dsh-livebench-panel 0.2.11 → 0.2.13

package/README.md

@@ -44,11 +44,11 @@ cd livebench
 - **标签页位置**：`conversation.view` 槽位 `id: "livebench"`、`order: 20`（对话 = 0，轨迹 = 10，LiveBench 排最右）。
 - **模型下拉框**：读取 `$DSH_HOME/settings.yaml` 的 `llm-pi-ai.providers` **并合并内置 `deepseek-official` 路由**（`@deepseek-ai/dsh-llm-deepseek` 注册的 DeepSeek-V4-Flash / V4-Pro / V4-Flash-Vision-Exp，默认 `https://api.deepseek.com` + `DEEPSEEK_API_KEY`，用户 `llm-deepseek` 设置节可覆盖），与 harness 模型选择同源，展示**全部 provider / 全部模型**；`openai-completions` 且配置了 `baseURL` 的 provider 会被标记为可直连 LiveBench（`--api-base` 路由）。
 - **推理强度下拉框**（模型右侧）：来自模型配置的 `reasoningEfforts` 映射（如 gpt-5.6 系 off/low/medium/high/xhigh/max，DeepSeek off/low/high/max）。选定后：
-  - 插件向 `livebench/model/model_configs/dsh_panel_generated.yaml` 写入一条模型配置，经 LiveBench 的 `api_kwargs.default.reasoning_effort` 透传给 API（`off` 表示不透传、由后端走默认）；
+  - 插件向 `livebench/model/model_configs/dsh_panel_generated__<display-name>.yaml` 写入一条模型配置（每个模型一个文件，避免并发写同名文件互相覆盖），经 LiveBench 的 `api_kwargs.default.reasoning_effort` 透传给 API（`off` 表示不透传、由后端走默认）；
   - 强度编码进 display-name（如 `code-gpt__gpt-5.6-sol@high`），**不同强度在成绩表中是独立条目**，可直接对比。
 - **参数下拉框**：题集 release（LiveBench 全部 releases）、分类（coding/math/reasoning/language/data_analysis/instruction_following）、任务（随分类联动）、题目序号范围、max-tokens。
-- **运行控制**：开始 / 停止 / 刷新；同一时间只允许一个评测；实时滚动日志（每 2.5s 轮询）。
-- **成绩表**：直接读取 `data/live_bench/**/model_judgment/ground_truth_judgment.jsonl` 计算 模型 × 任务 平均分（分数 = 正确率 ×100，附已判/总题数），无需等 LiveBench 自己出榜。
+- **运行控制**：开始 / 停止 / 刷新；最多 9 个模型并发评测（每个模型同时只跑 1 次）；实时滚动日志（每 2.5s 轮询）；「刷新」会清掉已结束的运行日志。
+- **成绩表**：直接读取 `data/live_bench/**/model_judgment/ground_truth_judgment.jsonl` 计算 模型 × 任务 平均分（分数 = 正确率 ×100，附已判/总题数），无需等 LiveBench 自己出榜。支持按模型名/时间排序、拖 ⠿ 自定义行序、勾选或单行按钮删除成绩（删除会同时清掉该模型的答案与判分行；只有空壳文件的“幽灵行”不会再被扫描出来，删除结果稳定持久）。
 
 ## 依赖
 
@@ -65,6 +65,9 @@ cd livebench
 | `/dsh-livebench-panel/api/status` | GET | 运行状态 + 日志尾部 |
 | `/dsh-livebench-panel/api/stop` | POST | 终止当前评测 |
 | `/dsh-livebench-panel/api/results` | GET | 汇总成绩行 |
+| `/dsh-livebench-panel/api/delete` | POST | 删除指定 模型×分类×任务 的答案与判分记录（正文 `{rows:[{model,category,task}]}`；正在评测的模型会被跳过并在 `skipped` 中返回） |
+| `/dsh-livebench-panel/api/clear` | POST | 清掉已结束的运行日志/记录 |
+| `/dsh-livebench-panel/api/home` | POST | 保存 LiveBench 检出路径 |
 
 ## 安装（本地开发，遵循 ~/.dsh/AGENTS.md）
 

package/lib/client.js

@@ -8,7 +8,7 @@ window.__ModuleLoader__.load({
 		const { createElement: h, useCallback, useEffect, useMemo, useRef, useState } = react;
 
