Merge pull request #2971 from sususu98/feat/antigravity-credits-fallback

feat(antigravity): conductor-level credits fallback for Claude models

config.example.yaml

@@ -98,7 +98,7 @@ disable-cooling: false
 quota-exceeded:
   switch-project: true # Whether to automatically switch to another project when a quota is exceeded
   switch-preview-model: true # Whether to automatically switch to a preview model when a quota is exceeded
-  antigravity-credits: true # Whether to retry Antigravity quota_exhausted 429s once with enabledCreditTypes=["GOOGLE_ONE_AI"]
+  antigravity-credits: true # Whether to use credits as last-resort fallback when all free-tier auths are exhausted for Claude models
 
 # Routing strategy for selecting credentials when multiple match.
 routing:

internal/config/config.go

@@ -206,8 +206,9 @@ type QuotaExceeded struct {
 	// SwitchPreviewModel indicates whether to automatically switch to a preview model when a quota is exceeded.
 	SwitchPreviewModel bool `yaml:"switch-preview-model" json:"switch-preview-model"`
 
-	// AntigravityCredits indicates whether to retry Antigravity quota_exhausted 429s once
+	// AntigravityCredits enables credits-based last-resort fallback for Claude models.
-	// on the same credential with enabledCreditTypes=["GOOGLE_ONE_AI"].
+	// When all free-tier auths are exhausted (429/503), the conductor retries with
+	// an auth that has available Google One AI credits.
 	AntigravityCredits bool `yaml:"antigravity-credits" json:"antigravity-credits"`
 }
 

internal/runtime/executor/antigravity_executor.go

@@ -52,8 +52,8 @@ const (
 	defaultAntigravityAgent                = "antigravity/1.21.9 darwin/arm64" // fallback only; overridden at runtime by misc.AntigravityUserAgent()
 	antigravityAuthType                    = "antigravity"
 	refreshSkew                            = 3000 * time.Second
-	antigravityCreditsRetryTTL             = 5 * time.Hour
+	antigravityCreditsHintRefreshInterval  = 10 * time.Minute
-	antigravityCreditsAutoDisableDuration  = 5 * time.Hour
+	antigravityCreditsHintRefreshTimeout   = 5 * time.Second
 	antigravityShortQuotaCooldownThreshold = 5 * time.Minute
 	antigravityInstantRetryThreshold       = 3 * time.Second
 	// systemInstruction              = "You are Antigravity, a powerful agentic AI coding assistant designed by the Google Deepmind team working on Advanced Agentic Coding.You are pair programming with a USER to solve their coding task. The task may require creating a new codebase, modifying or debugging an existing codebase, or simply answering a question.**Absolute paths only****Proactiveness**"
@@ -62,8 +62,6 @@ const (
 type antigravity429Category string
 
 type antigravityCreditsFailureState struct {
-	Count                    int
-	DisabledUntil            time.Time
 	PermanentlyDisabled      bool
 	ExplicitBalanceExhausted bool
 }
@@ -91,28 +89,85 @@ var (
 	randSource                        = rand.New(rand.NewSource(time.Now().UnixNano()))
 	randSourceMutex                   sync.Mutex
 	antigravityCreditsFailureByAuth   sync.Map
-	antigravityPreferCreditsByModel   sync.Map
 	antigravityShortCooldownByAuth    sync.Map
+	antigravityCreditsBalanceByAuth   sync.Map // auth.ID → antigravityCreditsBalance
+	antigravityCreditsHintRefreshByID sync.Map // auth.ID → *antigravityCreditsHintRefreshState
 	antigravityQuotaExhaustedKeywords = []string{
 		"quota_exhausted",
 		"quota exhausted",
 	}
-	antigravityCreditsExhaustedKeywords = []string{
-		"google_one_ai",
-		"insufficient credit",
-		"insufficient credits",
-		"not enough credit",
-		"not enough credits",
-		"credit exhausted",
-		"credits exhausted",
-		"credit balance",
-		"minimumcreditamountforusage",
-		"minimum credit amount for usage",
-		"minimum credit",
-		"resource has been exhausted",
-	}
 )
 
+type antigravityCreditsBalance struct {
+	CreditAmount    float64
+	MinCreditAmount float64
+	PaidTierID      string
+	Known           bool
+}
+
+type antigravityCreditsHintRefreshState struct {
+	mu          sync.Mutex
+	lastAttempt time.Time
+}
+
+func antigravityAuthHasCredits(auth *cliproxyauth.Auth) bool {
+	if auth == nil || strings.TrimSpace(auth.ID) == "" {
+		return false
+	}
+	if hint, ok := cliproxyauth.GetAntigravityCreditsHint(auth.ID); ok && hint.Known {
+		return hint.Available
+	}
+	val, ok := antigravityCreditsBalanceByAuth.Load(strings.TrimSpace(auth.ID))
+	if !ok {
+		return true // optimistic: assume credits available when balance unknown
+	}
+	bal, valid := val.(antigravityCreditsBalance)
+	if !valid {
+		antigravityCreditsBalanceByAuth.Delete(strings.TrimSpace(auth.ID))
+		return false
+	}
+	if !bal.Known {
+		return false
+	}
+	available := bal.CreditAmount >= bal.MinCreditAmount
+	cliproxyauth.SetAntigravityCreditsHint(strings.TrimSpace(auth.ID), cliproxyauth.AntigravityCreditsHint{
+		Known:           true,
+		Available:       available,
+		CreditAmount:    bal.CreditAmount,
+		MinCreditAmount: bal.MinCreditAmount,
+		PaidTierID:      bal.PaidTierID,
+		UpdatedAt:       time.Now(),
+	})
+	return available
+}
+
+// parseMetaFloat extracts a float64 from auth.Metadata (handles string and numeric types).
+func parseMetaFloat(metadata map[string]any, key string) (float64, bool) {
+	v, ok := metadata[key]
+	if !ok {
+		return 0, false
+	}
+	switch typed := v.(type) {
+	case float64:
+		return typed, true
+	case int:
+		return float64(typed), true
+	case int64:
+		return float64(typed), true
+	case uint64:
+		return float64(typed), true
+	case json.Number:
+		if f, err := typed.Float64(); err == nil {
+			return f, true
+		}
+	case string:
+		if f, err := strconv.ParseFloat(strings.TrimSpace(typed), 64); err == nil {
+			return f, true
+		}
+	}
+	return 0, false
+}
+
 // AntigravityExecutor proxies requests to the antigravity upstream.
 type AntigravityExecutor struct {
 	cfg *config.Config
@@ -189,7 +244,7 @@ func validateAntigravityRequestSignatures(from sdktranslator.Format, rawJSON []b
 	if from.String() != "claude" {
 		return rawJSON, nil
 	}
-	// Always strip thinking blocks with empty signatures (proxy-generated).
+	// Always strip thinking blocks with invalid signatures (empty or non-Claude-format).
 	rawJSON = antigravityclaude.StripEmptySignatureThinkingBlocks(rawJSON)
 	if cache.SignatureCacheEnabled() {
 		return rawJSON, nil
@@ -298,26 +353,13 @@ func decideAntigravity429(body []byte) antigravity429Decision {
 		decision.retryAfter = retryAfter
 	}
 
-	lowerBody := strings.ToLower(string(body))
-	for _, keyword := range antigravityQuotaExhaustedKeywords {
-		if strings.Contains(lowerBody, keyword) {
-			decision.kind = antigravity429DecisionFullQuotaExhausted
-			decision.reason = "quota_exhausted"
-			return decision
-		}
-	}
-
 	status := strings.TrimSpace(gjson.GetBytes(body, "error.status").String())
 	if !strings.EqualFold(status, "RESOURCE_EXHAUSTED") {
 		return decision
 	}
 
 	details := gjson.GetBytes(body, "error.details")
-	if !details.Exists() || !details.IsArray() {
+	if details.Exists() && details.IsArray() {
-		decision.kind = antigravity429DecisionSoftRetry
-		return decision
-	}
-
 		for _, detail := range details.Array() {
 			if detail.Get("@type").String() != "type.googleapis.com/google.rpc.ErrorInfo" {
 				continue
@@ -344,86 +386,25 @@ func decideAntigravity429(body []byte) antigravity429Decision {
 				return decision
 			}
 		}
+	}
+
+	lowerBody := strings.ToLower(string(body))
+	for _, keyword := range antigravityQuotaExhaustedKeywords {
+		if strings.Contains(lowerBody, keyword) {
+			decision.kind = antigravity429DecisionFullQuotaExhausted
+			decision.reason = "quota_exhausted"
+			return decision
+		}
+	}
 
 	decision.kind = antigravity429DecisionSoftRetry
 	return decision
 }
 
-func antigravityHasQuotaResetDelayOrModelInfo(body []byte) bool {
-	if len(body) == 0 {
-		return false
-	}
-	details := gjson.GetBytes(body, "error.details")
-	if !details.Exists() || !details.IsArray() {
-		return false
-	}
-	for _, detail := range details.Array() {
-		if detail.Get("@type").String() != "type.googleapis.com/google.rpc.ErrorInfo" {
-			continue
-		}
-		if strings.TrimSpace(detail.Get("metadata.quotaResetDelay").String()) != "" {
-			return true
-		}
-		if strings.TrimSpace(detail.Get("metadata.model").String()) != "" {
-			return true
-		}
-	}
-	return false
-}
-
 func antigravityCreditsRetryEnabled(cfg *config.Config) bool {
 	return cfg != nil && cfg.QuotaExceeded.AntigravityCredits
 }
 
-func antigravityCreditsFailureStateForAuth(auth *cliproxyauth.Auth) (string, antigravityCreditsFailureState, bool) {
-	if auth == nil || strings.TrimSpace(auth.ID) == "" {
-		return "", antigravityCreditsFailureState{}, false
-	}
-	authID := strings.TrimSpace(auth.ID)
-	value, ok := antigravityCreditsFailureByAuth.Load(authID)
-	if !ok {
-		return authID, antigravityCreditsFailureState{}, true
-	}
-	state, ok := value.(antigravityCreditsFailureState)
-	if !ok {
-		antigravityCreditsFailureByAuth.Delete(authID)
-		return authID, antigravityCreditsFailureState{}, true
-	}
-	return authID, state, true
-}
-
-func antigravityCreditsDisabled(auth *cliproxyauth.Auth, now time.Time) bool {
-	authID, state, ok := antigravityCreditsFailureStateForAuth(auth)
-	if !ok {
-		return false
-	}
-	if state.PermanentlyDisabled {
-		return true
-	}
-	if state.DisabledUntil.IsZero() {
-		return false
-	}
-	if state.DisabledUntil.After(now) {
-		return true
-	}
-	antigravityCreditsFailureByAuth.Delete(authID)
-	return false
-}
-
-func recordAntigravityCreditsFailure(auth *cliproxyauth.Auth, now time.Time) {
-	authID, state, ok := antigravityCreditsFailureStateForAuth(auth)
-	if !ok {
-		return
-	}
-	if state.PermanentlyDisabled {
-		antigravityCreditsFailureByAuth.Store(authID, state)
-		return
-	}
-	state.Count++
-	state.DisabledUntil = now.Add(antigravityCreditsAutoDisableDuration)
-	antigravityCreditsFailureByAuth.Store(authID, state)
-}
-
 func clearAntigravityCreditsFailureState(auth *cliproxyauth.Auth) {
 	if auth == nil || strings.TrimSpace(auth.ID) == "" {
 		return
@@ -440,6 +421,25 @@ func markAntigravityCreditsPermanentlyDisabled(auth *cliproxyauth.Auth) {
 		ExplicitBalanceExhausted: true,
 	}
 	antigravityCreditsFailureByAuth.Store(authID, state)
+	antigravityCreditsBalanceByAuth.Store(authID, antigravityCreditsBalance{
+		CreditAmount:    0,
+		MinCreditAmount: 1,
+		Known:           true,
+	})
+	cliproxyauth.SetAntigravityCreditsHint(authID, cliproxyauth.AntigravityCreditsHint{
+		Known:           true,
+		Available:       false,
+		CreditAmount:    0,
+		MinCreditAmount: 1,
+		UpdatedAt:       time.Now(),
+	})
+}
+
+func clearAntigravityCreditsPermanentlyDisabled(auth *cliproxyauth.Auth) {
+	if auth == nil || strings.TrimSpace(auth.ID) == "" {
+		return
+	}
+	antigravityCreditsFailureByAuth.Delete(strings.TrimSpace(auth.ID))
 }
 
 func antigravityHasExplicitCreditsBalanceExhaustedReason(body []byte) bool {
@@ -462,81 +462,6 @@ func antigravityHasExplicitCreditsBalanceExhaustedReason(body []byte) bool {
 	return false
 }
 
