dsh-eval-guard 0.2.2 → 0.3.0

package/lib/index.js

@@ -4,7 +4,7 @@ import fs from "node:fs";
 import path from "node:path";
 import os from "node:os";
 import { fileURLToPath } from "node:url";
-import { spawn } from "node:child_process";
+import { execFileSync, spawn } from "node:child_process";
 //#region src/config.ts
 /**
 * Plugin configuration. The Settings page writes this document through the
@@ -18,7 +18,13 @@ const Config = z.object({
 	timeLimitMinutes: z.number().default(90).volatile(),
 	forbiddenPaths: z.string().default("").volatile(),
 	toolBlacklistUseDefault: z.boolean().default(true).volatile(),
-	toolBlacklist: z.string().default("").volatile()
+	toolBlacklist: z.string().default("").volatile(),
+	cloneSource: z.string().default("github").volatile(),
+	customCloneUrl: z.string().default("").volatile(),
+	pipIndex: z.string().default("pypi").volatile(),
+	customPipUrl: z.string().default("").volatile(),
+	gitProxyUrl: z.string().default("").volatile(),
+	pipProxyUrl: z.string().default("").volatile()
 });
 /** Clamp and normalize a live configuration object. */
 function resolveConfig(value) {
@@ -29,7 +35,13 @@ function resolveConfig(value) {
 		timeLimitMinutes: Math.min(1440, Math.max(1, Math.round(minutes))),
 		forbiddenPaths: typeof value?.forbiddenPaths === "string" ? value.forbiddenPaths : "",
 		toolBlacklistUseDefault: value?.toolBlacklistUseDefault !== false,
-		toolBlacklist: typeof value?.toolBlacklist === "string" ? value.toolBlacklist : ""
+		toolBlacklist: typeof value?.toolBlacklist === "string" ? value.toolBlacklist : "",
+		cloneSource: typeof value?.cloneSource === "string" && value.cloneSource !== "" ? value.cloneSource : "github",
+		customCloneUrl: typeof value?.customCloneUrl === "string" ? value.customCloneUrl : "",
+		pipIndex: typeof value?.pipIndex === "string" && value.pipIndex !== "" ? value.pipIndex : "pypi",
+		customPipUrl: typeof value?.customPipUrl === "string" ? value.customPipUrl : "",
+		gitProxyUrl: typeof value?.gitProxyUrl === "string" ? value.gitProxyUrl : "",
+		pipProxyUrl: typeof value?.pipProxyUrl === "string" ? value.pipProxyUrl : ""
 	};
 }
 //#endregion
@@ -42,6 +54,10 @@ const PROMPT_ROUTE = "/api/eval-guard/prompt";
 const EVENTS_ROUTE = "/api/eval-guard/events";
 const REVEAL_ROUTE = "/api/eval-guard/reveal";
 const MANUAL_ROUTE = "/api/eval-guard/manual";
+const JOBS_ROUTE = "/api/eval-guard/jobs";
+const JOB_ROUTE = "/api/eval-guard/job";
+const JOB_CANCEL_ROUTE = "/api/eval-guard/job/cancel";
+const EVAL_ROUTE = "/api/eval-guard/evaluate";
 /** Supported task ids. */
 const TANK_TASK = "tank";
 const TANK_TASK_LABEL = "第 6 集 · 坦克大战";
@@ -52,6 +68,45 @@ const LAND_TASK_LABEL = "第 1 集 · 圈地";
 const BOMBERMAN_TASK = "bomberman";
 const BOMBERMAN_TASK_LABEL = "第 2 集 · 炸弹人";
 const CTF_TASK_LABEL = "第 3 集 · 夺旗";
+const COMMIT0_TASK = "commit0";
+const COMMIT0_TASK_LABEL = "Commit0 · 单库实现";
+const SWE_BENCH_PRO_TASK = "swe-bench-pro";
+const SWE_BENCH_PRO_TASK_LABEL = "SWE-Bench Pro · 单任务";
+const FEATUREBENCH_TASK = "featurebench";
+const FEATUREBENCH_TASK_LABEL = "FeatureBench · 单 feature";
+/** Common pip indexes offered by the settings page. */
+const PIP_MIRRORS = [
+	{
+		id: "pypi",
+		label: "PyPI",
+		url: "https://pypi.org/simple"
+	},
+	{
+		id: "tsinghua",
+		label: "清华 TUNA",
+		url: "https://pypi.tuna.tsinghua.edu.cn/simple"
+	},
+	{
+		id: "aliyun",
+		label: "阿里云",
+		url: "https://mirrors.aliyun.com/pypi/simple"
+	},
+	{
+		id: "tencent",
+		label: "腾讯云",
+		url: "https://mirrors.cloud.tencent.com/pypi/simple"
+	},
+	{
+		id: "bfsu",
+		label: "北外",
+		url: "https://mirrors.bfsu.edu.cn/pypi/web/simple"
+	},
+	{
+		id: "sjtu",
+		label: "上海交大",