 		//#region dsh-css:livebench-panel.css
-		const css = `.dlb_root{width:100%;max-width:1100px;margin:0 auto;padding:16px 20px 32px;color:var(--dsw-alias-label-primary);display:flex;flex-direction:column;gap:14px}
+		const css = `.dlb_root{width:100%;max-width:1400px;margin:0 auto;padding:16px 20px 32px;color:var(--dsw-alias-label-primary);display:flex;flex-direction:column;gap:14px}
 .dlb_head{display:flex;align-items:baseline;gap:10px}
 .dlb_head h2{margin:0;font-size:16px;font-weight:600}
 .dlb_sub{color:var(--dsw-alias-label-tertiary);font-size:12px}
@@ -21,6 +21,8 @@ window.__ModuleLoader__.load({
 .dlb_select,.dlb_input{height:34px;border:1px solid var(--dsw-alias-border-l2);background:var(--dsw-alias-bg-layer-1);color:var(--dsw-alias-label-primary);font:inherit;font-size:13px;border-radius:8px;padding:0 8px;outline:none;min-width:0}
 .dlb_select:focus-visible,.dlb_input:focus-visible{border-color:var(--dsw-alias-state-business-primary)}
 .dlb_row{display:flex;align-items:center;gap:8px;flex-wrap:wrap}
+.dlb_row{align-items:flex-start}
+.dlb_cardHead{display:flex;flex-direction:column;gap:2px;min-width:0;flex:1 1 auto}
 .dlb_btn{appearance:none;font:inherit;cursor:pointer;border-radius:8px;padding:6px 14px;font-size:13px;border:1px solid transparent;background:var(--dsw-alias-state-business-primary);color:var(--dsw-alias-label-primary)}
 .dlb_btn:disabled{cursor:default;opacity:.55}
 .dlb_btnGhost{background:transparent;border-color:var(--dsw-alias-border-l2);color:var(--dsw-alias-label-secondary)}
@@ -29,9 +31,31 @@ window.__ModuleLoader__.load({
 .dlb_badge[data-ok="1"]{color:var(--dsw-alias-state-success-primary);border-color:color-mix(in srgb,var(--dsw-alias-state-success-primary) 40%,transparent)}
 .dlb_badge[data-ok="0"]{color:var(--dsw-alias-state-error-primary);border-color:color-mix(in srgb,var(--dsw-alias-state-error-primary) 40%,transparent)}
 .dlb_log{background:var(--dsw-alias-bg-layer-1);border:1px solid var(--dsw-alias-border-l2);border-radius:8px;padding:10px;font:12px/1.5 ui-monospace,SFMono-Regular,Menlo,Consolas,monospace;max-height:260px;overflow:auto;white-space:pre-wrap;word-break:break-all;color:var(--dsw-alias-label-secondary)}
-.dlb_table{width:100%;border-collapse:collapse;font-size:12.5px}
-.dlb_table th,.dlb_table td{text-align:left;padding:6px 10px;border-bottom:1px solid var(--dsw-alias-border-l2)}
-.dlb_table th{color:var(--dsw-alias-label-tertiary);font-weight:500}
+.dlb_tableWrap{overflow:auto;max-height:min(64vh,620px);border:1px solid var(--dsw-alias-border-l2);border-radius:10px;background:var(--dsw-alias-bg-layer-2)}
+.dlb_table{table-layout:fixed;border-collapse:separate;border-spacing:0;font-size:11.5px}
+.dlb_table th,.dlb_table td{padding:4px 5px;border-bottom:1px solid var(--dsw-alias-border-l2);white-space:nowrap;vertical-align:middle;text-align:left;overflow:hidden;text-overflow:ellipsis}
+.dlb_table td{height:34px}
+.dlb_table thead th{color:var(--dsw-alias-label-tertiary);font-weight:500;background:var(--dsw-alias-bg-layer-2);position:sticky;top:0;z-index:3}
+.dlb_table thead tr.dlb_catRow th{top:0;height:20px;padding:0 6px;font-size:9.5px;letter-spacing:.04em;text-transform:uppercase;background:var(--dsw-alias-bg-layer-1);cursor:default}
+.dlb_table thead tr.dlb_thRow th{top:20px;height:26px;font-size:11px;box-shadow:inset 0 -1px 0 var(--dsw-alias-border-l2)}
+.dlb_table tbody tr:nth-child(even) td{background:color-mix(in srgb,var(--dsw-alias-fill-l1,rgba(255,255,255,.04)) 55%,transparent)}
+.dlb_table tbody tr:hover td{background:color-mix(in srgb,var(--dsw-alias-state-business-primary) 9%,transparent)}
+.dlb_stick{position:sticky;background:var(--dsw-alias-bg-layer-2);z-index:2}
+.dlb_table thead tr.dlb_catRow th.dlb_stick{z-index:5}
+.dlb_table thead tr.dlb_thRow th.dlb_stick{z-index:5}
+.dlb_table tbody tr:nth-child(even) .dlb_stick{background:color-mix(in srgb,var(--dsw-alias-bg-layer-2) 90%,var(--dsw-alias-fill-l1,rgba(255,255,255,.05)))}
+.dlb_table tbody tr:hover .dlb_stick{background:color-mix(in srgb,var(--dsw-alias-bg-layer-2) 88%,var(--dsw-alias-state-business-primary))}
+.dlb_grpStart{border-left:1px solid var(--dsw-alias-border-l2) !important}
+.dlb_cellScore{text-align:center !important;font-variant-numeric:tabular-nums;font-weight:600;padding-left:2px !important;padding-right:2px !important}
+.dlb_cellSub{font-size:9px;font-weight:400;color:var(--dsw-alias-label-tertiary);font-variant-numeric:tabular-nums}
+.dlb_thTask{font-size:10.5px !important;text-align:center !important;cursor:default}
+.dlb_modelCell{padding-left:8px !important}
+.dlb_modelName{font-weight:500;overflow:hidden;text-overflow:ellipsis}
+.dlb_dim{opacity:.35;font-weight:400}
+.dlb_timeCell{font-size:10.5px;color:var(--dsw-alias-label-secondary);font-variant-numeric:tabular-nums;text-align:center !important}
+.dlb_sHigh{color:var(--dsw-alias-state-success-primary)}
+.dlb_sMid{color:var(--dsw-alias-state-warning-primary,#d08700)}
+.dlb_sLow{color:var(--dsw-alias-state-error-primary)}
 .dlb_score{font-variant-numeric:tabular-nums;font-weight:600}
 .dlb_error{color:var(--dsw-alias-state-error-primary);font-size:12px;margin:0}
 .dlb_hint{color:var(--dsw-alias-label-tertiary);font-size:11.5px;margin:0;line-height:1.5}
@@ -309,20 +333,41 @@ window.__ModuleLoader__.load({
 					};
 
