Clean up the agent socket and working files when a signal ends the tool (closes #17)

Main ran the command tree on a background context, so SIGINT, SIGTERM or
SIGHUP killed the process before the deferred cleanup ran: the agent's
temporary directory and socket, and the install working directory, were
left behind.

Main now runs the tree on a signal.NotifyContext for those three signals.
A signal cancels the context, which ends the child ssh or sftp started
with exec.CommandContext, and the deferred cleanup then runs. The exit
status after a signal stays 1 unless ssh reported one of its own.

For "ssh to" the child is cancelled with SIGTERM rather than the default
kill, so ssh restores the terminal before it goes.

Model: opus-4-8

README.md

@@ -54,8 +54,14 @@ If none of these is available and standard input is not a terminal, the tool
 refuses and exits with status 1. A mnemonic that fails the BIP-39 checksum is
 refused with a message saying so.
 
+`KEYFUNC_MNEMONIC` and `KEYFUNC_MNEMONIC_COMMAND` are removed from the
+environment before the system `ssh` (`keyfunc ssh to`) and `sftp`
+(`keyfunc ssh install`) are started, so the mnemonic is never handed on to
+them.
+
 Every command takes `--index` / `-n` and `--mnemonic-command`, and has `--help`.
-`keyfunc --version` prints the version set at build time.
+`keyfunc --version` prints the version. `make build` stamps it; a binary
+installed with `go install` reports the module version instead.
 
 ## SSH keys: `keyfunc ssh`
 
@@ -140,7 +146,9 @@ Derives the key, serves it from an SSH agent that runs inside the tool on a unix
 socket in a new private `0700` temporary directory, then runs the system `ssh`
 with `-o IdentityAgent=<that socket>` followed by the host and all remaining
 arguments unchanged. The tool exits with `ssh`'s exit status and removes the
-socket and directory on the way out. The private key is never written to disk.
+socket and directory on the way out. The private key is never written to disk. A
+SIGINT, SIGTERM or SIGHUP ends `ssh` and still removes the socket and directory,
+and the tool then exits with status 1 unless `ssh` reported one of its own.
 
 ## age identities: `keyfunc age`
 

internal/cli/cli.go

@@ -2,9 +2,13 @@
 package cli
 
 import (
+	"context"
 	"errors"
 	"fmt"
 	"os"
+	"os/signal"
+	"runtime/debug"
+	"syscall"
 
 	"git.eeqj.de/sneak/keyfunc/internal/cli/age"
 	"git.eeqj.de/sneak/keyfunc/internal/cli/mnemonic"
@@ -13,20 +17,43 @@ import (
 	"github.com/spf13/cobra"
 )
 
-// Version is what --version prints. The build sets it.
+// devVersion is what Version holds until a build stamps a real one.
+const devVersion = "dev"
+
+// Version is what --version prints. make build stamps it with -ldflags.
 //
 //nolint:gochecknoglobals // set at build time with -ldflags
-var Version = "dev"
+var Version = devVersion
+
+// resolveVersion chooses what --version reports. A value stamped at
+// build time wins. Otherwise, for a binary from go install, the module
+// version recorded in the build info is used, unless that is empty or
+// the "(devel)" of a local build. When neither names a version, the
+// "dev" fallback stays.
+func resolveVersion(stamped string, info *debug.BuildInfo) string {
+	if stamped != devVersion {
+		return stamped
+	}
+
+	if info != nil && info.Main.Version != "" &&
+		info.Main.Version != "(devel)" {
+		return info.Main.Version
+	}
+
+	return devVersion
+}
 