-func antigravityPreferCreditsKey(auth *cliproxyauth.Auth, modelName string) string {
-	if auth == nil {
-		return ""
-	}
-	authID := strings.TrimSpace(auth.ID)
-	modelName = strings.TrimSpace(modelName)
-	if authID == "" || modelName == "" {
-		return ""
-	}
-	return authID + "|" + modelName
-}
-
-func antigravityShouldPreferCredits(auth *cliproxyauth.Auth, modelName string, now time.Time) bool {
-	key := antigravityPreferCreditsKey(auth, modelName)
-	if key == "" {
-		return false
-	}
-	value, ok := antigravityPreferCreditsByModel.Load(key)
-	if !ok {
-		return false
-	}
-	until, ok := value.(time.Time)
-	if !ok || until.IsZero() {
-		antigravityPreferCreditsByModel.Delete(key)
-		return false
-	}
-	if !until.After(now) {
-		antigravityPreferCreditsByModel.Delete(key)
-		return false
-	}
-	return true
-}
-
-func markAntigravityPreferCredits(auth *cliproxyauth.Auth, modelName string, now time.Time, retryAfter *time.Duration) {
-	key := antigravityPreferCreditsKey(auth, modelName)
-	if key == "" {
-		return
-	}
-	until := now.Add(antigravityCreditsRetryTTL)
-	if retryAfter != nil && *retryAfter > 0 {
-		until = now.Add(*retryAfter)
-	}
-	antigravityPreferCreditsByModel.Store(key, until)
-}
-
-func clearAntigravityPreferCredits(auth *cliproxyauth.Auth, modelName string) {
-	key := antigravityPreferCreditsKey(auth, modelName)
-	if key == "" {
-		return
-	}
-	antigravityPreferCreditsByModel.Delete(key)
-}
-
-func shouldMarkAntigravityCreditsExhausted(statusCode int, body []byte, reqErr error) bool {
-	if reqErr != nil || statusCode == 0 {
-		return false
-	}
-	if statusCode >= http.StatusInternalServerError || statusCode == http.StatusRequestTimeout {
-		return false
-	}
-	lowerBody := strings.ToLower(string(body))
-	for _, keyword := range antigravityCreditsExhaustedKeywords {
-		if strings.Contains(lowerBody, keyword) {
-			if keyword == "resource has been exhausted" &&
-				statusCode == http.StatusTooManyRequests &&
-				decideAntigravity429(body).kind == antigravity429DecisionSoftRetry &&
-				!antigravityHasQuotaResetDelayOrModelInfo(body) {
-				return false
-			}
-			return true
-		}
-	}
-	return false
-}
-
 func newAntigravityStatusErr(statusCode int, body []byte) statusErr {
 	err := statusErr{code: statusCode, msg: string(body)}
 	if statusCode == http.StatusTooManyRequests {
@@ -547,129 +472,6 @@ func newAntigravityStatusErr(statusCode int, body []byte) statusErr {
 	return err
 }
 
-func (e *AntigravityExecutor) attemptCreditsFallback(
-	ctx context.Context,
-	auth *cliproxyauth.Auth,
-	httpClient *http.Client,
-	token string,
-	modelName string,
-	payload []byte,
-	stream bool,
-	alt string,
-	baseURL string,
-	originalBody []byte,
-) (*http.Response, bool) {
-	if !antigravityCreditsRetryEnabled(e.cfg) {
-		return nil, false
-	}
-	if decideAntigravity429(originalBody).kind != antigravity429DecisionFullQuotaExhausted {
-		return nil, false
-	}
-	now := time.Now()
-	if shouldForcePermanentDisableCredits(originalBody) {
-		clearAntigravityPreferCredits(auth, modelName)
-		markAntigravityCreditsPermanentlyDisabled(auth)
-		return nil, false
-	}
-
-	if antigravityHasExplicitCreditsBalanceExhaustedReason(originalBody) {
-		clearAntigravityPreferCredits(auth, modelName)
-		markAntigravityCreditsPermanentlyDisabled(auth)
-		return nil, false
-	}
-
-	if antigravityCreditsDisabled(auth, now) {
-		return nil, false
-	}
-	creditsPayload := injectEnabledCreditTypes(payload)
-	if len(creditsPayload) == 0 {
-		return nil, false
-	}
-
-	httpReq, errReq := e.buildRequest(ctx, auth, token, modelName, creditsPayload, stream, alt, baseURL)
-	if errReq != nil {
-		helps.RecordAPIResponseError(ctx, e.cfg, errReq)
-		clearAntigravityPreferCredits(auth, modelName)
-		recordAntigravityCreditsFailure(auth, now)
-		return nil, true
-	}
-	httpResp, errDo := httpClient.Do(httpReq)
-	if errDo != nil {
-		helps.RecordAPIResponseError(ctx, e.cfg, errDo)
-		clearAntigravityPreferCredits(auth, modelName)
-		recordAntigravityCreditsFailure(auth, now)
-		return nil, true
-	}
-	if httpResp.StatusCode >= http.StatusOK && httpResp.StatusCode < http.StatusMultipleChoices {
-		retryAfter, _ := parseRetryDelay(originalBody)
-		markAntigravityPreferCredits(auth, modelName, now, retryAfter)
-		clearAntigravityCreditsFailureState(auth)
-		return httpResp, true
-	}
-
-	helps.RecordAPIResponseMetadata(ctx, e.cfg, httpResp.StatusCode, httpResp.Header.Clone())
-	bodyBytes, errRead := io.ReadAll(httpResp.Body)
-	if errClose := httpResp.Body.Close(); errClose != nil {
-		log.Errorf("antigravity executor: close credits fallback response body error: %v", errClose)
-	}
-	if errRead != nil {
-		helps.RecordAPIResponseError(ctx, e.cfg, errRead)
-		clearAntigravityPreferCredits(auth, modelName)
-		recordAntigravityCreditsFailure(auth, now)
-		return nil, true
-	}
-	helps.AppendAPIResponseChunk(ctx, e.cfg, bodyBytes)
-	if shouldForcePermanentDisableCredits(bodyBytes) {
-		clearAntigravityPreferCredits(auth, modelName)
-		markAntigravityCreditsPermanentlyDisabled(auth)
-		return nil, true
-	}
-
-	if antigravityHasExplicitCreditsBalanceExhaustedReason(bodyBytes) {
-		clearAntigravityPreferCredits(auth, modelName)
-		markAntigravityCreditsPermanentlyDisabled(auth)
-		return nil, true
-	}
-
-	clearAntigravityPreferCredits(auth, modelName)
-	recordAntigravityCreditsFailure(auth, now)
-	return nil, true
-}
-
-func (e *AntigravityExecutor) handleDirectCreditsFailure(ctx context.Context, auth *cliproxyauth.Auth, modelName string, reqErr error) {
-	if reqErr != nil {
-		if shouldForcePermanentDisableCredits(reqErrBody(reqErr)) {
-			clearAntigravityPreferCredits(auth, modelName)
-			markAntigravityCreditsPermanentlyDisabled(auth)
-			return
-		}
-
-		if antigravityHasExplicitCreditsBalanceExhaustedReason(reqErrBody(reqErr)) {
-			clearAntigravityPreferCredits(auth, modelName)
-			markAntigravityCreditsPermanentlyDisabled(auth)
-			return
-		}
-
-		helps.RecordAPIResponseError(ctx, e.cfg, reqErr)
-	}
-	clearAntigravityPreferCredits(auth, modelName)
-	recordAntigravityCreditsFailure(auth, time.Now())
-}
-func reqErrBody(reqErr error) []byte {
-	if reqErr == nil {
-		return nil
-	}
-	msg := reqErr.Error()
-	if strings.TrimSpace(msg) == "" {
-		return nil
-	}
-	return []byte(msg)
-}
-
-func shouldForcePermanentDisableCredits(body []byte) bool {
-	return antigravityHasExplicitCreditsBalanceExhaustedReason(body)
-}
-
 // Execute performs a non-streaming request to the Antigravity API.
 func (e *AntigravityExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (resp cliproxyexecutor.Response, err error) {
 	if opts.Alt == "responses/compact" {
@@ -721,6 +523,8 @@ func (e *AntigravityExecutor) Execute(ctx context.Context, auth *cliproxyauth.Au
 	requestedModel := helps.PayloadRequestedModel(opts, req.Model)
 	translated = helps.ApplyPayloadConfigWithRoot(e.cfg, baseModel, "antigravity", "request", translated, originalTranslated, requestedModel)
 
+	useCredits := cliproxyauth.AntigravityCreditsRequested(ctx) && antigravityCreditsRetryEnabled(e.cfg)
+
 	baseURLs := antigravityBaseURLFallbackOrder(auth)
 	httpClient := newAntigravityHTTPClient(ctx, e.cfg, auth, 0)
 	attempts := antigravityRetryAttempts(auth, e.cfg)
@@ -733,11 +537,10 @@ attemptLoop:
 
 		for idx, baseURL := range baseURLs {
 			requestPayload := translated
-			usedCreditsDirect := false
+			if useCredits {
-			if antigravityCreditsRetryEnabled(e.cfg) && antigravityShouldPreferCredits(auth, baseModel, time.Now()) {
+				if cp := injectEnabledCreditTypes(translated); len(cp) > 0 {
-				if creditsPayload := injectEnabledCreditTypes(translated); len(creditsPayload) > 0 {
+					requestPayload = cp
-					requestPayload = creditsPayload
+					helps.MarkCreditsUsed(ctx)
-					usedCreditsDirect = true
 				}
 			}
 
