Merge pull request #4983 from Wei-Shaw/fix/issue-4887-prompt-audit-recovery

fix(security-audit): 解密失败不再吞掉整份配置，修复升级后配置消失且无法保存的死锁

backend/cmd/server/wire_gen.go

@@ -255,7 +255,7 @@ func initializeApplication(buildInfo handler.BuildInfo) (*Application, error) {
 	contentModerationHashCache := repository.NewContentModerationHashCache(redisClient)
 	contentModerationService := service.NewContentModerationService(settingRepository, contentModerationRepository, contentModerationHashCache, groupRepository, userRepository, apiKeyAuthCacheInvalidator, emailService)
 	contentModerationHandler := admin.NewContentModerationHandler(contentModerationService)
-	configManager := securityaudit.NewConfigManager(db, settingRepository, redisClient, secretEncryptor)
+	configManager := securityaudit.NewConfigManager(db, settingRepository, redisClient, secretEncryptor, configConfig)
 	postgreSQLRepository := securityaudit.NewPostgreSQLRepository(db)
 	redisPayloadStore := securityaudit.NewRedisPayloadStore(redisClient)
 	openAICompatibleScanner := securityaudit.NewOpenAICompatibleScanner()

backend/internal/securityaudit/prompt_config.go

@@ -90,6 +90,11 @@ type ActiveEndpoint struct {
 	TimeoutMS  int
 	InputLimit int
 	Enabled    bool
+	// TokenInvalid marks an endpoint whose persisted token ciphertext cannot be
+	// decrypted with the current encryption key (key changed or auto-generated
+	// on restart). The endpoint is kept visible for admins but excluded from
+	// runtime use until the token is re-entered or cleared (issue #4887).
+	TokenInvalid bool
 }
 
 type ActiveConfig struct {
@@ -365,7 +370,23 @@ func (cfg ActiveConfig) EnabledEndpoints() []ActiveEndpoint {
 	return result
 }
 
-func PublicFromStorage(cfg storageConfig, riskControlEnabled bool) PublicConfig {
+// InvalidTokenEndpointIDs lists endpoints whose stored token could not be
+// decrypted with the current encryption key.
+func (cfg ActiveConfig) InvalidTokenEndpointIDs() []string {
+	ids := make([]string, 0)
+	for _, ep := range cfg.Endpoints {
+		if ep.TokenInvalid {
+			ids = append(ids, ep.ID)
+		}
+	}
+	return ids
+}
+
+func PublicFromStorage(cfg storageConfig, riskControlEnabled bool, invalidTokenEndpointIDs []string) PublicConfig {
+	invalid := make(map[string]struct{}, len(invalidTokenEndpointIDs))
+	for _, id := range invalidTokenEndpointIDs {
+		invalid[id] = struct{}{}
+	}
 	scanners := append([]string{}, cfg.Scanners...)
 	groupIDs := append([]int64{}, cfg.GroupIDs...)
 	endpoints := make([]PublicEndpoint, 0, len(cfg.Endpoints))
@@ -374,6 +395,9 @@ func PublicFromStorage(cfg storageConfig, riskControlEnabled bool) PublicConfig
 		status := "missing"
 		if hasToken {
 			status = "configured"
+			if _, ok := invalid[ep.ID]; ok {
+				status = "invalid"
+			}
 		}
 		endpoints = append(endpoints, PublicEndpoint{
 			ID: ep.ID, Name: ep.Name, Protocol: ep.Protocol, BaseURL: ep.BaseURL,
@@ -402,19 +426,27 @@ func ActiveFromStorage(cfg storageConfig, riskControlEnabled bool, encryptor Sec
 	}
 	for _, ep := range cfg.Endpoints {
 		token := ""
+		tokenInvalid := false
 		if ep.TokenCiphertext != "" {
 			if encryptor == nil {
 				return ActiveConfig{}, fmt.Errorf("prompt audit secret encryptor unavailable")
 			}
 			plain, err := encryptor.Decrypt(ep.TokenCiphertext)
 			if err != nil {
-				return ActiveConfig{}, fmt.Errorf("decrypt prompt audit endpoint token %q: %w", ep.ID, err)
+				// An undecryptable token (encryption key changed or regenerated)
+				// must not take the whole config down: admins would otherwise be
+				// locked out of the real config version and unable to recover
+				// (issue #4887). Keep the ciphertext persisted, but exclude the
+				// endpoint from runtime use until the token is re-entered.
+				tokenInvalid = true
+			} else {
+				token = plain
 			}
-			token = plain
 		}
 		active.Endpoints = append(active.Endpoints, ActiveEndpoint{
 			ID: ep.ID, Name: ep.Name, Protocol: ep.Protocol, BaseURL: ep.BaseURL, Model: ep.Model,
-			Token: token, TimeoutMS: ep.TimeoutMS, InputLimit: ep.InputLimit, Enabled: ep.Enabled,
+			Token: token, TimeoutMS: ep.TimeoutMS, InputLimit: ep.InputLimit,
+			Enabled: ep.Enabled && !tokenInvalid, TokenInvalid: tokenInvalid,
 		})
 	}
 	return active, nil

backend/internal/securityaudit/prompt_config_integration_test.go

@@ -148,8 +148,8 @@ func TestPromptAuditConfigCASSecretRoundTripInvalidationAndTTL(t *testing.T) {
 	t.Cleanup(func() { require.NoError(t, redisClient.Close()) })
 	require.NoError(t, redisClient.Ping(context.Background()).Err())
 
