perf(queue): 按策略动态扩缩异步 Worker

将 River 执行容量按平台模型与用户组的有效并发策略动态调整，并为长任务续租、并发租约原子抢占和限流退避补充保护。\n\n统一平台模型限流继承语义，兼容历史 platformLimits/modelLimits，并为三个迁移图像模型建立独立并发租约。补充管理端显式继承/覆盖配置、指标、单元测试及隔离 PostgreSQL 验收。\n\n验证：go test ./... -count=1；pnpm lint；pnpm test；pnpm build；隔离 PostgreSQL 并发原子性/续租测试；128 任务与三分钟长任务动态 Worker 验收。
2026-07-24 12:26:56 +08:00
parent 290b8c1854
commit 6c5daf29ca
37 changed files with 2365 additions and 205 deletions
@@ -29,6 +29,11 @@ IDENTITY_MODE=hybrid
 # - hold: reject new production generation before any upstream request; existing settlements continue.
 BILLING_ENGINE_MODE=observe

+# River 执行 worker 会按平台模型和活跃用户组中更严格的 concurrent 总量每 5 秒重算。
+# hard limit 是单进程安全边界；平台模型/用户组 concurrency lease 才是业务并发真值。
+AI_GATEWAY_ASYNC_WORKER_HARD_LIMIT=2048
+AI_GATEWAY_ASYNC_WORKER_REFRESH_INTERVAL_SECONDS=5
+
 # Unified identity business settings are managed in System Settings > Unified
 # Identity. Deployment only supplies the SecretStore and infrastructure timing.
 AI_GATEWAY_PUBLIC_BASE_URL=http://localhost:8088
@@ -159,6 +159,8 @@ AI_GATEWAY_DATABASE_URL=postgresql://easyai:easyai2025@localhost:5432/easyai_ai_

 如果现有 `easyai-pgvector` 没有把 `5432` 映射到宿主机，就需要补端口映射，或者把 AI Gateway 后端容器化后接入同一个 `easyai` Docker network。

+异步队列 worker 不使用固定并发。服务分别汇总启用平台模型和活跃用户组的有效 `concurrent` 策略，采用两者中更严格的容量，默认每 5 秒在线调整 River 执行容量。策略解析兼容历史 `platformLimits/modelLimits.max_concurrent_requests`，运行时统一转换为 `rules`。`AI_GATEWAY_ASYNC_WORKER_HARD_LIMIT` 默认 `2048`，仅作为单进程资源安全边界；平台模型和用户组的 PostgreSQL concurrency lease 才是业务并发真值。可通过 `AI_GATEWAY_ASYNC_WORKER_REFRESH_INTERVAL_SECONDS` 调整刷新周期。
+
 ## 迁移原则

 1. 新服务先并行运行，不直接删除 `easyai-server-main` 内现有模块。
@@ -12848,6 +12848,13 @@
                    "type": "object",
                    "additionalProperties": {}
                },
+                "rateLimitPolicyMode": {
+                    "type": "string",
+                    "enum": [
+                        "inherit",
+                        "override"
+                    ]
+                },
                "retryPolicy": {
                    "type": "object",
                    "additionalProperties": {}
@@ -13838,6 +13845,13 @@
                    "type": "object",
                    "additionalProperties": {}
                },
+                "rateLimitPolicyMode": {
+                    "type": "string",
+                    "enum": [
+                        "inherit",
+                        "override"
+                    ]
+                },
                "retryPolicy": {
                    "type": "object",
                    "additionalProperties": {}
@@ -2439,6 +2439,11 @@ definitions:
      rateLimitPolicy:
        additionalProperties: {}
        type: object
+      rateLimitPolicyMode:
+        enum:
+        - inherit
+        - override
+        type: string
      retryPolicy:
        additionalProperties: {}
        type: object
@@ -3108,6 +3113,11 @@ definitions:
      rateLimitPolicy:
        additionalProperties: {}
        type: object
+      rateLimitPolicyMode:
+        enum:
+        - inherit
+        - override
+        type: string
      retryPolicy:
        additionalProperties: {}
        type: object
@@ -48,6 +48,8 @@ type Config struct {
 	GlobalHTTPProxySource                         string
 	LogLevel                                      slog.Level
 	BillingEngineMode                             string
+	AsyncWorkerHardLimit                          int
+	AsyncWorkerRefreshIntervalSeconds             int
 }

 func Load() Config {
@@ -85,13 +87,15 @@ func Load() Config {
 		TaskProgressCallbackURL: env("TASK_PROGRESS_CALLBACK_URL",
 			strings.TrimRight(env("SERVER_MAIN_BASE_URL", "http://localhost:3000"), "/")+"/internal/platform/task-progress-callbacks",
 		),
-		TaskProgressCallbackTimeoutMS:   env("TASK_PROGRESS_CALLBACK_TIMEOUT_MS", "5000"),
-		TaskProgressCallbackMaxAttempts: env("TASK_PROGRESS_CALLBACK_MAX_ATTEMPTS", "10"),
-		CORSAllowedOrigin:               env("CORS_ALLOWED_ORIGIN", "http://localhost:5178,http://127.0.0.1:5178"),
-		GlobalHTTPProxy:                 globalProxy.HTTPProxy,
-		GlobalHTTPProxySource:           globalProxy.Source,
-		LogLevel:                        logLevel(env("LOG_LEVEL", "info")),
-		BillingEngineMode:               strings.ToLower(env("BILLING_ENGINE_MODE", "observe")),
+		TaskProgressCallbackTimeoutMS:     env("TASK_PROGRESS_CALLBACK_TIMEOUT_MS", "5000"),
+		TaskProgressCallbackMaxAttempts:   env("TASK_PROGRESS_CALLBACK_MAX_ATTEMPTS", "10"),
+		CORSAllowedOrigin:                 env("CORS_ALLOWED_ORIGIN", "http://localhost:5178,http://127.0.0.1:5178"),
+		GlobalHTTPProxy:                   globalProxy.HTTPProxy,
+		GlobalHTTPProxySource:             globalProxy.Source,
+		LogLevel:                          logLevel(env("LOG_LEVEL", "info")),
+		BillingEngineMode:                 strings.ToLower(env("BILLING_ENGINE_MODE", "observe")),
+		AsyncWorkerHardLimit:              envIntValidated("AI_GATEWAY_ASYNC_WORKER_HARD_LIMIT", 2048),
+		AsyncWorkerRefreshIntervalSeconds: envIntValidated("AI_GATEWAY_ASYNC_WORKER_REFRESH_INTERVAL_SECONDS", 5),
 	}
 }

@@ -101,6 +105,12 @@ func (c Config) Validate() error {
 	default:
 		return errors.New("BILLING_ENGINE_MODE must be observe, enforce, or hold")
 	}
+	if c.AsyncWorkerHardLimit < 1 || c.AsyncWorkerHardLimit > 10000 {
+		return errors.New("AI_GATEWAY_ASYNC_WORKER_HARD_LIMIT must be between 1 and 10000")
+	}
+	if c.AsyncWorkerRefreshIntervalSeconds < 1 {
+		return errors.New("AI_GATEWAY_ASYNC_WORKER_REFRESH_INTERVAL_SECONDS must be positive")
+	}
 	switch strings.ToLower(strings.TrimSpace(c.IdentitySecretStore)) {
 	case "":
 	case "file":
@@ -210,6 +220,18 @@ func envInt(key string, fallback int) int {
 	return parsed
 }

+func envIntValidated(key string, fallback int) int {
+	value := envValue(key)
+	if value == "" {
+		return fallback
+	}
+	parsed, err := strconv.Atoi(value)
+	if err != nil {
+		return 0
+	}
+	return parsed
+}
+
 func logLevel(value string) slog.Level {
 	switch strings.ToLower(value) {
 	case "debug":
@@ -26,7 +26,7 @@ func TestLoadIdentitySecretStoreUsesNewEnvironmentNamesAndIgnoresLegacyBusinessV
 }

 func TestValidateIdentityFileSecretStoreRequiresDirectory(t *testing.T) {
-	cfg := Config{IdentitySecretStore: "file"}
+	cfg := Config{IdentitySecretStore: "file", AsyncWorkerHardLimit: 2048, AsyncWorkerRefreshIntervalSeconds: 5}
 	if err := cfg.Validate(); err == nil || !strings.Contains(err.Error(), "IDENTITY_SECRET_DIR") {
 		t.Fatalf("Validate() error = %v, want missing identity secret directory", err)
 	}
@@ -37,7 +37,12 @@ func TestValidateIdentityFileSecretStoreRequiresDirectory(t *testing.T) {
 }

