perf(queue): 限制大媒体任务内存并消除准入惊群

将同步与异步非文本任务的准入唤醒改为按任务和全局队首推进，批量续租等待者，避免大量请求同时争抢 advisory lock 和数据库连接。

对 Base64 请求解析、素材解码、上游媒体执行和结果物化增加分层并发限制，复用请求素材并释放重复 Gemini wire 数据；生产 API 默认入口解析并发 16，媒体物化并发 8。

新增 1000 个同步 Gemini 图像编辑请求的模拟上游压力验收，10 MiB 输入和输出下全部成功，Heap 峰值增长约 2.58 GiB，并验证共享上传哈希、单次 attempt 与本地零落盘。

验证：go test ./... -count=1；go vet ./...；gofmt -l 无输出；kubectl kustomize deploy/kubernetes/production；10 MiB Gemini Base64 千任务压力测试通过。

apps/api/internal/clients/clients_test.go

@@ -1531,6 +1531,16 @@ func TestGeminiClientImageGenerateBuildsNativeImageBody(t *testing.T) {
 	if data[0].(map[string]any)["b64_json"] != "aW1hZ2U=" {
 		t.Fatalf("unexpected image response: %+v", response.Result)
 	}
+	if response.Wire == nil || len(response.Wire.RawJSON) != 0 {
+		t.Fatalf("Gemini image response retained duplicate raw JSON: %+v", response.Wire)
+	}
+	candidates, _ := response.Wire.Body["candidates"].([]any)
+	content, _ := candidates[0].(map[string]any)["content"].(map[string]any)
+	wireParts, _ := content["parts"].([]any)
+	inline, _ := wireParts[0].(map[string]any)["inlineData"].(map[string]any)
+	if inline["data"] != "aW1hZ2U=" {
+		t.Fatalf("decoded Gemini wire body lost Base64 output: %+v", response.Wire.Body)
+	}
 }
 
 func TestGeminiClientImageGenerateRejectsPromptBlockedWithoutImage(t *testing.T) {

apps/api/internal/clients/gemini.go

@@ -65,6 +65,13 @@ func (c GeminiClient) Run(ctx context.Context, request Request) (Response, error
 		}
 		return Response{}, annotateResponseError(err, requestID, responseStartedAt, responseFinishedAt)
 	}
+	if wire != nil && geminiImageModelType(request.ModelType) {
+		// Gemini image responses can contain multi-megabyte Base64 fields. The
+		// decoded wire body is still used for native synchronous passthrough, so
+		// retaining the original JSON bytes would keep a second full copy alive
+		// until the client response finishes.
+		wire.RawJSON = nil
+	}
 	if requestID == "" {
 		requestID = requestIDFromResult(output)
 	}

apps/api/internal/clients/helpers.go

@@ -65,12 +65,14 @@ func decodeHTTPResponseForProtocol(resp *http.Response, protocol string) (map[st
 		Protocol:   strings.TrimSpace(protocol),
 		StatusCode: resp.StatusCode,
 		Headers:    compatibleResponseHeaders(resp.Header),
-		RawJSON:    append([]byte(nil), raw...),
-	}
-	if len(raw) > 0 {
-		_ = json.Unmarshal(raw, &wire.Body)
+		// raw is already owned by this response. Reusing it avoids retaining a
+		// second copy of media responses that may be as large as 128 MiB.
+		RawJSON: raw,
 	}
 	if readErr != nil {
+		if len(raw) > 0 {
+			_ = json.Unmarshal(raw, &wire.Body)
+		}
 		return nil, wire, &ClientError{
 			Code:       "response_read_error",
 			Message:    readErr.Error(),
@@ -81,6 +83,9 @@ func decodeHTTPResponseForProtocol(resp *http.Response, protocol string) (map[st
 		}
 	}
 	if tooLarge {
+		if len(raw) > 0 {
+			_ = json.Unmarshal(raw, &wire.Body)
+		}
 		return nil, wire, &ClientError{
 			Code:       "response_too_large",
 			Message:    fmt.Sprintf("upstream JSON response exceeds %d bytes", maxBytes),
@@ -91,6 +96,9 @@ func decodeHTTPResponseForProtocol(resp *http.Response, protocol string) (map[st
 		}
 	}
 	if resp.StatusCode < 200 || resp.StatusCode >= 300 {
+		if len(raw) > 0 {
+			_ = json.Unmarshal(raw, &wire.Body)
+		}
 		return nil, wire, &ClientError{
 			Code:       statusCodeName(resp.StatusCode),
 			Message:    errorMessage(raw, resp.Status),
@@ -108,7 +116,10 @@ func decodeHTTPResponseForProtocol(resp *http.Response, protocol string) (map[st
 	if err := json.Unmarshal(raw, &out); err != nil {
 		return nil, wire, &ClientError{Code: "invalid_response", Message: err.Error(), StatusCode: resp.StatusCode, Retryable: false, Wire: wire}
 	}
-	wire.Body = cloneBody(out)
+	// Result and wire metadata are read-only until the runner replaces the
+	// canonical result, so they can share the decoded object instead of keeping
+	// another full copy of a large Base64 media response.
+	wire.Body = out
 	return out, wire, nil
 }
 