+		url: "https://mirror.sjtu.edu.cn/pypi/web/simple"
+	}
+];
 //#endregion
 //#region src/policy.ts
 /**
@@ -755,7 +810,13 @@ var Guard = class {
 			timeLimitMinutes: read("timeLimitMinutes"),
 			forbiddenPaths: read("forbiddenPaths"),
 			toolBlacklistUseDefault: read("toolBlacklistUseDefault"),
-			toolBlacklist: read("toolBlacklist")
+			toolBlacklist: read("toolBlacklist"),
+			cloneSource: read("cloneSource"),
+			customCloneUrl: read("customCloneUrl"),
+			pipIndex: read("pipIndex"),
+			customPipUrl: read("customPipUrl"),
+			gitProxyUrl: read("gitProxyUrl"),
+			pipProxyUrl: read("pipProxyUrl")
 		});
 	}
 	/** Apply a new configuration document. */
@@ -770,8 +831,6 @@ var Guard = class {
 		const wasEnabled = this.enabled;
 		if (!wasEnabled && next.enabled) this.enable();
 		else if (wasEnabled && !next.enabled) this.disable();
-		else if (this.enabled && !next.timeLimitEnabled) this.deadlines.clear();
-		else if (this.enabled && next.timeLimitEnabled) this.rearmRunning();
 	}
 	/** Whether the guard is active. */
 	get active() {
@@ -898,7 +957,6 @@ var Guard = class {
 			};
 		};
 		this.enabled = true;
-		this.rearmRunning();
 		for (const agent of this.services.agents.list()) if (this.isMain(agent.session)) this.forceApproval(agent.session);
 		this.record({
 			kind: "notice",
@@ -919,15 +977,6 @@ var Guard = class {
 			detail: "评测守护已停用。"
 		});
 	}
-	/** Re-arm deadlines for running main conversations after an enable/config change. */
-	rearmRunning() {
-		if (!this.enabled || !this.configuration.timeLimitEnabled) return;
-		for (const agent of this.services.agents.list()) {
-			if (!this.isMain(agent.session)) continue;
-			const jobs = (this.services.jobs?.list(agent.id) ?? []).filter((job) => job.status === "running" || job.status === "stopping");
-			if (agent.status === "running" || jobs.length > 0) this.arm(agent.session.id);
-		}
-	}
 	/** Set the deadline for one session. */
 	arm(sessionId) {
 		const config = this.configuration;
@@ -985,7 +1034,8 @@ var Guard = class {
 		if (event.type === "user/message") {
 			if ((event.data?.source)?.kind === "user") {
 				this.startedAt.set(session.id, Date.now());
-				this.arm(session.id);
+				if (this.configuration.timeLimitEnabled) this.arm(session.id);
+				else this.deadlines.delete(session.id);
 			}
 		} else if (event.type === "assistant/message") {
 			const usage = event.data?.usage;
@@ -1318,6 +1368,204 @@ var Guard = class {
 	}
 };
 //#endregion
+//#region src/jobs.ts
+/** Persistent background jobs for task setup and single-result evaluation. */
+function now() {
+	return Date.now();
+}
+function newId() {
+	return (/* @__PURE__ */ new Date()).toISOString().replace(/[-:.TZ]/g, "").slice(0, 14) + "-" + Math.random().toString(16).slice(2, 8);
+}
+/** Jobs live under the guard root and survive settings unmounts and page reloads. */
+var JobManager = class {
+	root;
+	jobs = /* @__PURE__ */ new Map();
+	children = /* @__PURE__ */ new Map();
+	constructor() {
+		this.root = path.join(guardRoot(), "jobs");
+		fs.mkdirSync(this.root, { recursive: true });
+		for (const name of fs.readdirSync(this.root)) {
+			const file = path.join(this.root, name, "job.json");
+			if (!fs.existsSync(file)) continue;
+			try {
+				const stored = JSON.parse(fs.readFileSync(file, "utf8"));
+				if (stored.status === "running" || stored.status === "queued") {
+					stored.status = "interrupted";
+					stored.error = "服务重启，任务已中断";
+					stored.endedAt = now();
+					this.write(stored);
+				}
+				this.jobs.set(stored.id, stored);
+			} catch {}
+		}
+	}
+	dir(idValue) {
+		return path.join(this.root, idValue);
+	}
+	write(job) {
+		fs.mkdirSync(this.dir(job.id), { recursive: true });
