fix(antigravity): refine 429 handling and credits fallback

Includes: restore SDK docs under docs/; update antigravity executor credits tests; gofmt.

config.example.yaml

@@ -109,7 +109,6 @@ enable-gemini-cli-endpoint: false
 
 # When > 0, emit blank lines every N seconds for non-streaming responses to prevent idle timeouts.
 nonstream-keepalive-interval: 0
-
 # Streaming behavior (SSE keep-alives + safe bootstrap retries).
 # streaming:
 #   keepalive-seconds: 15   # Default: 0 (disabled). <= 0 disables keep-alives.

internal/runtime/executor/antigravity_executor.go

@@ -38,34 +38,58 @@ import (
 )
 
 const (
-	antigravityBaseURLDaily        = "https://daily-cloudcode-pa.googleapis.com"
+	antigravityBaseURLDaily                = "https://daily-cloudcode-pa.googleapis.com"
-	antigravitySandboxBaseURLDaily = "https://daily-cloudcode-pa.sandbox.googleapis.com"
+	antigravitySandboxBaseURLDaily         = "https://daily-cloudcode-pa.sandbox.googleapis.com"
-	antigravityBaseURLProd         = "https://cloudcode-pa.googleapis.com"
+	antigravityBaseURLProd                 = "https://cloudcode-pa.googleapis.com"
-	antigravityCountTokensPath     = "/v1internal:countTokens"
+	antigravityCountTokensPath             = "/v1internal:countTokens"
-	antigravityStreamPath          = "/v1internal:streamGenerateContent"
+	antigravityStreamPath                  = "/v1internal:streamGenerateContent"
-	antigravityGeneratePath        = "/v1internal:generateContent"
+	antigravityGeneratePath                = "/v1internal:generateContent"
-	antigravityClientID            = "1071006060591-tmhssin2h21lcre235vtolojh4g403ep.apps.googleusercontent.com"
+	antigravityClientID                    = "1071006060591-tmhssin2h21lcre235vtolojh4g403ep.apps.googleusercontent.com"
-	antigravityClientSecret        = "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"
+	antigravityClientSecret                = "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"
-	defaultAntigravityAgent        = "antigravity/1.21.9 darwin/arm64" // fallback only; overridden at runtime by misc.AntigravityUserAgent()
+	defaultAntigravityAgent                = "antigravity/1.21.9 darwin/arm64" // fallback only; overridden at runtime by misc.AntigravityUserAgent()
-	antigravityAuthType            = "antigravity"
+	antigravityAuthType                    = "antigravity"
-	refreshSkew                    = 3000 * time.Second
+	refreshSkew                            = 3000 * time.Second
-	antigravityCreditsRetryTTL     = 5 * time.Hour
+	antigravityCreditsRetryTTL             = 5 * time.Hour
+	antigravityCreditsAutoDisableDuration  = 5 * time.Hour
+	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**"
 )
 
 type antigravity429Category string
 
+type antigravityCreditsFailureState struct {
+	Count                    int
+	DisabledUntil            time.Time
+	PermanentlyDisabled      bool
+	ExplicitBalanceExhausted bool
+}
+
+type antigravity429DecisionKind string
+
 const (
-	antigravity429Unknown        antigravity429Category = "unknown"
+	antigravity429Unknown                         antigravity429Category     = "unknown"
-	antigravity429RateLimited    antigravity429Category = "rate_limited"
+	antigravity429RateLimited                     antigravity429Category     = "rate_limited"
-	antigravity429QuotaExhausted antigravity429Category = "quota_exhausted"
+	antigravity429QuotaExhausted                  antigravity429Category     = "quota_exhausted"
+	antigravity429SoftRateLimit                   antigravity429Category     = "soft_rate_limit"
+	antigravity429DecisionSoftRetry               antigravity429DecisionKind = "soft_retry"
+	antigravity429DecisionInstantRetrySameAuth    antigravity429DecisionKind = "instant_retry_same_auth"
+	antigravity429DecisionShortCooldownSwitchAuth antigravity429DecisionKind = "short_cooldown_switch_auth"
+	antigravity429DecisionFullQuotaExhausted      antigravity429DecisionKind = "full_quota_exhausted"
 )
 
