diff --git a/interp/api.go b/interp/api.go
index e785d48b..2002689c 100644
--- a/interp/api.go
+++ b/interp/api.go
@@ -126,6 +126,10 @@ type Runner struct {
 	// such as the condition in a [syntax.IfClause].
 	noErrExit bool
 
+	// killProcessGroup controls whether to kill the entire process group
+	// instead of only a single process when terminating subprocesses.
+	killProcessGroup bool
+
 	// The current and last exit statuses. They can only be different if
 	// the interpreter is in the middle of running a statement. In that
 	// scenario, 'exit' is the status for the current statement being run,
@@ -348,10 +352,13 @@ func Dir(path string) RunnerOption {
 func Interactive(enabled bool) RunnerOption {
 	return func(r *Runner) error {
 		r.opts[optExpandAliases] = enabled
+		r.killProcessGroup = !enabled
 		return nil
 	}
 }
 
+
+
 // Params populates the shell options and parameters. For example, Params("-e",
 // "--", "foo") will set the "-e" option and the parameters ["foo"], and
 // Params("+e") will unset the "-e" option and leave the parameters untouched.
@@ -822,6 +829,8 @@ func (r *Runner) Reset() {
 
 		dirStack: r.dirStack[:0],
 		usedNew:  r.usedNew,
+
+		killProcessGroup: r.killProcessGroup,
 	}
 	// Ensure we stop referencing any pointers before we reuse bgProcs.
 	clear(r.bgProcs)
@@ -995,6 +1004,8 @@ func (r *Runner) subshell(background bool) *Runner {
 		lastExit:       r.lastExit,
 
 		origStdout: r.origStdout, // used for process substitutions
+
+		killProcessGroup: r.killProcessGroup,
 	}
 	r2.writeEnv = newOverlayEnviron(r.writeEnv, background)
 	// Funcs are copied, since they might be modified.
diff --git a/interp/handler.go b/interp/handler.go
index a60f6fd8..59294e4e 100644
--- a/interp/handler.go
+++ b/interp/handler.go
@@ -138,23 +138,29 @@ func DefaultExecHandler(killTimeout time.Duration) ExecHandlerFunc {
 			Stdout: hc.Stdout,
 			Stderr: hc.Stderr,
 		}
+		prepareCommand(&cmd, hc.runner.killProcessGroup)
 
 		err = cmd.Start()
 		if err == nil {
-			stopf := context.AfterFunc(ctx, func() {
-				if killTimeout <= 0 || runtime.GOOS == "windows" {
-					_ = cmd.Process.Signal(os.Kill)
-					return
-				}
-				_ = cmd.Process.Signal(os.Interrupt)
-				// TODO: don't sleep in this goroutine if the program
-				// stops itself with the interrupt above.
-				time.Sleep(killTimeout)
-				_ = cmd.Process.Signal(os.Kill)
-			})
-			defer stopf()
-
-			err = cmd.Wait()
+			if err = postStartCommand(&cmd, hc.runner.killProcessGroup); err == nil {
+				stopf := context.AfterFunc(ctx, func() {
+					pg := hc.runner.killProcessGroup
+					if killTimeout <= 0 {
+						_ = killCommand(&cmd, pg)
+						return
+					}
+					_ = interruptCommand(&cmd, pg)
+					// TODO: don't sleep in this goroutine if the program
+					// stops itself with the interrupt above.
+					time.Sleep(killTimeout)
+					_ = killCommand(&cmd, pg)
+				})
+				defer stopf()
+
+				err = cmd.Wait()
+			} else {
+				_ = cmd.Wait()
+			}
 		}
 
 		switch err := err.(type) {
diff --git a/interp/os_notunix.go b/interp/os_notunix.go
index e7ca682a..4c8ce9f2 100644
--- a/interp/os_notunix.go
+++ b/interp/os_notunix.go
@@ -1,13 +1,14 @@
 // Copyright (c) 2017, Andrey Nering <andrey.nering@gmail.com>
 // See LICENSE for licensing information
 