+		fs.writeFileSync(path.join(this.dir(job.id), "job.json"), JSON.stringify(job, void 0, 2));
+	}
+	/** Start one child process and stream its output to a persistent log. */
+	start(options) {
+		const jobId = newId();
+		const directory = this.dir(jobId);
+		fs.mkdirSync(directory, { recursive: true });
+		const logFile = path.join(directory, "log.txt");
+		fs.writeFileSync(logFile, "");
+		const job = {
+			id: jobId,
+			kind: options.kind,
+			task: options.task,
+			workspaceId: options.workspaceId,
+			status: "running",
+			createdAt: now(),
+			startedAt: now(),
+			endedAt: null,
+			updatedAt: now(),
+			command: options.command,
+			cwd: options.cwd,
+			logFile,
+			resultFile: path.join(directory, "result.json")
+		};
+		this.jobs.set(jobId, job);
+		this.write(job);
+		let child;
+		try {
+			child = spawn(options.command[0], options.command.slice(1), {
+				cwd: options.cwd,
+				env: {
+					...process.env,
+					...options.env ?? {},
+					PYTHONUNBUFFERED: "1",
+					EVAL_JOB_DIR: directory,
+					EVAL_RESULT_FILE: job.resultFile
+				},
+				detached: true,
+				stdio: [
+					"ignore",
+					"pipe",
+					"pipe"
+				]
+			});
+		} catch (error) {
+			job.status = "failed";
+			job.error = error instanceof Error ? error.message : String(error);
+			job.endedAt = now();
+			job.updatedAt = now();
+			this.write(job);
+			return this.info(job);
+		}
+		this.children.set(jobId, child);
+		const append = (chunk) => {
+			try {
+				fs.appendFileSync(logFile, chunk);
+			} catch {}
+		};
+		child.stdout?.on("data", (chunk) => append(chunk));
+		child.stderr?.on("data", (chunk) => append(chunk));
+		child.on("error", (error) => {
+			job.status = "failed";
+			job.error = error.message;
+			job.endedAt = now();
+			job.updatedAt = now();
+			this.write(job);
+		});
+		child.on("exit", (code, signal) => {
+			this.children.delete(jobId);
+			if (job.status === "cancelled") return;
+			if (code === 0) try {
+				const extra = options.onSuccess?.(this.detail(jobId));
+				if (extra !== void 0) {
+					if (extra.prompt !== void 0) job.prompt = extra.prompt;
+					if (extra.runId !== void 0) job.runId = extra.runId;
+					if (extra.result !== void 0) {
+						job.result = extra.result;
+						fs.writeFileSync(job.resultFile, JSON.stringify(extra.result, void 0, 2));
+					}
+				}
+				if (fs.existsSync(job.resultFile)) try {
+					job.result = JSON.parse(fs.readFileSync(job.resultFile, "utf8"));
+				} catch {}
+				job.status = "succeeded";
+			} catch (error) {
+				job.status = "failed";
+				job.error = error instanceof Error ? error.message : String(error);
+			}
+			else {
+				job.status = "failed";
+				job.error = signal !== null ? "进程被信号 " + signal + " 终止" : "退出码 " + String(code);
+			}
+			job.endedAt = now();
+			job.updatedAt = now();
+			this.write(job);
+		});
+		return this.info(job);
+	}
+	/** Cancel a running job and its process group. */
+	cancel(jobId) {
+		const job = this.jobs.get(jobId);
+		if (job === void 0) return void 0;
+		if (job.status !== "running" && job.status !== "queued") return this.info(job);
+		job.status = "cancelled";
+		job.endedAt = now();
+		job.updatedAt = now();
+		job.error = "用户取消";
+		this.write(job);
+		const child = this.children.get(jobId);
+		if (child?.pid !== void 0) try {
+			process.kill(-child.pid, "SIGTERM");
+		} catch {
+			try {
+				child.kill("SIGTERM");
+			} catch {}
+		}
+		return this.info(job);
+	}
+	/** Current state, including a bounded log tail and optional result. */
+	get(jobId) {
+		if (!this.jobs.has(jobId)) return void 0;
+		return this.detail(jobId);
+	}
+	/** Recent jobs, optionally filtered by workspace and task. */