-	managerOne := NewConfigManager(db, settingRepo, redisClient, encryptor)
-	managerTwo := NewConfigManager(db, settingRepo, redisClient, encryptor)
+	managerOne := NewConfigManager(db, settingRepo, redisClient, encryptor, testTotpKeyConfig())
+	managerTwo := NewConfigManager(db, settingRepo, redisClient, encryptor, testTotpKeyConfig())
 	ctx, cancel := context.WithCancel(context.Background())
 	t.Cleanup(cancel)
 	require.NoError(t, managerOne.Start(ctx))
@@ -220,7 +220,7 @@ func TestPromptAuditConfigCASSecretRoundTripInvalidationAndTTL(t *testing.T) {
 
 	// A manager without Redis subscriptions must still converge through the
 	// bounded five-second refresh loop.
-	ttlManager := NewConfigManager(db, settingRepo, nil, encryptor)
+	ttlManager := NewConfigManager(db, settingRepo, nil, encryptor, testTotpKeyConfig())
 	require.NoError(t, ttlManager.Start(ctx))
 	t.Cleanup(func() { require.NoError(t, ttlManager.Shutdown(context.Background())) })
 	waitForConfigVersion(t, ttlManager, 3, time.Second)
@@ -233,7 +233,7 @@ func TestPromptAuditConfigCASSecretRoundTripInvalidationAndTTL(t *testing.T) {
 	// successfully committed PostgreSQL config.
 	deadRedis := redis.NewClient(&redis.Options{Addr: "127.0.0.1:1", MaxRetries: 0, DialTimeout: 30 * time.Millisecond, ReadTimeout: 30 * time.Millisecond, WriteTimeout: 30 * time.Millisecond})
 	t.Cleanup(func() { _ = deadRedis.Close() })
-	degraded := NewConfigManager(db, settingRepo, deadRedis, encryptor)
+	degraded := NewConfigManager(db, settingRepo, deadRedis, encryptor, testTotpKeyConfig())
 	require.NoError(t, degraded.Reload(context.Background()))
 	degradedSaved, err := degraded.Save(context.Background(), promptAuditUpdateRequest(4, 6, ""), 401)
 	require.NoError(t, err)

backend/internal/securityaudit/prompt_config_store.go

@@ -12,6 +12,7 @@ import (
 	"sync/atomic"
 	"time"
 
+	"github.com/Wei-Shaw/sub2api/internal/config"
 	infraerrors "github.com/Wei-Shaw/sub2api/internal/pkg/errors"
 	"github.com/Wei-Shaw/sub2api/internal/service"
 	"github.com/redis/go-redis/v9"
@@ -29,6 +30,10 @@ type ConfigManager struct {
 	redis     *redis.Client
 	encryptor SecretEncryptor
 	clock     Clock
+	// encryptionKeyConfigured mirrors cfg.Totp.EncryptionKeyConfigured. With an
+	// auto-generated (per-boot) key, newly saved endpoint tokens would become
+	// undecryptable after the next restart, so Save rejects them (issue #4887).
+	encryptionKeyConfigured bool
 
 	snapshot atomic.Pointer[activeConfigSnapshot]
 	expected atomic.Int64
@@ -53,8 +58,11 @@ type ConfigManager struct {
 	wg          sync.WaitGroup
 }
 
-func NewConfigManager(db *sql.DB, settings service.SettingRepository, redisClient *redis.Client, encryptor service.SecretEncryptor) *ConfigManager {
-	return &ConfigManager{db: db, settings: settings, redis: redisClient, encryptor: encryptor, clock: realClock{}}
+func NewConfigManager(db *sql.DB, settings service.SettingRepository, redisClient *redis.Client, encryptor service.SecretEncryptor, cfg *config.Config) *ConfigManager {
+	return &ConfigManager{
+		db: db, settings: settings, redis: redisClient, encryptor: encryptor, clock: realClock{},
+		encryptionKeyConfigured: cfg != nil && cfg.Totp.EncryptionKeyConfigured,
+	}
 }
 
 func (m *ConfigManager) Start(ctx context.Context) error {
@@ -125,15 +133,45 @@ func (m *ConfigManager) Reload(ctx context.Context) error {
 		return err
 	}
 	now := m.clock.Now()
+	previous := m.snapshot.Load()
 	m.snapshot.Store(&activeConfigSnapshot{storage: cloneStorageConfig(storage), active: cloneActiveConfig(active), loadedAt: now})
 	m.configUntrusted.Store(false)
 	m.clearLoadError()
+	m.logInvalidTokenEndpoints(previous, active)
 	LogInfo(EventConfigLoaded, map[string]any{
 		"config_version": storage.ConfigVersion, "status": "loaded",
 	})
 	return nil
 }
 
+// logInvalidTokenEndpoints warns once per change (not on every 5s refresh)
+// when stored endpoint tokens cannot be decrypted with the current key.
+func (m *ConfigManager) logInvalidTokenEndpoints(previous *activeConfigSnapshot, active ActiveConfig) {
+	invalid := active.InvalidTokenEndpointIDs()
+	if len(invalid) == 0 {
+		return
+	}
+	if previous != nil {
+		prior := previous.active.InvalidTokenEndpointIDs()
+		if len(prior) == len(invalid) {
+			same := true
+			for i := range invalid {
+				if prior[i] != invalid[i] {
+					same = false
+					break
+				}
+			}
+			if same && previous.active.ConfigVersion == active.ConfigVersion {
+				return
+			}
+		}
+	}
+	LogWarn(EventConfigTokenInvalid, map[string]any{
+		"config_version": active.ConfigVersion, "status": "degraded",
+		"error_code": "endpoint_token_undecryptable", "guard_endpoint_id": strings.Join(invalid, ","),
+	})
+}
+
 func (m *ConfigManager) Active() (ActiveConfig, bool) {
 	if m == nil {
 		return ActiveConfig{}, false
@@ -202,7 +240,7 @@ func (m *ConfigManager) Public() (PublicConfig, error) {
 	if snapshot == nil {
 		return PublicConfig{}, infraerrors.ServiceUnavailable(ErrorCodeConfigUnavailable, "提示词审计配置暂不可用")
 	}
-	return PublicFromStorage(cloneStorageConfig(snapshot.storage), snapshot.active.RiskControlEnabled), nil
+	return PublicFromStorage(cloneStorageConfig(snapshot.storage), snapshot.active.RiskControlEnabled, snapshot.active.InvalidTokenEndpointIDs()), nil
 }
 
