Lock the state directory and write vault files atomically (closes #34)

Each command that changes the state directory holds one lock: flock(2)
on `lock` in the state directory, dropped by the kernel if the process
dies, or a process-wide mutex on the in-memory test filesystem. It
covers the state directory, not each vault, because `currentvault`,
`vault create` and cross-vault moves span vaults, and a lock file in a
vault would be deleted by `vault remove` under a waiting command.

Files go through `secret.WriteFileAtomic`; versions, new secrets and
cross-vault copies are built in a temporary directory and renamed into
place; removals rename out of the way first. Left for later: replacing
an unlocker (#71) and deleting
what an interrupted command leaves under a `.tmp-` name
(#75).

Model: opus-5-5

TODO.md

@@ -25,6 +25,41 @@ Bring the repo into policy compliance in one commit:
 
 # Completed Steps
 
+- 2026-10-03: Commands that change the state directory hold one lock
+  (`flock` on `lock` in the state directory; a mutex on the in-memory
+  test filesystem), so concurrent commands no longer lose versions or
+  race on the current pointers. Every file is written through
+  `secret.WriteFileAtomic` (temporary file, sync, rename), so no file
+  is ever half-written and `current`, `currentvault` and
+  `current-unlocker` never go missing. New versions, new secrets and
+  cross-vault copies are built in a temporary directory and renamed
+  into place, and removals rename out of the way first, so a version
+  or secret is never half-added and never half-removed. An
+  interrupted command can still leave:
+    - a broken unlocker, when it was replacing one: an unlocker added
+      under the directory name of an existing one is rewritten file by
+      file. That happens to a passphrase unlocker added to a vault that
+      has one, and to a PGP, keychain or Secure Enclave unlocker added
+      on the same host and day as another of its type
+      (https://git.eeqj.de/sneak/secret/issues/71);
+    - from `vault create` stopped at the passphrase prompt, a new vault
+      with no unlocker that is already the current vault; from `init`
+      stopped there, the default vault with no unlocker;
+    - from an unlocker add stopped before its metadata is written, a
+      directory that `unlocker list` warns about and `unlocker rm`
+      cannot remove;
+    - data under a `.tmp-` name in the state directory: a secret or
+      version being added, or the secret, version, unlocker or vault
+      being removed, encrypted keys included. Nothing deletes it; it
+      must be deleted by hand
+      (https://git.eeqj.de/sneak/secret/issues/75).
+- 2026-10-03: `version rm`, `version promote` and `get --version`
+  accept a version only if it is one of the versions `version list`
+  lists for that secret, compared as typed before any path is built
+  (`secret.VersionExists`), and touch nothing otherwise. An empty
+  `--version` is rejected instead of meaning the current version.
+  Before, `secret version rm x ../../..` deleted the whole vault,
+  `secret version rm x ..` the secret, and `.` or `""` every version.
 - 2026-10-03: Key material is wiped on every exit: `Entry()` returns
   the exit code after its deferred `memguard.Purge()` has run, and only
   `main` calls `os.Exit`. SIGINT and SIGTERM go through memguard's
@@ -112,8 +147,6 @@ Bring the repo into policy compliance in one commit:
       version.go:155); age secret key held in a plain string in
       cli/crypto.go:86,91,113; private keys exposed via buffer.Bytes()
       to GPGEncryptFunc and EncryptWithPassphrase.
-    - Race conditions: no file locking in vault/secrets.go:142-176;
-      non-atomic writes can leave the vault inconsistent.
     - Input validation: no maximum secret size (DoS).
     - Timing attacks: bytes.Equal passphrase compare (cli/init.go:
       209-216); non-constant-time public key compare (vault.go:95-100).

internal/cli/crypto.go

@@ -70,9 +70,44 @@ func newDecryptCmd() *cobra.Command {
 	)
 }
 
+// storeNewEncryptionKey generates an age secret key and stores it as the
+// named secret, holding the state directory lock while it does. It fails
+// with vault.ErrSecretExists if another command stored the secret first.
+// The caller must destroy the returned buffer.
+func (cli *Instance) storeNewEncryptionKey(
+	vlt *vault.Vault, secretName string,
+) (*memguard.LockedBuffer, error) {
+	release, err := vault.LockStateDir(cli.fs, cli.stateDir)
+	if err != nil {
+		return nil, err
+	}
+	defer release()
+
+	identity, err := age.GenerateX25519Identity()
+	if err != nil {
+		return nil, fmt.Errorf("failed to generate age key: %w", err)
+	}
+
+	// Store the generated key directly in a secure buffer
+	secureBuffer := memguard.NewBufferFromBytes([]byte(identity.String()))
+
+	err = vlt.AddSecret(secretName, secureBuffer, false)
+	if err != nil {
+		secureBuffer.Destroy()
+
+		return nil, fmt.Errorf("failed to store age key: %w", err)
+	}
+
+	return secureBuffer, nil
+}
+
 // resolveEncryptionKey returns a secure buffer holding the age secret key
 // for the named secret, generating and storing a new key if the secret
-// does not exist. The caller must destroy the returned buffer.
+// does not exist. The caller must destroy the returned buffer. Only storing
+// a new key takes the state directory lock, so that reading an existing key
+// works on a read-only state directory and keeps no other command waiting
+// at the passphrase prompt, and Encrypt streams its input and output
+// unlocked.
 func (cli *Instance) resolveEncryptionKey(
 	vlt *vault.Vault, secretName string,
 ) (*memguard.LockedBuffer, error) {
@@ -85,23 +120,11 @@ func (cli *Instance) resolveEncryptionKey(
 	}
 
 	if !exists {
-		// Secret doesn't exist, generate new age key and store it
-		identity, err := age.GenerateX25519Identity()
-		if err != nil {
-			return nil, fmt.Errorf("failed to generate age key: %w", err)
+		key, err := cli.storeNewEncryptionKey(vlt, secretName)
+		if !errors.Is(err, vault.ErrSecretExists) {
+			return key, err
 		}
-
-		// Store the generated key directly in a secure buffer
-		secureBuffer := memguard.NewBufferFromBytes([]byte(identity.String()))
-
-		err = vlt.AddSecret(secretName, secureBuffer, false)
-		if err != nil {
-			secureBuffer.Destroy()
-
-			return nil, fmt.Errorf("failed to store age key: %w", err)
-		}
-
-		return secureBuffer, nil
+		// Another command stored the key since the check above: read it
 	}
 
 	// Secret exists, get the age secret key from it

internal/cli/generate.go

@@ -155,6 +155,12 @@ func (cli *Instance) GenerateSecret(
 		return fmt.Errorf("failed to generate random secret: %w", err)
 	}
 
+	release, err := vault.LockStateDir(cli.fs, cli.stateDir)
+	if err != nil {
+		return err
+	}
+	defer release()
+
 	// Store the secret in the vault
 	vlt, err := vault.GetCurrentVault(cli.fs, cli.stateDir)
 	if err != nil {

internal/cli/init.go

@@ -103,8 +103,21 @@ func (cli *Instance) setupDefaultVault(
 	return vlt, ltIdentity, nil
 }
 
-// Init initializes the secret manager
+// Init initializes the secret manager, holding the state directory lock
+// while initialize runs
 func (cli *Instance) Init(cmd *cobra.Command) error {
+	release, err := vault.LockStateDir(cli.fs, cli.stateDir)
+	if err != nil {
+		return err
+	}
+	defer release()
+
+	return cli.initialize(cmd)
+}
+
+// initialize creates the state directory, the default vault and its first
+// unlocker
+func (cli *Instance) initialize(cmd *cobra.Command) error {
 	secret.Debug("Starting secret manager initialization")
 
 	// Create state directory

internal/cli/integration_test.go

@@ -829,6 +829,14 @@ func test09GetSpecificVersion(t *testing.T, tempDir, testMnemonic string, runSec
 
 	require.NoError(t, err, "get current version should succeed")
 	assert.Equal(t, "newpassword456", strings.TrimSpace(output), "should return new secret value without --version")
+
+	// An empty --version is not a version; it does not mean the current one
+	output, err = runSecretWithEnv(map[string]string{
+		secret.EnvMnemonic: testMnemonic,
+	}, "get", "--version", "", "database/password")
+
+	require.Error(t, err, "get with an empty version should fail")
+	assert.Contains(t, output, "version '' not found", "should reject the empty version")
 }
 
 func test10PromoteVersion(t *testing.T, tempDir, testMnemonic string, runSecret func(...string) (string, error), runSecretWithEnv func(map[string]string, ...string) (string, error)) {

