dsh-eval-guard 0.4.0 → 0.4.1

package/README.md

@@ -34,7 +34,9 @@
 - 3 个长时单任务：Commit0 单库实现、SWE-Bench Pro 单任务、FeatureBench 单 feature。
 - 1 个纯 C++ 项目：Neural-CPP 张量运算与反向传播后端。
 
-Neural-CPP 工作区只包含 TASK.md、CMakeLists.txt 和 include/nn/nn.hpp；实现在 src/ 中从零完成，纯 C++17，无 Python/OpenCL/第三方库。
+Neural-CPP 安装时自动把 workspace.zip 解压到工作区根目录，并在完成后移除 setup.sh；agent 从根目录的 CMakeLists.txt 开始。工作区只包含 TASK.md、CMakeLists.txt 和 include/nn/nn.hpp；实现在 src/ 中从零完成，纯 C++17，无 Python/OpenCL/第三方库。
+
+评测器同时兼容旧安装留下的嵌套布局：若根目录没有 CMakeLists.txt，会自动改用 workspace/CMakeLists.txt，避免已经产出的工作区无法评测。
 
 评测时在临时目录独立构建，然后从 assets/neural-cpp/hidden/ 注入 C++ 测试源码；测试数据由固定种子在进程内生成，不依赖或复制外部数据文件。
 

package/eval/neural-cpp.py

@@ -36,7 +36,16 @@ def main() -> int:
         root = Path(td)
         candidate = root / 'candidate'
         clean_copy(workspace, candidate)
-        build = candidate / 'build'
+        # The first task installation copied workspace.zip into the session
+        # root; some agents extracted it into workspace/ instead of the root.
+        # Accept both layouts so already-produced workspaces remain evaluable.
+        source = candidate
+        if not (source / 'CMakeLists.txt').is_file() and (candidate / 'workspace' / 'CMakeLists.txt').is_file():
+            source = candidate / 'workspace'
+        if not (source / 'CMakeLists.txt').is_file():
+            return fail({'task': 'neural-cpp', 'ok': False, 'error': 'CMakeLists.txt missing',
+                         'output': 'Expected CMakeLists.txt at workspace root or workspace/CMakeLists.txt'})
+        build = source / 'build'
         if build.exists():
             shutil.rmtree(build, ignore_errors=True)
         build.mkdir()
@@ -46,11 +55,11 @@ def main() -> int:
         if cmake is None or compiler is None:
             return fail({'task': 'neural-cpp', 'ok': False, 'error': 'cmake/g++ not available'})
 
-        cfg = run([cmake, '-S', '.', '-B', 'build', '-DCMAKE_BUILD_TYPE=Release'], candidate, 180)
+        cfg = run([cmake, '-S', '.', '-B', 'build', '-DCMAKE_BUILD_TYPE=Release'], source, 180)
         if cfg.returncode != 0:
             return fail({'task': 'neural-cpp', 'ok': False, 'error': 'cmake configure failed',
                          'output': (cfg.stdout + cfg.stderr)[-5000:]})
-        bld = run([cmake, '--build', 'build', '-j4'], candidate, 1800)
+        bld = run([cmake, '--build', 'build', '-j4'], source, 1800)
         if bld.returncode != 0:
             return fail({'task': 'neural-cpp', 'ok': False, 'error': 'build failed',
                          'output': (bld.stdout + bld.stderr)[-5000:]})
@@ -60,7 +69,7 @@ def main() -> int:
 
         exe = root / 'test_neural'
         cc = run([compiler, '-std=c++17', '-O3', '-pthread',
-                  '-I', str(candidate / 'include'), str(hidden), str(lib),
+                  '-I', str(source / 'include'), str(hidden), str(lib),
                   '-lm', '-o', str(exe)], root, 300)
         if cc.returncode != 0:
             return fail({'task': 'neural-cpp', 'ok': False, 'error': 'hidden test compile failed',

package/lib/index.js

@@ -1680,7 +1680,7 @@ const TASKS = [
 		description: "从零实现纯 C++ 张量运算与反向传播后端；隐藏测试统计正确性与速度。",
 		timeLimitMinutes: 1440,
 		timeLimitEnabled: false,
-		needsSetup: false
+		needsSetup: true
 	}
 ];
 /** The tank task definition (original prompt; only halftime mechanics removed). */
@@ -1909,12 +1909,12 @@ const NEURAL_CPP_DEFINITION = {
 	asset: path.join(packageRoot(), "assets", "neural-cpp", "workspace.zip"),
 	timeLimitMinutes: 1440,
 	timeLimitEnabled: false,
-	needsSetup: false,
+	needsSetup: true,
 	prompt() {
 		return [
 			"这是一个纯 C++ 张量运算后端任务：从零实现 DType、Tensor、视图/切片、逐元素与矩阵运算、自动求导和线程池；不提供 Python 或 OpenCL 支持。",
 			"",
-			"当前目录有 workspace.zip。请先解压，进入 workspace 目录，阅读 TASK.md 和 include/nn/nn.hpp，按接口约定实现 src/ 下的静态库，并用 CMake 构建。",
+			"当前 workspace 已经由安装器准备好：包含 TASK.md、CMakeLists.txt 和 include/nn/nn.hpp。请阅读 TASK.md 和公开头文件，按接口约定实现 src/ 下的静态库，并用 CMake 构建。",
 			"",
 			"要求：",
 			"- 仅使用 Linux + GCC + C++17 标准库；不允许引入第三方依赖；",

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "dsh-eval-guard",
-  "version": "0.4.0",
+  "version": "0.4.1",
   "type": "module",
   "description": "Evaluation guard for DSH: workspace-only filesystem, no network/subagents, wall-clock limits, and task installation for the meow-ai-arena tank task.",
   "license": "MIT",