 func (m *ConfigManager) Save(ctx context.Context, req UpdateConfigRequest, actorID int64) (PublicConfig, error) {
@@ -268,11 +306,13 @@ func (m *ConfigManager) Save(ctx context.Context, req UpdateConfigRequest, actor
 	}
 	m.expected.Store(next.ConfigVersion)
 	m.expectedBlocking.Store(active.RiskControlEnabled && next.Enabled && next.BlockingEnabled)
+	previous := m.snapshot.Load()
 	m.snapshot.Store(&activeConfigSnapshot{storage: cloneStorageConfig(next), active: cloneActiveConfig(active), loadedAt: m.clock.Now()})
 	// A successful admin save installs a trustworthy snapshot; clear any prior
 	// fail-closed degradation so disabling audit actually takes effect.
 	m.configUntrusted.Store(false)
 	m.clearLoadError()
+	m.logInvalidTokenEndpoints(previous, active)
 	LogInfo(EventConfigUpdated, map[string]any{
 		"config_version": next.ConfigVersion, "status": "updated",
 	})
@@ -283,7 +323,7 @@ func (m *ConfigManager) Save(ctx context.Context, req UpdateConfigRequest, actor
 			})
 		}
 	}
-	return PublicFromStorage(next, active.RiskControlEnabled), nil
+	return PublicFromStorage(next, active.RiskControlEnabled, active.InvalidTokenEndpointIDs()), nil
 }
 
 func (m *ConfigManager) buildNextStorage(current storageConfig, req UpdateConfigRequest, actorID int64) (storageConfig, error) {
@@ -317,6 +357,10 @@ func (m *ConfigManager) buildNextStorage(current storageConfig, req UpdateConfig
 		case endpoint.ClearToken:
 			stored.TokenCiphertext = ""
 		case strings.TrimSpace(endpoint.Token) != "":
+			if !m.encryptionKeyConfigured {
+				return storageConfig{}, infraerrors.BadRequest(ErrorCodeEncryptionKeyRequired,
+					"未配置固定加密密钥，审计节点 Token 将在服务重启后失效。请先设置 TOTP_ENCRYPTION_KEY 环境变量（64 位十六进制）并重启服务")
+			}
 			ciphertext, err := m.encryptor.Encrypt(strings.TrimSpace(endpoint.Token))
 			if err != nil {
 				return storageConfig{}, fmt.Errorf("encrypt prompt audit endpoint token: %w", err)

backend/internal/securityaudit/prompt_config_test.go

@@ -4,8 +4,10 @@ import (
 	"context"
 	"encoding/json"
 	"errors"
+	"strings"
 	"testing"
 
+	"github.com/Wei-Shaw/sub2api/internal/config"
 	infraerrors "github.com/Wei-Shaw/sub2api/internal/pkg/errors"
 	"github.com/stretchr/testify/require"
 )
@@ -13,7 +15,18 @@ import (
 type prefixEncryptor struct{}
 
 func (prefixEncryptor) Encrypt(value string) (string, error) { return "enc:" + value, nil }
-func (prefixEncryptor) Decrypt(value string) (string, error) { return value[4:], nil }
+func (prefixEncryptor) Decrypt(value string) (string, error) {
+	if !strings.HasPrefix(value, "enc:") {
+		return "", errors.New("cipher: message authentication failed")
+	}
+	return value[4:], nil
+}
+
+// testTotpKeyConfig mirrors a deployment with a fixed TOTP_ENCRYPTION_KEY so
+// unit tests may persist endpoint tokens.
+func testTotpKeyConfig() *config.Config {
+	return &config.Config{Totp: config.TotpConfig{EncryptionKeyConfigured: true}}
+}
 
 func TestDefaultConfigIsOff(t *testing.T) {
 	storage, err := ParseStorageConfig("")
@@ -23,7 +36,7 @@ func TestDefaultConfigIsOff(t *testing.T) {
 	require.NoError(t, err)
 	require.Equal(t, ModeOff, active.EffectiveMode())
 	require.Equal(t, AllScannerIDs, storage.Scanners)
-	publicJSON, err := json.Marshal(PublicFromStorage(storage, true))
+	publicJSON, err := json.Marshal(PublicFromStorage(storage, true, nil))
 	require.NoError(t, err)
 	require.Contains(t, string(publicJSON), `"group_ids":[]`)
 	require.Contains(t, string(publicJSON), `"endpoints":[]`)
@@ -38,7 +51,7 @@ func TestConfigRejectsBlockingWithoutAudit(t *testing.T) {
 func TestPublicConfigNeverMarshalsToken(t *testing.T) {
 	storage := DefaultStorageConfig()
 	storage.Endpoints = []StorageEndpoint{{ID: "one", Name: "One", Protocol: "openai_compatible", BaseURL: "http://127.0.0.1:8080", Model: DefaultGuardModel, TokenCiphertext: "GUARD_TOKEN_CANARY_SECRET", TimeoutMS: 1000, InputLimit: 1000, Enabled: true}}
-	public := PublicFromStorage(storage, true)
+	public := PublicFromStorage(storage, true, nil)
 	raw, err := json.Marshal(public)
 	require.NoError(t, err)
 	require.NotContains(t, string(raw), "GUARD_TOKEN_CANARY_SECRET")
@@ -61,7 +74,7 @@ func TestConfigManagerPublicRequiresSuccessfullyLoadedSnapshot(t *testing.T) {
 		manager := NewConfigManager(nil, staticSettingRepository{values: map[string]string{
 			SettingKeyPromptAuditConfig: "",
 			SettingKeyRiskControl:       "false",
-		}}, nil, prefixEncryptor{})
+		}}, nil, prefixEncryptor{}, testTotpKeyConfig())
 		require.NoError(t, manager.Reload(context.Background()))
 