internal/cli/lock_test.go

@@ -0,0 +1,529 @@
+//nolint:testpackage // sets the unexported fields of Instance
+package cli
+
+import (
+	"io"
+	"os"
+	"path/filepath"
+	"strconv"
+	"strings"
+	"sync"
+	"testing"
+	"time"
+
+	"git.eeqj.de/sneak/secret/internal/secret"
+	"git.eeqj.de/sneak/secret/internal/vault"
+	"github.com/awnumar/memguard"
+	"github.com/spf13/afero"
+	"github.com/spf13/cobra"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+const (
+	// lockWait is how long a test waits for something that must happen
+	// once the lock is free.
+	lockWait = 10 * time.Second
+
+	// heldWait is how long a test watches a command that must wait for the
+	// lock. A command that takes no lock changes the state directory well
+	// within it.
+	heldWait = 100 * time.Millisecond
+
+	// testPassphrase protects the passphrase unlockers the tests create.
+	testPassphrase = "test-passphrase"
+
+	// testInput is a file outside the state directory that commands read.
+	testInput = "/input"
+)
+
+// lockInBackground starts taking the state directory lock and returns a
+// channel that delivers the function releasing it once it has been taken.
+func lockInBackground(t *testing.T, fs afero.Fs) <-chan func() {
+	t.Helper()
+
+	taken := make(chan func(), 1)
+
+	go func() {
+		release, err := vault.LockStateDir(fs, testStateDir)
+		if assert.NoError(t, err) {
+			taken <- release
+		}
+	}()
+
+	return taken
+}
+
+// addAtOnce runs one add of the secret name per value, all at once, and
+// returns their errors.
+func addAtOnce(
+	fs afero.Fs, stateDir, name string, force bool, values []string,
+) []error {
+	errs := make(chan error, len(values))
+
+	for _, value := range values {
+		go func() {
+			cli := NewCLIInstanceWithStateDir(fs, stateDir)
+			cli.cmd = &cobra.Command{}
+			cli.cmd.SetIn(strings.NewReader(value))
+
+			errs <- cli.AddSecret(name, force)
+		}()
+	}
+
+	results := make([]error, 0, len(values))
+	for range values {
+		results = append(results, <-errs)
+	}
+
+	return results
+}
+
+// numbered returns count distinct values starting with prefix.
+func numbered(prefix string, count int) []string {
+	values := make([]string, 0, count)
+	for i := range count {
+		values = append(values, prefix+"-"+strconv.Itoa(i))
+	}
+
+	return values
+}
+
+// TestConcurrentAddsKeepEveryVersion runs adds of one secret at once, on
+// the in-memory and on the real filesystem. Without the state directory
+// lock, adds of a new secret all find it absent and replace each other, and
+// forced adds read the same highest version number and overwrite each
+// other's version. With it they behave as if run one after another.
+//
+//nolint:paralleltest // t.Setenv forbids parallel subtests
+func TestConcurrentAddsKeepEveryVersion(t *testing.T) {
+	t.Setenv(secret.EnvMnemonic, testMnemonic)
+
+	const adds = 8
+
+	for _, tc := range []struct {
+		name     string
+		fs       afero.Fs
+		stateDir string
+	}{
+		{"memory", afero.NewMemMapFs(), testStateDir},
+		{"real", afero.NewOsFs(), t.TempDir()},
+	} {
+		t.Run(tc.name, func(t *testing.T) {
+			_, err := vault.CreateVault(tc.fs, tc.stateDir, "default")
+			require.NoError(t, err)
+
+			// One add creates the secret; the others find that it exists
+			created := 0
+
+			for _, err := range addAtOnce(tc.fs, tc.stateDir, "shared", false,
+				numbered("create", adds)) {
+				if err == nil {
+					created++
+				} else {
+					require.ErrorIs(t, err, vault.ErrSecretExists)
+				}
+			}
+
+			require.Equal(t, 1, created, "exactly one add creates the secret")
+
+			// Every forced add stores a version of its own
+			for _, err := range addAtOnce(tc.fs, tc.stateDir, "shared", true,
+				numbered("force", adds)) {
+				require.NoError(t, err)
+			}
+
+			vlt, err := vault.GetCurrentVault(tc.fs, tc.stateDir)
+			require.NoError(t, err)
+
+			vaultDir, err := vlt.GetDirectory()
+			require.NoError(t, err)
+
+			versions, err := secret.ListVersions(tc.fs,
+				filepath.Join(vaultDir, "secrets.d", "shared"))
+			require.NoError(t, err)
+			require.Len(t, versions, adds+1, "one version per successful add")
+
+			values := make(map[string]bool, len(versions))
+
+			for _, version := range versions {
+				value, err := vlt.GetSecretVersion("shared", version)
+				require.NoError(t, err)
+
+				values[string(value)] = true
+			}
+
+			assert.Len(t, values, adds+1, "every add stored its own value")
+		})
+	}
+}
+
+// readNotifier passes reads through to Reader and closes reading at the
+// first one.
+type readNotifier struct {
+	io.Reader
+
+	reading chan struct{}
+	once    sync.Once
+}
+
+func (r *readNotifier) Read(p []byte) (int, error) {
+	r.once.Do(func() { close(r.reading) })
+
+	return r.Reader.Read(p)
+}
+
+// TestEncryptPipedIntoAdd runs `secret encrypt key | secret add name` in
+// one process, starting encrypt once add is reading its input. Had add
+// taken the state directory lock before reading, it would hold the lock
+// while waiting for encrypt's output, and encrypt would wait for the lock
+// to store its key: neither would finish.
+func TestEncryptPipedIntoAdd(t *testing.T) {
+	t.Setenv(secret.EnvMnemonic, testMnemonic)
+
+	fs := afero.NewMemMapFs()
+	_, err := vault.CreateVault(fs, testStateDir, "default")
+	require.NoError(t, err)
+	require.NoError(t, afero.WriteFile(fs, testInput, []byte("piped"), 0o600))
+
+	pipeReader, pipeWriter := io.Pipe()
+	// If the test gives up, this makes add's read fail, so that both
+	// commands return and release the lock the other tests use
+	t.Cleanup(func() { _ = pipeReader.Close() })
+
+	const commands = 2
+
+	input := &readNotifier{Reader: pipeReader, reading: make(chan struct{})}
+	results := make(chan error, commands)
+
+	go func() {
+		add := NewCLIInstanceWithStateDir(fs, testStateDir)
+		add.cmd = &cobra.Command{}
+		add.cmd.SetIn(input)
+
+		results <- add.AddSecret("encrypted", false)
+	}()
+
+	go func() {
+		<-input.reading
+
+		encrypt := NewCLIInstanceWithStateDir(fs, testStateDir)
+		encrypt.cmd = &cobra.Command{}
+		encrypt.cmd.SetOut(pipeWriter)
+
+		err := encrypt.Encrypt("key", testInput, "")
+		// Ends add's input, as the end of the pipe does
+		_ = pipeWriter.CloseWithError(err)
+
+		results <- err
+	}()
+
+	timeout := time.After(lockWait)
+
+	for range commands {
+		select {
+		case err := <-results:
+			require.NoError(t, err)
+		case <-timeout:
+			t.Fatal("secret encrypt piped into secret add never finished")
+		}
+	}
+}
+
+// TestFailedCommandReleasesLock checks that a command failing after it
+// took the state directory lock leaves the lock free for the next command.
+func TestFailedCommandReleasesLock(t *testing.T) {
+	t.Parallel()
+
+	fs := afero.NewMemMapFs()
+	cli := NewCLIInstanceWithStateDir(fs, testStateDir)
+
+	// Fails once it holds the lock: there is no current vault
+	err := cli.RemoveSecret(&cobra.Command{}, "missing", false)
+	require.Error(t, err)
+
+	select {
+	case release := <-lockInBackground(t, fs):
+		release()
+	case <-time.After(lockWait):
+		t.Fatal("the failed command left the state directory locked")
+	}
+}
+
+// stateDirModTimes returns the modification time of every file and
+// directory under the test state directory. Any change a command makes, even
+// rewriting a file with the same content, changes it.
+func stateDirModTimes(t *testing.T, fs afero.Fs) map[string]int64 {
+	t.Helper()
+
+	modTimes := make(map[string]int64)
+
+	err := afero.Walk(fs, testStateDir,
+		func(path string, info os.FileInfo, err error) error {
+			if err != nil {
+				return err
+			}
+
+			modTimes[path] = info.ModTime().UnixNano()
+
+			return nil
+		})
+	require.NoError(t, err)
+
+	return modTimes
+}
+
+// setupEveryCommand makes what each command in
+// TestChangingCommandsWaitForLock needs: the current vault "default" with
+// two versions of "test/secret", the vault "other" without a long-term key,
+// for vault import, and the file testInput. If withUnlocker is set, it also
+// gives "default" a passphrase unlocker, which is slow. It returns the older
+// version and the unlocker's ID.
+func setupEveryCommand(
+	t *testing.T, fs afero.Fs, withUnlocker bool,
+) (string, string) {
+	t.Helper()
+
+	other, err := vault.CreateVault(fs, testStateDir, "other")
+	require.NoError(t, err)
+
+	otherDir, err := other.GetDirectory()
+	require.NoError(t, err)
+	require.NoError(t, fs.Remove(filepath.Join(otherDir, "pub.age")))
+
+	vlt, err := vault.CreateVault(fs, testStateDir, "default")
+	require.NoError(t, err)
+
+	addTestSecret(t, vlt, []byte("older"), false)
+	addTestSecret(t, vlt, []byte("newer"), true)
+
+	vaultDir, err := vlt.GetDirectory()
+	require.NoError(t, err)
+
+	versions, err := secret.ListVersions(fs,
+		filepath.Join(vaultDir, "secrets.d", "test%secret"))
+	require.NoError(t, err)
+	require.Len(t, versions, 2)
+
+	unlockerID := ""
+
+	if withUnlocker {
+		passphrase := memguard.NewBufferFromBytes([]byte(testPassphrase))
+		defer passphrase.Destroy()
+
+		unlocker, err := vlt.CreatePassphraseUnlocker(passphrase)
+		require.NoError(t, err)
+
+		unlockerID = unlocker.GetID()
+	}
+
+	require.NoError(t, afero.WriteFile(fs, testInput, []byte("input"), 0o600))
+
+	// Newest first
+	return versions[1], unlockerID
+}
+
+// requireWaitsForLock runs a command, given what setupEveryCommand made,
+// while holding the state directory lock. The command must neither finish
+// nor change anything while the lock is held, and must succeed once it is
+// released.
+func requireWaitsForLock(
+	t *testing.T,
+	withUnlocker bool,
+	run func(cli *Instance, olderVersion, unlockerID string) error,
+) {
+	t.Helper()
+
+	fs := afero.NewMemMapFs()
+	olderVersion, unlockerID := setupEveryCommand(t, fs, withUnlocker)
+	before := stateDirModTimes(t, fs)
+
+	release, err := vault.LockStateDir(fs, testStateDir)
+	require.NoError(t, err)
+
+	// Released at most once, and also if the test fails while holding it,
+	// so that later tests can take it
+	release = sync.OnceFunc(release)
+	defer release()
+
+	cli := NewCLIInstanceWithStateDir(fs, testStateDir)
+	cli.cmd = &cobra.Command{}
+	cli.cmd.SetIn(strings.NewReader("value"))
+	cli.cmd.SetOut(io.Discard)
+
+	done := make(chan error, 1)
+
+	go func() { done <- run(cli, olderVersion, unlockerID) }()
+
+	select {
+	case err := <-done:
+		t.Fatalf("finished while the lock was held, with error %v", err)
+	case <-time.After(heldWait):
+	}
+
+	assert.Equal(t, before, stateDirModTimes(t, fs),
+		"changed the state directory while the lock was held")
+
+	release()
+
+	select {
+	case err := <-done:
+		require.NoError(t, err)
+	case <-time.After(lockWait):
+		t.Fatal("did not finish once the lock was released")
+	}
+}
+
+// TestChangingCommandsWaitForLock checks that each command that changes the
+// state directory waits for its lock.
+//
+//nolint:paralleltest // t.Setenv forbids parallel subtests
+func TestChangingCommandsWaitForLock(t *testing.T) {
+	t.Setenv(secret.EnvMnemonic, testMnemonic)
+	t.Setenv(secret.EnvUnlockPassphrase, testPassphrase)
+
+	for _, tc := range []struct {
+		name         string
+		withUnlocker bool
+		run          func(cli *Instance, olderVersion, unlockerID string) error
+	}{
+		{"add", false, func(cli *Instance, _, _ string) error {
+			return cli.AddSecret("added", false)
+		}},
+		{"import", false, func(cli *Instance, _, _ string) error {
+			return cli.ImportSecret(cli.cmd, "imported", testInput, false)
+		}},
+		{"generate secret", false, func(cli *Instance, _, _ string) error {
+			return cli.GenerateSecret(cli.cmd, "generated", 16, "base58", false)
+		}},
+		{"encrypt", false, func(cli *Instance, _, _ string) error {
+			return cli.Encrypt("key", testInput, "")
+		}},
+		{"rm", false, func(cli *Instance, _, _ string) error {
+			return cli.RemoveSecret(cli.cmd, "test/secret", false)
+		}},
+		{"move", false, func(cli *Instance, _, _ string) error {
+			return cli.MoveSecret(cli.cmd, "test/secret", "moved", false)
+		}},
+		{"version promote", false, func(cli *Instance, olderVersion, _ string) error {
+			return cli.PromoteVersion(cli.cmd, "test/secret", olderVersion)
+		}},
+		{"version rm", false, func(cli *Instance, olderVersion, _ string) error {
+			return cli.RemoveVersion(cli.cmd, "test/secret", olderVersion)
+		}},
+		{"vault create", false, func(cli *Instance, _, _ string) error {
+			return cli.CreateVault(cli.cmd, "created")
+		}},
+		{"vault select", false, func(cli *Instance, _, _ string) error {
+			return cli.SelectVault(cli.cmd, "other")
+		}},
+		{"vault import", false, func(cli *Instance, _, _ string) error {
+			return cli.VaultImport(cli.cmd, "other")
+		}},
+		{"vault rm", false, func(cli *Instance, _, _ string) error {
+			return cli.RemoveVault(cli.cmd, "other", false)
+		}},
+		{"unlocker add", false, func(cli *Instance, _, _ string) error {
+			return cli.UnlockersAdd("passphrase", cli.cmd)
+		}},
+		{"unlocker rm", true, func(cli *Instance, _, unlockerID string) error {
+			return cli.UnlockersRemove(unlockerID, true, cli.cmd)
+		}},
+		{"unlocker select", true, func(cli *Instance, _, unlockerID string) error {
+			return cli.UnlockerSelect(unlockerID)
+		}},
+		{"init", false, func(cli *Instance, _, _ string) error {
+			return cli.Init(cli.cmd)
+		}},
+	} {
+		t.Run(tc.name, func(t *testing.T) {
+			requireWaitsForLock(t, tc.withUnlocker, tc.run)
+		})
+	}
+}
+
+// TestEncryptWithExistingKeyTakesNoLock checks that secret encrypt with a
+// key that already exists, which only reads the state directory, finishes
+// while another command holds the state directory lock.
+func TestEncryptWithExistingKeyTakesNoLock(t *testing.T) {
+	t.Setenv(secret.EnvMnemonic, testMnemonic)
+
+	fs := afero.NewMemMapFs()
+	_, err := vault.CreateVault(fs, testStateDir, "default")
+	require.NoError(t, err)
+	require.NoError(t, afero.WriteFile(fs, testInput, []byte("input"), 0o600))
+
+	encrypt := NewCLIInstanceWithStateDir(fs, testStateDir)
+	encrypt.cmd = &cobra.Command{}
+	encrypt.cmd.SetOut(io.Discard)
+
+	// Stores the key
+	require.NoError(t, encrypt.Encrypt("key", testInput, ""))
+
+	release, err := vault.LockStateDir(fs, testStateDir)
+	require.NoError(t, err)
+	// Also frees a waiting encrypt if the test fails, so that it releases
+	// the lock the other tests use
+	defer release()
+
+	done := make(chan error, 1)
+
+	go func() { done <- encrypt.Encrypt("key", testInput, "") }()
+
+	select {
+	case err := <-done:
+		require.NoError(t, err)
+	case <-time.After(lockWait):
+		t.Fatal("secret encrypt with an existing key waited for the lock")
+	}
+}
+
+// TestEncryptStreamsUnlocked checks that secret encrypt has released the
+// state directory lock by the time it writes its output. Holding it while
+// streaming would stall every other changing command for as long as the
+// stream lasts, and forever when the other end of the pipe is one of them.
+func TestEncryptStreamsUnlocked(t *testing.T) {
+	t.Setenv(secret.EnvMnemonic, testMnemonic)
+
+	fs := afero.NewMemMapFs()
+	_, err := vault.CreateVault(fs, testStateDir, "default")
+	require.NoError(t, err)
+	require.NoError(t, afero.WriteFile(fs, testInput, []byte("streamed"), 0o600))
+
+	outputReader, outputWriter := io.Pipe()
+	done := make(chan error, 1)
+
+	go func() {
+		encrypt := NewCLIInstanceWithStateDir(fs, testStateDir)
+		encrypt.cmd = &cobra.Command{}
+		encrypt.cmd.SetOut(outputWriter)
+
+		err := encrypt.Encrypt("key", testInput, "")
+		_ = outputWriter.CloseWithError(err)
+
+		done <- err
+	}()
+
+	// The first byte of output: encrypt is streaming now, and blocked
+	// writing until it is read
+	_, err = io.ReadFull(outputReader, make([]byte, 1))
+	require.NoError(t, err)
+
+	taken := lockInBackground(t, fs)
+
+	select {
+	case release := <-taken:
+		release()
+	case <-time.After(lockWait):
+		// Let encrypt finish, so that it releases the lock, then free it
+		// again for the tests that follow
+		_, _ = io.Copy(io.Discard, outputReader)
+
+		(<-taken)()
+		t.Fatal("secret encrypt held the lock while streaming")
+	}
+
+	_, err = io.Copy(io.Discard, outputReader)
+	require.NoError(t, err)
+	require.NoError(t, <-done)
+}

internal/cli/path_traversal_test.go

@@ -1,6 +1,7 @@
 package cli_test
 
 import (
+	"fmt"
 	"maps"
 	"os"
 	"slices"
@@ -119,13 +120,13 @@ func newFsFromSnapshot(t *testing.T, tree map[string]string) afero.Fs {
 }
 
 // requireRejectedAndUnchanged runs a command on a copy of the state
-// directory recorded in before. It requires exactly the error
-// vault.ValidateSecretName gives for the rejected name, so that a later
-// check rejecting the name does not count, and everything under the state
-// directory as it was: the error alone proves nothing, since it could come
-// after the vault had already been deleted.
+// directory recorded in before. It requires an error with exactly the
+// message of want, so that a later check rejecting the argument does not
+// count, and everything under the state directory as it was: the error
+// alone proves nothing, since it could come after the vault had already
+// been deleted.
 func requireRejectedAndUnchanged(
-	t *testing.T, before map[string]string, rejected string,
+	t *testing.T, before map[string]string, want error,
 	run func(c *cli.Instance) error,
 ) {
 	t.Helper()
@@ -135,8 +136,7 @@ func requireRejectedAndUnchanged(
 	err := run(cli.NewCLIInstanceWithStateDir(fs, testStateDir))
 
 	require.Equal(t, before, snapshotStateDir(t, fs))
-	require.ErrorIs(t, err, vault.ErrInvalidSecretName)
-	require.EqualError(t, err, vault.ValidateSecretName(rejected).Error())
+	require.EqualError(t, err, want.Error())
 }
 
 // TestInvalidSecretNameLeavesVaultsUnchanged is a regression test for
@@ -229,11 +229,92 @@ func TestInvalidSecretNameLeavesVaultsUnchanged(t *testing.T) {
 
 	for _, tt := range tests {
 		t.Run(tt.command, func(t *testing.T) {
-			requireRejectedAndUnchanged(t, before, tt.rejected, tt.run)
+			requireRejectedAndUnchanged(t, before, vault.ValidateSecretName(tt.rejected), tt.run)
 		})
 	}
 }
 
+// TestInvalidVersionLeavesVaultsUnchanged is a regression test for
+// https://git.eeqj.de/sneak/secret/issues/67, where
+// `secret version rm x ../../..` deleted the whole vault,
+// `secret version rm x ..` the secret x, and `secret version rm x .` or
+// `secret version rm x ""` every version of x. A version argument is
+// accepted only if it is one of the versions `secret version list` lists.
+//
+//nolint:paralleltest // newTwoVaultFs uses t.Setenv
+func TestInvalidVersionLeavesVaultsUnchanged(t *testing.T) {
+	before := snapshotStateDir(t, newTwoVaultFs(t))
+
+	cmd := &cobra.Command{}
+
+	commands := []struct {
+		command string
+		run     func(c *cli.Instance, version string) error
+	}{
+		{"version rm x", func(c *cli.Instance, version string) error {
+			return c.RemoveVersion(cmd, "x", version)
+		}},
+		{"version promote x", func(c *cli.Instance, version string) error {
+			return c.PromoteVersion(cmd, "x", version)
+		}},
+		{"get x --version", func(c *cli.Instance, version string) error {
+			return c.GetSecretWithVersion(cmd, "x", version)
+		}},
+	}
+
+	for _, tt := range commands {
+		for _, version := range []string{"", ".", "..", "../../..", "a/b"} {
+			t.Run(fmt.Sprintf("%s %q", tt.command, version), func(t *testing.T) {
+				want := fmt.Errorf("version '%s' %w '%s'",
+					version, vault.ErrVersionNotFound, "x")
+				requireRejectedAndUnchanged(t, before, want,
+					func(c *cli.Instance) error { return tt.run(c, version) })
+			})
+		}
+	}
+}
+
+// TestRemoveVersionRemovesOnlyThatVersion checks that `secret version rm`
+// with a version that is not the current one removes that version and
+// changes nothing else.
+//
+//nolint:paralleltest // newTwoVaultFs uses t.Setenv
+func TestRemoveVersionRemovesOnlyThatVersion(t *testing.T) {
+	fs := newTwoVaultFs(t)
+
+	vlt, err := vault.GetCurrentVault(fs, testStateDir)
+	require.NoError(t, err)
+
+	// A second version of "x" becomes the current one.
+	err = vlt.AddSecret("x", memguard.NewBufferFromBytes([]byte("new")), true)
+	require.NoError(t, err)
+
+	secretDir := testStateDir + "/vaults.d/default/secrets.d/x"
+	versions, err := secret.ListVersions(fs, secretDir)
+	require.NoError(t, err)
+	require.Len(t, versions, 2)
+
+	// ListVersions lists the newest version first.
+	oldDir := secretDir + "/versions/" + versions[1] + "/"
+	before := snapshotStateDir(t, fs)
+	require.Contains(t, before, oldDir)
+
+	c := cli.NewCLIInstanceWithStateDir(fs, testStateDir)
+	err = c.RemoveVersion(&cobra.Command{}, "x", versions[1])
+	require.NoError(t, err)
+
+	// Expected: the state as before without everything under oldDir.
+	want := map[string]string{}
+
+	for path, content := range before {
+		if !strings.HasPrefix(path, oldDir) {
+			want[path] = content
+		}
+	}
+
+	require.Equal(t, want, snapshotStateDir(t, fs))
+}
+
 // TestMoveToVaultNameRenamesInCurrentVault checks that `secret mv x work`,
 // where "work" is also the name of a vault, renames the secret "x" to "work"
 // in the current vault and changes nothing else.