 					// ---- 成绩矩阵：model 为行标识、category/task 为列标识 ----
-					// time 列 = 该模型评测开始时间（行内最早判分时间）
+					// time 列 = 该模型这次评测的开始时间：
+					//   1) 评测元数据 startedAt；2) displayName 尾部的运行戳 __rYYYYMMDD-HHMMSS；3) 最早的判分时间
+					const runStampSeconds = (modelName) => {
+						const m = /__r(\d{4})(\d{2})(\d{2})-(\d{2})(\d{2})(\d{2})$/.exec(String(modelName));
+						if (!m) return null;
+						const t = new Date(Number(m[1]), Number(m[2]) - 1, Number(m[3]), Number(m[4]), Number(m[5]), Number(m[6])).getTime();
+						return Number.isFinite(t) ? Math.floor(t / 1000) : null;
+					};
 					const taskKey = (category, task) => `${category}/${task}`;
 					const matrix = useMemo(() => {
 						if (!results) return null;
 						const models = [...new Set(results.rows.map((r) => r.model))].sort();
-						const tasks = [...new Set(results.rows.map((r) => taskKey(r.category, r.task)))].sort();
+						const taskSet = new Set(results.rows.map((r) => taskKey(r.category, r.task)));
+						const tasks = [...taskSet].sort();
+						// 按 category 分组（同一 category 的 task 相邻），列上以分隔线分组
+						const groups = [];
+						const byCat = new Map();
+						for (const key of tasks) {
+							const category = key.split("/")[0];
+							if (!byCat.has(category)) { byCat.set(category, []); groups.push({ category, tasks: byCat.get(category) }); }
+							byCat.get(category).push(key);
+						}
+						groups.sort((a, b) => a.category.localeCompare(b.category));
+						const groupStart = new Set(groups.map((g) => g.tasks[0]));
 						const cells = new Map();
 						const startTimes = new Map();
 						for (const row of results.rows) {
 							cells.set(`${row.model}__@__${taskKey(row.category, row.task)}`, row);
+							const fromMeta = row.runStartedAt ? Math.floor(Date.parse(row.runStartedAt) / 1000) : null;
+							const candidate = (Number.isFinite(fromMeta) ? fromMeta : null) ?? runStampSeconds(row.model) ?? row.time ?? null;
 							const prev = startTimes.get(row.model);
-							if (row.time && (prev === undefined || row.time < prev)) startTimes.set(row.model, row.time);
+							if (candidate && (prev === undefined || candidate < prev)) startTimes.set(row.model, candidate);
 						}
-						return { models, tasks, cells, startTimes };
+						const taskTotals = (results && results.taskTotals) || {};
+						return { models, tasks, groups, groupStart, cells, startTimes, taskTotals };
 					}, [results]);
 					const cellOf = (model, key) => matrix.cells.get(`${model}__@__${key}`);
 					// 排序：sortBy=null 时用拖拽自定义顺序；点击表头在 正序/倒序 间切换
@@ -392,17 +437,29 @@ window.__ModuleLoader__.load({
 						}
 						if (serverRows.length === 0) return;
 						if (!window.confirm(`确认删除所选 ${modelsToDelete.length} 个模型的全部评测成绩（共 ${serverRows.length} 条记录，含答案）？不可恢复。`)) return;
-						await api("/delete", {
+						const delRes = await api("/delete", {
 							method: "POST",
 							headers: { "content-type": "application/json" },
 							body: JSON.stringify({ rows: serverRows }),
 						});
 						setPickedModels(new Set());
+						if (delRes && delRes.skipped > 0) {
+							window.alert(`${delRes.skipped} 个模型正在评测中，已跳过未删除；请等评测结束后再删。`);
+						}
 						persistModelOrder(modelOrder.filter((model) => !modelsToDelete.includes(model)));
 						await loadResults();
 					};
 