 		public, err := manager.Public()
@@ -70,12 +83,14 @@ func TestConfigManagerPublicRequiresSuccessfullyLoadedSnapshot(t *testing.T) {
 		require.False(t, public.Enabled)
 	})
 
-	t.Run("persisted config activation failure is unavailable", func(t *testing.T) {
+	t.Run("unparseable persisted config is unavailable", func(t *testing.T) {
 		const canary = "persisted-token-canary"
 		manager := NewConfigManager(nil, staticSettingRepository{values: map[string]string{
+			// Endpoint without id/name fails validation, so no trustworthy
+			// snapshot can be installed from this raw value.
 			SettingKeyPromptAuditConfig: `{"enabled":true,"config_version":9,"endpoints":[{"token_ciphertext":"` + canary + `"}]}`,
 			SettingKeyRiskControl:       "true",
-		}}, nil, prefixEncryptor{})
+		}}, nil, prefixEncryptor{}, testTotpKeyConfig())
 		require.Error(t, manager.Reload(context.Background()))
 
 		public, err := manager.Public()
@@ -95,7 +110,7 @@ func TestConfigManagerPublicRequiresSuccessfullyLoadedSnapshot(t *testing.T) {
 			SettingKeyPromptAuditConfig: string(raw),
 			SettingKeyRiskControl:       "false",
 		}}}
-		manager := NewConfigManager(nil, repository, nil, prefixEncryptor{})
+		manager := NewConfigManager(nil, repository, nil, prefixEncryptor{}, testTotpKeyConfig())
 		require.NoError(t, manager.Reload(context.Background()))
 		repository.loadErr = errors.New("settings unavailable")
 		require.Error(t, manager.Reload(context.Background()))
@@ -107,8 +122,66 @@ func TestConfigManagerPublicRequiresSuccessfullyLoadedSnapshot(t *testing.T) {
 	})
 }
 
+// Regression coverage for issue #4887: a persisted config whose endpoint token
+// can no longer be decrypted (encryption key changed or auto-generated per
+// boot) must stay visible and editable for admins instead of falling back to a
+// default v1 config that makes every save fail the CAS version check.
+func TestConfigManagerUndecryptableTokenKeepsConfigVisibleAndRecoverable(t *testing.T) {
+	const canary = "persisted-token-canary"
+	persisted := `{"enabled":true,"blocking_enabled":false,"config_version":9,"endpoints":[{"id":"g1","name":"Guard","protocol":"openai_compatible","base_url":"http://127.0.0.1:8080","model":"m","token_ciphertext":"` + canary + `","timeout_ms":1000,"input_limit":1000,"enabled":true}]}`
+	manager := NewConfigManager(nil, staticSettingRepository{values: map[string]string{
+		SettingKeyPromptAuditConfig: persisted,
+		SettingKeyRiskControl:       "true",
+	}}, nil, prefixEncryptor{}, testTotpKeyConfig())
+	require.NoError(t, manager.Reload(context.Background()), "an undecryptable token must not fail the whole config load")
+
+	public, err := manager.Public()
+	require.NoError(t, err)
+	require.Equal(t, int64(9), public.ConfigVersion, "admins must see the real persisted version so CAS saves can succeed")
+	require.Len(t, public.Endpoints, 1)
+	require.True(t, public.Endpoints[0].HasToken)
+	require.Equal(t, "invalid", public.Endpoints[0].TokenStatus)
+	raw, err := json.Marshal(public)
+	require.NoError(t, err)
+	require.NotContains(t, string(raw), canary)
+
+	active, ok := manager.Active()
+	require.True(t, ok)
+	require.Len(t, active.Endpoints, 1)
+	require.False(t, active.Endpoints[0].Enabled, "an endpoint with an undecryptable token must not be used at runtime")
+	require.True(t, active.Endpoints[0].TokenInvalid)
+	require.Empty(t, active.Endpoints[0].Token)
+	require.Empty(t, active.EnabledEndpoints())
+	require.Equal(t, []string{"g1"}, active.InvalidTokenEndpointIDs())
+
+	expected, activeVersion, _, _ := manager.RuntimeState()
+	require.Equal(t, int64(9), expected)
+	require.Equal(t, int64(9), activeVersion)
+}
+
+func TestConfigManagerUndecryptableTokenStillFailsClosedForBlockingIntent(t *testing.T) {
+	persisted := `{"enabled":true,"blocking_enabled":true,"config_version":9,"endpoints":[{"id":"g1","name":"Guard","protocol":"openai_compatible","base_url":"http://127.0.0.1:8080","model":"m","token_ciphertext":"undecryptable","timeout_ms":1000,"input_limit":1000,"enabled":true}]}`
+	manager := NewConfigManager(nil, staticSettingRepository{values: map[string]string{
+		SettingKeyPromptAuditConfig: persisted,
+		SettingKeyRiskControl:       "true",
+	}}, nil, prefixEncryptor{}, testTotpKeyConfig())
+	require.NoError(t, manager.Reload(context.Background()))
+	require.Equal(t, ModeBlocking, manager.EffectiveMode())
+
+	service := &PromptService{config: manager, evaluator: NewGuardEvaluator(NewOpenAICompatibleScanner(), nil, nil)}
+	decision, err := service.Evaluate(context.Background(), Request{
+		Protocol: "openai_chat_completions",
+		Body:     []byte(`{"messages":[{"role":"user","content":"hi"}]}`),
+	})
+	require.Error(t, err, "blocking intent with no usable endpoint must not let requests pass unaudited")
+	require.Nil(t, decision)
+	var guardErr *GuardError
+	require.ErrorAs(t, err, &guardErr)
+	require.Equal(t, ErrorCodeUnavailable, guardErr.Code)
+}
+
 func TestBuildNextStoragePreserveReplaceAndClearToken(t *testing.T) {
-	manager := &ConfigManager{encryptor: prefixEncryptor{}}
+	manager := &ConfigManager{encryptor: prefixEncryptor{}, encryptionKeyConfigured: true}
 	current := DefaultStorageConfig()
 	current.Endpoints = []StorageEndpoint{{ID: "one", Name: "One", Protocol: "openai_compatible", BaseURL: "http://127.0.0.1:8080", Model: DefaultGuardModel, TokenCiphertext: "enc:old", TimeoutMS: 1000, InputLimit: 1000}}
 	base := UpdateConfigRequest{ExpectedConfigVersion: 1, Strategy: "priority", WorkerCount: 1, QueueCapacity: 10, Scanners: []string{"PII"}, AllGroups: true,
@@ -130,6 +203,37 @@ func TestBuildNextStoragePreserveReplaceAndClearToken(t *testing.T) {
 	require.Empty(t, cleared.Endpoints[0].TokenCiphertext)
 }
 