internal/cli/secrets.go

@@ -109,6 +109,12 @@ func newGetCmd() *cobra.Command {
 				return fmt.Errorf("failed to initialize CLI: %w", err)
 			}
 
+			// Without --version, get the current version. A given
+			// --version is checked as typed, so an empty one is rejected.
+			if !cmd.Flags().Changed("version") {
+				return cli.GetSecret(cmd, args[0])
+			}
+
 			return cli.GetSecretWithVersion(cmd, args[0], version)
 		},
 	}
@@ -377,6 +383,15 @@ func (cli *Instance) AddSecret(secretName string, force bool) error {
 	valueBuffer := combineBuffers(buffers, totalSize)
 	defer valueBuffer.Destroy()
 
+	// Locked only now that stdin has been read: in `secret encrypt key |
+	// secret add name`, holding the lock while reading would leave each
+	// command waiting for the other.
+	release, err := vault.LockStateDir(cli.fs, cli.stateDir)
+	if err != nil {
+		return err
+	}
+	defer release()
+
 	// Add the secret to the vault
 	secret.Debug("Calling vault.AddSecret", "secret_name", secretName,
 		"value_length", valueBuffer.Size(), "force", force)
@@ -393,12 +408,32 @@ func (cli *Instance) AddSecret(secretName string, force bool) error {
 	return nil
 }
 
-// GetSecret retrieves and prints a secret from the current vault
+// GetSecret retrieves and prints the current version of a secret
 func (cli *Instance) GetSecret(cmd *cobra.Command, secretName string) error {
-	return cli.GetSecretWithVersion(cmd, secretName, "")
+	secret.Debug("GetSecret called", "secretName", secretName)
+
+	// Store the command for output
+	cli.cmd = cmd
+
+	// Get current vault
+	vlt, err := vault.GetCurrentVault(cli.fs, cli.stateDir)
+	if err != nil {
+		return err
+	}
+
+	value, err := vlt.GetSecret(secretName)
+	if err != nil {
+		return err
+	}
+
+	// Print the secret value to stdout
+	_, _ = cli.Print(string(value))
+
+	return nil
 }
 
-// GetSecretWithVersion retrieves and prints a specific version of a secret
+// GetSecretWithVersion retrieves and prints a specific version of a secret.
+// The version must be one of the secret's versions.
 func (cli *Instance) GetSecretWithVersion(
 	cmd *cobra.Command, secretName string, version string,
 ) error {
@@ -417,13 +452,7 @@ func (cli *Instance) GetSecretWithVersion(
 	}
 
 	// Get the secret value
-	var value []byte
-	if version == "" {
-		value, err = vlt.GetSecret(secretName)
-	} else {
-		value, err = vlt.GetSecretVersion(secretName, version)
-	}
-
+	value, err := vlt.GetSecretVersion(secretName, version)
 	if err != nil {
 		secret.Debug("Failed to get secret", "error", err)
 
@@ -640,6 +669,14 @@ func (cli *Instance) ImportSecret(
 	valueBuffer := combineBuffers(buffers, totalSize)
 	defer valueBuffer.Destroy()
 
+	// Locked only now that the file has been read, as in AddSecret: the
+	// file may be a pipe written by another secret command.
+	release, err := vault.LockStateDir(cli.fs, cli.stateDir)
+	if err != nil {
+		return err
+	}
+	defer release()
+
 	// Store the secret in the vault
 	err = vlt.AddSecret(secretName, valueBuffer, force)
 	if err != nil {
@@ -659,6 +696,12 @@ func (cli *Instance) RemoveSecret(cmd *cobra.Command, secretName string, _ bool)
 		return err
 	}
 
+	release, err := vault.LockStateDir(cli.fs, cli.stateDir)
+	if err != nil {
+		return err
+	}
+	defer release()
+
 	// Get current vault
 	currentVlt, err := vault.GetCurrentVault(cli.fs, cli.stateDir)
 	if err != nil {
@@ -693,7 +736,7 @@ func (cli *Instance) RemoveSecret(cmd *cobra.Command, secretName string, _ bool)
 	}
 
 	// Remove the secret directory
-	err = cli.fs.RemoveAll(secretDir)
+	err = secret.RemoveDirAtomic(cli.fs, secretDir)
 	if err != nil {
 		return fmt.Errorf("failed to remove secret: %w", err)
 	}
@@ -704,9 +747,24 @@ func (cli *Instance) RemoveSecret(cmd *cobra.Command, secretName string, _ bool)
 	return nil
 }
 
-// MoveSecret moves or renames a secret (within or across vaults)
+// MoveSecret moves or renames a secret (within or across vaults), holding
+// the state directory lock.
 func (cli *Instance) MoveSecret(
 	cmd *cobra.Command, source, dest string, force bool,
+) error {
+	release, err := vault.LockStateDir(cli.fs, cli.stateDir)
+	if err != nil {
+		return err
+	}
+	defer release()
+
+	return cli.moveSecret(cmd, source, dest, force)
+}
+
+// moveSecret does the work of MoveSecret. The caller holds the state
+// directory lock.
+func (cli *Instance) moveSecret(
+	cmd *cobra.Command, source, dest string, force bool,
 ) error {
 	// Parse source and destination
 	srcVaultName, srcSecretName, srcQualified := ParseVaultSecretRef(source)
@@ -813,7 +871,7 @@ func (cli *Instance) moveSecretWithinVault(
 			return fmt.Errorf("secret '%s' %w", dest, errSecretExistsNoForce)
 		}
 
-		err = cli.fs.RemoveAll(destDir)
+		err = secret.RemoveDirAtomic(cli.fs, destDir)
 		if err != nil {
 			return fmt.Errorf("failed to remove existing destination: %w", err)
 		}
@@ -894,7 +952,7 @@ func (cli *Instance) moveSecretCrossVault(
 	}
 
 	// Delete source secret
-	err = cli.fs.RemoveAll(srcSecretDir)
+	err = secret.RemoveDirAtomic(cli.fs, srcSecretDir)
 	if err != nil {
 		// Copy succeeded but delete failed - warn but don't fail
 		cmd.Printf("Warning: copied secret but failed to remove source: %v\n", err)

internal/cli/unlockers.go

@@ -534,6 +534,12 @@ func (cli *Instance) printUnlockersTable(unlockers []UnlockerInfo) error {
 
 // UnlockersAdd adds a new unlocker
 func (cli *Instance) UnlockersAdd(unlockerType string, cmd *cobra.Command) error {
+	release, err := vault.LockStateDir(cli.fs, cli.stateDir)
+	if err != nil {
+		return err
+	}
+	defer release()
+
 	switch unlockerType {
 	case unlockerTypePassphrase:
 		return cli.addPassphraseUnlocker(cmd)
@@ -714,6 +720,12 @@ func (cli *Instance) addPGPUnlocker(cmd *cobra.Command) error {
 func (cli *Instance) UnlockersRemove(
 	unlockerID string, force bool, cmd *cobra.Command,
 ) error {
+	release, err := vault.LockStateDir(cli.fs, cli.stateDir)
+	if err != nil {
+		return err
+	}
+	defer release()
+
 	// Get current vault
 	vlt, err := vault.GetCurrentVault(cli.fs, cli.stateDir)
 	if err != nil {
@@ -763,6 +775,12 @@ func (cli *Instance) UnlockersRemove(
 
 // UnlockerSelect selects an unlocker as current
 func (cli *Instance) UnlockerSelect(unlockerID string) error {
+	release, err := vault.LockStateDir(cli.fs, cli.stateDir)
+	if err != nil {
+		return err
+	}
+	defer release()
+
 	// Get current vault
 	vlt, err := vault.GetCurrentVault(cli.fs, cli.stateDir)
 	if err != nil {

internal/cli/vault.go

@@ -267,6 +267,12 @@ func resolvePassphrase() (*memguard.LockedBuffer, error) {
 func (cli *Instance) CreateVault(cmd *cobra.Command, name string) error {
 	secret.Debug("Creating new vault", "name", name, "state_dir", cli.stateDir)
 
+	release, err := vault.LockStateDir(cli.fs, cli.stateDir)
+	if err != nil {
+		return err
+	}
+	defer release()
+
 	// Get or prompt for mnemonic
 	var mnemonicStr string
 
@@ -354,7 +360,13 @@ func (cli *Instance) CreateVault(cmd *cobra.Command, name string) error {
 
 // SelectVault selects a vault as the current one
 func (cli *Instance) SelectVault(cmd *cobra.Command, name string) error {
-	err := vault.SelectVault(cli.fs, cli.stateDir, name)
+	release, err := vault.LockStateDir(cli.fs, cli.stateDir)
+	if err != nil {
+		return err
+	}
+	defer release()
+
+	err = vault.SelectVault(cli.fs, cli.stateDir, name)
 	if err != nil {
 		return err
 	}
@@ -442,8 +454,21 @@ func updateVaultImportMetadata(
 	return nil
 }
 
-// VaultImport imports a mnemonic into a specific vault
+// VaultImport imports a mnemonic into a specific vault, holding the state
+// directory lock while importMnemonic runs
 func (cli *Instance) VaultImport(cmd *cobra.Command, vaultName string) error {
+	release, err := vault.LockStateDir(cli.fs, cli.stateDir)
+	if err != nil {
+		return err
+	}
+	defer release()
+
+	return cli.importMnemonic(cmd, vaultName)
+}
+
+// importMnemonic gives the vault a long-term key derived from the mnemonic
+// and a passphrase unlocker
+func (cli *Instance) importMnemonic(cmd *cobra.Command, vaultName string) error {
 	secret.Debug("Importing mnemonic into vault",
 		"vault_name", vaultName, "state_dir", cli.stateDir)
 
@@ -478,7 +503,7 @@ func (cli *Instance) VaultImport(cmd *cobra.Command, vaultName string) error {
 	secret.Debug("Storing long-term public key",
 		"pubkey", ltPublicKey, "vault_dir", vaultDir)
 
-	err = afero.WriteFile(cli.fs, pubKeyPath, []byte(ltPublicKey), secret.FilePerms)
+	err = secret.WriteFileAtomic(cli.fs, pubKeyPath, []byte(ltPublicKey))
 	if err != nil {
 		return fmt.Errorf("failed to store long-term public key: %w", err)
 	}
@@ -577,6 +602,12 @@ func (cli *Instance) switchAwayFromVault(
 
 // RemoveVault removes a vault with safety checks
 func (cli *Instance) RemoveVault(cmd *cobra.Command, name string, force bool) error {
+	release, err := vault.LockStateDir(cli.fs, cli.stateDir)
+	if err != nil {
+		return err
+	}
+	defer release()
+
 	// Get list of all vaults
 	vaults, err := vault.ListVaults(cli.fs, cli.stateDir)
 	if err != nil {
@@ -626,7 +657,7 @@ func (cli *Instance) RemoveVault(cmd *cobra.Command, name string, force bool) er
 	}
 
 	// Remove the vault directory
-	err = cli.fs.RemoveAll(vaultDir)
+	err = secret.RemoveDirAtomic(cli.fs, vaultDir)
 	if err != nil {
 		return fmt.Errorf("failed to remove vault directory: %w", err)
 	}

internal/cli/version.go

@@ -249,6 +249,12 @@ func (cli *Instance) PromoteVersion(
 		return err
 	}
 
+	release, err := vault.LockStateDir(cli.fs, cli.stateDir)
+	if err != nil {
+		return err
+	}
+	defer release()
+
 	// Get current vault
 	vlt, err := vault.GetCurrentVault(cli.fs, cli.stateDir)
 	if err != nil {
@@ -265,9 +271,7 @@ func (cli *Instance) PromoteVersion(
 	secretDir := filepath.Join(vaultDir, "secrets.d", encodedName)
 
 	// Check if version exists
-	versionDir := filepath.Join(secretDir, "versions", version)
-
-	exists, err := afero.DirExists(cli.fs, versionDir)
+	exists, err := secret.VersionExists(cli.fs, secretDir, version)
 	if err != nil {
 		return fmt.Errorf("failed to check if version exists: %w", err)
 	}
@@ -297,6 +301,12 @@ func (cli *Instance) RemoveVersion(
 		return err
 	}
 
+	release, err := vault.LockStateDir(cli.fs, cli.stateDir)
+	if err != nil {
+		return err
+	}
+	defer release()
+
 	// Get current vault
 	vlt, err := vault.GetCurrentVault(cli.fs, cli.stateDir)
 	if err != nil {
@@ -323,9 +333,7 @@ func (cli *Instance) RemoveVersion(
 	}
 
 	// Check if version exists
-	versionDir := filepath.Join(secretDir, "versions", version)
-
-	exists, err = afero.DirExists(cli.fs, versionDir)
+	exists, err = secret.VersionExists(cli.fs, secretDir, version)
 	if err != nil {
 		return fmt.Errorf("failed to check if version exists: %w", err)
 	}
@@ -348,7 +356,9 @@ func (cli *Instance) RemoveVersion(
 	}
 
 	// Remove the version directory
-	err = cli.fs.RemoveAll(versionDir)
+	versionDir := filepath.Join(secretDir, "versions", version)
+
+	err = secret.RemoveDirAtomic(cli.fs, versionDir)
 	if err != nil {
 		return fmt.Errorf("failed to remove version: %w", err)
 	}

internal/cli/version_test.go

@@ -276,8 +276,8 @@ func TestGetSecretWithVersion(t *testing.T) {
 	var buf bytes.Buffer
 	cmd.SetOut(&buf)
 
