diff --git a/frontend/src/run/__tests__/fixtures.json b/frontend/src/run/__tests__/fixtures.json
new file mode 100644
index 00000000..aba76276
--- /dev/null
+++ b/frontend/src/run/__tests__/fixtures.json
@@ -0,0 +1 @@
+{"db": [{"seq": 1, "type": "run.started", "node_id": null, "ts": 1789630273.8135118, "data": {"nodes": 3, "resumed": false}}, {"seq": 2, "type": "node.started", "node_id": "input", "ts": 1789630273.820915, "data": {"node_type": "input", "label": "输入"}}, {"seq": 3, "type": "node.finished", "node_id": "input", "ts": 1789630273.8260312, "data": {"duration_ms": 0, "attempt": 1, "preview": {}, "artifact": "44136fa355b3678a1146ad16f7e8649e94fb4fc21fe77e8310c060f61caaff8a"}}, {"seq": 4, "type": "node.started", "node_id": "query_kpi", "ts": 1789630273.8285508, "data": {"node_type": "tool", "label": "查询 KPI 信息表"}}, {"seq": 5, "type": "tool.start", "node_id": "query_kpi", "ts": 1789630273.834141, "data": {"tool": "db_query__warehouse", "args": {"sql": "SELECT * FROM ANALYTICS.v_device_kpi FETCH FIRST 50 ROWS ONLY"}}}, {"seq": 6, "type": "tool.end", "node_id": "query_kpi", "ts": 1789630273.929114, "data": {"tool": "db_query__warehouse", "duration_ms": 95, "preview": "\"{\\\"columns\\\": [\\\"factory_code\\\", \\\"object_type\\\", \\\"object_id\\\", \\\"start_date_time\\\", \\\"end_date_time\\\", \\\"insert_time\\\", \\\"attribute_group\\\", \\\"attribute01\\\", \\\"attribute02\\\", \\\"attribute03\\\", \\\"attribute04\\\", \\\"attribute05\\\", \\\"attribute06\\\", \\\"attribute07\\\", \\\"attribute08\\\", \\\"attribute09\\\", \\\"attribute10\\\", \\\"attribute11\\\", \\\"attribute12\\\", \\\"attribute13\\\", \\\"attribute14\\\", \\\"attribute15\\\", \\\"attribute16\\\", \\\"attribute17\\\", \\\"attribute18\\\", \\\"attribute19\\\", \\\"attribute20\\\", \\\"load_time\\\", \\\"attribute21\\\", \\\"attribute22\\\", \\\"attribute23\\\", \\\"attribute24\\\", \\\"attribute25\\\", \\\"attribute26\\\", \\\"attribute27\\\", \\\"attribute28\\\", \\\"attribute29\\\", \\\"attribute30\\\", \\\"attribute31\\\", \\\"attribute32\\\", \\\"attribute33\\\", \\\"attribute34\\\", \\\"attribute35\\\", \\\"attribute36\\\", \\\"attribute37\\\", \\\"attribute38\\\", \\\"attribute39\\\", \\\"attribute40\\\", \\\"attribute41\\\", \\\"attribute42\\\", \\\"attribute43\\\", \\\"attribute44\\\", \\\"attribute45\\\", \\\"attribute46\\\", \\\"attribute47\\\", \\\"attribute48\\\", \\\"attribute49\\\", \\\"attribute50\\\", \\\"attribute51\\\", \\\"attribute52\\\", \\\"attribute53\\\", \\\"attribute54\\\", \\\"attribute55\\\", \\\"attribute56\\\", \\\"attribute57\\\", \\\"attribute58\\\", \\\"attribute59\\\", \\\"attribute60\\\", \\\"attribute61\\\", \\\"attribute62\\\", \\\"attribute63\\\", \\\"attribute64\\\", \\\"attribute65\\\", \\\"attribute66\\\", \\\"attribute67\\\", \\\"attribute68\\\", \\\"attribute69\\\", \\\"attribute70\\\"], \\\"rows\\\": [[\\\"1063\\\", \\\"设备\\\", \\\"EQ2023SWMZZZZZ0008\\\", \\\"2023-04-04T12:00:00\\\", \\\"2023-04-04T12:30:00\\\", \\\"2023-04-05T07:13:02.055000\\\", \\\"001\\\", 30, 30, \\\"1.4249\\\", \\\"7.7251\\\", 0, null, null, null, null, null, null, null, null, null, null, \\\"在线\\\", null, null, null, null, \\\"2023-04-05T07:15:01\\\", null, null, null, null, null, null, null, null, null, null, null, null, null, null, null, null, null, null, null, null, null, null, null, null, null, null, null, null, null, null, null, null, null, null, null, null, null, null, null, null, null, null, null, null, null, null, null, null, null, null], [\\\"1063\\\", \\\"设备\\\", \\\"EQ2023SWMZZZZZ0009\\\", \\\"2023-04", "artifact": "49efce484256d8667b853cd69b3b421cdc7a705d6cb36e0ad4534b084b3ea177"}}, {"seq": 7, "type": "node.finished", "node_id": "query_kpi", "ts": 1789630273.9320729, "data": {"duration_ms": 99, "attempt": 1, "preview": "{\"columns\": [\"factory_code\", \"object_type\", \"object_id\", \"start_date_time\", \"end_date_time\", \"insert_time\", \"attribute_group\", \"attribute01\", \"attribute02\", \"attribute03\", \"attribute04\", \"attribute05\", \"attribute06\", \"attribute07\", \"attribute08\", \"attribute09\", \"attribute10\", \"attribute11\", \"attribute12\", \"attribute13\", \"attribute14\", \"attribute15\", \"attribute16\", \"attribute17\", \"attribute18\", \"attribute19\", \"attribute20\", \"load_time\", \"attribute21\", \"attribute22\", \"attribute23\", \"attribute24\", \"attribute25\", \"attribute26\", \"attribute27\", \"attribute28\", \"attribute29\", \"attribute30\", \"attribute31\", \"attribute32\", \"attribute33\", \"attribute34\", \"attribute35\", \"attribute36\", \"attribute37\", \"attribute38\", \"attribute39\", \"attribute40\", \"attribute41\", \"attribute42\", \"attribute43\", \"attribute44\", \"attribute45\", \"attribute46\", \"attribute47\", \"attribute48\", \"attribute49\", \"attribute50\", \"attribute51\", \"attribute52\", \"attribute53\", \"attribute54\", \"attribute55\", \"attribute56\", \"attribute57\", \"attribute58\", \"attribute59\", \"attribute60\", \"attribute61\", \"attribute62\", \"attribute63\", \"attribute64\", \"attribute65\", \"attribute66\", \"attribute67\", \"attribute68\", \"attribute69\", \"attribute70\"], \"rows\": [", "artifact": "9c915393eb335c8cae9300e969c9580fc1ab092613ab2d8a16d5281da6ea0f5e"}}, {"seq": 8, "type": "node.started", "node_id": "output", "ts": 1789630273.9351192, "data": {"node_type": "output", "label": "输出 KPI 数据"}}, {"seq": 9, "type": "node.finished", "node_id": "output", "ts": 1789630273.937645, "data": {"duration_ms": 0, "attempt": 1, "preview": {"kpi_list": "\"{\\\"columns\\\": [\\\"factory_code\\\", \\\"object_type\\\", \\\"object_id\\\", \\\"start_date_time\\\", \\\"end_date_time\\\", \\\"insert_time\\\", \\\"attribute_group\\\", \\\"attribute01\\\", \\\"attribute02\\\", \\\"attribute03\\\", \\\"attribute04\\\", \\\"attribute05\\\", \\\"attribute06\\\", \\\"attribute07\\\", \\\"attribute08\\\", \\\"attribute09\\\", \\\"attribute10\\\", \\\"attribute11\\\", \\\"attribute12\\\", \\\"attribute13\\\", \\\"attribute14\\\", \\\"attribute15\\\", \\\"attribute16\\\", \\\"attribute17\\\", \\\"attribute18\\\", \\\"attribute19\\\", \\\"attribute20\\\", \\\"load_time\\\", \\\"attribute21\\\", \\\"attribute22\\\", \\\"attribute23\\\", \\\"attribute24\\\", \\\"attribute25\\\", \\\"attribute26\\\", \\\"attribute27\\\", \\\"attribute28\\\", \\\"attribute29\\\", \\\"attribute30\\\", \\\"attribute31\\\", \\\"attribute32\\\", \\\"attribute33\\\", \\\"attribute34\\\", \\\"attribute35\\\", \\\"attribute36\\\", \\\"attribute37\\\", \\\"attribute38\\\", \\\"attribute39\\\", \\\"attribute40\\\", \\\"attribute41\\\", \\\"attribute42\\\", \\\"attribute43\\\", \\\"attribute44\\\", \\\"attribute45\\\", \\\"attribute46\\\", \\\"attribute47\\\", \\\"attribute48\\\", \\\"attribute49\\\", \\\"attribute50\\\", \\\"attribute51\\\", \\\"attribute52\\\", \\\"attribute53\\\", \\\"attribute54\\\", \\\"attribute55\\\", \\\"attribute56\\\", \\\"attribute57\\\", \\\"attribute58\\\", \\\"attribute59\\\", \\\"attribute60\\\", \\\"attribute61\\\", \\\"at"}, "artifact": "4581db6b040a22c80de69ee3d5f3ef5f07928721ce90de4212ec13e02fdf939c"}}, {"seq": 