-					const fmtTime = (t) => {
+					// 显示用短名：去掉运行时间戳后缀（运行时间单列显示在模型名下方）
+			const shortModelName = (m) => String(m).replace(/__r[0-9]{8}-[0-9]{6}$/, "");
+			const fmtTimeShort = (t) => {
+				if (!t) return "—";
+				const dt = new Date(t * 1000);
+				const pad = (n) => String(n).padStart(2, "0");
+				return `${pad(dt.getMonth() + 1)}-${pad(dt.getDate())} ${pad(dt.getHours())}:${pad(dt.getMinutes())}`;
+			};
+
+			const fmtTime = (t) => {
 						if (!t) return "—";
 						const dt = new Date(t * 1000);
 						const pad = (n) => String(n).padStart(2, "0");
@@ -542,74 +599,125 @@ window.__ModuleLoader__.load({
 						),
 						sortedModels.length > 0 && h("div", { className: c("card") },
 							h("div", { className: c("row") },
-								h("span", { className: c("label") }, "评测成绩（model 为行；分数 = 平均分 ×100；行首可拖动排序，顺序保存在本机）"),
+								h("div", { className: c("cardHead") },
+									h("span", { className: c("label") }, `评测成绩（行 = 模型，列 = 任务）· ${sortedModels.length} 个模型 × ${matrix.tasks.length} 个任务`),
+									h("span", { className: c("hint") }, "分数 = 该任务平均分 ×100；括号 (-没做或没做完/本次设定/该任务总题数)；× = 未执行该任务；— = 无有效判分（如全部访问失败，不计入正确率）；颜色 ≥75 绿 / 40–75 橙 / <40 红；拖 ⠿ 换行序（存在本机），点表头按模型名或时间排序。"),
+								),
 								h("button", {
 									className: c("btn") + " " + c("btnDanger"),
 									disabled: pickedModels.size === 0,
 									onClick: () => deleteModels([...pickedModels]),
 								}, `删除所选（${pickedModels.size}）`),
 							),
-							h("table", { className: c("table") },
-								h("thead", null, h("tr", null,
-									h("th", { style: { width: "26px" } }, "⠿"),
-									h("th", { style: { width: "30px" } },
-										h("input", { type: "checkbox", checked: pickedModels.size > 0 && pickedModels.size === sortedModels.length,
-											onChange: (event) => setPickedModels(event.target.checked ? new Set(sortedModels) : new Set()) })),
-									h("th", {
-										style: { cursor: "pointer", userSelect: "none" },
-										title: "点击按模型名排序（再次点击切换正序/倒序）",
-										onClick: () => toggleSort("model"),
-									}, "model" + sortMark("model")),
-									matrix.tasks.map((taskKeyCol) => {
-										const [cat, task] = taskKeyCol.split("/");
-										return h("th", { key: taskKeyCol, title: `${cat} / ${task}` }, task);
-									}),
-									h("th", null, "删除"),
-									h("th", {
-										style: { cursor: "pointer", userSelect: "none" },
-										title: "点击按时间排序（再次点击切换正序/倒序）",
-										onClick: () => toggleSort("time"),
-									}, "time" + sortMark("time")),
-								)),
-								h("tbody", null,
-									sortedModels.map((model) => {
-										const checked = pickedModels.has(model);
-										const startTime = matrix.startTimes.get(model);
-										return h("tr", {
-											key: model,
-											draggable: true,
-											onDragStart: () => { dragKey.current = model; },
-											onDragOver: (event) => event.preventDefault(),
-											onDrop: () => onRowDrop(model),
-										},
-											h("td", { className: c("drag"), title: "拖动排序" }, "⠿"),
-											h("td", null, h("input", { type: "checkbox", checked, onChange: () => togglePicked(model) })),
-											h("td", null, model),
+							h("div", { className: c("tableWrap") },
+								h("table", {
+									className: c("table"),
+									style: { width: `${20 + 26 + 190 + matrix.tasks.length * 74 + 30 + 86}px` },
+								},
+									h("colgroup", null,
+										h("col", { style: { width: "20px" } }),
+										h("col", { style: { width: "26px" } }),
+										h("col", { style: { width: "190px" } }),
+										matrix.tasks.map((taskKeyCol) => h("col", { key: taskKeyCol, style: { width: "74px" } })),
+										h("col", { style: { width: "30px" } }),
+										h("col", { style: { width: "86px" } }),
+									),
+									h("thead", null,
+										// 第一行：category 分组（同一 category 的 task 列相邻，组间有分隔线）
+										h("tr", { className: c("catRow") },
+											h("th", { className: c("stick"), style: { left: 0 }, colSpan: 3 }, "分类 ↓"),
+											matrix.groups.map((group) => h("th", {
+												key: group.category,
+												className: c("grpStart"),
+												colSpan: group.tasks.length,
+												style: { textAlign: "center" },
+												title: `${group.category}（${group.tasks.length} 个任务）`,
+											}, `${group.category} · ${group.tasks.length}`)),
+											h("th", null, ""),
+											h("th", null, ""),
+										),
+										// 第二行：task 列头 + model / time 排序
+										h("tr", { className: c("thRow") },
+											h("th", { className: c("stick"), style: { left: 0 } }, "⠿"),
+											h("th", { className: c("stick"), style: { left: "20px" } },
+												h("input", { type: "checkbox", checked: pickedModels.size > 0 && pickedModels.size === sortedModels.length,
+													onChange: (event) => setPickedModels(event.target.checked ? new Set(sortedModels) : new Set()) })),
+											h("th", {
+												className: c("stick"),
+												style: { left: "46px", cursor: "pointer", userSelect: "none", paddingLeft: "8px" },
+												title: "点击按模型名排序（再次点击切换正序/倒序）",
+												onClick: () => toggleSort("model"),
+											}, "模型 / 运行时间" + sortMark("model")),
 											matrix.tasks.map((taskKeyCol) => {
-												const row = matrix.cells.get(`${model}__@__${taskKeyCol}`);