+// Without a fixed encryption key the per-boot auto-generated key would make a
+// freshly saved token undecryptable after the next restart (issue #4887), so
+// saving a new token must be rejected with an actionable error. Preserving or
+// clearing an existing ciphertext stays allowed so admins can still edit or
+// disable the feature.
+func TestBuildNextStorageRejectsNewTokenWithoutConfiguredEncryptionKey(t *testing.T) {
+	manager := &ConfigManager{encryptor: prefixEncryptor{}, encryptionKeyConfigured: false}
+	current := DefaultStorageConfig()
+	current.Endpoints = []StorageEndpoint{{ID: "one", Name: "One", Protocol: "openai_compatible", BaseURL: "http://127.0.0.1:8080", Model: DefaultGuardModel, TokenCiphertext: "enc:old", TimeoutMS: 1000, InputLimit: 1000}}
+	base := UpdateConfigRequest{ExpectedConfigVersion: 1, Strategy: "priority", WorkerCount: 1, QueueCapacity: 10, Scanners: []string{"PII"}, AllGroups: true,
+		Endpoints: []UpdateEndpoint{{ID: "one", Name: "One", Protocol: "openai_compatible", BaseURL: "http://127.0.0.1:8080", TimeoutMS: 1000, InputLimit: 1000}}}
+
+	newTokenReq := base
+	newTokenReq.Endpoints = append([]UpdateEndpoint(nil), base.Endpoints...)
+	newTokenReq.Endpoints[0].Token = "fresh-token"
+	_, err := manager.buildNextStorage(current, newTokenReq, 9)
+	require.Error(t, err)
+	require.Equal(t, ErrorCodeEncryptionKeyRequired, infraerrors.Reason(err))
+
+	preserved, err := manager.buildNextStorage(current, base, 9)
+	require.NoError(t, err)
+	require.Equal(t, "enc:old", preserved.Endpoints[0].TokenCiphertext)
+
+	clearedReq := base
+	clearedReq.Endpoints = append([]UpdateEndpoint(nil), base.Endpoints...)
+	clearedReq.Endpoints[0].ClearToken = true
+	cleared, err := manager.buildNextStorage(current, clearedReq, 9)
+	require.NoError(t, err)
+	require.Empty(t, cleared.Endpoints[0].TokenCiphertext)
+}
+
 func TestEffectiveModeTruthTable(t *testing.T) {
 	tests := []struct {
 		risk, enabled, blocking bool
@@ -203,7 +307,7 @@ func (r *switchableSettingRepository) GetMultiple(ctx context.Context, keys []st
 func TestConfigManagerStartupLoadFailureDoesNotBlockWhenBlockingNotIntended(t *testing.T) {
 	// Settings unavailable and no prior blocking intent: stay ModeOff so the
 	// gateway remains usable and admins can still disable/configure Prompt Audit.
-	manager := NewConfigManager(nil, errorSettingRepository{}, nil, prefixEncryptor{})
+	manager := NewConfigManager(nil, errorSettingRepository{}, nil, prefixEncryptor{}, testTotpKeyConfig())
 	err := manager.Start(context.Background())
 	require.Error(t, err)
 	require.True(t, manager.configUntrusted.Load())
@@ -222,7 +326,7 @@ func TestConfigManagerStartupLoadFailureDoesNotBlockWhenBlockingNotIntended(t *t
 }
 