10, "type": "run.finished", "node_id": null, "ts": 1789630273.942965, "data": {"output": {"kpi_list": "\"{\\\"columns\\\": [\\\"factory_code\\\", \\\"object_type\\\", \\\"object_id\\\", \\\"start_date_time\\\", \\\"end_date_time\\\", \\\"insert_time\\\", \\\"attribute_group\\\", \\\"attribute01\\\", \\\"attribute02\\\", \\\"attribute03\\\", \\\"attribute04\\\", \\\"attribute05\\\", \\\"attribute06\\\", \\\"attribute07\\\", \\\"attribute08\\\", \\\"attribute09\\\", \\\"attribute10\\\", \\\"attribute11\\\", \\\"attribute12\\\", \\\"attribute13\\\", \\\"attribute14\\\", \\\"attribute15\\\", \\\"attribute16\\\", \\\"attribute17\\\", \\\"attribute18\\\", \\\"attribute19\\\", \\\"attribute20\\\", \\\"load_time\\\", \\\"attribute21\\\", \\\"attribute22\\\", \\\"attribute23\\\", \\\"attribute24\\\", \\\"attribute25\\\", \\\"attribute26\\\", \\\"attribute27\\\", \\\"attribute28\\\", \\\"attribute29\\\", \\\"attribute30\\\", \\\"attribute31\\\", \\\"attribute32\\\", \\\"attribute33\\\", \\\"attribute34\\\", \\\"attribute35\\\", \\\"attribute36\\\", \\\"attribute37\\\", \\\"attribute38\\\", \\\"attribute39\\\", \\\"attribute40\\\", \\\"attribute41\\\", \\\"attribute42\\\", \\\"attribute43\\\", \\\"attribute44\\\", \\\"attribute45\\\", \\\"attribute46\\\", \\\"attribute47\\\", \\\"attribute48\\\", \\\"attribute49\\\", \\\"attribute50\\\", \\\"attribute51\\\", \\\"attribute52\\\", \\\"attribute53\\\", \\\"attribute54\\\", \\\"attribute55\\\", \\\"attribute56\\\", \\\"attribute57\\\", \\\"attribute58\\\", \\\"attribute59\\\", \\\"attribute60\\\", \\\"attribute61\\\", \\\"attribute62\\\", \\\"attribute63\\\", \\\"attribute64\\\", \\\"attribute65\\\", \\\"attribute66\\\", \\\"attribute67\\\", \\\"attribute68\\\", \\\"attribute69\\\", \\\"attribute70\\\"], \\\"rows\\\": [[\\\"1063\\\", \\\"设备\\\", \\\"EQ2023SWMZZZZZ0008\\\", \\\"2023-04-04T12:00:00\\\", \\\"2023-04-04T12:30:00\\\", \\\"2023-04-05T07:13:02.055000\\\", \\\"001\\\", 30, 30, \\\"1.4249\\\", \\\"7.7251\\\", 0, null, null, null, null, null, null, null, null, null, null, \\\"在线\\\", null, null, null, null, \\\"2023-04-05T07:15:01\\\", null, null, null, null, null, null, null, null, null, null, null, null, null, null, null, null, null, null, null, null, null, null, null, null, null, null, null, null, null, null, null, null, null, null, null, null, null, null, null, null, null, null, null, null, null, null, null, null, null, null], [\\\"1063\\\", \\\"设备\\\", \\\"EQ2023SWMZZZZZ0009\\\", \\\"2023-04"}, "usage": {"duration_ms": 130}, "duration_ms": 130}}], "human": [{"seq": 1, "type": "run.started", "node_id": null, "ts": 1789632951.2781608, "data": {"nodes": 3, "resumed": false}}, {"seq": 2, "type": "node.started", "node_id": "in", "ts": 1789632951.281773, "data": {"node_type": "input", "label": "in"}}, {"seq": 3, "type": "node.finished", "node_id": "in", "ts": 1789632951.284436, "data": {"duration_ms": 0, "attempt": 1, "preview": {}, "artifact": "44136fa355b3678a1146ad16f7e8649e94fb4fc21fe77e8310c060f61caaff8a"}}, {"seq": 4, "type": "node.started", "node_id": "h", "ts": 1789632951.28604, "data": {"node_type": "human", "label": "h"}}, {"seq": 5, "type": "human.requested", "node_id": "h", "ts": 1789632951.287274, "data": {"kind": "human_node", "node_id": "h", "mode": "approve", "title": "确认", "message": "", "schema": {}, "context": {}}}, {"seq": 6, "type": "run.interrupted", "node_id": "h", "ts": 1789632951.289222, "data": {"interrupt_id": "5b55e4440cbb6ea8fc041f3285e64356", "payload": {"kind": "human_node", "node_id": "h", "mode": "approve", "title": "确认", "message": "", "schema": {}, "context": {}}}}, {"seq": 7, "type": "run.resumed", "node_id": null, "ts": 1789632952.2943501, "data": {"response": {"approved": true, "note": ""}}}, {"seq": 8, "type": "run.started", "node_id": null, "ts": 1789632952.295941, "data": {"nodes": 3, "resumed": true}}, {"seq": 9, "type": "node.started", "node_id": "h", "ts": 1789632952.298927, "data": {"node_type": "human", "label": "h"}}, {"seq": 10, "type": "human.requested", "node_id": "h", "ts": 1789632952.299924, "data": {"kind": "human_node", "node_id": "h", "mode": "approve", "title": "确认", "message": "", "schema": {}, "context": {}}}, {"seq": 11, "type": "human.resolved", "node_id": "h", "ts": 1789632952.300632, "data": {"response": {"approved": true, "note": ""}}}, {"seq": 12, "type": "node.finished", "node_id": "h", "ts": 1789632952.302637, "data": {"duration_ms": 0, "attempt": 1, "preview": {"approved": true, "note": "", "__decision__": "approved"}, "artifact": "37c6c8932b33766e3422b1f6ba5cb0548613f0b734b2ac3a22a3decd3cf053b6"}}, {"seq": 13, "type": "node.started", "node_id": "out", "ts": 1789632952.303895, "data": {"node_type": "output", "label": "out"}}, {"seq": 14, "type": "node.finished", "node_id": "out", "ts": 1789632952.306143, "data": {"duration_ms": 0, "attempt": 1, "preview": {"result": ""}, "artifact": "f58d9226d7c4bc9e7b20258dffae67e24b8507ca86f4bab783533d6bb5810a16"}}, {"seq": 15, "type": "run.finished", "node_id": null, "ts": 1789632952.3088498, "data": {"output": {"result": ""}, "usage": {"duration_ms": 12}, "duration_ms": 12}}], "issue": [{"seq": 1, "type": "run.started", "node_id": null, "ts": 1789371953.1991081, "data": {"nodes": 3, "resumed": false}}, {"seq": 2, "type": "node.started", "node_id": "in", "ts": 1789371953.205951, "data": {"node_type": "input", "label": "in"}}, {"seq": 3, "type": "node.finished", "node_id": "in", "ts": 1789371953.2096052, "data": {"duration_ms": 0, "attempt": 1, "preview": {}, "artifact": "44136fa355b3678a1146ad16f7e8649e94fb4fc21fe77e8310c060f61caaff8a"}}, {"seq": 4, "type": "node.started", "node_id": "m", "ts": 