apps/api/internal/config/config.go

@@ -59,6 +59,8 @@ type Config struct {
 	BillingEngineMode                             string
 	ProcessRole                                   string
 	DatabaseMaxConns                              int
+	MediaRequestConcurrency                       int
+	MediaMaterializationConcurrency               int
 	AsyncQueueWorkerEnabled                       bool
 	AsyncWorkerHardLimit                          int
 	AsyncWorkerRefreshIntervalSeconds             int
@@ -117,6 +119,8 @@ func Load() Config {
 		BillingEngineMode:                 strings.ToLower(env("BILLING_ENGINE_MODE", "observe")),
 		ProcessRole:                       strings.ToLower(strings.TrimSpace(env("AI_GATEWAY_PROCESS_ROLE", "all"))),
 		DatabaseMaxConns:                  envInt("AI_GATEWAY_DATABASE_MAX_CONNS", 0),
+		MediaRequestConcurrency:           envIntValidated("AI_GATEWAY_MEDIA_REQUEST_CONCURRENCY", 16),
+		MediaMaterializationConcurrency:   envIntValidated("AI_GATEWAY_MEDIA_MATERIALIZATION_CONCURRENCY", 8),
 		AsyncQueueWorkerEnabled:           env("AI_GATEWAY_ASYNC_QUEUE_WORKER_ENABLED", "true") == "true",
 		AsyncWorkerHardLimit:              envIntValidated("AI_GATEWAY_ASYNC_WORKER_HARD_LIMIT", 2048),
 		AsyncWorkerRefreshIntervalSeconds: envIntValidated("AI_GATEWAY_ASYNC_WORKER_REFRESH_INTERVAL_SECONDS", 5),
@@ -132,6 +136,12 @@ func (c Config) Validate() error {
 	if c.DatabaseMaxConns < 0 || c.DatabaseMaxConns > 1000 {
 		return errors.New("AI_GATEWAY_DATABASE_MAX_CONNS must be between 1 and 1000 when configured")
 	}
+	if c.MediaMaterializationConcurrency != 0 && (c.MediaMaterializationConcurrency < 1 || c.MediaMaterializationConcurrency > 256) {
+		return errors.New("AI_GATEWAY_MEDIA_MATERIALIZATION_CONCURRENCY must be between 1 and 256")
+	}
+	if c.MediaRequestConcurrency != 0 && (c.MediaRequestConcurrency < 1 || c.MediaRequestConcurrency > 1024) {
+		return errors.New("AI_GATEWAY_MEDIA_REQUEST_CONCURRENCY must be between 1 and 1024")
+	}
 	switch strings.ToLower(strings.TrimSpace(c.BillingEngineMode)) {
 	case "", "observe", "enforce", "hold":
 	default:

apps/api/internal/config/config_test.go

@@ -131,6 +131,24 @@ func TestValidateProcessRoleAndDatabasePool(t *testing.T) {
 	if err := cfg.Validate(); err == nil || !strings.Contains(err.Error(), "DATABASE_MAX_CONNS") {
 		t.Fatalf("Validate() error = %v, want invalid database max conns", err)
 	}
+	cfg.DatabaseMaxConns = 16
+	cfg.MediaMaterializationConcurrency = 257
+	if err := cfg.Validate(); err == nil || !strings.Contains(err.Error(), "MEDIA_MATERIALIZATION_CONCURRENCY") {
+		t.Fatalf("Validate() error = %v, want invalid media materialization concurrency", err)
+	}
+	t.Setenv("AI_GATEWAY_MEDIA_MATERIALIZATION_CONCURRENCY", "12")
+	if got := Load().MediaMaterializationConcurrency; got != 12 {
+		t.Fatalf("media materialization concurrency = %d, want 12", got)
+	}
+	cfg = Load()
+	cfg.MediaRequestConcurrency = 1025
+	if err := cfg.Validate(); err == nil || !strings.Contains(err.Error(), "MEDIA_REQUEST_CONCURRENCY") {
+		t.Fatalf("Validate() error = %v, want invalid media request concurrency", err)
+	}
+	t.Setenv("AI_GATEWAY_MEDIA_REQUEST_CONCURRENCY", "96")
+	if got := Load().MediaRequestConcurrency; got != 96 {
+		t.Fatalf("media request concurrency = %d, want 96", got)
+	}
 }
 