 func TestConfigManagerStartupLoadFailureFailsClosedWhenBlockingIntended(t *testing.T) {
-	manager := NewConfigManager(nil, errorSettingRepository{}, nil, prefixEncryptor{})
+	manager := NewConfigManager(nil, errorSettingRepository{}, nil, prefixEncryptor{}, testTotpKeyConfig())
 	// Simulate intent observed before a later load failure (e.g. decrypt error).
 	manager.observeExpectedState(`{"enabled":true,"blocking_enabled":true,"config_version":3}`, true)
 	manager.markConfigUntrusted()

backend/internal/securityaudit/prompt_handler_test.go

@@ -161,7 +161,7 @@ func TestPromptAdminConfigRequiresVersionMapsConflictAndNeverEchoesToken(t *test
 func TestPromptAdminGetConfigReturnsSecretFreeUnavailableError(t *testing.T) {
 	const canary = "persisted-config-secret-canary"
 	repository := &switchableSettingRepository{loadErr: errors.New("failed to load token " + canary)}
-	manager := NewConfigManager(nil, repository, nil, prefixEncryptor{})
+	manager := NewConfigManager(nil, repository, nil, prefixEncryptor{}, testTotpKeyConfig())
 	require.Error(t, manager.Reload(context.Background()))
 	service := &PromptService{config: manager}
 

backend/internal/securityaudit/prompt_logging.go

@@ -10,6 +10,7 @@ const (
 	EventConfigUpdated        = "prompt_audit.config_updated"
 	EventConfigLoaded         = "prompt_guard.config_loaded"
 	EventConfigReloadDegraded = "prompt_guard.config_reload_degraded"
+	EventConfigTokenInvalid   = "prompt_guard.config_token_invalid"
 	EventProbeStarted         = "prompt_audit.endpoint_probe_started"
 	EventProbeFinished        = "prompt_audit.endpoint_probe_finished"
 	EventProbeFailed          = "prompt_audit.endpoint_probe_failed"
@@ -37,7 +38,7 @@ const (
 )
 
 var knownLogEvents = map[string]struct{}{
-	EventConfigUpdated: {}, EventConfigLoaded: {}, EventConfigReloadDegraded: {},
+	EventConfigUpdated: {}, EventConfigLoaded: {}, EventConfigReloadDegraded: {}, EventConfigTokenInvalid: {},
 	EventProbeStarted: {}, EventProbeFinished: {}, EventProbeFailed: {},
 	EventJobEnqueued: {}, EventEnqueueSkipped: {}, EventEnqueueDropped: {},
 	EventAuditStarted: {}, EventProcessingReclaimed: {}, EventProcessed: {}, EventProcessFailed: {}, EventFindingRecorded: {},

backend/internal/securityaudit/prompt_logging_test.go

@@ -33,7 +33,7 @@ func TestPromptAuditLogAllowlistAndErrorsDoNotLeakCanarySecrets(t *testing.T) {
 	beforeUnknown := output.Len()
 	LogWarn("prompt_audit.typo_event", map[string]any{"status": "failed"})
 	require.Equal(t, beforeUnknown, output.Len(), "events outside the stable dictionary must not be emitted")
-	require.Len(t, knownLogEvents, 27)
+	require.Len(t, knownLogEvents, 28)
 
 	_, err := NormalizeBaseURL("https://guard.example.test/path?token=" + canary)
 	require.Error(t, err)

backend/internal/securityaudit/prompt_service_test.go

@@ -39,7 +39,7 @@ func TestPromptServiceHasExplicitIdempotentLifecycle(t *testing.T) {
 	config := NewConfigManager(nil, staticSettingRepository{values: map[string]string{
 		SettingKeyPromptAuditConfig: "",
 		SettingKeyRiskControl:       "false",
-	}}, nil, prefixEncryptor{})
+	}}, nil, prefixEncryptor{}, testTotpKeyConfig())
 	service := NewPromptService(
 		config,
 		NewPostgreSQLRepository(nil),

backend/internal/securityaudit/prompt_types.go

@@ -12,12 +12,13 @@ const (
 	ConfigInvalidationChannel = "sub2api:prompt_guard:config:invalidate"
 	PayloadKeyPrefix          = "sub2api:prompt_audit:payload:"
 
-	ErrorCodeBlocked           = "prompt_guard_blocked"
-	ErrorCodeUnavailable       = "prompt_guard_unavailable"
-	ErrorCodeInvalidResponse   = "prompt_guard_invalid_response"
-	ErrorCodeConfigConflict    = "prompt_audit_config_conflict"
-	ErrorCodeConfigUnavailable = "prompt_audit_config_unavailable"
-	ErrorCodeRequiresEnabled   = "prompt_guard_requires_audit_enabled"
+	ErrorCodeBlocked               = "prompt_guard_blocked"
+	ErrorCodeUnavailable           = "prompt_guard_unavailable"
+	ErrorCodeInvalidResponse       = "prompt_guard_invalid_response"
+	ErrorCodeConfigConflict        = "prompt_audit_config_conflict"
+	ErrorCodeConfigUnavailable     = "prompt_audit_config_unavailable"
+	ErrorCodeEncryptionKeyRequired = "prompt_audit_encryption_key_required"
+	ErrorCodeRequiresEnabled       = "prompt_guard_requires_audit_enabled"
 
 	DefaultGuardModel = "sileader/qwen3guard:0.6b"
 )

frontend/src/features/prompt-audit/__tests__/components.spec.ts

@@ -50,6 +50,21 @@ describe('Prompt Audit components', () => {
     expect(wrapper.emitted('probe')?.[0]?.[0]).toMatchObject({ id: 'guard-1' })
   })
 