 func TestValidateIdentityKubernetesSecretStore(t *testing.T) {
-	cfg := Config{IdentitySecretStore: "kubernetes", IdentityKubernetesSecretName: "easyai-gateway-identity"}
+	cfg := Config{
+		IdentitySecretStore:               "kubernetes",
+		IdentityKubernetesSecretName:      "easyai-gateway-identity",
+		AsyncWorkerHardLimit:              2048,
+		AsyncWorkerRefreshIntervalSeconds: 5,
+	}
 	if err := cfg.Validate(); err == nil || !strings.Contains(err.Error(), "namespace") {
 		t.Fatalf("Validate() error = %v, want missing namespace", err)
 	}
@@ -52,8 +57,27 @@ func TestValidateIdentitySecurityEventTiming(t *testing.T) {
 		IdentitySecurityEventHeartbeatIntervalSeconds: 60,
 		IdentitySecurityEventStaleAfterSeconds:        60,
 		IdentitySecurityEventClockSkewSeconds:         60,
+		AsyncWorkerHardLimit:                          2048,
+		AsyncWorkerRefreshIntervalSeconds:             5,
 	}
 	if err := cfg.Validate(); err == nil || !strings.Contains(err.Error(), "heartbeat") {
 		t.Fatalf("Validate() error = %v, want invalid stale threshold", err)
 	}
 }
+
+func TestValidateAsyncWorkerSettings(t *testing.T) {
+	cfg := Config{AsyncWorkerHardLimit: 10001, AsyncWorkerRefreshIntervalSeconds: 5}
+	if err := cfg.Validate(); err == nil || !strings.Contains(err.Error(), "HARD_LIMIT") {
+		t.Fatalf("Validate() error = %v, want invalid hard limit", err)
+	}
+	cfg.AsyncWorkerHardLimit = 2048
+	cfg.AsyncWorkerRefreshIntervalSeconds = 0
+	if err := cfg.Validate(); err == nil || !strings.Contains(err.Error(), "REFRESH_INTERVAL") {
+		t.Fatalf("Validate() error = %v, want invalid refresh interval", err)
+	}
+	t.Setenv("AI_GATEWAY_ASYNC_WORKER_HARD_LIMIT", "not-an-integer")
+	loaded := Load()
+	if err := loaded.Validate(); err == nil || !strings.Contains(err.Error(), "HARD_LIMIT") {
+		t.Fatalf("Validate() error = %v, want invalid non-integer hard limit", err)
+	}
+}
@@ -0,0 +1,479 @@
+package httpapi
+
+import (
+	"bytes"
+	"context"
+	"encoding/json"
+	"fmt"
+	"io"
+	"log/slog"
+	"net/http"
+	"net/http/httptest"
+	"os"
+	"strconv"
+	"strings"
+	"sync"
+	"testing"
+	"time"
+
+	"github.com/easyai/easyai-ai-gateway/apps/api/internal/config"
+	"github.com/easyai/easyai-ai-gateway/apps/api/internal/store"
+	"github.com/jackc/pgx/v5/pgxpool"
+)
+
+func TestAsyncWorkerDynamicConcurrencyAcceptance(t *testing.T) {
+	databaseURL := strings.TrimSpace(os.Getenv("AI_GATEWAY_TEST_DATABASE_URL"))
+	if databaseURL == "" {
+		t.Skip("set AI_GATEWAY_TEST_DATABASE_URL to run dynamic worker acceptance tests")
+	}
+	ctx, cancel := context.WithTimeout(context.Background(), 7*time.Minute)
+	defer cancel()
+	applyMigration(t, ctx, databaseURL)
+
+	db, err := store.Connect(ctx, databaseURL)
+	if err != nil {
+		t.Fatalf("connect store: %v", err)
+	}
+	defer db.Close()
+	pool, err := pgxpool.New(ctx, databaseURL)
+	if err != nil {
+		t.Fatalf("connect acceptance pool: %v", err)
+	}
+	defer pool.Close()
+	if _, err := pool.Exec(ctx, `UPDATE integration_platforms SET status = 'disabled' WHERE deleted_at IS NULL`); err != nil {
+		t.Fatalf("isolate acceptance platforms: %v", err)
+	}
+
+	serverCtx, cancelServer := context.WithCancel(ctx)
+	defer cancelServer()
+	server := httptest.NewServer(NewServerWithContext(serverCtx, config.Config{
+		AppEnv:                            "test",
+		HTTPAddr:                          ":0",
+		DatabaseURL:                       databaseURL,
+		IdentityMode:                      "hybrid",
+		JWTSecret:                         "test-secret",
+		BillingEngineMode:                 "observe",
+		CORSAllowedOrigin:                 "*",
+		AsyncWorkerHardLimit:              256,
+		AsyncWorkerRefreshIntervalSeconds: 1,
+	}, db, slog.New(slog.NewTextHandler(io.Discard, nil))))
+	defer server.Close()
+
+	adminToken := createAsyncAcceptanceAdmin(t, ctx, pool, server.URL)
+	if _, err := pool.Exec(ctx, `
+UPDATE gateway_user_groups
+SET rate_limit_policy = '{"rules":[{"metric":"concurrent","limit":256,"leaseTtlSeconds":120}]}'::jsonb
+WHERE status = 'active'`); err != nil {
+		t.Fatalf("raise acceptance user-group concurrency: %v", err)
+	}
+	suffix := strconv.FormatInt(time.Now().UnixNano(), 10)
+
+	t.Run("超过旧64上限的高并发", func(t *testing.T) {
+		model := "worker-burst-" + suffix
+		platform := createAsyncAcceptancePlatform(t, server.URL, adminToken, "burst-"+suffix, "Burst Simulation", 128)
+		defer updateAsyncAcceptancePlatform(t, server.URL, adminToken, platform, 128, "disabled")
+		modelID := createAsyncAcceptanceModel(t, server.URL, adminToken, platform.ID, model, "video_generate", "override", 96, 120)
+
+		waitForAsyncWorkerMetric(t, server.URL, "easyai_gateway_async_worker_capacity", 96, 15*time.Second)
+		startedAt := time.Now()
+		taskIDs := submitAsyncSimulationTasks(t, server.URL, adminToken, "/api/v1/videos/generations", model, 128, 15*time.Second)
+		peakRunning, peakLeases := waitForAcceptanceTasks(t, ctx, pool, taskIDs, modelID, 60*time.Second)
+		elapsed := time.Since(startedAt)
+		if peakRunning < 80 || peakLeases < 80 {
+			t.Fatalf("running peak=%d lease peak=%d, want both >=80", peakRunning, peakLeases)
+		}
+		if peakLeases > 96 {
+			t.Fatalf("lease peak=%d exceeded model concurrent limit 96", peakLeases)
+		}
+		assertAcceptanceAttempts(t, ctx, pool, taskIDs, len(taskIDs))
+		if elapsed >= 60*time.Second {
+			t.Fatalf("128 tasks completed in %s, want <60s", elapsed)
+		}
+		t.Logf("高并发证据: tasks=128 duration=%s running_peak=%d lease_peak=%d capacity=96", elapsed.Round(time.Millisecond), peakRunning, peakLeases)
+	})
+
+	t.Run("三分钟视频任务在线扩缩容与续租", func(t *testing.T) {
+		model := "worker-video-long-" + suffix
+		platform := createAsyncAcceptancePlatform(t, server.URL, adminToken, "video-long-"+suffix, "Long Video Simulation", 1)
+		modelID := createAsyncAcceptanceModel(t, server.URL, adminToken, platform.ID, model, "video_generate", "inherit", 0, 120)
+		waitForAsyncWorkerMetric(t, server.URL, "easyai_gateway_async_worker_capacity", 1, 15*time.Second)
+
+		longTaskIDs := submitAsyncSimulationTasks(t, server.URL, adminToken, "/api/v1/videos/generations", model, 3, 180*time.Second)
+		waitForTaskCount(t, ctx, pool, longTaskIDs, "running", 1, 10*time.Second)
+		updateStartedAt := time.Now()
+		updateAsyncAcceptancePlatform(t, server.URL, adminToken, platform, 3, "enabled")
+		waitForAsyncWorkerMetric(t, server.URL, "easyai_gateway_async_worker_desired_capacity", 3, 15*time.Second)
+		waitForAsyncWorkerMetric(t, server.URL, "easyai_gateway_async_worker_capacity", 3, 15*time.Second)
+		waitForActiveLeaseCount(t, ctx, pool, modelID, 3, 15*time.Second)
+		waitForTaskCount(t, ctx, pool, longTaskIDs, "running", 3, 5*time.Second)
+		if time.Since(updateStartedAt) > 15*time.Second {
+			t.Fatalf("capacity expansion took %s, want <=15s", time.Since(updateStartedAt))
+		}
+
+		initialExpiry := activeLeaseExpiry(t, ctx, pool, modelID, 3)
+		updateAsyncAcceptancePlatform(t, server.URL, adminToken, platform, 1, "enabled")
+		waitForAsyncWorkerMetric(t, server.URL, "easyai_gateway_async_worker_capacity", 1, 15*time.Second)
+		shortTaskIDs := submitAsyncSimulationTasks(t, server.URL, adminToken, "/api/v1/videos/generations", model, 1, time.Second)
+		time.Sleep(3 * time.Second)
+		if attempts := taskAttemptCount(t, ctx, pool, shortTaskIDs); attempts != 0 {
+			t.Fatalf("short task created %d upstream attempts while three long leases exceeded reduced limit", attempts)
+		}
+		waitForTaskCount(t, ctx, pool, longTaskIDs, "running", 3, 5*time.Second)
+
+		time.Sleep(130 * time.Second)
+		renewedExpiry := activeLeaseExpiry(t, ctx, pool, modelID, 3)
+		if !renewedExpiry.After(initialExpiry.Add(30*time.Second)) || !renewedExpiry.After(time.Now()) {
+			t.Fatalf("lease was not renewed: initial=%s renewed=%s now=%s", initialExpiry, renewedExpiry, time.Now())
+		}
+
+		waitForTaskCount(t, ctx, pool, longTaskIDs, "succeeded", 3, 75*time.Second)
+		shortStartedAt := time.Now()
+		waitForTaskAttempts(t, ctx, pool, shortTaskIDs, 1, 5*time.Second)
+		waitForTaskCount(t, ctx, pool, shortTaskIDs, "succeeded", 1, 10*time.Second)
+		if time.Since(shortStartedAt) > 5*time.Second {
+			t.Fatalf("short task started after %s once long tasks completed, want <=5s", time.Since(shortStartedAt))
+		}
+		allTaskIDs := append(append([]string{}, longTaskIDs...), shortTaskIDs...)
+		assertAcceptanceAttempts(t, ctx, pool, allTaskIDs, 4)
+		assertSimulationOutputsAndEvents(t, ctx, pool, allTaskIDs)
+		t.Logf("长任务证据: initial_capacity=1 expanded_capacity=3 reduced_capacity=1 initial_expiry=%s renewed_expiry=%s tasks=4", initialExpiry.UTC().Format(time.RFC3339Nano), renewedExpiry.UTC().Format(time.RFC3339Nano))
+	})
+}
+
+type asyncAcceptancePlatform struct {
+	ID          string `json:"id"`
+	Provider    string `json:"provider"`
+	PlatformKey string `json:"platformKey"`
+	Name        string `json:"name"`
+	BaseURL     string `json:"baseUrl"`
+	AuthType    string `json:"authType"`
+	Priority    int    `json:"priority"`
+}
+
+func createAsyncAcceptanceAdmin(t *testing.T, ctx context.Context, pool *pgxpool.Pool, baseURL string) string {
+	t.Helper()
+	suffix := strconv.FormatInt(time.Now().UnixNano(), 10)
+	username := "async_acceptance_" + suffix
+	password := "password123"
+	doJSON(t, baseURL, http.MethodPost, "/api/v1/auth/register", "", map[string]any{
+		"username": username,
+		"email":    username + "@example.com",
+		"password": password,
+	}, http.StatusCreated, &map[string]any{})
+	if _, err := pool.Exec(ctx, `UPDATE gateway_users SET roles = '["admin"]'::jsonb WHERE username = $1`, username); err != nil {
+		t.Fatalf("promote acceptance user: %v", err)
+	}
+	var login struct {
+		AccessToken string `json:"accessToken"`
+	}
+	doJSON(t, baseURL, http.MethodPost, "/api/v1/auth/login", "", map[string]any{
+		"account": username, "password": password,
+	}, http.StatusOK, &login)
+	if login.AccessToken == "" {
+		t.Fatal("acceptance login returned no access token")
+	}
+	return login.AccessToken
+}
+
+func createAsyncAcceptancePlatform(t *testing.T, baseURL, token, key, name string, concurrent int) asyncAcceptancePlatform {
+	t.Helper()
+	var platform asyncAcceptancePlatform
+	doJSON(t, baseURL, http.MethodPost, "/api/admin/platforms", token, asyncAcceptancePlatformPayload(key, name, concurrent, "enabled"), http.StatusCreated, &platform)
+	if platform.ID == "" {
+		t.Fatal("acceptance platform returned no id")
+	}
+	return platform
+}
+
+func updateAsyncAcceptancePlatform(t *testing.T, baseURL, token string, platform asyncAcceptancePlatform, concurrent int, status string) {
+	t.Helper()
+	doJSON(t, baseURL, http.MethodPatch, "/api/admin/platforms/"+platform.ID, token,
+		asyncAcceptancePlatformPayload(platform.PlatformKey, platform.Name, concurrent, status), http.StatusOK, &asyncAcceptancePlatform{})
+}
+
+func asyncAcceptancePlatformPayload(key, name string, concurrent int, status string) map[string]any {
+	return map[string]any{
+		"provider":    "openai",
+		"platformKey": key,
+		"name":        name,
+		"baseUrl":     "https://api.openai.com/v1",
+		"authType":    "bearer",
+		"credentials": map[string]any{"mode": "simulation"},
+		"config":      map[string]any{"testMode": true},
+		"priority":    1,
+		"status":      status,
+		"rateLimitPolicy": map[string]any{"rules": []any{
+			map[string]any{"metric": "concurrent", "limit": concurrent, "leaseTtlSeconds": 120},
+		}},
+	}
+}
+
+func createAsyncAcceptanceModel(t *testing.T, baseURL, token, platformID, model, modelType, mode string, concurrent, ttl int) string {
+	t.Helper()
+	payload := map[string]any{
+		"modelName":           model,
+		"providerModelName":   model,
+		"modelAlias":          model,
+		"modelType":           []string{modelType},
+		"displayName":         model,
+		"rateLimitPolicyMode": mode,
+	}
+	if mode == "override" {
+		payload["rateLimitPolicy"] = map[string]any{"rules": []any{
+			map[string]any{"metric": "concurrent", "limit": concurrent, "leaseTtlSeconds": ttl},
+		}}
+	}
+	var response struct {
+		ID string `json:"id"`
+	}
+	doJSON(t, baseURL, http.MethodPost, "/api/admin/platforms/"+platformID+"/models", token, payload, http.StatusCreated, &response)
+	if response.ID == "" {
+		t.Fatal("acceptance model returned no id")
+	}
+	return response.ID
+}
+
+func submitAsyncSimulationTasks(t *testing.T, baseURL, token, path, model string, count int, duration time.Duration) []string {
+	t.Helper()
+	ids := make([]string, count)
+	errs := make(chan error, count)
+	var wg sync.WaitGroup
+	for index := 0; index < count; index++ {
+		wg.Add(1)
+		go func(index int) {
+			defer wg.Done()
+			payload := map[string]any{
+				"model":                model,
+				"runMode":              "simulation",
+				"simulation":           true,
+				"simulationDurationMs": duration.Milliseconds(),
+				"prompt":               fmt.Sprintf("async acceptance %d", index),
+				"input":                fmt.Sprintf("async acceptance %d", index),
+			}
+			raw, _ := json.Marshal(payload)
+			req, err := http.NewRequest(http.MethodPost, baseURL+path, bytes.NewReader(raw))
+			if err != nil {
+				errs <- err
+				return
+			}
+			req.Header.Set("Content-Type", "application/json")
+			req.Header.Set("Authorization", "Bearer "+token)
+			req.Header.Set("X-Async", "true")
+			resp, err := http.DefaultClient.Do(req)
+			if err != nil {
+				errs <- err
+				return
+			}
+			defer resp.Body.Close()
+			responseBody, _ := io.ReadAll(resp.Body)
+			if resp.StatusCode != http.StatusAccepted {
+				errs <- fmt.Errorf("submit task %d status=%d body=%s", index, resp.StatusCode, responseBody)
+				return
+			}
+			var response struct {
+				TaskID string `json:"taskId"`
+			}
+			if err := json.Unmarshal(responseBody, &response); err != nil || response.TaskID == "" {
+				errs <- fmt.Errorf("decode task %d: %w body=%s", index, err, responseBody)
+				return
+			}
+			ids[index] = response.TaskID
+		}(index)
+	}
+	wg.Wait()
+	close(errs)
+	for err := range errs {
+		t.Fatal(err)
+	}
+	return ids
+}
+
+func waitForAcceptanceTasks(t *testing.T, ctx context.Context, pool *pgxpool.Pool, taskIDs []string, modelID string, timeout time.Duration) (int64, int64) {
+	t.Helper()
+	deadline := time.Now().Add(timeout)
+	var peakRunning, peakLeases int64
+	for time.Now().Before(deadline) {
+		var running, succeeded, failed, leases int64
+		if err := pool.QueryRow(ctx, `
+SELECT COUNT(*) FILTER (WHERE status = 'running'),
+       COUNT(*) FILTER (WHERE status = 'succeeded'),
+       COUNT(*) FILTER (WHERE status IN ('failed', 'cancelled', 'manual_review'))
+FROM gateway_tasks
+WHERE id = ANY($1::uuid[])`, taskIDs).Scan(&running, &succeeded, &failed); err != nil {
+			t.Fatalf("read acceptance tasks: %v", err)
+		}
+		if err := pool.QueryRow(ctx, `
+SELECT COUNT(*)
+FROM gateway_concurrency_leases
+WHERE scope_type = 'platform_model'
+  AND scope_key = $1
+  AND released_at IS NULL
+  AND expires_at > now()`, modelID).Scan(&leases); err != nil {
+			t.Fatalf("read acceptance leases: %v", err)
+		}
+		peakRunning = maxInt64(peakRunning, running)
+		peakLeases = maxInt64(peakLeases, leases)
+		if failed > 0 {
+			t.Fatalf("acceptance tasks entered failed/manual status: %d", failed)
+		}
+		if succeeded == int64(len(taskIDs)) {
+			return peakRunning, peakLeases
+		}
+		time.Sleep(50 * time.Millisecond)
+	}
+	t.Fatalf("tasks did not finish within %s", timeout)
+	return 0, 0
+}
+
+func waitForTaskCount(t *testing.T, ctx context.Context, pool *pgxpool.Pool, taskIDs []string, status string, count int, timeout time.Duration) {
+	t.Helper()
+	deadline := time.Now().Add(timeout)
+	for time.Now().Before(deadline) {
+		var got int
+		if err := pool.QueryRow(ctx, `SELECT COUNT(*) FROM gateway_tasks WHERE id = ANY($1::uuid[]) AND status = $2`, taskIDs, status).Scan(&got); err != nil {
+			t.Fatalf("read task count: %v", err)
+		}
+		if got == count {
+			return
+		}
+		time.Sleep(100 * time.Millisecond)
+	}
+	t.Fatalf("task status %s count did not reach %d within %s", status, count, timeout)
+}
+
+func activeLeaseExpiry(t *testing.T, ctx context.Context, pool *pgxpool.Pool, modelID string, wantCount int) time.Time {
+	t.Helper()
+	var count int
+	var earliest time.Time
+	if err := pool.QueryRow(ctx, `
+SELECT COUNT(*), MIN(expires_at)
+FROM gateway_concurrency_leases
+WHERE scope_type = 'platform_model'
+  AND scope_key = $1
+  AND released_at IS NULL
+  AND expires_at > now()`, modelID).Scan(&count, &earliest); err != nil {
+		t.Fatalf("read active lease expiry: %v", err)
+	}
+	if count != wantCount {
+		t.Fatalf("active lease count=%d, want=%d", count, wantCount)
+	}
+	return earliest
+}
+
+func waitForActiveLeaseCount(t *testing.T, ctx context.Context, pool *pgxpool.Pool, modelID string, count int, timeout time.Duration) {
+	t.Helper()
+	deadline := time.Now().Add(timeout)
+	for time.Now().Before(deadline) {
+		var got int
+		if err := pool.QueryRow(ctx, `
+SELECT COUNT(*)
+FROM gateway_concurrency_leases
+WHERE scope_type = 'platform_model'
+  AND scope_key = $1
+  AND released_at IS NULL
+  AND expires_at > now()`, modelID).Scan(&got); err != nil {
+			t.Fatalf("read active lease count: %v", err)
+		}
+		if got == count {
+			return
+		}
+		time.Sleep(100 * time.Millisecond)
+	}
+	t.Fatalf("active lease count did not reach %d within %s", count, timeout)
+}
+
+func taskAttemptCount(t *testing.T, ctx context.Context, pool *pgxpool.Pool, taskIDs []string) int {
+	t.Helper()
+	var count int
+	if err := pool.QueryRow(ctx, `SELECT COUNT(*) FROM gateway_task_attempts WHERE task_id = ANY($1::uuid[])`, taskIDs).Scan(&count); err != nil {
+		t.Fatalf("count task attempts: %v", err)
+	}
+	return count
+}
+
+func waitForTaskAttempts(t *testing.T, ctx context.Context, pool *pgxpool.Pool, taskIDs []string, count int, timeout time.Duration) {
+	t.Helper()
+	deadline := time.Now().Add(timeout)
+	for time.Now().Before(deadline) {
+		if taskAttemptCount(t, ctx, pool, taskIDs) == count {
+			return
+		}
+		time.Sleep(50 * time.Millisecond)
+	}
+	t.Fatalf("attempt count did not reach %d within %s", count, timeout)
+}
+
+func assertAcceptanceAttempts(t *testing.T, ctx context.Context, pool *pgxpool.Pool, taskIDs []string, want int) {
+	t.Helper()
+	var attempts, duplicateTasks, manualReview int
+	if err := pool.QueryRow(ctx, `
+SELECT COUNT(*),
+       COUNT(DISTINCT task_id) FILTER (WHERE per_task.attempts > 1),
+       COUNT(*) FILTER (WHERE status = 'manual_review')
+FROM (
+  SELECT task_id, COUNT(*) AS attempts, MAX(status) AS status
+  FROM gateway_task_attempts
+  WHERE task_id = ANY($1::uuid[])
+  GROUP BY task_id
+) per_task`, taskIDs).Scan(&attempts, &duplicateTasks, &manualReview); err != nil {
+		t.Fatalf("read attempt acceptance evidence: %v", err)
+	}
+	if attempts != want || duplicateTasks != 0 || manualReview != 0 {
+		t.Fatalf("attempts=%d duplicate_tasks=%d manual_review=%d, want %d/0/0", attempts, duplicateTasks, manualReview, want)
+	}
+}
+
+func assertSimulationOutputsAndEvents(t *testing.T, ctx context.Context, pool *pgxpool.Pool, taskIDs []string) {
+	t.Helper()
+	var complete int
+	if err := pool.QueryRow(ctx, `
+SELECT COUNT(*)
+FROM gateway_tasks task
+WHERE task.id = ANY($1::uuid[])
+  AND task.status = 'succeeded'
+  AND task.finished_at IS NOT NULL
+  AND task.result IS NOT NULL
+  AND task.result <> '{}'::jsonb
+  AND task.metrics IS NOT NULL
+  AND jsonb_typeof(task.billings) = 'array'
+  AND jsonb_array_length(task.billings) > 0
+  AND EXISTS (SELECT 1 FROM gateway_task_events event WHERE event.task_id = task.id AND event.event_type = 'task.completed')`,
+		taskIDs).Scan(&complete); err != nil {
+		t.Fatalf("read simulation output evidence: %v", err)
+	}
+	if complete != len(taskIDs) {
+		t.Fatalf("complete simulation outputs/events=%d, want=%d", complete, len(taskIDs))
+	}
+}
+
+func waitForAsyncWorkerMetric(t *testing.T, baseURL, metric string, want int64, timeout time.Duration) {
+	t.Helper()
+	deadline := time.Now().Add(timeout)
+	for time.Now().Before(deadline) {
+		resp, err := http.Get(baseURL + "/metrics")
+		if err == nil {
+			raw, _ := io.ReadAll(resp.Body)
+			resp.Body.Close()
+			if resp.StatusCode == http.StatusOK {
+				for _, line := range strings.Split(string(raw), "\n") {
+					fields := strings.Fields(line)
+					if len(fields) == 2 && fields[0] == metric {
+						value, _ := strconv.ParseInt(fields[1], 10, 64)
+						if value == want {
+							return
+						}
+					}
+				}
+			}
+		}
+		time.Sleep(100 * time.Millisecond)
+	}
+	t.Fatalf("metric %s did not reach %d within %s", metric, want, timeout)
+}
+
+func maxInt64(left, right int64) int64 {
+	if right > left {
+		return right
+	}
+	return left
+}
@@ -487,18 +487,20 @@ func discountTitle(label string, discount float64) string {
 }

 func effectiveModelRateLimits(model store.PlatformModel, platform store.Platform, runtimePolicyMap map[string]store.RuntimePolicySet) ModelCatalogRateLimits {
-	overridePolicyRaw, hasOverridePolicy := model.RuntimePolicyOverride["rateLimitPolicy"]
-	overridePolicy := objectValue(overridePolicyRaw)
 	runtimePolicy := map[string]any(nil)
 	if model.RuntimePolicySetID != "" {
 		runtimePolicy = runtimePolicyMap[model.RuntimePolicySetID].RateLimitPolicy
 	}
-	policies := []rateLimitPolicySource{
-		{policy: overridePolicy, authoritative: hasOverridePolicy},
-		{policy: model.RateLimitPolicy, authoritative: len(model.RateLimitPolicy) > 0},
-		{policy: runtimePolicy, authoritative: strings.TrimSpace(model.RuntimePolicySetID) != ""},
-		{policy: platform.RateLimitPolicy},
-	}
+	policy := store.EffectiveRateLimitPolicy(store.EffectiveRateLimitPolicyInput{
+		BasePolicy:            model.BaseRateLimitPolicy,
+		PlatformPolicy:        platform.RateLimitPolicy,
+		RuntimePolicy:         runtimePolicy,
+		RuntimePolicyExplicit: model.RuntimePolicySetID != "",
+		RuntimePolicyOverride: model.RuntimePolicyOverride,
+		ModelPolicy:           model.RateLimitPolicy,
+		ModelPolicyMode:       model.RateLimitPolicyMode,
+	})
+	policies := []rateLimitPolicySource{{policy: policy, authoritative: policy != nil}}
 	limits := ModelCatalogRateLimits{
 		RPM:        firstRateLimit(policies, "rpm"),
 		TPM:        firstRateLimit(policies, "tpm_total"),
@@ -74,7 +74,7 @@ func (s *Service) rateLimitReservations(ctx context.Context, user *auth.User, ca
 				"groupKey": group.GroupKey,
 				"name":     group.Name,
 			},
-			group.RateLimitPolicy,
+			store.NormalizeRateLimitPolicy(group.RateLimitPolicy),
 			body,
 		)...)
 	}
@@ -82,27 +82,15 @@ func (s *Service) rateLimitReservations(ctx context.Context, user *auth.User, ca
 }

 func effectiveRateLimitPolicy(candidate store.RuntimeModelCandidate) map[string]any {
-	policy := candidate.PlatformRateLimitPolicy
-	if strings.TrimSpace(candidate.RuntimePolicySetID) != "" {
-		policy = candidate.RuntimeRateLimitPolicy
-	} else if hasRules(candidate.RuntimeRateLimitPolicy) {
-		policy = mergeMap(policy, candidate.RuntimeRateLimitPolicy)
-	}
-	if _, hasOverride := candidate.RuntimePolicyOverride["rateLimitPolicy"]; hasOverride {
-		nested, _ := candidate.RuntimePolicyOverride["rateLimitPolicy"].(map[string]any)
-		if len(nested) == 0 {
-			policy = nil
-		} else {
-			policy = mergeMap(policy, nested)
-		}
-	}
-	if hasRules(candidate.ModelRateLimitPolicy) {
-		policy = mergeMap(policy, candidate.ModelRateLimitPolicy)
-	}
-	if hasRules(policy) {
-		return policy
-	}
-	return nil
+	return store.EffectiveRateLimitPolicy(store.EffectiveRateLimitPolicyInput{
+		BasePolicy:            candidate.BaseRateLimitPolicy,
+		PlatformPolicy:        candidate.PlatformRateLimitPolicy,
+		RuntimePolicy:         candidate.RuntimeRateLimitPolicy,
+		RuntimePolicyExplicit: candidate.RuntimePolicyExplicit,
+		RuntimePolicyOverride: candidate.RateLimitRuntimeOverride,
+		ModelPolicy:           candidate.ModelRateLimitPolicy,
+		ModelPolicyMode:       candidate.ModelRateLimitPolicyMode,
+	})
 }

 func effectiveRetryPolicy(candidate store.RuntimeModelCandidate) map[string]any {
@@ -102,8 +102,10 @@ func TestEffectiveRateLimitPolicyTreatsEmptyRuntimePolicyAsUnlimited(t *testing.
 		PlatformRateLimitPolicy: map[string]any{"rules": []any{
 			map[string]any{"metric": "rpm", "limit": 500},
 		}},
-		RuntimePolicySetID:     "runtime-policy-1",
-		RuntimeRateLimitPolicy: map[string]any{"rules": []any{}},
+		RuntimePolicySetID:       "runtime-policy-1",
+		RuntimePolicyExplicit:    true,
+		RuntimeRateLimitPolicy:   map[string]any{"rules": []any{}},
+		ModelRateLimitPolicyMode: "inherit",
 	})

 	if hasRules(policy) {
@@ -11,6 +11,7 @@ import (
 	"github.com/easyai/easyai-ai-gateway/apps/api/internal/auth"
 	"github.com/easyai/easyai-ai-gateway/apps/api/internal/store"
 	"github.com/google/uuid"
+	"github.com/jackc/pgx/v5"
 	"github.com/riverqueue/river"
 	"github.com/riverqueue/river/riverdriver/riverpgxv5"
 	"github.com/riverqueue/river/rivermigrate"
@@ -23,6 +24,12 @@ type asyncTaskArgs struct {
 	TaskID string `json:"task_id" river:"unique"`
 }

+type asyncExecutionClient interface {
+	Start(context.Context) error
+	Stop(context.Context) error
+	StopAndCancel(context.Context) error
+}
+
 func (asyncTaskArgs) Kind() string { return "gateway_task_run" }

 type asyncTaskWorker struct {
@@ -87,21 +94,70 @@ func (s *Service) startRiverQueue(ctx context.Context) error {
 		return err
 	}

