Fix tests

Author: HuanCheng65

grader.py

@@ -560,7 +560,7 @@ def _find_pattern(
             匹配结束的位置，未找到则返回-1
         """
         if regex_mode:
-            match = re.search(pattern, text[start_pos:])
+            match = re.search(pattern, text[start_pos:], re.MULTILINE)
             if match:
                 match_end = start_pos + match.end()
                 return match_end
@@ -1243,8 +1243,12 @@ def _execute_interactive_steps(
         # 获取步骤描述基础信息
         step_name = step.get("name", step["command"])
         base_description = current_description if current_description else f"Running {test.meta['name']}: {step_name}"
+        result = None
 
         for i, interaction_step in enumerate(step.get("steps", []), 1):
+            if result:
+                break
+            
             step_type = interaction_step.get("type", "")
 
             # 更新进度显示，显示内部步骤进度
@@ -1305,8 +1309,6 @@ def _execute_interactive_steps(
                     cmd,
                     args,
                 )
-                if result:
-                    return result
             elif step_type == "sleep":
                 self._handle_sleep_step(interaction_step)
             elif step_type == "signal":
@@ -1326,8 +1328,6 @@ def _execute_interactive_steps(
                     cmd,
                     args,
                 )
-                if result:
-                    return result
 
             self.previous_step_type = step_type
 
@@ -1367,6 +1367,9 @@ def _execute_interactive_steps(
         interactive_process.terminate()
         if ref_process:
             ref_process.terminate()
+            
+        if result:
+            return result
 
         # 如果没有分步评分，使用整体分数
         if not self.step_scores:
@@ -1818,17 +1821,14 @@ def _resolve_relative_path(self, path: str, cwd: Path = os.getcwd()) -> str:
     def _resolve_path(
         self, path: str, test_dir: Path, cwd: Path = os.getcwd(), relative: bool = True
     ) -> str:
-        # 检查是否是预定义的可执行文件
         if path.startswith("${") and path.endswith("}"):
-            # 提取可执行文件名称
             exec_name = path[2:-1]  # 去掉 ${ 和 }
-            # 检查是否在预定义的可执行文件列表中
             if exec_name in self.config.executables:
-                # 替换为预定义的可执行文件路径
                 path = self.config.executables[exec_name]
 
         build_dir = test_dir / "build"
         build_dir.mkdir(exist_ok=True)
+        
         if relative:
             replacements = {
                 "${test_dir}": self._resolve_relative_path(test_dir, cwd),
@@ -2483,6 +2483,11 @@ def _resolve_relative_path(self, path: str, cwd: Path = os.getcwd()) -> str:
         return result
 
     def _resolve_path(self, path: str, test_dir: Path, cwd: Path = os.getcwd()) -> str:
+        if path.startswith("${") and path.endswith("}"):
+            exec_name = path[2:-1]
+            if exec_name in self.config.executables:
+                path = self.config.executables[exec_name]
+                
         build_dir = test_dir / "build"
         build_dir.mkdir(exist_ok=True)
 

tests/cases/10-fg/config.toml

@@ -66,7 +66,6 @@ type = "close"
 # Wait for shell to finish processing
 [[run.steps]]
 type = "wait"
-timeout = 3.0
 
 # Validate output against reference
 [[run.steps]]

tests/cases/17-cd/judge.py

@@ -10,27 +10,50 @@ def judge():
     stdout = input_data["stdout"]
     max_score = input_data["max_score"]
 
-    # 查找 shell 提示符模式，以便正确识别命令输出
-    prompt_pattern = r"[\w\d\-_]+[@:][\w\d\-_/]+[#$%>]"
-    prompt_matches = re.findall(prompt_pattern, stdout)
-    prompt = None
-    if prompt_matches:
-        # 使用最常见的提示符
-        from collections import Counter
+    prompt = "> "
+    success = True
+    score = 0
+
+    # 分析输出，查找命令和它们的输出
+    commands_with_output = []
+
+    # 使用提示符分割输出
+    lines = stdout.split("\n")
+    prompt_line_indices = []
+
+    # 找出所有提示符所在行的索引
+    for i, line in enumerate(lines):
+        if line.startswith(prompt.strip()):
+            prompt_line_indices.append(i)
+
+    # 解析每个命令和它的输出
+    for i, prompt_index in enumerate(prompt_line_indices):
+        # 获取命令
+        cmd_line = lines[prompt_index]
+        cmd = cmd_line[len(prompt.strip()) :].strip()
 
