fix(envd): support opt-in descendant signals

packages/envd/internal/services/process/handler/handler.go

@@ -227,6 +227,7 @@ func New(
 			Groups: groups,
 		},
 	}
+	configureProcessGroup(cmd.SysProcAttr, req.GetPty() != nil)
 	applyCgroupFD(cmd.SysProcAttr, cgroupFD, ok)
 
 	resolvedPath, err := permissions.ExpandAndResolve(req.GetProcess().GetCwd(), user, defaults.Workdir)
@@ -462,25 +463,50 @@ func getProcType(req *rpc.StartRequest) cgroups.ProcessType {
 	return cgroups.ProcessTypeUser
 }
 
-func (p *Handler) SendSignal(signal syscall.Signal) error {
-	if signal == syscall.SIGKILL || signal == syscall.SIGTERM {
-		p.outCancel()
+func configureProcessGroup(attr *syscall.SysProcAttr, hasPTY bool) {
+	// PTY startup creates a new session (and therefore a new process group).
+	// Non-PTY commands need an explicit process group so callers can opt into
+	// signalling the command and its descendants without affecting envd.
+	if !hasPTY {
+		attr.Setpgid = true
 	}
+}
 
-	// Re-adopted handler (post live-upgrade): no cmd, signal by stored pid.
-	if p.cmd == nil {
-		if p.pid == 0 {
-			return errors.New("process not started")
+func (p *Handler) SendSignal(signal syscall.Signal, descendants bool) error {
+	pid := int(p.Pid())
+	if pid == 0 {
+		return errors.New("process not started")
+	}
+
+	var err error
+	switch {
+	case descendants:
+		pgid, groupErr := syscall.Getpgid(pid)
+		if groupErr != nil {
+			return fmt.Errorf("get process group for pid %d: %w", pid, groupErr)
+		}
+		if pgid != pid {
+			return fmt.Errorf("process %d does not own a process group", pid)
 		}
 
-		return syscall.Kill(int(p.pid), signal)
+		err = syscall.Kill(-pgid, signal)
+	case p.cmd == nil:
+		// Re-adopted handler (post live-upgrade): no cmd, signal by stored pid.
+		err = syscall.Kill(pid, signal)
+	default:
+		err = p.cmd.Process.Signal(signal)
+	}
+	if err != nil {
+		return err
 	}
 
-	if p.cmd.Process == nil {
-		return errors.New("process not started")
+	// Keep delivering output when signal validation or delivery fails. Once a terminal signal has
+	// actually been sent, stop the pumps promptly instead of waiting for every inherited pipe to close.
+	if signal == syscall.SIGKILL || signal == syscall.SIGTERM {
+		p.outCancel()
 	}
 
-	return p.cmd.Process.Signal(signal)
+	return nil
 }
 
 func (p *Handler) ResizeTty(size *pty.Winsize) error {

packages/envd/internal/services/process/handler/handler_test.go

@@ -2,9 +2,16 @@ package handler
 
 import (
 	"errors"
+	"fmt"
+	"os"
+	"os/exec"
+	"path/filepath"
+	"syscall"
 	"testing"
+	"time"
 
 	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
 )
 