 func TestValidateTaskHistorySettings(t *testing.T) {

apps/api/internal/httpapi/async_worker_acceptance_integration_test.go

@@ -21,6 +21,173 @@ import (
 	"github.com/jackc/pgx/v5/pgxpool"
 )
 
+func TestAsyncSubmissionReservesConcurrencyOnlyAfterWorkerClaim(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 async admission timing tests")
+	}
+	ctx, cancel := context.WithTimeout(context.Background(), 90*time.Second)
+	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:                 "*",
+		AsyncQueueWorkerEnabled:           true,
+		AsyncWorkerHardLimit:              1,
+		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)
+	model := "worker-claim-admission-" + suffix
+	platform := createAsyncAcceptancePlatform(t, server.URL, adminToken, "claim-admission-"+suffix, "Claim Admission Simulation", 3)
+	defer updateAsyncAcceptancePlatform(t, server.URL, adminToken, platform, 3, "disabled")
+	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)
+	taskIDs := submitAsyncSimulationTasks(t, server.URL, adminToken, "/api/v1/videos/generations", model, 3, 5*time.Second)
+	waitForTaskCount(t, ctx, pool, taskIDs, "running", 1, 10*time.Second)
+	waitForActiveLeaseCount(t, ctx, pool, modelID, 1, 10*time.Second)
+
+	var admitted, attempts int
+	if err := pool.QueryRow(ctx, `
+SELECT
+  (SELECT COUNT(*)
+   FROM gateway_task_admissions
+   WHERE task_id = ANY($1::uuid[]) AND status = 'admitted'),
+  (SELECT COUNT(*)
+   FROM gateway_task_attempts
+   WHERE task_id = ANY($1::uuid[]))`, taskIDs).Scan(&admitted, &attempts); err != nil {
+		t.Fatalf("read queued admission state: %v", err)
+	}
+	if admitted != 1 || attempts != 1 {
+		t.Fatalf("claimed admission state admitted=%d attempts=%d, want 1/1", admitted, attempts)
+	}
+
+	waitForTaskCount(t, ctx, pool, taskIDs, "succeeded", len(taskIDs), 30*time.Second)
+	assertAcceptanceAttempts(t, ctx, pool, taskIDs, len(taskIDs))
+}
+
+func TestAsyncWorkerThousandConcurrentSchedulingAcceptance(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 thousand-task worker acceptance tests")
+	}
+	ctx, cancel := context.WithTimeout(context.Background(), 3*time.Minute)
+	defer cancel()
+	applyMigration(t, ctx, databaseURL)
+
+	apiDB, err := store.ConnectWithMaxConns(ctx, databaseURL, 16)
+	if err != nil {
+		t.Fatalf("connect API store: %v", err)
+	}
+	defer apiDB.Close()
+	workerDB, err := store.ConnectWithMaxConns(ctx, databaseURL, 24)
+	if err != nil {
+		t.Fatalf("connect Worker store: %v", err)
+	}
+	defer workerDB.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:                 "*",
+		ProcessRole:                       "api",
+		AsyncQueueWorkerEnabled:           true,
+		AsyncWorkerHardLimit:              1000,
+		AsyncWorkerRefreshIntervalSeconds: 1,
+	}, apiDB, slog.New(slog.NewTextHandler(os.Stderr, &slog.HandlerOptions{Level: slog.LevelError}))))
+	defer server.Close()
+	workerServer := httptest.NewServer(NewServerWithContext(serverCtx, config.Config{
+		AppEnv:                            "test",
+		HTTPAddr:                          ":0",
+		DatabaseURL:                       databaseURL,
+		IdentityMode:                      "hybrid",
+		JWTSecret:                         "test-secret",
+		BillingEngineMode:                 "observe",
