fix: a closed terminal hangs its shell up instead of killing it, so the history is saved; idle shells end with their session, and on agent stop

cmd/agent/main.go

@@ -38,7 +38,11 @@ func main() {
 	ctx, stop := signal.NotifyContext(context.Background(), syscall.SIGINT, syscall.SIGTERM)
 	defer stop()
 
-	if err := agent.New(cfg).Run(ctx); err != nil {
+	err := agent.New(cfg).Run(ctx)
+	// Stopping (an update, a restart): hang the terminals up, so their shells
+	// save their history instead of being killed.
+	agent.CloseTerminals()
+	if err != nil {
 		log.Fatalf("agent: %v", err)
 	}
 }

internal/agent/session.go

@@ -54,6 +54,10 @@ func (a *Agent) bridgeStream(parent context.Context, sid string, rw io.ReadWrite
 	ctx, cancel := context.WithCancel(parent)
 	defer cancel()
 	defer rw.Close()
+	// A session closed from the panel (a tab reloaded, a window shut) ends
+	// the stream at once, not on its next output: an idle shell would live
+	// on until the agent stopped.
+	go func() { <-ctx.Done(); rw.Close() }()
 
 	// Downstream: pull input frames and write them to the stream.
 	go func() {

internal/agent/terminal.go

@@ -4,6 +4,9 @@ import (
 	"fmt"
 	"os"
 	"os/exec"
+	"sync"
+	"syscall"
+	"time"
 
 	"github.com/creack/pty"
 )
@@ -12,6 +15,27 @@ import (
 type PTY struct {
 	cmd    *exec.Cmd
 	f      *os.File
+	exited chan struct{} // closed once the shell has exited
+	once   sync.Once
+}
+
+// hangupGrace is how long a shell gets after the hangup to save its history
+// and exit before it is killed.
+const hangupGrace = 3 * time.Second
+
+// openPTYs are the terminals running now, to close on shutdown.
+var openPTYs sync.Map // *PTY -> struct{}
+
+// CloseTerminals hangs up every open terminal and waits for them, so the
+// shells save their history before the agent exits.
+func CloseTerminals() {
+	var wg sync.WaitGroup
+	openPTYs.Range(func(k, _ any) bool {
+		wg.Add(1)
+		go func() { defer wg.Done(); k.(*PTY).Close() }()
+		return true
+	})
+	wg.Wait()
 }
 
 // StartPTY launches a login shell attached to a new PTY. With user empty (or
@@ -56,7 +80,10 @@ func StartPTY(user string) (*PTY, error) {
 	if err != nil {
 		return nil, err
 	}
-	return &PTY{cmd: cmd, f: f}, nil
+	p := &PTY{cmd: cmd, f: f, exited: make(chan struct{})}
+	go func() { _ = cmd.Wait(); close(p.exited) }()
+	openPTYs.Store(p, struct{}{})
+	return p, nil
 }
 
 func isDir(p string) bool {
@@ -78,12 +105,24 @@ func (p *PTY) Resize(cols, rows uint16) {
 	_ = pty.Setsize(p.f, &pty.Winsize{Rows: rows, Cols: cols})
 }
 
-// Close terminates the shell and releases the PTY.
+// Close hangs the terminal up as a closed window would: the shell gets
+// SIGHUP, saves its history and exits. One that does not within hangupGrace
+// is killed. Safe to call more than once.
 func (p *PTY) Close() error {
-	err := p.f.Close()
-	if p.cmd.Process != nil {
-		_ = p.cmd.Process.Kill()
+	var err error
+	p.once.Do(func() {
+		// Closing the master hangs up the session; the signal to the
+		// shell's process group makes sure it hears, through su as well.
+		err = p.f.Close()
+		pid := p.cmd.Process.Pid
+		_ = syscall.Kill(-pid, syscall.SIGHUP)
+		select {
+		case <-p.exited:
+		case <-time.After(hangupGrace):
+			_ = syscall.Kill(-pid, syscall.SIGKILL)
+			<-p.exited
 		}
-	_ = p.cmd.Wait()
+		openPTYs.Delete(p)
+	})
 	return err
 }