feat(auth): active credential health prober

internal/config/config.go

@@ -25,6 +25,9 @@ type Config struct {
 	// CredentialInFlight configures credential observation snapshots.
 	CredentialInFlight CredentialInFlightConfig `yaml:"credential-in-flight" json:"credential-in-flight"`
 
+	// CredentialProber configures optional active credential health probing.
+	CredentialProber CredentialProberConfig `yaml:"credential-prober" json:"credential-prober"`
+
 	// RemoteManagement nests management-related options under 'remote-management'.
 	RemoteManagement RemoteManagement `yaml:"remote-management" json:"-"`
 

internal/config/prober.go

@@ -0,0 +1,49 @@
+package config
+
+import "time"
+
+const (
+	defaultCredentialProberInterval       = 60 * time.Second
+	defaultCredentialProberTimeout        = 10 * time.Second
+	defaultCredentialProberMaxConcurrency = 4
+	defaultCredentialProberRatePerMinute  = 60
+	defaultCredentialProberBackoffBase    = 5 * time.Second
+	defaultCredentialProberBackoffMax     = 5 * time.Minute
+	defaultCredentialProberPath           = "/v1/models"
+)
+
+// CredentialProberConfig controls optional active health probing for registered credentials.
+// When enabled, the conductor periodically issues a lightweight HTTP probe per credential
+// and feeds failures into the existing cooldown/suspension machinery.
+type CredentialProberConfig struct {
+	// Enabled turns active credential health probing on. Default false.
+	Enabled bool `yaml:"enabled" json:"enabled"`
+	// Interval is the period between probe sweeps. Default 60s.
+	Interval time.Duration `yaml:"interval" json:"interval"`
+	// Timeout is the maximum duration a single probe request may take. Default 10s.
+	Timeout time.Duration `yaml:"timeout" json:"timeout"`
+	// MaxConcurrency limits the number of in-flight probes. Default 4.
+	MaxConcurrency int `yaml:"max-concurrency" json:"max-concurrency"`
+	// RateLimitPerMinute caps the number of probe requests across all credentials per minute. Default 60.
+	RateLimitPerMinute int `yaml:"rate-limit-per-minute" json:"rate-limit-per-minute"`
+	// BackoffBase is the initial cooldown applied when a probe fails. Default 5s.
+	BackoffBase time.Duration `yaml:"backoff-base" json:"backoff-base"`
+	// BackoffMax is the maximum probe-induced cooldown. Default 5m.
+	BackoffMax time.Duration `yaml:"backoff-max" json:"backoff-max"`
+	// DefaultProbePath is the HTTP path appended to the credential base_url for the probe. Default /v1/models.
+	DefaultProbePath string `yaml:"default-probe-path" json:"default-probe-path"`
+}
+
+// DefaultCredentialProberConfig returns the prober default configuration.
+func DefaultCredentialProberConfig() CredentialProberConfig {
+	return CredentialProberConfig{
+		Enabled:            false,
+		Interval:           defaultCredentialProberInterval,
+		Timeout:            defaultCredentialProberTimeout,
+		MaxConcurrency:     defaultCredentialProberMaxConcurrency,
+		RateLimitPerMinute: defaultCredentialProberRatePerMinute,
+		BackoffBase:        defaultCredentialProberBackoffBase,
+		BackoffMax:         defaultCredentialProberBackoffMax,
+		DefaultProbePath:   defaultCredentialProberPath,
+	}
+}

sdk/cliproxy/auth/conductor.go

@@ -157,6 +157,10 @@ type Manager struct {
 	refreshCancel context.CancelFunc
 	refreshLoop   *authAutoRefreshLoop
 
+	// Active credential prober state
+	proberCancel context.CancelFunc
+	proberLoop   *authProberLoop
+
 	requestPrepareLocks sync.Map
 	// refreshLocks serializes credential refresh per auth ID so concurrent
 	// 401 recoveries and auto-refresh workers do not race the same refresh_token.

sdk/cliproxy/auth/conductor_cooldown.go

@@ -110,10 +110,12 @@ func (m *Manager) SetConfig(cfg *internalconfig.Config) {
 		return
 	}
 	m.configCooldownMu.Lock()
-	defer m.configCooldownMu.Unlock()
-	if m.setConfigSnapshotLocked(cfg) {
+	cleared := m.setConfigSnapshotLocked(cfg)
+	if cleared {
 		m.persistCooldownStatesLocked(context.Background())
 	}
+	m.configCooldownMu.Unlock()
+	m.restartProber(cfg)
 }
 
 // SetConfigSnapshot updates only in-memory configuration state. It reports whether
@@ -123,8 +125,10 @@ func (m *Manager) SetConfigSnapshot(cfg *internalconfig.Config) bool {
 		return false
 	}
 	m.configCooldownMu.Lock()
-	defer m.configCooldownMu.Unlock()
-	return m.setConfigSnapshotLocked(cfg)
+	cleared := m.setConfigSnapshotLocked(cfg)
+	m.configCooldownMu.Unlock()
+	m.restartProber(cfg)
+	return cleared
 }
 
 func (m *Manager) setConfigSnapshotLocked(cfg *internalconfig.Config) bool {
@@ -171,26 +175,31 @@ func (m *Manager) ApplyConfigWithCooldownStateStore(ctx context.Context, cfg *in
 	}
 
 	m.configCooldownMu.Lock()
-	defer m.configCooldownMu.Unlock()
 	m.mu.RLock()
 	oldStore := m.cooldownStore
 	m.mu.RUnlock()
 	m.setConfigSnapshotLocked(cfg)
 	if oldStore != nil && !m.persistCooldownStatesToLocked(ctx, oldStore) {
+		m.configCooldownMu.Unlock()
 		return false
 	}
 	if errContext := ctx.Err(); errContext != nil {
+		m.configCooldownMu.Unlock()
 		return false
 	}
 	m.mu.Lock()
-	defer m.mu.Unlock()
 	if m.cooldownStore != oldStore {
+		m.mu.Unlock()
+		m.configCooldownMu.Unlock()
 		return false
 	}
 	if m.pendingCooldownStateStore == oldStore {
 		m.pendingCooldownStateStore = nil
 	}
 	m.cooldownStore = store
+	m.mu.Unlock()
+	m.configCooldownMu.Unlock()
+	m.restartProber(cfg)
 	return true
 }