-//go:build !unix
+//go:build !unix && !windows
 
 package interp
 
 import (
 	"context"
 	"fmt"
+	"os/exec"
 
 	"mvdan.cc/sh/v3/syntax"
 )
@@ -56,3 +57,19 @@ type waitStatus struct{}
 
 func (waitStatus) Signaled() bool { return false }
 func (waitStatus) Signal() int    { return 0 }
+
+// prepareCommand is a no-op.
+func prepareCommand(cmd *exec.Cmd, processGroup bool) {}
+
+// postStartCommand is a no-op.
+func postStartCommand(cmd *exec.Cmd, processGroup bool) error { return nil }
+
+// interruptCommand interrupts the process by killing it.
+func interruptCommand(cmd *exec.Cmd, processGroup bool) error {
+	return cmd.Process.Kill()
+}
+
+// killCommand kills the process by killing it.
+func killCommand(cmd *exec.Cmd, processGroup bool) error {
+	return cmd.Process.Kill()
+}
diff --git a/interp/os_unix.go b/interp/os_unix.go
index 214f7e0f..b57224e6 100644
--- a/interp/os_unix.go
+++ b/interp/os_unix.go
@@ -7,6 +7,7 @@ package interp
 
 import (
 	"context"
+	"os/exec"
 	"os/user"
 	"strconv"
 	"syscall"
@@ -46,3 +47,30 @@ func (r *Runner) unTestOwnOrGrp(ctx context.Context, op syntax.UnTestOperator, x
 }
 