@@ -785,7 +588,6 @@ attemptLoop:
 							wait := antigravityInstantRetryDelay(*decision.retryAfter)
 							log.Debugf("antigravity executor: instant retry for model %s, waiting %s", baseModel, wait)
 							if errWait := antigravityWait(ctx, wait); errWait != nil {
-
 								return resp, errWait
 							}
 						}
@@ -794,34 +596,13 @@ attemptLoop:
 				case antigravity429DecisionShortCooldownSwitchAuth:
 					if decision.retryAfter != nil && *decision.retryAfter > 0 {
 						markAntigravityShortCooldown(auth, baseModel, time.Now(), *decision.retryAfter)
-						log.Debugf("antigravity executor: short quota cooldown (%s) for model %s, recorded cooldown and skipping credits fallback", *decision.retryAfter, baseModel)
+						log.Debugf("antigravity executor: short quota cooldown (%s) for model %s, recorded cooldown", *decision.retryAfter, baseModel)
 					}
 				case antigravity429DecisionFullQuotaExhausted:
-					if usedCreditsDirect {
+					if useCredits && antigravityHasExplicitCreditsBalanceExhaustedReason(bodyBytes) {
-						clearAntigravityPreferCredits(auth, baseModel)
+						markAntigravityCreditsPermanentlyDisabled(auth)
-						recordAntigravityCreditsFailure(auth, time.Now())
-					} else {
-						creditsResp, _ := e.attemptCreditsFallback(ctx, auth, httpClient, token, baseModel, translated, false, opts.Alt, baseURL, bodyBytes)
-						if creditsResp != nil {
-							helps.RecordAPIResponseMetadata(ctx, e.cfg, creditsResp.StatusCode, creditsResp.Header.Clone())
-							creditsBody, errCreditsRead := io.ReadAll(creditsResp.Body)
-							if errClose := creditsResp.Body.Close(); errClose != nil {
-								log.Errorf("antigravity executor: close credits success response body error: %v", errClose)
-							}
-							if errCreditsRead != nil {
-								helps.RecordAPIResponseError(ctx, e.cfg, errCreditsRead)
-								err = errCreditsRead
-								return resp, err
-							}
-							helps.AppendAPIResponseChunk(ctx, e.cfg, creditsBody)
-							reporter.Publish(ctx, helps.ParseAntigravityUsage(creditsBody))
-							var param any
-							converted := sdktranslator.TranslateNonStream(ctx, to, from, req.Model, opts.OriginalRequest, translated, creditsBody, &param)
-							resp = cliproxyexecutor.Response{Payload: converted, Headers: creditsResp.Header.Clone()}
-							reporter.EnsurePublished(ctx)
-							return resp, nil
-						}
 					}
+					// No credits logic - just fall through to error return below
 				}
 			}
 
@@ -870,6 +651,10 @@ attemptLoop:
 				return resp, err
 			}
 
+			// Success
+			if useCredits {
+				clearAntigravityCreditsFailureState(auth)
+			}
 			reporter.Publish(ctx, helps.ParseAntigravityUsage(bodyBytes))
 			var param any
 			converted := sdktranslator.TranslateNonStream(ctx, to, from, req.Model, opts.OriginalRequest, translated, bodyBytes, &param)
@@ -935,6 +720,8 @@ func (e *AntigravityExecutor) executeClaudeNonStream(ctx context.Context, auth *
 	requestedModel := helps.PayloadRequestedModel(opts, req.Model)
 	translated = helps.ApplyPayloadConfigWithRoot(e.cfg, baseModel, "antigravity", "request", translated, originalTranslated, requestedModel)
 
+	useCredits := cliproxyauth.AntigravityCreditsRequested(ctx) && antigravityCreditsRetryEnabled(e.cfg)
+
 	baseURLs := antigravityBaseURLFallbackOrder(auth)
 	httpClient := newAntigravityHTTPClient(ctx, e.cfg, auth, 0)
 
@@ -948,11 +735,10 @@ attemptLoop:
 
 		for idx, baseURL := range baseURLs {
 			requestPayload := translated
-			usedCreditsDirect := false
+			if useCredits {
-			if antigravityCreditsRetryEnabled(e.cfg) && antigravityShouldPreferCredits(auth, baseModel, time.Now()) {
+				if cp := injectEnabledCreditTypes(translated); len(cp) > 0 {
-				if creditsPayload := injectEnabledCreditTypes(translated); len(creditsPayload) > 0 {
+					requestPayload = cp
-					requestPayload = creditsPayload
+					helps.MarkCreditsUsed(ctx)
-					usedCreditsDirect = true
 				}
 			}
 			httpReq, errReq := e.buildRequest(ctx, auth, token, baseModel, requestPayload, true, opts.Alt, baseURL)
@@ -1014,7 +800,6 @@ attemptLoop:
 								wait := antigravityInstantRetryDelay(*decision.retryAfter)
 								log.Debugf("antigravity executor: instant retry for model %s, waiting %s", baseModel, wait)
 								if errWait := antigravityWait(ctx, wait); errWait != nil {
-
 									return resp, errWait
 								}
 							}
@@ -1023,25 +808,16 @@ attemptLoop:
 					case antigravity429DecisionShortCooldownSwitchAuth:
 						if decision.retryAfter != nil && *decision.retryAfter > 0 {
 							markAntigravityShortCooldown(auth, baseModel, time.Now(), *decision.retryAfter)
-							log.Debugf("antigravity executor: short quota cooldown (%s) for model %s, recorded cooldown and skipping credits fallback", *decision.retryAfter, baseModel)
+							log.Debugf("antigravity executor: short quota cooldown (%s) for model %s, recorded cooldown", *decision.retryAfter, baseModel)
 						}
 					case antigravity429DecisionFullQuotaExhausted:
-						if usedCreditsDirect {
+						if useCredits && antigravityHasExplicitCreditsBalanceExhaustedReason(bodyBytes) {
-							clearAntigravityPreferCredits(auth, baseModel)
+							markAntigravityCreditsPermanentlyDisabled(auth)
-							recordAntigravityCreditsFailure(auth, time.Now())
-						} else {
-							creditsResp, _ := e.attemptCreditsFallback(ctx, auth, httpClient, token, baseModel, translated, true, opts.Alt, baseURL, bodyBytes)
-							if creditsResp != nil {
-								httpResp = creditsResp
-								helps.RecordAPIResponseMetadata(ctx, e.cfg, httpResp.StatusCode, httpResp.Header.Clone())
-							}
 						}
+						// No credits logic - just fall through to error return below
 					}
 				}
 
-				if httpResp.StatusCode >= http.StatusOK && httpResp.StatusCode < http.StatusMultipleChoices {
-					goto streamSuccessClaudeNonStream
-				}
 				lastStatus = httpResp.StatusCode
 				lastBody = append([]byte(nil), bodyBytes...)
 				lastErr = nil
@@ -1085,7 +861,10 @@ attemptLoop:
 				return resp, err
 			}
 
-		streamSuccessClaudeNonStream:
+			// Stream success
+			if useCredits {
+				clearAntigravityCreditsFailureState(auth)
+			}
 			out := make(chan cliproxyexecutor.StreamChunk)
 			go func(resp *http.Response) {
 				defer close(out)
@@ -1389,6 +1168,7 @@ func (e *AntigravityExecutor) ExecuteStream(ctx context.Context, auth *cliproxya
 	if updatedAuth != nil {
 		auth = updatedAuth
 	}
+
 	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, true)
 	translated := sdktranslator.TranslateRequest(from, to, baseModel, req.Payload, true)
 
@@ -1400,6 +1180,8 @@ func (e *AntigravityExecutor) ExecuteStream(ctx context.Context, auth *cliproxya
 	requestedModel := helps.PayloadRequestedModel(opts, req.Model)
 	translated = helps.ApplyPayloadConfigWithRoot(e.cfg, baseModel, "antigravity", "request", translated, originalTranslated, requestedModel)
 
+	useCredits := cliproxyauth.AntigravityCreditsRequested(ctx) && antigravityCreditsRetryEnabled(e.cfg)
+
 	baseURLs := antigravityBaseURLFallbackOrder(auth)
 	httpClient := newAntigravityHTTPClient(ctx, e.cfg, auth, 0)
 
@@ -1413,11 +1195,10 @@ attemptLoop:
 
 		for idx, baseURL := range baseURLs {
 			requestPayload := translated
-			usedCreditsDirect := false
+			if useCredits {
-			if antigravityCreditsRetryEnabled(e.cfg) && antigravityShouldPreferCredits(auth, baseModel, time.Now()) {
+				if cp := injectEnabledCreditTypes(translated); len(cp) > 0 {
-				if creditsPayload := injectEnabledCreditTypes(translated); len(creditsPayload) > 0 {
+					requestPayload = cp
-					requestPayload = creditsPayload
+					helps.MarkCreditsUsed(ctx)
-					usedCreditsDirect = true
 				}
 			}
 			httpReq, errReq := e.buildRequest(ctx, auth, token, baseModel, requestPayload, true, opts.Alt, baseURL)
@@ -1478,7 +1259,6 @@ attemptLoop:
 								wait := antigravityInstantRetryDelay(*decision.retryAfter)
 								log.Debugf("antigravity executor: instant retry for model %s, waiting %s", baseModel, wait)
 								if errWait := antigravityWait(ctx, wait); errWait != nil {
-
 									return nil, errWait
 								}
 							}
@@ -1487,25 +1267,16 @@ attemptLoop:
 					case antigravity429DecisionShortCooldownSwitchAuth:
 						if decision.retryAfter != nil && *decision.retryAfter > 0 {
 							markAntigravityShortCooldown(auth, baseModel, time.Now(), *decision.retryAfter)
-							log.Debugf("antigravity executor: short quota cooldown (%s) for model %s, recorded cooldown and skipping credits fallback", *decision.retryAfter, baseModel)
+							log.Debugf("antigravity executor: short quota cooldown (%s) for model %s recorded", *decision.retryAfter, baseModel)
 						}
 					case antigravity429DecisionFullQuotaExhausted:
-						if usedCreditsDirect {
+						if useCredits && antigravityHasExplicitCreditsBalanceExhaustedReason(bodyBytes) {
-							clearAntigravityPreferCredits(auth, baseModel)
+							markAntigravityCreditsPermanentlyDisabled(auth)
-							recordAntigravityCreditsFailure(auth, time.Now())
-						} else {
-							creditsResp, _ := e.attemptCreditsFallback(ctx, auth, httpClient, token, baseModel, translated, true, opts.Alt, baseURL, bodyBytes)
-							if creditsResp != nil {
-								httpResp = creditsResp
-								helps.RecordAPIResponseMetadata(ctx, e.cfg, httpResp.StatusCode, httpResp.Header.Clone())
-							}
 						}
+						// No credits logic - just fall through to error return below
 					}
 				}
 
-				if httpResp.StatusCode >= http.StatusOK && httpResp.StatusCode < http.StatusMultipleChoices {
-					goto streamSuccessExecuteStream
-				}
 				lastStatus = httpResp.StatusCode
 				lastBody = append([]byte(nil), bodyBytes...)
 				lastErr = nil
@@ -1549,7 +1320,10 @@ attemptLoop:
 				return nil, err
 			}
 
-		streamSuccessExecuteStream:
+			// Stream success
+			if useCredits {
+				clearAntigravityCreditsFailureState(auth)
+			}
 			out := make(chan cliproxyexecutor.StreamChunk)
 			go func(resp *http.Response) {
 				defer close(out)
@@ -1792,6 +1566,7 @@ func (e *AntigravityExecutor) ensureAccessToken(ctx context.Context, auth *clipr
 	accessToken := metaStringValue(auth.Metadata, "access_token")
 	expiry := tokenExpiry(auth.Metadata)
 	if accessToken != "" && expiry.After(time.Now().Add(refreshSkew)) {
+		e.maybeRefreshAntigravityCreditsHint(ctx, auth, accessToken)
 		return accessToken, nil, nil
 	}
 	refreshCtx := context.Background()
@@ -1807,6 +1582,63 @@ func (e *AntigravityExecutor) ensureAccessToken(ctx context.Context, auth *clipr
 	return metaStringValue(updated.Metadata, "access_token"), updated, nil
 }
 