-	// Test getting current version (empty version string)
-	err = cli.GetSecretWithVersion(cmd, "test/secret", "")
+	// Test getting the current version
+	err = cli.GetSecret(cmd, "test/secret")
 	require.NoError(t, err)
 	assert.Equal(t, "version-2", buf.String())
 

internal/secret/atomic.go

@@ -0,0 +1,86 @@
+package secret
+
+import (
+	"fmt"
+	"path/filepath"
+
+	"github.com/spf13/afero"
+)
+
+// WriteFileAtomic replaces the file at path with data so that a reader, or
+// a crash at any moment, finds either the old content or the new, never a
+// partial file. The data goes into a temporary file that afero.TempFile
+// creates with mode 0600 in the same directory (a rename is only atomic
+// within one filesystem), is synced to disk, and is renamed over path. The
+// temporary file is removed if any step fails.
+func WriteFileAtomic(fs afero.Fs, path string, data []byte) error {
+	tmp, err := afero.TempFile(fs, filepath.Dir(path),
+		"."+filepath.Base(path)+".tmp-*")
+	if err != nil {
+		return fmt.Errorf("failed to create temporary file for %s: %w", path, err)
+	}
+
+	_, err = tmp.Write(data)
+	if err == nil {
+		err = tmp.Sync()
+	}
+
+	closeErr := tmp.Close()
+	if err == nil {
+		err = closeErr
+	}
+
+	if err == nil {
+		err = fs.Rename(tmp.Name(), path)
+	}
+
+	if err != nil {
+		_ = fs.Remove(tmp.Name())
+
+		return fmt.Errorf("failed to write %s: %w", path, err)
+	}
+
+	return nil
+}
+
+// TempDirFor creates an empty temporary directory in which to build the
+// directory target before renaming it into place, or into which to move
+// target before deleting it. It is made in target's grandparent: on the
+// same filesystem, so the rename is atomic, and outside target's parent,
+// the directory that is listed to find vaults, secrets, versions and
+// unlockers, so one left behind by a crash is never taken for one of them.
+// Its name leaves out target's, which may already be as long as a file name
+// can be.
+func TempDirFor(fs afero.Fs, target string) (string, error) {
+	dir, err := afero.TempDir(fs, filepath.Dir(filepath.Dir(target)), ".tmp-")
+	if err != nil {
+		return "", fmt.Errorf(
+			"failed to create temporary directory for %s: %w", target, err)
+	}
+
+	return dir, nil
+}
+
+// RemoveDirAtomic deletes the directory dir so that it disappears in one
+// rename: dir is moved into a new directory from TempDirFor, which is then
+// deleted. A crash part-way leaves only that temporary directory behind.
+func RemoveDirAtomic(fs afero.Fs, dir string) error {
+	tmp, err := TempDirFor(fs, dir)
+	if err != nil {
+		return err
+	}
+
+	err = fs.Rename(dir, filepath.Join(tmp, filepath.Base(dir)))
+	if err != nil {
+		_ = fs.Remove(tmp)
+
+		return fmt.Errorf("failed to remove %s: %w", dir, err)
+	}
+
+	err = fs.RemoveAll(tmp)
+	if err != nil {
+		return fmt.Errorf("failed to remove %s: %w", dir, err)
+	}
+
+	return nil
+}

internal/secret/atomic_test.go

@@ -0,0 +1,672 @@
+package secret_test
+
+import (
+	"errors"
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+
+	"filippo.io/age"
+	"git.eeqj.de/sneak/secret/internal/secret"
+	"git.eeqj.de/sneak/secret/internal/vault"
+	"github.com/awnumar/memguard"
+	"github.com/spf13/afero"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+var errInjected = errors.New("injected failure")
+
+// The kinds of change hookFs passes to before.
+const (
+	opCreate = "create"
+	opOpen   = "open"
+	opSync   = "sync"
+	opMkdir  = "mkdir"
+	opRemove = "remove"
+	opRename = "rename"
+)
+
+// currentFile is the file in a secret's directory that names its current
+// version.
+const currentFile = "current"
+
+// unlockerMetadataFile is the file a new unlocker writes last.
+const unlockerMetadataFile = "unlocker-metadata.json"
+
+// unlockerPassphrase protects the passphrase unlockers the tests create.
+//
+//nolint:gosec // G101: test data, not a real credential
+const unlockerPassphrase = "unlocker passphrase"
+
+// hookFs passes every call through to Fs, but first calls before for each
+// call that changes the filesystem, and for each Sync of a file opened
+// through it, with the path it changes (the new path, for Rename). A test
+// uses before to inspect the tree at every point where a crash could stop
+// the code under test, or returns an error from it to make that call fail.
+// If opened is set, OpenFile also tells it the mode it opens each file with.
+type hookFs struct {
+	afero.Fs
+
+	before func(op, path string) error
+	opened func(path string, perm os.FileMode)
+}
+
+// hookFile is a file opened through hookFs.
+type hookFile struct {
+	afero.File
+
+	before func(op, path string) error
+}
+
+func (f hookFile) Sync() error {
+	err := f.before(opSync, f.Name())
+	if err != nil {
+		return err
+	}
+
+	return f.File.Sync()
+}
+
+//nolint:ireturn // implements afero.Fs
+func (h hookFs) Create(name string) (afero.File, error) {
+	err := h.before(opCreate, name)
+	if err != nil {
+		return nil, err
+	}
+
+	file, err := h.Fs.Create(name)
+	if err != nil {
+		return nil, err
+	}
+
+	return hookFile{File: file, before: h.before}, nil
+}
+
+//nolint:ireturn // implements afero.Fs
+func (h hookFs) OpenFile(
+	name string, flag int, perm os.FileMode,
+) (afero.File, error) {
+	err := h.before(opOpen, name)
+	if err != nil {
+		return nil, err
+	}
+
+	if h.opened != nil {
+		h.opened(name, perm)
+	}
+
+	file, err := h.Fs.OpenFile(name, flag, perm)
+	if err != nil {
+		return nil, err
+	}
+
+	return hookFile{File: file, before: h.before}, nil
+}
+
+func (h hookFs) Mkdir(name string, perm os.FileMode) error {
+	err := h.before(opMkdir, name)
+	if err != nil {
+		return err
+	}
+
+	return h.Fs.Mkdir(name, perm)
+}
+
+func (h hookFs) MkdirAll(path string, perm os.FileMode) error {
+	err := h.before(opMkdir, path)
+	if err != nil {
+		return err
+	}
+
+	return h.Fs.MkdirAll(path, perm)
+}
+
+func (h hookFs) Remove(name string) error {
+	err := h.before(opRemove, name)
+	if err != nil {
+		return err
+	}
+
+	return h.Fs.Remove(name)
+}
+
+func (h hookFs) RemoveAll(path string) error {
+	err := h.before(opRemove, path)
+	if err != nil {
+		return err
+	}
+
+	return h.Fs.RemoveAll(path)
+}
+
+func (h hookFs) Rename(oldname, newname string) error {
+	err := h.before(opRename, newname)
+	if err != nil {
+		return err
+	}
+
+	return h.Fs.Rename(oldname, newname)
+}
+
+// testFilesystem is a filesystem to run a test on, with a directory in it
+// to work in.
+type testFilesystem struct {
+	name string
+	open func(t *testing.T) (afero.Fs, string)
+}
+
+// testFilesystems are the in-memory filesystem that most tests use and the
+// real one: every rename-based guarantee is checked on both.
+//
+//nolint:gochecknoglobals // read-only table shared by the tests below
+var testFilesystems = []testFilesystem{
+	{"memory", func(*testing.T) (afero.Fs, string) {
+		return afero.NewMemMapFs(), "/test"
+	}},
+	{"real", func(t *testing.T) (afero.Fs, string) {
+		t.Helper()
+
+		return afero.NewOsFs(), t.TempDir()
+	}},
+}
+
+// dirNames lists the names in dir.
+func dirNames(t *testing.T, fs afero.Fs, dir string) []string {
+	t.Helper()
+
+	entries, err := afero.ReadDir(fs, dir)
+	require.NoError(t, err)
+
+	names := make([]string, 0, len(entries))
+	for _, entry := range entries {
+		names = append(names, entry.Name())
+	}
+
+	return names
+}
+
+// writeLongTermKey gives the test vault under stateDir a new long-term key
+// and returns it.
+func writeLongTermKey(
+	t *testing.T, fs afero.Fs, stateDir string,
+) *age.X25519Identity {
+	t.Helper()
+
+	vault := &MockVersionVault{Name: testVaultName, fs: fs, stateDir: stateDir}
+
+	vaultDir, err := vault.GetDirectory()
+	require.NoError(t, err)
+	require.NoError(t, fs.MkdirAll(vaultDir, 0o700))
+
+	ltIdentity, err := age.GenerateX25519Identity()
+	require.NoError(t, err)
+	require.NoError(t, afero.WriteFile(fs, filepath.Join(vaultDir, "pub.age"),
+		[]byte(ltIdentity.Recipient().String()), 0o600))
+
+	return ltIdentity
+}
+
+// newVaultWithSecret creates the vault name under stateDir from the test
+// mnemonic, with a secret "shared" in it that holds value.
+func newVaultWithSecret(
+	t *testing.T, fs afero.Fs, stateDir, name, value string,
+) *vault.Vault {
+	t.Helper()
+
+	vlt, err := vault.CreateVault(fs, stateDir, name)
+	require.NoError(t, err)
+
+	buffer := memguard.NewBufferFromBytes([]byte(value))
+	defer buffer.Destroy()
+
+	require.NoError(t, vlt.AddSecret("shared", buffer, false))
+
+	return vlt
+}
+
+func TestWriteFileAtomicReplacesFile(t *testing.T) {
+	t.Parallel()
+
+	for _, tfs := range testFilesystems {
+		t.Run(tfs.name, func(t *testing.T) {
+			t.Parallel()
+
+			fs, dir := tfs.open(t)
+			path := filepath.Join(dir, currentFile)
+
+			require.NoError(t, secret.WriteFileAtomic(fs, path, []byte("old")))
+			require.NoError(t, secret.WriteFileAtomic(fs, path, []byte("new")))
+
+			data, err := afero.ReadFile(fs, path)
+			require.NoError(t, err)
+			assert.Equal(t, "new", string(data))
+
+			info, err := fs.Stat(path)
+			require.NoError(t, err)
+			assert.Equal(t, secret.FilePerms, info.Mode().Perm())
+
+			// No temporary file is left next to it
+			assert.Equal(t, []string{currentFile}, dirNames(t, fs, dir))
+		})
+	}
+}
+
+func TestWriteFileAtomicFailureKeepsOldFile(t *testing.T) {
+	t.Parallel()
+
+	for _, tfs := range testFilesystems {
+		t.Run(tfs.name, func(t *testing.T) {
+			t.Parallel()
+
+			base, dir := tfs.open(t)
+			path := filepath.Join(dir, currentFile)
+			require.NoError(t, secret.WriteFileAtomic(base, path, []byte("old")))
+
+			fs := hookFs{Fs: base, before: func(op, _ string) error {
+				if op == opRename {
+					return errInjected
+				}
+
+				return nil
+			}}
+
+			err := secret.WriteFileAtomic(fs, path, []byte("new"))
+			require.ErrorIs(t, err, errInjected)
+
+			data, err := afero.ReadFile(base, path)
+			require.NoError(t, err)
+			assert.Equal(t, "old", string(data))
+
+			// The temporary file is removed again
+			assert.Equal(t, []string{currentFile}, dirNames(t, base, dir))
+		})
+	}
+}
+
+// TestRemoveDirAtomic checks that RemoveDirAtomic deletes nothing where the
+// directory stands, which a crash could stop half-way, and that it leaves
+// nothing behind.
+func TestRemoveDirAtomic(t *testing.T) {
+	t.Parallel()
+
+	for _, tfs := range testFilesystems {
+		t.Run(tfs.name, func(t *testing.T) {
+			t.Parallel()
+
+			base, dir := tfs.open(t)
+			listed := filepath.Join(dir, "secrets.d")
+			target := filepath.Join(listed, "doomed")
+
+			require.NoError(t, base.MkdirAll(filepath.Join(target, "versions"), 0o700))
+			require.NoError(t, secret.WriteFileAtomic(base,
+				filepath.Join(target, currentFile), []byte("20231216.001")))
+
+			fs := hookFs{Fs: base, before: func(op, path string) error {
+				if op == opRemove && strings.HasPrefix(path, target) {
+					t.Errorf("deleted %s where it stands", path)
+				}
+
+				return nil
+			}}
+
+			require.NoError(t, secret.RemoveDirAtomic(fs, target))
+
+			// Gone, and no temporary directory is left in the directory
+			// that is listed or in the one above it
+			assert.Empty(t, dirNames(t, base, listed))
+			assert.Equal(t, []string{"secrets.d"}, dirNames(t, base, dir))
+		})
+	}
+}
+
+// TestLongestNames adds a secret to a vault and removes the vault, both
+// named with 255 bytes, the most a file name may have, on the real
+// filesystem: the temporary directories they use must fit that limit too.
+func TestLongestNames(t *testing.T) {
+	t.Setenv(secret.EnvMnemonic, testMnemonic)
+
+	const longestName = 255
+
+	fs := afero.NewOsFs()
+	name := strings.Repeat("a", longestName)
+
+	vlt, err := vault.CreateVault(fs, t.TempDir(), name)
+	require.NoError(t, err)
+
+	value := memguard.NewBufferFromBytes([]byte("long"))
+	defer value.Destroy()
+
+	require.NoError(t, vlt.AddSecret(name, value, false))
+
+	got, err := vlt.GetSecret(name)
+	require.NoError(t, err)
+	assert.Equal(t, "long", string(got))
+
+	vaultDir, err := vlt.GetDirectory()
+	require.NoError(t, err)
+	require.NoError(t, secret.RemoveDirAtomic(fs, vaultDir))
+	assert.NoDirExists(t, vaultDir)
+}
+
+// TestForcedCopyKeepsDestinationUntilReplaced copies a secret over one in
+// another vault, as a forced move between vaults does, and makes the last
+// step that completes the copy fail. The secret it was to replace must
+// still be there unchanged: it may go only once its replacement is whole.
+//
+//nolint:paralleltest // t.Setenv forbids t.Parallel
+func TestForcedCopyKeepsDestinationUntilReplaced(t *testing.T) {
+	t.Setenv(secret.EnvMnemonic, testMnemonic)
+
+	for _, tfs := range testFilesystems {
+		t.Run(tfs.name, func(t *testing.T) {
+			base, stateDir := tfs.open(t)
+			src := newVaultWithSecret(t, base, stateDir, "source", "new")
+			dest := newVaultWithSecret(t, base, stateDir, "dest", "old")
+
+			// The copy is complete once its current file is written
+			fs := hookFs{Fs: base, before: func(op, path string) error {
+				if op == opRename && filepath.Base(path) == currentFile {
+					return errInjected
+				}
+
+				return nil
+			}}
+
+			err := vault.NewVault(fs, stateDir, "dest").
+				CopySecretAllVersions(src, "shared", "shared", true)
+			require.ErrorIs(t, err, errInjected)
+
+			value, err := dest.GetSecret("shared")
+			require.NoError(t, err)
+			assert.Equal(t, "old", string(value))
+		})
+	}
+}
+
+// TestTempDirsStayOutOfListings adds a version, adds a secret, copies a
+// secret over another and removes one, and checks that none of them makes a
+// directory directly in secrets.d or in a versions directory. Those are
+// listed to find secrets and versions, so a temporary directory made there
+// would be listed while half-built, and one left by a crash would stay.
+//
+//nolint:paralleltest // t.Setenv forbids t.Parallel
+func TestTempDirsStayOutOfListings(t *testing.T) {
+	t.Setenv(secret.EnvMnemonic, testMnemonic)
+
+	for _, tfs := range testFilesystems {
+		t.Run(tfs.name, func(t *testing.T) {
+			base, stateDir := tfs.open(t)
+			newVaultWithSecret(t, base, stateDir, "default", "first")
+
+			fs := hookFs{Fs: base, before: func(op, path string) error {
+				parent := filepath.Base(filepath.Dir(path))
+				if op == opMkdir && (parent == "secrets.d" || parent == "versions") {
+					t.Errorf("made %s where it is listed", path)
+				}
+
+				return nil
+			}}
+			vlt := vault.NewVault(fs, stateDir, "default")
+
+			value := memguard.NewBufferFromBytes([]byte("second"))
+			defer value.Destroy()
+
+			require.NoError(t, vlt.AddSecret("shared", value, true))
+			require.NoError(t, vlt.AddSecret("other", value, false))
+			require.NoError(t, vlt.CopySecretAllVersions(vlt, "shared", "other", true))
+
+			vaultDir, err := vlt.GetDirectory()
+			require.NoError(t, err)
+			require.NoError(t, secret.RemoveDirAtomic(fs,
+				filepath.Join(vaultDir, "secrets.d", "shared")))
+		})
+	}
+}
+
+// TestVersionSaveIsWholeOrAbsent checks, before every change Save makes and
+// once after it returns, that the version directory either does not exist
+// or holds all of its files: a crash at any point leaves no version that
+// cannot be decrypted.
+func TestVersionSaveIsWholeOrAbsent(t *testing.T) {
+	t.Parallel()
+
+	for _, tfs := range testFilesystems {
+		t.Run(tfs.name, func(t *testing.T) {
+			t.Parallel()
+
+			base, stateDir := tfs.open(t)
+			ltIdentity := writeLongTermKey(t, base, stateDir)
+
+			var versionDir string
+
+			checkVersionDir := func(string, string) error {
+				exists, err := afero.DirExists(base, versionDir)
+				require.NoError(t, err)
+
+				if exists {
+					assert.ElementsMatch(t,
+						[]string{"pub.age", "value.age", "priv.age", "metadata.age"},
+						dirNames(t, base, versionDir),
+						"version directory visible before it was complete")
+				}
+
+				return nil
+			}
+
+			fs := hookFs{Fs: base, before: checkVersionDir}
+			vault := &MockVersionVault{Name: testVaultName, fs: fs, stateDir: stateDir}
+			sv := secret.NewVersion(vault, "test/secret", "20231215.001")
+			versionDir = sv.Directory
+
+			value := memguard.NewBufferFromBytes([]byte("whole or nothing"))
+			defer value.Destroy()
+
+			require.NoError(t, sv.Save(value))
+			require.NoError(t, checkVersionDir("", ""))
+
+			got, err := sv.GetValue(ltIdentity)
+			require.NoError(t, err)
+
+			defer got.Destroy()
+
+			assert.Equal(t, "whole or nothing", got.String())
+		})
+	}
+}
+
+// TestVersionSaveFailureLeavesNothing makes the write of the encrypted
+// private key fail, after the value has been written, and checks that
+// neither the version nor its temporary directory is left behind.
+func TestVersionSaveFailureLeavesNothing(t *testing.T) {
+	t.Parallel()
+
+	for _, tfs := range testFilesystems {
+		t.Run(tfs.name, func(t *testing.T) {
+			t.Parallel()
+
+			base, stateDir := tfs.open(t)
+			writeLongTermKey(t, base, stateDir)
+
+			fs := hookFs{Fs: base, before: func(op, path string) error {
+				if op == opRename && filepath.Base(path) == "priv.age" {
+					return errInjected
+				}
+
+				return nil
+			}}
+			vault := &MockVersionVault{Name: testVaultName, fs: fs, stateDir: stateDir}
+			sv := secret.NewVersion(vault, "test/secret", "20231215.001")
+
+			value := memguard.NewBufferFromBytes([]byte("never stored"))
+			defer value.Destroy()
+
+			require.ErrorIs(t, sv.Save(value), errInjected)
+
+			// The secret directory holds only the empty versions directory
+			versionsDir := filepath.Dir(sv.Directory)
+			assert.Equal(t, []string{"versions"},
+				dirNames(t, base, filepath.Dir(versionsDir)))
+			assert.Empty(t, dirNames(t, base, versionsDir))
+		})
+	}
+}
+
+// TestCurrentFilesNeverMissing selects the current version, vault and
+// unlocker again and checks, before each change this makes, that the file
+// naming the current one exists: a reader or a crash never finds it
+// missing.
+//
+//nolint:paralleltest // t.Setenv forbids t.Parallel
+func TestCurrentFilesNeverMissing(t *testing.T) {
+	t.Setenv(secret.EnvMnemonic, testMnemonic)
+
+	for _, tfs := range testFilesystems {
+		t.Run(tfs.name, func(t *testing.T) {
+			base, stateDir := tfs.open(t)
+			vlt := newVaultWithSecret(t, base, stateDir, testVaultName, "value")
+
+			passphrase := memguard.NewBufferFromBytes([]byte(unlockerPassphrase))
+			defer passphrase.Destroy()
+
+			// Created as the current unlocker
+			unlocker, err := vlt.CreatePassphraseUnlocker(passphrase)
+			require.NoError(t, err)
+
+			vaultDir, err := vlt.GetDirectory()
+			require.NoError(t, err)
+
+			secretDir := filepath.Join(vaultDir, "secrets.d", "shared")
+			version, err := secret.GetCurrentVersion(base, secretDir)
+			require.NoError(t, err)
+
+			for _, tc := range []struct {
+				path     string
+				reselect func(fs afero.Fs) error
+			}{
+				{filepath.Join(secretDir, currentFile), func(fs afero.Fs) error {
+					return secret.SetCurrentVersion(fs, secretDir, version)
+				}},
+				{filepath.Join(stateDir, "currentvault"), func(fs afero.Fs) error {
+					return vault.SelectVault(fs, stateDir, testVaultName)
+				}},
+				{filepath.Join(vaultDir, "current-unlocker"), func(fs afero.Fs) error {
+					return vault.NewVault(fs, stateDir, testVaultName).
+						SelectUnlocker(unlocker.GetID())
+				}},
+			} {
+				fs := hookFs{Fs: base, before: func(string, string) error {
+					exists, err := afero.Exists(base, tc.path)
+					require.NoError(t, err)
+					assert.True(t, exists, "%s is missing", filepath.Base(tc.path))
+
+					return nil
+				}}
+
+				require.NoError(t, tc.reselect(fs))
+			}
+		})
+	}
+}
+
+// TestWriteFileAtomicTempFile checks that WriteFileAtomic creates its
+// temporary file with mode 0600, rather than wider and narrowed later, so
+// that no other user can ever read it, and syncs it before renaming it into
+// place, so that a crash cannot leave the file named but its data lost.
+func TestWriteFileAtomicTempFile(t *testing.T) {
+	t.Parallel()
+
+	for _, tfs := range testFilesystems {
+		t.Run(tfs.name, func(t *testing.T) {
+			t.Parallel()
+
+			base, dir := tfs.open(t)
+
+			var modes []os.FileMode
+
+			synced := false
+			fs := hookFs{
+				Fs: base,
+				before: func(op, _ string) error {
+					switch op {
+					case opSync:
+						synced = true
+					case opRename:
+						assert.True(t, synced, "renamed before syncing")
+					}
+
+					return nil
+				},
+				opened: func(_ string, perm os.FileMode) {
+					modes = append(modes, perm)
+				},
+			}
+
+			require.NoError(t, secret.WriteFileAtomic(fs,
+				filepath.Join(dir, currentFile), []byte("new")))
+			assert.Equal(t, []os.FileMode{secret.FilePerms}, modes)
+		})
+	}
+}
+
+// TestPassphraseUnlockerGetsKeyFirst creates a passphrase unlocker in a
+// vault whose long-term key cannot be had: it must fail without writing
+// anything, so that it never leaves a partial unlocker, nor breaks the one
+// it would replace.
+func TestPassphraseUnlockerGetsKeyFirst(t *testing.T) {
+	// No mnemonic, and no current unlocker to get the key from
+	t.Setenv(secret.EnvMnemonic, "")
+
+	base := afero.NewMemMapFs()
+	_, err := vault.CreateVault(base, testVaultStateDir, testVaultName)
+	require.NoError(t, err)
+
+	fs := hookFs{Fs: base, before: func(_, path string) error {
+		t.Errorf("changed %s before getting the long-term key", path)
+
+		return nil
+	}}
+
+	passphrase := memguard.NewBufferFromBytes([]byte(unlockerPassphrase))
+	defer passphrase.Destroy()
+
+	_, err = vault.NewVault(fs, testVaultStateDir, testVaultName).
+		CreatePassphraseUnlocker(passphrase)
+	require.Error(t, err)
+}
+
+// TestPassphraseUnlockerWritesMetadataLast checks that the last file a new
+// passphrase unlocker writes in its directory is its metadata: an unlocker
+// directory without metadata is never used, so one interrupted earlier
+// cannot be.
+func TestPassphraseUnlockerWritesMetadataLast(t *testing.T) {
+	t.Setenv(secret.EnvMnemonic, testMnemonic)
+
+	base := afero.NewMemMapFs()
+	vlt, err := vault.CreateVault(base, testVaultStateDir, testVaultName)
+	require.NoError(t, err)
+
+	vaultDir, err := vlt.GetDirectory()
+	require.NoError(t, err)
+
+	unlockerDir := filepath.Join(vaultDir, "unlockers.d", "passphrase")
+
+	var last string
+
+	fs := hookFs{Fs: base, before: func(_, path string) error {
+		if filepath.Dir(path) == unlockerDir {
+			last = filepath.Base(path)
+		}
+
+		return nil
+	}}
+
+	passphrase := memguard.NewBufferFromBytes([]byte(unlockerPassphrase))
+	defer passphrase.Destroy()
+
+	_, err = vault.NewVault(fs, testVaultStateDir, testVaultName).
+		CreatePassphraseUnlocker(passphrase)
+	require.NoError(t, err)
+	assert.Equal(t, unlockerMetadataFile, last)
+}