-												let cellText = "×";
-												let title = "该模型未执行此任务";
-												if (row) {
-													if (row.answered > 0 && row.errors >= row.answered) {
-														cellText = `E (${row.errors})`;
-														title = "该任务全部回答均 API 失败（$ERROR$），计 0 分";
-													} else if (row.judged === 0) {
-														cellText = "未判分";
-														title = "有答案但尚未判分";
-													} else {
-														cellText = `${row.score.toFixed(1)} (${row.judged}/${row.total})`;
-														if (row.errors > 0) title += ` · 其中 ${row.errors} 题 API 失败计 0 分`;
-													}
-												}
-												return h("td", { key: taskKeyCol, className: c("score"), title }, cellText);
+												const [cat, task] = taskKeyCol.split("/");
+												return h("th", {
+													key: taskKeyCol,
+													className: c("thTask") + (matrix.groupStart.has(taskKeyCol) ? " " + c("grpStart") : ""),
+													title: `${cat} / ${task}`,
+												}, task);
 											}),
-											h("td", null, h("button", {
-												className: c("btnGhost") + " " + c("btn"), title: "删除该模型的所有评测成绩",
-												onClick: () => deleteModels([model]),
-											}, "删除")),
-											h("td", null, fmtTime(startTime)),
-										);
-									})),
+											h("th", { title: "删除该模型的这一行成绩" }, "✕"),
+											h("th", {
+												style: { cursor: "pointer", userSelect: "none" },
+												title: "点击按运行时间排序（再次点击切换正序/倒序）",
+												onClick: () => toggleSort("time"),
+											}, "time" + sortMark("time")),
+										),
+									),
+									h("tbody", null,
+										sortedModels.map((model) => {
+											const checked = pickedModels.has(model);
+											const startTime = matrix.startTimes.get(model);
+											return h("tr", {
+												key: model,
+												draggable: true,
+												onDragStart: () => { dragKey.current = model; },
+												onDragOver: (event) => event.preventDefault(),
+												onDrop: () => onRowDrop(model),
+											},
+												h("td", { className: c("stick") + " " + c("drag"), style: { left: 0 }, title: "按住拖动排序" }, "⠿"),
+												h("td", { className: c("stick"), style: { left: "20px" } },
+													h("input", { type: "checkbox", checked, onChange: () => togglePicked(model) })),
+												h("td", { className: c("stick") + " " + c("modelCell"), style: { left: "46px" }, title: model },
+													h("div", { className: c("modelName") }, shortModelName(model)),
+													h("div", { className: c("cellSub") }, fmtTimeShort(startTime))),
+												matrix.tasks.map((taskKeyCol) => {
+													const row = matrix.cells.get(`${model}__@__${taskKeyCol}`);
+													// 单元格：上方正确率，下方 (-没做/设定/总)。
+													// 没做/没做完（访问失败、网络中断）不计入正确率分母。
+													const totalCol = row ? row.total : (matrix.taskTotals[taskKeyCol] ?? 0);
+												const sub = `(-${row ? (row.notDone ?? 0) : totalCol}/${row ? (row.configured ?? 0) : totalCol}/${totalCol})`;
+													const cls = c("cellScore") + (matrix.groupStart.has(taskKeyCol) ? " " + c("grpStart") : "");
+													if (!row) {
+														return h("td", { key: taskKeyCol, className: cls, title: `未执行该任务：${sub}` },
+															h("div", { className: c("dim") }, "×"),
+															h("div", { className: c("cellSub") }, sub));
+													}
+													const allFailed = row.answered > 0 && row.errors >= row.answered;
+													const title = allFailed
+														? `该任务 ${row.errors} 题全部访问失败（未做完），不计入正确率 ${sub}`
+														: `做完 ${row.done ?? 0} 题，没做/失败 ${row.notDone ?? 0} 题 ${sub}` +
+															(row.errors > 0 ? `（其中 ${row.errors} 题 API 失败）` : "");
+													if (allFailed || row.judged === 0) {
+														return h("td", { key: taskKeyCol, className: cls, title },
+															h("div", { className: c("dim") }, "—"),
+															h("div", { className: c("cellSub") }, sub));
+													}
+													const scoreCls = row.score >= 75 ? c("sHigh") : row.score >= 40 ? c("sMid") : c("sLow");
+													return h("td", { key: taskKeyCol, className: cls, title },
+														h("div", { className: scoreCls }, row.score.toFixed(1)),
+														h("div", { className: c("cellSub") }, sub));
+												}),
+												h("td", { style: { textAlign: "center" } }, h("button", {
+													className: c("btnGhost") + " " + c("btn"), style: { padding: "1px 7px", fontSize: "11px", lineHeight: "16px" },
+													title: "删除该模型的所有评测成绩",
+													onClick: () => deleteModels([model]),
+												}, "✕")),
+												h("td", { className: c("timeCell") }, fmtTimeShort(startTime)),
+											);
+										}),
+									),
+								),
 							),
 						),
 						config !== null && config.available === false && h("div", { className: c("card") + " " + c("notice") },
@@ -658,7 +766,7 @@ window.__ModuleLoader__.load({
 								"提示：回答为 $ERROR$ 表示该题 API 调用失败（网络/鉴权/参数问题）计 0 分，可在上方日志区查看具体错误；zebra_puzzle（逻辑谜题）是公认最难的任务，低分属正常现象。"),
 						),
 						h("p", { className: c("hint") },
-							"评分读取 LiveBench 的 ground_truth_judgment.jsonl；正式榜单可用 release 2024-11-25。多选题集请在「分类/任务」中缩小范围，避免长时间运行。"),
+							"格式说明：单元格上方为正确率（= 做对 / 做完），下方括号 (-没做/设定/总)：没做或没做完的题（访问失败、网络中断、未跑到）不计入正确率。× 表示该模型完全没跑过该任务。正式榜单可用 release 2024-11-25。"),
 					);
 				}
 		/**