+type antigravity429Decision struct {
+	kind       antigravity429DecisionKind
+	retryAfter *time.Duration
+	reason     string
+}
+
 var (
 	randSource                        = rand.New(rand.NewSource(time.Now().UnixNano()))
 	randSourceMutex                   sync.Mutex
-	antigravityCreditsExhaustedByAuth sync.Map
+	antigravityCreditsFailureByAuth   sync.Map
 	antigravityPreferCreditsByModel   sync.Map
+	antigravityShortCooldownByAuth    sync.Map
 	antigravityQuotaExhaustedKeywords = []string{
 		"quota_exhausted",
 		"quota exhausted",
@@ -229,36 +253,77 @@ func injectEnabledCreditTypes(payload []byte) []byte {
 }
 
 func classifyAntigravity429(body []byte) antigravity429Category {
-	if len(body) == 0 {
+	switch decideAntigravity429(body).kind {
+	case antigravity429DecisionInstantRetrySameAuth, antigravity429DecisionShortCooldownSwitchAuth:
+		return antigravity429RateLimited
+	case antigravity429DecisionFullQuotaExhausted:
+		return antigravity429QuotaExhausted
+	case antigravity429DecisionSoftRetry:
+		return antigravity429SoftRateLimit
+	default:
 		return antigravity429Unknown
 	}
+}
+
+func decideAntigravity429(body []byte) antigravity429Decision {
+	decision := antigravity429Decision{kind: antigravity429DecisionSoftRetry}
+	if len(body) == 0 {
+		return decision
+	}
+
+	if retryAfter, parseErr := parseRetryDelay(body); parseErr == nil && retryAfter != nil {
+		decision.retryAfter = retryAfter
+	}
+
 	lowerBody := strings.ToLower(string(body))
 	for _, keyword := range antigravityQuotaExhaustedKeywords {
 		if strings.Contains(lowerBody, keyword) {
-			return antigravity429QuotaExhausted
+			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 antigravity429Unknown
+		return decision
 	}
+
 	details := gjson.GetBytes(body, "error.details")
 	if !details.Exists() || !details.IsArray() {
-		return antigravity429Unknown
+		decision.kind = antigravity429DecisionSoftRetry
+		return decision
 	}
+
 	for _, detail := range details.Array() {
 		if detail.Get("@type").String() != "type.googleapis.com/google.rpc.ErrorInfo" {
 			continue
 		}
 		reason := strings.TrimSpace(detail.Get("reason").String())
-		if strings.EqualFold(reason, "QUOTA_EXHAUSTED") {
+		decision.reason = reason
-			return antigravity429QuotaExhausted
+		switch {
-		}
+		case strings.EqualFold(reason, "QUOTA_EXHAUSTED"):
-		if strings.EqualFold(reason, "RATE_LIMIT_EXCEEDED") {
+			decision.kind = antigravity429DecisionFullQuotaExhausted
-			return antigravity429RateLimited
+			return decision
+		case strings.EqualFold(reason, "RATE_LIMIT_EXCEEDED"):
+			if decision.retryAfter == nil {
+				decision.kind = antigravity429DecisionSoftRetry
+				return decision
+			}
+			switch {
+			case *decision.retryAfter < antigravityInstantRetryThreshold:
+				decision.kind = antigravity429DecisionInstantRetrySameAuth
+			case *decision.retryAfter < antigravityShortQuotaCooldownThreshold:
+				decision.kind = antigravity429DecisionShortCooldownSwitchAuth
+			default:
+				decision.kind = antigravity429DecisionFullQuotaExhausted
+			}
+			return decision
 		}
 	}
-	return antigravity429Unknown
+
+	decision.kind = antigravity429DecisionSoftRetry
+	return decision
 }
 
 func antigravityHasQuotaResetDelayOrModelInfo(body []byte) bool {
@@ -287,38 +352,91 @@ func antigravityCreditsRetryEnabled(cfg *config.Config) bool {
 	return cfg != nil && cfg.QuotaExceeded.AntigravityCredits
 }
 