 // Root returns the whole command tree.
 func Root() *cobra.Command {
+	info, _ := debug.ReadBuildInfo()
+
 	root := &cobra.Command{
 		Use:   "keyfunc",
 		Short: "derive key pairs from a BIP-39 mnemonic",
 		Long: "keyfunc turns a BIP-39 mnemonic into key pairs that can " +
 			"be recreated from that mnemonic at any time. The same " +
 			"mnemonic, key type and index always give the same key.",
-		Version:       Version,
+		Version:       resolveVersion(Version, info),
 		SilenceUsage:  true,
 		SilenceErrors: true,
 	}
@@ -42,8 +69,19 @@ func Root() *cobra.Command {
 // status of its own, which "ssh to" uses to hand on the status ssh
 // ended with. ssh has already said whatever it had to say in that
 // case, so nothing more is printed.
+//
+// SIGINT, SIGTERM and SIGHUP cancel the command's context instead of
+// killing the process outright, so the child ssh or sftp ends and the
+// deferred cleanup that removes the agent socket and the install
+// working directory still runs.
 func Main() int {
-	err := Root().Execute()
+	ctx, stop := signal.NotifyContext(
+		context.Background(),
+		syscall.SIGINT, syscall.SIGTERM, syscall.SIGHUP,
+	)
+	defer stop()
+
+	err := Root().ExecuteContext(ctx)
 	if err == nil {
 		return 0
 	}

internal/cli/ssh/install.go

@@ -166,6 +166,7 @@ func session(
 
 	//nolint:gosec // the options are the user's own, meant for sftp
 	command := exec.CommandContext(cmd.Context(), "sftp", argv...)
+	command.Env = childEnv()
 	command.Stdin = strings.NewReader(strings.Join(batch, "\n") + "\n")
 	command.Stdout = &said
 	command.Stderr = &said

internal/cli/ssh/ssh.go

@@ -3,9 +3,12 @@ package ssh
 
 import (
 	"fmt"
+	"os"
+	"strings"
 
 	"git.eeqj.de/sneak/keyfunc/internal/cli/options"
 	"git.eeqj.de/sneak/keyfunc/internal/derive"
+	"git.eeqj.de/sneak/keyfunc/internal/mnemonic"
 	"git.eeqj.de/sneak/keyfunc/internal/sshkey"
 	"github.com/spf13/cobra"
 )
@@ -84,6 +87,26 @@ func write(cmd *cobra.Command, text string) error {
 	return nil
 }
 
+// childEnv is the tool's environment with the mnemonic variables taken
+// out, for the ssh and sftp children it starts. "ssh to" exists so the
+// private key never leaves the tool; the mnemonic, from either variable,
+// must not leave it either.
+func childEnv() []string {
+	environ := os.Environ()
+	kept := make([]string, 0, len(environ))
+
+	for _, entry := range environ {
+		name, _, _ := strings.Cut(entry, "=")
+		if name == mnemonic.Variable || name == mnemonic.CommandVariable {
+			continue
+		}
+
+		kept = append(kept, entry)
+	}
+
+	return kept
+}
+
 // addComment gives a command its comment flag.
 func addComment(cmd *cobra.Command) {
 	cmd.Flags().String(

internal/cli/ssh/to.go

@@ -7,6 +7,7 @@ import (
 	"os"
 	"os/exec"
 	"slices"
+	"syscall"
 
 	"github.com/spf13/cobra"
 )
@@ -70,10 +71,18 @@ func to() *cobra.Command {
 func connect(ctx context.Context, argv []string) error {
 	//nolint:gosec // the arguments are the user's own, meant for ssh
 	command := exec.CommandContext(ctx, "ssh", argv...)
+	command.Env = childEnv()
 	command.Stdin = os.Stdin
 	command.Stdout = os.Stdout
 	command.Stderr = os.Stderr
 
+	// A cancelled context means a signal ended the tool. Send ssh a
+	// SIGTERM rather than the default kill, so it puts the terminal
+	// back the way it found it before it goes.
+	command.Cancel = func() error {
+		return command.Process.Signal(syscall.SIGTERM)
+	}
+
 	err := command.Run()
 	if err == nil {
 		return nil

internal/cli/ssh_test.go

@@ -3,11 +3,14 @@ package cli_test
 import (
 	"bytes"
 	"os"
+	"os/exec"
 	"path/filepath"
 	"slices"
 	"strconv"
 	"strings"
+	"syscall"
 	"testing"
+	"time"
 
 	"git.eeqj.de/sneak/keyfunc/internal/cli"
 	"git.eeqj.de/sneak/keyfunc/internal/cli/ssh"
@@ -15,6 +18,23 @@ import (
 	"github.com/stretchr/testify/require"
 )
 