1789371953.212884, "data": {"node_type": "metrics", "label": "m"}}, {"seq": 5, "type": "log", "node_id": "m", "ts": 1789371953.214304, "data": {"level": "info", "message": "口径卡 口径@v1：产出 1 个受控指标"}}, {"seq": 6, "type": "node.finished", "node_id": "m", "ts": 1789371953.2180061, "data": {"duration_ms": 0, "attempt": 1, "preview": {"kind": "metric_set", "caliber": "口径", "caliber_version": "v1", "metrics": "[1 项]", "text": "指标清单（口径 口径 @ v1）：\n- a（营收）= 100"}, "artifact": "a5a29ed14620d0a2d45c7481ed2e0c38f5ae0a33f2e0bc9a310663cae39c8dae"}}, {"seq": 7, "type": "node.started", "node_id": "out", "ts": 1789371953.224318, "data": {"node_type": "output", "label": "out"}}, {"seq": 8, "type": "issuance", "node_id": "out", "ts": 1789371953.225867, "data": {"tier": "formal", "missing_required": [], "missing_expected": [], "unmatched": 0, "calibers": [{"node": "m", "caliber": "口径", "version": "v1"}]}}, {"seq": 9, "type": "node.finished", "node_id": "out", "ts": 1789371953.228422, "data": {"duration_ms": 3, "attempt": 1, "preview": {"result": "", "_issuance": "{'tier': 'formal', 'calibers': [{'node': 'm', 'caliber': '口径', 'version': 'v1'}], 'metrics_checked': 1, 'missing_required': [], 'missing_expected': [], 'unmatched_numbers': [], 'matched_numbers': 1, 'gaps': [], 'declared_at': '2026-09-14T07:45:53.223082+00:00'}"}, "artifact": "3496d28212f7919a0320209c49e967256075ebacb7ec73797eda119742044723"}}, {"seq": 10, "type": "run.finished", "node_id": null, "ts": 1789371953.232348, "data": {"output": {"result": "", "_issuance": {"tier": "formal", "calibers": [{"node": "m", "caliber": "口径", "version": "v1"}], "metrics_checked": 1, "missing_required": [], "missing_expected": [], "unmatched_numbers": [], "matched_numbers": 1, "gaps": [], "declared_at": "2026-09-14T07:45:53.223082+00:00"}}, "usage": {"duration_ms": 33}, "duration_ms": 33}}], "failed": [{"seq": 1, "type": "run.started", "node_id": null, "ts": 1789459923.142487, "data": {"nodes": 5, "resumed": false}}, {"seq": 2, "type": "node.started", "node_id": "start", "ts": 1789459923.155498, "data": {"node_type": "input", "label": "输入"}}, {"seq": 3, "type": "node.finished", "node_id": "start", "ts": 1789459923.160125, "data": {"duration_ms": 0, "attempt": 1, "preview": {"question": "111"}, "artifact": "33cce89d8426526c7100b83d4f47b5dd04be226b8ccbbfa1d6dc915ab84698ed"}}, {"seq": 4, "type": "node.started", "node_id": "init_state", "ts": 1789459923.171509, "data": {"node_type": "code", "label": "初始化状态"}}, {"seq": 5, "type": "sandbox.start", "node_id": "init_state", "ts": 1789459923.1733382, "data": {"language": "python", "limits": {"timeout": 30, "memory_mb": 512, "cpus": 1.0, "network": false, "max_output": 20000}, "isolation": "auto"}}, {"seq": 6, "type": "sandbox.end", "node_id": "init_state", "ts": 1789459923.517236, "data": {"ok": true, "exit_code": 0, "duration_ms": 352, "backend": "microvm", "stdout": "", "stderr": ""}}, {"seq": 7, "type": "node.finished", "node_id": "init_state", "ts": 1789459923.520313, "data": {"duration_ms": 353, "attempt": 1, "preview": {"ok": true, "exit_code": 0, "stdout": "", "stderr": "", "duration_ms": 352, "timed_out": false, "truncated": false, "backend": "microvm", "files_written": "[0 项]", "error": null, "session_reset": false, "data": null, "text": ""}, "artifact": "71b0561366b2cfafee35957ecdc488e91cc9f7d55378987fdce3696b4cbbbe2f"}}, {"seq": 8, "type": "node.started", "node_id": "loop_sum", "ts": 1789459923.522104, "data": {"node_type": "loop", "label": "循环累加"}}, {"seq": 9, "type": "node.failed", "node_id": "loop_sum", "ts": 1789459923.523083, "data": {"error": "循环条件写错了：表达式里不允许出现 Set", "duration_ms": 0, "node_id": "loop_sum"}}, {"seq": 10, "type": "run.failed", "node_id": null, "ts": 1789459923.5266151, "data": {"error": "循环条件写错了：表达式里不允许出现 Set"}}], "think": [{"seq": 1, "type": "run.started", "node_id": null, "ts": 1789354107.535924, "data": {"nodes": 5, "resumed": false}}, {"seq": 2, "type": "node.started", "node_id": "start", "ts": 1789354107.544369, "data": {"node_type": "input", "label": "条目列表"}}, {"seq": 3, "type": "node.finished", "node_id": "start", "ts": 1789354107.547843, "data": {"duration_ms": 0, "attempt": 1, "preview": {"items": "[3 项]"}, "artifact": "381bf97db4e270c59c4fded9e1ef27d7bf1f398a2807824b9921ab5f727b6839"}}, {"seq": 4, "type": "node.started", "node_id": "each", "ts": 1789354107.5504498, "data": {"node_type": "loop", "label": "逐条循环"}}, {"seq": 5, "type": "edge.taken", "node_id": "each", "ts": 1789354107.5522308, "data": {"branch": "body", "iteration": 0, "total": 3, "mode": "foreach"}}, {"seq": 6, "type": "node.finished", "node_id": "each", "ts": 1789354107.5559611, "data": {"duration_ms": 0, "attempt": 1, "preview": {"__decision__": "body", "iteration": 0, "total": 3, "item": "降低获客成本", "finished": false}, "artifact": "455052b94bfefdd7929de8a9f61ddca958735e20861ce61abc54b46f9943f8bc"}}, {"seq": 7, "type": "node.started", "node_id": "handle", "ts": 1789354107.5582778, "data": {"node_type": "llm", "label": "处理单条"}}, {"seq": 8, "type": "llm.start", "node_id": "handle", "ts": 1789354107.562285, "data": {"message_count": 1}}, {"seq": 39, "type": "llm.thinking", "node_id": "handle", "ts": 1789354113.069164, "data": {"text": "I need one concrete, actionable CAC-reduction tip under 40 characters. I'm considering splitting spend by channel and cutting the lowest-ROI 20% while boosting the best performer, or alternatively launching a referral reward program for existing customers.\n\n", "chars": 258, "truncated": false}}, {"seq": 40, "type": "llm.end", "node_id": "handle", "ts": 1789354113.075391, "data": {"model": "claude-opus-5", "duration_ms": 5504, "input_tokens": 59, "output_tokens": 236, "total_tokens": 295, "cost_usd": 0.006195, "calls": 1}}, {"seq": 41, "type": "node.finished", "node_id": "handle", "ts": 1789354113.0791411, "data": {"duration_ms": 5507, "attempt": 1, "preview": {"text": "按渠道核算获客成本，每月停掉 ROI 最低的 20% 预算，转投转化最高的渠道。", "thinking": "I need one concrete, actionable CAC-reduction tip under 40 characters. I'm considering splitting spend by channel and cutting the lowest-ROI 20% while boosting the best performer, or alternatively launching a referral reward program for existing customers.