-	workers := river.NewWorkers()
-	if err := river.AddWorkerSafely(workers, &asyncTaskWorker{service: s}); err != nil {
+	controlClient, err := river.NewClient(driver, &river.Config{
+		ID:       asyncWorkerID() + "-control",
+		Logger:   s.logger,
+		TestOnly: s.cfg.AppEnv == "test",
+	})
+	if err != nil {
 		return err
 	}
-	riverClient, err := river.NewClient(driver, &river.Config{
-		ID:                          asyncWorkerID(),
+	snapshot, err := s.loadAsyncWorkerCapacity(ctx)
+	if err != nil {
+		return fmt.Errorf("calculate initial async worker capacity: %w", err)
+	}
+	executionClient, err := s.makeAsyncExecutionClient(snapshot.Capacity)
+	if err != nil {
+		return err
+	}
+	if err := executionClient.Start(ctx); err != nil {
+		cleanupCtx, cancel := context.WithTimeout(context.Background(), 10*time.Second)
+		_ = executionClient.StopAndCancel(cleanupCtx)
+		cancel()
+		return err
+	}
+	s.riverMu.Lock()
+	s.riverControlClient = controlClient
+	s.riverExecutionClient = executionClient
+	s.riverWorkerCapacity = snapshot.Capacity
+	s.riverDrainingClients = make(map[asyncExecutionClient]struct{})
+	s.riverMu.Unlock()
+	s.observeAsyncWorkerCapacity(snapshot)
+	s.logger.Info("async worker capacity initialized",
+		"capacity", snapshot.Capacity,
+		"desiredCapacity", snapshot.Desired,
+		"hardLimit", snapshot.HardLimit,
+		"enabledModels", snapshot.EnabledModels,
+		"unlimitedModels", snapshot.UnlimitedModels,
+		"modelDesired", snapshot.ModelDesired,
+		"enabledGroups", snapshot.EnabledGroups,
+		"unlimitedGroups", snapshot.UnlimitedGroups,
+		"groupDesired", snapshot.GroupDesired,
+		"capped", snapshot.Capped,
+	)
+	if err := s.recoverAsyncRiverJobs(ctx); err != nil {
+		cleanupCtx, cancel := context.WithTimeout(context.Background(), 10*time.Second)
+		_ = executionClient.StopAndCancel(cleanupCtx)
+		cancel()
+		return err
+	}
+	go s.refreshAsyncWorkerCapacity(ctx)
+	go s.stopAsyncWorkersOnShutdown(ctx)
+	return nil
+}
+
+func (s *Service) newRiverAsyncExecutionClient(capacity int) (*river.Client[pgx.Tx], error) {
+	workers := river.NewWorkers()
+	if err := river.AddWorkerSafely(workers, &asyncTaskWorker{service: s}); err != nil {
+		return nil, err
+	}
+	return river.NewClient(riverpgxv5.New(s.store.Pool()), &river.Config{
+		ID:                          fmt.Sprintf("%s-exec-%d-%d", asyncWorkerID(), capacity, time.Now().UnixNano()),
 		JobTimeout:                  -1,
 		Logger:                      s.logger,
 		CompletedJobRetentionPeriod: 24 * time.Hour,
 		Queues: map[string]river.QueueConfig{
-			// Image providers may hold a worker while polling for several
-			// minutes. Keep enough workers available for production bursts so
-			// unrelated models do not remain queued behind long-running media
-			// tasks.
-			asyncTaskQueueName: {MaxWorkers: 96},
+			asyncTaskQueueName: {MaxWorkers: capacity},
 		},
 		// Provider-backed media jobs commonly poll for 10-20 minutes. River may
 		// execute a still-running job again once this window elapses, so keep the
@@ -110,32 +166,160 @@ func (s *Service) startRiverQueue(ctx context.Context) error {
 		TestOnly:             s.cfg.AppEnv == "test",
 		Workers:              workers,
 	})
+}
+
+func (s *Service) makeAsyncExecutionClient(capacity int) (asyncExecutionClient, error) {
+	if s.asyncClientFactory != nil {
+		return s.asyncClientFactory(capacity)
+	}
+	return s.newRiverAsyncExecutionClient(capacity)
+}
+
+func (s *Service) loadAsyncWorkerCapacity(ctx context.Context) (store.AsyncWorkerCapacitySnapshot, error) {
+	if s.asyncCapacityLoader != nil {
+		return s.asyncCapacityLoader(ctx, s.cfg.AsyncWorkerHardLimit)
+	}
+	return s.store.AsyncWorkerCapacity(ctx, s.cfg.AsyncWorkerHardLimit)
+}
+
+func (s *Service) refreshAsyncWorkerCapacity(ctx context.Context) {
+	ticker := time.NewTicker(time.Duration(s.cfg.AsyncWorkerRefreshIntervalSeconds) * time.Second)
+	defer ticker.Stop()
+	for {
+		select {
+		case <-ctx.Done():
+			return
+		case <-ticker.C:
+			s.resizeAsyncWorkerCapacity(ctx)
+		}
+	}
+}
+
+func (s *Service) resizeAsyncWorkerCapacity(ctx context.Context) {
+	snapshot, err := s.loadAsyncWorkerCapacity(ctx)
 	if err != nil {
-		return err
+		s.observeAsyncWorkerResize("refresh_failed")
+		s.logger.Warn("refresh async worker capacity failed; keeping current client", "error", err)
+		return
 	}
-	s.riverClient = riverClient
-	if err := riverClient.Start(ctx); err != nil {
-		return err
+	s.riverMu.RLock()
+	currentCapacity := s.riverWorkerCapacity
+	s.riverMu.RUnlock()
+	s.observeAsyncWorkerCapacity(snapshot)
+	if snapshot.Capacity == currentCapacity {
+		return
 	}
-	if err := s.recoverAsyncRiverJobs(ctx); err != nil {
-		return err
+	newClient, err := s.makeAsyncExecutionClient(snapshot.Capacity)
+	if err != nil {
+		s.observeAsyncWorkerResize("create_failed")
+		s.logger.Warn("create replacement async worker client failed; keeping current client",
+			"error", err, "currentCapacity", currentCapacity, "desiredCapacity", snapshot.Capacity)
+		return
 	}
-	go func() {
-		<-ctx.Done()
+	if err := newClient.Start(ctx); err != nil {
+		cleanupCtx, cancel := context.WithTimeout(context.Background(), 10*time.Second)
+		_ = newClient.StopAndCancel(cleanupCtx)
+		cancel()
+		s.observeAsyncWorkerResize("start_failed")
+		s.logger.Warn("start replacement async worker client failed; keeping current client",
+			"error", err, "currentCapacity", currentCapacity, "desiredCapacity", snapshot.Capacity)
+		return
+	}
+	if ctx.Err() != nil {
+		cleanupCtx, cancel := context.WithTimeout(context.Background(), 10*time.Second)
+		_ = newClient.StopAndCancel(cleanupCtx)
+		cancel()
+		return
+	}
+	s.riverMu.Lock()
+	oldClient := s.riverExecutionClient
+	s.riverExecutionClient = newClient
+	s.riverWorkerCapacity = snapshot.Capacity
+	if oldClient != nil {
+		if s.riverDrainingClients == nil {
+			s.riverDrainingClients = make(map[asyncExecutionClient]struct{})
+		}
+		s.riverDrainingClients[oldClient] = struct{}{}
+	}
+	s.riverMu.Unlock()
+	s.observeAsyncWorkerResize("success")
+	s.logger.Info("async worker capacity resized",
+		"previousCapacity", currentCapacity,
+		"capacity", snapshot.Capacity,
+		"desiredCapacity", snapshot.Desired,
+		"hardLimit", snapshot.HardLimit,
+		"modelDesired", snapshot.ModelDesired,
+		"groupDesired", snapshot.GroupDesired,
+		"capped", snapshot.Capped,
+	)
+	if oldClient != nil {
+		go s.drainAsyncWorkerClient(oldClient)
+	}
+}
+
+func (s *Service) drainAsyncWorkerClient(client asyncExecutionClient) {
+	stopCtx, cancel := context.WithTimeout(context.Background(), time.Hour)
+	defer cancel()
+	if err := client.Stop(stopCtx); err != nil {
+		s.logger.Warn("gracefully drain previous async worker client failed", "error", err)
+	}
+	s.riverMu.Lock()
+	delete(s.riverDrainingClients, client)
+	s.riverMu.Unlock()
+}
+
+func (s *Service) stopAsyncWorkersOnShutdown(ctx context.Context) {
+	<-ctx.Done()
+	s.riverMu.Lock()
+	clients := make([]asyncExecutionClient, 0, 1+len(s.riverDrainingClients))
+	if s.riverExecutionClient != nil {
+		clients = append(clients, s.riverExecutionClient)
+	}
+	for client := range s.riverDrainingClients {
+		clients = append(clients, client)
+	}
+	s.riverExecutionClient = nil
+	s.riverDrainingClients = nil
+	s.riverMu.Unlock()
+	for _, client := range clients {
 		stopCtx, cancel := context.WithTimeout(context.Background(), 10*time.Second)
-		defer cancel()
-		if err := riverClient.StopAndCancel(stopCtx); err != nil {
+		if err := client.StopAndCancel(stopCtx); err != nil {
 			s.logger.Warn("stop river async queue failed", "error", err)
 		}
-	}()
-	return nil
+		cancel()
+	}
+}
+
+func (s *Service) asyncControlClient() *river.Client[pgx.Tx] {
+	s.riverMu.RLock()
+	defer s.riverMu.RUnlock()
+	return s.riverControlClient
+}
+
+func (s *Service) observeAsyncWorkerCapacity(snapshot store.AsyncWorkerCapacitySnapshot) {
+	observer, ok := s.billingMetrics.(interface {
+		SetAsyncWorkerCapacity(current, desired, hardLimit int, capped bool)
+	})
+	if ok {
+		observer.SetAsyncWorkerCapacity(snapshot.Capacity, snapshot.Desired, snapshot.HardLimit, snapshot.Capped)
+	}
+}
+
+func (s *Service) observeAsyncWorkerResize(outcome string) {
+	observer, ok := s.billingMetrics.(interface {
+		ObserveAsyncWorkerResize(string)
+	})
+	if ok {
+		observer.ObserveAsyncWorkerResize(outcome)
+	}
 }

 func (s *Service) EnqueueAsyncTask(ctx context.Context, task store.GatewayTask) error {
-	if s.riverClient == nil {
+	riverClient := s.asyncControlClient()
+	if riverClient == nil {
 		return errors.New("river async queue is not started")
 	}
-	result, err := s.riverClient.Insert(ctx, asyncTaskArgs{TaskID: task.ID}, asyncTaskInsertOpts(task))
+	result, err := riverClient.Insert(ctx, asyncTaskArgs{TaskID: task.ID}, asyncTaskInsertOpts(task))
 	if err != nil {
 		return err
 	}
@@ -155,12 +339,16 @@ func (s *Service) RunAsyncTask(ctx context.Context, task store.GatewayTask, user
 }

 func (s *Service) recoverAsyncRiverJobs(ctx context.Context) error {
+	riverClient := s.asyncControlClient()
+	if riverClient == nil {
+		return errors.New("river async queue is not started")
+	}
 	items, err := s.store.ListRecoverableAsyncTasks(ctx, 1000)
 	if err != nil {
 		return err
 	}
 	for _, item := range items {
-		result, err := s.riverClient.Insert(ctx, asyncTaskArgs{TaskID: item.ID}, asyncTaskRecoveryInsertOpts(item, time.Now()))
+		result, err := riverClient.Insert(ctx, asyncTaskArgs{TaskID: item.ID}, asyncTaskRecoveryInsertOpts(item, time.Now()))
 		if err != nil {
 			return err
 		}
@@ -0,0 +1,189 @@
+package runner
+
+import (
+	"context"
+	"errors"
+	"io"
+	"log/slog"
+	"sync/atomic"
+	"testing"
+	"time"
+
+	"github.com/easyai/easyai-ai-gateway/apps/api/internal/config"
+	"github.com/easyai/easyai-ai-gateway/apps/api/internal/store"
+)
+
+type fakeAsyncExecutionClient struct {
+	startErr      error
+	stopGate      <-chan struct{}
+	started       atomic.Int64
+	stopped       atomic.Int64
+	stopCancelled atomic.Int64
+}
+
+func (c *fakeAsyncExecutionClient) Start(context.Context) error {
+	c.started.Add(1)
+	return c.startErr
+}
+
+func (c *fakeAsyncExecutionClient) Stop(ctx context.Context) error {
+	c.stopped.Add(1)
+	if c.stopGate == nil {
+		return nil
+	}
+	select {
+	case <-c.stopGate:
+		return nil
+	case <-ctx.Done():
+		return ctx.Err()
+	}
+}
+
+func (c *fakeAsyncExecutionClient) StopAndCancel(context.Context) error {
+	c.stopCancelled.Add(1)
+	return nil
+}
+
+func TestResizeAsyncWorkerCapacityStartsReplacementBeforeGracefulDrain(t *testing.T) {
+	oldClient := &fakeAsyncExecutionClient{}
+	drainGate := make(chan struct{})
+	oldClient.stopGate = drainGate
+	newClient := &fakeAsyncExecutionClient{}
+	service := asyncWorkerManagerTestService(1)
+	service.riverExecutionClient = oldClient
+	service.asyncCapacityLoader = fixedAsyncCapacity(3)
+	service.asyncClientFactory = func(capacity int) (asyncExecutionClient, error) {
+		if capacity != 3 {
+			t.Fatalf("factory capacity=%d, want=3", capacity)
+		}
+		return newClient, nil
+	}
+
+	service.resizeAsyncWorkerCapacity(context.Background())
+	if newClient.started.Load() != 1 {
+		t.Fatalf("replacement start count=%d, want=1", newClient.started.Load())
+	}
+	if service.riverExecutionClient != newClient || service.riverWorkerCapacity != 3 {
+		t.Fatalf("replacement was not installed: capacity=%d client=%T", service.riverWorkerCapacity, service.riverExecutionClient)
+	}
+	waitForAtomicValue(t, &oldClient.stopped, 1)
+	if oldClient.stopCancelled.Load() != 0 {
+		t.Fatal("graceful drain cancelled an already running old task")
+	}
+	close(drainGate)
+	waitForDrainingClients(t, service, 0)
+}
+
+func TestResizeAsyncWorkerCapacityFailuresKeepCurrentClient(t *testing.T) {
+	t.Run("refresh failure", func(t *testing.T) {
+		oldClient := &fakeAsyncExecutionClient{}
+		service := asyncWorkerManagerTestService(4)
+		service.riverExecutionClient = oldClient
+		service.asyncCapacityLoader = func(context.Context, int) (store.AsyncWorkerCapacitySnapshot, error) {
+			return store.AsyncWorkerCapacitySnapshot{}, errors.New("database unavailable")
+		}
+		service.asyncClientFactory = func(int) (asyncExecutionClient, error) {
+			t.Fatal("factory must not run after refresh failure")
+			return nil, nil
+		}
+		service.resizeAsyncWorkerCapacity(context.Background())
+		if service.riverExecutionClient != oldClient || service.riverWorkerCapacity != 4 {
+			t.Fatal("refresh failure replaced the current client")
+		}
+	})
+
+	t.Run("create failure", func(t *testing.T) {
+		oldClient := &fakeAsyncExecutionClient{}
+		service := asyncWorkerManagerTestService(4)
+		service.riverExecutionClient = oldClient
+		service.asyncCapacityLoader = fixedAsyncCapacity(8)
+		service.asyncClientFactory = func(int) (asyncExecutionClient, error) {
+			return nil, errors.New("factory failed")
+		}
+		service.resizeAsyncWorkerCapacity(context.Background())
+		if service.riverExecutionClient != oldClient || service.riverWorkerCapacity != 4 {
+			t.Fatal("create failure replaced the current client")
+		}
+	})
+
+	t.Run("start failure", func(t *testing.T) {
+		oldClient := &fakeAsyncExecutionClient{}
+		failedClient := &fakeAsyncExecutionClient{startErr: errors.New("start failed")}
+		service := asyncWorkerManagerTestService(4)
+		service.riverExecutionClient = oldClient
+		service.asyncCapacityLoader = fixedAsyncCapacity(8)
+		service.asyncClientFactory = func(int) (asyncExecutionClient, error) {
+			return failedClient, nil
+		}
+		service.resizeAsyncWorkerCapacity(context.Background())
+		if service.riverExecutionClient != oldClient || service.riverWorkerCapacity != 4 {
+			t.Fatal("start failure replaced the current client")
+		}
+		if failedClient.stopCancelled.Load() != 1 {
+			t.Fatal("failed replacement client was not cleaned up")
+		}
+	})
+}
+
+func TestStopAsyncWorkersIncludesCurrentAndDrainingClients(t *testing.T) {
+	current := &fakeAsyncExecutionClient{}
+	draining := &fakeAsyncExecutionClient{}
+	service := asyncWorkerManagerTestService(2)
+	service.riverExecutionClient = current
+	service.riverDrainingClients[draining] = struct{}{}
+	ctx, cancel := context.WithCancel(context.Background())
+	cancel()
+	service.stopAsyncWorkersOnShutdown(ctx)
+	if current.stopCancelled.Load() != 1 || draining.stopCancelled.Load() != 1 {
+		t.Fatalf("shutdown cancellations current=%d draining=%d, want 1/1", current.stopCancelled.Load(), draining.stopCancelled.Load())
+	}
+}
+
+func asyncWorkerManagerTestService(capacity int) *Service {
+	return &Service{
+		cfg: config.Config{
+			AsyncWorkerHardLimit:              2048,
+			AsyncWorkerRefreshIntervalSeconds: 5,
+		},
+		logger:               slog.New(slog.NewTextHandler(io.Discard, nil)),
+		riverDrainingClients: make(map[asyncExecutionClient]struct{}),
+		riverWorkerCapacity:  capacity,
+		riverExecutionClient: nil,
+	}
+}
+
+func fixedAsyncCapacity(capacity int) func(context.Context, int) (store.AsyncWorkerCapacitySnapshot, error) {
+	return func(context.Context, int) (store.AsyncWorkerCapacitySnapshot, error) {
+		return store.AsyncWorkerCapacitySnapshot{Capacity: capacity, Desired: capacity, HardLimit: 2048}, nil
+	}
+}
+
+func waitForAtomicValue(t *testing.T, value *atomic.Int64, want int64) {
+	t.Helper()
+	deadline := time.Now().Add(time.Second)
+	for time.Now().Before(deadline) {
+		if value.Load() == want {
+			return
+		}
+		time.Sleep(time.Millisecond)
+	}
+	t.Fatalf("atomic value=%d, want=%d", value.Load(), want)
+}
+
+func waitForDrainingClients(t *testing.T, service *Service, want int) {
+	t.Helper()
+	deadline := time.Now().Add(time.Second)
+	for time.Now().Before(deadline) {
+		service.riverMu.RLock()
+		count := len(service.riverDrainingClients)
+		service.riverMu.RUnlock()
+		if count == want {
+			return
+		}
+		time.Sleep(time.Millisecond)
+	}
+	service.riverMu.RLock()
+	count := len(service.riverDrainingClients)
+	service.riverMu.RUnlock()
+	t.Fatalf("draining client count=%d, want=%d", count, want)
+}
@@ -10,6 +10,7 @@ import (
 	"regexp"
 	"strconv"
 	"strings"
+	"sync"
 	"time"