package/lib/index.js

@@ -25,7 +25,7 @@
  * @module dsh-livebench-panel
  */
 import { spawn } from "node:child_process";
-import { readdirSync, readFileSync, existsSync, statSync, writeFileSync, mkdirSync } from "node:fs";
+import { readdirSync, readFileSync, existsSync, statSync, writeFileSync, mkdirSync, unlinkSync } from "node:fs";
 import { createRequire } from "node:module";
 import { delimiter, dirname, join } from "node:path";
 import { homedir } from "node:os";
@@ -307,6 +307,22 @@ function readDirSafe(dir) {
 	}
 }
 
+/** 列出目录下的【文件】（可按扩展名过滤）——注意 readDirSafe 只列目录。 */
+function readFilesSafe(dir, ext) {
+	try {
+		return readdirSync(dir).filter((f) => {
+			try {
+				if (!statSync(join(dir, f)).isFile()) return false;
+				return ext === undefined || f.endsWith(ext);
+			} catch {
+				return false;
+			}
+		});
+	} catch {
+		return [];
+	}
+}
+
 function readdirNames(dir) {
 	try {
 		return readdirSync(dir).filter((entry) => {
@@ -506,11 +522,22 @@ async function readTextFileAsync(path) {
 async function computeResults(dataDir) {
 	const now = Date.now();
 	if (resultsCache.data !== null && now - resultsCache.at < 5000) return resultsCache.data;
-	const judged = new Map(); // `${model}${task}` -> {model, category, task, sum, n, time}
+	const judged = new Map(); // `${model}\u0000${task}` -> {model, category, task, sum, n, time}
 	const totals = new Map(); // task -> question count
-	const answers = new Map(); // `${model}${task}` -> {answered, errors}
+	const answers = new Map(); // `${model}\u0000${task}` -> {answered, errors}
 	const catByTask = new Map();
 	if (!existsSync(dataDir)) return { rows: [], taskTotals: {} };
+	// 读取每次评测的配置范围（displayName -> {benchNames, begin, end}）
+	const runMeta = new Map();
+	try {
+		const metaDir = join(dataDir, ".dsh_runs");
+		for (const f of readdirSync(metaDir)) {
+			if (!f.endsWith(".json")) continue;
+			try {
+				runMeta.set(f.slice(0, -5), JSON.parse(readFileSync(join(metaDir, f), "utf8")));
+			} catch { /* 单个元数据损坏忽略 */ }
+		}
+	} catch { /* 无元数据目录（历史数据） */ }
 	for (const category of readDirSafe(dataDir)) {
 		const catDir = join(dataDir, category);
 		for (const task of readDirSafe(catDir)) {
@@ -519,19 +546,20 @@ async function computeResults(dataDir) {
 			const qText = await readTextFileAsync(join(taskDir, "question.jsonl"));
 			totals.set(task, qText === null ? 0 : qText.split("\n").reduce((n, l) => n + (l.trim().length > 0 ? 1 : 0), 0));
 			const answerDir = join(taskDir, "model_answer");
-			for (const file of readDirSafe(answerDir)) {
-				if (!file.endsWith(".jsonl")) continue;
+			for (const file of readFilesSafe(answerDir, ".jsonl")) {
 				const model = file.slice(0, -".jsonl".length);
 				const text = await readTextFileAsync(join(answerDir, file));
 				if (text === null) continue;
-				const key = `${model}${task}`;
+				const key = `${model}\u0000${task}`;
 				const stat = answers.get(key) ?? { answered: 0, errors: 0 };
 				for (const line of text.split("\n")) {
 					if (line.trim().length === 0) continue;
 					stat.answered += 1;
 					if (line.includes('"$ERROR$"')) stat.errors += 1;
 				}
-				answers.set(key, stat);
+				// 空文件（尚未产出答案，或成绩已被删除只留下 0 字节壳）不计入结果：
+				// 否则删掉的记录会在下一次扫描时凭“文件名仍在”重新冒出来。
+				if (stat.answered > 0) answers.set(key, stat);
 			}
 			const jText = await readTextFileAsync(join(taskDir, "model_judgment", "ground_truth_judgment.jsonl"));
 			if (jText !== null) {
@@ -542,7 +570,7 @@ async function computeResults(dataDir) {