\n\n", "model": "claude-opus-5", "duration_ms": 5504}, "artifact": "cf412b9c7d3ed3df930e27d76710448b926cfa343e49c29e51cd32f4bab9b7dd"}}, {"seq": 42, "type": "node.started", "node_id": "collect", "ts": 1789354113.084119, "data": {"node_type": "transform", "label": "累积结果"}}, {"seq": 43, "type": "node.finished", "node_id": "collect", "ts": 1789354113.0874162, "data": {"duration_ms": 0, "attempt": 1, "preview": "· 按渠道核算获客成本，每月停掉 ROI 最低的 20% 预算，转投转化最高的渠道。\n", "artifact": "acc84091b73979eb7e09ab041698618f5dfb4373469fe483aa934df26ea182ab"}}, {"seq": 44, "type": "node.started", "node_id": "each", "ts": 1789354113.089597, "data": {"node_type": "loop", "label": "逐条循环"}}, {"seq": 45, "type": "edge.taken", "node_id": "each", "ts": 1789354113.090702, "data": {"branch": "body", "iteration": 1, "total": 3, "mode": "foreach"}}, {"seq": 46, "type": "node.finished", "node_id": "each", "ts": 1789354113.092248, "data": {"duration_ms": 0, "attempt": 1, "preview": {"__decision__": "body", "iteration": 1, "total": 3, "item": "提升续费率", "finished": false}, "artifact": "590540b5c92efaa4636f4c0c83f2d1508b3699b7047e6547621b3236f8937460"}}, {"seq": 47, "type": "node.started", "node_id": "handle", "ts": 1789354113.093664, "data": {"node_type": "llm", "label": "处理单条"}}, {"seq": 48, "type": "llm.start", "node_id": "handle", "ts": 1789354113.09708, "data": {"message_count": 1}}, {"seq": 104, "type": "llm.thinking", "node_id": "handle", "ts": 1789354125.2800598, "data": {"text": "I need something concise and actionable, under 40 characters. I'm drafting a suggestion around pushing usage-value reports before expiration paired with a renewal discount, and checking the character count to keep it within limits.\n\nActually, I think a higher-leverage suggestion would be identifying customers whose engagement drops in the 90 days before renewal and triggering proactive customer success outreach — this targets risk earlier. I'll go with that one since it's more impactful than the simple discount nudge.\n\n", "chars": 525, "truncated": false}}, {"seq": 105, "type": "llm.end", "node_id": "handle", "ts": 1789354125.286401, "data": {"model": "claude-opus-5", "duration_ms": 12177, "input_tokens": 58, "output_tokens": 637, "total_tokens": 695, "cost_usd": 0.016215, "calls": 1}}, {"seq": 106, "type": "node.finished", "node_id": "handle", "ts": 1789354125.291695, "data": {"duration_ms": 12180, "attempt": 1, "preview": {"text": "到期前 60 天识别活跃度下滑客户，由客户成功主动介入，附价值报告与续费方案。", "thinking": "I need something concise and actionable, under 40 characters. I'm drafting a suggestion around pushing usage-value reports before expiration paired with a renewal discount, and checking the character count to keep it within limits.\n\nActually, I think a higher-leverage suggestion would be identifying customers whose engagement drops in the 90 days before renewal and triggering proactive customer success outreach — this targets risk earlier. I'll go with that one since it's more impactful than the simple discount nudge.\n\n", "model": "claude-opus-5", "duration_ms": 12177}, "artifact": "db2eccae90c3baa4135f7f9cb8a987e7c07f5ef4f17783f4ba7b5d991971203f"}}, {"seq": 107, "type": "node.started", "node_id": "collect", "ts": 1789354125.296695, "data": {"node_type": "transform", "label": "累积结果"}}, {"seq": 108, "type": "node.finished", "node_id": "collect", "ts": 1789354125.300266, "data": {"duration_ms": 0, "attempt": 1, "preview": "· 按渠道核算获客成本，每月停掉 ROI 最低的 20% 预算，转投转化最高的渠道。\n· 到期前 60 天识别活跃度下滑客户，由客户成功主动介入，附价值报告与续费方案。\n", "artifact": "ca327d4774f9631dcf16e42191e827d1de9ed3409eac4844a138d87bd811714d"}}, {"seq": 109, "type": "node.started", "node_id": "each", "ts": 1789354125.304375, "data": {"node_type": "loop", "label": "逐条循环"}}, {"seq": 110, "type": "edge.taken", "node_id": "each", "ts": 1789354125.307219, "data": {"branch": "body", "iteration": 2, "total": 3, "mode": "foreach"}}, {"seq": 111, "type": "node.finished", "node_id": "each", "ts": 1789354125.310409, "data": {"duration_ms": 0, "attempt": 1, "preview": {"__decision__": "body", "iteration": 2, "total": 3, "item": "缩短上线周期", "finished": false}, "artifact": "aa1bd1a9f8c20746ec2b0cde1cef9b445ca4dba46fcf49cef19b9c87db80adc6"}}, {"seq": 112, "type": "node.started", "node_id": "handle", "ts": 1789354125.313403, "data": {"node_type": "llm", "label": "处理单条"}}, {"seq": 113, "type": "llm.start", "node_id": "handle", "ts": 1789354125.319797, "data": {"message_count": 1}}, {"seq": 134, "type": "llm.end", "node_id": "handle", "ts": 1789354128.034537, "data": {"model": "claude-opus-5", "duration_ms": 2710, "input_tokens": 60, "output_tokens": 49, "total_tokens": 109, "cost_usd": 0.001525, "calls": 1}}, {"seq": 135, "type": "node.finished", "node_id": "handle", "ts": 1789354128.042933, "data": {"duration_ms": 2715, "attempt": 1, "preview": {"text": "建议：将部署流程自动化，代码合并后自动触发构建、测试与灰度发布，把人工审批环节压缩到单点。", "thinking": "", "model": "claude-opus-5", "duration_ms": 2710}, "artifact": "1dd859e26cadc45a1ae451b141c0b52fbf55eaa544cc2b644692c9aa999e27d7"}}, {"seq": 136, "type": "node.started", "node_id": "collect", "ts": 1789354128.0519838, "data": {"node_type": "transform", "label": "累积结果"}}, {"seq": 137, "type": "node.finished", "node_id": "collect", "ts": 1789354128.061368, "data": {"duration_ms": 0, "attempt": 1, "preview": "· 按渠道核算获客成本，每月停掉 ROI 最低的 20% 预算，转投转化最高的渠道。\n· 到期前 60 天识别活跃度下滑客户，由客户成功主动介入，附价值报告与续费方案。\n· 建议：将部署流程自动化，代码合并后自动触发构建、测试与灰度发布，把人工审批环节压缩到单点。\n", "artifact": "65a6ab813cc4b27d3d4b87eb5a043cfb496f7c11edb1ec4f5c4a613e34e469e4"}}, {"seq": 138, "type": "node.started", "node_id": "each", "ts": 1789354128.0655859, "data": {"node_type": "loop", "label": "逐条循环"}}, {"seq": 139, "type": "edge.taken", "node_id": "each", "ts": 1789354128.068269, "data": {"branch": "done", "iteration": 3, "total": 3, "mode": "foreach"}}, {"seq": 140, "type": "node.finished", "node_id": "each", "ts": 1789354128.0713658, "data": {"duration_ms": 0, "attempt": 1, "preview": {"__decision__": "done", "iteration": 3, "total": 3, "item": null, "finished": true}, "artifact": "3d1328a816ace27d9b66281108eb1132835835eb06a38b9deb49f3389e3e7ec0"}}, {"seq": 141, "type": "node.started", "node_id": "done", "ts": 1789354128.073812, "data": {"node_type": "output", "label": "汇总"}}, {"seq": 142, "type": "node.finished", "node_id": "done", "ts": 1789354128.07654, "data": {"duration_ms": 0, "attempt": 1, "preview": {"逐条建议": "· 按渠道核算获客成本，每月停掉 ROI 最低的 20% 预算，转投转化最高的渠道。\n· 到期前 60 天识别活跃度下滑客户，由客户成功主动介入，附价值报告与续费方案。\n· 建议：将部署流程自动化，代码合并后自动触发构建、测试与灰度发布，把人工审批环节压缩到单点。\n", "处理条数": "3"}, "artifact": "668932bde6139ccac8af5eff0756698a1e1fd1881c7e53eb7e2f437deae5d655"}}, {"seq": 143, "type": "run.finished", "node_id": null, "ts": 1789354128.081253, "data": {"output": {"逐条建议": "· 按渠道核算获客成本，每月停掉 ROI 最低的 20% 预算，转投转化最高的渠道。\n· 到期前 60 天识别活跃度下滑客户，由客户成功主动介入，附价值报告与续费方案。\n· 建议：将部署流程自动化，代码合并后自动触发构建、测试与灰度发布，把人工审批环节压缩到单点。\n", "处理条数": "3"}, "usage": {"input_tokens": 177, "output_tokens": 922, "total_tokens": 1099, "cost_usd": 0.023934999999999998, "calls": 3, "duration_ms": 20544}, "duration_ms": 20544}}]}