+	list(workspaceId, task) {
+		return [...this.jobs.values()].filter((job) => workspaceId === void 0 || job.workspaceId === workspaceId).filter((job) => task === void 0 || job.task === task).sort((a, b) => b.createdAt - a.createdAt).slice(0, 20).map((job) => this.info(job));
+	}
+	info(job) {
+		return {
+			id: job.id,
+			kind: job.kind,
+			task: job.task,
+			workspaceId: job.workspaceId,
+			status: job.status,
+			createdAt: job.createdAt,
+			startedAt: job.startedAt,
+			endedAt: job.endedAt,
+			updatedAt: job.updatedAt,
+			error: job.error
+		};
+	}
+	detail(jobId) {
+		const job = this.jobs.get(jobId);
+		let log = "";
+		try {
+			log = fs.readFileSync(job.logFile, "utf8").slice(-16e3);
+		} catch {}
+		let result;
+		if (job.result !== void 0) result = job.result;
+		else if (fs.existsSync(job.resultFile)) try {
+			result = JSON.parse(fs.readFileSync(job.resultFile, "utf8"));
+		} catch {}
+		return {
+			...this.info(job),
+			log,
+			prompt: job.prompt,
+			runId: job.runId,
+			result
+		};
+	}
+};
+//#endregion
 //#region src/task.ts
 /**
 * Task packs: packaged workspace.zip install, original prompt construction, and
@@ -1329,27 +1577,100 @@ const TASKS = [
 	{
 		id: TANK_TASK,
 		label: TANK_TASK_LABEL,
-		description: "1v1 坦克大战 bot；解压 workspace.zip，提交 workspace/submission/bot.cpp。"
+		description: "1v1 坦克大战 bot；解压 workspace.zip，提交 workspace/submission/bot.cpp。",
+		timeLimitMinutes: 90,
+		timeLimitEnabled: true,
+		needsSetup: false
 	},
 	{
 		id: WAREHOUSE_TASK,
 		label: WAREHOUSE_TASK_LABEL,
-		description: "单人单次仓储调度优化；解压 workspace.zip，提交 workspace/src/。"
+		description: "单人单次仓储调度优化；解压 workspace.zip，提交 workspace/src/。",
+		timeLimitMinutes: 90,
+		timeLimitEnabled: true,
+		needsSetup: false
 	},
 	{
 		id: LAND_TASK,
 		label: LAND_TASK_LABEL,
-		description: "圈地 1v1/4 人混战 bot；解压 workspace.zip，提交 workspace/submission/bot.cpp。"
+		description: "圈地 1v1/4 人混战 bot；解压 workspace.zip，提交 workspace/submission/bot.cpp。",
+		timeLimitMinutes: 90,
+		timeLimitEnabled: true,
+		needsSetup: false
 	},
 	{
 		id: BOMBERMAN_TASK,
 		label: BOMBERMAN_TASK_LABEL,
-		description: "炸弹人 2–10 人 bot；解压 workspace.zip，提交 workspace/submission/bot.cpp。"
+		description: "炸弹人 2–10 人 bot；解压 workspace.zip，提交 workspace/submission/bot.cpp。",
+		timeLimitMinutes: 90,
+		timeLimitEnabled: true,
+		needsSetup: false
 	},
 	{
 		id: "ctf",
 		label: CTF_TASK_LABEL,
-		description: "夺旗 2–15 人 bot；解压 workspace.zip，提交 workspace/submission/bot.cpp。"
+		description: "夺旗 2–15 人 bot；解压 workspace.zip，提交 workspace/submission/bot.cpp。",
+		timeLimitMinutes: 90,
+		timeLimitEnabled: true,
+		needsSetup: false
+	},
+	{
+		id: COMMIT0_TASK,
+		label: COMMIT0_TASK_LABEL,
+		description: "Commit0 单库实现：按规格从零实现一个 Python 库，单元测试评分。",
+		timeLimitMinutes: 480,
+		timeLimitEnabled: false,
+		needsSetup: true,
+		repoUrl: "https://github.com/msiemens/tinydb",
+		remotes: [{
+			id: "github",
+			label: "GitHub",
+			url: "https://github.com/msiemens/tinydb.git"
+		}]
+	},
+	{
+		id: SWE_BENCH_PRO_TASK,
+		label: SWE_BENCH_PRO_TASK_LABEL,
+		description: "SWE-Bench Pro 单任务：在真实仓库中修复一个多文件 issue，隐藏测试评分。",
+		timeLimitMinutes: 1440,
+		timeLimitEnabled: false,
+		needsSetup: true,
+		repoUrl: "https://github.com/ansible/ansible",
+		remotes: [{
+			id: "github",
+			label: "GitHub",
+			url: "https://github.com/ansible/ansible.git"
+		}, {
+			id: "gitee",
+			label: "Gitee 镜像",
+			url: "https://gitee.com/mirrors/ansible.git"
+		}]
+	},
+	{
+		id: FEATUREBENCH_TASK,
+		label: FEATUREBENCH_TASK_LABEL,
+		description: "FeatureBench 单 feature：实现一个跨文件功能，隐藏测试评分。",
+		timeLimitMinutes: 1440,
+		timeLimitEnabled: false,
+		needsSetup: true,
+		repoUrl: "https://github.com/mwaskom/seaborn",