-func antigravityCreditsExhausted(auth *cliproxyauth.Auth, now time.Time) bool {
+func antigravityCreditsFailureStateForAuth(auth *cliproxyauth.Auth) (string, antigravityCreditsFailureState, bool) {
 	if auth == nil || strings.TrimSpace(auth.ID) == "" {
-		return false
+		return "", antigravityCreditsFailureState{}, false
 	}
-	value, ok := antigravityCreditsExhaustedByAuth.Load(auth.ID)
+	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
 	}
-	until, ok := value.(time.Time)
+	if state.PermanentlyDisabled {
-	if !ok || until.IsZero() {
+		return true
-		antigravityCreditsExhaustedByAuth.Delete(auth.ID)
+	}
+	if state.DisabledUntil.IsZero() {
 		return false
 	}
-	if !until.After(now) {
+	if state.DisabledUntil.After(now) {
-		antigravityCreditsExhaustedByAuth.Delete(auth.ID)
+		return true
-		return false
 	}
-	return true
+	antigravityCreditsFailureByAuth.Delete(authID)
+	return false
 }
 
-func markAntigravityCreditsExhausted(auth *cliproxyauth.Auth, now time.Time) {
+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
 	}
-	antigravityCreditsExhaustedByAuth.Store(auth.ID, now.Add(antigravityCreditsRetryTTL))
+	antigravityCreditsFailureByAuth.Delete(strings.TrimSpace(auth.ID))
 }
-
+func markAntigravityCreditsPermanentlyDisabled(auth *cliproxyauth.Auth) {
-func clearAntigravityCreditsExhausted(auth *cliproxyauth.Auth) {
 	if auth == nil || strings.TrimSpace(auth.ID) == "" {
 		return
 	}
-	antigravityCreditsExhaustedByAuth.Delete(auth.ID)
+	authID := strings.TrimSpace(auth.ID)
+	state := antigravityCreditsFailureState{
+		PermanentlyDisabled:      true,
+		ExplicitBalanceExhausted: true,
+	}
+	antigravityCreditsFailureByAuth.Store(authID, state)
+}
+
+func antigravityHasExplicitCreditsBalanceExhaustedReason(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
+		}
+		reason := strings.TrimSpace(detail.Get("reason").String())
+		if strings.EqualFold(reason, "INSUFFICIENT_G1_CREDITS_BALANCE") {
+			return true
+		}
+	}
+	return false
 }
 
 func antigravityPreferCreditsKey(auth *cliproxyauth.Auth, modelName string) string {
@@ -386,7 +504,7 @@ func shouldMarkAntigravityCreditsExhausted(statusCode int, body []byte, reqErr e
 		if strings.Contains(lowerBody, keyword) {
 			if keyword == "resource has been exhausted" &&
 				statusCode == http.StatusTooManyRequests &&
-				classifyAntigravity429(body) == antigravity429Unknown &&
+				decideAntigravity429(body).kind == antigravity429DecisionSoftRetry &&
 				!antigravityHasQuotaResetDelayOrModelInfo(body) {
 				return false
 			}
@@ -421,11 +539,23 @@ func (e *AntigravityExecutor) attemptCreditsFallback(
 	if !antigravityCreditsRetryEnabled(e.cfg) {
 		return nil, false
 	}
-	if classifyAntigravity429(originalBody) != antigravity429QuotaExhausted {
+	if decideAntigravity429(originalBody).kind != antigravity429DecisionFullQuotaExhausted {
 		return nil, false
 	}
 	now := time.Now()
-	if antigravityCreditsExhausted(auth, 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)
@@ -436,17 +566,21 @@ func (e *AntigravityExecutor) attemptCreditsFallback(
 	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)
-		clearAntigravityCreditsExhausted(auth)
+		clearAntigravityCreditsFailureState(auth)
 		return httpResp, true
 	}
 