+// runAsTool, set in the environment of a re-executed test binary, tells
+// TestMain to run the tool through Main rather than the suite, so the
+// signal test can drive the real signal path in a process it can send a
+// signal to.
+const runAsTool = "KEYFUNC_TEST_RUN_AS_TOOL"
+
+// TestMain re-executes the test binary as the tool when runAsTool is
+// set, and otherwise runs the suite. The signal test starts the tool
+// this way, as a subprocess it can signal and watch clean up.
+func TestMain(m *testing.M) {
+	if os.Getenv(runAsTool) == "1" {
+		os.Exit(cli.Main())
+	}
+
+	os.Exit(m.Run())
+}
+
 // The modes the host is supposed to end up with, and the mode the
 // stand-ins need so that they can be run at all.
 const (
@@ -47,6 +67,9 @@ const (
 	keptIn    = "authorized_keys"
 )
 
+// remoteCommand is the command the "to" tests hand ssh after the host.
+const remoteCommand = "uptime"
+
 // The tool's own name, as it stands in the arguments a test hands to
 // Main, the ssh subcommand both commands the tests here drive live
 // under, and the one of those two these tests name most.
@@ -60,13 +83,19 @@ const (
 // an authorized_keys file.
 const keyLine = vectorZero + " keyfunc/ssh/0\n"
 
+// marker is a variable set beside the mnemonic ones and expected to
+// reach the stand-in, so a scrubbed environment is told apart from an
+// empty one.
+const marker = "KEYFUNC_TEST_MARKER"
+
 // installer is a stand-in for the system sftp for the install
 // command. It writes down the arguments and every command of the
-// batch it is given, echoes each command as sftp does, and carries
+// batch it is given, echoes each command as sftp does, writes down its
-// the commands out against a directory standing in for the host's
+// own environment when a test asks for it, and carries the commands out
-// home directory, so that what keyfunc sends can be watched doing its
+// against a directory standing in for the host's home directory, so that
-// work. A command that begins with a dash may fail; any other failure
+// what keyfunc sends can be watched doing its work. A command that
-// ends the session, as it does in sftp's own batch mode.
+// begins with a dash may fail; any other failure ends the session, as it
+// does in sftp's own batch mode.
 //
 // The listing and the two ways a get can fail are worded as the
 // OpenSSH client words them, each naming the path the server expanded.
@@ -81,6 +110,7 @@ const keyLine = vectorZero + " keyfunc/ssh/0\n"
 // draws the warning ssh writes for it, which carries the wording of a
 // missing file into a session that goes on to authenticate.
 const installer = `
+[ -n "$KEYFUNC_TEST_ENVIRONMENT" ] && env > "$KEYFUNC_TEST_ENVIRONMENT"
 previous=
 for argument in "$@"; do
 	printf '%s\n' "$argument" >> "$KEYFUNC_TEST_ARGUMENTS"
@@ -144,9 +174,11 @@ done
 
 // caller is a stand-in for the system ssh for the to command. It
 // writes down the arguments it was given, notes the agent socket if
-// there really is one at the path it was handed, and ends with the
+// there really is one at the path it was handed, writes down its own
-// status the test asked for.
+// environment when a test asks for it, and ends with the status the
+// test asked for.
 const caller = `
+[ -n "$KEYFUNC_TEST_ENVIRONMENT" ] && env > "$KEYFUNC_TEST_ENVIRONMENT"
 for argument in "$@"; do
 	printf '%s\n' "$argument" >> "$KEYFUNC_TEST_ARGUMENTS"
 done
@@ -157,6 +189,18 @@ fi
 exit "$KEYFUNC_TEST_STATUS"
 `
 