+  it('surfaces an undecryptable saved credential and prompts for re-entry', async () => {
+    const invalidEndpoint = { ...endpoint(), token_status: 'invalid' }
+    const wrapper = mount(EndpointPool, {
+      props: { endpoints: [invalidEndpoint], probeResults: {}, probingIds: [] },
+      global: { stubs: { BaseDialog: DialogStub } },
+    })
+    expect(wrapper.text()).toContain('admin.promptAudit.pool.invalid')
+    expect(wrapper.text()).not.toContain('admin.promptAudit.pool.configured')
+
+    const edit = wrapper.findAll('button').find((button) => button.text().includes('common.edit'))
+    await edit!.trigger('click')
+    const token = wrapper.get<HTMLInputElement>('[aria-label="admin.promptAudit.pool.apiKey"]')
+    expect(token.attributes('placeholder')).toContain('admin.promptAudit.pool.reenterSecret')
+  })
+
   it('supports group search, stale configured groups, nine scanners, and bounded worker inputs', async () => {
     const draft: PromptAuditDraft = {
       enabled: true, blocking_enabled: false, store_pass_events: false, effective_mode: 'async_audit', strategy: 'priority',

frontend/src/features/prompt-audit/components/EndpointPool.vue

@@ -68,9 +68,9 @@
 
           <div class="min-w-0">
             <p class="mb-1 text-[10px] font-semibold uppercase tracking-wider text-gray-400 xl:hidden">{{ t('admin.promptAudit.pool.credential') }}</p>
-            <div class="flex items-center gap-1.5 text-xs font-medium" :class="hasCredential(endpoint) ? 'text-emerald-700 dark:text-emerald-300' : 'text-gray-500 dark:text-dark-400'">
-              <span class="h-1.5 w-1.5 rounded-full" :class="hasCredential(endpoint) ? 'bg-emerald-500' : 'bg-gray-300 dark:bg-dark-500'" aria-hidden="true" />
-              {{ hasCredential(endpoint) ? t('admin.promptAudit.pool.configured') : t('admin.promptAudit.pool.missing') }}
+            <div class="flex items-center gap-1.5 text-xs font-medium" :class="credentialInvalid(endpoint) ? 'text-red-600 dark:text-red-300' : hasCredential(endpoint) ? 'text-emerald-700 dark:text-emerald-300' : 'text-gray-500 dark:text-dark-400'">
+              <span class="h-1.5 w-1.5 rounded-full" :class="credentialInvalid(endpoint) ? 'bg-red-500' : hasCredential(endpoint) ? 'bg-emerald-500' : 'bg-gray-300 dark:bg-dark-500'" aria-hidden="true" />
+              {{ credentialInvalid(endpoint) ? t('admin.promptAudit.pool.invalid') : hasCredential(endpoint) ? t('admin.promptAudit.pool.configured') : t('admin.promptAudit.pool.missing') }}
             </div>
             <p v-if="probingIds.includes(endpoint.id)" class="mt-1.5 text-xs text-primary-600 dark:text-primary-300">
               {{ t('admin.promptAudit.pool.probeProgress') }}
@@ -108,7 +108,7 @@
         </label>
         <label class="space-y-1 text-sm text-gray-700 dark:text-dark-200 sm:col-span-2">
           <span>{{ t('admin.promptAudit.pool.apiKey') }}</span>
-          <input v-model="editing.token" class="input w-full" type="password" autocomplete="new-password" :placeholder="editing.has_token ? t('admin.promptAudit.pool.keepSecret') : ''" :aria-label="t('admin.promptAudit.pool.apiKey')" />
+          <input v-model="editing.token" class="input w-full" type="password" autocomplete="new-password" :placeholder="editing.has_token ? (editing.token_status === 'invalid' ? t('admin.promptAudit.pool.reenterSecret') : t('admin.promptAudit.pool.keepSecret')) : ''" :aria-label="t('admin.promptAudit.pool.apiKey')" />
           <span class="block text-xs text-gray-500 dark:text-dark-400">{{ t('admin.promptAudit.pool.secretHint') }}</span>
         </label>
         <label v-if="editing.has_token" class="flex items-center gap-2 text-sm text-red-600 dark:text-red-300 sm:col-span-2">
@@ -190,4 +190,7 @@ function removeEndpoint(endpoint: PromptAuditEndpointDraft) {
 function hasCredential(endpoint: PromptAuditEndpointDraft): boolean {
   return Boolean(endpoint.token.trim() || (endpoint.has_token && !endpoint.clear_token))
 }
+function credentialInvalid(endpoint: PromptAuditEndpointDraft): boolean {
+  return endpoint.token_status === 'invalid' && !endpoint.token.trim() && !endpoint.clear_token
+}
 </script>

frontend/src/features/prompt-audit/types.ts

@@ -12,7 +12,7 @@ export interface PromptAuditEndpoint {
   input_limit: number
   enabled: boolean
   has_token: boolean
-  token_status: 'configured' | 'missing' | string
+  token_status: 'configured' | 'missing' | 'invalid' | string
 }
 