 type waitStatus = syscall.WaitStatus
+
+// prepareCommand sets the SysProcAttr for the command.
+// If processGroup is true, a new process group is created.
+func prepareCommand(cmd *exec.Cmd, processGroup bool) {
+	if processGroup {
+		cmd.SysProcAttr = &syscall.SysProcAttr{Setpgid: true}
+	}
+}
+
+// postStartCommand is a no-op on Unix.
+func postStartCommand(cmd *exec.Cmd, processGroup bool) error { return nil }
+
+// interruptCommand sends SIGINT to the process or its process group.
+func interruptCommand(cmd *exec.Cmd, processGroup bool) error {
+	if processGroup {
+		return unix.Kill(-cmd.Process.Pid, unix.SIGINT)
+	}
+	return cmd.Process.Signal(syscall.SIGINT)
+}
+
+// killCommand sends SIGKILL to the process or its process group.
+func killCommand(cmd *exec.Cmd, processGroup bool) error {
+	if processGroup {
+		return unix.Kill(-cmd.Process.Pid, unix.SIGKILL)
+	}
+	return cmd.Process.Kill()
+}
diff --git a/interp/os_windows.go b/interp/os_windows.go
new file mode 100644
index 00000000..53c5294f
--- /dev/null
+++ b/interp/os_windows.go
@@ -0,0 +1,229 @@
+// Copyright (c) 2017, Andrey Nering <andrey.nering@gmail.com>
+// See LICENSE for licensing information
+
+//go:build windows
+
+package interp
+
+import (
+	"context"
+	"fmt"
+	"os"
+	"os/exec"
+	"sync"
+	"syscall"
+	"unsafe"
+
+	"golang.org/x/sys/windows"
+	"mvdan.cc/sh/v3/syntax"
+)
+
+func mkfifo(path string, mode uint32) error {
+	return fmt.Errorf("unsupported")
+}
+
+// access attempts to emulate [unix.Access] on Windows.
+// Windows seems to have a different system of permissions than Unix,
+// so for now just rely on what [io/fs.FileInfo] gives us.
+func (r *Runner) access(ctx context.Context, path string, mode uint32) error {
+	info, err := r.lstat(ctx, path)
+	if err != nil {
+		return err
+	}
+	m := info.Mode()
+	switch mode {
+	case access_R_OK:
+		if m&0o400 == 0 {
+			return fmt.Errorf("file is not readable")
+		}
+	case access_W_OK:
+		if m&0o200 == 0 {
+			return fmt.Errorf("file is not writable")
+		}
+	case access_X_OK:
+		if m&0o100 == 0 {
+			return fmt.Errorf("file is not executable")
+		}
+	}
+	return nil
+}
+
+// unTestOwnOrGrp panics. Under Unix, it implements the -O and -G unary tests,
+// but under Windows, it's unclear how to implement those tests, since Windows
+// doesn't have the concept of a file owner, just ACLs, and it's unclear how
+// to map the one to the other.
+func (r *Runner) unTestOwnOrGrp(ctx context.Context, op syntax.UnTestOperator, x string) bool {
+	panic(fmt.Sprintf("unhandled unary test op: %v", op))
+}
+
+// waitStatus is a no-op on windows.
+type waitStatus struct{}
+
+func (waitStatus) Signaled() bool { return false }
+func (waitStatus) Signal() int    { return 0 }
+
+// jobEntry holds the job object handle and a flag indicating whether it has
+// already been terminated, so that the cleanup goroutine and killCommand /
+// interruptCommand do not race to close the same handle.
+type jobEntry struct {
+	mu          sync.Mutex
+	handle      windows.Handle
+	closed      bool
+}
+
+// terminate terminates the job object and closes the handle exactly once.
+// It is safe to call from multiple goroutines concurrently.
+// When called from the cleanup goroutine after natural process exit, the
+// TerminateJobObject call is a no-op because the job is already empty.
+func (e *jobEntry) terminate(exitCode uint32) error {
+	e.mu.Lock()
+	defer e.mu.Unlock()
+	if e.closed {
+		return nil
+	}
+	e.closed = true
+	err := windows.TerminateJobObject(e.handle, exitCode)
+	windows.CloseHandle(e.handle)
+	return err
+}
+
+// jobHandles maps process IDs to their *jobEntry.
+var jobHandles sync.Map
+
+// ntResumeProcess is the NtResumeProcess syscall from ntdll.dll.
+// Unlike ResumeThread (which requires a thread handle), NtResumeProcess
+// accepts a process handle and resumes all suspended threads in the process.
+// The required access right is PROCESS_SUSPEND_RESUME.
+var ntResumeProcess = windows.NewLazyDLL("ntdll.dll").NewProc("NtResumeProcess")
+
+// prepareCommand sets the SysProcAttr for the command.
+// If processGroup is true, the process is created suspended and will be
+// assigned to a job object in postStartCommand before being resumed.
+func prepareCommand(cmd *exec.Cmd, processGroup bool) {
+	if processGroup {
+		cmd.SysProcAttr = &syscall.SysProcAttr{
+			CreationFlags: windows.CREATE_SUSPENDED | windows.CREATE_NEW_PROCESS_GROUP,
+		}
+	}
+}
+
+// postStartCommand assigns the process to a job object if processGroup is true,