 	"github.com/easyai/easyai-ai-gateway/apps/api/internal/auth"
@@ -23,14 +24,20 @@ import (
 )

 type Service struct {
-	cfg            config.Config
-	store          *store.Store
-	logger         *slog.Logger
-	clients        map[string]clients.Client
-	scriptExecutor *scriptengine.Executor
-	httpClients    *httpClientCache
-	riverClient    *river.Client[pgx.Tx]
-	billingMetrics billingMetricsObserver
+	cfg                  config.Config
+	store                *store.Store
+	logger               *slog.Logger
+	clients              map[string]clients.Client
+	scriptExecutor       *scriptengine.Executor
+	httpClients          *httpClientCache
+	riverMu              sync.RWMutex
+	riverControlClient   *river.Client[pgx.Tx]
+	riverExecutionClient asyncExecutionClient
+	riverDrainingClients map[asyncExecutionClient]struct{}
+	riverWorkerCapacity  int
+	asyncCapacityLoader  func(context.Context, int) (store.AsyncWorkerCapacitySnapshot, error)
+	asyncClientFactory   func(int) (asyncExecutionClient, error)
+	billingMetrics       billingMetricsObserver
 }

 type billingMetricsObserver interface {
@@ -75,6 +82,12 @@ func (e *TaskQueuedError) Is(target error) bool {
 }

 func New(cfg config.Config, db *store.Store, logger *slog.Logger, observers ...billingMetricsObserver) *Service {
+	if cfg.AsyncWorkerHardLimit == 0 {
+		cfg.AsyncWorkerHardLimit = 2048
+	}
+	if cfg.AsyncWorkerRefreshIntervalSeconds == 0 {
+		cfg.AsyncWorkerRefreshIntervalSeconds = 5
+	}
 	httpClients := newHTTPClientCache()
 	scriptExecutor := &scriptengine.Executor{Logger: logger}
 	service := &Service{
@@ -891,7 +904,7 @@ func (s *Service) runCandidate(ctx context.Context, task store.GatewayTask, user
 			_ = s.store.ReleaseRateLimitReservations(context.WithoutCancel(ctx), limitResult.Reservations, "attempt_failed")
 		}
 	}()
-	defer s.store.ReleaseConcurrencyLeases(context.WithoutCancel(ctx), limitResult.LeaseIDs)
+	defer s.store.ReleaseConcurrencyLeases(context.WithoutCancel(ctx), limitResult.Leases)

 	attemptID, err := s.store.CreateTaskAttempt(ctx, store.CreateTaskAttemptInput{
 		TaskID:             task.ID,
@@ -1031,7 +1044,8 @@ func (s *Service) runCandidate(ctx context.Context, task store.GatewayTask, user
 		return nil
 	}
 	var submissionWire *clients.WireResponse
-	response, err := client.Run(ctx, clients.Request{
+	runCtx, stopLeaseRenewal := s.startConcurrencyLeaseRenewal(ctx, task.ID, limitResult.Leases)
+	response, err := client.Run(runCtx, clients.Request{
 		Kind:              task.Kind,
 		ModelType:         candidate.ModelType,
 		Model:             task.Model,
@@ -1083,6 +1097,13 @@ func (s *Service) runCandidate(ctx context.Context, task store.GatewayTask, user
 		UpstreamPreviousResponseID: responseExecution.UpstreamPreviousResponseID,
 		PreviousResponseTurns:      responseExecution.PreviousTurns,
 	})
+	if leaseErr := stopLeaseRenewal(); leaseErr != nil {
+		err = &clients.ClientError{
+			Code:      "concurrency_lease_lost",
+			Message:   leaseErr.Error(),
+			Retryable: true,
+		}
+	}
 	callFinishedAt := time.Now()
 	if err == nil {
 		if markErr := setSubmissionStatus("response_received"); markErr != nil {
@@ -1500,6 +1521,7 @@ func (s *Service) requeueRateLimitedTask(ctx context.Context, task store.Gateway
 	if delay <= 0 {
 		delay = 5 * time.Second
 	}
+	delay += taskRetryJitter(task.ID)
 	queued, err := s.store.RequeueTask(ctx, task.ID, task.ExecutionToken, delay, candidate.QueueKey)
 	if err != nil {
 		return store.GatewayTask{}, 0, err
@@ -1515,6 +1537,81 @@ func (s *Service) requeueRateLimitedTask(ctx context.Context, task store.Gateway
 	return queued, delay, nil
 }

+func taskRetryJitter(taskID string) time.Duration {
+	var sum uint32
+	for _, value := range []byte(taskID) {
+		sum = sum*33 + uint32(value)
+	}
+	return time.Duration(sum%251) * time.Millisecond
+}
+
+func (s *Service) startConcurrencyLeaseRenewal(ctx context.Context, taskID string, leases []store.ConcurrencyLease) (context.Context, func() error) {
+	if len(leases) == 0 {
+		return ctx, func() error { return nil }
+	}
+	interval := 30 * time.Second
+	for _, lease := range leases {
+		ttl := lease.TTL
+		if ttl <= 0 {
+			ttl = 120 * time.Second
+		}
+		candidate := ttl / 3
+		if candidate < time.Second {
+			candidate = time.Second
+		}
+		if candidate < interval {
+			interval = candidate
+		}
+	}
+	runCtx, cancelRun := context.WithCancel(ctx)
+	renewCtx, cancelRenew := context.WithCancel(ctx)
+	done := make(chan error, 1)
+	go func() {
+		ticker := time.NewTicker(interval)
+		defer ticker.Stop()
+		for {
+			select {
+			case <-renewCtx.Done():
+				done <- nil
+				return
+			case <-ticker.C:
+				if err := s.store.RenewConcurrencyLeases(renewCtx, leases); err != nil {
+					if renewCtx.Err() != nil {
+						done <- nil
+						return
+					}
+					outcome := "failure"
+					if errors.Is(err, store.ErrConcurrencyLeaseLost) {
+						outcome = "lost"
+					}
+					s.observeConcurrencyLeaseRenewal(outcome)
+					s.logger.Error("concurrency lease renewal failed; cancelling upstream execution",
+						"taskID", taskID, "leaseCount", len(leases), "outcome", outcome, "error", err)
+					done <- err
+					cancelRun()
+					return
+				}
+				s.observeConcurrencyLeaseRenewal("success")
+			}
+		}
+	}()
+	return runCtx, func() error {
+		cancelRenew()
+		renewalErr := <-done
+		cancelRun()
+		return renewalErr
+	}
+}
+
+func (s *Service) observeConcurrencyLeaseRenewal(outcome string) {
+	observer, ok := s.billingMetrics.(interface {
+		ObserveConcurrencyLeaseRenewal(string)
+	})
+	if ok {
+		observer.ObserveConcurrencyLeaseRenewal(outcome)
+	}
+}
+
 func (s *Service) requeueInterruptedAsyncTask(ctx context.Context, task store.GatewayTask) (store.GatewayTask, error) {
 	queued, err := s.store.RequeueTask(ctx, task.ID, task.ExecutionToken, 0, "")
 	if err != nil {
@@ -58,10 +58,11 @@ func (s *Service) CancelTask(ctx context.Context, taskID string, user *auth.User
 		return taskCancelUnavailable(task, "任务已开始执行，当前阶段不可取消，请继续查询结果"), nil
 	}
 	if task.RiverJobID > 0 {
-		if s.riverClient == nil {
+		riverClient := s.asyncControlClient()
+		if riverClient == nil {
 			return taskCancelUnavailable(task, "任务取消队列未就绪，请继续查询结果"), nil
 		}
-		job, err := s.riverClient.JobGet(ctx, task.RiverJobID)
+		job, err := riverClient.JobGet(ctx, task.RiverJobID)
 		if errors.Is(err, rivertype.ErrNotFound) {
 			return taskCancelUnavailable(task, "任务已不在本地排队队列，可能已提交上游，当前不可取消，请继续查询结果"), nil
 		}
@@ -71,7 +72,7 @@ func (s *Service) CancelTask(ctx context.Context, taskID string, user *auth.User
 		if job == nil || !riverJobStateCancellable(job.State) {
 			return taskCancelUnavailable(task, "任务已不在可取消队列状态，请继续查询结果"), nil
 		}
-		if _, err := s.riverClient.JobDelete(ctx, task.RiverJobID); err != nil {
+		if _, err := riverClient.JobDelete(ctx, task.RiverJobID); err != nil {
 			if errors.Is(err, rivertype.ErrJobRunning) || errors.Is(err, rivertype.ErrNotFound) {
 				return taskCancelUnavailable(task, "任务已被工作进程领取，当前不可取消，请继续查询结果"), nil
 			}
@@ -44,6 +44,17 @@ type Metrics struct {
 	billingEstimateFailed      atomic.Uint64
 	billingIdempotentReplay    atomic.Uint64
 	billingPricingUnavailable  atomic.Uint64
+	asyncWorkerCapacity        atomic.Int64
+	asyncWorkerDesiredCapacity atomic.Int64
+	asyncWorkerHardLimit       atomic.Int64
+	asyncWorkerCapacityCapped  atomic.Int64
+	asyncWorkerResizeSuccess   atomic.Uint64
+	asyncWorkerRefreshFailed   atomic.Uint64
+	asyncWorkerCreateFailed    atomic.Uint64
+	asyncWorkerStartFailed     atomic.Uint64
+	leaseRenewalSuccess        atomic.Uint64
+	leaseRenewalFailure        atomic.Uint64
+	leaseRenewalLost           atomic.Uint64
 }

 var processingDurationBounds = [...]time.Duration{
@@ -123,6 +134,41 @@ func (m *Metrics) ObserveBillingEvent(event string) {
 	}
 }

+func (m *Metrics) SetAsyncWorkerCapacity(current, desired, hardLimit int, capped bool) {
+	m.asyncWorkerCapacity.Store(int64(current))
+	m.asyncWorkerDesiredCapacity.Store(int64(desired))
+	m.asyncWorkerHardLimit.Store(int64(hardLimit))
+	cappedValue := int64(0)
+	if capped {
+		cappedValue = 1
+	}
+	m.asyncWorkerCapacityCapped.Store(cappedValue)
+}
+
+func (m *Metrics) ObserveAsyncWorkerResize(outcome string) {
+	switch outcome {
+	case "success":
+		m.asyncWorkerResizeSuccess.Add(1)
+	case "refresh_failed":
+		m.asyncWorkerRefreshFailed.Add(1)
+	case "create_failed":
+		m.asyncWorkerCreateFailed.Add(1)
+	case "start_failed":
+		m.asyncWorkerStartFailed.Add(1)
+	}
+}
+
+func (m *Metrics) ObserveConcurrencyLeaseRenewal(outcome string) {
+	switch outcome {
+	case "success":
+		m.leaseRenewalSuccess.Add(1)
+	case "lost":
+		m.leaseRenewalLost.Add(1)
+	default:
+		m.leaseRenewalFailure.Add(1)
+	}
+}
+
 func (m *Metrics) Handler(provider MetricsSnapshotProvider, issuer, audience string, enabled bool) http.Handler {
 	return http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
 		mode := "disabled"
@@ -212,6 +258,21 @@ func (m *Metrics) Handler(provider MetricsSnapshotProvider, issuer, audience str
 		plainCounter(w, "easyai_gateway_billing_estimate_failures_total", "Pricing estimate requests that failed.", m.billingEstimateFailed.Load())
 		plainCounter(w, "easyai_gateway_billing_idempotent_replays_total", "Generation requests replayed idempotently.", m.billingIdempotentReplay.Load())
 		plainCounter(w, "easyai_gateway_billing_pricing_unavailable_total", "Pricing requests rejected because no effective price was available.", m.billingPricingUnavailable.Load())
+		plainGauge(w, "easyai_gateway_async_worker_capacity", "Current River asynchronous worker capacity.", m.asyncWorkerCapacity.Load())
+		plainGauge(w, "easyai_gateway_async_worker_desired_capacity", "Uncapped asynchronous worker capacity derived from model and user-group policies.", m.asyncWorkerDesiredCapacity.Load())
+		plainGauge(w, "easyai_gateway_async_worker_hard_limit", "Per-process asynchronous worker safety limit.", m.asyncWorkerHardLimit.Load())
+		plainGauge(w, "easyai_gateway_async_worker_capacity_capped", "Whether desired asynchronous worker capacity is capped by the hard limit.", m.asyncWorkerCapacityCapped.Load())
+		outcomeCounters(w, "easyai_gateway_async_worker_resizes_total", "Asynchronous worker resize attempts by bounded outcome.", []outcomeValue{
+			{"success", m.asyncWorkerResizeSuccess.Load()},
+			{"refresh_failed", m.asyncWorkerRefreshFailed.Load()},
+			{"create_failed", m.asyncWorkerCreateFailed.Load()},
+			{"start_failed", m.asyncWorkerStartFailed.Load()},
+		})
+		outcomeCounters(w, "easyai_gateway_concurrency_lease_renewals_total", "Concurrency lease renewals by bounded outcome.", []outcomeValue{
+			{"success", m.leaseRenewalSuccess.Load()},
+			{"failure", m.leaseRenewalFailure.Load()},
+			{"lost", m.leaseRenewalLost.Load()},
+		})
 	})
 }

@@ -249,3 +310,7 @@ func outcomeCounters(w http.ResponseWriter, name, help string, values []outcomeV
 func plainCounter(w http.ResponseWriter, name, help string, value uint64) {
 	fmt.Fprintf(w, "# HELP %s %s\n# TYPE %s counter\n%s %d\n", name, help, name, name, value)
 }
+
+func plainGauge(w http.ResponseWriter, name, help string, value int64) {
+	fmt.Fprintf(w, "# HELP %s %s\n# TYPE %s gauge\n%s %d\n", name, help, name, name, value)
+}
@@ -26,6 +26,10 @@ func TestMetricsExposeBoundedOutcomesAndState(t *testing.T) {
 	metrics.ObserveHeartbeat("accepted")
 	metrics.ObserveIntrospection("failed")
 	metrics.ObserveJWKSRefreshFailure("ssf")
+	metrics.SetAsyncWorkerCapacity(96, 128, 96, true)
+	metrics.ObserveAsyncWorkerResize("success")
+	metrics.ObserveConcurrencyLeaseRenewal("success")
+	metrics.ObserveConcurrencyLeaseRenewal("lost")