 					const model = typeof row.model === "string" ? row.model : null;
 					const score = typeof row.score === "number" ? row.score : Number(row.score);
 					if (model === null || !Number.isFinite(score) || score < 0) continue;
-					const key = `${model}${task}`;
+					const key = `${model}\u0000${task}`;
 					const entry = judged.get(key) ?? { model, category, task, sum: 0, n: 0, time: 0 };
 					entry.sum += score;
 					entry.n += 1;
@@ -559,21 +587,54 @@ async function computeResults(dataDir) {
 	const rows = [...keys].map((key) => {
 		const entry = judged.get(key);
 		const stat = answers.get(key) ?? { answered: 0, errors: 0 };
-		const [model, task] = key.split("");
+		const [model, task] = key.split("\u0000");
 		const category = entry?.category ?? catByTask.get(task) ?? "";
+		const total = totals.get(task) ?? 0;
+		const judgedCount = entry?.n ?? 0;
+		// 做完的题 = 实际产出的有效答案（尝试数 - 访问失败数）
+		const done = Math.max(0, stat.answered - stat.errors);
+		// 设定题数 = 本次评测设定的范围（end 含端点）；没有元数据（历史数据）时用该任务的题目总数兜底
+		let configured = total;
+		const meta = runMeta.get(model);
+		if (meta !== undefined) {
+			const inScope = (meta.benchNames || []).some((bn) => {
+				const parts = String(bn).split("/").filter((s) => s.length > 0);
+				if (parts.length >= 3) return parts[1] === category && parts[2] === task;
+				if (parts.length === 2) return parts[1] === category;
+				return true;
+			});
+			if (inScope) {
+				const hasRange = meta.begin !== null && meta.begin !== undefined && meta.end !== null && meta.end !== undefined;
+				if (hasRange) configured = Math.min(total, Number(meta.end) - Number(meta.begin) + 1);
+			}
+		}
+		// 实际尝试数比设定还多（断点重跑等）时，以实际为准
+		if (configured < stat.answered) configured = stat.answered;
+		const notDone = Math.max(0, configured - done);
+		// 正确率 = 做对的题 / 已判分的做完题（访问失败/未做的题不计入分母）
+		const judgedDone = Math.max(0, judgedCount - stat.errors);
+		const score = entry && judgedDone > 0 ? (entry.sum / judgedDone) * 100 : null;
 		return {
 			model,
 			category,
 			task,
-			score: entry && entry.n > 0 ? (entry.sum / entry.n) * 100 : null,
+			score,
 			judged: entry?.n ?? 0,
-			total: totals.get(task) ?? 0,
+			total,
 			time: entry && entry.time > 0 ? entry.time : null,
 			answered: stat.answered,
 			errors: stat.errors,
+			done,
+			configured,
+			notDone,
+			// 本次运行的开始时间（ISO）；旧数据没有元数据时为 null，前端退回解析 displayName 里的运行戳
+			runStartedAt: meta !== undefined && typeof meta.startedAt === "string" ? meta.startedAt : null,
 		};
 	});
-	const result = { rows, taskTotals: Object.fromEntries(totals) };
+	// taskTotals 同时给出裸任务名与 `category/task` 两种键，便于前端按列取题目总数
+	const taskTotals = Object.fromEntries(totals);
+	for (const [t, n] of totals) taskTotals[`${catByTask.get(t) ?? ""}/${t}`] = n;
+	const result = { rows, taskTotals };
 	resultsCache = { at: Date.now(), data: result };
 	return result;
 }
@@ -625,7 +686,18 @@ function filterJsonlByModel(path, field, model) {
 		if (match) removed += 1;
 		else kept.push(line);
 	}
-	if (removed > 0) writeFileSync(path, kept.length > 0 ? kept.join("\n") + "\n" : "", "utf8");
+	if (removed > 0) {
+		if (kept.length > 0) writeFileSync(path, kept.join("\n") + "\n", "utf8");
+		else {
+			// 全部清空时直接删掉文件，别留 0 字节壳让扫描把它当成一行成绩。
+			// （若文件被评测进程占用导致 unlink 失败，退化为写空文件。）
+			try { unlinkSync(path); }
+			catch { writeFileSync(path, "", "utf8"); }
+		}
+	} else if (kept.length === 0) {
+		// 本来就是 0 字节空壳（早期删除留下的）：一并清掉
+		try { unlinkSync(path); } catch { /* 被占用则保留 */ }
+	}
 	return removed;
 }
 