internal/secret/keychainunlocker.go

@@ -189,7 +189,7 @@ func (k *KeychainUnlocker) Remove() error {
 
 	// Step 3: Remove directory
 	Debug("Removing keychain unlocker directory", "directory", k.Directory)
-	if err := k.fs.RemoveAll(k.Directory); err != nil {
+	if err := RemoveDirAtomic(k.fs, k.Directory); err != nil {
 		Debug("Failed to remove keychain unlocker directory", "error", err, "directory", k.Directory)
 
 		return fmt.Errorf("failed to remove keychain unlocker directory: %w", err)
@@ -368,7 +368,7 @@ func CreateKeychainUnlocker(fs afero.Fs, stateDir string) (*KeychainUnlocker, er
 	// Step 3: Store age recipient as plaintext
 	ageRecipient := ageIdentity.Recipient().String()
 	recipientPath := filepath.Join(unlockerDir, "pub.txt")
-	if err := afero.WriteFile(fs, recipientPath, []byte(ageRecipient), FilePerms); err != nil {
+	if err := WriteFileAtomic(fs, recipientPath, []byte(ageRecipient)); err != nil {
 		return nil, fmt.Errorf("failed to write age recipient: %w", err)
 	}
 
@@ -384,7 +384,7 @@ func CreateKeychainUnlocker(fs afero.Fs, stateDir string) (*KeychainUnlocker, er
 	}
 
 	agePrivKeyPath := filepath.Join(unlockerDir, "priv.age")
-	if err := afero.WriteFile(fs, agePrivKeyPath, encryptedAgePrivKey, FilePerms); err != nil {
+	if err := WriteFileAtomic(fs, agePrivKeyPath, encryptedAgePrivKey); err != nil {
 		return nil, fmt.Errorf("failed to write encrypted age private key: %w", err)
 	}
 
@@ -403,7 +403,7 @@ func CreateKeychainUnlocker(fs afero.Fs, stateDir string) (*KeychainUnlocker, er
 
 	// Write encrypted long-term private key
 	ltPrivKeyPath := filepath.Join(unlockerDir, "longterm.age")
-	if err := afero.WriteFile(fs, ltPrivKeyPath, encryptedLtPrivKeyToAge, FilePerms); err != nil {
+	if err := WriteFileAtomic(fs, ltPrivKeyPath, encryptedLtPrivKeyToAge); err != nil {
 		return nil, fmt.Errorf("failed to write encrypted long-term private key: %w", err)
 	}
 
@@ -440,9 +440,9 @@ func CreateKeychainUnlocker(fs afero.Fs, stateDir string) (*KeychainUnlocker, er
 		return nil, fmt.Errorf("failed to marshal unlocker metadata: %w", err)
 	}
 
-	if err := afero.WriteFile(fs,
+	if err := WriteFileAtomic(fs,
 		filepath.Join(unlockerDir, "unlocker-metadata.json"),
-		metadataBytes, FilePerms); err != nil {
+		metadataBytes); err != nil {
 		return nil, fmt.Errorf("failed to write unlocker metadata: %w", err)
 	}
 

internal/secret/passphraseunlocker.go

@@ -127,7 +127,7 @@ func (p *PassphraseUnlocker) Remove() error {
 
 	// For passphrase unlockers, we just need to remove the directory
 	// No external resources (like keychain items) to clean up
-	err := p.fs.RemoveAll(p.Directory)
+	err := RemoveDirAtomic(p.fs, p.Directory)
 	if err != nil {
 		return fmt.Errorf("failed to remove passphrase unlocker directory: %w", err)
 	}

internal/secret/pgpunlocker.go

@@ -172,7 +172,7 @@ func (p *PGPUnlocker) GetID() string {
 func (p *PGPUnlocker) Remove() error {
 	// For PGP unlockers, we just need to remove the directory
 	// No external resources (like keychain items) to clean up
-	err := p.fs.RemoveAll(p.Directory)
+	err := RemoveDirAtomic(p.fs, p.Directory)
 	if err != nil {
 		return fmt.Errorf("failed to remove PGP unlocker directory: %w", err)
 	}
@@ -275,7 +275,7 @@ func CreatePGPUnlocker(
 	ageRecipient := ageIdentity.Recipient().String()
 	recipientPath := filepath.Join(unlockerDir, "pub.txt")
 
-	err = afero.WriteFile(fs, recipientPath, []byte(ageRecipient), FilePerms)
+	err = WriteFileAtomic(fs, recipientPath, []byte(ageRecipient))
 	if err != nil {
 		return nil, fmt.Errorf("failed to write age recipient: %w", err)
 	}
@@ -298,7 +298,7 @@ func CreatePGPUnlocker(
 	// Write encrypted long-term private key
 	ltPrivKeyPath := filepath.Join(unlockerDir, "longterm.age")
 