 	recorder := httptest.NewRecorder()
 	metrics.Handler(metricsSnapshot{mode: "push_healthy", last: time.Now().Add(-time.Minute)}, "issuer", "audience", true).
@@ -42,6 +46,12 @@ func TestMetricsExposeBoundedOutcomesAndState(t *testing.T) {
 		`easyai_gateway_ssf_mode{mode="push_healthy"} 1`,
 		`easyai_gateway_oidc_introspection_total{outcome="failed"} 1`,
 		`easyai_gateway_jwks_refresh_failures_total{outcome="ssf"} 1`,
+		`easyai_gateway_async_worker_capacity 96`,
+		`easyai_gateway_async_worker_desired_capacity 128`,
+		`easyai_gateway_async_worker_capacity_capped 1`,
+		`easyai_gateway_async_worker_resizes_total{outcome="success"} 1`,
+		`easyai_gateway_concurrency_lease_renewals_total{outcome="success"} 1`,
+		`easyai_gateway_concurrency_lease_renewals_total{outcome="lost"} 1`,
 	} {
 		if !strings.Contains(recorder.Body.String(), expected) {
 			t.Fatalf("missing metric %q in:\n%s", expected, recorder.Body.String())
@@ -0,0 +1,152 @@
+package store
+
+import (
+	"context"
+	"fmt"
+)
+
+type AsyncWorkerCapacitySnapshot struct {
+	Capacity        int
+	Desired         int
+	HardLimit       int
+	Capped          bool
+	EnabledModels   int
+	UnlimitedModels int
+	EnabledGroups   int
+	UnlimitedGroups int
+	ModelDesired    int
+	GroupDesired    int
+}
+
+func (s *Store) AsyncWorkerCapacity(ctx context.Context, hardLimit int) (AsyncWorkerCapacitySnapshot, error) {
+	if hardLimit < 1 {
+		return AsyncWorkerCapacitySnapshot{}, fmt.Errorf("async worker hard limit must be positive")
+	}
+	rows, err := s.pool.Query(ctx, `
+SELECT COALESCE(b.default_rate_limit_policy, '{}'::jsonb),
+       p.rate_limit_policy,
+       COALESCE(rp.rate_limit_policy, '{}'::jsonb),
+       (m.runtime_policy_set_id IS NOT NULL),
+       COALESCE(m.runtime_policy_override, '{}'::jsonb),
+       m.rate_limit_policy,
+       m.rate_limit_policy_mode
+FROM platform_models m
+JOIN integration_platforms p ON p.id = m.platform_id
+LEFT JOIN base_model_catalog b ON b.id = m.base_model_id
+LEFT JOIN model_runtime_policy_sets rp ON rp.id = COALESCE(m.runtime_policy_set_id, b.runtime_policy_set_id)
+WHERE p.status = 'enabled'
+  AND p.deleted_at IS NULL
+  AND m.enabled = true`)
+	if err != nil {
+		return AsyncWorkerCapacitySnapshot{}, err
+	}
+	defer rows.Close()
+
+	modelPolicies := make([]map[string]any, 0)
+	for rows.Next() {
+		var basePolicyBytes, platformPolicyBytes, runtimePolicyBytes []byte
+		var runtimeOverrideBytes, modelPolicyBytes []byte
+		var runtimeExplicit bool
+		var modelPolicyMode string
+		if err := rows.Scan(
+			&basePolicyBytes,
+			&platformPolicyBytes,
+			&runtimePolicyBytes,
+			&runtimeExplicit,
+			&runtimeOverrideBytes,
+			&modelPolicyBytes,
+			&modelPolicyMode,
+		); err != nil {
+			return AsyncWorkerCapacitySnapshot{}, err
+		}
+		modelPolicies = append(modelPolicies, EffectiveRateLimitPolicy(EffectiveRateLimitPolicyInput{
+			BasePolicy:            decodeObject(basePolicyBytes),
+			PlatformPolicy:        decodeObject(platformPolicyBytes),
+			RuntimePolicy:         decodeObject(runtimePolicyBytes),
+			RuntimePolicyExplicit: runtimeExplicit,
+			RuntimePolicyOverride: decodeObject(runtimeOverrideBytes),
+			ModelPolicy:           decodeObject(modelPolicyBytes),
+			ModelPolicyMode:       modelPolicyMode,
+		}))
+	}
+	if err := rows.Err(); err != nil {
+		return AsyncWorkerCapacitySnapshot{}, err
+	}
+	groupRows, err := s.pool.Query(ctx, `
+SELECT rate_limit_policy
+FROM gateway_user_groups
+WHERE status = 'active'`)
+	if err != nil {
+		return AsyncWorkerCapacitySnapshot{}, err
+	}
+	defer groupRows.Close()
+	groupPolicies := make([]map[string]any, 0)
+	for groupRows.Next() {
+		var policyBytes []byte
+		if err := groupRows.Scan(&policyBytes); err != nil {
+			return AsyncWorkerCapacitySnapshot{}, err
+		}
+		groupPolicies = append(groupPolicies, NormalizeRateLimitPolicy(decodeObject(policyBytes)))
+	}
+	if err := groupRows.Err(); err != nil {
+		return AsyncWorkerCapacitySnapshot{}, err
+	}
+	return asyncWorkerCapacityFromPolicySets(modelPolicies, groupPolicies, hardLimit), nil
+}
+
+func asyncWorkerCapacityFromPolicies(policies []map[string]any, hardLimit int) AsyncWorkerCapacitySnapshot {
+	return asyncWorkerCapacityFromPolicySets(policies, nil, hardLimit)
+}
+
+func asyncWorkerCapacityFromPolicySets(modelPolicies []map[string]any, groupPolicies []map[string]any, hardLimit int) AsyncWorkerCapacitySnapshot {
+	snapshot := AsyncWorkerCapacitySnapshot{
+		HardLimit:     hardLimit,
+		EnabledModels: len(modelPolicies),
+		EnabledGroups: len(groupPolicies),
+	}
+	modelDesired, modelFinite, unlimitedModels := concurrentPolicySetCapacity(modelPolicies)
+	groupDesired, groupFinite, unlimitedGroups := concurrentPolicySetCapacity(groupPolicies)
+	snapshot.ModelDesired = modelDesired
+	snapshot.GroupDesired = groupDesired
+	snapshot.UnlimitedModels = unlimitedModels
+	snapshot.UnlimitedGroups = unlimitedGroups
+
+	desired := hardLimit
+	switch {
+	case snapshot.EnabledModels == 0:
+		desired = 1
+	case modelFinite && groupFinite:
+		desired = min(modelDesired, groupDesired)
+	case modelFinite:
+		desired = modelDesired
+	case groupFinite:
+		desired = groupDesired
+	}
+	if desired < 1 {
+		desired = 1
+	}
+	snapshot.Desired = desired
+	snapshot.Capacity = desired
+	if snapshot.Capacity > hardLimit {
+		snapshot.Capacity = hardLimit
+		snapshot.Capped = true
+	}
+	return snapshot
+}
+
+func concurrentPolicySetCapacity(policies []map[string]any) (total int, finite bool, unlimited int) {
+	if len(policies) == 0 {
+		return 0, false, 0
+	}
+	finite = true
+	for _, policy := range policies {
+		capacity, policyFinite := ConcurrentPolicyCapacity(policy)
+		if !policyFinite {
+			unlimited++
+			finite = false
+			continue
+		}
+		total += capacity
+	}
+	return total, finite, unlimited
+}
@@ -25,7 +25,7 @@ func (s *Store) ListModelCandidates(ctx context.Context, model string, modelType
       COALESCE(p.base_url, ''),
       p.auth_type, p.credentials, p.config, p.default_pricing_mode,
       p.default_discount_factor::float8, COALESCE(p.pricing_rule_set_id::text, ''),
-       p.retry_policy, p.rate_limit_policy,
+       p.retry_policy, p.rate_limit_policy, COALESCE(b.default_rate_limit_policy, '{}'::jsonb),
       COALESCE(p.dynamic_priority, p.priority) AS effective_priority,
       m.id::text, COALESCE(m.base_model_id::text, ''), COALESCE(b.canonical_model_key, ''),
       COALESCE(NULLIF(m.provider_model_name, ''), m.model_name), m.model_name, COALESCE(m.model_alias, ''),
@@ -54,8 +54,11 @@ func (s *Store) ListModelCandidates(ctx context.Context, model string, modelType
       COALESCE(b.base_billing_config, '{}'::jsonb), m.billing_config, m.billing_config_override,
       m.pricing_mode, COALESCE(m.discount_factor, 0)::float8, COALESCE(m.pricing_rule_set_id::text, ''),
       COALESCE(b.pricing_rule_set_id::text, ''),
-       m.permission_config, m.retry_policy, m.rate_limit_policy, COALESCE(m.runtime_policy_set_id::text, COALESCE(b.runtime_policy_set_id::text, '')),
+       m.permission_config, m.retry_policy, m.rate_limit_policy, m.rate_limit_policy_mode,
+       COALESCE(m.runtime_policy_set_id::text, COALESCE(b.runtime_policy_set_id::text, '')),
+       (m.runtime_policy_set_id IS NOT NULL),
       COALESCE(NULLIF(m.runtime_policy_override, '{}'::jsonb), b.runtime_policy_override, '{}'::jsonb),
+       COALESCE(m.runtime_policy_override, '{}'::jsonb),
       COALESCE(rp.retry_policy, '{}'::jsonb), COALESCE(rp.rate_limit_policy, '{}'::jsonb),
       COALESCE(rp.auto_disable_policy, '{}'::jsonb), COALESCE(rp.degrade_policy, '{}'::jsonb),
       COALESCE(con.active, 0)::float8,
@@ -177,6 +180,7 @@ WHERE p.status = 'enabled'
 		var platformConfig []byte
 		var platformRetryPolicy []byte
 		var platformRateLimitPolicy []byte
+		var baseRateLimitPolicy []byte
 		var capabilities []byte
 		var capabilityOverride []byte
 		var baseBilling []byte
@@ -186,6 +190,7 @@ WHERE p.status = 'enabled'
 		var modelRetryPolicy []byte
 		var modelRateLimitPolicy []byte
 		var runtimePolicyOverride []byte
+		var rateLimitRuntimeOverride []byte
 		var runtimeRetryPolicy []byte
 		var runtimeRateLimitPolicy []byte
 		var autoDisablePolicy []byte
@@ -222,6 +227,7 @@ WHERE p.status = 'enabled'
 			&item.PlatformPricingRuleSetID,
 			&platformRetryPolicy,
 			&platformRateLimitPolicy,
+			&baseRateLimitPolicy,
 			&item.PlatformPriority,
 			&item.PlatformModelID,
 			&item.BaseModelID,
@@ -244,8 +250,11 @@ WHERE p.status = 'enabled'
 			&permissionConfig,
 			&modelRetryPolicy,
 			&modelRateLimitPolicy,
+			&item.ModelRateLimitPolicyMode,
 			&item.RuntimePolicySetID,
+			&item.RuntimePolicyExplicit,
 			&runtimePolicyOverride,
+			&rateLimitRuntimeOverride,
 			&runtimeRetryPolicy,
 			&runtimeRateLimitPolicy,
 			&autoDisablePolicy,
@@ -274,6 +283,7 @@ WHERE p.status = 'enabled'
 		item.PlatformConfig = decodeObject(platformConfig)
 		item.PlatformRetryPolicy = decodeObject(platformRetryPolicy)
 		item.PlatformRateLimitPolicy = decodeObject(platformRateLimitPolicy)
+		item.BaseRateLimitPolicy = decodeObject(baseRateLimitPolicy)
 		item.Capabilities = decodeObject(capabilities)
 		item.CapabilityOverride = decodeObject(capabilityOverride)
 		item.BaseBillingConfig = decodeObject(baseBilling)
@@ -283,6 +293,7 @@ WHERE p.status = 'enabled'
 		item.ModelRetryPolicy = decodeObject(modelRetryPolicy)
 		item.ModelRateLimitPolicy = decodeObject(modelRateLimitPolicy)
 		item.RuntimePolicyOverride = decodeObject(runtimePolicyOverride)
+		item.RateLimitRuntimeOverride = decodeObject(rateLimitRuntimeOverride)
 		item.RuntimeRetryPolicy = decodeObject(runtimeRetryPolicy)
 		item.RuntimeRateLimitPolicy = decodeObject(runtimeRateLimitPolicy)
 		item.AutoDisablePolicy = decodeObject(autoDisablePolicy)
@@ -303,7 +314,15 @@ WHERE p.status = 'enabled'
 			LastObservedUnix:  cacheLastObservedUnix,
 		})
 		applyRuntimeCandidateLoad(&item, runtimeCandidateLoadInput{
-			Policy:            effectiveModelRateLimitPolicy(item.PlatformRateLimitPolicy, item.RuntimeRateLimitPolicy, item.RuntimePolicySetID, item.RuntimePolicyOverride, item.ModelRateLimitPolicy),
+			Policy: EffectiveRateLimitPolicy(EffectiveRateLimitPolicyInput{
+				BasePolicy:            item.BaseRateLimitPolicy,
+				PlatformPolicy:        item.PlatformRateLimitPolicy,
+				RuntimePolicy:         item.RuntimeRateLimitPolicy,
+				RuntimePolicyExplicit: item.RuntimePolicyExplicit,
+				RuntimePolicyOverride: item.RateLimitRuntimeOverride,
+				ModelPolicy:           item.ModelRateLimitPolicy,
+				ModelPolicyMode:       item.ModelRateLimitPolicyMode,
+			}),
 			ConcurrentActive:  concurrentActive,
 			QueuedWaiting:     queuedWaiting,
 			RPMUsed:           rpmUsed,
@@ -143,18 +143,13 @@ SELECT EXISTS (
 	// soon as the base pricing rule changes and can mask the authoritative rule.
 	billingConfig := input.BillingConfig
 	explicitRuntimePolicySetID := strings.TrimSpace(input.RuntimePolicySetID)
+	rateLimitPolicyMode := NormalizeRateLimitPolicyMode(input.RateLimitPolicyMode, input.RateLimitPolicy != nil)
 	rateLimitPolicy := input.RateLimitPolicy
-	if len(rateLimitPolicy) == 0 && explicitRuntimePolicySetID == "" {
-		rateLimitPolicy = base.DefaultRateLimitPolicy
+	if rateLimitPolicyMode == RateLimitPolicyModeInherit {
+		rateLimitPolicy = nil
 	}
 	runtimePolicySetID := explicitRuntimePolicySetID
-	if runtimePolicySetID == "" {
-		runtimePolicySetID = base.RuntimePolicySetID
-	}
 	runtimePolicyOverride := input.RuntimePolicyOverride
-	if len(runtimePolicyOverride) == 0 {
-		runtimePolicyOverride = base.RuntimePolicyOverride
-	}