+// sleeper is a stand-in for the system ssh that notes the agent socket
+// and then blocks, so a test can cancel the context while it is running
+// and watch the tool take the agent down. The wait ends on its own only
+// as a backstop, well after the test has cancelled and looked.
+const sleeper = `
+socket=${2#IdentityAgent=}
+if [ -S "$socket" ]; then
+	printf '%s\n' "$socket" > "$KEYFUNC_TEST_SOCKET"
+fi
+sleep 5
+`
+
 // pretended is where a stand-in writes down what it was asked to do.
 type pretended struct {
 	// home stands in for the home directory on the host.
@@ -412,7 +456,7 @@ func TestSSHIsPointedAtTheAgentAndItsStatusIsHandedOn(t *testing.T) {
 
 	arguments, noted := pretendCall(t)
 
-	_, err := execute(t, subcommand, "to", host, "uptime")
+	_, err := execute(t, subcommand, "to", host, remoteCommand)
 
 	var passed ssh.StatusError
 
@@ -421,7 +465,7 @@ func TestSSHIsPointedAtTheAgentAndItsStatusIsHandedOn(t *testing.T) {
 
 	given := recorded(t, arguments)
 	require.Equal(t, "-o", given[0])
-	require.Equal(t, []string{host, "uptime"}, given[2:])
+	require.Equal(t, []string{host, remoteCommand}, given[2:])
 
 	// The stand-in wrote the path down only because there really was
 	// a socket there while it ran.
@@ -439,11 +483,130 @@ func TestTheToolEndsWithTheStatusSSHEndedWith(t *testing.T) {
 
 	t.Cleanup(func() { os.Args = given })
 
-	os.Args = []string{tool, subcommand, "to", host, "uptime"}
+	os.Args = []string{tool, subcommand, "to", host, remoteCommand}
 
 	require.Equal(t, failingStatus, cli.Main())
 }
 
+func TestASignalTakesTheAgentDirectoryDown(t *testing.T) {
+	t.Setenv(mnemonic.Variable, example())
+
+	// The three signals the tool handles, checked one after another.
+	signals := []struct {
+		name   string
+		signal os.Signal
+	}{
+		{"SIGTERM", syscall.SIGTERM},
+		{"SIGINT", syscall.SIGINT},
+		{"SIGHUP", syscall.SIGHUP},
+	}
+
+	for _, ending := range signals {
+		signalEndsTheTool(t, ending.name, ending.signal)
+	}
+}
+
+// signalEndsTheTool runs the tool as a subprocess against a stand-in
+// ssh that blocks, waits until the agent is up and ssh is running
+// against it, sends the tool the signal, and requires the agent socket
+// and its directory to be gone once the tool has ended. The subprocess
+// goes through Main and its signal handling, so with that handling
+// removed the signal kills the tool outright, no deferred cleanup runs,
+// the directory is left behind, and the check fails.
+func signalEndsTheTool(t *testing.T, name string, signal os.Signal) {
+	t.Helper()
+
+	noted := filepath.Join(t.TempDir(), "socket")
+	t.Setenv("KEYFUNC_TEST_SOCKET", noted)
+	standIn(t, "ssh", sleeper)
+
+	//nolint:gosec // the binary is this test's own, re-run as the tool
+	command := exec.CommandContext(
+		t.Context(), os.Args[0], subcommand, "to", host, remoteCommand,
+	)
+
+	command.Env = append(os.Environ(), runAsTool+"=1")
+	require.NoError(t, command.Start())
+
+	// The stand-in notes the socket only once the agent is up and ssh
+	// is running against it, so this is where the signal lands.
+	socket := waitForSocket(t, noted)
+
+	require.NoError(t, command.Process.Signal(signal))
+	waitForTool(t, name, command)
+
+	// The signal ended the tool, and its deferred cleanup still ran:
+	// the agent socket and its directory are gone.
+	require.NoDirExists(t, filepath.Dir(socket), name)
+}
+
+// waitForTool waits for the subprocess to end, and fails the test if it
+// does not end in time.
+func waitForTool(t *testing.T, name string, command *exec.Cmd) {
+	t.Helper()
+
+	done := make(chan error, 1)
+	go func() { done <- command.Wait() }()
+
+	select {
+	case <-done:
+	case <-time.After(10 * time.Second):
+		t.Fatalf("the tool did not end after %s", name)
+	}
+}
+
+// waitForSocket waits for the stand-in to write down the agent socket
+// and gives back the path, which means the agent is up and ssh is
+// running against it.
+func waitForSocket(t *testing.T, noted string) string {
+	t.Helper()
+
+	var socket string
+
+	require.Eventually(t, func() bool {
+		content, err := os.ReadFile(noted) //nolint:gosec // test path
+		if err != nil {
+			return false
+		}
+
+		socket = strings.TrimSpace(string(content))
+
+		return socket != ""
+	}, 5*time.Second, 5*time.Millisecond)
+
+	return socket
+}
+
+func TestTheMnemonicIsNotHandedToSFTP(t *testing.T) {
+	t.Setenv(mnemonic.CommandVariable, "echo "+example())
+	t.Setenv(mnemonic.Variable, example())
+	t.Setenv(marker, "reaches the stand-in")
+
+	pretendHost(t)
+	environment := recordEnvironment(t)
+
+	install(t, host)
+
+	mnemonicWithheld(t, read(t, environment))
+}
+
+func TestTheMnemonicIsNotHandedToSSH(t *testing.T) {
+	t.Setenv(mnemonic.CommandVariable, "echo "+example())
+	t.Setenv(mnemonic.Variable, example())
+	t.Setenv(marker, "reaches the stand-in")
+
+	pretendCall(t)
+	environment := recordEnvironment(t)
+
+	_, err := execute(t, subcommand, "to", host, remoteCommand)
+
+	var passed ssh.StatusError
+
+	require.ErrorAs(t, err, &passed)
+
+	mnemonicWithheld(t, read(t, environment))
+}
+
 // pretendHost puts the install stand-in on the path and gives back the
 // places it writes to.
 func pretendHost(t *testing.T) pretended {
@@ -592,6 +755,28 @@ func standIn(t *testing.T, name, body string) {
 	)
 }
 
