代码修改

offline-backend-20260512/backend/app/gateway/routers/intent.py

@@ -133,6 +133,70 @@ def _normalize_options(raw: object) -> tuple[str, ...]:
     return tuple(sorted(o for o in out if o))
 
 
+def _count_past_clarifications(messages: list[dict[str, Any]]) -> int:
+    """统计当前 thread 内已经发出过的 ``ask_clarification`` 次数。
+
+    用来给 ``stream_intent`` 做"3 轮硬上限"兜底:
+      - 0-1 次:正常流程,模型自行决定继续问还是收敛
+      - 2 次(即将进入第 3 轮):prompt 追加系统提示,要求本轮必须 INTENT_READY
+      - ≥3 次:工具白名单移除 ``ask_clarification``,模型物理上无法再问
+
+    只数"已经成型"的 tool_call,不区分用户是否回答过——LangGraph 把 ask
+    工具的调用永久记录在 AIMessage.tool_calls 里,数它即可。
+    """
+    if not isinstance(messages, list):
+        return 0
+    count = 0
+    for msg in messages:
+        if not isinstance(msg, dict):
+            continue
+        if (msg.get("type") or "").lower() not in ("ai", "aimessage", "aimessagechunk"):
+            continue
+        for tc in msg.get("tool_calls") or []:
+            if (tc.get("name") or "").lower() == "ask_clarification":
+                count += 1
+    return count
+
+
+async def _fetch_thread_messages(
+    client: httpx.AsyncClient,
+    base: str,
+    headers: dict[str, str],
+    thread_id: str,
+) -> list[dict[str, Any]]:
+    """读取当前 thread checkpoint 的 messages 列表(best-effort)。
+
+    用 ASGITransport loopback 调 ``GET /api/threads/{tid}/state``,失败时返回
+    空列表——计数兜底是个"刹车",失败时退化成原有的纯 SOUL 自律行为,不阻断主流程。
+    """
+    try:
+        resp = await client.get(f"{base}/api/threads/{thread_id}/state", headers=headers)
+    except Exception as exc:
+        logger.warning("intent: get thread state failed (%s) — falling back to soul-only flow", exc)
+        return []
+    if resp.status_code != 200:
+        logger.warning(
+            "intent: get thread state %s returned %s — falling back to soul-only flow",
+            thread_id,
+            resp.status_code,
+        )
+        return []
+    try:
+        payload = resp.json()
+    except Exception:
+        return []
+    values = payload.get("values") if isinstance(payload, dict) else None
+    if not isinstance(values, dict):
+        return []
+    messages = values.get("messages")
+    return messages if isinstance(messages, list) else []
+
+
+# 硬上限:超过这个轮数,无论模型怎么想,都强制移除 ask_clarification 工具。
+# 与 SOUL.md §0 「硬规则 2」对齐——"绝不存在第 4 轮"。
+_MAX_CLARIFICATION_ROUNDS = 3
+
+
 def _is_repeated_clarification(
     messages: list[dict[str, Any]],
     *,
@@ -279,33 +343,70 @@ async def stream_intent(request: Request, payload: dict = Body(...)) -> Streamin
         # 因此既无须 trust_env=False 防代理，也不存在 loopback 40s 超时这种坑。
         async with _make_loopback_client(request) as client:
             try:
+                # —— 3 轮硬上限兜底 ——
+                # 在发起 run 之前先读一次 thread state,数当前 ask_clarification 已发出次数:
+                #   past=0/1 → 正常,不动 message,不动工具白名单
+                #   past=2(本轮即将变成第 3 轮)→ 给消息加前缀,强提示"必须 INTENT_READY"
+                #   past≥3 → 直接把 ask_clarification 也加进 excluded_tools,工具物理消失
+                # 这一步是个"刹车":SOUL 让模型自律,这里让系统兜底。
+                past_clarifications = _count_past_clarifications(
+                    await _fetch_thread_messages(client, base, headers, thread_id)
+                )
+
+                # 1. ask_clarification 的工具白名单——超过硬上限就移除。
+                base_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",
+                ]
+                if past_clarifications >= _MAX_CLARIFICATION_ROUNDS:
+                    base_excluded_tools.append("ask_clarification")
+                    logger.info(
+                        "intent stream: past_clarifications=%d >= %d, force-disabling "
+                        "ask_clarification to break the loop",
+                        past_clarifications,
+                        _MAX_CLARIFICATION_ROUNDS,
+                    )
+
+                # 2. message 前缀——给模型一个明确的"本轮必须收敛"信号。
+                effective_message = message
+                if past_clarifications >= _MAX_CLARIFICATION_ROUNDS:
+                    effective_message = (
+                        f"[系统强制] 已澄清 {past_clarifications} 轮,触发硬上限。"
+                        f"本轮请直接输出 [INTENT_READY],缺的字段用 assumptions 兜底,"
+                        f"绝对不要再调任何工具,也不要再询问任何问题。\n\n"
+                        f"用户最新输入:{message}"
+                    )
+                elif past_clarifications >= _MAX_CLARIFICATION_ROUNDS - 1:
+                    effective_message = (
+                        f"[系统提示] 已澄清 {past_clarifications} 轮,这是最后一轮。"
+                        f"本轮请直接输出 [INTENT_READY],未明确的字段用 assumptions 兜底。\n\n"
+                        f"用户最新输入:{message}"
+                    )
+
                 body = _run_payload(
                     agent_name=INTENT_AGENT_ID,
                     model_name=model_name,
                     thread_id=thread_id,
-                    new_message=message,
-                    # Intent agent only ever needs ask_clarification. Block
+                    new_message=effective_message,
+                    # Intent agent only ever needs ask_clarification (and even that
+                    # is yanked once past_clarifications hits the cap above). 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.
-                    # 仅保留 ask_clarification；其余工具见模块 docstring「工具白名单」。
-                    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",
-                    ],
+                    # (e.g. saving a design doc with write_file) — SOUL prose is
+                    # not reliable on its own, hard exclusion is.
+                    excluded_tools=base_excluded_tools,
                     skill_stop_names=[],
                 )
 
@@ -338,7 +439,13 @@ async def stream_intent(request: Request, payload: dict = Body(...)) -> Streamin
                         except json.JSONDecodeError as exc:
                             logger.warning("intent summary JSON parse failed: %s", exc)
                             summary = None
-                        ready_text = msg_text
+                        # SOUL §0 硬规则 3 要求 ```json``` 反引号闭合后不得再输出任何字符,
+                        # 但模型仍偶尔会在后面接 "那么我帮你列一下大纲……" 之类的尾巴。
+                        # 终态帧只回到 marker 块结尾 + 一个换行,把废话从 content 里裁掉——
+                        # 草稿持久化 / 跨设备重放 / 前端 splitIntentReadyBlock 的 after
+                        # 段都因此干净。流式期间用户仍会逐 token 看到那段废话,前端
+                        # Step1Panel 已经把 `parts.after` 隐藏掉,体验闭环。
+                        ready_text = msg_text[: m.end()].rstrip() + "\n"
                         break
                 if summary and ready_text:
                     final_frame = {