 	capabilityOverrideJSON, _ := json.Marshal(emptyObjectIfNil(input.CapabilityOverride))
 	capabilitiesJSON, _ := json.Marshal(emptyObjectIfNil(capabilities))
@@ -189,14 +184,14 @@ SELECT EXISTS (
 INSERT INTO platform_models (
  platform_id, base_model_id, model_name, provider_model_name, model_alias, model_type, display_name,
  capability_override, capabilities, pricing_mode, discount_factor,
-  pricing_rule_set_id, billing_config_override, billing_config, permission_config, retry_policy, rate_limit_policy,
+  pricing_rule_set_id, billing_config_override, billing_config, permission_config, retry_policy, rate_limit_policy, rate_limit_policy_mode,
  runtime_policy_set_id, runtime_policy_override, enabled
 )
 VALUES (
  $1::uuid, $2::uuid, $3, NULLIF($4, ''), NULLIF($5, ''), $6::jsonb, $7,
  $8::jsonb, $9::jsonb, $10, $11::numeric,
-  NULLIF($12, '')::uuid, $13::jsonb, $14::jsonb, $15::jsonb, $16::jsonb, $17::jsonb,
-  NULLIF($18, '')::uuid, $19::jsonb, true
+  NULLIF($12, '')::uuid, $13::jsonb, $14::jsonb, $15::jsonb, $16::jsonb, $17::jsonb, $18,
+  NULLIF($19, '')::uuid, $20::jsonb, true
 )
 ON CONFLICT (platform_id, model_name) DO UPDATE
 SET base_model_id = EXCLUDED.base_model_id,
@@ -213,6 +208,7 @@ SET base_model_id = EXCLUDED.base_model_id,
    permission_config = EXCLUDED.permission_config,
    retry_policy = EXCLUDED.retry_policy,
    rate_limit_policy = EXCLUDED.rate_limit_policy,
+    rate_limit_policy_mode = EXCLUDED.rate_limit_policy_mode,
    runtime_policy_set_id = EXCLUDED.runtime_policy_set_id,
    runtime_policy_override = EXCLUDED.runtime_policy_override,
    enabled = true,
@@ -221,7 +217,7 @@ RETURNING id::text, platform_id::text, COALESCE(base_model_id::text, ''), model_
          COALESCE(NULLIF(provider_model_name, ''), model_name), COALESCE(model_alias, ''), model_type, display_name, capability_override,
          capabilities, pricing_mode, COALESCE(discount_factor, 0)::float8,
          COALESCE(pricing_rule_set_id::text, ''), billing_config_override, billing_config,
-          permission_config, retry_policy, rate_limit_policy, COALESCE(runtime_policy_set_id::text, ''), runtime_policy_override,
+          permission_config, retry_policy, rate_limit_policy, rate_limit_policy_mode, COALESCE(runtime_policy_set_id::text, ''), runtime_policy_override,
          COALESCE(to_char(cooldown_until AT TIME ZONE 'UTC', 'YYYY-MM-DD"T"HH24:MI:SS.MS"Z"'), ''),
          enabled, created_at, updated_at`,
 		input.PlatformID,
@@ -241,6 +237,7 @@ RETURNING id::text, platform_id::text, COALESCE(base_model_id::text, ''), model_
 		string(permissionJSON),
 		string(retryJSON),
 		string(rateLimitJSON),
+		rateLimitPolicyMode,
 		runtimePolicySetID,
 		string(runtimePolicyOverrideJSON),
 	).Scan(
@@ -262,6 +259,7 @@ RETURNING id::text, platform_id::text, COALESCE(base_model_id::text, ''), model_
 		&permissionBytes,
 		&retryPolicyBytes,
 		&rateLimitPolicyBytes,
+		&model.RateLimitPolicyMode,
 		&model.RuntimePolicySetID,
 		&runtimePolicyOverrideBytes,
 		&model.CooldownUntil,
@@ -235,6 +235,8 @@ type PlatformModel struct {
 	Capabilities             map[string]any `json:"capabilities,omitempty"`
 	BaseCapabilities         map[string]any `json:"-"`
 	BaseBillingConfig        map[string]any `json:"-"`
+	BaseRateLimitPolicy      map[string]any `json:"-"`
+	BaseRuntimePolicySetID   string         `json:"-"`
 	BasePricingRuleSetID     string         `json:"-"`
 	PlatformPricingRuleSetID string         `json:"-"`
 	PricingMode              string         `json:"pricingMode"`
@@ -245,6 +247,7 @@ type PlatformModel struct {
 	PermissionConfig         map[string]any `json:"permissionConfig,omitempty"`
 	RetryPolicy              map[string]any `json:"retryPolicy,omitempty"`
 	RateLimitPolicy          map[string]any `json:"rateLimitPolicy,omitempty"`
+	RateLimitPolicyMode      string         `json:"rateLimitPolicyMode" enums:"inherit,override"`
 	RuntimePolicySetID       string         `json:"runtimePolicySetId,omitempty"`
 	RuntimePolicyOverride    map[string]any `json:"runtimePolicyOverride,omitempty"`
 	CooldownUntil            string         `json:"cooldownUntil,omitempty"`
@@ -956,7 +959,9 @@ SELECT m.id::text, m.platform_id::text, COALESCE(m.base_model_id::text, ''), p.p
       COALESCE(b.base_billing_config, '{}'::jsonb), COALESCE(b.pricing_rule_set_id::text, ''),
       COALESCE(p.pricing_rule_set_id::text, ''), m.pricing_mode, COALESCE(m.discount_factor, 0)::float8,
       COALESCE(m.pricing_rule_set_id::text, ''), m.billing_config_override, m.billing_config,
-       m.permission_config, m.retry_policy, m.rate_limit_policy, COALESCE(m.runtime_policy_set_id::text, ''), m.runtime_policy_override,
+       m.permission_config, m.retry_policy, m.rate_limit_policy, m.rate_limit_policy_mode,
+       COALESCE(m.runtime_policy_set_id::text, ''), m.runtime_policy_override,
+       COALESCE(b.default_rate_limit_policy, '{}'::jsonb), COALESCE(b.runtime_policy_set_id::text, ''),
       COALESCE(to_char(m.cooldown_until AT TIME ZONE 'UTC', 'YYYY-MM-DD"T"HH24:MI:SS.MS"Z"'), ''),
       m.enabled, m.created_at, m.updated_at
 FROM platform_models m
@@ -964,6 +969,7 @@ JOIN integration_platforms p ON p.id = m.platform_id
 LEFT JOIN LATERAL (
  SELECT catalog.invocation_name, catalog.display_name, catalog.model_type,
         catalog.capabilities, catalog.base_billing_config, catalog.pricing_rule_set_id,
+         catalog.default_rate_limit_policy, catalog.runtime_policy_set_id,
         COALESCE((
           SELECT jsonb_agg(DISTINCT compatibility_alias.alias ORDER BY compatibility_alias.alias)
           FROM model_compatibility_aliases compatibility_alias
@@ -1004,6 +1010,7 @@ ORDER BY m.model_type ASC, m.model_name ASC`, args...)
 		var retryPolicy []byte
 		var rateLimitPolicy []byte
 		var runtimePolicyOverride []byte
+		var baseRateLimitPolicy []byte
 		var modelTypeBytes []byte
 		var legacyAliasesBytes []byte
 		if err := rows.Scan(
@@ -1032,8 +1039,11 @@ ORDER BY m.model_type ASC, m.model_name ASC`, args...)
 			&permissionConfig,
 			&retryPolicy,
 			&rateLimitPolicy,
+			&model.RateLimitPolicyMode,
 			&model.RuntimePolicySetID,
 			&runtimePolicyOverride,
+			&baseRateLimitPolicy,
+			&model.BaseRuntimePolicySetID,
 			&model.CooldownUntil,
 			&model.Enabled,
 			&model.CreatedAt,
@@ -1053,6 +1063,7 @@ ORDER BY m.model_type ASC, m.model_name ASC`, args...)
 		model.RetryPolicy = decodeObject(retryPolicy)
 		model.RateLimitPolicy = decodeObject(rateLimitPolicy)
 		model.RuntimePolicyOverride = decodeObject(runtimePolicyOverride)
+		model.BaseRateLimitPolicy = decodeObject(baseRateLimitPolicy)
 		models = append(models, model)
 	}
 	return models, rows.Err()
@@ -0,0 +1,167 @@
+package store
+
+import (
+	"math"
+	"strings"
+)
+
+const (
+	RateLimitPolicyModeInherit  = "inherit"
+	RateLimitPolicyModeOverride = "override"
+)
+
+type EffectiveRateLimitPolicyInput struct {
+	BasePolicy            map[string]any
+	PlatformPolicy        map[string]any
+	RuntimePolicy         map[string]any
+	RuntimePolicyExplicit bool
+	RuntimePolicyOverride map[string]any
+	ModelPolicy           map[string]any
+	ModelPolicyMode       string
+}
+
+// EffectiveRateLimitPolicy resolves one authoritative policy. Platform
+// policies are defaults for every bound model; explicit model/runtime settings
+// replace the complete policy instead of merging individual metrics.
+func EffectiveRateLimitPolicy(input EffectiveRateLimitPolicyInput) map[string]any {
+	policy := input.BasePolicy
+	if policySpecified(input.PlatformPolicy) {
+		policy = input.PlatformPolicy
+	}
+	if input.RuntimePolicyExplicit {
+		policy = input.RuntimePolicy
+	}
+	if raw, ok := input.RuntimePolicyOverride["rateLimitPolicy"]; ok {
+		policy, _ = raw.(map[string]any)
+	}
+	mode := NormalizeRateLimitPolicyMode(input.ModelPolicyMode, input.ModelPolicy != nil)
+	if mode == RateLimitPolicyModeOverride {
+		policy = input.ModelPolicy
+	}
+	return NormalizeRateLimitPolicy(policy)
+}
+
+func NormalizeRateLimitPolicyMode(mode string, policyProvided bool) string {
+	switch strings.ToLower(strings.TrimSpace(mode)) {
+	case RateLimitPolicyModeOverride:
+		return RateLimitPolicyModeOverride
+	case RateLimitPolicyModeInherit:
+		return RateLimitPolicyModeInherit
+	default:
+		if policyProvided {
+			return RateLimitPolicyModeOverride
+		}
+		return RateLimitPolicyModeInherit
+	}
+}
+
+func RateLimitPolicyMetric(policy map[string]any, metric string) (float64, bool) {
+	rules, _ := NormalizeRateLimitPolicy(policy)["rules"].([]any)
+	for _, rawRule := range rules {
+		rule, _ := rawRule.(map[string]any)
+		if strings.TrimSpace(stringValue(rule["metric"])) != metric {
+			continue
+		}
+		value := floatValue(rule["limit"])
+		return value, true
+	}
+	return 0, false
+}
+
+// NormalizeRateLimitPolicy keeps the canonical rules contract while accepting
+// policies imported from server-main before that contract existed. Runtime
+// enforcement and worker sizing must agree on these legacy limits; otherwise a
+// configured max_concurrent_requests silently becomes unlimited.
+func NormalizeRateLimitPolicy(policy map[string]any) map[string]any {
+	if policy == nil {
+		return nil
+	}
+	out := clonePolicy(policy)
+	rules, _ := out["rules"].([]any)
+	if len(rules) > 0 {
+		return out
+	}
+
+	legacyScopes := []map[string]any{policy}
+	for _, key := range []string{"platformLimits", "modelLimits", "platform_limits", "model_limits"} {
+		if nested, ok := policy[key].(map[string]any); ok {
+			legacyScopes = append(legacyScopes, nested)
+		}
+	}
+	if limit, ok := lowestPositiveLegacyLimit(legacyScopes,
+		"max_concurrent_requests", "maxConcurrentRequests", "concurrent"); ok {
+		rules = append(rules, map[string]any{
+			"metric":          "concurrent",
+			"limit":           limit,
+			"leaseTtlSeconds": 120,
+		})
+	}
+	if limit, ok := lowestPositiveLegacyLimit(legacyScopes,
+		"max_request_per_minute", "maxRequestsPerMinute", "rpm"); ok {
+		rules = append(rules, map[string]any{
+			"metric":        "rpm",
+			"limit":         limit,
+			"windowSeconds": 60,
+		})
+	}
+	if limit, ok := lowestPositiveLegacyLimit(legacyScopes,
+		"max_tokens_per_minute", "maxTokensPerMinute", "tpm_total"); ok {
+		rules = append(rules, map[string]any{
+			"metric":        "tpm_total",
+			"limit":         limit,
+			"windowSeconds": 60,
+		})
+	}
+	if len(rules) > 0 {
+		out["rules"] = rules
+	}
+	return out
+}
+
+func ConcurrentPolicyCapacity(policy map[string]any) (int, bool) {
+	limit, ok := RateLimitPolicyMetric(policy, "concurrent")
+	if !ok || limit <= 0 {
+		return 0, false
+	}
+	if limit < 1 {
+		return 1, true
+	}
+	if limit >= float64(math.MaxInt) {
+		return math.MaxInt, true
+	}
+	return int(math.Floor(limit)), true
+}
+
+func policySpecified(policy map[string]any) bool {
+	rules, _ := NormalizeRateLimitPolicy(policy)["rules"].([]any)
+	return len(rules) > 0
+}
+
+func lowestPositiveLegacyLimit(scopes []map[string]any, keys ...string) (float64, bool) {
+	limit := 0.0
+	found := false
+	for _, scope := range scopes {
+		for _, key := range keys {
+			value := floatValue(scope[key])
+			if value <= 0 {
+				continue
+			}
+			if !found || value < limit {
+				limit = value
+				found = true
+			}
+		}
+	}
+	return limit, found
+}
+
+func clonePolicy(policy map[string]any) map[string]any {
+	if policy == nil {
+		return nil
+	}
+	out := make(map[string]any, len(policy))
+	for key, value := range policy {
+		out[key] = value
+	}
+	return out
+}
@@ -0,0 +1,120 @@
+package store
+
+import (
+	"context"
+	"fmt"
+	"net/url"
+	"os"
+	"path/filepath"
+	"runtime"
+	"strings"
+	"testing"
+	"time"
+
+	"github.com/jackc/pgx/v5/pgxpool"
+)
+
+func TestRateLimitPolicyModeMigrationClassifiesExistingRows(t *testing.T) {
+	databaseURL := strings.TrimSpace(os.Getenv("AI_GATEWAY_TEST_DATABASE_URL"))
+	if databaseURL == "" {
+		t.Skip("set AI_GATEWAY_TEST_DATABASE_URL to run rate-limit migration PostgreSQL integration tests")
+	}
+	ctx, cancel := context.WithTimeout(context.Background(), 20*time.Second)
+	defer cancel()
+	schema := fmt.Sprintf("rate_limit_mode_%d", time.Now().UnixNano())
+	adminPool, err := pgxpool.New(ctx, databaseURL)
+	if err != nil {
+		t.Fatalf("connect migration test database: %v", err)
+	}
+	defer adminPool.Close()
+	if _, err := adminPool.Exec(ctx, `CREATE SCHEMA `+schema); err != nil {
+		t.Fatalf("create migration test schema: %v", err)
+	}
+	defer adminPool.Exec(context.Background(), `DROP SCHEMA IF EXISTS `+schema+` CASCADE`)
+
+	schemaURL, err := url.Parse(databaseURL)
+	if err != nil {
+		t.Fatalf("parse test database url: %v", err)
+	}
+	query := schemaURL.Query()
+	query.Set("search_path", schema)
+	schemaURL.RawQuery = query.Encode()
+	pool, err := pgxpool.New(ctx, schemaURL.String())
+	if err != nil {
+		t.Fatalf("connect migration test schema: %v", err)
+	}
+	defer pool.Close()
+	if _, err := pool.Exec(ctx, `
+CREATE TABLE base_model_catalog (
+  id uuid PRIMARY KEY,
+  default_rate_limit_policy jsonb NOT NULL DEFAULT '{}'::jsonb,
+  runtime_policy_set_id uuid,
+  runtime_policy_override jsonb NOT NULL DEFAULT '{}'::jsonb
+);
+CREATE TABLE integration_platforms (
+  id uuid PRIMARY KEY,
+  rate_limit_policy jsonb NOT NULL DEFAULT '{}'::jsonb,
+  updated_at timestamptz NOT NULL DEFAULT now()
+);
+CREATE TABLE platform_models (
+  id uuid PRIMARY KEY,
+  base_model_id uuid,
+  rate_limit_policy jsonb NOT NULL DEFAULT '{}'::jsonb,
+  runtime_policy_set_id uuid,
+  runtime_policy_override jsonb NOT NULL DEFAULT '{}'::jsonb,
+  updated_at timestamptz NOT NULL DEFAULT now()
+);
+INSERT INTO integration_platforms (id, rate_limit_policy)
+VALUES ('20000000-0000-0000-0000-000000000001', '{"rules":[]}');
+INSERT INTO base_model_catalog (id, default_rate_limit_policy)
+VALUES ('00000000-0000-0000-0000-000000000001', '{"rules":[{"metric":"concurrent","limit":4}]}');
+INSERT INTO platform_models (id, base_model_id, rate_limit_policy, runtime_policy_override) VALUES
+  ('10000000-0000-0000-0000-000000000001', '00000000-0000-0000-0000-000000000001', '{"rules":[{"metric":"concurrent","limit":4}]}', '{}'),
+  ('10000000-0000-0000-0000-000000000002', '00000000-0000-0000-0000-000000000001', '{"rules":[{"metric":"concurrent","limit":8}]}', '{}'),
+  ('10000000-0000-0000-0000-000000000003', '00000000-0000-0000-0000-000000000001', '{"rules":[{"metric":"concurrent","limit":4}]}', '{"rateLimitPolicy":{"rules":[]}}'),
+  ('10000000-0000-0000-0000-000000000004', NULL, '{}', '{}');`); err != nil {
+		t.Fatalf("seed pre-migration rows: %v", err)
+	}
+	_, currentFile, _, _ := runtime.Caller(0)
+	migrationPath := filepath.Join(filepath.Dir(currentFile), "..", "..", "migrations", "0080_platform_model_rate_limit_policy_mode.sql")
+	migration, err := os.ReadFile(migrationPath)
+	if err != nil {
+		t.Fatalf("read migration: %v", err)
+	}
+	if _, err := pool.Exec(ctx, string(migration)); err != nil {
+		t.Fatalf("apply migration: %v", err)
+	}
+	rows, err := pool.Query(ctx, `SELECT id::text, rate_limit_policy_mode FROM platform_models ORDER BY id`)
+	if err != nil {
+		t.Fatalf("read classified rows: %v", err)
+	}
+	defer rows.Close()
+	want := []string{"inherit", "override", "override", "inherit"}
+	index := 0
+	for rows.Next() {
+		var id, mode string
+		if err := rows.Scan(&id, &mode); err != nil {
+			t.Fatalf("scan classified row: %v", err)
+		}
+		if index >= len(want) || mode != want[index] {
+			t.Fatalf("row %s mode=%s, want=%s", id, mode, want[index])
+		}
+		index++
+	}
+	if err := rows.Err(); err != nil {
+		t.Fatalf("read classified rows: %v", err)
+	}
+	if index != len(want) {
+		t.Fatalf("classified %d rows, want %d", index, len(want))
+	}
+	var normalizedPlatformPolicy string
+	if err := pool.QueryRow(ctx, `
+SELECT rate_limit_policy::text
+FROM integration_platforms
+WHERE id = '20000000-0000-0000-0000-000000000001'`).Scan(&normalizedPlatformPolicy); err != nil {
+		t.Fatalf("read normalized platform policy: %v", err)
+	}
+	if normalizedPlatformPolicy != "{}" {
+		t.Fatalf("normalized platform policy=%s, want={}", normalizedPlatformPolicy)
+	}
+}
@@ -0,0 +1,169 @@
+package store
+
+import "testing"
+
+func TestEffectiveRateLimitPolicyPrecedence(t *testing.T) {
+	policy := func(limit float64) map[string]any {
+		return map[string]any{"rules": []any{map[string]any{"metric": "concurrent", "limit": limit}}}
+	}
+	tests := []struct {
+		name  string
+		input EffectiveRateLimitPolicyInput
+		want  float64
+		ok    bool
+	}{
+		{name: "base", input: EffectiveRateLimitPolicyInput{BasePolicy: policy(2)}, want: 2, ok: true},
+		{name: "platform", input: EffectiveRateLimitPolicyInput{BasePolicy: policy(2), PlatformPolicy: policy(4)}, want: 4, ok: true},
+		{name: "empty platform inherits base", input: EffectiveRateLimitPolicyInput{BasePolicy: policy(2), PlatformPolicy: map[string]any{"rules": []any{}}}, want: 2, ok: true},
+		{name: "runtime", input: EffectiveRateLimitPolicyInput{PlatformPolicy: policy(4), RuntimePolicy: policy(8), RuntimePolicyExplicit: true}, want: 8, ok: true},
+		{name: "runtime override", input: EffectiveRateLimitPolicyInput{PlatformPolicy: policy(4), RuntimePolicyOverride: map[string]any{"rateLimitPolicy": policy(16)}}, want: 16, ok: true},
+		{name: "model override", input: EffectiveRateLimitPolicyInput{PlatformPolicy: policy(4), ModelPolicy: policy(32), ModelPolicyMode: "override"}, want: 32, ok: true},
+		{name: "model explicit unlimited", input: EffectiveRateLimitPolicyInput{PlatformPolicy: policy(4), ModelPolicy: map[string]any{}, ModelPolicyMode: "override"}, ok: false},
+		{name: "model inherit", input: EffectiveRateLimitPolicyInput{PlatformPolicy: policy(4), ModelPolicy: policy(32), ModelPolicyMode: "inherit"}, want: 4, ok: true},
+	}
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			got, ok := ConcurrentPolicyCapacity(EffectiveRateLimitPolicy(tt.input))
+			if ok != tt.ok || (ok && got != int(tt.want)) {
+				t.Fatalf("capacity = (%d, %v), want (%d, %v)", got, ok, int(tt.want), tt.ok)
+			}
+		})
+	}
+}
+
+func TestNormalizeRateLimitPolicyLegacyShapes(t *testing.T) {
+	tests := []struct {
+		name   string
+		policy map[string]any
+		metric string
+		want   float64
+	}{
+		{
+			name:   "platform concurrent",
+			policy: map[string]any{"platformLimits": map[string]any{"max_concurrent_requests": 5.0}},
+			metric: "concurrent",
+			want:   5,
+		},
+		{
+			name:   "model concurrent camel case",
+			policy: map[string]any{"modelLimits": map[string]any{"maxConcurrentRequests": 10.0}},
+			metric: "concurrent",
+			want:   10,
+		},
+		{
+			name: "stricter duplicate wins",
+			policy: map[string]any{
+				"platformLimits": map[string]any{"max_concurrent_requests": 8.0},
+				"modelLimits":    map[string]any{"max_concurrent_requests": 3.0},
+			},
+			metric: "concurrent",
+			want:   3,
+		},
+		{
+			name:   "requests per minute",
+			policy: map[string]any{"model_limits": map[string]any{"max_request_per_minute": 60.0}},
+			metric: "rpm",
+			want:   60,
+		},
+	}
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			got, ok := RateLimitPolicyMetric(tt.policy, tt.metric)
+			if !ok || got != tt.want {
+				t.Fatalf("metric %s = (%v, %v), want (%v, true)", tt.metric, got, ok, tt.want)
+			}
+		})
+	}
+}
+
+func TestConcurrentPolicyCapacityRoundsDown(t *testing.T) {
+	policy := map[string]any{"rules": []any{map[string]any{"metric": "concurrent", "limit": 96.9}}}
+	got, ok := ConcurrentPolicyCapacity(policy)
+	if !ok || got != 96 {
+		t.Fatalf("capacity = (%d, %v), want (96, true)", got, ok)
+	}
+}
+
+func TestAsyncWorkerCapacityAggregation(t *testing.T) {
+	policy := func(limit float64) map[string]any {
+		return map[string]any{"rules": []any{map[string]any{"metric": "concurrent", "limit": limit}}}
+	}
+	tests := []struct {
+		name         string
+		policies     []map[string]any
+		hardLimit    int
+		wantCapacity int
+		wantDesired  int
+		wantCapped   bool
+	}{
+		{name: "no enabled models", hardLimit: 2048, wantCapacity: 1, wantDesired: 1},
+		{name: "finite sum", policies: []map[string]any{policy(64), policy(32)}, hardLimit: 2048, wantCapacity: 96, wantDesired: 96},
+		{name: "unlimited model", policies: []map[string]any{policy(64), {}}, hardLimit: 2048, wantCapacity: 2048, wantDesired: 2048},
+		{name: "hard limit cap", policies: []map[string]any{policy(80), policy(80)}, hardLimit: 96, wantCapacity: 96, wantDesired: 160, wantCapped: true},
+	}
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			got := asyncWorkerCapacityFromPolicies(tt.policies, tt.hardLimit)
+			if got.Capacity != tt.wantCapacity || got.Desired != tt.wantDesired || got.Capped != tt.wantCapped {
+				t.Fatalf("snapshot=%+v, want capacity=%d desired=%d capped=%v", got, tt.wantCapacity, tt.wantDesired, tt.wantCapped)
+			}
+		})
+	}
+}
+
+func TestAsyncWorkerCapacityRespectsUserGroupCeiling(t *testing.T) {
+	policy := func(limit float64) map[string]any {
+		return map[string]any{"rules": []any{map[string]any{"metric": "concurrent", "limit": limit}}}
+	}
+	tests := []struct {
+		name         string
+		models       []map[string]any
+		groups       []map[string]any
+		hardLimit    int
+		wantCapacity int
+		wantDesired  int
+		wantCapped   bool
+	}{
+		{
+			name:         "group ceiling prevents worker oversubscription",
+			models:       []map[string]any{{}},
+			groups:       []map[string]any{policy(3), policy(10)},
+			hardLimit:    2048,
+			wantCapacity: 13,
+			wantDesired:  13,
+		},
+		{
+			name:         "model ceiling is stricter",
+			models:       []map[string]any{policy(5), policy(7)},
+			groups:       []map[string]any{policy(300)},
+			hardLimit:    2048,
+			wantCapacity: 12,
+			wantDesired:  12,
+		},
+		{
+			name:         "both policy sets unlimited use hard limit",
+			models:       []map[string]any{{}},
+			groups:       []map[string]any{{}},
+			hardLimit:    256,
+			wantCapacity: 256,
+			wantDesired:  256,
+		},
+		{
+			name:         "finite group desired still reports hard cap",
+			models:       []map[string]any{{}},
+			groups:       []map[string]any{policy(500)},
+			hardLimit:    256,
+			wantCapacity: 256,
+			wantDesired:  500,
+			wantCapped:   true,
+		},
+	}
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			got := asyncWorkerCapacityFromPolicySets(tt.models, tt.groups, tt.hardLimit)
+			if got.Capacity != tt.wantCapacity || got.Desired != tt.wantDesired || got.Capped != tt.wantCapped {
+				t.Fatalf("snapshot=%+v, want capacity=%d desired=%d capped=%v", got, tt.wantCapacity, tt.wantDesired, tt.wantCapped)
+			}
+		})
+	}
+}
@@ -0,0 +1,19 @@
+package store
+
+import (
+	"testing"
+	"time"
+)
+
+func TestConcurrencyRetryAfterBoundsPolling(t *testing.T) {
+	if got := concurrencyRetryAfter(time.Time{}); got != 5*time.Second {
+		t.Fatalf("zero expiry retry=%s, want=5s", got)
+	}
+	if got := concurrencyRetryAfter(time.Now().Add(30 * time.Second)); got != 5*time.Second {
+		t.Fatalf("distant expiry retry=%s, want=5s", got)
+	}
+	got := concurrencyRetryAfter(time.Now().Add(2500 * time.Millisecond))
+	if got < 2*time.Second || got > 2500*time.Millisecond {
+		t.Fatalf("near expiry retry=%s, want within [2s,2.5s]", got)
+	}
+}
@@ -143,8 +143,10 @@ func (s *Store) ListModelRateLimitStatuses(ctx context.Context) ([]ModelRateLimi
 		       p.priority, p.dynamic_priority, COALESCE(p.dynamic_priority, p.priority),
 		       m.model_name, COALESCE(NULLIF(m.provider_model_name, ''), m.model_name), COALESCE(m.model_alias, ''),
 	       m.model_type, m.display_name, m.enabled,
-	       p.rate_limit_policy, COALESCE(rp.rate_limit_policy, '{}'::jsonb), COALESCE(m.runtime_policy_set_id::text, b.runtime_policy_set_id::text, ''),
-	       COALESCE(NULLIF(m.runtime_policy_override, '{}'::jsonb), b.runtime_policy_override, '{}'::jsonb), m.rate_limit_policy,
+	       COALESCE(b.default_rate_limit_policy, '{}'::jsonb), p.rate_limit_policy,
+	       COALESCE(rp.rate_limit_policy, '{}'::jsonb),
+	       (m.runtime_policy_set_id IS NOT NULL),
+	       COALESCE(m.runtime_policy_override, '{}'::jsonb), m.rate_limit_policy, m.rate_limit_policy_mode,
 	       COALESCE(to_char(p.cooldown_until AT TIME ZONE 'UTC', 'YYYY-MM-DD"T"HH24:MI:SS.MS"Z"'), ''),
 	       COALESCE(to_char(m.cooldown_until AT TIME ZONE 'UTC', 'YYYY-MM-DD"T"HH24:MI:SS.MS"Z"'), ''),
 	       COALESCE(con.active, 0)::float8,
@@ -217,11 +219,13 @@ ORDER BY p.priority ASC, m.model_name ASC`)
 	for rows.Next() {
 		var item ModelRateLimitStatus
 		var modelTypeBytes []byte
+		var basePolicyBytes []byte
 		var platformPolicyBytes []byte
 		var runtimePolicyBytes []byte
-		var runtimePolicySetID string
+		var runtimePolicyExplicit bool
 		var runtimeOverrideBytes []byte
 		var modelPolicyBytes []byte
+		var modelPolicyMode string
 		var platformDynamicPriority sql.NullInt64
 		var platformCooldownUntil string
 		var modelCooldownUntil string
@@ -248,11 +252,13 @@ ORDER BY p.priority ASC, m.model_name ASC`)
 			&modelTypeBytes,
 			&item.DisplayName,
 			&item.Enabled,
+			&basePolicyBytes,
 			&platformPolicyBytes,
 			&runtimePolicyBytes,
-			&runtimePolicySetID,
+			&runtimePolicyExplicit,
 			&runtimeOverrideBytes,
 			&modelPolicyBytes,
+			&modelPolicyMode,
 			&platformCooldownUntil,
 			&modelCooldownUntil,
 			&concurrentCurrent,
@@ -268,13 +274,15 @@ ORDER BY p.priority ASC, m.model_name ASC`)
 		}
 		item.PlatformDynamicPriority = intPointerFromNull(platformDynamicPriority)
 		item.ModelType = decodeStringArray(modelTypeBytes)
-		policy := effectiveModelRateLimitPolicy(
-			decodeObject(platformPolicyBytes),
-			decodeObject(runtimePolicyBytes),
-			runtimePolicySetID,
-			decodeObject(runtimeOverrideBytes),
-			decodeObject(modelPolicyBytes),
-		)
+		policy := EffectiveRateLimitPolicy(EffectiveRateLimitPolicyInput{
+			BasePolicy:            decodeObject(basePolicyBytes),
+			PlatformPolicy:        decodeObject(platformPolicyBytes),
+			RuntimePolicy:         decodeObject(runtimePolicyBytes),
+			RuntimePolicyExplicit: runtimePolicyExplicit,
+			RuntimePolicyOverride: decodeObject(runtimeOverrideBytes),
+			ModelPolicy:           decodeObject(modelPolicyBytes),
+			ModelPolicyMode:       modelPolicyMode,
+		})
 		item.PlatformCooldownUntil = platformCooldownUntil
 		item.ModelCooldownUntil = modelCooldownUntil
 		item.RateLimitPolicy = policy
@@ -491,46 +499,6 @@ func platformPolicyEventFromPayload(id string, taskID string, eventType string,
 	}
 }

