fix: 修复 Windows Codex 进程完成后不自动退出问题 (v1.7.21)

问题根因:
1. messageTimerCh 中阻塞等待 waitCh 导致无限挂起
2. Windows 上 proc.Kill() 只杀主进程不杀子进程

解决方案:
1. 移除 messageTimerCh case 中的阻塞等待
2. 新增 killProcessTree() 使用 taskkill /T /F /PID 终止进程树
3. 修改 terminateCommand/terminateProcess/forwardSignals 使用进程树终止

CHANGELOG.md

@@ -7,6 +7,75 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ---
 
+## [1.7.21] - 2026-01-13
+
+### 🐛 Bug 修复
+
+**彻底修复 Windows 系统 Codex 进程无法自动终止问题**
+
+#### 问题背景
+
+v1.7.18 的修复不完整，Windows 上 Codex 完成后进程仍然卡住：
+- Web UI 显示完成，但 codeagent-wrapper 进程一直不退出
+- 只有手动杀死进程才能继续
+- Mac 没问题，Windows 的 Gemini 也没问题
+
+#### 根本原因分析
+
+1. **进程树问题**：`proc.Kill()` 只终止主进程，Codex 启动的子进程（Node.js workers）仍然运行，持有 stdout handle
+2. **阻塞等待问题**：`messageTimerCh` case 里直接 `waitErr = <-waitCh` 阻塞等待，如果进程没被杀死就永远卡住
+
+#### 核心修复
+
+1. **使用 `taskkill /T` 终止整个进程树**：
+   - 新增 `killProcessTree()` 函数，使用 `taskkill /T /F /PID` 递归杀死所有子进程
+   - `terminateCommand()`、`terminateProcess()`、`forwardSignals()` 在 Windows 上调用进程树终止
+
+2. **移除阻塞等待**：
+   - `messageTimerCh` case 里不再直接阻塞等待 `waitCh`
+   - 让循环继续，下一轮通过 `case waitErr = <-waitCh` 正常退出
+   - 即使 `taskkill` 失败，也不会永远卡住
+
+#### 技术细节
+
+修改文件：`codeagent-wrapper/executor.go`
+
+```go
+// 新增：Windows 进程树终止
+func killProcessTree(pid int) error {
+    if !isWindows() {
+        return nil
+    }
+    cmd := exec.Command("taskkill", "/T", "/F", "/PID", fmt.Sprintf("%d", pid))
+    return cmd.Run()
+}
+
+// 修改：terminateCommand 使用进程树终止
+if isWindows() {
+    if err := killProcessTree(proc.Pid()); err != nil {
+        _ = proc.Kill() // fallback
+    }
+}
+
+// 修改：messageTimerCh case 不再阻塞
+case <-messageTimerCh:
+    // ...terminate logic...
+    // Do NOT block here - let loop continue
+```
+
+#### 影响范围
+
+- ✅ **修复前**：Windows 用户 Codex 完成后进程卡住
+- ✅ **修复后**：所有平台进程正确终止
+- ✅ **向后兼容**：Unix 平台行为不变
+
+#### codeagent-wrapper 版本
+
+- 升级至 v5.6.0
+- 重新编译所有平台二进制文件
+
+---
+
 ## [1.7.20] - 2026-01-13
 
 ### 🐛 Bug 修复

codeagent-wrapper/executor.go

@@ -1126,8 +1126,9 @@ waitLoop:
 					terminated = true
 				}
 			}
-			waitErr = <-waitCh
-			break waitLoop
+			// Do NOT block here waiting for waitCh - let the loop continue
+			// This is critical for Windows where process termination may fail silently
+			// The next iteration will catch waitCh if the process exits
 		case <-completeSeen:
 			completeSeenObserved = true
 			if messageTimer != nil {
@@ -1256,9 +1257,11 @@ func forwardSignals(ctx context.Context, cmd commandRunner, logErrorFn func(stri
 		case sig := <-sigCh:
 			logErrorFn(fmt.Sprintf("Received signal: %v", sig))
 			if proc := cmd.Process(); proc != nil {
-				// Windows does not support SIGTERM - use Kill() directly
+				// Windows does not support SIGTERM - use taskkill /T to kill process tree
 				if isWindows() {
-					_ = proc.Kill()
+					if err := killProcessTree(proc.Pid()); err != nil {
+						_ = proc.Kill()
+					}
 				} else {
 					_ = proc.Signal(syscall.SIGTERM)
 					time.AfterFunc(time.Duration(forceKillDelay.Load())*time.Second, func() {
@@ -1322,6 +1325,26 @@ func (t *forceKillTimer) Stop() {
 	t.stopped.Store(true)
 }
 
