优化测试脚本stage分别输出

tools/test_validate/modules/stage_debugger.py

@@ -39,6 +39,52 @@ def _select_stage():
         print("❌ 请输入 1-6 之间的数字")
 
 
+def _summarize_system_prompt_origin(obj: dict, upstream: dict) -> str:
+    """
+    根据上游 stage1 推断 system_prompt 的拼接顺序，仅展示文件名，不输出全文。
+    """
+    stage1 = upstream.get("stage1") or {}
+    scene_mode = stage1.get("scene_mode", "scene4")
+    drone_state = stage1.get("drone_state", "on_ground")
+    intent_type = stage1.get("intent_type", "generic_mission")
+
+    if scene_mode == "simple":
+        return "【system_prompt 来源】simple_mode_prompt.txt"
+
+    parts = [
+        "macro_header.txt",
+        "core_nodes.json (裁剪后)",
+        f"template_{'ground' if drone_state == 'on_ground' else 'air'}.txt",
+        "common_rules.txt",
+    ]
+    if intent_type == "generic_mission":
+        parts.append("[可选] system_extra_examples.txt")
+    parts.append("任务意图标签 (代码生成)")
+    return "【system_prompt 来源】" + " → ".join(parts)
+
+
+def _summarize_output_for_display(obj: dict, upstream: dict) -> dict:
+    """将 output/upstream 中的 system_prompt、final_prompt 替换为简要说明，便于终端查看"""
+    if not isinstance(obj, dict):
+        return obj
+    out = {}
+    for k, v in obj.items():
+        if k == "system_prompt" and isinstance(v, str) and len(v) > 200:
+            out[k] = _summarize_system_prompt_origin(obj, upstream) + f"\n(全文约 {len(v)} 字符，已保存至 response.json)"
+        elif k == "final_prompt" and isinstance(v, str) and len(v) > 200:
+            out[k] = f"【final_prompt】= system_prompt + user_prompt (全文约 {len(v)} 字符，已保存至 response.json)"
+        elif isinstance(v, dict):
+            out[k] = _summarize_output_for_display(v, upstream)
+        elif isinstance(v, list):
+            out[k] = [
+                _summarize_output_for_display(item, upstream) if isinstance(item, dict) else item
+                for item in v
+            ]
+        else:
+            out[k] = v
+    return out
+
+
 def _infer_drone_state_from_prompt(prompt: str) -> str:
     """从指令中简单推断 drone_state（批量模式下若未单独指定则用此）"""
     if "在空中" in prompt or "已起飞" in prompt:
@@ -75,8 +121,23 @@ def run_stage_debug_single():
                 print(f"✅ 请求成功 (耗时: {result['latency']:.2f}s)")
                 data = result["data"]
                 output = data.get("output", data)
+                upstream = data.get("upstream") or {}
+                target_stage = data.get("target_stage", 0)
+                if target_stage >= 2 and not upstream:
+                    print("\n⚠️  未获取到上游 Stage 输出。请重启后端服务 (如 start_all.sh) 后重试。")
+                elif upstream:
+                    print("\n" + "=" * 60)
+                    print("【上游 Stage 输出】")
+                    print("=" * 60)
+                    for k, v in sorted(upstream.items()):
+                        print(f"\n--- {k} ---")
+                        summarized = _summarize_output_for_display(v, upstream)
+                        print(json.dumps(summarized, ensure_ascii=False, indent=2))
                 print("\n" + "=" * 60)
-                print(json.dumps(output, ensure_ascii=False, indent=2))
+                print(f"【Stage{data.get('target_stage', '?')} 输出】")
+                print("=" * 60)
+                summarized_output = _summarize_output_for_display(output, upstream)
+                print(json.dumps(summarized_output, ensure_ascii=False, indent=2))
                 print("=" * 60)
 
                 # 保存到 validation/temporary