\ No newline at end of file
diff --git a/frontend/src/run/decode.ts b/frontend/src/run/decode.ts
new file mode 100644
index 00000000..9856f347
--- /dev/null
+++ b/frontend/src/run/decode.ts
@@ -0,0 +1,566 @@
+import type { RunEvent } from '../types'
+
+/**
+ * 事件 → 人能看懂的步骤。
+ *
+ * 这是全站唯一的翻译层：画布时间线、助手栏、运行页都用它。分两套翻译的话，
+ * 同一次运行在不同地方会读出不同的故事，而用户没法判断哪个是真的。
+ *
+ * 三条原则：
+ *
+ * 1. **不回写事件**。事件表是审计凭证——run 终态时对全部已落库事件算
+ *    manifest_hash，事后增删改会对不上。翻译只发生在前端内存里。
+ * 2. **成对事件合并成一行**。tool.start + tool.end 是同一件事的两个时刻，
+ *    拆成两行等于把"查询数据"和"查完了"当成两件事讲。
+ * 3. **节点是天然的分组单位**。一个图节点内部可能有好几轮模型调用和工具
+ *    调用，它们属于同一个"步骤"——这也正是用户说的"简要思考节点"：
+ *    节点本身就是那个节点，里面的动作是它的过程。
+ */
+
+export type StepKind =
+  | 'node'      // 一个图节点的执行（可能带子步骤）
+  | 'think'     // 模型的思考
+  | 'llm'       // 模型调用
+  | 'query'     // 数据库查询
+  | 'schema'    // 看表结构
+  | 'tool'      // 其他工具
+  | 'code'      // 沙箱代码
+  | 'branch'    // 分支走向
+  | 'human'     // 人工介入
+  | 'issuance'  // 出具判定
+  | 'note'      // 日志/提示
+  | 'error'
+  | 'lifecycle' // 开始/结束
+
+export type StepStatus = 'running' | 'done' | 'failed' | 'waiting'
+
+export interface Step {
+  id: string
+  seq: number
+  kind: StepKind
+  /** 一句人话，不带技术黑话 */
+  title: string
+  /** 展开后看的东西：SQL 原文、工具参数、结果预览、思考全文 */
+  detail?: string
+  /** 行尾的次要信息：耗时、行数 */
+  meta?: string
+  status?: StepStatus
+  level?: 'info' | 'warn' | 'error'
+  nodeId?: string
+  /** 工件 id，可下钻到完整证据 */
+  artifact?: string
+  children?: Step[]
+}
+
+// 这两个是流式增量，后端根本不落库（_EPHEMERAL）。单次运行的 delta 量级会
+// 把列表淹掉，而且刷新页面后它们不会回来——UI 不能建立在它们之上。
+const EPHEMERAL = new Set(['llm.token', 'llm.thinking.delta'])
+
+// 节点类型的中文名。刻意内联而不是 import canvas/nodeDefs：解码器是纯粹的
+// 翻译层，画布只是它的消费者之一，反过来依赖画布会让运行页也被迫加载
+// 整套节点定义（含图标、表单 schema）。这张表只有 16 行，重复得起。
+const TYPE_LABEL: Record<string, string> = {
+  input: '输入', output: '成果', llm: '模型调用', agent: 'Agent',
+  supervisor: '多 Agent 协作', tool: '调用工具', code: '沙箱代码',
+  branch: '条件分支', loop: '循环', subgraph: '子工作流',
+  memory: '长期记忆', retrieve: '知识检索', transform: '数据整形',
+  human: '人工介入', validate: '结构校验', metrics: '口径卡',
+}
+
+// 只对开发者有意义、对使用者是噪音的。不是丢弃事件本身（原始事件另有视图），
+// 只是不进这条给人看的流。
+const SILENT = new Set(['usage', 'node.skipped'])
+
+const num = (v: unknown): number | undefined =>
+  typeof v === 'number' && Number.isFinite(v) ? v : undefined
+
+export function formatDuration(ms?: number): string {
+  if (ms == null) return ''
+  if (ms < 1000) return `${Math.round(ms)}ms`
+  if (ms < 60_000) return `${(ms / 1000).toFixed(1)}s`
+  return `${Math.floor(ms / 60_000)}m${Math.round((ms % 60_000) / 1000)}s`
+}
+
+/** 从 db_query 的结果预览里抠出行数——用户关心的是"查到多少"，不是 JSON。 */
+function rowCountOf(preview: string): number | undefined {
+  const m = preview.match(/"row_count":\s*(\d+)/)
+  return m ? Number(m[1]) : undefined
+}
+
+/** 把查询结果预览转成小表格能用的形状；解析不了就算了，原样展示。 */
+export function parseQueryResult(preview: string):
+  { columns: string[]; rows: unknown[][]; truncated: boolean } | null {
+  try {
+    const raw = preview.trim().startsWith('"') ? JSON.parse(preview) : preview
+    const data = typeof raw === 'string' ? JSON.parse(raw) : raw
+    if (Array.isArray(data?.columns) && Array.isArray(data?.rows)) {
+      return { columns: data.columns, rows: data.rows, truncated: !!data.truncated }
+    }
+  } catch {
+    /* 不是标准查询结果，当普通文本 */
+  }
+  return null
+}
+
+function toolStep(seq: number, tool: string, args: Record<string, any>): Step {
+  const base = { id: `tool-${seq}`, seq, status: 'running' as StepStatus }
+  // 数据库工具单独认：它是这个产品最主要的取数方式，"调用 db_query__warehouse"
+  // 这种说法等于没翻译
+  if (tool.startsWith('db_query')) {
+    const source = tool.replace(/^db_query__/, '')
+    return { ...base, kind: 'query', title: `在 ${source} 上查询数据`,
+             detail: String(args?.sql ?? '') }
+  }
+  if (tool.startsWith('db_schema')) {
+    const table = String(args?.table ?? '')
+    return { ...base, kind: 'schema',
+             title: table ? `查看 ${table} 的字段` : '列出有哪些表' }
+  }
+  const argText = Object.keys(args ?? {}).length
+    ? JSON.stringify(args, null, 2) : ''
+  return { ...base, kind: 'tool', title: `调用 ${tool}`, detail: argText }
+}
+
+/**
+ * 把一次运行的事件解码成步骤树。
+ *
+ * 顶层是节点，节点内部的模型/工具调用是子步骤。没有 node_id 的事件
+ * （run.*、issuance 这类）留在顶层，按时间顺序穿插。
+ */
+export function decodeRun(events: RunEvent[]): Step[] {
+  const out: Step[] = []
+  /** node_id → 该节点的顶层 Step，用于把子步骤挂进去 */
+  const nodeSteps = new Map<string, Step>()
+  /** call_id（或工具名兜底）→ 待配对的工具 Step */
+  const pendingTools = new Map<string, Step>()
+  /** node_id → 待配对的模型调用 Step */
+  const pendingLlm = new Map<string, Step>()
+  /** 已见过的 interrupt_id，避免 human.requested 与 run.interrupted 各出一行 */
+  const seenInterrupts = new Set<string>()
+
+  const push = (step: Step, nodeId?: string) => {
+    const parent = nodeId ? nodeSteps.get(nodeId) : undefined
+    if (parent) (parent.children ??= []).push(step)
+    else out.push(step)
+  }
+
+  for (const event of events) {
+    const type = String(event.type)
+    if (EPHEMERAL.has(type) || SILENT.has(type)) continue
+    const d: any = event.data ?? {}
+    const seq = event.seq ?? 0
+    const nodeId = event.node_id ?? undefined
+
+    switch (type) {
+      case 'run.started':
+      case 'run.resumed': {
+        const resumed = type === 'run.resumed' || !!d.resumed
+        // 恢复一次会同时来 run.resumed 和 run.started(resumed=true)，
+        // 说两遍"继续执行"只会让人以为恢复了两次
+        if (resumed && out.some((s) => s.kind === 'lifecycle' && s.title === '继续执行')
+            && !out.some((s) => s.status === 'waiting')) {
+          break
+        }
+        if (resumed) {
+          // 恢复意味着之前那条"等你确认"已经过去了
+          out.forEach((s) => { if (s.status === 'waiting') s.status = 'done' })
+          nodeSteps.forEach((s) => {
+            s.children?.forEach((c) => { if (c.status === 'waiting') c.status = 'done' })
+          })
+        }
+        out.push({
+          id: `s-${seq}`, seq, kind: 'lifecycle', status: 'running',
+          title: resumed ? '继续执行' : `开始执行（${d.nodes ?? '?'} 步）`,