+// killProcessTree terminates a process and all its child processes on Windows.
+// Uses taskkill /T /F /PID which recursively kills the entire process tree.
+// This is necessary because Codex CLI may spawn child processes (Node.js workers, etc.)
+// that inherit stdout handles and prevent the parent from exiting cleanly.
+// Returns nil on non-Windows platforms or if the process has already exited.
+func killProcessTree(pid int) error {
+	if !isWindows() {
+		return nil
+	}
+	if pid <= 0 {
+		return nil
+	}
+
+	// taskkill /T = kill process tree, /F = force, /PID = process ID
+	cmd := exec.Command("taskkill", "/T", "/F", "/PID", fmt.Sprintf("%d", pid))
+	cmd.Stdout = nil
+	cmd.Stderr = nil
+	return cmd.Run()
+}
+
 func terminateCommand(cmd commandRunner) *forceKillTimer {
 	if cmd == nil {
 		return nil
@@ -1333,9 +1356,14 @@ func terminateCommand(cmd commandRunner) *forceKillTimer {
 
 	// Windows does not support SIGTERM - it silently fails without terminating the process.
 	// This causes codeagent-wrapper to hang waiting for the process to exit.
-	// On Windows, we directly use Kill() which calls TerminateProcess().
+	// On Windows, we use taskkill /T to kill the entire process tree, not just the main process.
+	// This is critical because Codex CLI spawns child processes that hold stdout handles open.
 	if isWindows() {
-		_ = proc.Kill()
+		pid := proc.Pid()
+		if err := killProcessTree(pid); err != nil {
+			// Fallback to direct Kill() if taskkill fails
+			_ = proc.Kill()
+		}
 	} else {
 		_ = proc.Signal(syscall.SIGTERM)
 	}
@@ -1343,7 +1371,11 @@ func terminateCommand(cmd commandRunner) *forceKillTimer {
 	done := make(chan struct{}, 1)
 	timer := time.AfterFunc(time.Duration(forceKillDelay.Load())*time.Second, func() {
 		if p := cmd.Process(); p != nil {
-			_ = p.Kill()
+			if isWindows() {
+				_ = killProcessTree(p.Pid())
+			} else {
+				_ = p.Kill()
+			}
 		}
 		close(done)
 	})
@@ -1360,16 +1392,23 @@ func terminateProcess(cmd commandRunner) *time.Timer {
 		return nil
 	}
 
-	// Windows does not support SIGTERM - use Kill() directly
+	// Windows does not support SIGTERM - use taskkill /T to kill process tree
 	if isWindows() {
-		_ = proc.Kill()
+		pid := proc.Pid()
+		if err := killProcessTree(pid); err != nil {
+			_ = proc.Kill()
+		}
 	} else {
 		_ = proc.Signal(syscall.SIGTERM)
 	}
 
 	return time.AfterFunc(time.Duration(forceKillDelay.Load())*time.Second, func() {
 		if p := cmd.Process(); p != nil {
-			_ = p.Kill()
+			if isWindows() {
+				_ = killProcessTree(p.Pid())
+			} else {
+				_ = p.Kill()
+			}
 		}
 	})
 }

codeagent-wrapper/main.go

@@ -14,7 +14,7 @@ import (
 )
 
 const (
-	version               = "5.5.0"
+	version               = "5.6.0"
 	defaultWorkdir        = "."
 	defaultTimeout        = 7200 // seconds (2 hours)
 	defaultCoverageTarget = 90.0

codeagent-wrapper/main_test.go

@@ -3067,7 +3067,7 @@ func TestVersionFlag(t *testing.T) {
 		}
 	})
 
-	want := "codeagent-wrapper version 5.4.0\n"
+	want := "codeagent-wrapper version 5.6.0\n"
 
 	if output != want {
 		t.Fatalf("output = %q, want %q", output, want)
@@ -3083,7 +3083,7 @@ func TestVersionShortFlag(t *testing.T) {
 		}
 	})
 
-	want := "codeagent-wrapper version 5.4.0\n"
+	want := "codeagent-wrapper version 5.6.0\n"
 
 	if output != want {
 		t.Fatalf("output = %q, want %q", output, want)
@@ -3099,7 +3099,7 @@ func TestVersionLegacyAlias(t *testing.T) {
 		}
 	})
 
-	want := "codex-wrapper version 5.4.0\n"
+	want := "codex-wrapper version 5.6.0\n"
 
 	if output != want {
 		t.Fatalf("output = %q, want %q", output, want)

package.json

@@ -1,6 +1,6 @@
 {
   "name": "ccg-workflow",
-  "version": "1.7.20",
+  "version": "1.7.21",
   "description": "Claude-Codex-Gemini 多模型协作系统 - 智能路由多模型开发工作流",
   "type": "module",
   "packageManager": "pnpm@10.17.1",