+		remotes: [
+			{
+				id: "github",
+				label: "GitHub",
+				url: "https://github.com/mwaskom/seaborn.git"
+			},
+			{
+				id: "gitee",
+				label: "Gitee 镜像",
+				url: "https://gitee.com/mirrors/seaborn.git"
+			},
+			{
+				id: "gitcode",
+				label: "GitCode 镜像",
+				url: "https://gitcode.com/gh_mirrors/se/seaborn.git"
+			}
+		]
 	}
 ];
 /** The tank task definition (original prompt; only halftime mechanics removed). */
@@ -1358,6 +1679,9 @@ const TANK_DEFINITION = {
 	label: TANK_TASK_LABEL,
 	description: TASKS[0].description,
 	asset: path.join(packageRoot(), "assets", "tank", "workspace.zip"),
+	timeLimitMinutes: 90,
+	timeLimitEnabled: true,
+	needsSetup: false,
 	prompt() {
 		return [
 			"按照要求编写尽可能强的坦克大战 bot。你的对手是其他顶尖 AI 模型写的 bot，它们可能主动设计战术。基准是一个中等强度、思路常规的 bot，它在正式比赛中是 0 分线；战胜基准不代表足够强，正式对手远强于基准。正式赛是 1 对 1 循环赛，用 Bradley-Terry 模型拟合实力值。",
@@ -1384,6 +1708,9 @@ const WAREHOUSE_DEFINITION = {
 	label: WAREHOUSE_TASK_LABEL,
 	description: TASKS[1].description,
 	asset: path.join(packageRoot(), "assets", "warehouse", "workspace.zip"),
+	timeLimitMinutes: 90,
+	timeLimitEnabled: true,
+	needsSetup: false,
 	prompt() {
 		return [
 			"这是一场 AI 编程赛。你接手了一个代码混乱的老系统：src/ 里是一个仓储机器人调度仿真系统，功能正确，但运行很慢。你的任务：在输出完全不变的前提下，让它尽可能快。",
@@ -1407,6 +1734,9 @@ const LAND_DEFINITION = {
 	label: LAND_TASK_LABEL,
 	description: TASKS[2].description,
 	asset: path.join(packageRoot(), "assets", "land", "workspace.zip"),
+	timeLimitMinutes: 90,
+	timeLimitEnabled: true,
+	needsSetup: false,
 	prompt() {
 		return [
 			"这是一场正式的 AI 编程对抗赛，你的对手是其他顶尖 AI 模型写的 bot，它们会主动切断你、抢你的地。比赛结果会公开排名，请拿出你的最高水平。",
@@ -1421,6 +1751,9 @@ const BOMBERMAN_DEFINITION = {
 	label: BOMBERMAN_TASK_LABEL,
 	description: TASKS[3].description,
 	asset: path.join(packageRoot(), "assets", "bomberman", "workspace.zip"),
+	timeLimitMinutes: 90,
+	timeLimitEnabled: true,
+	needsSetup: false,
 	prompt() {
 		return ["按照要求编写尽可能强bot，你的对手是全球其他顶尖ai，可能主动设计战术。基准只是很弱的bot，它在正式比赛为0分线。正式赛会对战多轮，用Bradley-Terry模型拟合实力值。", "解压 workspace.zip，阅读 README.md 和 RULES.md，编写支持2–10人的炸弹人 bot。尽量60分钟完成，90分钟必须交卷。允许本地工具、对战和复盘，禁止联网、调用其他AI或查看其他选手资料。提交 submission/bot.cpp，简述策略。"].join("\n");
 	}
@@ -1431,25 +1764,173 @@ const CTF_DEFINITION = {
 	label: CTF_TASK_LABEL,
 	description: TASKS[4].description,
 	asset: path.join(packageRoot(), "assets", "ctf", "workspace.zip"),
+	timeLimitMinutes: 90,
+	timeLimitEnabled: true,
+	needsSetup: false,
 	prompt() {
 		return ["按照要求编写尽可能强的 bot，你的对手是全球其他顶尖 AI，可能主动设计战术。基准是一个中等强度、思路常规的 bot，它在正式比赛是 0 分线；战胜基准不代表足够强，正式对手远强于基准。正式赛会对战多轮，用 Bradley-Terry 模型拟合实力值。", "解压 workspace.zip，阅读 README.md 和 RULES.md，编写支持 2–15 方的夺旗 bot。你有 60 分钟，只看最终 bot 的强度，提前交卷不加分，请充分利用时间持续改进；可以适当超时，但到 90 分钟会强制交卷。允许本地工具、对战和复盘，禁止联网、调用其他 AI 或查看其他选手资料。提交 workspace/submission/bot.cpp，简述策略。"].join("\n");
 	}
 };
+/** Run the task's host-side setup script with network access. */
+function runSetup(workspacePath) {
+	execFileSync("bash", ["setup.sh"], {
+		cwd: workspacePath,
+		stdio: "inherit",
+		env: {
+			...process.env,
+			GIT_TERMINAL_PROMPT: "0"
+		}
+	});
+}
+/** Commit0 single-library task. */
+const COMMIT0_DEFINITION = {
+	id: COMMIT0_TASK,
+	label: COMMIT0_TASK_LABEL,
+	description: TASKS[5].description,
+	asset: path.join(packageRoot(), "assets", "commit0", "workspace.zip"),
+	prepare: runSetup,
+	timeLimitMinutes: 480,
+	timeLimitEnabled: false,
+	needsSetup: true,
+	repoUrl: "https://github.com/msiemens/tinydb",
+	remotes: [{
+		id: "github",
+		label: "GitHub",
+		url: "https://github.com/msiemens/tinydb.git"
+	}],
+	prompt() {
+		return [
+			"这是 Commit0 的一个单库实现任务：从零实现一个 Python 库，使其通过仓库自带的完整单元测试。",
+			"",
+			"当前 workspace 已经由安装器准备好：包含该库的测试、规格文档和项目配置，但库的实现源码已被移除。请阅读 TASK.md 和规格文档，自己实现库，直到单元测试全部通过。",