 export interface PromptAuditEndpointDraft extends PromptAuditEndpoint {

frontend/src/i18n/locales/en/admin/promptAudit.ts

@@ -44,9 +44,9 @@ export default {
     pool: {
       title: 'Audit pool', description: 'Enabled OpenAI-compatible nodes are tried in order. Probes run from the server network.',
       add: 'Add node', edit: 'Edit node', empty: 'No audit nodes configured.', node: 'Node', model: 'Model', limits: 'Timeout / chunk limit', credential: 'Credential and probe',
-      configured: 'API Key configured', missing: 'API Key missing', probe: 'Test connection', probing: 'Probing…',
+      configured: 'API Key configured', missing: 'API Key missing', invalid: 'API Key cannot be decrypted; re-enter it', probe: 'Test connection', probing: 'Probing…',
       probeProgress: 'Config validated ✓ · request sent · awaiting service response…', probeResult: 'Config ✓ · request ✓ · HTTP {http} · {status} · {latency} ms',
-      name: 'Node name', id: 'Stable node ID', baseUrl: 'Base URL', apiKey: 'API Key', keepSecret: 'Leave blank to keep the saved API Key',
+      name: 'Node name', id: 'Stable node ID', baseUrl: 'Base URL', apiKey: 'API Key', keepSecret: 'Leave blank to keep the saved API Key', reenterSecret: 'The saved API Key cannot be decrypted (encryption key changed); enter a new one',
       secretHint: 'Plaintext exists only in this editor and is cleared immediately after a successful save.', clearSecret: 'Explicitly clear the saved API Key', timeout: 'Total timeout (ms)', inputLimit: 'Unicode characters per chunk',
       toggleNode: 'Toggle node {name}', deleteConfirm: 'Remove “{name}” from the draft? It takes effect after saving.',
     },
@@ -93,6 +93,7 @@ export default {
     errors: {
       loadConfig: 'Unable to load Prompt Audit configuration.', loadRuntime: 'Unable to load Prompt Audit runtime.', loadGroups: 'Unable to load groups.', loadEvents: 'Unable to load audit events.', loadDetail: 'Unable to load event details.', saveConfig: 'Unable to save the configuration.', probe: 'Node probe failed.', delete: 'Unable to delete events.', previewDelete: 'Unable to create a deletion preview. Check the time range.', deleteConfirmation: 'The deletion confirmation is invalid or expired. Preview again.',
       prompt_audit_config_conflict: 'Another administrator updated this configuration. Reload the server version before deciding how to merge your draft.',
+      prompt_audit_encryption_key_required: 'No fixed encryption key is configured, so audit node API Keys would be lost on restart. Set the TOTP_ENCRYPTION_KEY environment variable and restart the service first.',
       prompt_guard_requires_audit_enabled: 'Enable Prompt Audit before synchronous blocking.', prompt_audit_invalid_endpoint: 'The audit node configuration is invalid.', prompt_audit_endpoint_required: 'Enable at least one audit node before enabling Prompt Audit.', prompt_audit_groups_required: 'Select at least one group in selected-group mode.', prompt_audit_scanners_required: 'Enable at least one risk category.',
     },
   },

frontend/src/i18n/locales/zh/admin/promptAudit.ts

@@ -44,9 +44,9 @@ export default {
     pool: {
       title: '审计池', description: '按顺序使用启用的 OpenAI 兼容节点；探测由服务端真实网络环境发起。',
       add: '新增节点', edit: '编辑节点', empty: '尚未配置审计节点。', node: '节点', model: '模型', limits: '超时 / 单片上限', credential: '凭据与探测',
-      configured: 'API Key 已配置', missing: '未配置 API Key', probe: '连接测试', probing: '探测中…',
+      configured: 'API Key 已配置', missing: '未配置 API Key', invalid: 'API Key 无法解密，请重新输入', probe: '连接测试', probing: '探测中…',
       probeProgress: '配置校验 ✓ · 请求已发送 · 等待服务响应…', probeResult: '配置校验 ✓ · 请求 ✓ · HTTP {http} · {status} · {latency} ms',
-      name: '节点名称', id: '稳定节点 ID', baseUrl: 'Base URL', apiKey: 'API Key', keepSecret: '留空以保留已保存的 API Key',
+      name: '节点名称', id: '稳定节点 ID', baseUrl: 'Base URL', apiKey: 'API Key', keepSecret: '留空以保留已保存的 API Key', reenterSecret: '已保存的 API Key 无法解密（加密密钥已变更），请重新输入',
       secretHint: '明文只在本次编辑内存中存在；保存成功后会立即清除。', clearSecret: '显式清除已保存的 API Key', timeout: '总超时（毫秒）', inputLimit: '单片 Unicode 字符上限',
       toggleNode: '切换节点 {name}', deleteConfirm: '从草稿中删除节点“{name}”？保存配置后生效。',
     },
@@ -93,6 +93,7 @@ export default {
     errors: {
       loadConfig: '无法加载提示词审计配置。', loadRuntime: '无法加载提示词审计运行态。', loadGroups: '无法加载分组列表。', loadEvents: '无法加载审计事件。', loadDetail: '无法加载事件详情。', saveConfig: '配置保存失败。', probe: '节点探测失败。', delete: '事件删除失败。', previewDelete: '无法生成删除预览，请检查时间范围。', deleteConfirmation: '删除确认无效或已过期，请重新预览。',
       prompt_audit_config_conflict: '配置已被其他管理员更新。请重新加载服务端配置，再决定如何合并本地草稿。',
+      prompt_audit_encryption_key_required: '未配置固定加密密钥，审计节点 API Key 将在服务重启后失效。请先设置 TOTP_ENCRYPTION_KEY 环境变量并重启服务。',
       prompt_guard_requires_audit_enabled: '开启同步阻止前必须先启用提示词审计。', prompt_audit_invalid_endpoint: '审计节点配置无效。', prompt_audit_endpoint_required: '启用审计前至少需要一个启用节点。', prompt_audit_groups_required: '指定分组模式至少需要选择一个分组。', prompt_audit_scanners_required: '至少需要启用一个风险分类。',
     },
   },