@@ -457,24 +591,75 @@ func (e *AntigravityExecutor) attemptCreditsFallback(
 	}
 	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 shouldMarkAntigravityCreditsExhausted(httpResp.StatusCode, bodyBytes, nil) {
+	if shouldForcePermanentDisableCredits(bodyBytes) {
 		clearAntigravityPreferCredits(auth, modelName)
-		markAntigravityCreditsExhausted(auth, now)
+		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" {
 		return resp, statusErr{code: http.StatusNotImplemented, msg: "/responses/compact not supported"}
 	}
 	baseModel := thinking.ParseSuffix(req.Model).ModelName
-	isClaude := strings.Contains(strings.ToLower(baseModel), "claude")
+	if inCooldown, remaining := antigravityIsInShortCooldown(auth, baseModel, time.Now()); inCooldown {
+		log.Debugf("antigravity executor: auth %s in short cooldown for model %s (%s remaining), returning 429 to switch auth", auth.ID, baseModel, remaining)
+		d := remaining
+		return resp, statusErr{code: http.StatusTooManyRequests, msg: fmt.Sprintf("auth in short cooldown, %s remaining", remaining), retryAfter: &d}
+	}
 
+	isClaude := strings.Contains(strings.ToLower(baseModel), "claude")
 	if isClaude || strings.Contains(baseModel, "gemini-3-pro") || strings.Contains(baseModel, "gemini-3.1-flash-image") {
 		return e.executeClaudeNonStream(ctx, auth, req, opts)
 	}
@@ -511,7 +696,6 @@ func (e *AntigravityExecutor) Execute(ctx context.Context, auth *cliproxyauth.Au
 
 	baseURLs := antigravityBaseURLFallbackOrder(auth)
 	httpClient := newAntigravityHTTPClient(ctx, e.cfg, auth, 0)
-
 	attempts := antigravityRetryAttempts(auth, e.cfg)
 
 attemptLoop:
@@ -529,6 +713,7 @@ attemptLoop:
 					usedCreditsDirect = true
 				}
 			}
+
 			httpReq, errReq := e.buildRequest(ctx, auth, token, baseModel, requestPayload, false, opts.Alt, baseURL)
 			if errReq != nil {
 				err = errReq
@@ -565,31 +750,50 @@ attemptLoop:
 			helps.AppendAPIResponseChunk(ctx, e.cfg, bodyBytes)
 
 			if httpResp.StatusCode == http.StatusTooManyRequests {
-				if usedCreditsDirect {
+				decision := decideAntigravity429(bodyBytes)
-					if shouldMarkAntigravityCreditsExhausted(httpResp.StatusCode, bodyBytes, nil) {
+				switch decision.kind {
-						clearAntigravityPreferCredits(auth, baseModel)
+				case antigravity429DecisionInstantRetrySameAuth:
-						markAntigravityCreditsExhausted(auth, time.Now())
+					if attempt+1 < attempts {
+						if decision.retryAfter != nil && *decision.retryAfter > 0 {
+							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
+							}
+						}
+						continue attemptLoop
 					}
-				} else {
+				case antigravity429DecisionShortCooldownSwitchAuth:
-					creditsResp, _ := e.attemptCreditsFallback(ctx, auth, httpClient, token, baseModel, translated, false, opts.Alt, baseURL, bodyBytes)
+					if decision.retryAfter != nil && *decision.retryAfter > 0 {
-					if creditsResp != nil {
+						markAntigravityShortCooldown(auth, baseModel, time.Now(), *decision.retryAfter)
-						helps.RecordAPIResponseMetadata(ctx, e.cfg, creditsResp.StatusCode, creditsResp.Header.Clone())
+						log.Debugf("antigravity executor: short quota cooldown (%s) for model %s, recorded cooldown and skipping credits fallback", *decision.retryAfter, baseModel)
-						creditsBody, errCreditsRead := io.ReadAll(creditsResp.Body)
+					}
-						if errClose := creditsResp.Body.Close(); errClose != nil {
+				case antigravity429DecisionFullQuotaExhausted:
-							log.Errorf("antigravity executor: close credits success response body error: %v", errClose)
+					if usedCreditsDirect {
+						clearAntigravityPreferCredits(auth, baseModel)
+						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
 						}
-						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
 					}
 				}
 			}
@@ -625,6 +829,16 @@ attemptLoop:
 						continue attemptLoop
 					}
 				}
+				if antigravityShouldRetrySoftRateLimit(httpResp.StatusCode, bodyBytes) {
+					if attempt+1 < attempts {
+						delay := antigravitySoftRateLimitDelay(attempt)
+						log.Debugf("antigravity executor: soft rate limit for model %s, retrying in %s (attempt %d/%d)", baseModel, delay, attempt+1, attempts)
+						if errWait := antigravityWait(ctx, delay); errWait != nil {
+							return resp, errWait
+						}
+						continue attemptLoop
+					}
+				}
 				err = newAntigravityStatusErr(httpResp.StatusCode, bodyBytes)
 				return resp, err
 			}
@@ -654,6 +868,11 @@ attemptLoop:
 // executeClaudeNonStream performs a claude non-streaming request to the Antigravity API.
 func (e *AntigravityExecutor) executeClaudeNonStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (resp cliproxyexecutor.Response, err error) {
 	baseModel := thinking.ParseSuffix(req.Model).ModelName
+	if inCooldown, remaining := antigravityIsInShortCooldown(auth, baseModel, time.Now()); inCooldown {
+		log.Debugf("antigravity executor: auth %s in short cooldown for model %s (%s remaining), returning 429 to switch auth", auth.ID, baseModel, remaining)
+		d := remaining
+		return resp, statusErr{code: http.StatusTooManyRequests, msg: fmt.Sprintf("auth in short cooldown, %s remaining", remaining), retryAfter: &d}
+	}
 