+			"",
+			"要求：",
+			"- 只能修改当前工作区内的文件；工作区外不可访问，包括 /tmp；",
+			"- 禁止联网、禁止调用其他 AI 或子代理；",
+			"- 使用工作区里的 .venv 运行测试，例如 .venv/bin/python -m pytest tests；",
+			"- 不要删除、修改或伪造测试；评分以隐藏的测试执行为准；",
+			"- 完成后用几句话总结实现思路。"
+		].join("\n");
+	}
+};
+/** SWE-Bench Pro single-task task. */
+const SWE_BENCH_PRO_DEFINITION = {
+	id: SWE_BENCH_PRO_TASK,
+	label: SWE_BENCH_PRO_TASK_LABEL,
+	description: TASKS[6].description,
+	asset: path.join(packageRoot(), "assets", "swe-bench-pro", "workspace.zip"),
+	prepare: runSetup,
+	timeLimitMinutes: 1440,
+	timeLimitEnabled: false,
+	needsSetup: true,
+	repoUrl: "https://github.com/ansible/ansible",
+	remotes: [{
+		id: "github",
+		label: "GitHub",
+		url: "https://github.com/ansible/ansible.git"
+	}, {
+		id: "gitee",
+		label: "Gitee 镜像",
+		url: "https://gitee.com/mirrors/ansible.git"
+	}],
+	prompt() {
+		return [
+			"这是 SWE-Bench Pro 的一个单任务：在真实的 Python 仓库中完成一个多文件 issue 修复。",
+			"",
+			"当前 workspace 已经由安装器准备好：包含目标仓库在指定 commit 的源码和依赖环境。请阅读 TASK.md 中的 issue 描述与要求，修改源码完成修复。",
+			"",
+			"要求：",
+			"- 只能修改当前工作区内的文件；工作区外不可访问，包括 /tmp；",
+			"- 禁止联网、禁止调用其他 AI 或子代理；",
+			"- 使用工作区里的 .venv 运行测试，例如 .venv/bin/python -m pytest <测试文件>；",
+			"- 不要删除、修改或伪造测试；评分会在你的代码上应用隐藏测试；",
+			"- 完成后用几句话总结修改。"
+		].join("\n");
+	}
+};
+/** FeatureBench single-feature task. */
+const FEATUREBENCH_DEFINITION = {
+	id: FEATUREBENCH_TASK,
+	label: FEATUREBENCH_TASK_LABEL,
+	description: TASKS[7].description,
+	asset: path.join(packageRoot(), "assets", "featurebench", "workspace.zip"),
+	prepare: runSetup,
+	timeLimitMinutes: 1440,
+	timeLimitEnabled: false,
+	needsSetup: true,
+	repoUrl: "https://github.com/mwaskom/seaborn",
+	remotes: [
+		{
+			id: "github",
+			label: "GitHub",
+			url: "https://github.com/mwaskom/seaborn.git"
+		},
+		{
+			id: "gitee",
+			label: "Gitee 镜像",
+			url: "https://gitee.com/mirrors/seaborn.git"
+		},
+		{
+			id: "gitcode",
+			label: "GitCode 镜像",
+			url: "https://gitcode.com/gh_mirrors/se/seaborn.git"
+		}
+	],
+	prompt() {
+		return [
+			"这是 FeatureBench 的一个单 feature 任务：在真实 Python 仓库中实现一个跨文件功能，并保持已有功能不退化。",
+			"",
+			"当前 workspace 已经由安装器准备好：包含目标仓库源码和依赖环境。请阅读 TASK.md 中的功能需求，修改源码完成实现。",
+			"",
+			"要求：",
+			"- 只能修改当前工作区内的文件；工作区外不可访问，包括 /tmp；",
+			"- 禁止联网、禁止调用其他 AI 或子代理；",
+			"- 使用工作区里的 .venv 运行测试；",
+			"- 不要删除、修改或伪造测试；评分会在你的代码上运行隐藏测试和回归测试；",
+			"- 完成后用几句话总结实现思路。"
+		].join("\n");
+	}
+};
 /** All task definitions keyed by id. */
 const TASK_DEFINITIONS = {
 	[TANK_TASK]: TANK_DEFINITION,
 	[WAREHOUSE_TASK]: WAREHOUSE_DEFINITION,
 	[LAND_TASK]: LAND_DEFINITION,
 	[BOMBERMAN_TASK]: BOMBERMAN_DEFINITION,
-	["ctf"]: CTF_DEFINITION
+	["ctf"]: CTF_DEFINITION,
+	[COMMIT0_TASK]: COMMIT0_DEFINITION,
+	[SWE_BENCH_PRO_TASK]: SWE_BENCH_PRO_DEFINITION,
+	[FEATUREBENCH_TASK]: FEATUREBENCH_DEFINITION
 };
-/** Install, check, and prompt services for the task packs. */
 var TaskService = class {
 	guard;
 	registry;
+	jobs;
 	constructor(guard, registry) {
 		this.guard = guard;
 		this.registry = registry;
+		this.jobs = new JobManager();
+	}
+	/** The clone URL for a task, honouring the selected source. */
+	cloneUrl(definition, options) {
+		const repo = definition.repoUrl ?? "";
+		const source = options.cloneSource ?? "github";
+		if (source === "custom" && options.customCloneUrl !== void 0 && options.customCloneUrl.trim() !== "") return options.customCloneUrl.trim().replace(/\/$/, "") + "/" + repo.replace(/^https?:\/\/github\.com\//, "");
+		const remote = definition.remotes?.find((item) => item.id === source);
+		if (remote !== void 0) return remote.url;
+		return definition.remotes?.[0]?.url ?? repo;
+	}
+	/** Normalize a user-entered proxy address, or return an empty string. */