+func (e *AntigravityExecutor) maybeRefreshAntigravityCreditsHint(ctx context.Context, auth *cliproxyauth.Auth, accessToken string) {
+	if e == nil || auth == nil || !antigravityCreditsRetryEnabled(e.cfg) {
+		return
+	}
+	if ctx != nil && ctx.Err() != nil {
+		return
+	}
+	authID := strings.TrimSpace(auth.ID)
+	if authID == "" {
+		return
+	}
+	if hint, ok := cliproxyauth.GetAntigravityCreditsHint(authID); ok && hint.Known {
+		return
+	}
+	if strings.TrimSpace(accessToken) == "" {
+		accessToken = metaStringValue(auth.Metadata, "access_token")
+	}
+	if strings.TrimSpace(accessToken) == "" {
+		return
+	}
+
+	state := &antigravityCreditsHintRefreshState{}
+	if existing, loaded := antigravityCreditsHintRefreshByID.LoadOrStore(authID, state); loaded {
+		if cast, ok := existing.(*antigravityCreditsHintRefreshState); ok && cast != nil {
+			state = cast
+		} else {
+			antigravityCreditsHintRefreshByID.Delete(authID)
+			antigravityCreditsHintRefreshByID.Store(authID, state)
+		}
+	}
+
+	now := time.Now()
+	if !state.mu.TryLock() {
+		return
+	}
+	if !state.lastAttempt.IsZero() && now.Sub(state.lastAttempt) < antigravityCreditsHintRefreshInterval {
+		state.mu.Unlock()
+		return
+	}
+	state.lastAttempt = now
+
+	refreshCtx := context.Background()
+	if ctx != nil {
+		if rt, ok := ctx.Value("cliproxy.roundtripper").(http.RoundTripper); ok && rt != nil {
+			refreshCtx = context.WithValue(refreshCtx, "cliproxy.roundtripper", rt)
+		}
+	}
+	refreshCtx, cancel := context.WithTimeout(refreshCtx, antigravityCreditsHintRefreshTimeout)
+	authCopy := auth.Clone()
+
+	go func(state *antigravityCreditsHintRefreshState, auth *cliproxyauth.Auth, token string) {
+		defer cancel()
+		defer state.mu.Unlock()
+		e.updateAntigravityCreditsBalance(refreshCtx, auth, token)
+	}(state, authCopy, accessToken)
+}
+
 func (e *AntigravityExecutor) refreshToken(ctx context.Context, auth *cliproxyauth.Auth) (*cliproxyauth.Auth, error) {
 	if auth == nil {
 		return nil, statusErr{code: http.StatusUnauthorized, msg: "missing auth"}
@@ -1882,6 +1714,7 @@ func (e *AntigravityExecutor) refreshToken(ctx context.Context, auth *cliproxyau
 	if errProject := e.ensureAntigravityProjectID(ctx, auth, tokenResp.AccessToken); errProject != nil {
 		log.Warnf("antigravity executor: ensure project id failed: %v", errProject)
 	}
+	e.updateAntigravityCreditsBalance(ctx, auth, tokenResp.AccessToken)
 	return auth, nil
 }
 
@@ -1918,6 +1751,94 @@ func (e *AntigravityExecutor) ensureAntigravityProjectID(ctx context.Context, au
 	return nil
 }
 
+func (e *AntigravityExecutor) updateAntigravityCreditsBalance(ctx context.Context, auth *cliproxyauth.Auth, accessToken string) {
+	if auth == nil || strings.TrimSpace(auth.ID) == "" {
+		return
+	}
+	token := strings.TrimSpace(accessToken)
+	if token == "" {
+		token = metaStringValue(auth.Metadata, "access_token")
+	}
+	if token == "" {
+		return
+	}
+
+	loadReqBody := `{"metadata":{"ideType":"ANTIGRAVITY","platform":"PLATFORM_UNSPECIFIED","pluginType":"GEMINI"}}`
+	endpointURL := "https://cloudcode-pa.googleapis.com/v1internal:loadCodeAssist"
+	httpReq, errReq := http.NewRequestWithContext(ctx, http.MethodPost, endpointURL, strings.NewReader(loadReqBody))
+	if errReq != nil {
+		log.Debugf("antigravity executor: create loadCodeAssist request error: %v", errReq)
+		return
+	}
+	httpReq.Header.Set("Authorization", "Bearer "+token)
+	httpReq.Header.Set("Content-Type", "application/json")
+	httpReq.Header.Set("User-Agent", "google-api-nodejs-client/9.15.1")
+
+	httpClient := newAntigravityHTTPClient(ctx, e.cfg, auth, 0)
+	httpResp, errDo := httpClient.Do(httpReq)
+	if errDo != nil {
+		log.Debugf("antigravity executor: loadCodeAssist request error: %v", errDo)
+		return
+	}
+	defer func() {
+		if errClose := httpResp.Body.Close(); errClose != nil {
+			log.Errorf("antigravity executor: close loadCodeAssist response body error: %v", errClose)
+		}
+	}()
+
+	bodyBytes, errRead := io.ReadAll(httpResp.Body)
+	if errRead != nil || httpResp.StatusCode < http.StatusOK || httpResp.StatusCode >= http.StatusMultipleChoices {
+		log.Debugf("antigravity executor: loadCodeAssist returned status %d, err=%v", httpResp.StatusCode, errRead)
+		return
+	}
+
+	authID := strings.TrimSpace(auth.ID)
+	paidTierID := strings.TrimSpace(gjson.GetBytes(bodyBytes, "paidTier.id").String())
+
+	credits := gjson.GetBytes(bodyBytes, "paidTier.availableCredits")
+	if !credits.IsArray() {
+		cliproxyauth.SetAntigravityCreditsHint(authID, cliproxyauth.AntigravityCreditsHint{
+			Known:      true,
+			Available:  false,
+			PaidTierID: paidTierID,
+			UpdatedAt:  time.Now(),
+		})
+		return
+	}
+	for _, credit := range credits.Array() {
+		if !strings.EqualFold(credit.Get("creditType").String(), "GOOGLE_ONE_AI") {
+			continue
+		}
+		creditAmount, errCA := strconv.ParseFloat(strings.TrimSpace(credit.Get("creditAmount").String()), 64)
+		if errCA != nil {
+			continue
+		}
+		minAmount, errMA := strconv.ParseFloat(strings.TrimSpace(credit.Get("minimumCreditAmountForUsage").String()), 64)
+		if errMA != nil {
+			continue
+		}
+		bal := antigravityCreditsBalance{
+			CreditAmount:    creditAmount,
+			MinCreditAmount: minAmount,
+			PaidTierID:      paidTierID,
+			Known:           true,
+		}
+		antigravityCreditsBalanceByAuth.Store(authID, bal)
+		cliproxyauth.SetAntigravityCreditsHint(authID, cliproxyauth.AntigravityCreditsHint{
+			Known:           true,
+			Available:       creditAmount >= minAmount,
+			CreditAmount:    creditAmount,
+			MinCreditAmount: minAmount,
+			PaidTierID:      paidTierID,
+			UpdatedAt:       time.Now(),
+		})
+		if creditAmount >= minAmount {
+			clearAntigravityCreditsPermanentlyDisabled(auth)
+		}
+		return
+	}
+}
+
 func (e *AntigravityExecutor) buildRequest(ctx context.Context, auth *cliproxyauth.Auth, token, modelName string, payload []byte, stream bool, alt, baseURL string) (*http.Request, error) {
 	if token == "" {
 		return nil, statusErr{code: http.StatusUnauthorized, msg: "missing access token"}

internal/runtime/executor/antigravity_executor_credits_test.go

@@ -18,8 +18,9 @@ import (
 
 func resetAntigravityCreditsRetryState() {
 	antigravityCreditsFailureByAuth = sync.Map{}
-	antigravityPreferCreditsByModel = sync.Map{}
 	antigravityShortCooldownByAuth = sync.Map{}
+	antigravityCreditsBalanceByAuth = sync.Map{}
+	antigravityCreditsHintRefreshByID = sync.Map{}
 }
 
 func TestClassifyAntigravity429(t *testing.T) {
@@ -30,6 +31,43 @@ func TestClassifyAntigravity429(t *testing.T) {
 		}
 	})
 
+	t.Run("standard antigravity rate limit with ui message stays rate limited", func(t *testing.T) {
+		body := []byte(`{
+			"error": {
+				"code": 429,
+				"message": "You have exhausted your capacity on this model. Your quota will reset after 0s.",
+				"status": "RESOURCE_EXHAUSTED",
+				"details": [
+					{
+						"@type": "type.googleapis.com/google.rpc.ErrorInfo",
+						"reason": "RATE_LIMIT_EXCEEDED",
+						"domain": "cloudcode-pa.googleapis.com",
+						"metadata": {
+							"model": "claude-opus-4-6-thinking",
+							"quotaResetDelay": "479.417207ms",
+							"quotaResetTimeStamp": "2026-04-20T09:19:49Z",
+							"uiMessage": "true"
+						}
+					},
+					{
+						"@type": "type.googleapis.com/google.rpc.RetryInfo",
+						"retryDelay": "0.479417207s"
+					}
+				]
+			}
+		}`)
+		if got := classifyAntigravity429(body); got != antigravity429RateLimited {
+			t.Fatalf("classifyAntigravity429() = %q, want %q", got, antigravity429RateLimited)
+		}
+		decision := decideAntigravity429(body)
+		if decision.kind != antigravity429DecisionInstantRetrySameAuth {
+			t.Fatalf("decideAntigravity429().kind = %q, want %q", decision.kind, antigravity429DecisionInstantRetrySameAuth)
+		}
+		if decision.retryAfter == nil {
+			t.Fatal("decideAntigravity429().retryAfter = nil")
+		}
+	})
+
 	t.Run("structured rate limit", func(t *testing.T) {
 		body := []byte(`{
 			"error": {
@@ -67,8 +105,31 @@ func TestClassifyAntigravity429(t *testing.T) {
 	})
 }
 
+func TestAntigravityShouldRetryNoCapacity_Standard503(t *testing.T) {
+	body := []byte(`{
+		"error": {
+			"code": 503,
+			"message": "No capacity available for model gemini-3.1-flash-image on the server",
+			"status": "UNAVAILABLE",
+			"details": [
+				{
+					"@type": "type.googleapis.com/google.rpc.ErrorInfo",
+					"reason": "MODEL_CAPACITY_EXHAUSTED",
+					"domain": "cloudcode-pa.googleapis.com",
+					"metadata": {
+						"model": "gemini-3.1-flash-image"
+					}
+				}
+			]
+		}
+	}`)
+	if !antigravityShouldRetryNoCapacity(http.StatusServiceUnavailable, body) {
+		t.Fatal("antigravityShouldRetryNoCapacity() = false, want true")
+	}
+}
+
 func TestInjectEnabledCreditTypes(t *testing.T) {
-	body := []byte(`{"model":"gemini-2.5-flash","request":{}}`)
+	body := []byte(`{"model":"claude-sonnet-4-6","request":{}}`)
 	got := injectEnabledCreditTypes(body)
 	if got == nil {
 		t.Fatal("injectEnabledCreditTypes() returned nil")
@@ -82,34 +143,18 @@ func TestInjectEnabledCreditTypes(t *testing.T) {
 	}
 }
 