-	err = afero.WriteFile(fs, ltPrivKeyPath, encryptedLtPrivKeyToAge, FilePerms)
+	err = WriteFileAtomic(fs, ltPrivKeyPath, encryptedLtPrivKeyToAge)
 	if err != nil {
 		return nil, fmt.Errorf("failed to write encrypted long-term private key: %w", err)
 	}
@@ -315,7 +315,7 @@ func CreatePGPUnlocker(
 
 	agePrivKeyPath := filepath.Join(unlockerDir, "priv.age.gpg")
 
-	err = afero.WriteFile(fs, agePrivKeyPath, encryptedAgePrivKey, FilePerms)
+	err = WriteFileAtomic(fs, agePrivKeyPath, encryptedAgePrivKey)
 	if err != nil {
 		return nil, fmt.Errorf("failed to write encrypted age private key: %w", err)
 	}
@@ -357,9 +357,8 @@ func writePGPUnlockerMetadata(
 		return nil, fmt.Errorf("failed to marshal unlocker metadata: %w", err)
 	}
 
-	err = afero.WriteFile(fs,
-		filepath.Join(unlockerDir, "unlocker-metadata.json"),
-		metadataBytes, FilePerms)
+	err = WriteFileAtomic(fs,
+		filepath.Join(unlockerDir, "unlocker-metadata.json"), metadataBytes)
 	if err != nil {
 		return nil, fmt.Errorf("failed to write unlocker metadata: %w", err)
 	}

internal/secret/seunlocker_darwin.go

@@ -148,7 +148,7 @@ func (s *SecureEnclaveUnlocker) Remove() error {
 	}
 
 	Debug("Removing SE unlocker directory", "directory", s.Directory)
-	if err := s.fs.RemoveAll(s.Directory); err != nil {
+	if err := RemoveDirAtomic(s.fs, s.Directory); err != nil {
 		return fmt.Errorf("failed to remove SE unlocker directory: %w", err)
 	}
 
@@ -271,7 +271,7 @@ func CreateSecureEnclaveUnlocker(
 
 	// Write SE-encrypted long-term key
 	ltKeyPath := filepath.Join(unlockerDir, seLongtermFilename)
-	if err := afero.WriteFile(fs, ltKeyPath, encryptedLtKey, FilePerms); err != nil {
+	if err := WriteFileAtomic(fs, ltKeyPath, encryptedLtKey); err != nil {
 		return nil, fmt.Errorf(
 			"failed to write SE-encrypted long-term key: %w",
 			err,
@@ -295,7 +295,7 @@ func CreateSecureEnclaveUnlocker(
 	}
 
 	metadataPath := filepath.Join(unlockerDir, "unlocker-metadata.json")
-	if err := afero.WriteFile(fs, metadataPath, metadataBytes, FilePerms); err != nil {
+	if err := WriteFileAtomic(fs, metadataPath, metadataBytes); err != nil {
 		return nil, fmt.Errorf("failed to write metadata: %w", err)
 	}
 

internal/secret/version.go

@@ -6,6 +6,7 @@ import (
 	"fmt"
 	"log/slog"
 	"path/filepath"
+	"slices"
 	"sort"
 	"strings"
 	"time"
@@ -131,7 +132,10 @@ func GenerateVersionName(fs afero.Fs, secretDir string) (string, error) {
 	return fmt.Sprintf("%s.%03d", today, newSerial), nil
 }
 
-// Save saves the version metadata and value
+// Save saves the version metadata and value. The files are written into a
+// temporary directory that is renamed to sv.Directory once all of them are
+// complete, so the version directory is either whole or absent, even if the
+// process dies part-way.
 func (sv *Version) Save(value *memguard.LockedBuffer) error {
 	if value == nil {
 		return errNilValueBuffer
@@ -145,14 +149,22 @@ func (sv *Version) Save(value *memguard.LockedBuffer) error {
 
 	fs := sv.vault.GetFilesystem()
 
-	// Create version directory
-	err := fs.MkdirAll(sv.Directory, DirPerms)
+	// Create the versions directory the finished version is renamed into
+	err := fs.MkdirAll(filepath.Dir(sv.Directory), DirPerms)
 	if err != nil {
-		Debug("Failed to create version directory", "error", err, "dir", sv.Directory)
+		Debug("Failed to create versions directory", "error", err, "dir", sv.Directory)
 
-		return fmt.Errorf("failed to create version directory: %w", err)
+		return fmt.Errorf("failed to create versions directory: %w", err)
 	}
 
+	tmpDir, err := TempDirFor(fs, sv.Directory)
+	if err != nil {
+		return err
+	}
+
+	// Once the rename below has moved it into place, this finds nothing.
+	defer func() { _ = fs.RemoveAll(tmpDir) }()
+
 	// Generate a new keypair for this version
 	Debug("Generating version-specific keypair", "version", sv.Version)
 
@@ -173,21 +185,28 @@ func (sv *Version) Save(value *memguard.LockedBuffer) error {
 		slog.String("public_key", versionIdentity.Recipient().String()),
 	)
 
-	err = sv.writePublicKeyAndValue(fs, versionIdentity, value)
+	err = sv.writePublicKeyAndValue(fs, tmpDir, versionIdentity, value)
 	if err != nil {
 		return err
 	}
 
-	err = sv.writeEncryptedPrivateKey(fs, versionPrivateKeyBuffer)
+	err = sv.writeEncryptedPrivateKey(fs, tmpDir, versionPrivateKeyBuffer)
 	if err != nil {
 		return err
 	}
 
-	err = sv.writeEncryptedMetadata(fs, versionIdentity)
+	err = sv.writeEncryptedMetadata(fs, tmpDir, versionIdentity)
 	if err != nil {
 		return err
 	}
 
+	err = fs.Rename(tmpDir, sv.Directory)
+	if err != nil {
+		Debug("Failed to move version into place", "error", err, "dir", sv.Directory)
+
+		return fmt.Errorf("failed to move version into place: %w", err)
+	}
+
 	Debug("Successfully saved secret version",
 		"version", sv.Version, "secret_name", sv.SecretName)
 
@@ -358,17 +377,18 @@ func (sv *Version) GetValue(
 }
 
 // writePublicKeyAndValue stores the version's public key and the value
-// encrypted to it.
+// encrypted to it in dir.
 func (sv *Version) writePublicKeyAndValue(
 	fs afero.Fs,
+	dir string,
 	versionIdentity *age.X25519Identity,
 	value *memguard.LockedBuffer,
 ) error {
 	versionPublicKey := versionIdentity.Recipient().String()
-	pubKeyPath := filepath.Join(sv.Directory, "pub.age")
+	pubKeyPath := filepath.Join(dir, "pub.age")
 	Debug("Writing version public key", "path", pubKeyPath)
 
-	err := afero.WriteFile(fs, pubKeyPath, []byte(versionPublicKey), FilePerms)
+	err := WriteFileAtomic(fs, pubKeyPath, []byte(versionPublicKey))
 	if err != nil {
 		Debug("Failed to write version public key", "error", err, "path", pubKeyPath)
 
@@ -385,10 +405,10 @@ func (sv *Version) writePublicKeyAndValue(
 		return fmt.Errorf("failed to encrypt version value: %w", err)
 	}
 
-	valuePath := filepath.Join(sv.Directory, "value.age")
+	valuePath := filepath.Join(dir, "value.age")
 	Debug("Writing encrypted version value", "path", valuePath)
 
-	err = afero.WriteFile(fs, valuePath, encryptedValue, FilePerms)
+	err = WriteFileAtomic(fs, valuePath, encryptedValue)
 	if err != nil {
 		Debug("Failed to write encrypted version value", "error", err, "path", valuePath)
 
@@ -399,9 +419,10 @@ func (sv *Version) writePublicKeyAndValue(
 }
 
 // writeEncryptedPrivateKey encrypts the version's private key to the
-// vault's long-term public key and stores it.
+// vault's long-term public key and stores it in dir.
 func (sv *Version) writeEncryptedPrivateKey(
 	fs afero.Fs,
+	dir string,
 	versionPrivateKeyBuffer *memguard.LockedBuffer,
 ) error {
 	vaultDir, _ := sv.vault.GetDirectory()
@@ -435,10 +456,10 @@ func (sv *Version) writeEncryptedPrivateKey(
 		return fmt.Errorf("failed to encrypt version private key: %w", err)
 	}
 
-	privKeyPath := filepath.Join(sv.Directory, "priv.age")
+	privKeyPath := filepath.Join(dir, "priv.age")
 	Debug("Writing encrypted version private key", "path", privKeyPath)
 
-	err = afero.WriteFile(fs, privKeyPath, encryptedPrivKey, FilePerms)
+	err = WriteFileAtomic(fs, privKeyPath, encryptedPrivKey)
 	if err != nil {
 		Debug("Failed to write encrypted version private key",
 			"error", err, "path", privKeyPath)
@@ -450,9 +471,10 @@ func (sv *Version) writeEncryptedPrivateKey(
 }
 
 // writeEncryptedMetadata encrypts the version metadata to the version's
-// public key and stores it.
+// public key and stores it in dir.
 func (sv *Version) writeEncryptedMetadata(
 	fs afero.Fs,
+	dir string,
 	versionIdentity *age.X25519Identity,
 ) error {
 	Debug("Encrypting version metadata", "version", sv.Version)
@@ -476,10 +498,10 @@ func (sv *Version) writeEncryptedMetadata(
 		return fmt.Errorf("failed to encrypt version metadata: %w", err)
 	}
 
-	metadataPath := filepath.Join(sv.Directory, "metadata.age")
+	metadataPath := filepath.Join(dir, "metadata.age")
 	Debug("Writing encrypted version metadata", "path", metadataPath)
 
-	err = afero.WriteFile(fs, metadataPath, encryptedMetadata, FilePerms)
+	err = WriteFileAtomic(fs, metadataPath, encryptedMetadata)
 	if err != nil {
 		Debug("Failed to write encrypted version metadata",
 			"error", err, "path", metadataPath)
@@ -524,6 +546,18 @@ func ListVersions(fs afero.Fs, secretDir string) ([]string, error) {
 	return versions, nil
 }
 
+// VersionExists reports whether version is one of the versions ListVersions
+// lists for the secret in secretDir. It only compares names, so a version
+// the user typed can be checked with it before any path is built from it.
+func VersionExists(fs afero.Fs, secretDir string, version string) (bool, error) {
+	versions, err := ListVersions(fs, secretDir)
+	if err != nil {
+		return false, err
+	}
+
+	return slices.Contains(versions, version), nil
+}
+
 // GetCurrentVersion returns the version that the "current" file points to
 // The file contains just the version name (e.g., "20231215.001")
 func GetCurrentVersion(fs afero.Fs, secretDir string) (string, error) {
@@ -540,15 +574,12 @@ func GetCurrentVersion(fs afero.Fs, secretDir string) (string, error) {
 }
 
 // SetCurrentVersion updates the "current" file to point to a specific version
-// The file contains just the version name (e.g., "20231215.001")
+// The file contains just the version name (e.g., "20231215.001"). It is
+// replaced in one rename, so once written it always exists.
 func SetCurrentVersion(fs afero.Fs, secretDir string, version string) error {
 	currentPath := filepath.Join(secretDir, "current")
 
-	// Remove existing file if it exists
-	_ = fs.Remove(currentPath)
-
-	// Write just the version name to the file
-	err := afero.WriteFile(fs, currentPath, []byte(version), FilePerms)
+	err := WriteFileAtomic(fs, currentPath, []byte(version))
 	if err != nil {
 		return fmt.Errorf("failed to create current version file: %w", err)
 	}

internal/vault/errors.go

@@ -49,7 +49,7 @@ var (
 
 	// ErrVersionNotFound indicates the requested secret version does not
 	// exist. Composed as
-	// "version <version> not found for secret <name>".
+	// "version '<version>' not found for secret '<name>'".
 	ErrVersionNotFound = errors.New("not found for secret")
 
 	// ErrNoVersions indicates the source secret has no versions. Composed
@@ -63,4 +63,10 @@ var (
 	// ErrUnlockerNotFound indicates no unlocker with the given ID exists.
 	// Composed as "unlocker with ID <id> not found".
 	ErrUnlockerNotFound = errors.New("not found")
+
+	// ErrNoLockForFilesystem indicates LockStateDir was given a filesystem
+	// it cannot lock. Composed as "cannot lock the state directory on
+	// filesystem <type>".
+	ErrNoLockForFilesystem = errors.New(
+		"cannot lock the state directory on filesystem")
 )

internal/vault/integration_version_test.go

@@ -235,10 +235,10 @@ func testRetrieveSpecificVersions(
 	require.NoError(t, err)
 	assert.Equal(t, []byte("version-3-data"), value3)
 
-	// Empty version should return current
-	valueCurrent, err := vault.GetSecretVersion(secretName, "")
-	require.NoError(t, err)
-	assert.Equal(t, []byte("version-3-data"), valueCurrent)
+	// An empty version is not one of the versions; GetSecret gets the
+	// current one
+	_, err = vault.GetSecretVersion(secretName, "")
+	require.ErrorIs(t, err, ErrVersionNotFound)
 }
 
 func testPromoteOldVersion(

internal/vault/lock.go

@@ -0,0 +1,73 @@
+package vault
+
+import (
+	"fmt"
+	"os"
+	"path/filepath"
+	"sync"
+	"syscall"
+
+	"git.eeqj.de/sneak/secret/internal/secret"
+	"github.com/spf13/afero"
+)
+
+// lockFileName is the file in the state directory that LockStateDir locks.
+const lockFileName = "lock"
+
+// memFsLock stands in for the lock file on the in-memory filesystem, which
+// has no file locks. Every in-memory filesystem in the process shares it.
+//
+//nolint:gochecknoglobals // must outlive the call that takes it
+var memFsLock sync.Mutex
+
+// LockStateDir takes the lock that a command changing anything under
+// stateDir holds until it returns, and returns the function that releases
+// it. While one command holds it, the next one waits here. Reads take no
+// lock: each file or directory a command changes is replaced in a single
+// rename, so a reader finds it as it was before or after, never half-made.
+//
+// On the real filesystem the lock is flock(2) on the file "lock" in
+// stateDir, which the kernel releases when the process dies, so a killed
+// command never leaves the tool locked. The in-memory filesystem the tests
+// use has no file locks, so a process-wide mutex stands in for flock there.
+// Any other filesystem is refused rather than left unlocked.
+func LockStateDir(fs afero.Fs, stateDir string) (func(), error) {
+	switch fs.(type) {
+	case *afero.OsFs:
+		return flockStateDir(stateDir)
+	case *afero.MemMapFs:
+		memFsLock.Lock()
+
+		return memFsLock.Unlock, nil
+	default:
+		return nil, fmt.Errorf("%w %T", ErrNoLockForFilesystem, fs)
+	}
+}
+
+// flockStateDir takes flock(2) on the lock file in stateDir, creating the
+// directory and the file if needed. Go opens files close-on-exec, so
+// programs the command runs, such as gpg, do not inherit the lock.
+func flockStateDir(stateDir string) (func(), error) {
+	err := os.MkdirAll(stateDir, secret.DirPerms)
+	if err != nil {
+		return nil, fmt.Errorf("failed to create state directory: %w", err)
+	}
+
+	lockPath := filepath.Join(stateDir, lockFileName)
+
+	//nolint:gosec // G304: the path is the lock file in the state directory
+	file, err := os.OpenFile(lockPath, os.O_RDWR|os.O_CREATE, secret.FilePerms)
+	if err != nil {
+		return nil, fmt.Errorf("failed to open lock file: %w", err)
+	}
+
+	err = syscall.Flock(int(file.Fd()), syscall.LOCK_EX)
+	if err != nil {
+		_ = file.Close()
+
+		return nil, fmt.Errorf("failed to lock %s: %w", lockPath, err)
+	}
+
+	// Closing the file releases the lock.
+	return func() { _ = file.Close() }, nil
+}