+	proxyValue(raw) {
+		const value = (raw ?? "").trim();
+		if (value === "") return "";
+		return /^[a-z][a-z0-9+.-]*:\/\//i.test(value) ? value : "http://" + value;
+	}
+	/** The pip index URL for a task, honouring the selected mirror. */
+	pipIndexUrl(options) {
+		const index = options.pipIndex ?? "pypi";
+		if (index === "custom") return (options.customPipUrl ?? "").trim() || "https://pypi.org/simple";
+		return PIP_MIRRORS.find((item) => item.id === index)?.url ?? "https://pypi.org/simple";
 	}
 	/** Check one workspace before install. */
 	check(workspaceId) {
@@ -1487,6 +1968,118 @@ var TaskService = class {
 	definition(taskId) {
 		return TASK_DEFINITIONS[taskId];
 	}
+	/** Start a non-blocking install. The settings page polls the returned job. */
+	startInstall(workspaceId, taskId, options = {}) {
+		const definition = TASK_DEFINITIONS[taskId];
+		if (definition === void 0) return {
+			ok: false,
+			error: "未知任务：" + taskId
+		};
+		const checked = this.check(workspaceId);
+		if (!checked.ok) return {
+			ok: false,
+			error: checked.error
+		};
+		if (!checked.empty) return {
+			ok: false,
+			error: "工作区不是空的；未执行任何操作。"
+		};
+		if (!this.guard.active) return {
+			ok: false,
+			error: "请先启用评测守护，再安装评测任务。"
+		};
+		try {
+			fs.copyFileSync(this.materialZip(taskId), path.join(checked.path, "workspace.zip"));
+		} catch (error) {
+			return {
+				ok: false,
+				error: "复制素材失败：" + (error instanceof Error ? error.message : String(error))
+			};
+		}
+		const env = {
+			EVAL_CLONE_URL: this.cloneUrl(definition, options),
+			EVAL_PIP_INDEX_URL: this.pipIndexUrl(options),
+			GIT_TERMINAL_PROMPT: "0",
+			EVAL_GIT_PROXY: this.proxyValue(options.gitProxyUrl),
+			EVAL_PIP_PROXY: this.proxyValue(options.pipProxyUrl)
+		};
+		const command = definition.needsSetup ? ["bash", "setup.sh"] : [
+			"bash",
+			"-c",
+			"true"
+		];
+		return {
+			ok: true,
+			jobId: this.jobs.start({
+				kind: "install",
+				task: taskId,
+				workspaceId,
+				command,
+				cwd: checked.path,
+				env,
+				onSuccess: () => {
+					const runId = this.newRunId();
+					const prompt = this.guard.promptFor(checked.path, this.materialZip(taskId), definition);
+					const run = {
+						runId,
+						task: definition.id,
+						workspaceId,
+						workspacePath: checked.path,
+						sessionId: null,
+						createdAt: Date.now(),
+						startedAt: null,
+						endedAt: null,
+						status: "prepared",
+						prompt,
+						timeLimitMinutes: this.guard.configuration.timeLimitMinutes
+					};
+					this.guard.registerRun(run);
+					return {
+						runId,
+						prompt
+					};
+				}
+			}).id
+		};
+	}
+	/** Start a single-workspace evaluation job. */
+	startEval(workspaceId, taskId) {
+		if (TASK_DEFINITIONS[taskId] === void 0) return {
+			ok: false,
+			error: "未知任务：" + taskId
+		};
+		const checked = this.check(workspaceId);
+		if (!checked.ok) return {
+			ok: false,
+			error: checked.error
+		};
+		const script = path.join(packageRoot(), "eval", taskId + ".py");
+		if (!fs.existsSync(script)) return {
+			ok: false,
+			error: "评测脚本不存在：" + script
+		};
+		return {
+			ok: true,
+			jobId: this.jobs.start({
+				kind: "eval",
+				task: taskId,
+				workspaceId,
+				command: [
+					"python3",
+					"-X",
+					"utf8",
+					script,
+					"--workspace",
+					checked.path
+				],
+				cwd: packageRoot(),
+				env: {
+					PYTHONPATH: "",
+					PYTHONHOME: ""
+				}
+			}).id
+		};
+	}
 	/** Copy the packaged workspace.zip to a stable material path for install/manual use. */
 	materialZip(taskId = TANK_TASK) {
 		const definition = TASK_DEFINITIONS[taskId];
@@ -1533,12 +2126,13 @@ var TaskService = class {
 		};
 		try {