-func TestShouldMarkAntigravityCreditsExhausted(t *testing.T) {
+func TestParseRetryDelay_HumanReadableDuration(t *testing.T) {
-	t.Run("credit errors are marked", func(t *testing.T) {
+	body := []byte(`{"error":{"message":"You have exhausted your capacity on this model. Your quota will reset after 1h43m56s."}}`)
-		for _, body := range [][]byte{
+	retryAfter, err := parseRetryDelay(body)
-			[]byte(`{"error":{"message":"Insufficient GOOGLE_ONE_AI credits"}}`),
+	if err != nil {
-			[]byte(`{"error":{"message":"minimumCreditAmountForUsage requirement not met"}}`),
+		t.Fatalf("parseRetryDelay() error = %v", err)
-		} {
-			if !shouldMarkAntigravityCreditsExhausted(http.StatusForbidden, body, nil) {
-				t.Fatalf("shouldMarkAntigravityCreditsExhausted(%s) = false, want true", string(body))
 	}
+	if retryAfter == nil {
+		t.Fatal("parseRetryDelay() returned nil")
 	}
-	})
+	want := time.Hour + 43*time.Minute + 56*time.Second
-
+	if *retryAfter != want {
-	t.Run("transient 429 resource exhausted is not marked", func(t *testing.T) {
+		t.Fatalf("parseRetryDelay() = %v, want %v", *retryAfter, want)
-		body := []byte(`{"error":{"code":429,"message":"Resource has been exhausted (e.g. check quota).","status":"RESOURCE_EXHAUSTED"}}`)
-		if shouldMarkAntigravityCreditsExhausted(http.StatusTooManyRequests, body, nil) {
-			t.Fatalf("shouldMarkAntigravityCreditsExhausted(%s) = true, want false", string(body))
-		}
-	})
-
-	t.Run("resource exhausted with quota metadata is still marked", func(t *testing.T) {
-		body := []byte(`{"error":{"code":429,"message":"Resource has been exhausted","status":"RESOURCE_EXHAUSTED","details":[{"@type":"type.googleapis.com/google.rpc.ErrorInfo","metadata":{"quotaResetDelay":"1h","model":"claude-sonnet-4-6"}}]}}`)
-		if !shouldMarkAntigravityCreditsExhausted(http.StatusTooManyRequests, body, nil) {
-			t.Fatalf("shouldMarkAntigravityCreditsExhausted(%s) = false, want true", string(body))
-		}
-	})
-
-	if shouldMarkAntigravityCreditsExhausted(http.StatusServiceUnavailable, []byte(`{"error":{"message":"credits exhausted"}}`), nil) {
-		t.Fatal("shouldMarkAntigravityCreditsExhausted() = true for 5xx, want false")
 	}
 }
 
@@ -147,7 +192,7 @@ func TestAntigravityExecute_RetriesTransient429ResourceExhausted(t *testing.T) {
 	}
 
 	resp, err := exec.Execute(context.Background(), auth, cliproxyexecutor.Request{
-		Model:   "gemini-2.5-flash",
+		Model:   "claude-sonnet-4-6",
 		Payload: []byte(`{"request":{"contents":[{"role":"user","parts":[{"text":"hi"}]}]}}`),
 	}, cliproxyexecutor.Options{
 		SourceFormat: sdktranslator.FormatAntigravity,
@@ -163,32 +208,18 @@ func TestAntigravityExecute_RetriesTransient429ResourceExhausted(t *testing.T) {
 	}
 }
 
-func TestAntigravityExecute_RetriesQuotaExhaustedWithCredits(t *testing.T) {
+func TestAntigravityExecute_CreditsInjectedWhenConductorRequests(t *testing.T) {
 	resetAntigravityCreditsRetryState()
 	t.Cleanup(resetAntigravityCreditsRetryState)
 
-	var (
+	var requestBodies []string
-		mu            sync.Mutex
-		requestBodies []string
-	)
-
 	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
 		body, _ := io.ReadAll(r.Body)
 		_ = r.Body.Close()
-
-		mu.Lock()
 		requestBodies = append(requestBodies, string(body))
-		reqNum := len(requestBodies)
-		mu.Unlock()
-
-		if reqNum == 1 {
-			w.WriteHeader(http.StatusTooManyRequests)
-			_, _ = w.Write([]byte(`{"error":{"status":"RESOURCE_EXHAUSTED","message":"QUOTA_EXHAUSTED"}}`))
-			return
-		}
 
 		if !strings.Contains(string(body), `"enabledCreditTypes":["GOOGLE_ONE_AI"]`) {
-			t.Fatalf("second request body missing enabledCreditTypes: %s", string(body))
+			t.Fatalf("request body missing enabledCreditTypes: %s", string(body))
 		}
 		w.Header().Set("Content-Type", "application/json")
 		_, _ = w.Write([]byte(`{"response":{"candidates":[{"content":{"role":"model","parts":[{"text":"ok"}]}}],"usageMetadata":{"promptTokenCount":1,"candidatesTokenCount":1,"totalTokenCount":2}}}`))
@@ -199,7 +230,7 @@ func TestAntigravityExecute_RetriesQuotaExhaustedWithCredits(t *testing.T) {
 		QuotaExceeded: config.QuotaExceeded{AntigravityCredits: true},
 	})
 	auth := &cliproxyauth.Auth{
-		ID: "auth-credits-ok",
+		ID: "auth-credits-conductor",
 		Attributes: map[string]string{
 			"base_url": server.URL,
 		},
@@ -210,8 +241,11 @@ func TestAntigravityExecute_RetriesQuotaExhaustedWithCredits(t *testing.T) {
 		},
 	}
 
-	resp, err := exec.Execute(context.Background(), auth, cliproxyexecutor.Request{
+	// Simulate conductor setting credits requested flag in context
-		Model:   "gemini-2.5-flash",
+	ctx := cliproxyauth.WithAntigravityCredits(context.Background())
+
+	resp, err := exec.Execute(ctx, auth, cliproxyexecutor.Request{
+		Model:   "claude-sonnet-4-6",
 		Payload: []byte(`{"request":{"contents":[{"role":"user","parts":[{"text":"hi"}]}]}}`),
 	}, cliproxyexecutor.Options{
 		SourceFormat: sdktranslator.FormatAntigravity,
@@ -222,227 +256,12 @@ func TestAntigravityExecute_RetriesQuotaExhaustedWithCredits(t *testing.T) {
 	if len(resp.Payload) == 0 {
 		t.Fatal("Execute() returned empty payload")
 	}
-
+	if len(requestBodies) != 1 {
-	mu.Lock()
+		t.Fatalf("request count = %d, want 1", len(requestBodies))
-	defer mu.Unlock()
-	if len(requestBodies) != 2 {
-		t.Fatalf("request count = %d, want 2", len(requestBodies))
 	}
 }
 