+        })
+        break
+      }
+
+      case 'node.started': {
+        // data.label 是节点标题**随事件传递的唯一来源**。靠画布 nodes 反查的话，
+        // 助手栏和运行页没加载画布，标题会退化成裸 node_id。
+        // label 是节点标题随事件传递的唯一来源，但没起过名字的节点它等于
+        // 节点 id（后端兜底），"in"/"h"/"out" 对用户毫无意义——退到节点
+        // 类型的中文名，至少能看出这步在干嘛
+        const label = String(d.label ?? '')
+        const typeName = TYPE_LABEL[String(d.node_type ?? '')]
+        const title = (label && label !== nodeId ? label : typeName) || nodeId || '执行步骤'
+        // 人工审批恢复后 LangGraph 会重放该节点，node.started 因此来第二遍。
+        // 那不是"又执行了一个步骤"，是同一步继续——复用原来那条，
+        // 否则时间线上每审批一次就多一个同名节点。
+        const existing = nodeId ? nodeSteps.get(nodeId) : undefined
+        if (existing) {
+          existing.status = 'running'
+          break
+        }
+        const step: Step = {
+          id: `n-${nodeId ?? seq}`, seq, kind: 'node', nodeId, title, status: 'running',
+        }
+        if (nodeId) nodeSteps.set(nodeId, step)
+        out.push(step)
+        break
+      }
+
+      case 'node.finished': {
+        const step = nodeId ? nodeSteps.get(nodeId) : undefined
+        if (step) {
+          step.status = 'done'
+          step.meta = formatDuration(num(d.duration_ms))
+          step.artifact = d.artifact || step.artifact
+          // 空对象/空串不是"详情"，显示出来只是噪音
+          const raw = d.preview
+          const empty = raw == null || raw === '' ||
+            (typeof raw === 'object' && Object.keys(raw).length === 0)
+          if (!empty) {
+            const preview = typeof raw === 'string' ? raw : JSON.stringify(raw, null, 2)
+            step.detail = preview.slice(0, 4000)
+          }
+        }
+        break
+      }
+
+      case 'node.failed': {
+        const step = nodeId ? nodeSteps.get(nodeId) : undefined
+        if (step) {
+          step.status = 'failed'
+          step.level = 'error'
+          step.detail = String(d.error ?? '')
+          step.meta = formatDuration(num(d.duration_ms))
+        } else {
+          push({ id: `e-${seq}`, seq, kind: 'error', level: 'error',
+                 title: String(d.error ?? '这一步失败了'), nodeId }, nodeId)
+        }
+        break
+      }
+
+      case 'llm.start':
+        pendingLlm.set(nodeId ?? '_', {
+          id: `llm-${seq}`, seq, kind: 'llm', title: '思考并作答',
+          status: 'running', nodeId,
+        })
+        break
+
+      case 'llm.end': {
+        const step = pendingLlm.get(nodeId ?? '_')
+        if (step) {
+          step.status = 'done'
+          // 不给 token 数和美元——那是账单视角，不是"它干了什么"。
+          // 成本在运行详情的用量区单独看。
+          step.meta = formatDuration(num(d.duration_ms))
+          push(step, nodeId)
+          pendingLlm.delete(nodeId ?? '_')
+        }
+        break
+      }
+
+      case 'llm.thinking': {
+        // 只有 Anthropic 系模型会产出（thinking_text 认的是 thinking 块）。
+        // 换个 provider 就没有——所以它是增强信息，主干靠上面那些步骤撑着。
+        const text = String(d.text ?? d.delta ?? '')
+        if (!text.trim()) break
+        push({
+          id: `th-${seq}`, seq, kind: 'think', nodeId, status: 'done',
+          title: text.replace(/\s+/g, ' ').slice(0, 60) + (text.length > 60 ? '…' : ''),
+          detail: text + (d.truncated ? '\n\n（已截断）' : ''),
+        }, nodeId)
+        break
+      }
+
+      case 'tool.start': {
+        const step = toolStep(seq, String(d.tool ?? ''), d.args ?? {})
+        step.nodeId = nodeId
+        // call_id 才是可靠的配对键：同一节点并发调同名工具时，按名字配会错位。
+        // 独立 Tool 节点不带 call_id，回退到工具名。
+        pendingTools.set(String(d.call_id || d.tool || seq), step)
+        push(step, nodeId)
+        break
+      }
+
+      case 'tool.end':
+      case 'tool.error': {
+        const key = String(d.call_id || d.tool || '')
+        const step = pendingTools.get(key)
+        const preview = String(d.preview ?? d.error ?? '')
+        if (step) {
+          const failed = type === 'tool.error' || preview.startsWith('SQL 被拒绝')
+            || preview.startsWith('查询失败')
+          step.status = failed ? 'failed' : 'done'
+          if (failed) step.level = 'error'
+          // tool.error 只有 {tool,error}，没有 duration_ms/preview——不能假设统一形状
+          const rows = rowCountOf(preview)
+          const parts = [
+            rows != null ? `${rows} 行` : '',
+            formatDuration(num(d.duration_ms)),
+          ].filter(Boolean)
+          step.meta = parts.join(' · ')
+          step.artifact = d.artifact
+          if (step.kind === 'query' || step.kind === 'schema' || failed) {
+            // 查询保留 SQL 作为 detail，结果另挂；失败时结果就是错误原因
+            step.detail = failed ? `${step.detail ?? ''}\n\n${preview}`.trim() : step.detail
+            ;(step as any).result = preview
+          } else {
+            step.detail = preview.slice(0, 4000)
+          }
+          pendingTools.delete(key)
+        }
+        break
+      }
+
+      case 'sandbox.start':
+        pendingTools.set(`sandbox-${nodeId ?? seq}`, (() => {
+          const step: Step = {
+            id: `sb-${seq}`, seq, kind: 'code', nodeId, status: 'running',
+            title: `运行${d.language === 'python' ? ' Python' : d.language ? ` ${d.language}` : ''}代码`,
+          }
+          push(step, nodeId)
+          return step
+        })())
+        break
+
+      case 'sandbox.end': {
+        const step = pendingTools.get(`sandbox-${nodeId ?? seq}`)
+        if (step) {
+          step.status = d.ok ? 'done' : 'failed'
+          if (!d.ok) step.level = 'error'
+          step.meta = formatDuration(num(d.duration_ms))
+          const body = [d.stdout, d.stderr].filter(Boolean).join('\n').slice(0, 4000)
+          step.detail = body || (d.ok ? '（无输出）' : `退出码 ${d.exit_code}`)
+          pendingTools.delete(`sandbox-${nodeId ?? seq}`)
+        }
+        break
+      }
+
+      case 'edge.taken':
+        push({
+          id: `br-${seq}`, seq, kind: 'branch', nodeId, status: 'done',
+          title: `走「${d.branch}」这条路`
+            + (d.iteration != null ? `（第 ${Number(d.iteration) + 1} 轮）` : ''),
+          detail: d.reason ? String(d.reason) : undefined,
+        }, nodeId)
+        break
+
+      case 'human.requested':
+      case 'run.interrupted': {
+        // 同一次中断会来两条事件，而且审批恢复后节点重放还会再来一遍
+        // （LangGraph 的重放语义，不是 bug）。三条都指向同一次"等你确认"，
+        // 界面上只该有一个。
+        //
+        // 不能用 interrupt_id 做键：human.requested 压根没这个字段，
