diff --git a/offline-backend-20260512/backend/app/gateway/routers/intent.py b/offline-backend-20260512/backend/app/gateway/routers/intent.py
index 8001d2e4..0835bb73 100644
--- a/offline-backend-20260512/backend/app/gateway/routers/intent.py
+++ b/offline-backend-20260512/backend/app/gateway/routers/intent.py
@@ -111,13 +111,26 @@ async def stream_intent(request: Request, payload: dict = Body(...)) -> Streamin
                     model_name=model_name,
                     thread_id=thread_id,
                     new_message=message,
-                    # Intent agent shouldn't dispatch, search, write files, or
-                    # view images — its only tool of note is ask_clarification.
+                    # Intent agent only ever needs ask_clarification. Block
+                    # everything else so the model can't drift into solutioning
+                    # (e.g. saving a design doc with write_file) — the model
+                    # tends to ignore SOUL prose, hard exclusion is reliable.
                     excluded_tools=[
                         "web_search",
                         "present_files",
                         "view_image",
                         "agent_orchestration",
+                        "bash",
+                        "ls",
+                        "read_file",
+                        "write_file",
+                        "str_replace",
+                        "memory",
+                        "hindsight_recall",
+                        "hindsight_reflect",
+                        "hindsight_retain",
+                        "task",
+                        "write_todos",
                     ],
                     skill_stop_names=[],
                 )
@@ -163,14 +176,33 @@ async def stream_intent(request: Request, payload: dict = Body(...)) -> Streamin
                         parts.append(question)
                     content = "\n\n".join(parts) if parts else content
 
+                # Scan ALL AI messages for the [INTENT_READY] marker, not just
+                # dispatch_msg.content. The model often emits the summary in an
+                # earlier AI message and then keeps yapping (or doing extra
+                # tool calls) — the marker is the source of truth, wherever it
+                # appears in the run.
                 summary: dict[str, Any] | None = None
-                match = _INTENT_READY_PATTERN.search(content)
-                if match:
-                    try:
-                        summary = json.loads(match.group(1))
-                    except json.JSONDecodeError as exc:
-                        logger.warning("intent summary JSON parse failed: %s", exc)
-                        summary = None
+                ready_text: str = ""
+                for msg in reversed(messages):
+                    if not isinstance(msg, dict):
+                        continue
+                    if (msg.get("type") or "").lower() not in ("ai", "aimessage", "aimessagechunk"):
+                        continue
+                    msg_text = _flatten_content(msg.get("content", ""))
+                    m = _INTENT_READY_PATTERN.search(msg_text)
+                    if m:
+                        try:
+                            summary = json.loads(m.group(1))
+                        except json.JSONDecodeError as exc:
+                            logger.warning("intent summary JSON parse failed: %s", exc)
+                            summary = None
+                        ready_text = msg_text
+                        break
+                # When the marker was found in an earlier message, surface that
+                # message's text (which contains the user-facing 📋 汇总 + JSON)
+                # instead of the final filler AI text like "好的，收到".
+                if summary and ready_text:
+                    content = ready_text
 
                 status_value = "done" if summary else "asking"
                 final_frame: dict[str, Any] = {