internal/vault/lock_test.go

@@ -0,0 +1,134 @@
+package vault_test
+
+import (
+	"testing"
+	"time"
+
+	"git.eeqj.de/sneak/secret/internal/vault"
+	"github.com/spf13/afero"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+const (
+	// lockWait is how long a test waits for the lock before deciding it
+	// will never come free.
+	lockWait = 10 * time.Second
+
+	// heldWait is how long a test watches a second holder fail to take a
+	// lock that is held. Broken exclusion lets it in at once.
+	heldWait = 100 * time.Millisecond
+)
+
+// lockFilesystem is a filesystem LockStateDir can lock, with a state
+// directory on it.
+type lockFilesystem struct {
+	name     string
+	fs       afero.Fs
+	stateDir string
+}
+
+// lockFilesystems returns the real filesystem, locked with flock, and the
+// in-memory one, locked with a mutex.
+func lockFilesystems(t *testing.T) []lockFilesystem {
+	t.Helper()
+
+	return []lockFilesystem{
+		{"memory", afero.NewMemMapFs(), testStateDir},
+		{"real", afero.NewOsFs(), t.TempDir()},
+	}
+}
+
+// lockInBackground starts taking the lock and returns a channel that
+// delivers the function releasing it once it has been taken.
+func lockInBackground(
+	t *testing.T, fs afero.Fs, stateDir string,
+) <-chan func() {
+	t.Helper()
+
+	taken := make(chan func(), 1)
+
+	go func() {
+		release, err := vault.LockStateDir(fs, stateDir)
+		if assert.NoError(t, err) {
+			taken <- release
+		}
+	}()
+
+	return taken
+}
+
+// TestLockStateDirExcludes checks that while the lock is held a second
+// holder, with its own open lock file on the real filesystem, waits, and
+// that it gets the lock once the first releases it.
+func TestLockStateDirExcludes(t *testing.T) {
+	t.Parallel()
+
+	for _, lfs := range lockFilesystems(t) {
+		t.Run(lfs.name, func(t *testing.T) {
+			t.Parallel()
+
+			release, err := vault.LockStateDir(lfs.fs, lfs.stateDir)
+			require.NoError(t, err)
+
+			taken := lockInBackground(t, lfs.fs, lfs.stateDir)
+
+			select {
+			case second := <-taken:
+				second()
+				release()
+				t.Fatal("a second holder took the lock while it was held")
+			case <-time.After(heldWait):
+			}
+
+			release()
+
+			select {
+			case second := <-taken:
+				second()
+			case <-time.After(lockWait):
+				t.Fatal("the second holder never got the lock")
+			}
+		})
+	}
+}
+
+// TestLockStateDirFreeAfterPanic checks that a holder that panics, and
+// releases the lock with defer as every command does, leaves it free.
+func TestLockStateDirFreeAfterPanic(t *testing.T) {
+	t.Parallel()
+
+	for _, lfs := range lockFilesystems(t) {
+		t.Run(lfs.name, func(t *testing.T) {
+			t.Parallel()
+
+			assert.Panics(t, func() {
+				release, err := vault.LockStateDir(lfs.fs, lfs.stateDir)
+				require.NoError(t, err)
+
+				defer release()
+
+				panic("the command failed")
+			})
+
+			select {
+			case release := <-lockInBackground(t, lfs.fs, lfs.stateDir):
+				release()
+			case <-time.After(lockWait):
+				t.Fatal("the lock was still held after its holder panicked")
+			}
+		})
+	}
+}
+
+// TestLockStateDirRefusesOtherFilesystems checks that a filesystem with no
+// lock implementation is refused instead of being used unlocked.
+func TestLockStateDirRefusesOtherFilesystems(t *testing.T) {
+	t.Parallel()
+
+	fs := afero.NewReadOnlyFs(afero.NewMemMapFs())
+
+	release, err := vault.LockStateDir(fs, testStateDir)
+	require.ErrorIs(t, err, vault.ErrNoLockForFilesystem)
+	assert.Nil(t, release)
+}

internal/vault/management.go

@@ -169,7 +169,7 @@ func processMnemonicForVault(
 
 	ltPubKeyPath := filepath.Join(vaultDir, "pub.age")
 
-	err = afero.WriteFile(fs, ltPubKeyPath, []byte(ltPubKey), secret.FilePerms)
+	err = secret.WriteFileAtomic(fs, ltPubKeyPath, []byte(ltPubKey))
 	if err != nil {
 		return 0, "", "", fmt.Errorf("failed to write long-term public key: %w", err)
 	}
@@ -295,21 +295,13 @@ func SelectVault(fs afero.Fs, stateDir string, name string) error {
 		return fmt.Errorf("vault %s %w", name, ErrVaultNotFound)
 	}
 
-	// Create or update the currentvault file with just the vault name
+	// Create or replace the currentvault file with just the vault name. It
+	// is replaced in one rename, so it never goes missing.
 	currentVaultPath := filepath.Join(stateDir, "currentvault")
 
-	// Remove existing file if it exists
-	_, err = fs.Stat(currentVaultPath)
-	if err == nil {
-		secret.Debug("Removing existing currentvault file", "path", currentVaultPath)
-
-		_ = fs.Remove(currentVaultPath)
-	}
-
-	// Write just the vault name to the file
 	secret.Debug("Writing currentvault file", "vault_name", name)
 
-	err = afero.WriteFile(fs, currentVaultPath, []byte(name), secret.FilePerms)
+	err = secret.WriteFileAtomic(fs, currentVaultPath, []byte(name))
 	if err != nil {
 		return fmt.Errorf("failed to select vault: %w", err)
 	}

internal/vault/metadata.go

@@ -113,7 +113,7 @@ func SaveVaultMetadata(fs afero.Fs, vaultDir string, metadata *Metadata) error {
 		return fmt.Errorf("failed to marshal vault metadata: %w", err)
 	}
 
-	err = afero.WriteFile(fs, metadataPath, metadataBytes, secret.FilePerms)
+	err = secret.WriteFileAtomic(fs, metadataPath, metadataBytes)
 	if err != nil {
 		return fmt.Errorf("failed to write vault metadata: %w", err)
 	}

internal/vault/secrets.go

@@ -171,17 +171,59 @@ func (v *Vault) AddSecret(name string, value *memguard.LockedBuffer, force bool)
 		slog.String("secret_dir", secretDir),
 	)
 
-	// Check for an existing secret and prepare its directory
-	exists, previousVersion, err := v.prepareSecretDir(name, secretDir, force)
+	// Check for an existing secret and the version the new one supersedes
+	exists, previousVersion, err := v.checkExistingSecret(name, secretDir, force)
 	if err != nil {
 		return err
 	}
 
+	if exists {
+		return v.addVersion(name, secretDir, value, previousVersion)
+	}
+
+	return v.addNewSecret(name, secretDir, value)
+}
+
+// addNewSecret creates a secret by assembling its first version and current
+// pointer in a temporary directory, then renaming that directory to
+// secretDir, so an interrupted add leaves no half-made secret behind.
+func (v *Vault) addNewSecret(
+	name, secretDir string, value *memguard.LockedBuffer,
+) error {
+	buildDir, err := secret.TempDirFor(v.fs, secretDir)
+	if err != nil {
+		return err
+	}
+
+	// Once the rename below has moved it into place, this finds nothing.
+	defer func() { _ = v.fs.RemoveAll(buildDir) }()
+
+	err = v.addVersion(name, buildDir, value, nil)
+	if err != nil {
+		return err
+	}
+
+	err = v.fs.Rename(buildDir, secretDir)
+	if err != nil {
+		return fmt.Errorf("failed to move new secret into place: %w", err)
+	}
+
+	return nil
+}
+
+// addVersion saves value as a new version under secretDir, sets the
+// notAfter timestamp of the version it supersedes, if any, and then points
+// current at the new version. Until that last step, current still names the
+// previous version, which stays readable.
+func (v *Vault) addVersion(
+	name, secretDir string, value *memguard.LockedBuffer,
+	previousVersion *secret.Version,
+) error {
 	now := time.Now()
 
 	// Create the new version and save the encrypted value
 	versionName, err := v.createAndSaveVersion(
-		name, secretDir, value, previousVersion, &now, exists)
+		name, secretDir, value, previousVersion, &now)
 	if err != nil {
 		return err
 	}
@@ -251,7 +293,7 @@ func updateVersionMetadata(
 	// Write encrypted metadata
 	metadataPath := filepath.Join(version.Directory, "metadata.age")
 
-	err = afero.WriteFile(fs, metadataPath, encryptedMetadata, secret.FilePerms)
+	err = secret.WriteFileAtomic(fs, metadataPath, encryptedMetadata)
 	if err != nil {
 		return fmt.Errorf("failed to write encrypted version metadata: %w", err)
 	}
@@ -259,18 +301,31 @@ func updateVersionMetadata(
 	return nil
 }
 
-// GetSecret retrieves a secret from this vault
+// GetSecret retrieves the current version of a secret from this vault
 func (v *Vault) GetSecret(name string) ([]byte, error) {
 	secret.DebugWith("Getting secret from vault",
 		slog.String("vault_name", v.Name),
 		slog.String("secret_name", name),
 	)
 
-	return v.GetSecretVersion(name, "")
+	// GetSecretObject validates the name and checks that the secret exists
+	secretObj, err := v.GetSecretObject(name)
+	if err != nil {
+		return nil, err
+	}
+
+	currentVersion, err := secret.GetCurrentVersion(v.fs, secretObj.Directory)
+	if err != nil {
+		secret.Debug("Failed to get current version", "error", err, "secret_name", name)
+
+		return nil, fmt.Errorf("failed to get current version: %w", err)
+	}
+
+	return v.GetSecretVersion(name, currentVersion)
 }
 
-// GetSecretVersion retrieves a specific version of a secret (empty version
-// means current)
+// GetSecretVersion retrieves a specific version of a secret. The version
+// must be one of the secret's versions; GetSecret gets the current one.
 func (v *Vault) GetSecretVersion(name string, version string) ([]byte, error) {
 	secret.DebugWith("Getting secret version from vault",
 		slog.String("vault_name", v.Name),
@@ -278,8 +333,8 @@ func (v *Vault) GetSecretVersion(name string, version string) ([]byte, error) {
 		slog.String("version", version),
 	)
 
-	// Validate the name and resolve the version to fetch
-	version, err := v.resolveSecretVersion(name, version)
+	// Validate the name and check that the version exists
+	err := v.checkSecretVersion(name, version)
 	if err != nil {
 		return nil, err
 	}
@@ -410,12 +465,14 @@ func (v *Vault) GetSecretObject(name string) (*secret.Secret, error) {
 	return secretObj, nil
 }
 
-// CopySecretVersion copies a single version from source to this vault
-// It decrypts the value using srcIdentity and re-encrypts for this vault
+// CopySecretVersion copies a single version from source into destSecretDir
+// in this vault. It decrypts the value using srcIdentity and re-encrypts
+// for this vault.
 func (v *Vault) CopySecretVersion(
 	srcVersion *secret.Version,
 	srcIdentity *age.X25519Identity,
 	destSecretName string,
+	destSecretDir string,
 	destVersionName string,
 ) error {
 	secret.DebugWith("Copying secret version to vault",
@@ -441,6 +498,7 @@ func (v *Vault) CopySecretVersion(
 
 	// Create destination version with same name
 	destVersion := secret.NewVersion(v, destSecretName, destVersionName)
+	destVersion.Directory = filepath.Join(destSecretDir, "versions", destVersionName)
 
 	// Copy metadata (preserve original timestamps)
 	destVersion.Metadata = srcVersion.Metadata
@@ -481,11 +539,11 @@ func (v *Vault) CopySecretAllVersions(
 		return fmt.Errorf("failed to get destination vault directory: %w", err)
 	}
 
-	// Check if destination secret already exists and clear it if forced
+	// Refuse to replace an existing destination secret unless forced
 	destStorageName := strings.ReplaceAll(destSecretName, "/", "%")
 	destSecretDir := filepath.Join(destVaultDir, "secrets.d", destStorageName)
 
-	err = v.prepareCopyDestination(destSecretDir, destSecretName, force)
+	err = v.checkCopyDestination(destSecretDir, destSecretName, force)
 	if err != nil {
 		return err
 	}
@@ -521,14 +579,8 @@ func (v *Vault) CopySecretAllVersions(
 		return fmt.Errorf("failed to get current version: %w", err)
 	}
 
-	// Create destination secret directory
-	err = v.fs.MkdirAll(destSecretDir, secret.DirPerms)
-	if err != nil {
-		return fmt.Errorf("failed to create destination secret directory: %w", err)
-	}
-
-	// Copy each version and set the current pointer, rolling back on error
-	err = v.copyVersionsWithRollback(srcVault, srcIdentity,
+	// Copy each version and the current pointer, then move the copy into place
+	err = v.copyVersions(srcVault, srcIdentity,
 		srcSecretName, destSecretName, destSecretDir, versions, currentVersion)
 	if err != nil {
 		return err
@@ -543,10 +595,10 @@ func (v *Vault) CopySecretAllVersions(
 	return nil
 }
 
-// prepareSecretDir checks for an existing secret directory and prepares it
-// for a new version. It returns whether the secret already existed and the
-// current version to be superseded, if any.
-func (v *Vault) prepareSecretDir(
+// checkExistingSecret reports whether the secret already exists, refuses to
+// overwrite it unless force is set, and returns its current version, which
+// the new version supersedes, if any.
+func (v *Vault) checkExistingSecret(
 	name, secretDir string, force bool,
 ) (bool, *secret.Version, error) {
 	// Check if secret already exists
@@ -563,19 +615,6 @@ func (v *Vault) prepareSecretDir(
 	secret.Debug("Secret existence check complete", "exists", exists)
 
 	if !exists {
-		// Create secret directory for new secret
-		secret.Debug("Creating secret directory", "secret_dir", secretDir)
-
-		err = v.fs.MkdirAll(secretDir, secret.DirPerms)
-		if err != nil {
-			secret.Debug("Failed to create secret directory",
-				"error", err, "secret_dir", secretDir)
-
-			return false, nil, fmt.Errorf("failed to create secret directory: %w", err)
-		}
-
-		secret.Debug("Created secret directory successfully")
-
 		return false, nil, nil
 	}
 