 			fs.copyFileSync(this.materialZip(taskId), path.join(checked.path, "workspace.zip"));
+			if (definition.prepare !== void 0) definition.prepare(checked.path);
 		} catch (error) {
 			return {
 				ok: false,
 				workspaceId,
 				workspacePath: checked.path,
-				error: "复制素材失败：" + (error instanceof Error ? error.message : String(error))
+				error: "任务准备失败：" + (error instanceof Error ? error.message : String(error))
 			};
 		}
 		const runId = this.newRunId();
@@ -1645,7 +2239,8 @@ function registerRoutes(ctx, guard, tasks) {
 					defaultBlacklist: guard.defaultBlacklist(),
 					generatedDefault: guard.configuration.toolBlacklistUseDefault,
 					workspaceRoot: dshHome(),
-					runsRoot: path.join(guardRoot(), "runs")
+					runsRoot: path.join(guardRoot(), "runs"),
+					pipMirrors: [...PIP_MIRRORS]
 				});
 			}
 		});
@@ -1673,7 +2268,69 @@ function registerRoutes(ctx, guard, tasks) {
 				const workspaceId = typeof body.workspaceId === "string" ? body.workspaceId : "";
 				const taskId = typeof body.task === "string" && body.task !== "" ? body.task : TANK_TASK;
 				if (workspaceId === "") return json({ error: "workspaceId is required" }, 400);
-				return json(tasks.install(workspaceId, taskId));
+				const config = guard.configuration;
+				return json(tasks.startInstall(workspaceId, taskId, {
+					cloneSource: typeof body.cloneSource === "string" ? body.cloneSource : config.cloneSource,
+					customCloneUrl: typeof body.customCloneUrl === "string" ? body.customCloneUrl : config.customCloneUrl,
+					pipIndex: typeof body.pipIndex === "string" ? body.pipIndex : config.pipIndex,
+					customPipUrl: typeof body.customPipUrl === "string" ? body.customPipUrl : config.customPipUrl,
+					gitProxyUrl: typeof body.gitProxyUrl === "string" ? body.gitProxyUrl : config.gitProxyUrl,
+					pipProxyUrl: typeof body.pipProxyUrl === "string" ? body.pipProxyUrl : config.pipProxyUrl
+				}));
+			}
+		});
+		add({
+			path: EVAL_ROUTE,
+			methods: ["POST"],
+			requestBody: "buffered",
+			fetch: async (request) => {
+				let body = {};
+				try {
+					body = await request.json();
+				} catch {
+					return json({ error: "invalid body" }, 400);
+				}
+				const workspaceId = typeof body.workspaceId === "string" ? body.workspaceId : "";
+				const taskId = typeof body.task === "string" && body.task !== "" ? body.task : TANK_TASK;
+				if (workspaceId === "") return json({ error: "workspaceId is required" }, 400);
+				return json(tasks.startEval(workspaceId, taskId));
+			}
+		});
+		add({
+			path: JOBS_ROUTE,
+			methods: ["GET"],
+			requestBody: "buffered",
+			fetch: async (request) => {
+				const workspaceId = query(request, "workspaceId");
+				const taskId = query(request, "task");
+				return json(tasks.jobs.list(workspaceId === "" ? void 0 : workspaceId, taskId === "" ? void 0 : taskId));
+			}
+		});
+		add({
+			path: JOB_ROUTE,
+			methods: ["GET"],
+			requestBody: "buffered",
+			fetch: async (request) => {
+				const jobId = query(request, "id");
+				const job = tasks.jobs.get(jobId);
+				return job === void 0 ? json({ error: "job not found" }, 404) : json(job);
+			}
+		});
+		add({
+			path: JOB_CANCEL_ROUTE,
+			methods: ["POST"],
+			requestBody: "buffered",
+			fetch: async (request) => {
+				let body = {};
+				try {
+					body = await request.json();
+				} catch {
+					return json({ error: "invalid body" }, 400);
+				}
+				const jobId = typeof body.id === "string" ? body.id : "";
+				if (jobId === "") return json({ error: "id is required" }, 400);
+				const job = tasks.jobs.cancel(jobId);
+				return job === void 0 ? json({ error: "job not found" }, 404) : json(job);
 			}
 		});
 		add({