-        prompt = Counter(prompt_matches).most_common(1)[0][0]
+        # 获取输出: 从命令行的下一行开始，到下一个提示符行（或文件结束）
+        output_start = prompt_index + 1
+        output_end = (
+            prompt_line_indices[i + 1]
+            if i + 1 < len(prompt_line_indices)
+            else len(lines)
+        )
+
+        # 如果没有输出行或者下一行就是提示符，则输出为空
+        if output_start >= output_end:
+            output = ""
+        else:
+            output = "\n".join(lines[output_start:output_end])
+
+        commands_with_output.append((cmd, output))
 
     # 提取 pwd 命令的输出
     pwd_outputs = []
-    if prompt:
-        # 查找 pwd 命令的输出
-        pwd_pattern = rf"{re.escape(prompt)} pwd\s*\n(.*?)(?={re.escape(prompt)}|$)"
-        pwd_matches = re.findall(pwd_pattern, stdout, re.DOTALL)
-        pwd_outputs = [match.strip() for match in pwd_matches]
-    else:
-        # 如果找不到提示符，尝试直接提取可能的路径
-        path_pattern = r"/[a-zA-Z0-9_\-./]+"
-        pwd_outputs = re.findall(path_pattern, stdout)
+    
+    for cmd, output in commands_with_output:
+        if cmd.strip() == "pwd":
+            pwd_outputs.append(output.strip())
 
     # 检查是否有至少两个不同的目录输出（初始目录和 cd 后的目录）
     success = False
@@ -44,17 +67,17 @@ def judge():
             try:
                 first_dir = os.path.normpath(pwd_outputs[0])
                 second_dir = os.path.normpath(pwd_outputs[1])
-                parent_dir = os.path.normpath(os.path.join(first_dir, ".."))
+                
+                # 支持中文路径，确保正确处理路径分隔符
+                parent_dir = os.path.normpath(os.path.dirname(first_dir))
 
                 if second_dir == parent_dir:
                     success = True
                     message = "Directory change verified successfully"
                     score = max_score
                 else:
                     message = f"Second directory ({second_dir}) is not the parent of first directory ({first_dir})"
-                    score = (
-                        max_score * 0.5
-                    )  # 部分得分，因为目录确实改变了，但不是预期的父目录
+                    score = max_score * 0.5  # 部分得分，因为目录确实改变了，但不是预期的父目录
             except Exception as e:
                 message = f"Error comparing directories: {str(e)}"
         else:

tests/cases/3-foreground-job/config.toml

@@ -27,9 +27,7 @@ echo = false
 
 # Quit the shell
 [[run.steps]]
-type = "input"
-content = "quit"
-echo = false
+type = "close"
 
 # Wait for the shell to terminate
 [[run.steps]]
@@ -40,7 +38,7 @@ type = "check"
 
 [run.steps.check.sequence]
 patterns = [
-    { pattern = "^Hello from foreground job!\n$", required = true, description = "First message" },
-    { pattern = "^Another message from tsh!\n$", required = true, description = "Second message" },
+    { pattern = "^Hello from foreground job!", required = true, description = "First message" },
+    { pattern = "^Another message from tsh!", required = true, description = "Second message" },
 ]
 verify_end = true

tests/cases/4-background-job/config.toml

@@ -41,9 +41,7 @@ echo = false
 
 # Quit the shell
 [[run.steps]]
-type = "input"
-content = "quit"
-echo = false
+type = "close"
 
 # Wait for the shell to terminate
 [[run.steps]]
@@ -54,8 +52,7 @@ type = "check"
 
 [run.steps.check.sequence]
 patterns = [
-    { pattern = '''^\[\d+\] \(\d+\) \.\/myspin 1 &\n$''', required = true, description = "Background job started" },
-    { pattern = '''^\[\d+\] \(\d+\) Running \.\/myspin 1 &\n$''', required = true, description = "Background job running" },
-    { pattern = "^$", required = true, description = "Background job completed" },
+    { pattern = '''^\[\d+\] \(\d+\) \.\/myspin 2 &''', required = true, description = "Background job started" },
+    { pattern = '''^\[\d+\] \(\d+\) Running \.\/myspin 2 &''', required = true, description = "Background job running" },
 ]
 verify_end = true