if refusal == "stop":
+        assert response["context_handoff_receipt"]["status"] == "delivered"
+        assert response["context_execution"]["submitted"] is False
+        assert "本次未提交受控执行" in response["message"]
+    else:
+        assert "尚未转交" in response["message"]
+        assert "context_handoff_receipt" not in response
+
+
+def test_replayed_handoff_keeps_one_answer_and_one_original_operation(flow):
+    root, registry, (_, session, turn, request, _, _), started = flow
+    answer = "**检查结果**\n\n```text\nretain\n```\n\n原约束仍适用。"
+    original = {"context_handoff": request, "message": answer}
+    results = [execution.handoff_response(root, registry, session=session, turn=turn,
+        response=original, source_authorized=lambda: True, execution_allowed=lambda: True)
+        for _ in range(2)]
+    assert original["message"] == answer
+    assert len(started) == 1
+    assert results[0]["context_execution"]["operation_id"] == results[1]["context_execution"]["operation_id"]
+    assert results[1]["context_execution"]["replayed"]
+    for result in results:
+        assert result["message"].startswith(answer + "\n\n")
+        assert result["message"].count(answer) == 1
+        assert request["brief"]["context"] not in result["message"]
+
+
 def test_handoff_scope_revocation_stops_before_inbox_delivery(flow, monkeypatch):
     root, registry, (_, session, turn, request, _, _), started = flow
     from loopx.capabilities import manager_context