+// recordEnvironment asks the stand-in to write its environment down and
+// gives back the file it writes it to.
+func recordEnvironment(t *testing.T) string {
+	t.Helper()
+
+	path := filepath.Join(t.TempDir(), "environment")
+	t.Setenv("KEYFUNC_TEST_ENVIRONMENT", path)
+
+	return path
+}
+
+// mnemonicWithheld requires that neither mnemonic variable reached the
+// stand-in and that the marker set beside them did, so an empty
+// environment does not pass for a scrubbed one.
+func mnemonicWithheld(t *testing.T, environment string) {
+	t.Helper()
+
+	require.NotContains(t, environment, mnemonic.Variable+"=")
+	require.NotContains(t, environment, mnemonic.CommandVariable+"=")
+	require.Contains(t, environment, marker+"=")
+}
+
 // read returns what is in a file.
 func read(t *testing.T, path string) string {
 	t.Helper()

internal/cli/version_internal_test.go

@@ -0,0 +1,37 @@
+package cli
+
+import (
+	"runtime/debug"
+	"testing"
+
+	"github.com/stretchr/testify/require"
+)
+
+func TestResolveVersion(t *testing.T) {
+	t.Parallel()
+
+	release := &debug.BuildInfo{Main: debug.Module{Version: "v1.2.3"}}
+	local := &debug.BuildInfo{Main: debug.Module{Version: "(devel)"}}
+	empty := &debug.BuildInfo{}
+
+	t.Run("stamped value wins over build info", func(t *testing.T) {
+		t.Parallel()
+		require.Equal(t, "v0.1.0", resolveVersion("v0.1.0", release))
+	})
+
+	t.Run("go install reports the module version", func(t *testing.T) {
+		t.Parallel()
+		require.Equal(t, "v1.2.3", resolveVersion(devVersion, release))
+	})
+
+	t.Run("a local build stays dev", func(t *testing.T) {
+		t.Parallel()
+		require.Equal(t, devVersion, resolveVersion(devVersion, local))
+	})
+
+	t.Run("no version anywhere stays dev", func(t *testing.T) {
+		t.Parallel()
+		require.Equal(t, devVersion, resolveVersion(devVersion, empty))
+		require.Equal(t, devVersion, resolveVersion(devVersion, nil))
+	})
+}