 // The process wrapper must degrade cleanly when the priority helpers are absent
@@ -50,3 +57,87 @@ func TestWrapperPrefix(t *testing.T) {
 		assert.Equal(t, "/bin/ionice -c 1 -n 6 ", ioniceNicePrefix(1, 6, 0, only("ionice")))
 	})
 }
+
+func TestSendSignalTargetsOnlyLeaderByDefault(t *testing.T) {
+	t.Parallel()
+
+	cmd := startCommandGroup(t)
+
+	h := &Handler{cmd: cmd, outCancel: func() {}}
+	t.Cleanup(func() {
+		_ = syscall.Kill(-cmd.Process.Pid, syscall.SIGKILL)
+		_, _ = cmd.Process.Wait()
+	})
+
+	require.NoError(t, h.SendSignal(syscall.SIGTERM, false))
+	state, err := cmd.Process.Wait()
+	require.NoError(t, err)
+	assert.False(t, state.Success())
+	require.NoError(t, syscall.Kill(-cmd.Process.Pid, syscall.Signal(0)), "child should remain in the group")
+}
+
+func TestSendSignalTargetsProcessGroupWhenDescendantsRequested(t *testing.T) {
+	t.Parallel()
+
+	cmd := startCommandGroup(t)
+
+	h := &Handler{cmd: cmd, outCancel: func() {}}
+	t.Cleanup(func() {
+		_ = syscall.Kill(-cmd.Process.Pid, syscall.SIGKILL)
+		_, _ = cmd.Process.Wait()
+	})
+
+	require.NoError(t, h.SendSignal(syscall.SIGTERM, true))
+	state, err := cmd.Process.Wait()
+	require.NoError(t, err)
+	assert.False(t, state.Success())
+	require.Eventually(t, func() bool {
+		return syscall.Kill(-cmd.Process.Pid, syscall.Signal(0)) != nil
+	}, 2*time.Second, 10*time.Millisecond)
+}
+
+func TestSendSignalRejectsDescendantsWithoutOwnedProcessGroup(t *testing.T) {
+	t.Parallel()
+
+	cmd := exec.CommandContext(t.Context(), "sleep", "60")
+	require.NoError(t, cmd.Start())
+	t.Cleanup(func() {
+		_ = cmd.Process.Kill()
+		_, _ = cmd.Process.Wait()
+	})
+
+	outputCancelled := false
+	h := &Handler{cmd: cmd, outCancel: func() { outputCancelled = true }}
+	err := h.SendSignal(syscall.SIGKILL, true)
+	require.ErrorContains(t, err, "does not own a process group")
+	require.NoError(t, cmd.Process.Signal(syscall.Signal(0)), "leader should remain alive")
+	assert.False(t, outputCancelled, "a rejected signal must keep the live process output connected")
+}
+
+func TestConfigureProcessGroup(t *testing.T) {
+	t.Parallel()
+
+	nonPTY := &syscall.SysProcAttr{}
+	configureProcessGroup(nonPTY, false)
+	assert.True(t, nonPTY.Setpgid)
+
+	pty := &syscall.SysProcAttr{}
+	configureProcessGroup(pty, true)
+	assert.False(t, pty.Setpgid, "PTY startup creates its own session")
+}
+
+func startCommandGroup(t *testing.T) *exec.Cmd {
+	t.Helper()
+
+	childReady := filepath.Join(t.TempDir(), "child.pid")
+	cmd := exec.CommandContext(t.Context(), "sh", "-c", fmt.Sprintf("sleep 60 & echo $! > %q; exec sleep 60", childReady))
+	cmd.SysProcAttr = &syscall.SysProcAttr{Setpgid: true}
+	require.NoError(t, cmd.Start())
+	require.Eventually(t, func() bool {
+		_, err := os.Stat(childReady)
+
+		return err == nil
+	}, 2*time.Second, 10*time.Millisecond)
+
+	return cmd
+}

packages/envd/internal/services/process/handler/readopt.go

@@ -151,7 +151,7 @@ func (p *Handler) BeginReaping() {
 
 			select {
 			case <-t.C:
-				_ = p.SendSignal(syscall.SIGKILL)
+				_ = p.SendSignal(syscall.SIGKILL, false)
 			case <-p.outCtx.Done():
 			}
 		}()

packages/envd/internal/services/process/signal.go

@@ -29,7 +29,7 @@ func (s *Service) SendSignal(
 		return nil, connect.NewError(connect.CodeUnimplemented, fmt.Errorf("invalid signal: %s", req.Msg.GetSignal()))
 	}
 
-	err = handler.SendSignal(signal)
+	err = handler.SendSignal(signal, req.Msg.GetDescendants())
 	if err != nil {
 		return nil, connect.NewError(connect.CodeInternal, fmt.Errorf("error sending signal: %w", err))
 	}