+        // run.interrupted 才有。两者共有的是 payload 里的 node_id + mode + title，
+        // 用它们才能让三条算出同一个键。
+        const payload = d.payload ?? d
+        const key = [payload.node_id ?? nodeId, payload.mode ?? '', payload.title ?? ''].join('|')
+        if (seenInterrupts.has(key)) break
+        seenInterrupts.add(key)
+        push({
+          id: `hm-${seq}`, seq, kind: 'human', nodeId, status: 'waiting', level: 'warn',
+          title: String(payload.title || d.title || '等你确认'),
+          detail: [payload.message, payload.tool ? `工具：${payload.tool}` : '']
+            .filter(Boolean).join('\n') || undefined,
+        }, nodeId)
+        break
+      }
+
+      case 'human.resolved': {
+        const r = d.response ?? {}
+        const approved = typeof r === 'object' ? r.approved : undefined
+        push({
+          id: `hr-${seq}`, seq, kind: 'human', nodeId, status: 'done',
+          title: approved === false ? '你驳回了' : '你放行了',
+          detail: typeof r === 'object' && r.note ? String(r.note) : undefined,
+        }, nodeId)
+        break
+      }
+
+      case 'issuance': {
+        const tier = String(d.tier ?? '')
+        const label = { formal: '正式出具', degraded: '降档出具', withheld: '不予出具' }[tier] ?? tier
+        const gaps: string[] = []
+        if (d.missing_required?.length) gaps.push(`缺必需指标 ${d.missing_required.join('、')}`)
+        if (d.missing_expected?.length) gaps.push(`缺期望指标 ${d.missing_expected.join('、')}`)
+        if (num(d.unmatched)) gaps.push(`${d.unmatched} 个数字无法回指指标集`)
+        out.push({
+          id: `is-${seq}`, seq, kind: 'issuance', status: 'done',
+          level: tier === 'formal' ? 'info' : tier === 'withheld' ? 'error' : 'warn',
+          title: gaps.length ? `${label}：${gaps.join('；')}` : label,
+        })
+        break
+      }
+
+      case 'caliber.upgrade':
+        out.push({
+          id: `cu-${seq}`, seq, kind: 'note', level: 'warn', status: 'done',
+          title: `口径卡 v${d.pinned} → v${d.latest} 有新版，按「${d.policy_label ?? d.policy}」处置`,
+        })
+        break
+
+      case 'agent.step.start':
+        push({
+          id: `as-${seq}`, seq, kind: 'note', nodeId, status: 'running',
+          title: `${d.agent}：${String(d.instruction ?? '').slice(0, 80)}`,
+        }, nodeId)
+        break
+
+      case 'agent.step.end':
+        push({
+          id: `ae-${seq}`, seq, kind: 'note', nodeId, status: 'done',
+          title: `${d.agent} 回复`, meta: formatDuration(num(d.duration_ms)),
+          detail: String(d.preview ?? '').slice(0, 2000),
+        }, nodeId)
+        break
+
+      case 'log': {
+        const level = String(d.level ?? 'info')
+        // info 级日志是给排查用的，不进主流程——但 warn/error 用户必须看到
+        if (level === 'info') break
+        push({
+          id: `lg-${seq}`, seq, kind: 'note', nodeId, status: 'done',
+          level: level === 'error' ? 'error' : 'warn',
+          title: String(d.message ?? ''),
+        }, nodeId)
+        break
+      }
+
+      case 'run.failed': {
+        const msg = String(d.error ?? '运行失败')
+        // 节点失败会先报一次，run.failed 往往是同一句话——说两遍不会让人
+        // 更明白，只会让人以为出了两个错
+        const dup = out.some((s) => s.status === 'failed' &&
+          (s.detail === msg || s.title === msg))
+        out.forEach((s) => { if (s.kind === 'lifecycle' && s.status === 'running') s.status = 'failed' })
+        if (!dup) {
+          out.push({
+            id: `rf-${seq}`, seq, kind: 'error', level: 'error', status: 'failed', title: msg,
+          })
+        }
+        break
+      }
+
+      case 'run.cancelled':
+        out.push({ id: `rc-${seq}`, seq, kind: 'lifecycle', status: 'failed', title: '已取消' })
+        break
+
+      case 'run.finished': {
+        // 开头那条"开始执行"要收尾，否则跑完了还挂着一个转圈的图标
+        const opener = out.find((s) => s.kind === 'lifecycle' && s.status === 'running')
+        if (opener) opener.status = 'done'
+        out.push({
+          id: `rd-${seq}`, seq, kind: 'lifecycle', status: 'done',
+          title: '完成', meta: formatDuration(num(d.duration_ms)),
+        })
+        break
+      }
+
+      default:
+        // 新增事件类型时不要静默吞掉——宁可显示一条原始的，也比让人觉得
+        // "什么都没发生"好。加了新事件记得回来补一条映射。
+        push({
+          id: `x-${seq}`, seq, kind: 'note', nodeId, status: 'done',
+          title: type, detail: JSON.stringify(d, null, 2).slice(0, 1000),
+        }, nodeId)
+    }
+  }
+
+  // 还没收到 end 的工具/模型调用保持 running——它们正在进行，不是丢了
+  return out
+}
+
+// -------------------------------------------------------------------------
+// Copilot 建图阶段：操作流 → 同一套 Step
+// -------------------------------------------------------------------------
+
+export interface CopilotOp {
+  op: string
+  [k: string]: any
+}
+
+const PHASE_LABEL: Record<string, string> = {
+  connecting: '正在连接模型',
+  planning: '正在理解需求、规划步骤',
+  building: '正在搭建流程',
+  wiring: '正在连接数据流',
+  finalizing: '正在排版和校验',
+}
+
+/**
+ * Copilot 的操作流解码。
+ *
+ * 和运行事件是两套完全独立的协议（不共享 seq、node_id、不落库），只在这里
+ * 统一成同一种 Step，让用户看到的是一条连续的过程，而不是"建图"和"执行"
+ * 两段风格迥异的日志。
+ */
+export function decodeCopilot(ops: CopilotOp[]): Step[] {
+  const out: Step[] = []
+  let i = 0
+  let nodeCount = 0
+  for (const op of ops) {
+    i += 1
+    switch (op.op) {
+      case 'thinking': {
+        const text = String(op.delta ?? '')
+        if (!text.trim()) break
+        const last = out[out.length - 1]
+        // 思考是连续流，一片 delta 一行会碎成几十条。同一段连续思考并成一条，
+        // 标题取开头、详情是全文——这才是"一次思考是一个节点"。
+        if (last?.kind === 'think') {
+          last.detail = (last.detail ?? '') + text
+          last.title = last.detail.replace(/\s+/g, ' ').slice(0, 60)
+            + (last.detail.length > 60 ? '…' : '')
+        } else {
+          out.push({ id: `ct-${i}`, seq: i, kind: 'think', status: 'done',
+                     title: text.replace(/\s+/g, ' ').slice(0, 60), detail: text })
+        }
+        break
+      }
+      case 'heartbeat': {
+        // 心跳只更新"还活着"的状态，不该每 3 秒堆一行
+        const last = out[out.length - 1]
+        const label = PHASE_LABEL[String(op.phase)] ?? '正在处理'
+        const elapsed = num(op.elapsed_ms)
+        if (last?.kind === 'lifecycle' && last.status === 'running') {
+          last.title = label
+          last.meta = elapsed ? formatDuration(elapsed) : last.meta
+        } else {
+          out.push({ id: `hb-${i}`, seq: i, kind: 'lifecycle', status: 'running',
+                     title: label, meta: elapsed ? formatDuration(elapsed) : undefined })