+		CORSAllowedOrigin:                 "*",
+		ProcessRole:                       "worker",
+		AsyncQueueWorkerEnabled:           true,
+		AsyncWorkerHardLimit:              1000,
+		AsyncWorkerRefreshIntervalSeconds: 1,
+	}, workerDB, slog.New(slog.NewTextHandler(os.Stderr, &slog.HandlerOptions{Level: slog.LevelError}))))
+	defer workerServer.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":1024,"leaseTtlSeconds":120}]}'::jsonb
+WHERE status = 'active'`); err != nil {
+		t.Fatalf("raise acceptance user-group concurrency: %v", err)
+	}
+	suffix := strconv.FormatInt(time.Now().UnixNano(), 10)
+	model := "worker-thousand-" + suffix
+	platform := createAsyncAcceptancePlatform(t, server.URL, adminToken, "thousand-"+suffix, "Thousand Task Simulation", 1000)
+	defer updateAsyncAcceptancePlatform(t, server.URL, adminToken, platform, 1000, "disabled")
+	modelID := createAsyncAcceptanceModel(t, server.URL, adminToken, platform.ID, model, "video_generate", "override", 1000, 120)
+
+	waitForAsyncWorkerMetric(t, workerServer.URL, "easyai_gateway_async_worker_capacity", 1000, 15*time.Second)
+	taskIDs := submitAsyncSimulationTasks(t, server.URL, adminToken, "/api/v1/videos/generations", model, 1000, 45*time.Second)
+	peakRunning, peakLeases := waitForAcceptanceTasks(t, ctx, pool, taskIDs, modelID, 2*time.Minute)
+	if peakRunning < 800 || peakLeases < 800 {
+		t.Fatalf("thousand-task peak running=%d leases=%d, want both >=800", peakRunning, peakLeases)
+	}
+	if peakLeases > 1000 {
+		t.Fatalf("thousand-task lease peak=%d exceeded model concurrent limit 1000", peakLeases)
+	}
+	assertAcceptanceAttempts(t, ctx, pool, taskIDs, len(taskIDs))
+	t.Logf(
+		"千级并发证据: tasks=%d duration=%s running_peak=%d lease_peak=%d api_database_max_connections=16 worker_database_max_connections=24",
+		len(taskIDs),
+		45*time.Second,
+		peakRunning,
+		peakLeases,
+	)
+}
+
 func TestAsyncWorkerDynamicConcurrencyAcceptance(t *testing.T) {
 	databaseURL := strings.TrimSpace(os.Getenv("AI_GATEWAY_TEST_DATABASE_URL"))
 	if databaseURL == "" {
@@ -295,6 +462,7 @@ func waitForAcceptanceTasks(t *testing.T, ctx context.Context, pool *pgxpool.Poo
 	t.Helper()
 	deadline := time.Now().Add(timeout)
 	var peakRunning, peakLeases int64
+	var lastRunning, lastSucceeded, lastFailed, lastLeases int64
 	for time.Now().Before(deadline) {
 		var running, succeeded, failed, leases int64
 		if err := pool.QueryRow(ctx, `
@@ -316,6 +484,10 @@ WHERE scope_type = 'platform_model'
 		}
 		peakRunning = maxInt64(peakRunning, running)
 		peakLeases = maxInt64(peakLeases, leases)
+		lastRunning = running
+		lastSucceeded = succeeded
+		lastFailed = failed
+		lastLeases = leases
 		if failed > 0 {
 			t.Fatalf("acceptance tasks entered failed/manual status: %d", failed)
 		}
@@ -324,7 +496,17 @@ WHERE scope_type = 'platform_model'
 		}
 		time.Sleep(50 * time.Millisecond)
 	}
-	t.Fatalf("tasks did not finish within %s", timeout)
+	t.Fatalf(
+		"tasks did not finish within %s: running=%d succeeded=%d failed=%d leases=%d peak_running=%d peak_leases=%d total=%d",
+		timeout,
+		lastRunning,
+		lastSucceeded,
+		lastFailed,
+		lastLeases,
+		peakRunning,
+		peakLeases,
+		len(taskIDs),
+	)
 	return 0, 0
 }
 