 	token, updatedAuth, errToken := e.ensureAccessToken(ctx, auth)
 	if errToken != nil {
@@ -755,19 +974,40 @@ attemptLoop:
 				}
 				helps.AppendAPIResponseChunk(ctx, e.cfg, bodyBytes)
 				if httpResp.StatusCode == http.StatusTooManyRequests {
-					if usedCreditsDirect {
+					decision := decideAntigravity429(bodyBytes)
-						if shouldMarkAntigravityCreditsExhausted(httpResp.StatusCode, bodyBytes, nil) {
+
-							clearAntigravityPreferCredits(auth, baseModel)
+					switch decision.kind {
-							markAntigravityCreditsExhausted(auth, time.Now())
+					case antigravity429DecisionInstantRetrySameAuth:
+						if attempt+1 < attempts {
+							if decision.retryAfter != nil && *decision.retryAfter > 0 {
+								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
+								}
+							}
+							continue attemptLoop
 						}
-					} else {
+					case antigravity429DecisionShortCooldownSwitchAuth:
-						creditsResp, _ := e.attemptCreditsFallback(ctx, auth, httpClient, token, baseModel, translated, true, opts.Alt, baseURL, bodyBytes)
+						if decision.retryAfter != nil && *decision.retryAfter > 0 {
-						if creditsResp != nil {
+							markAntigravityShortCooldown(auth, baseModel, time.Now(), *decision.retryAfter)
-							httpResp = creditsResp
+							log.Debugf("antigravity executor: short quota cooldown (%s) for model %s, recorded cooldown and skipping credits fallback", *decision.retryAfter, baseModel)
-							helps.RecordAPIResponseMetadata(ctx, e.cfg, httpResp.StatusCode, httpResp.Header.Clone())
+						}
+					case antigravity429DecisionFullQuotaExhausted:
+						if usedCreditsDirect {
+							clearAntigravityPreferCredits(auth, baseModel)
+							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())
+							}
 						}
 					}
 				}
+
 				if httpResp.StatusCode >= http.StatusOK && httpResp.StatusCode < http.StatusMultipleChoices {
 					goto streamSuccessClaudeNonStream
 				}
@@ -800,6 +1040,16 @@ attemptLoop:
 						continue attemptLoop
 					}
 				}
+				if antigravityShouldRetrySoftRateLimit(httpResp.StatusCode, bodyBytes) {
+					if attempt+1 < attempts {
+						delay := antigravitySoftRateLimitDelay(attempt)
+						log.Debugf("antigravity executor: soft rate limit for model %s, retrying in %s (attempt %d/%d)", baseModel, delay, attempt+1, attempts)
+						if errWait := antigravityWait(ctx, delay); errWait != nil {
+							return resp, errWait
+						}
+						continue attemptLoop
+					}
+				}
 				err = newAntigravityStatusErr(httpResp.StatusCode, bodyBytes)
 				return resp, err
 			}
@@ -1079,6 +1329,11 @@ func (e *AntigravityExecutor) ExecuteStream(ctx context.Context, auth *cliproxya
 	baseModel := thinking.ParseSuffix(req.Model).ModelName
 
