fix: preserve posix pytest environment

Author: JayLin

docs/architecture/governed-test-execution.md

@@ -17,6 +17,7 @@ The capability has four boundaries:
 ## Composition
 
 ```python
+import sys
 from pathlib import Path
 
 from mini_code_agent.policy import (
@@ -38,7 +39,7 @@ workspace = WorkspaceBoundary(root)
 runner = PytestRunner(
     root,
     profile=PytestProfile(
-        python_executable=Path(".venv/Scripts/python.exe").resolve(),
+        python_executable=Path(sys.executable),
         default_targets=("tests",),
         trusted_plugins=("pytest_asyncio.plugin",),
     ),
@@ -63,8 +64,9 @@ executor = GovernedToolExecutor(
 )
 ```
 
-Windows and POSIX environments need different virtual-environment executable paths. The path is
-host configuration, never a ToolCall argument.
+Use the active environment's `sys.executable` path without resolving symlinks. On POSIX, resolving
+`.venv/bin/python` can erase virtual-environment identity and select the base interpreter. The path
+is host configuration, never a ToolCall argument.
 
 ## Model and Host Control
 

docs/learning/knowledge-map.md

@@ -558,6 +558,8 @@
   配置，模型不能构造命令。
 - `python -I` 隔离用户 site 与 `PYTHON*` 启动影响；
   `PYTEST_DISABLE_PLUGIN_AUTOLOAD=1` 禁止环境中的 entry-point 插件自动扩展执行面。
+- POSIX 的 `.venv/bin/python` 可能是 symlink；不能对 `sys.executable` 调用
+  `resolve()`，否则会丢失 venv identity，`-I` 下会落到不含 Pytest 的 base Python。
 - `-p no:cacheprovider` 避免 Harness 自己创建 `.pytest_cache`；不代表测试代码只读。
 - model target 经 `WorkspaceBoundary` 解析为现存文件/目录，并在 `--` 后传给 Pytest。
 - 进程状态和报告状态拆成两个 failure domain：测试 exit code 仍可保留，即使 JUnit

docs/learning/progress.md

@@ -482,6 +482,8 @@
   plugins. Profile models are immutable and validated before execution.
 - Fixed argv uses `python -I -m pytest`, disables ambient plugin autoload and Pytest cache writes,
   and places `--` before validated targets.
+- The host preserves the active `sys.executable` path. Resolving a POSIX venv symlink selects the
+  base interpreter and loses environment-local Pytest under isolated startup.
 - Execute remains denied by default. Interactive `ASK` requires independent approval;
   non-interactive `ASK`, rejection, and approval exceptions start no process.
 - Pytest exit classification is separate from JUnit report status. A missing or invalid report

src/mini_code_agent/testing/junit.py

@@ -116,6 +116,15 @@ def parse_junit_report(
 
 
 def _read_report(path: Path, limit: int) -> bytes:
+    try:
+        initial_mode = path.lstat().st_mode
+    except FileNotFoundError:
+        raise _report_error(PytestReportErrorCode.MISSING) from None
+    except OSError:
+        raise _report_error(PytestReportErrorCode.UNSAFE) from None
+    if not stat.S_ISREG(initial_mode):
+        raise _report_error(PytestReportErrorCode.UNSAFE)
+
     flags = os.O_RDONLY | getattr(os, "O_BINARY", 0) | getattr(os, "O_NOFOLLOW", 0)
     try:
         descriptor = os.open(path, flags)

tests/integration/test_governed_pytest_agent.py

@@ -100,7 +100,7 @@ def executor_for(
     runner = PytestRunner(
         root,
         profile=PytestProfile(
-            python_executable=Path(sys.executable).resolve(),
+            python_executable=Path(sys.executable),
             timeout_seconds=30,
             max_failures=5,
         ),