-func TestAntigravityExecute_SkipsCreditsRetryWhenAlreadyExhausted(t *testing.T) {
+func TestAntigravityExecute_NoCreditsWithoutConductorFlag(t *testing.T) {
-	resetAntigravityCreditsRetryState()
-	t.Cleanup(resetAntigravityCreditsRetryState)
-
-	var requestCount int
-	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
-		requestCount++
-		w.WriteHeader(http.StatusTooManyRequests)
-		_, _ = w.Write([]byte(`{"error":{"status":"RESOURCE_EXHAUSTED","message":"QUOTA_EXHAUSTED"}}`))
-	}))
-	defer server.Close()
-
-	exec := NewAntigravityExecutor(&config.Config{
-		QuotaExceeded: config.QuotaExceeded{AntigravityCredits: true},
-	})
-	auth := &cliproxyauth.Auth{
-		ID: "auth-credits-exhausted",
-		Attributes: map[string]string{
-			"base_url": server.URL,
-		},
-		Metadata: map[string]any{
-			"access_token": "token",
-			"project_id":   "project-1",
-			"expired":      time.Now().Add(1 * time.Hour).Format(time.RFC3339),
-		},
-	}
-	recordAntigravityCreditsFailure(auth, time.Now())
-
-	_, err := exec.Execute(context.Background(), auth, cliproxyexecutor.Request{
-		Model:   "gemini-2.5-flash",
-		Payload: []byte(`{"request":{"contents":[{"role":"user","parts":[{"text":"hi"}]}]}}`),
-	}, cliproxyexecutor.Options{
-		SourceFormat: sdktranslator.FormatAntigravity,
-	})
-	if err == nil {
-		t.Fatal("Execute() error = nil, want 429")
-	}
-	sErr, ok := err.(statusErr)
-	if !ok {
-		t.Fatalf("Execute() error type = %T, want statusErr", err)
-	}
-	if got := sErr.StatusCode(); got != http.StatusTooManyRequests {
-		t.Fatalf("Execute() status code = %d, want %d", got, http.StatusTooManyRequests)
-	}
-	if requestCount != 1 {
-		t.Fatalf("request count = %d, want 1", requestCount)
-	}
-}
-
-func TestAntigravityExecute_PrefersCreditsAfterSuccessfulFallback(t *testing.T) {
-	resetAntigravityCreditsRetryState()
-	t.Cleanup(resetAntigravityCreditsRetryState)
-
-	var (
-		mu            sync.Mutex
-		requestBodies []string
-	)
-
-	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
-		body, _ := io.ReadAll(r.Body)
-		_ = r.Body.Close()
-
-		mu.Lock()
-		requestBodies = append(requestBodies, string(body))
-		reqNum := len(requestBodies)
-		mu.Unlock()
-
-		switch reqNum {
-		case 1:
-			w.WriteHeader(http.StatusTooManyRequests)
-			_, _ = w.Write([]byte(`{"error":{"status":"RESOURCE_EXHAUSTED","details":[{"@type":"type.googleapis.com/google.rpc.ErrorInfo","reason":"QUOTA_EXHAUSTED"},{"@type":"type.googleapis.com/google.rpc.RetryInfo","retryDelay":"10s"}]}}`))
-		case 2, 3:
-			if !strings.Contains(string(body), `"enabledCreditTypes":["GOOGLE_ONE_AI"]`) {
-				t.Fatalf("request %d body missing enabledCreditTypes: %s", reqNum, string(body))
-			}
-			w.Header().Set("Content-Type", "application/json")
-			_, _ = w.Write([]byte(`{"response":{"candidates":[{"content":{"role":"model","parts":[{"text":"OK"}]}}],"usageMetadata":{"promptTokenCount":1,"candidatesTokenCount":1,"totalTokenCount":2}}}`))
-		default:
-			t.Fatalf("unexpected request count %d", reqNum)
-		}
-	}))
-	defer server.Close()
-
-	exec := NewAntigravityExecutor(&config.Config{
-		QuotaExceeded: config.QuotaExceeded{AntigravityCredits: true},
-	})
-	auth := &cliproxyauth.Auth{
-		ID: "auth-prefer-credits",
-		Attributes: map[string]string{
-			"base_url": server.URL,
-		},
-		Metadata: map[string]any{
-			"access_token": "token",
-			"project_id":   "project-1",
-			"expired":      time.Now().Add(1 * time.Hour).Format(time.RFC3339),
-		},
-	}
-
-	request := cliproxyexecutor.Request{
-		Model:   "gemini-2.5-flash",
-		Payload: []byte(`{"request":{"contents":[{"role":"user","parts":[{"text":"hi"}]}]}}`),
-	}
-	opts := cliproxyexecutor.Options{SourceFormat: sdktranslator.FormatAntigravity}
-
-	if _, err := exec.Execute(context.Background(), auth, request, opts); err != nil {
-		t.Fatalf("first Execute() error = %v", err)
-	}
-	if _, err := exec.Execute(context.Background(), auth, request, opts); err != nil {
-		t.Fatalf("second Execute() error = %v", err)
-	}
-
-	mu.Lock()
-	defer mu.Unlock()
-	if len(requestBodies) != 3 {
-		t.Fatalf("request count = %d, want 3", len(requestBodies))
-	}
-	if strings.Contains(requestBodies[0], `"enabledCreditTypes":["GOOGLE_ONE_AI"]`) {
-		t.Fatalf("first request unexpectedly used credits: %s", requestBodies[0])
-	}
-	if !strings.Contains(requestBodies[1], `"enabledCreditTypes":["GOOGLE_ONE_AI"]`) {
-		t.Fatalf("fallback request missing credits: %s", requestBodies[1])
-	}
-	if !strings.Contains(requestBodies[2], `"enabledCreditTypes":["GOOGLE_ONE_AI"]`) {
-		t.Fatalf("preferred request missing credits: %s", requestBodies[2])
-	}
-}
-
-func TestAntigravityExecute_PreservesBaseURLFallbackAfterCreditsRetryFailure(t *testing.T) {
-	resetAntigravityCreditsRetryState()
-	t.Cleanup(resetAntigravityCreditsRetryState)
-
-	var (
-		mu          sync.Mutex
-		firstCount  int
-		secondCount int
-	)
-
-	firstServer := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
-		body, _ := io.ReadAll(r.Body)
-		_ = r.Body.Close()
-
-		mu.Lock()
-		firstCount++
-		reqNum := firstCount
-		mu.Unlock()
-
-		switch reqNum {
-		case 1:
-			w.WriteHeader(http.StatusTooManyRequests)
-			_, _ = w.Write([]byte(`{"error":{"status":"RESOURCE_EXHAUSTED","details":[{"@type":"type.googleapis.com/google.rpc.ErrorInfo","reason":"QUOTA_EXHAUSTED"}]}}`))
-		case 2:
-			if !strings.Contains(string(body), `"enabledCreditTypes":["GOOGLE_ONE_AI"]`) {
-				t.Fatalf("credits retry missing enabledCreditTypes: %s", string(body))
-			}
-			w.WriteHeader(http.StatusForbidden)
-			_, _ = w.Write([]byte(`{"error":{"message":"permission denied"}}`))
-		default:
-			t.Fatalf("unexpected first server request count %d", reqNum)
-		}
-	}))
-	defer firstServer.Close()
-
-	secondServer := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
-		mu.Lock()
-		secondCount++
-		mu.Unlock()
-		w.Header().Set("Content-Type", "application/json")
-		_, _ = w.Write([]byte(`{"response":{"candidates":[{"content":{"role":"model","parts":[{"text":"ok"}]}}],"usageMetadata":{"promptTokenCount":1,"candidatesTokenCount":1,"totalTokenCount":2}}}`))
-	}))
-	defer secondServer.Close()
-
-	exec := NewAntigravityExecutor(&config.Config{
-		QuotaExceeded: config.QuotaExceeded{AntigravityCredits: true},
-	})
-	auth := &cliproxyauth.Auth{
-		ID: "auth-baseurl-fallback",
-		Attributes: map[string]string{
-			"base_url": firstServer.URL,
-		},
-		Metadata: map[string]any{
-			"access_token": "token",
-			"project_id":   "project-1",
-			"expired":      time.Now().Add(1 * time.Hour).Format(time.RFC3339),
-		},
-	}
-
-	originalOrder := antigravityBaseURLFallbackOrder
-	defer func() { antigravityBaseURLFallbackOrder = originalOrder }()
-	antigravityBaseURLFallbackOrder = func(auth *cliproxyauth.Auth) []string {
-		return []string{firstServer.URL, secondServer.URL}
-	}
-
-	resp, err := exec.Execute(context.Background(), auth, cliproxyexecutor.Request{
-		Model:   "gemini-2.5-flash",
-		Payload: []byte(`{"request":{"contents":[{"role":"user","parts":[{"text":"hi"}]}]}}`),
-	}, cliproxyexecutor.Options{
-		SourceFormat: sdktranslator.FormatAntigravity,
-	})
-	if err != nil {
-		t.Fatalf("Execute() error = %v", err)
-	}
-	if len(resp.Payload) == 0 {
-		t.Fatal("Execute() returned empty payload")
-	}
-	if firstCount != 2 {
-		t.Fatalf("first server request count = %d, want 2", firstCount)
-	}
-	if secondCount != 1 {
-		t.Fatalf("second server request count = %d, want 1", secondCount)
-	}
-}
-
-func TestAntigravityExecute_DoesNotDirectInjectCreditsWhenFlagDisabled(t *testing.T) {
 	resetAntigravityCreditsRetryState()
 	t.Cleanup(resetAntigravityCreditsRetryState)
 
@@ -457,10 +276,10 @@ func TestAntigravityExecute_DoesNotDirectInjectCreditsWhenFlagDisabled(t *testin
 	defer server.Close()
 
 	exec := NewAntigravityExecutor(&config.Config{
-		QuotaExceeded: config.QuotaExceeded{AntigravityCredits: false},
+		QuotaExceeded: config.QuotaExceeded{AntigravityCredits: true},
 	})
 	auth := &cliproxyauth.Auth{
-		ID: "auth-flag-disabled",
+		ID: "auth-no-conductor-flag",
 		Attributes: map[string]string{
 			"base_url": server.URL,
 		},
@@ -470,10 +289,10 @@ func TestAntigravityExecute_DoesNotDirectInjectCreditsWhenFlagDisabled(t *testin
 			"expired":      time.Now().Add(1 * time.Hour).Format(time.RFC3339),
 		},
 	}
-	markAntigravityPreferCredits(auth, "gemini-2.5-flash", time.Now(), nil)
 
+	// No conductor credits flag set in context
 	_, err := exec.Execute(context.Background(), auth, cliproxyexecutor.Request{
-		Model:   "gemini-2.5-flash",
+		Model:   "claude-sonnet-4-6",
 		Payload: []byte(`{"request":{"contents":[{"role":"user","parts":[{"text":"hi"}]}]}}`),
 	}, cliproxyexecutor.Options{
 		SourceFormat: sdktranslator.FormatAntigravity,
@@ -484,7 +303,156 @@ func TestAntigravityExecute_DoesNotDirectInjectCreditsWhenFlagDisabled(t *testin
 	if len(requestBodies) != 1 {
 		t.Fatalf("request count = %d, want 1", len(requestBodies))
 	}
-	if strings.Contains(requestBodies[0], `"enabledCreditTypes":["GOOGLE_ONE_AI"]`) {
+	// Should NOT contain credits since conductor didn't request them
-		t.Fatalf("request unexpectedly used enabledCreditTypes with flag disabled: %s", requestBodies[0])
+	if strings.Contains(requestBodies[0], `"enabledCreditTypes"`) {
+		t.Fatalf("request should not contain enabledCreditTypes without conductor flag: %s", requestBodies[0])
+	}
+}
+
+func TestAntigravityAuthHasCredits(t *testing.T) {
+	t.Run("sufficient balance", func(t *testing.T) {
+		resetAntigravityCreditsRetryState()
+		auth := &cliproxyauth.Auth{ID: "test-sufficient"}
+		antigravityCreditsBalanceByAuth.Store("test-sufficient", antigravityCreditsBalance{
+			CreditAmount:    25000,
+			MinCreditAmount: 50,
+			Known:           true,
+		})
+		if !antigravityAuthHasCredits(auth) {
+			t.Fatal("antigravityAuthHasCredits() = false, want true")
+		}
+	})
+
+	t.Run("insufficient balance", func(t *testing.T) {
+		resetAntigravityCreditsRetryState()
+		auth := &cliproxyauth.Auth{ID: "test-insufficient"}
+		antigravityCreditsBalanceByAuth.Store("test-insufficient", antigravityCreditsBalance{
+			CreditAmount:    30,
+			MinCreditAmount: 50,
+			Known:           true,
+		})
+		if antigravityAuthHasCredits(auth) {
+			t.Fatal("antigravityAuthHasCredits() = true, want false")
+		}
+	})
+
+	t.Run("no balance stored returns true (optimistic)", func(t *testing.T) {
+		resetAntigravityCreditsRetryState()
+		auth := &cliproxyauth.Auth{ID: "test-no-balance"}
+		if !antigravityAuthHasCredits(auth) {
+			t.Fatal("antigravityAuthHasCredits() = false with no balance stored, want true (optimistic default)")
+		}
+	})
+
+	t.Run("nil auth returns false", func(t *testing.T) {
+		if antigravityAuthHasCredits(nil) {
+			t.Fatal("antigravityAuthHasCredits(nil) = true, want false")
+		}
+	})
+
+	t.Run("empty ID returns false", func(t *testing.T) {
+		auth := &cliproxyauth.Auth{}
+		if antigravityAuthHasCredits(auth) {
+			t.Fatal("antigravityAuthHasCredits(empty ID) = true, want false")
+		}
+	})
+
+	t.Run("unknown balance returns false", func(t *testing.T) {
+		resetAntigravityCreditsRetryState()
+		auth := &cliproxyauth.Auth{ID: "test-unknown"}
+		antigravityCreditsBalanceByAuth.Store("test-unknown", antigravityCreditsBalance{
+			Known: false,
+		})
+		if antigravityAuthHasCredits(auth) {
+			t.Fatal("antigravityAuthHasCredits() = true for unknown balance, want false")
+		}
+	})
+}
+
+type roundTripperFunc func(*http.Request) (*http.Response, error)
+
+func (f roundTripperFunc) RoundTrip(req *http.Request) (*http.Response, error) {
+	return f(req)
+}
+
+func TestEnsureAccessToken_WarmTokenLoadsCreditsHint(t *testing.T) {
+	resetAntigravityCreditsRetryState()
+	t.Cleanup(resetAntigravityCreditsRetryState)
+
+	exec := NewAntigravityExecutor(&config.Config{
+		QuotaExceeded: config.QuotaExceeded{AntigravityCredits: true},
+	})
+	auth := &cliproxyauth.Auth{
+		ID: "auth-warm-token-credits",
+		Metadata: map[string]any{
+			"access_token": "token",
+			"expired":      time.Now().Add(1 * time.Hour).Format(time.RFC3339),
+		},
+	}
+	ctx := context.WithValue(context.Background(), "cliproxy.roundtripper", roundTripperFunc(func(req *http.Request) (*http.Response, error) {
+		if req.URL.String() != "https://cloudcode-pa.googleapis.com/v1internal:loadCodeAssist" {
+			t.Fatalf("unexpected request url %s", req.URL.String())
+		}
+		return &http.Response{
+			StatusCode: http.StatusOK,
+			Header:     make(http.Header),
+			Body:       io.NopCloser(strings.NewReader(`{"paidTier":{"id":"tier-1","availableCredits":[{"creditType":"GOOGLE_ONE_AI","creditAmount":"25000","minimumCreditAmountForUsage":"50"}]}}`)),
+		}, nil
+	}))
+
+	token, updatedAuth, err := exec.ensureAccessToken(ctx, auth)
+	if err != nil {
+		t.Fatalf("ensureAccessToken() error = %v", err)
+	}
+	if token != "token" {
+		t.Fatalf("ensureAccessToken() token = %q, want %q", token, "token")
+	}
+	if updatedAuth != nil {
+		t.Fatalf("ensureAccessToken() updatedAuth = %v, want nil", updatedAuth)
+	}
+	deadline := time.Now().Add(2 * time.Second)
+	for time.Now().Before(deadline) && !cliproxyauth.HasKnownAntigravityCreditsHint(auth.ID) {
+		time.Sleep(10 * time.Millisecond)
+	}
+	if !cliproxyauth.HasKnownAntigravityCreditsHint(auth.ID) {
+		t.Fatal("expected credits hint to be populated for warm token auth")
+	}
+	hint, ok := cliproxyauth.GetAntigravityCreditsHint(auth.ID)
+	if !ok {
+		t.Fatal("expected credits hint lookup to succeed")
+	}
+	if !hint.Available {
+		t.Fatalf("hint.Available = %v, want true", hint.Available)
+	}
+	if hint.CreditAmount != 25000 || hint.MinCreditAmount != 50 {
+		t.Fatalf("hint amounts = (%v, %v), want (25000, 50)", hint.CreditAmount, hint.MinCreditAmount)
+	}
+}
+
+func TestParseMetaFloat(t *testing.T) {
+	tests := []struct {
+		name    string
+		value   any
+		wantVal float64
+		wantOK  bool
+	}{
+		{"string", "25000", 25000, true},
+		{"float64", float64(100), 100, true},
+		{"int", int(50), 50, true},
+		{"int64", int64(75), 75, true},
+		{"empty string", "", 0, false},
+		{"invalid string", "abc", 0, false},
+	}
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			meta := map[string]any{"key": tt.value}
+			got, ok := parseMetaFloat(meta, "key")
+			if ok != tt.wantOK {
+				t.Fatalf("parseMetaFloat() ok = %v, want %v", ok, tt.wantOK)
+			}
+			if ok && got != tt.wantVal {
+				t.Fatalf("parseMetaFloat() = %f, want %f", got, tt.wantVal)
+			}
+		})
 	}
 }