 	ctx = context.WithValue(ctx, "alt", "")
+	if inCooldown, remaining := antigravityIsInShortCooldown(auth, baseModel, time.Now()); inCooldown {
+		log.Debugf("antigravity executor: auth %s in short cooldown for model %s (%s remaining), returning 429 to switch auth", auth.ID, baseModel, remaining)
+		d := remaining
+		return nil, statusErr{code: http.StatusTooManyRequests, msg: fmt.Sprintf("auth in short cooldown, %s remaining", remaining), retryAfter: &d}
+	}
 
 	token, updatedAuth, errToken := e.ensureAccessToken(ctx, auth)
 	if errToken != nil {
@@ -1179,19 +1434,40 @@ attemptLoop:
 				}
 				helps.AppendAPIResponseChunk(ctx, e.cfg, bodyBytes)
 				if httpResp.StatusCode == http.StatusTooManyRequests {
-					if usedCreditsDirect {
+					decision := decideAntigravity429(bodyBytes)
-						if shouldMarkAntigravityCreditsExhausted(httpResp.StatusCode, bodyBytes, nil) {
+
-							clearAntigravityPreferCredits(auth, baseModel)
+					switch decision.kind {
-							markAntigravityCreditsExhausted(auth, time.Now())
+					case antigravity429DecisionInstantRetrySameAuth:
+						if attempt+1 < attempts {
+							if decision.retryAfter != nil && *decision.retryAfter > 0 {
+								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
+								}
+							}
+							continue attemptLoop
 						}
-					} else {
+					case antigravity429DecisionShortCooldownSwitchAuth:
-						creditsResp, _ := e.attemptCreditsFallback(ctx, auth, httpClient, token, baseModel, translated, true, opts.Alt, baseURL, bodyBytes)
+						if decision.retryAfter != nil && *decision.retryAfter > 0 {
-						if creditsResp != nil {
+							markAntigravityShortCooldown(auth, baseModel, time.Now(), *decision.retryAfter)
-							httpResp = creditsResp
+							log.Debugf("antigravity executor: short quota cooldown (%s) for model %s, recorded cooldown and skipping credits fallback", *decision.retryAfter, baseModel)
-							helps.RecordAPIResponseMetadata(ctx, e.cfg, httpResp.StatusCode, httpResp.Header.Clone())
+						}
+					case antigravity429DecisionFullQuotaExhausted:
+						if usedCreditsDirect {
+							clearAntigravityPreferCredits(auth, baseModel)
+							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())
+							}
 						}
 					}
 				}
+
 				if httpResp.StatusCode >= http.StatusOK && httpResp.StatusCode < http.StatusMultipleChoices {
 					goto streamSuccessExecuteStream
 				}
@@ -1224,6 +1500,16 @@ attemptLoop:
 						continue attemptLoop
 					}
 				}
+				if antigravityShouldRetrySoftRateLimit(httpResp.StatusCode, bodyBytes) {
+					if attempt+1 < attempts {
+						delay := antigravitySoftRateLimitDelay(attempt)
+						log.Debugf("antigravity executor: soft rate limit for model %s, retrying in %s (attempt %d/%d)", baseModel, delay, attempt+1, attempts)
+						if errWait := antigravityWait(ctx, delay); errWait != nil {
+							return nil, errWait
+						}
+						continue attemptLoop
+					}
+				}
 				err = newAntigravityStatusErr(httpResp.StatusCode, bodyBytes)
 				return nil, err
 			}
@@ -1844,6 +2130,66 @@ func antigravityShouldRetryTransientResourceExhausted429(statusCode int, body []
 	return strings.Contains(msg, "resource has been exhausted")
 }
 