+// then resumes the process. This must be called immediately after cmd.Start().
+//
+// The process is started suspended (via CREATE_SUSPENDED in prepareCommand) to
+// eliminate the race between process start and job object assignment: the process
+// cannot spawn children until it is resumed, so all descendants will inherit the job.
+func postStartCommand(cmd *exec.Cmd, processGroup bool) error {
+	if !processGroup || cmd.Process == nil {
+		return nil
+	}
+
+	pid := cmd.Process.Pid
+
+	// Open the process with the rights needed for job assignment, termination,
+	// and resuming (PROCESS_SUSPEND_RESUME).
+	procHandle, err := windows.OpenProcess(
+		windows.PROCESS_SET_QUOTA|windows.PROCESS_TERMINATE|windows.PROCESS_SUSPEND_RESUME|windows.SYNCHRONIZE,
+		false, uint32(pid))
+	if err != nil {
+		cmd.Process.Kill()
+		return fmt.Errorf("OpenProcess: %w", err)
+	}
+	// procHandle is kept open for the lifetime of postStartCommand; the cleanup
+	// goroutine below takes ownership and closes it when the process exits.
+
+	// Create a job object.
+	job, err := windows.CreateJobObject(nil, nil)
+	if err != nil {
+		windows.CloseHandle(procHandle)
+		cmd.Process.Kill()
+		return fmt.Errorf("CreateJobObject: %w", err)
+	}
+
+	// Set the job to kill all processes when the last job handle is closed.
+	info := windows.JOBOBJECT_EXTENDED_LIMIT_INFORMATION{}
+	info.BasicLimitInformation.LimitFlags = windows.JOB_OBJECT_LIMIT_KILL_ON_JOB_CLOSE
+	_, err = windows.SetInformationJobObject(
+		job,
+		windows.JobObjectExtendedLimitInformation,
+		uintptr(unsafe.Pointer(&info)),
+		uint32(unsafe.Sizeof(info)),
+	)
+	if err != nil {
+		windows.CloseHandle(job)
+		windows.CloseHandle(procHandle)
+		cmd.Process.Kill()
+		return fmt.Errorf("SetInformationJobObject: %w", err)
+	}
+
+	// Assign the process to the job before resuming it, so all child processes
+	// it spawns will also belong to the job.
+	if err = windows.AssignProcessToJobObject(job, procHandle); err != nil {
+		windows.CloseHandle(job)
+		windows.CloseHandle(procHandle)
+		cmd.Process.Kill()
+		return fmt.Errorf("AssignProcessToJobObject: %w", err)
+	}
+
+	entry := &jobEntry{handle: job}
+
+	// Store the job entry before resuming so interruptCommand/killCommand can find it.
+	jobHandles.Store(pid, entry)
+
+	// Resume the suspended process. NtResumeProcess accepts a process handle
+	// directly, unlike ResumeThread which requires a thread handle.
+	// NTSTATUS 0 == STATUS_SUCCESS.
+	//
+	// We must pin procHandle as a uintptr passed to a syscall only during the
+	// call itself; storing it before and using it here is correct because
+	// procHandle is a windows.Handle (uintptr) that we own.
+	if status, _, _ := ntResumeProcess.Call(uintptr(procHandle)); status != 0 {
+		jobHandles.Delete(pid)
+		entry.terminate(1)
+		windows.CloseHandle(procHandle)
+		return fmt.Errorf("NtResumeProcess: NTSTATUS 0x%x", status)
+	}
+
+	// Release the job handle when the process exits naturally.
+	// procHandle was opened with SYNCHRONIZE so we can wait on it directly,
+	// avoiding a second OpenProcess call.
+	go func() {
+		defer windows.CloseHandle(procHandle)
+		windows.WaitForSingleObject(procHandle, windows.INFINITE)
+		if e, ok := jobHandles.LoadAndDelete(pid); ok {
+			e.(*jobEntry).terminate(0)
+		}
+	}()
+	return nil
+}
+
+// interruptCommand sends CTRL_BREAK_EVENT to the process group if processGroup
+// is true, otherwise signals the individual process.
+//
+// Note: processes created with CREATE_NEW_PROCESS_GROUP ignore CTRL_C_EVENT,
+// so CTRL_BREAK_EVENT must be used instead. If the console event fails,
+// the job object is terminated as a fallback.
+func interruptCommand(cmd *exec.Cmd, processGroup bool) error {
+	if processGroup {
+		err := windows.GenerateConsoleCtrlEvent(windows.CTRL_BREAK_EVENT, uint32(cmd.Process.Pid))
+		if err == nil {
+			return nil
+		}
+		// Fall back to terminating the job object if the console event fails.
+		if e, ok := jobHandles.LoadAndDelete(cmd.Process.Pid); ok {
+			return e.(*jobEntry).terminate(1)
+		}
+	}
+	return cmd.Process.Signal(os.Interrupt)
+}
+
+// killCommand terminates the job object if processGroup is true,
+// otherwise kills the individual process.
+func killCommand(cmd *exec.Cmd, processGroup bool) error {
+	if processGroup {
+		if e, ok := jobHandles.LoadAndDelete(cmd.Process.Pid); ok {
+			return e.(*jobEntry).terminate(1)
+		}
+	}
+	return cmd.Process.Kill()
+}