internal/runtime/executor/gemini_cli_executor.go

@@ -898,7 +898,14 @@ func parseRetryDelay(errorBody []byte) (*time.Duration, error) {
 		if matches := re.FindStringSubmatch(message); len(matches) > 1 {
 			seconds, err := strconv.Atoi(matches[1])
 			if err == nil {
-				return new(time.Duration(seconds) * time.Second), nil
+				duration := time.Duration(seconds) * time.Second
+				return &duration, nil
+			}
+		}
+		reHuman := regexp.MustCompile(`after\s+((?:\d+h)?(?:\d+m)?(?:\d+s)?)\.?`)
+		if matches := reHuman.FindStringSubmatch(strings.ToLower(message)); len(matches) > 1 {
+			if duration, err := time.ParseDuration(matches[1]); err == nil && duration > 0 {
+				return &duration, nil
 			}
 		}
 	}

internal/runtime/executor/helps/logging_helpers.go

@@ -24,6 +24,7 @@ const (
 	apiRequestKey           = "API_REQUEST"
 	apiResponseKey          = "API_RESPONSE"
 	apiWebsocketTimelineKey = "API_WEBSOCKET_TIMELINE"
+	creditsUsedKey          = "__antigravity_credits_used__"
 )
 
 // UpstreamRequestLog captures the outbound upstream request details for logging.
@@ -568,3 +569,24 @@ func LogWithRequestID(ctx context.Context) *log.Entry {
 	}
 	return log.WithField("request_id", requestID)
 }
+
+// MarkCreditsUsed flags the request as having used AI credits for billing.
+func MarkCreditsUsed(ctx context.Context) {
+	ginCtx := ginContextFrom(ctx)
+	if ginCtx != nil {
+		ginCtx.Set(creditsUsedKey, true)
+	}
+}
+
+// CreditsUsed returns true if the request used AI credits.
+func CreditsUsed(ctx context.Context) bool {
+	ginCtx := ginContextFrom(ctx)
+	if ginCtx != nil {
+		if val, exists := ginCtx.Get(creditsUsedKey); exists {
+			if b, ok := val.(bool); ok {
+				return b
+			}
+		}
+	}
+	return false
+}

sdk/cliproxy/auth/antigravity_credits.go

@@ -0,0 +1,90 @@
+package auth
+
+import (
+	"context"
+	"strings"
+	"sync"
+	"time"
+)
+
+type antigravityUseCreditsContextKey struct{}
+
+// WithAntigravityCredits returns a child context that signals the executor to
+// inject enabledCreditTypes into the request payload.
+func WithAntigravityCredits(ctx context.Context) context.Context {
+	return context.WithValue(ctx, antigravityUseCreditsContextKey{}, true)
+}
+
+// AntigravityCreditsRequested reports whether the context carries the credits flag.
+func AntigravityCreditsRequested(ctx context.Context) bool {
+	if ctx == nil {
+		return false
+	}
+	v, _ := ctx.Value(antigravityUseCreditsContextKey{}).(bool)
+	return v
+}
+
+// AntigravityCreditsHint stores the latest known AI credits state for one auth.
+type AntigravityCreditsHint struct {
+	Known           bool
+	Available       bool
+	CreditAmount    float64
+	MinCreditAmount float64
+	PaidTierID      string
+	UpdatedAt       time.Time
+}
+
+var antigravityCreditsHintByAuth sync.Map
+
+// SetAntigravityCreditsHint updates the latest known AI credits state for an auth.
+func SetAntigravityCreditsHint(authID string, hint AntigravityCreditsHint) {
+	authID = strings.TrimSpace(authID)
+	if authID == "" {
+		return
+	}
+	if hint.UpdatedAt.IsZero() {
+		hint.UpdatedAt = time.Now()
+	}
+	antigravityCreditsHintByAuth.Store(authID, hint)
+}
+
+// GetAntigravityCreditsHint returns the latest known AI credits state for an auth.
+func GetAntigravityCreditsHint(authID string) (AntigravityCreditsHint, bool) {
+	authID = strings.TrimSpace(authID)
+	if authID == "" {
+		return AntigravityCreditsHint{}, false
+	}
+	value, ok := antigravityCreditsHintByAuth.Load(authID)
+	if !ok {
+		return AntigravityCreditsHint{}, false
+	}
+	hint, ok := value.(AntigravityCreditsHint)
+	if !ok {
+		antigravityCreditsHintByAuth.Delete(authID)
+		return AntigravityCreditsHint{}, false
+	}
+	return hint, true
+}
+
+// HasKnownAntigravityCreditsHint reports whether credits state has been discovered for an auth.
+func HasKnownAntigravityCreditsHint(authID string) bool {
+	hint, ok := GetAntigravityCreditsHint(authID)
+	return ok && hint.Known
+}
+
+func antigravityCreditsAvailableForModel(auth *Auth, model string) bool {
+	if auth == nil {
+		return false
+	}
+	if !strings.EqualFold(strings.TrimSpace(auth.Provider), "antigravity") {
+		return false
+	}
+	if !strings.Contains(strings.ToLower(strings.TrimSpace(model)), "claude") {
+		return false
+	}
+	hint, ok := GetAntigravityCreditsHint(auth.ID)
+	if !ok || !hint.Known {
+		return false
+	}
+	return hint.Available
+}

sdk/cliproxy/auth/antigravity_credits_test.go

@@ -0,0 +1,62 @@
+package auth
+
+import (
+	"testing"
+	"time"
+)
+
+func TestIsAuthBlockedForModel_ClaudeWithCreditsStillBlockedDuringCooldown(t *testing.T) {
+	auth := &Auth{
+		ID:       "ag-1",
+		Provider: "antigravity",
+		ModelStates: map[string]*ModelState{
+			"claude-sonnet-4-6": {
+				Unavailable:    true,
+				NextRetryAfter: time.Now().Add(10 * time.Minute),
+				Quota: QuotaState{
+					Exceeded:      true,
+					NextRecoverAt: time.Now().Add(10 * time.Minute),
+				},
+			},
+		},
+	}
+
+	SetAntigravityCreditsHint(auth.ID, AntigravityCreditsHint{
+		Known:     true,
+		Available: true,
+		UpdatedAt: time.Now(),
+	})
+
+	blocked, reason, _ := isAuthBlockedForModel(auth, "claude-sonnet-4-6", time.Now())
+	if !blocked || reason != blockReasonCooldown {
+		t.Fatalf("expected auth to be blocked during cooldown even with credits, got blocked=%v reason=%v", blocked, reason)
+	}
+}
+
+func TestIsAuthBlockedForModel_KeepsGeminiBlockedWithoutCreditsBypass(t *testing.T) {
+	auth := &Auth{
+		ID:       "ag-2",
+		Provider: "antigravity",
+		ModelStates: map[string]*ModelState{
+			"gemini-3-flash": {
+				Unavailable:    true,
+				NextRetryAfter: time.Now().Add(10 * time.Minute),
+				Quota: QuotaState{
+					Exceeded:      true,
+					NextRecoverAt: time.Now().Add(10 * time.Minute),
+				},
+			},
+		},
+	}
+
+	SetAntigravityCreditsHint(auth.ID, AntigravityCreditsHint{
+		Known:     true,
+		Available: true,
+		UpdatedAt: time.Now(),
+	})
+
+	blocked, reason, _ := isAuthBlockedForModel(auth, "gemini-3-flash", time.Now())
+	if !blocked || reason != blockReasonCooldown {
+		t.Fatalf("expected gemini auth to remain blocked, got blocked=%v reason=%v", blocked, reason)
+	}
+}

sdk/cliproxy/auth/conductor.go

@@ -1202,12 +1202,16 @@ func (m *Manager) Execute(ctx context.Context, providers []string, req cliproxye
 		}
 	}
 	if lastErr != nil {
+		if shouldAttemptAntigravityCreditsFallback(m, lastErr, normalized) {
+			if resp, ok := m.tryAntigravityCreditsExecute(ctx, req, opts); ok {
+				return resp, nil
+			}
+		}
 		return cliproxyexecutor.Response{}, lastErr
 	}
 	return cliproxyexecutor.Response{}, &Error{Code: "auth_not_found", Message: "no auth available"}
 }
 