+func antigravityShouldRetrySoftRateLimit(statusCode int, body []byte) bool {
+	if statusCode != http.StatusTooManyRequests {
+		return false
+	}
+	return decideAntigravity429(body).kind == antigravity429DecisionSoftRetry
+}
+
+func antigravitySoftRateLimitDelay(attempt int) time.Duration {
+	if attempt < 0 {
+		attempt = 0
+	}
+	base := time.Duration(attempt+1) * 500 * time.Millisecond
+	if base > 3*time.Second {
+		base = 3 * time.Second
+	}
+	return base
+}
+
+func antigravityShortCooldownKey(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 + "|sc"
+}
+
+func antigravityIsInShortCooldown(auth *cliproxyauth.Auth, modelName string, now time.Time) (bool, time.Duration) {
+	key := antigravityShortCooldownKey(auth, modelName)
+	if key == "" {
+		return false, 0
+	}
+	value, ok := antigravityShortCooldownByAuth.Load(key)
+	if !ok {
+		return false, 0
+	}
+	until, ok := value.(time.Time)
+	if !ok || until.IsZero() {
+		antigravityShortCooldownByAuth.Delete(key)
+		return false, 0
+	}
+	remaining := until.Sub(now)
+	if remaining <= 0 {
+		antigravityShortCooldownByAuth.Delete(key)
+		return false, 0
+	}
+	return true, remaining
+}
+
+func markAntigravityShortCooldown(auth *cliproxyauth.Auth, modelName string, now time.Time, duration time.Duration) {
+	key := antigravityShortCooldownKey(auth, modelName)
+	if key == "" {
+		return
+	}
+	antigravityShortCooldownByAuth.Store(key, now.Add(duration))
+}
+
 func antigravityNoCapacityRetryDelay(attempt int) time.Duration {
 	if attempt < 0 {
 		attempt = 0
@@ -1866,6 +2212,13 @@ func antigravityTransient429RetryDelay(attempt int) time.Duration {
 	return delay
 }
 
+func antigravityInstantRetryDelay(wait time.Duration) time.Duration {
+	if wait <= 0 {
+		return 0
+	}
+	return wait + 800*time.Millisecond
+}
+
 func antigravityWait(ctx context.Context, wait time.Duration) error {
 	if wait <= 0 {
 		return nil

internal/runtime/executor/antigravity_executor_credits_test.go

@@ -17,8 +17,9 @@ import (
 )
 
 func resetAntigravityCreditsRetryState() {
-	antigravityCreditsExhaustedByAuth = sync.Map{}
+	antigravityCreditsFailureByAuth = sync.Map{}
 	antigravityPreferCreditsByModel = sync.Map{}
+	antigravityShortCooldownByAuth = sync.Map{}
 }
 
 func TestClassifyAntigravity429(t *testing.T) {
@@ -58,10 +59,10 @@ func TestClassifyAntigravity429(t *testing.T) {
 		}
 	})
 
-	t.Run("unknown", func(t *testing.T) {
+	t.Run("unstructured 429 defaults to soft rate limit", func(t *testing.T) {
 		body := []byte(`{"error":{"message":"too many requests"}}`)
-		if got := classifyAntigravity429(body); got != antigravity429Unknown {
+		if got := classifyAntigravity429(body); got != antigravity429SoftRateLimit {
-			t.Fatalf("classifyAntigravity429() = %q, want %q", got, antigravity429Unknown)
+			t.Fatalf("classifyAntigravity429() = %q, want %q", got, antigravity429SoftRateLimit)
 		}
 	})
 }
@@ -255,7 +256,7 @@ func TestAntigravityExecute_SkipsCreditsRetryWhenAlreadyExhausted(t *testing.T)
 			"expired":      time.Now().Add(1 * time.Hour).Format(time.RFC3339),
 		},
 	}
-	markAntigravityCreditsExhausted(auth, time.Now())
+	recordAntigravityCreditsFailure(auth, time.Now())
 
 	_, err := exec.Execute(context.Background(), auth, cliproxyexecutor.Request{
 		Model:   "gemini-2.5-flash",