@@ -808,6 +880,19 @@ function apply(ctx) {
 			}
 		}
 
+		// 记录本次评测的配置范围：/results 用它计算每个任务的"设定题数"
+		// （未做/失败的题数 = 设定题数 - 实际产出的有效答案数）。
+		try {
+			const metaDir = join(layout.dataDir, ".dsh_runs");
+			if (!existsSync(metaDir)) mkdirSync(metaDir, { recursive: true });
+			writeFileSync(join(metaDir, displayName + ".json"), JSON.stringify({
+				benchNames: benchNames !== null ? benchNames : [benchParts.join("/")],
+				begin: body.begin ?? null,
+				end: body.end ?? null,
+				startedAt: new Date().toISOString(),
+			}), "utf8");
+		} catch { /* 元数据写入失败不影响评测 */ }
+
 		const child = spawn(layout.pythonExe, args, {
 			cwd: layout.livebenchDir,
 			env,
@@ -865,7 +950,7 @@ function apply(ctx) {
 								startRun(retryBody, record.autoRetries + 1).catch((e) => {
 									appendLog(record, `[自动补跑启动失败] ${e.message}`);
 								});
-							}, 15000);
+							}, [15000, 60000, 180000][Math.min(record.autoRetries, 2)]);
 						}
 					} catch (e) {
 						record.log.push(`[自动补跑检查失败] ${e.message}`);
@@ -1105,7 +1190,11 @@ function apply(ctx) {
 			const rows = Array.isArray(body.rows) ? body.rows.slice(0, 200) : [];
 			const layout = livebenchLayout();
 			const tasks = scanCategoryTasks(layout.dataDir);
+			// 正在跑的模型不删：文件还被评测进程持有，删了会损坏本次结果
+			const activeModels = new Set();
+			for (const r of runs.values()) if (r.exitCode === null) activeModels.add(r.displayName);
 			let removed = 0;
+			let skipped = 0;
 			for (const row of rows) {
 				if (!row || typeof row !== "object") continue;
 				const model = typeof row.model === "string" ? row.model : "";
@@ -1115,12 +1204,13 @@ function apply(ctx) {
 				if (!/^[A-Za-z0-9._@-]{1,160}$/.test(model)) continue;
 				if (!/^[A-Za-z0-9_]{1,80}$/.test(category) || !/^[A-Za-z0-9_]{1,80}$/.test(task)) continue;
 				if (!(tasks[category] && tasks[category][task])) continue;
+				if (activeModels.has(model)) { skipped += 1; continue; }
 				const taskDir = join(layout.dataDir, category, task);
 				removed += filterJsonlByModel(join(taskDir, "model_answer", `${model}.jsonl`), "model_id", model);
 				removed += filterJsonlByModel(join(taskDir, "model_judgment", "ground_truth_judgment.jsonl"), "model", model);
 			}
 			resultsCache = { at: 0, data: null };
-			sendJson(res, 200, { ok: true, removed });
+			sendJson(res, 200, { ok: true, removed, skipped });
 		},
 	}), `${name}: delete route`);
 }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "dsh-livebench-panel",
-  "version": "0.2.11",
+  "version": "0.2.13",
   "description": "DSH web plugin: a LiveBench tab in the Trajectory view (right of 对话/轨迹). Run LiveBench evaluations against every model configured in the DeepSeek Harness — pick provider/model, category, task, release and question range from dropdowns, watch progress, and read scores in place.",
   "license": "MIT",
   "type": "module",