-// ExecuteCount performs a non-streaming execution using the configured selector and executor.
 // It supports multiple providers for the same model and round-robins the starting provider per model.
 func (m *Manager) ExecuteCount(ctx context.Context, providers []string, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (cliproxyexecutor.Response, error) {
 	normalized := m.normalizeProviders(providers)
@@ -1264,6 +1268,11 @@ func (m *Manager) ExecuteStream(ctx context.Context, providers []string, req cli
 		}
 	}
 	if lastErr != nil {
+		if shouldAttemptAntigravityCreditsFallback(m, lastErr, normalized) {
+			if result, ok := m.tryAntigravityCreditsExecuteStream(ctx, req, opts); ok {
+				return result, nil
+			}
+		}
 		return nil, lastErr
 	}
 	return nil, &Error{Code: "auth_not_found", Message: "no auth available"}
@@ -2319,7 +2328,8 @@ func retryAfterFromError(err error) *time.Duration {
 	if retryAfter == nil {
 		return nil
 	}
-	return new(*retryAfter)
+	value := *retryAfter
+	return &value
 }
 
 func statusCodeFromResult(err *Error) int {
@@ -2409,11 +2419,18 @@ func isRequestInvalidError(err error) bool {
 	status := statusCodeFromError(err)
 	switch status {
 	case http.StatusBadRequest:
-		return strings.Contains(err.Error(), "invalid_request_error")
+		msg := err.Error()
+		return strings.Contains(msg, "invalid_request_error") ||
+			strings.Contains(msg, "INVALID_ARGUMENT") ||
+			strings.Contains(msg, "FAILED_PRECONDITION")
 	case http.StatusNotFound:
 		return isRequestScopedNotFoundMessage(err.Error())
 	case http.StatusUnprocessableEntity:
 		return true
+	case http.StatusInternalServerError:
+		msg := err.Error()
+		return strings.Contains(msg, "\"status\":\"UNKNOWN\"") ||
+			strings.Contains(msg, "\"status\": \"UNKNOWN\"")
 	default:
 		return false
 	}
@@ -2886,6 +2903,175 @@ func (m *Manager) pickNextMixed(ctx context.Context, providers []string, model s
 	return authCopy, executor, providerKey, nil
 }
 
+func (m *Manager) findAllAntigravityCreditsCandidateAuths(routeModel string, opts cliproxyexecutor.Options) []creditsCandidateEntry {
+	if m == nil {
+		return nil
+	}
+	pinnedAuthID := pinnedAuthIDFromMetadata(opts.Metadata)
+	m.mu.RLock()
+	defer m.mu.RUnlock()
+	var known []creditsCandidateEntry
+	var unknown []creditsCandidateEntry
+	for _, auth := range m.auths {
+		if auth == nil || auth.Disabled || auth.Status == StatusDisabled {
+			continue
+		}
+		if pinnedAuthID != "" && auth.ID != pinnedAuthID {
+			continue
+		}
+		if !strings.EqualFold(strings.TrimSpace(auth.Provider), "antigravity") {
+			continue
+		}
+		if !strings.Contains(strings.ToLower(strings.TrimSpace(routeModel)), "claude") {
+			continue
+		}
+		providerKey := strings.TrimSpace(strings.ToLower(auth.Provider))
+		executor, ok := m.executors[providerKey]
+		if !ok {
+			continue
+		}
+
+		hint, okHint := GetAntigravityCreditsHint(auth.ID)
+		if okHint && hint.Known {
+			if !hint.Available {
+				continue
+			}
+			known = append(known, creditsCandidateEntry{
+				auth:     auth.Clone(),
+				executor: executor,
+				provider: providerKey,
+			})
+			continue
+		}
+		unknown = append(unknown, creditsCandidateEntry{
+			auth:     auth.Clone(),
+			executor: executor,
+			provider: providerKey,
+		})
+	}
+	sort.Slice(known, func(i, j int) bool {
+		return known[i].auth.ID < known[j].auth.ID
+	})
+	sort.Slice(unknown, func(i, j int) bool {
+		return unknown[i].auth.ID < unknown[j].auth.ID
+	})
+	return append(known, unknown...)
+}
+
+type creditsCandidateEntry struct {
+	auth     *Auth
+	executor ProviderExecutor
+	provider string
+}
+
+func shouldAttemptAntigravityCreditsFallback(m *Manager, lastErr error, providers []string) bool {
+	if m == nil || lastErr == nil {
+		return false
+	}
+	if len(providers) > 0 {
+		hasAntigravity := false
+		for _, p := range providers {
+			if strings.EqualFold(strings.TrimSpace(p), "antigravity") {
+				hasAntigravity = true
+				break
+			}
+		}
+		if !hasAntigravity {
+			return false
+		}
+	}
+	cfg, _ := m.runtimeConfig.Load().(*internalconfig.Config)
+	if cfg == nil || !cfg.QuotaExceeded.AntigravityCredits {
+		return false
+	}
+	status := statusCodeFromError(lastErr)
+	switch status {
+	case http.StatusTooManyRequests, http.StatusServiceUnavailable:
+		return true
+	case 0:
+		var authErr *Error
+		if errors.As(lastErr, &authErr) && authErr != nil {
+			return authErr.Code == "auth_not_found" || authErr.Code == "auth_unavailable" || authErr.Code == "model_cooldown"
+		}
+		var cooldownErr *modelCooldownError
+		if errors.As(lastErr, &cooldownErr) {
+			return true
+		}
+		return false
+	default:
+		return false
+	}
+}
+
+func (m *Manager) tryAntigravityCreditsExecute(ctx context.Context, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (cliproxyexecutor.Response, bool) {
+	routeModel := req.Model
+	candidates := m.findAllAntigravityCreditsCandidateAuths(routeModel, opts)
+	for _, c := range candidates {
+		if ctx.Err() != nil {
+			return cliproxyexecutor.Response{}, false
+		}
+		creditsCtx := WithAntigravityCredits(ctx)
+		if rt := m.roundTripperFor(c.auth); rt != nil {
+			creditsCtx = context.WithValue(creditsCtx, roundTripperContextKey{}, rt)
+			creditsCtx = context.WithValue(creditsCtx, "cliproxy.roundtripper", rt)
+		}
+		creditsOpts := ensureRequestedModelMetadata(opts, routeModel)
+		publishSelectedAuthMetadata(creditsOpts.Metadata, c.auth.ID)
+		models := m.executionModelCandidates(c.auth, routeModel)
+		if len(models) == 0 {
+			continue
+		}
+		for _, upstreamModel := range models {
+			resultModel := m.stateModelForExecution(c.auth, routeModel, upstreamModel, len(models) > 1)
+			execReq := req
+			execReq.Model = upstreamModel
+			resp, errExec := c.executor.Execute(creditsCtx, c.auth, execReq, creditsOpts)
+			result := Result{AuthID: c.auth.ID, Provider: c.provider, Model: resultModel, Success: errExec == nil}
+			if errExec != nil {
+				result.Error = &Error{Message: errExec.Error()}
+				if se, ok := errors.AsType[cliproxyexecutor.StatusError](errExec); ok && se != nil {
+					result.Error.HTTPStatus = se.StatusCode()
+				}
+				if ra := retryAfterFromError(errExec); ra != nil {
+					result.RetryAfter = ra
+				}
+				m.MarkResult(creditsCtx, result)
+				continue
+			}
+			m.MarkResult(creditsCtx, result)
+			return resp, true
+		}
+	}
+	return cliproxyexecutor.Response{}, false
+}
+
+func (m *Manager) tryAntigravityCreditsExecuteStream(ctx context.Context, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (*cliproxyexecutor.StreamResult, bool) {
+	routeModel := req.Model
+	candidates := m.findAllAntigravityCreditsCandidateAuths(routeModel, opts)
+	for _, c := range candidates {
+		if ctx.Err() != nil {
+			return nil, false
+		}
+		creditsCtx := WithAntigravityCredits(ctx)
+		if rt := m.roundTripperFor(c.auth); rt != nil {
+			creditsCtx = context.WithValue(creditsCtx, roundTripperContextKey{}, rt)
+			creditsCtx = context.WithValue(creditsCtx, "cliproxy.roundtripper", rt)
+		}
+		creditsOpts := ensureRequestedModelMetadata(opts, routeModel)
+		publishSelectedAuthMetadata(creditsOpts.Metadata, c.auth.ID)
+		models := m.executionModelCandidates(c.auth, routeModel)
+		if len(models) == 0 {
+			continue
+		}
+		result, errStream := m.executeStreamWithModelPool(creditsCtx, c.executor, c.auth, c.provider, req, creditsOpts, routeModel, models, len(models) > 1)
+		if errStream != nil {
+			continue
+		}
+		return result, true
+	}
+	return nil, false
+}
+
 func (m *Manager) persist(ctx context.Context, auth *Auth) error {
 	if m.store == nil || auth == nil {
 		return nil
@@ -3200,14 +3386,15 @@ func (m *Manager) refreshAuth(ctx context.Context, id string) {
 	m.mu.RLock()
 	auth := m.auths[id]
 	var exec ProviderExecutor
+	var cloned *Auth
 	if auth != nil {
 		exec = m.executors[auth.Provider]
+		cloned = auth.Clone()
 	}
 	m.mu.RUnlock()
 	if auth == nil || exec == nil {
 		return
 	}
-	cloned := auth.Clone()
 	updated, err := exec.Refresh(ctx, cloned)
 	if err != nil && errors.Is(err, context.Canceled) {
 		log.Debugf("refresh canceled for %s, %s", auth.Provider, auth.ID)

sdk/cliproxy/auth/conductor_credits_candidates_test.go

@@ -0,0 +1,61 @@
+package auth
+
+import (
+	"testing"
+	"time"
+
+	cliproxyexecutor "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/executor"
+)
+
+func TestFindAllAntigravityCreditsCandidateAuths_PrefersKnownCreditsThenUnknown(t *testing.T) {
+	m := &Manager{
+		auths: map[string]*Auth{
+			"zz-credits": {ID: "zz-credits", Provider: "antigravity"},
+			"aa-unknown": {ID: "aa-unknown", Provider: "antigravity"},
+			"mm-no":      {ID: "mm-no", Provider: "antigravity"},
+		},
+		executors: map[string]ProviderExecutor{
+			"antigravity": schedulerTestExecutor{},
+		},
+	}
+
+	SetAntigravityCreditsHint("zz-credits", AntigravityCreditsHint{
+		Known:     true,
+		Available: true,
+		UpdatedAt: time.Now(),
+	})
+	SetAntigravityCreditsHint("mm-no", AntigravityCreditsHint{
+		Known:     true,
+		Available: false,
+		UpdatedAt: time.Now(),
+	})
+
+	opts := cliproxyexecutor.Options{}
+
+	candidates := m.findAllAntigravityCreditsCandidateAuths("claude-sonnet-4-6", opts)
+	if len(candidates) != 2 {
+		t.Fatalf("candidates len = %d, want 2", len(candidates))
+	}
+	if candidates[0].auth.ID != "zz-credits" {
+		t.Fatalf("candidates[0].auth.ID = %q, want %q", candidates[0].auth.ID, "zz-credits")
+	}
+	if candidates[1].auth.ID != "aa-unknown" {
+		t.Fatalf("candidates[1].auth.ID = %q, want %q", candidates[1].auth.ID, "aa-unknown")
+	}
+
+	nonClaude := m.findAllAntigravityCreditsCandidateAuths("gemini-3-flash", opts)
+	if len(nonClaude) != 0 {
+		t.Fatalf("nonClaude len = %d, want 0", len(nonClaude))
+	}
+
+	pinnedOpts := cliproxyexecutor.Options{
+		Metadata: map[string]any{cliproxyexecutor.PinnedAuthMetadataKey: "aa-unknown"},
+	}
+	pinned := m.findAllAntigravityCreditsCandidateAuths("claude-sonnet-4-6", pinnedOpts)
+	if len(pinned) != 1 {
+		t.Fatalf("pinned len = %d, want 1", len(pinned))
+	}
+	if pinned[0].auth.ID != "aa-unknown" {
+		t.Fatalf("pinned[0].auth.ID = %q, want %q", pinned[0].auth.ID, "aa-unknown")
+	}
+}