@@ -640,15 +679,15 @@ func (v *Vault) updatePreviousVersion(
 	return nil
 }
 
-// resolveSecretVersion validates the secret name, verifies the secret and
-// version exist, and resolves an empty version to the current one.
-func (v *Vault) resolveSecretVersion(name, version string) (string, error) {
+// checkSecretVersion validates the secret name and verifies that the secret
+// exists and that version is one of its versions.
+func (v *Vault) checkSecretVersion(name, version string) error {
 	// Validate secret name to prevent path traversal
 	err := ValidateSecretName(name)
 	if err != nil {
 		secret.Debug("Invalid secret name provided", "secret_name", name)
 
-		return "", err
+		return err
 	}
 
 	// Get vault directory
@@ -656,7 +695,7 @@ func (v *Vault) resolveSecretVersion(name, version string) (string, error) {
 	if err != nil {
 		secret.Debug("Failed to get vault directory", "error", err, "vault_name", v.Name)
 
-		return "", err
+		return err
 	}
 
 	// Convert slashes to percent signs for storage
@@ -668,58 +707,38 @@ func (v *Vault) resolveSecretVersion(name, version string) (string, error) {
 	if err != nil {
 		secret.Debug("Failed to check if secret exists", "error", err, "secret_name", name)
 
-		return "", fmt.Errorf("failed to check if secret exists: %w", err)
+		return fmt.Errorf("failed to check if secret exists: %w", err)
 	}
 
 	if !exists {
 		secret.Debug("Secret not found in vault", "secret_name", name, "vault_name", v.Name)
 
-		return "", fmt.Errorf("secret %s %w", name, ErrSecretNotFound)
-	}
-
-	// Determine which version to get
-	if version == "" {
-		// Get current version
-		currentVersion, err := secret.GetCurrentVersion(v.fs, secretDir)
-		if err != nil {
-			secret.Debug("Failed to get current version", "error", err, "secret_name", name)
-
-			return "", fmt.Errorf("failed to get current version: %w", err)
-		}
-
-		version = currentVersion
-
-		secret.Debug("Using current version", "version", version, "secret_name", name)
+		return fmt.Errorf("secret %s %w", name, ErrSecretNotFound)
 	}
 
 	// Check if version exists
-	versionPath := filepath.Join(secretDir, "versions", version)
-
-	exists, err = afero.DirExists(v.fs, versionPath)
+	exists, err = secret.VersionExists(v.fs, secretDir, version)
 	if err != nil {
 		secret.Debug("Failed to check if version exists", "error", err, "version", version)
 
-		return "", fmt.Errorf("failed to check if version exists: %w", err)
+		return fmt.Errorf("failed to check if version exists: %w", err)
 	}
 
 	if !exists {
 		secret.Debug("Version not found", "version", version, "secret_name", name)
 
-		return "", fmt.Errorf(
-			"version %s %w %s",
-			version, ErrVersionNotFound, name,
-		)
+		return fmt.Errorf("version '%s' %w '%s'", version, ErrVersionNotFound, name)
 	}
 
-	return version, nil
+	return nil
 }
 
 // createAndSaveVersion generates a new version name, sets the version
-// timestamps, and saves the encrypted value. When saving fails for a newly
-// created secret, the secret directory is removed again.
+// timestamps, and saves the encrypted value under secretDir, which is a
+// temporary directory while a new secret is being assembled.
 func (v *Vault) createAndSaveVersion(
 	name, secretDir string, value *memguard.LockedBuffer,
-	previousVersion *secret.Version, now *time.Time, exists bool,
+	previousVersion *secret.Version, now *time.Time,
 ) (string, error) {
 	// Generate new version name
 	versionName, err := secret.GenerateVersionName(v.fs, secretDir)
@@ -733,6 +752,7 @@ func (v *Vault) createAndSaveVersion(
 
 	// Create new version
 	newVersion := secret.NewVersion(v, name, versionName)
+	newVersion.Directory = filepath.Join(secretDir, "versions", versionName)
 
 	// Set version timestamps
 	if previousVersion == nil {
@@ -752,57 +772,73 @@ func (v *Vault) createAndSaveVersion(
 	if err != nil {
 		secret.Debug("Failed to save new version", "error", err, "version", versionName)
 
-		// Clean up the secret directory if this was a new secret
-		if !exists {
-			secret.Debug("Cleaning up secret directory due to save failure",
-				"secret_dir", secretDir)
-
-			_ = v.fs.RemoveAll(secretDir)
-		}
-
 		return "", fmt.Errorf("failed to save version: %w", err)
 	}
 
 	return versionName, nil
 }
 
-// copyVersionsWithRollback copies each version of the source secret into the
-// destination directory and sets the current version pointer, removing the
-// partial copy when any step fails.
-func (v *Vault) copyVersionsWithRollback(
+// copyVersions copies each version of the source secret and its current
+// pointer into a temporary directory, then moves that directory to
+// destSecretDir, replacing a secret already there. Nothing in this vault
+// changes until the copy is complete, so an interrupted copy leaves only a
+// temporary directory behind.
+func (v *Vault) copyVersions(
 	srcVault *Vault, srcIdentity *age.X25519Identity,
 	srcSecretName, destSecretName, destSecretDir string,
 	versions []string, currentVersion string,
 ) error {
-	// Copy each version
+	buildDir, err := secret.TempDirFor(v.fs, destSecretDir)
+	if err != nil {
+		return err
+	}
+
+	// Once the rename below has moved it into place, this finds nothing.
+	defer func() { _ = v.fs.RemoveAll(buildDir) }()
+
 	for _, versionName := range versions {
 		srcVersion := secret.NewVersion(srcVault, srcSecretName, versionName)
 
-		err := v.CopySecretVersion(srcVersion, srcIdentity, destSecretName, versionName)
+		err = v.CopySecretVersion(
+			srcVersion, srcIdentity, destSecretName, buildDir, versionName)
 		if err != nil {
-			// Rollback: remove partial copy
-			secret.Debug("Rolling back partial copy due to error", "error", err)
-
-			_ = v.fs.RemoveAll(destSecretDir)
-
 			return fmt.Errorf("failed to copy version %s: %w", versionName, err)
 		}
 	}
 
-	// Set current version
-	err := secret.SetCurrentVersion(v.fs, destSecretDir, currentVersion)
+	err = secret.SetCurrentVersion(v.fs, buildDir, currentVersion)
 	if err != nil {
-		_ = v.fs.RemoveAll(destSecretDir)
-
 		return fmt.Errorf("failed to set current version: %w", err)
 	}
 
+	// With --force, the secret being replaced goes only now that its
+	// replacement is complete
+	exists, err := afero.DirExists(v.fs, destSecretDir)
+	if err != nil {
+		return fmt.Errorf("failed to check destination: %w", err)
+	}
+
+	if exists {
+		secret.Debug("Removing existing destination secret", "path", destSecretDir)
+
+		err = secret.RemoveDirAtomic(v.fs, destSecretDir)
+		if err != nil {
+			return fmt.Errorf("failed to remove existing destination secret: %w", err)
+		}
+	}
+
+	err = v.fs.Rename(buildDir, destSecretDir)
+	if err != nil {
+		return fmt.Errorf("failed to move copied secret into place: %w", err)
+	}
+
 	return nil
 }
 
-// prepareCopyDestination ensures the destination secret directory can be
-// created, removing an existing secret when force is set.
-func (v *Vault) prepareCopyDestination(
+// checkCopyDestination refuses to copy over an existing secret unless force
+// is set. A secret being replaced is removed by copyVersions, once its
+// replacement is complete.
+func (v *Vault) checkCopyDestination(
 	destSecretDir, destSecretName string, force bool,
 ) error {
 	exists, err := afero.DirExists(v.fs, destSecretDir)
@@ -817,15 +853,5 @@ func (v *Vault) prepareCopyDestination(
 		)
 	}
 
-	if exists && force {
-		// Remove existing secret
-		secret.Debug("Removing existing destination secret", "path", destSecretDir)
-
-		err = v.fs.RemoveAll(destSecretDir)
-		if err != nil {
-			return fmt.Errorf("failed to remove existing destination secret: %w", err)
-		}
-	}
-
 	return nil
 }

internal/vault/secrets_version_test.go

@@ -202,10 +202,10 @@ func TestVaultGetSecretVersion(t *testing.T) {
 	require.NoError(t, err)
 	assert.Equal(t, []byte("version-2"), value)
 
-	// Get current (empty version)
-	value, err = vault.GetSecretVersion(testSecretPath, "")
-	require.NoError(t, err)
-	assert.Equal(t, []byte("version-2"), value)
+	// An empty version is not one of the versions; GetSecret gets the
+	// current one
+	_, err = vault.GetSecretVersion(testSecretPath, "")
+	require.ErrorIs(t, err, ErrVersionNotFound)
 }
 
 //nolint:paralleltest // createTestVaultWithKey uses t.Setenv

internal/vault/unlockers.go

@@ -310,30 +310,16 @@ func (v *Vault) SelectUnlocker(unlockerID string) error {
 		return fmt.Errorf("unlocker with ID %s %w", unlockerID, ErrUnlockerNotFound)
 	}
 
-	// Create/update current-unlocker file with just the unlocker name
+	// Create or replace the current-unlocker file with just the unlocker
+	// name. It is replaced in one rename, so it never goes missing.
 	currentUnlockerPath := filepath.Join(vaultDir, "current-unlocker")
 
-	// Remove existing file if it exists
-	exists, err := afero.Exists(v.fs, currentUnlockerPath)
-	if err != nil {
-		return fmt.Errorf("failed to check if current-unlocker file exists: %w", err)
-	}
-
-	if exists {
-		err = v.fs.Remove(currentUnlockerPath)
-		if err != nil {
-			return fmt.Errorf("failed to remove existing current-unlocker file: %w", err)
-		}
-	}
-
 	// Get just the unlocker name (basename of the directory)
 	unlockerName := filepath.Base(targetUnlockerDir)
 
-	// Write just the unlocker name to the file
 	secret.Debug("Writing current-unlocker file", "unlocker_name", unlockerName)
 
-	err = afero.WriteFile(v.fs, currentUnlockerPath, []byte(unlockerName),
-		secret.FilePerms)
+	err = secret.WriteFileAtomic(v.fs, currentUnlockerPath, []byte(unlockerName))
 	if err != nil {
 		return fmt.Errorf("failed to create current-unlocker file: %w", err)
 	}
@@ -351,6 +337,14 @@ func (v *Vault) CreatePassphraseUnlocker(
 		return nil, fmt.Errorf("failed to get vault directory: %w", err)
 	}
 
+	// We need to get the long-term key (either from memory if unlocked, or
+	// derive it). Getting it before anything is written means failing to
+	// get it changes nothing, even when replacing the current unlocker.
+	ltIdentity, err := v.GetOrDeriveLongTermKey()
+	if err != nil {
+		return nil, fmt.Errorf("failed to get long-term key: %w", err)
+	}
+
 	// Create unlocker directory
 	unlockerDir := filepath.Join(vaultDir, "unlockers.d", unlockerTypePassphrase)
 
@@ -371,33 +365,7 @@ func (v *Vault) CreatePassphraseUnlocker(
 		return nil, err
 	}
 
-	// Create metadata
-	metadata := UnlockerMetadata{
-		Type:      unlockerTypePassphrase,
-		CreatedAt: time.Now(),
-		Flags:     []string{},
-	}
-
-	// Write metadata
-	metadataBytes, err := json.MarshalIndent(metadata, "", "  ")
-	if err != nil {
-		return nil, fmt.Errorf("failed to marshal metadata: %w", err)
-	}
-
-	metadataPath := filepath.Join(unlockerDir, "unlocker-metadata.json")
-
-	err = afero.WriteFile(v.fs, metadataPath, metadataBytes, secret.FilePerms)
-	if err != nil {
-		return nil, fmt.Errorf("failed to write unlocker metadata: %w", err)
-	}
-
 	// Encrypt long-term private key to this unlocker
-	// We need to get the long-term key (either from memory if unlocked, or derive it)
-	ltIdentity, err := v.GetOrDeriveLongTermKey()
-	if err != nil {
-		return nil, fmt.Errorf("failed to get long-term key: %w", err)
-	}
-
 	ltPrivKeyBuffer := memguard.NewBufferFromBytes([]byte(ltIdentity.String()))
 	defer ltPrivKeyBuffer.Destroy()
 
@@ -409,11 +377,31 @@ func (v *Vault) CreatePassphraseUnlocker(
 
 	ltPrivKeyPath := filepath.Join(unlockerDir, "longterm.age")
 
-	err = afero.WriteFile(v.fs, ltPrivKeyPath, encryptedLtPrivKey, secret.FilePerms)
+	err = secret.WriteFileAtomic(v.fs, ltPrivKeyPath, encryptedLtPrivKey)
 	if err != nil {
 		return nil, fmt.Errorf("failed to write encrypted long-term private key: %w", err)
 	}
 
+	// Write the metadata last: readers skip an unlocker directory without
+	// it, so an unlocker interrupted before this point is never used.
+	metadata := UnlockerMetadata{
+		Type:      unlockerTypePassphrase,
+		CreatedAt: time.Now(),
+		Flags:     []string{},
+	}
+
+	metadataBytes, err := json.MarshalIndent(metadata, "", "  ")
+	if err != nil {
+		return nil, fmt.Errorf("failed to marshal metadata: %w", err)
+	}
+
+	metadataPath := filepath.Join(unlockerDir, "unlocker-metadata.json")
+
+	err = secret.WriteFileAtomic(v.fs, metadataPath, metadataBytes)
+	if err != nil {
+		return nil, fmt.Errorf("failed to write unlocker metadata: %w", err)
+	}
+
 	// Create the unlocker instance
 	unlocker := secret.NewPassphraseUnlocker(v.fs, unlockerDir, metadata)
 
@@ -467,9 +455,8 @@ func (v *Vault) writeUnlockerKeypair(
 	// Write public key
 	pubKeyPath := filepath.Join(unlockerDir, "pub.age")
 
-	err := afero.WriteFile(v.fs, pubKeyPath,
-		[]byte(unlockerIdentity.Recipient().String()),
-		secret.FilePerms)
+	err := secret.WriteFileAtomic(v.fs, pubKeyPath,
+		[]byte(unlockerIdentity.Recipient().String()))
 	if err != nil {
 		return fmt.Errorf("failed to write unlocker public key: %w", err)
 	}
@@ -488,7 +475,7 @@ func (v *Vault) writeUnlockerKeypair(
 	// Write encrypted private key
 	privKeyPath := filepath.Join(unlockerDir, "priv.age")
 
-	err = afero.WriteFile(v.fs, privKeyPath, encryptedPrivKey, secret.FilePerms)
+	err = secret.WriteFileAtomic(v.fs, privKeyPath, encryptedPrivKey)
 	if err != nil {
 		return fmt.Errorf("failed to write encrypted unlocker private key: %w", err)
 	}

internal/vault/vault_error_test.go

@@ -90,4 +90,10 @@ func TestAddSecretCleansUpOnFailure(t *testing.T) {
 	secretDir := filepath.Join(vaultDir, "secrets.d", testSecretName)
 	exists, _ := afero.DirExists(fs, secretDir)
 	assert.False(t, exists, "Secret directory should not exist after failed AddSecret")
+
+	// Nor is the temporary directory the secret was assembled in left behind
+	entries, err := afero.ReadDir(fs, vaultDir)
+	require.NoError(t, err)
+	require.Len(t, entries, 1)
+	assert.Equal(t, "pub.age", entries[0].Name())
 }