-func effectiveModelRateLimitPolicy(platformPolicy map[string]any, runtimePolicy map[string]any, runtimePolicySetID string, runtimeOverride map[string]any, modelPolicy map[string]any) map[string]any {
-	policy := platformPolicy
-	if strings.TrimSpace(runtimePolicySetID) != "" {
-		policy = runtimePolicy
-	} else if hasRateLimitRules(runtimePolicy) {
-		policy = shallowMergeMap(policy, runtimePolicy)
-	}
-	if _, hasOverride := runtimeOverride["rateLimitPolicy"]; hasOverride {
-		nested, _ := runtimeOverride["rateLimitPolicy"].(map[string]any)
-		if len(nested) == 0 {
-			policy = nil
-		} else {
-			policy = shallowMergeMap(policy, nested)
-		}
-	}
-	if hasRateLimitRules(modelPolicy) {
-		policy = shallowMergeMap(policy, modelPolicy)
-	}
-	if hasRateLimitRules(policy) {
-		return policy
-	}
-	return nil
-}
-
-func hasRateLimitRules(policy map[string]any) bool {
-	rules, _ := policy["rules"].([]any)
-	return len(rules) > 0
-}
-
-func shallowMergeMap(base map[string]any, override map[string]any) map[string]any {
-	out := map[string]any{}
-	for key, value := range base {
-		out[key] = value
-	}
-	for key, value := range override {
-		out[key] = value
-	}
-	return out
-}
-
 func rateLimitForMetric(policy map[string]any, metric string) float64 {
 	rules, _ := policy["rules"].([]any)
 	for _, rawRule := range rules {
@@ -6,23 +6,23 @@ import (
 )

 func TestEffectiveModelRateLimitPolicyTreatsModelRulesAsAuthoritative(t *testing.T) {
-	policy := effectiveModelRateLimitPolicy(
-		map[string]any{"rules": []any{
+	policy := EffectiveRateLimitPolicy(EffectiveRateLimitPolicyInput{
+		PlatformPolicy: map[string]any{"rules": []any{
 			map[string]any{"metric": "rpm", "limit": 500},
 			map[string]any{"metric": "tpm_total", "limit": 100000},
 		}},
-		map[string]any{"rules": []any{
+		RuntimePolicy: map[string]any{"rules": []any{
 			map[string]any{"metric": "rpm", "limit": 120},
 			map[string]any{"metric": "tpm_total", "limit": 240000},
 			map[string]any{"metric": "concurrent", "limit": 6},
 		}},
-		"runtime-policy-1",
-		map[string]any{},
-		map[string]any{"rules": []any{
+		RuntimePolicyExplicit: true,
+		ModelPolicy: map[string]any{"rules": []any{
 			map[string]any{"metric": "rpm", "limit": 30},
 			map[string]any{"metric": "concurrent", "limit": 2},
 		}},
-	)
+		ModelPolicyMode: "override",
+	})

 	if got := rateLimitForMetric(policy, "rpm"); got != 30 {
 		t.Fatalf("expected model rpm limit to win, got %v", got)
@@ -36,16 +36,15 @@ func TestEffectiveModelRateLimitPolicyTreatsModelRulesAsAuthoritative(t *testing
 }

 func TestEffectiveModelRateLimitPolicyTreatsEmptyRuntimePolicyAsUnlimited(t *testing.T) {
-	policy := effectiveModelRateLimitPolicy(
-		map[string]any{"rules": []any{
+	policy := EffectiveRateLimitPolicy(EffectiveRateLimitPolicyInput{
+		PlatformPolicy: map[string]any{"rules": []any{
 			map[string]any{"metric": "rpm", "limit": 500},
 			map[string]any{"metric": "tpm_total", "limit": 100000},
 		}},
-		map[string]any{"rules": []any{}},
-		"runtime-policy-1",
-		map[string]any{},
-		map[string]any{},
-	)
+		RuntimePolicy:         map[string]any{"rules": []any{}},
+		RuntimePolicyExplicit: true,
+		ModelPolicyMode:       "inherit",
+	})

 	if got := rateLimitForMetric(policy, "rpm"); got != 0 {
 		t.Fatalf("expected empty runtime policy rpm to mean unlimited, got %v", got)
@@ -56,17 +55,16 @@ func TestEffectiveModelRateLimitPolicyTreatsEmptyRuntimePolicyAsUnlimited(t *tes
 }

 func TestEffectiveModelRateLimitPolicyTreatsNegativeLimitAsUnlimited(t *testing.T) {
-	policy := effectiveModelRateLimitPolicy(
-		map[string]any{"rules": []any{
+	policy := EffectiveRateLimitPolicy(EffectiveRateLimitPolicyInput{
+		PlatformPolicy: map[string]any{"rules": []any{
 			map[string]any{"metric": "rpm", "limit": 500},
 		}},
-		map[string]any{"rules": []any{
+		RuntimePolicy: map[string]any{"rules": []any{
 			map[string]any{"metric": "rpm", "limit": -1},
 		}},
-		"runtime-policy-1",
-		map[string]any{},
-		map[string]any{},
-	)
+		RuntimePolicyExplicit: true,
+		ModelPolicyMode:       "inherit",
+	})

 	if got := rateLimitForMetric(policy, "rpm"); got != -1 {
 		t.Fatalf("expected negative runtime rpm marker to be preserved, got %v", got)
@@ -4,6 +4,7 @@ import (
 	"context"
 	"errors"
 	"fmt"
+	"sort"
 	"time"

 	"github.com/jackc/pgx/v5"
@@ -17,6 +18,8 @@ type RuntimeRecoveryResult struct {
 	RequeuedAsyncTasks        int64 `json:"requeuedAsyncTasks"`
 }

+var ErrConcurrencyLeaseLost = errors.New("concurrency lease lost")
+
 func (s *Store) ReserveRateLimits(ctx context.Context, taskID string, attemptID string, reservations []RateLimitReservation) (RateLimitResult, error) {
 	tx, err := s.pool.Begin(ctx)
 	if err != nil {
@@ -24,6 +27,26 @@ func (s *Store) ReserveRateLimits(ctx context.Context, taskID string, attemptID
 	}
 	defer tx.Rollback(ctx)

+	lockKeys := make([]string, 0)
+	lockKeySet := make(map[string]struct{})
+	for _, reservation := range reservations {
+		if reservation.Metric != "concurrent" || reservation.Limit <= 0 || reservation.Amount <= 0 {
+			continue
+		}
+		key := fmt.Sprintf("%d:%s%d:%s", len(reservation.ScopeType), reservation.ScopeType, len(reservation.ScopeKey), reservation.ScopeKey)
+		if _, exists := lockKeySet[key]; exists {
+			continue
+		}
+		lockKeySet[key] = struct{}{}
+		lockKeys = append(lockKeys, key)
+	}
+	sort.Strings(lockKeys)
+	for _, key := range lockKeys {
+		if _, err := tx.Exec(ctx, `SELECT pg_advisory_xact_lock(hashtextextended($1, 0))`, key); err != nil {
+			return RateLimitResult{}, err
+		}
+	}
+
 	result := RateLimitResult{}
 	for _, reservation := range reservations {
 		if reservation.Limit <= 0 || reservation.Amount <= 0 {
@@ -49,11 +72,11 @@ func (s *Store) ReserveRateLimits(ctx context.Context, taskID string, attemptID
 			reservation.WindowSeconds = 60
 		}
 		if reservation.Metric == "concurrent" {
-			leaseID, err := reserveConcurrencyLease(ctx, tx, taskID, attemptID, reservation)
+			lease, err := reserveConcurrencyLease(ctx, tx, taskID, attemptID, reservation)
 			if err != nil {
 				return RateLimitResult{}, err
 			}
-			result.LeaseIDs = append(result.LeaseIDs, leaseID)
+			result.Leases = append(result.Leases, lease)
 			continue
 		}
 		normalized, err := reserveCounterWindow(ctx, tx, taskID, attemptID, reservation)
@@ -65,7 +88,7 @@ func (s *Store) ReserveRateLimits(ctx context.Context, taskID string, attemptID
 	return result, tx.Commit(ctx)
 }

-func reserveConcurrencyLease(ctx context.Context, tx pgx.Tx, taskID string, attemptID string, reservation RateLimitReservation) (string, error) {
+func reserveConcurrencyLease(ctx context.Context, tx pgx.Tx, taskID string, attemptID string, reservation RateLimitReservation) (ConcurrencyLease, error) {
 	if reservation.LeaseTTLSeconds <= 0 {
 		reservation.LeaseTTLSeconds = 120
 	}
@@ -83,10 +106,10 @@ WHERE scope_type = $1
 		reservation.ScopeKey,
 		reservation.LeaseTTLSeconds,
 	).Scan(&active, &nextAvailableAt); err != nil {
-		return "", err
+		return ConcurrencyLease{}, err
 	}
 	if active+reservation.Amount > reservation.Limit {
-		return "", &RateLimitExceededError{
+		return ConcurrencyLease{}, &RateLimitExceededError{
 			ScopeType:     reservation.ScopeType,
 			ScopeKey:      reservation.ScopeKey,
 			ScopeName:     reservation.ScopeName,
@@ -117,9 +140,9 @@ RETURNING id::text`,
 		reservation.Amount,
 		reservation.LeaseTTLSeconds,
 	).Scan(&leaseID); err != nil {
-		return "", err
+		return ConcurrencyLease{}, err
 	}
-	return leaseID, nil
+	return ConcurrencyLease{ID: leaseID, TTL: time.Duration(reservation.LeaseTTLSeconds) * time.Second}, nil
 }

 func reserveCounterWindow(ctx context.Context, tx pgx.Tx, taskID string, attemptID string, reservation RateLimitReservation) (RateLimitReservation, error) {
@@ -232,13 +255,16 @@ func retryAfterUntil(when time.Time) time.Duration {

 func concurrencyRetryAfter(leaseExpiresAt time.Time) time.Duration {
 	if leaseExpiresAt.IsZero() {
-		return time.Second
+		return 5 * time.Second
 	}
 	duration := time.Until(leaseExpiresAt)
 	if duration <= time.Second {
 		return time.Second
 	}
-	return time.Second
+	if duration > 5*time.Second {
+		return 5 * time.Second
+	}
+	return duration
 }

 func (s *Store) CommitRateLimitReservations(ctx context.Context, reservations []RateLimitReservation, actualByMetric map[string]float64) error {
@@ -249,26 +275,57 @@ func (s *Store) ReleaseRateLimitReservations(ctx context.Context, reservations [
 	return s.finishRateLimitReservations(ctx, reservations, nil, "released", reason)
 }

-func (s *Store) ReleaseConcurrencyLeases(ctx context.Context, leaseIDs []string) error {
+func (s *Store) ReleaseConcurrencyLeases(ctx context.Context, leases []ConcurrencyLease) error {
+	leaseIDs := concurrencyLeaseIDs(leases)
 	if len(leaseIDs) == 0 {
 		return nil
 	}
-	for _, leaseID := range leaseIDs {
-		if leaseID == "" {
-			continue
-		}
-		if _, err := s.pool.Exec(ctx, `
+	_, err := s.pool.Exec(ctx, `
 UPDATE gateway_concurrency_leases
 SET released_at = now()
-WHERE id = $1::uuid AND released_at IS NULL`, leaseID); err != nil && !errors.Is(err, ErrRateLimited) {
-			return err
+WHERE id = ANY($1::uuid[]) AND released_at IS NULL`, leaseIDs)
+	return err
+}
+
+func (s *Store) RenewConcurrencyLeases(ctx context.Context, leases []ConcurrencyLease) error {
+	leaseIDs := make([]string, 0, len(leases))
+	ttlSeconds := make([]int32, 0, len(leases))
+	for _, lease := range leases {
+		if lease.ID == "" {
+			continue
 		}
+		ttl := lease.TTL
+		if ttl <= 0 {
+			ttl = 120 * time.Second
+		}
+		seconds := int32(ttl / time.Second)
+		if seconds < 1 {
+			seconds = 1
+		}
+		leaseIDs = append(leaseIDs, lease.ID)
+		ttlSeconds = append(ttlSeconds, seconds)
+	}
+	if len(leaseIDs) == 0 {
+		return nil
+	}
+	tag, err := s.pool.Exec(ctx, `
+UPDATE gateway_concurrency_leases lease
+SET expires_at = now() + (renewal.ttl_seconds * interval '1 second')
+FROM unnest($1::uuid[], $2::int[]) AS renewal(id, ttl_seconds)
+WHERE lease.id = renewal.id
+  AND lease.released_at IS NULL
+  AND lease.expires_at > now()`, leaseIDs, ttlSeconds)
+	if err != nil {
+		return err
+	}
+	if tag.RowsAffected() != int64(len(leaseIDs)) {
+		return fmt.Errorf("%w: renewed %d of %d leases", ErrConcurrencyLeaseLost, tag.RowsAffected(), len(leaseIDs))
 	}
 	return nil
 }

 func (s *Store) AttachRateLimitResultToAttempt(ctx context.Context, attemptID string, result RateLimitResult) error {
-	if attemptID == "" || (len(result.Reservations) == 0 && len(result.LeaseIDs) == 0) {
+	if attemptID == "" || (len(result.Reservations) == 0 && len(result.Leases) == 0) {
 		return nil
 	}
 	tx, err := s.pool.Begin(ctx)
@@ -289,7 +346,8 @@ WHERE id = $1::uuid`, reservation.ReservationID, attemptID); err != nil {
 			return err
 		}
 	}
-	for _, leaseID := range result.LeaseIDs {
+	for _, lease := range result.Leases {
+		leaseID := lease.ID
 		if leaseID == "" {
 			continue
 		}
@@ -303,6 +361,16 @@ WHERE id = $1::uuid`, leaseID, attemptID); err != nil {
 	return tx.Commit(ctx)
 }

+func concurrencyLeaseIDs(leases []ConcurrencyLease) []string {
+	ids := make([]string, 0, len(leases))
+	for _, lease := range leases {
+		if lease.ID != "" {
+			ids = append(ids, lease.ID)
+		}
+	}
+	return ids
+}
+
 func (s *Store) RecoverInterruptedRuntimeState(ctx context.Context) (RuntimeRecoveryResult, error) {
 	tx, err := s.pool.Begin(ctx)
 	if err != nil {
@@ -0,0 +1,198 @@
+package store
+
+import (
+	"context"
+	"errors"
+	"os"
+	"strings"
+	"sync"
+	"sync/atomic"
+	"testing"
+	"time"
+)
+
+func TestConcurrencyLeaseReservationIsAtomicAcrossPools(t *testing.T) {
+	databaseURL := strings.TrimSpace(os.Getenv("AI_GATEWAY_TEST_DATABASE_URL"))
+	if databaseURL == "" {
+		t.Skip("set AI_GATEWAY_TEST_DATABASE_URL to run concurrency lease PostgreSQL integration tests")
+	}
+	ctx, cancel := context.WithTimeout(context.Background(), 45*time.Second)
+	defer cancel()
+	first, err := Connect(ctx, databaseURL)
+	if err != nil {
+		t.Fatalf("connect first store: %v", err)
+	}
+	defer first.Close()
+	second, err := Connect(ctx, databaseURL)
+	if err != nil {
+		t.Fatalf("connect second store: %v", err)
+	}
+	defer second.Close()
+
+	scopeKey := "atomic-" + time.Now().UTC().Format("20060102150405.000000000")
+	taskIDs := createLeaseTestTasks(t, ctx, first, 256, scopeKey)
+	defer deleteLeaseTestTasks(t, first, taskIDs)
+
+	var successes atomic.Int64
+	var peak atomic.Int64
+	monitorCtx, stopMonitor := context.WithCancel(ctx)
+	var monitorWG sync.WaitGroup
+	monitorWG.Add(1)
+	go func() {
+		defer monitorWG.Done()
+		ticker := time.NewTicker(2 * time.Millisecond)
+		defer ticker.Stop()
+		for {
+			select {
+			case <-monitorCtx.Done():
+				return
+			case <-ticker.C:
+				var active int64
+				if err := first.Pool().QueryRow(monitorCtx, `
+SELECT COUNT(*)
+FROM gateway_concurrency_leases
+WHERE scope_type = 'platform_model'
+  AND scope_key = $1
+  AND released_at IS NULL
+  AND expires_at > now()`, scopeKey).Scan(&active); err == nil {
+					for active > peak.Load() && !peak.CompareAndSwap(peak.Load(), active) {
+					}
+				}
+			}
+		}
+	}()
+
+	var wg sync.WaitGroup
+	errs := make(chan error, len(taskIDs))
+	for index, taskID := range taskIDs {
+		wg.Add(1)
+		go func(index int, taskID string) {
+			defer wg.Done()
+			target := first
+			if index%2 == 1 {
+				target = second
+			}
+			_, err := target.ReserveRateLimits(ctx, taskID, "", []RateLimitReservation{{
+				ScopeType:       "platform_model",
+				ScopeKey:        scopeKey,
+				Metric:          "concurrent",
+				Limit:           64,
+				Amount:          1,
+				LeaseTTLSeconds: 30,
+			}})
+			if err == nil {
+				successes.Add(1)
+				return
+			}
+			if !errors.Is(err, ErrRateLimited) {
+				errs <- err
+			}
+		}(index, taskID)
+	}
+	wg.Wait()
+	stopMonitor()
+	monitorWG.Wait()
+	close(errs)
+	for err := range errs {
+		t.Fatalf("unexpected reservation error: %v", err)
+	}
+
+	var active int64
+	if err := first.Pool().QueryRow(ctx, `
+SELECT COUNT(*)
+FROM gateway_concurrency_leases
+WHERE scope_type = 'platform_model'
+  AND scope_key = $1
+  AND released_at IS NULL
+  AND expires_at > now()`, scopeKey).Scan(&active); err != nil {
+		t.Fatalf("count active leases: %v", err)
+	}
+	if successes.Load() != 64 || active != 64 {
+		t.Fatalf("successful reservations=%d active leases=%d, want exactly 64", successes.Load(), active)
+	}
+	if peak.Load() > 64 {
+		t.Fatalf("active lease peak=%d, want <=64", peak.Load())
+	}
+}
+
+func TestConcurrencyLeaseRenewalExtendsAndReleases(t *testing.T) {
+	databaseURL := strings.TrimSpace(os.Getenv("AI_GATEWAY_TEST_DATABASE_URL"))
+	if databaseURL == "" {
+		t.Skip("set AI_GATEWAY_TEST_DATABASE_URL to run concurrency lease PostgreSQL integration tests")
+	}
+	ctx, cancel := context.WithTimeout(context.Background(), 15*time.Second)
+	defer cancel()
+	db, err := Connect(ctx, databaseURL)
+	if err != nil {
+		t.Fatalf("connect store: %v", err)
+	}
+	defer db.Close()
+	scopeKey := "renew-" + time.Now().UTC().Format("20060102150405.000000000")
+	taskIDs := createLeaseTestTasks(t, ctx, db, 1, scopeKey)
+	defer deleteLeaseTestTasks(t, db, taskIDs)
+
+	result, err := db.ReserveRateLimits(ctx, taskIDs[0], "", []RateLimitReservation{{
+		ScopeType:       "platform_model",
+		ScopeKey:        scopeKey,
+		Metric:          "concurrent",
+		Limit:           1,
+		Amount:          1,
+		LeaseTTLSeconds: 2,
+	}})
+	if err != nil {
+		t.Fatalf("reserve short lease: %v", err)
+	}
+	time.Sleep(time.Second)
+	if err := db.RenewConcurrencyLeases(ctx, result.Leases); err != nil {
+		t.Fatalf("renew short lease: %v", err)
+	}
+	time.Sleep(1500 * time.Millisecond)
+	var active bool
+	if err := db.Pool().QueryRow(ctx, `
+SELECT EXISTS (
+  SELECT 1 FROM gateway_concurrency_leases
+  WHERE id = $1::uuid AND released_at IS NULL AND expires_at > now()
+)`, result.Leases[0].ID).Scan(&active); err != nil {
+		t.Fatalf("read renewed lease: %v", err)
+	}
+	if !active {
+		t.Fatal("renewed lease expired at its original TTL")
+	}
+	if err := db.ReleaseConcurrencyLeases(ctx, result.Leases); err != nil {
+		t.Fatalf("release renewed lease: %v", err)
+	}
+}
+
+func createLeaseTestTasks(t *testing.T, ctx context.Context, db *Store, count int, marker string) []string {
+	t.Helper()
+	rows, err := db.Pool().Query(ctx, `
+INSERT INTO gateway_tasks (kind, run_mode, user_id, model, model_type, request, status, queue_key)
+SELECT 'lease-test', 'simulation', $2, 'lease-test', 'text_generate', '{}'::jsonb, 'queued', $2
+FROM generate_series(1, $1)
+RETURNING id::text`, count, marker)
+	if err != nil {
+		t.Fatalf("create lease test tasks: %v", err)
+	}
+	defer rows.Close()
+	ids := make([]string, 0, count)
+	for rows.Next() {
+		var id string
+		if err := rows.Scan(&id); err != nil {
+			t.Fatalf("scan lease test task: %v", err)
+		}
+		ids = append(ids, id)
+	}
+	if err := rows.Err(); err != nil {
+		t.Fatalf("create lease test tasks: %v", err)
+	}
+	return ids
+}
+
+func deleteLeaseTestTasks(t *testing.T, db *Store, taskIDs []string) {
+	t.Helper()
+	cleanupCtx, cancel := context.WithTimeout(context.Background(), 10*time.Second)
+	defer cancel()
+	if _, err := db.Pool().Exec(cleanupCtx, `DELETE FROM gateway_tasks WHERE id = ANY($1::uuid[])`, taskIDs); err != nil {
+		t.Errorf("delete lease test tasks: %v", err)
+	}
+}
@@ -111,6 +111,7 @@ type CreatePlatformModelInput struct {
 	PermissionConfig      map[string]any `json:"permissionConfig"`
 	RetryPolicy           map[string]any `json:"retryPolicy"`
 	RateLimitPolicy       map[string]any `json:"rateLimitPolicy"`
+	RateLimitPolicyMode   string         `json:"rateLimitPolicyMode" enums:"inherit,override"`
 	RuntimePolicySetID    string         `json:"runtimePolicySetId"`
 	RuntimePolicyOverride map[string]any `json:"runtimePolicyOverride"`
 	Enabled               bool           `json:"enabled"`
@@ -130,6 +131,7 @@ type RuntimeModelCandidate struct {
 	DefaultDiscountFactor    float64
 	PlatformRetryPolicy      map[string]any
 	PlatformRateLimitPolicy  map[string]any
+	BaseRateLimitPolicy      map[string]any
 	PlatformPriority         int
 	PlatformModelID          string
 	BaseModelID              string
@@ -153,8 +155,11 @@ type RuntimeModelCandidate struct {
 	ModelPricingRuleSetID    string
 	ModelRetryPolicy         map[string]any
 	ModelRateLimitPolicy     map[string]any
+	ModelRateLimitPolicyMode string
 	RuntimePolicySetID       string
+	RuntimePolicyExplicit    bool
 	RuntimePolicyOverride    map[string]any
+	RateLimitRuntimeOverride map[string]any
 	RuntimeRetryPolicy       map[string]any
 	RuntimeRateLimitPolicy   map[string]any
 	AutoDisablePolicy        map[string]any
@@ -219,10 +224,15 @@ type RateLimitReservation struct {
 }