+        }
+        break
+      }
+      case 'plan':
+        out.push({ id: `cp-${i}`, seq: i, kind: 'note', status: 'done',
+                   title: String(op.summary ?? '想好了怎么做') })
+        break
+      case 'add_node':
+        nodeCount += 1
+        out.push({
+          id: `cn-${i}`, seq: i, kind: 'node', status: 'done',
+          title: String(op.node?.data?.label || op.node?.label || op.node?.id || '新步骤'),
+          meta: String(op.node?.type ?? ''),
+        })
+        break
+      case 'update_node':
+        out.push({ id: `cu-${i}`, seq: i, kind: 'note', status: 'done',
+                   title: `调整了「${op.id}」` })
+        break
+      case 'remove_node':
+        out.push({ id: `cr-${i}`, seq: i, kind: 'note', status: 'done',
+                   title: `去掉了「${op.id}」` })
+        break
+      // 连线不单独成行：用户关心有哪些步骤，不关心箭头
+      case 'add_edge':
+      case 'remove_edge':
+        break
+      case 'done':
+        out.push({ id: `cd-${i}`, seq: i, kind: 'lifecycle', status: 'done',
+                   title: `流程搭好了，共 ${nodeCount} 步`,
+                   detail: String(op.explanation ?? '') || undefined })
+        break
+      case 'error':
+        out.push({ id: `ce-${i}`, seq: i, kind: 'error', level: 'error', status: 'failed',
+                   title: String(op.message ?? '生成失败') })
+        break
+      default:
+        break
+    }
+  }
+  return out
+}
diff --git a/scripts/check-decode.mjs b/scripts/check-decode.mjs
new file mode 100644
index 00000000..300ec0e0
--- /dev/null
+++ b/scripts/check-decode.mjs
@@ -0,0 +1,116 @@
+// 事件解码器的回归检查。
+//
+// 用**真实运行**导出的事件跑（fixtures.json 取自 data/agentlab.db），不是构造的样本：
+// 这里要守的恰恰是真实事件的脏细节——审批会重放、成对事件字段不齐、
+// 空 preview、节点没起过名字。构造的样本永远不会长成那样。
+//
+// 转译借 vite dev server（它 serve 的就是应用实际运行的那份），所以跑之前
+// 前端得起着：./scripts/dev.sh
+import { readFileSync } from 'node:fs'
+
+const WEB = process.env.AGENTLAB_WEB ?? 'http://localhost:5273'
+const root = new URL('..', import.meta.url).pathname
+const fixtures = JSON.parse(
+  readFileSync(`${root}frontend/src/run/__tests__/fixtures.json`, 'utf8'))
+
+const res = await fetch(`${WEB}/src/run/decode.ts?t=${Date.now()}`)
+if (!res.ok) {
+  console.error(`✗ 拿不到转译结果（${WEB}）——前端没起？先跑 ./scripts/dev.sh`)
+  process.exit(1)
+}
+const mod = await import(
+  'data:text/javascript;base64,' + Buffer.from(await res.text()).toString('base64'))
+
+let failed = 0
+const check = (name, cond, detail = '') => {
+  console.log(`  ${cond ? '✓' : '✗'} ${name}${detail ? ` — ${detail}` : ''}`)
+  if (!cond) failed++
+}
+const flatten = (steps) =>
+  steps.flatMap((s) => [s, ...(s.children ? flatten(s.children) : [])])
+
+console.log('=== 数据库查询 ===')
+{
+  const steps = mod.decodeRun(fixtures.db)
+  const all = flatten(steps)
+  const q = all.find((s) => s.kind === 'query')
+  check('查询步骤说人话，不是 db_query__warehouse(...)',
+    !!q && q.title.includes('查询数据') && !q.title.includes('db_query__'), q?.title)
+  check('SQL 原文留在详情里可展开', !!q?.detail?.includes('SELECT'))
+  check('工具调用挂在节点下，不是平铺',
+    steps.some((s) => s.kind === 'node' && s.children?.some((c) => c.kind === 'query')))
+}
+
+console.log('\n=== 人工审批（含重放）===')
+{
+  const steps = mod.decodeRun(fixtures.human)
+  const all = flatten(steps)
+  // 一次审批会产生 human.requested ×2（重放）+ run.interrupted，界面上只该有一条
+  check('"等你确认"只出现一次',
+    all.filter((s) => s.kind === 'human' && s.title.includes('确认')).length === 1,
+    `${all.filter((s) => s.kind === 'human' && s.title.includes('确认')).length} 条`)
+  check('"继续执行"只出现一次',
+    steps.filter((s) => s.title === '继续执行').length === 1,
+    `${steps.filter((s) => s.title === '继续执行').length} 条`)
+  // 节点重放不该让同一个节点在时间线上出现两遍
+  const nodeIds = steps.filter((s) => s.kind === 'node').map((s) => s.nodeId)
+  check('节点不因重放而重复', new Set(nodeIds).size === nodeIds.length,
+    nodeIds.join(','))
+  check('审批结果有交代', all.some((s) => s.title.includes('放行') || s.title.includes('驳回')))
+}
+
+console.log('\n=== 出具判定 ===')
+{
+  const steps = mod.decodeRun(fixtures.issue)
+  check('出具档位翻译成中文', steps.some((s) => s.kind === 'issuance' && s.title.includes('出具')),
+    steps.find((s) => s.kind === 'issuance')?.title)
+}
+
+console.log('\n=== 失败运行 ===')
+{
+  const steps = mod.decodeRun(fixtures.failed)
+  const all = flatten(steps)
+  const errs = all.filter((s) => s.level === 'error')
+  check('错误只说一遍（节点失败与 run.failed 是同一件事）',
+    new Set(errs.map((s) => s.detail ?? s.title)).size === errs.length,
+    `${errs.length} 条错误`)
+  check('开头那条不会一直转圈',
+    !steps.some((s) => s.kind === 'lifecycle' && s.status === 'running'))
+}
+
+console.log('\n=== 通用规则 ===')
+{
+  const all = Object.values(fixtures).flatMap((evts) => flatten(mod.decodeRun(evts)))
+  check('没有裸节点 id 当标题',
+    !all.some((s) => s.kind === 'node' && /^(in|out|h|m|t\d|n\d)$/.test(s.title)),
+    all.filter((s) => s.kind === 'node' && /^(in|out|h|m)$/.test(s.title)).map(s=>s.title).join(','))
+  check('没有空详情', !all.some((s) => s.detail === '{}' || s.detail === ''))
+  check('没有原始事件名泄漏到标题',
+    !all.some((s) => /^(node|run|llm|tool)\./.test(s.title)),
+    all.filter((s) => /^(node|run|llm|tool)\./.test(s.title)).map(s=>s.title).join(','))
+  check('耗时是人类单位', all.filter((s) => s.meta).every((s) => /(\d+ms|\d+\.\d+s|\d+m\d+s|\d+ 行)/.test(s.meta)))
+}
+
+console.log('\n=== Copilot 操作流 ===')
+{
+  const ops = [
+    { op: 'heartbeat', phase: 'planning', elapsed_ms: 3000 },
+    { op: 'heartbeat', phase: 'planning', elapsed_ms: 6000 },
+    { op: 'thinking', delta: '用户要查销量。' },
+    { op: 'thinking', delta: '先看表结构。' },
+    { op: 'plan', summary: '三步：取数→汇总→出具' },
+    { op: 'add_node', node: { id: 'q', type: 'tool', data: { label: '取数' } } },
+    { op: 'add_edge', edge: { source: 'a', target: 'q' } },
+    { op: 'done', explanation: '完成' },
+  ]
+  const steps = mod.decodeCopilot(ops)
+  check('心跳不堆行', steps.filter((s) => s.kind === 'lifecycle' && s.status === 'running').length === 1)
+  check('连续思考并成一条', steps.filter((s) => s.kind === 'think').length === 1,
+    `${steps.filter((s) => s.kind === 'think').length} 条`)
+  check('思考全文保留在详情', steps.find((s) => s.kind === 'think')?.detail?.includes('先看表结构'))
+  check('连线不单独成行', !steps.some((s) => s.title.includes('edge')))
+  check('节点用标签不用 id', steps.some((s) => s.kind === 'node' && s.title === '取数'))
+}
+
+console.log(failed ? `\n✗ ${failed} 项未通过` : '\n✓ 解码器全部通过')
+process.exit(failed ? 1 : 0)