 type RateLimitResult struct {
-	LeaseIDs     []string
+	Leases       []ConcurrencyLease
 	Reservations []RateLimitReservation
 }

+type ConcurrencyLease struct {
+	ID  string
+	TTL time.Duration
+}
+
 type CreateTaskAttemptInput struct {
 	TaskID             string
 	AttemptNo          int
@@ -0,0 +1,58 @@
+ALTER TABLE platform_models
+  ADD COLUMN IF NOT EXISTS rate_limit_policy_mode text;
+
+-- Historical server-main imports wrote {"rules": []} as a placeholder. It
+-- means "no override", not an explicit unlimited policy.
+UPDATE integration_platforms
+SET rate_limit_policy = '{}'::jsonb,
+    updated_at = now()
+WHERE rate_limit_policy = '{"rules":[]}'::jsonb;
+
+UPDATE platform_models
+SET rate_limit_policy = '{}'::jsonb,
+    updated_at = now()
+WHERE rate_limit_policy = '{"rules":[]}'::jsonb;
+
+UPDATE platform_models model
+SET runtime_policy_set_id = NULL,
+    runtime_policy_override = '{}'::jsonb
+FROM base_model_catalog base
+WHERE model.base_model_id = base.id
+  AND model.runtime_policy_set_id IS NOT DISTINCT FROM base.runtime_policy_set_id
+  AND model.runtime_policy_override IS NOT DISTINCT FROM base.runtime_policy_override;
+
+UPDATE platform_models model
+SET rate_limit_policy_mode = CASE
+  WHEN model.runtime_policy_override ? 'rateLimitPolicy' THEN 'override'
+  WHEN model.rate_limit_policy <> '{}'::jsonb
+    AND (
+      base.id IS NULL
+      OR model.rate_limit_policy IS DISTINCT FROM base.default_rate_limit_policy
+    )
+    THEN 'override'
+  ELSE 'inherit'
+END
+FROM (
+  SELECT platform_model.id AS platform_model_id, catalog.id, catalog.default_rate_limit_policy
+  FROM platform_models platform_model
+  LEFT JOIN base_model_catalog catalog ON catalog.id = platform_model.base_model_id
+) base
+WHERE model.id = base.platform_model_id
+  AND model.rate_limit_policy_mode IS NULL;
+
+UPDATE platform_models
+SET rate_limit_policy_mode = 'inherit'
+WHERE rate_limit_policy_mode IS NULL;
+
+ALTER TABLE platform_models
+  ALTER COLUMN rate_limit_policy_mode SET DEFAULT 'inherit';
+
+ALTER TABLE platform_models
+  ADD CONSTRAINT platform_models_rate_limit_policy_mode_check
+  CHECK (
+    COALESCE(rate_limit_policy_mode, '') IN ('inherit', 'override')
+  )
+  NOT VALID;
+
+ALTER TABLE platform_models
+  VALIDATE CONSTRAINT platform_models_rate_limit_policy_mode_check;
@@ -0,0 +1,44 @@
+-- The three migrated image identities previously relied on a global
+-- user-group limit. Give every upstream binding its own canonical concurrency
+-- lease so raising the service group no longer removes provider protection.
+WITH limits(invocation_name, policy) AS (
+  VALUES
+    (
+      'gemini-3-pro-image',
+      '{"rules":[{"metric":"concurrent","limit":10,"leaseTtlSeconds":600}]}'::jsonb
+    ),
+    (
+      'gemini-3.1-flash-image',
+      '{"rules":[{"metric":"concurrent","limit":10,"leaseTtlSeconds":600}]}'::jsonb
+    ),
+    (
+      'gpt-image-2',
+      '{"rules":[{"metric":"concurrent","limit":5,"leaseTtlSeconds":300}]}'::jsonb
+    )
+),
+updated_base_models AS (
+  UPDATE base_model_catalog base_model
+  SET default_rate_limit_policy = limits.policy,
+      metadata = COALESCE(base_model.metadata, '{}'::jsonb) || jsonb_build_object(
+        'rateLimitSource', 'image-gateway-migration',
+        'rateLimitUpdatedAt', '2026-07-24'
+      ),
+      updated_at = now()
+  FROM limits
+  WHERE base_model.invocation_name = limits.invocation_name
+  RETURNING base_model.id
+)
+UPDATE platform_models platform_model
+SET rate_limit_policy = '{}'::jsonb,
+    rate_limit_policy_mode = 'inherit',
+    updated_at = now()
+FROM updated_base_models base_model
+WHERE platform_model.base_model_id = base_model.id
+  AND (
+    platform_model.rate_limit_policy = '{}'::jsonb
+    OR platform_model.rate_limit_policy = '{"rules":[]}'::jsonb
+    OR platform_model.rate_limit_policy ? 'platformLimits'
+    OR platform_model.rate_limit_policy ? 'modelLimits'
+    OR platform_model.rate_limit_policy ? 'platform_limits'
+    OR platform_model.rate_limit_policy ? 'model_limits'
+  );
@@ -1513,6 +1513,7 @@ function mergeExistingPlatformModelInput(input: PlatformModelBindingInput, curre
    discountFactor: (input.discountFactor ?? existing.discountFactor) || undefined,
    pricingRuleSetId: input.pricingRuleSetId ?? existing.pricingRuleSetId,
    rateLimitPolicy: input.rateLimitPolicy ?? existing.rateLimitPolicy,
+    rateLimitPolicyMode: input.rateLimitPolicyMode ?? existing.rateLimitPolicyMode,
    retryPolicy: input.retryPolicy ?? existing.retryPolicy,
    runtimePolicyOverride: input.runtimePolicyOverride ?? (existing.runtimePolicyOverride as Record<string, unknown> | undefined),
    runtimePolicySetId: input.runtimePolicySetId ?? existing.runtimePolicySetId,
@@ -409,13 +409,13 @@ function ModelBindingPolicy(props: { form: PlatformWizardForm; onChange: (value:
          </Label>
        </>
      )}
-      <ToggleField checked={form.modelOverrideRateLimit} label="覆盖模型限流策略" onChange={(checked) => onChange({ ...form, modelOverrideRateLimit: checked })} />
+      <ToggleField checked={form.modelOverrideRateLimit} label="覆盖模型限流策略" onChange={(checked) => onChange({ ...form, modelOverrideRateLimit: checked, modelRateLimitOverrideTouched: true })} />
      {form.modelOverrideRateLimit && (
        <>
-          <Label>模型 RPM<Input value={form.modelRpmLimit} placeholder="不填则不限制" inputMode="numeric" onChange={(event) => onChange({ ...form, modelRpmLimit: event.target.value })} /></Label>
-          <Label>模型 RPS<Input value={form.modelRpsLimit} placeholder="不填则不限制" inputMode="numeric" onChange={(event) => onChange({ ...form, modelRpsLimit: event.target.value })} /></Label>
-          <Label>模型 TPM<Input value={form.modelTpmLimit} placeholder="不填则不限制" inputMode="numeric" onChange={(event) => onChange({ ...form, modelTpmLimit: event.target.value })} /></Label>
-          <Label>模型并发<Input value={form.modelConcurrencyLimit} placeholder="不填则不限制" inputMode="numeric" onChange={(event) => onChange({ ...form, modelConcurrencyLimit: event.target.value })} /></Label>
+          <Label>模型 RPM<Input value={form.modelRpmLimit} placeholder="不填则不限制" inputMode="numeric" onChange={(event) => onChange({ ...form, modelRpmLimit: event.target.value, modelRateLimitOverrideTouched: true })} /></Label>
+          <Label>模型 RPS<Input value={form.modelRpsLimit} placeholder="不填则不限制" inputMode="numeric" onChange={(event) => onChange({ ...form, modelRpsLimit: event.target.value, modelRateLimitOverrideTouched: true })} /></Label>
+          <Label>模型 TPM<Input value={form.modelTpmLimit} placeholder="不填则不限制" inputMode="numeric" onChange={(event) => onChange({ ...form, modelTpmLimit: event.target.value, modelRateLimitOverrideTouched: true })} /></Label>
+          <Label>模型并发<Input value={form.modelConcurrencyLimit} placeholder="不填则不限制" inputMode="numeric" onChange={(event) => onChange({ ...form, modelConcurrencyLimit: event.target.value, modelRateLimitOverrideTouched: true })} /></Label>
        </>
      )}
    </div>
@@ -1004,6 +1004,7 @@ function platformToForm(
  const rateLimitPolicy = platform.rateLimitPolicy ?? {};
  const networkProxy = readNetworkProxyConfig(config);
  const currentModels = platformModels.filter((model) => model.platformId === platform.id);
+  const modelRateLimitPolicy = currentModels.find((model) => model.rateLimitPolicyMode === 'override')?.rateLimitPolicy ?? {};
  return {
    ...createEmptyPlatformForm(platform.provider, defaults),
    provider: platform.provider,
@@ -1034,6 +1035,12 @@ function platformToForm(
    selectedModelIds: platformModelBaseIds(platform, baseModels, currentModels),
    modelDiscountFactors: platformModelDiscountFactors(platform, baseModels, currentModels),
    modelNameMappings: platformModelNameMappings(platform, baseModels, currentModels),
+    modelOverrideRateLimit: currentModels.length > 0 && currentModels.every((model) => model.rateLimitPolicyMode === 'override'),
+    modelRateLimitOverrideTouched: false,
+    modelRpmLimit: readLimit(modelRateLimitPolicy, 'rpm'),
+    modelRpsLimit: readLimit(modelRateLimitPolicy, 'rps'),
+    modelTpmLimit: readLimit(modelRateLimitPolicy, 'tpm_total'),
+    modelConcurrencyLimit: readLimit(modelRateLimitPolicy, 'concurrent'),
    selectionMode: 'partial',
  };
 }
@@ -0,0 +1,56 @@
+import type { BaseModelCatalogItem } from '@easyai-ai-gateway/contracts';
+import { describe, expect, it } from 'vitest';
+import { createEmptyPlatformForm, platformModelPayloads } from './platform-form';
+
+const baseModel = {
+  id: 'model-1',
+  providerKey: 'simulation',
+  canonicalModelKey: 'simulation:model-1',
+  invocationName: 'model-1',
+  providerModelName: 'model-1',
+  modelType: ['video_generate'],
+  modelAlias: 'model-1',
+  displayName: 'Model 1',
+  referenceCount: 0,
+  pricingVersion: 1,
+  status: 'active',
+  createdAt: '2026-07-24T00:00:00Z',
+  updatedAt: '2026-07-24T00:00:00Z',
+} satisfies BaseModelCatalogItem;
+
+function selectedForm() {
+  return {
+    ...createEmptyPlatformForm('simulation'),
+    selectedModelIds: [baseModel.id],
+  };
+}
+
+describe('platform model rate limit payload', () => {
+  it('preserves an existing model policy when the edit form was not touched', () => {
+    const [payload] = platformModelPayloads([baseModel], selectedForm());
+
+    expect(payload.rateLimitPolicyMode).toBeUndefined();
+    expect(payload.rateLimitPolicy).toBeUndefined();
+  });
+
+  it('writes inherit only after the override is explicitly disabled', () => {
+    const [payload] = platformModelPayloads([baseModel], {
+      ...selectedForm(),
+      modelRateLimitOverrideTouched: true,
+    });
+
+    expect(payload.rateLimitPolicyMode).toBe('inherit');
+    expect(payload.rateLimitPolicy).toBeUndefined();
+  });
+
+  it('writes an explicit unlimited override when the enabled policy has no rules', () => {
+    const [payload] = platformModelPayloads([baseModel], {
+      ...selectedForm(),
+      modelOverrideRateLimit: true,
+      modelRateLimitOverrideTouched: true,
+    });
+
+    expect(payload.rateLimitPolicyMode).toBe('override');
+    expect(payload.rateLimitPolicy).toEqual({ rules: [] });
+  });
+});
@@ -41,6 +41,7 @@ export interface PlatformWizardForm {
  modelRetryEnabled: boolean;
  modelRetryMaxAttempts: string;
  modelOverrideRateLimit: boolean;
+  modelRateLimitOverrideTouched: boolean;
  modelRpmLimit: string;
  modelRpsLimit: string;
  modelTpmLimit: string;
@@ -100,6 +101,7 @@ export function createEmptyPlatformForm(provider = '', defaults?: ProviderConnec
    modelRetryEnabled: true,
    modelRetryMaxAttempts: '2',
    modelOverrideRateLimit: false,
+    modelRateLimitOverrideTouched: false,
    modelRpmLimit: '',
    modelRpsLimit: '',
    modelTpmLimit: '',
@@ -187,12 +189,15 @@ export function platformModelPayloads(models: BaseModelCatalogItem[], form: Plat
      retryPolicy: form.modelOverrideRetry
        ? { enabled: form.modelRetryEnabled, maxAttempts: form.modelRetryEnabled ? positiveInt(form.modelRetryMaxAttempts, 2) : 1 }
        : undefined,
-      rateLimitPolicy: form.modelOverrideRateLimit ? rateLimitPolicyPayload({
-        rpmLimit: form.modelRpmLimit,
-        rpsLimit: form.modelRpsLimit,
-        tpmLimit: form.modelTpmLimit,
-        concurrencyLimit: form.modelConcurrencyLimit,
-      }) : undefined,
+      rateLimitPolicy: form.modelRateLimitOverrideTouched && form.modelOverrideRateLimit ? rateLimitPolicyPayload({
+          rpmLimit: form.modelRpmLimit,
+          rpsLimit: form.modelRpsLimit,
+          tpmLimit: form.modelTpmLimit,
+          concurrencyLimit: form.modelConcurrencyLimit,
+        }) : undefined,
+      rateLimitPolicyMode: form.modelRateLimitOverrideTouched
+        ? (form.modelOverrideRateLimit ? 'override' : 'inherit')
+        : undefined,
      runtimePolicyOverride: platformModelRuntimeOverride(form),
    };
  });
@@ -304,14 +309,6 @@ function platformModelRuntimeOverride(form: PlatformWizardForm) {
      maxAttempts: form.modelRetryEnabled ? positiveInt(form.modelRetryMaxAttempts, 2) : 1,
    };
  }
-  if (form.modelOverrideRateLimit) {
-    override.rateLimitPolicy = rateLimitPolicyPayload({
-      rpmLimit: form.modelRpmLimit,
-      rpsLimit: form.modelRpsLimit,
-      tpmLimit: form.modelTpmLimit,
-      concurrencyLimit: form.modelConcurrencyLimit,
-    });
-  }
  return Object.keys(override).length ? override : undefined;
 }

@@ -152,6 +152,7 @@ export interface PlatformModelBindingInput {
  pricingMode?: string;
  retryPolicy?: Record<string, unknown>;
  rateLimitPolicy?: Record<string, unknown>;
+  rateLimitPolicyMode?: 'inherit' | 'override';
  runtimePolicySetId?: string;
  runtimePolicyOverride?: Record<string, unknown>;
  pricingRuleSetId?: string;
@@ -723,6 +723,7 @@ export interface PlatformModel {
  permissionConfig?: Record<string, unknown>;
  retryPolicy?: Record<string, unknown>;
  rateLimitPolicy?: RateLimitPolicy;
+  rateLimitPolicyMode: 'inherit' | 'override';
  runtimePolicySetId?: string;
  runtimePolicyOverride?: RuntimePolicyOverride;
  cooldownUntil?: string;