apps/api/internal/httpapi/gemini_base64_stress_integration_test.go

@@ -0,0 +1,753 @@
+package httpapi
+
+import (
+	"bytes"
+	"context"
+	"crypto/sha256"
+	"encoding/base64"
+	"encoding/hex"
+	"encoding/json"
+	"fmt"
+	"io"
+	"log/slog"
+	"net/http"
+	"net/http/httptest"
+	"os"
+	"path/filepath"
+	"runtime"
+	"strconv"
+	"strings"
+	"sync"
+	"sync/atomic"
+	"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 TestGeminiBase64ImageEditThousandSynchronousRequests(t *testing.T) {
+	if os.Getenv("AI_GATEWAY_RUN_GEMINI_BASE64_STRESS") != "1" {
+		t.Skip("set AI_GATEWAY_RUN_GEMINI_BASE64_STRESS=1 to run the Gemini Base64 stress acceptance")
+	}
+	databaseURL := strings.TrimSpace(os.Getenv("AI_GATEWAY_TEST_DATABASE_URL"))
+	if databaseURL == "" {
+		t.Skip("set AI_GATEWAY_TEST_DATABASE_URL to run the Gemini Base64 stress acceptance")
+	}
+
+	const (
+		taskCount          = 1000
+		upstreamConcurrent = 64
+		inputVariants      = 16
+		mediaConcurrent    = 8
+	)
+	imageBytes := stressEnvInt(t, "AI_GATEWAY_GEMINI_STRESS_IMAGE_BYTES", 256<<10)
+	upstreamDelay := time.Duration(stressEnvInt(t, "AI_GATEWAY_GEMINI_STRESS_UPSTREAM_DELAY_MS", 4000)) * time.Millisecond
+	maxHeapGrowth := uint64(stressEnvInt(t, "AI_GATEWAY_GEMINI_STRESS_MAX_HEAP_BYTES", 1<<30))
+	clientConnections := stressEnvInt(t, "AI_GATEWAY_GEMINI_STRESS_CLIENT_CONNECTIONS", 256)
+	testTimeout := time.Duration(stressEnvInt(t, "AI_GATEWAY_GEMINI_STRESS_TIMEOUT_SECONDS", 1200)) * time.Second
+
+	ctx, cancel := context.WithTimeout(context.Background(), testTimeout)
+	defer cancel()
+	applyMigration(t, ctx, databaseURL)
+
+	apiDB, err := store.ConnectWithMaxConns(ctx, databaseURL, 16)
+	if err != nil {
+		t.Fatalf("connect API store: %v", err)
+	}
+	defer apiDB.Close()
+	poolConfig, err := pgxpool.ParseConfig(databaseURL)
+	if err != nil {
+		t.Fatalf("parse observation pool config: %v", err)
+	}
+	poolConfig.MaxConns = 4
+	pool, err := pgxpool.NewWithConfig(ctx, poolConfig)
+	if err != nil {
+		t.Fatalf("connect observation pool: %v", err)
+	}
+	defer pool.Close()
+	if _, err := pool.Exec(ctx, `
+UPDATE integration_platforms SET status = 'disabled' WHERE deleted_at IS NULL;
+UPDATE file_storage_channels SET status = 'disabled' WHERE deleted_at IS NULL;
+UPDATE gateway_user_groups
+SET rate_limit_policy = '{"rules":[
+  {"metric":"concurrent","limit":1024,"leaseTtlSeconds":120},
+  {"metric":"queue_size","limit":1000,"maxWaitSeconds":900}
+]}'::jsonb
+WHERE status = 'active'`); err != nil {
+		t.Fatalf("isolate Gemini stress acceptance: %v", err)
+	}
+
+	inputPayloads := make([][]byte, inputVariants)
+	inputBase64 := make([]string, inputVariants)
+	validInputs := make(map[string]struct{}, inputVariants)
+	validInputHashes := make(map[string]struct{}, inputVariants)
+	for index := range inputPayloads {
+		inputPayloads[index] = stressImagePayload(imageBytes, byte(index+1))
+		inputBase64[index] = base64.StdEncoding.EncodeToString(inputPayloads[index])
+		validInputs[inputBase64[index]] = struct{}{}
+		validInputHashes[stressPayloadSHA256(inputPayloads[index])] = struct{}{}
+	}
+	outputPayload := stressImagePayload(imageBytes, 0xf3)
+	outputBase64 := base64.StdEncoding.EncodeToString(outputPayload)
+	outputHash := stressPayloadSHA256(outputPayload)
+
+	uploadService := newGeminiStressUploadService(imageBytes, validInputHashes, outputHash)
+	uploadServer := httptest.NewServer(uploadService)
+	defer uploadServer.Close()
+
+	var upstreamActive atomic.Int64
+	var upstreamPeak atomic.Int64
+	var upstreamCalls atomic.Int64
+	var invalidInputs atomic.Int64
+	upstream := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		upstreamCalls.Add(1)
+		active := upstreamActive.Add(1)
+		defer upstreamActive.Add(-1)
+		updateAtomicPeak(&upstreamPeak, active)
+
+		var body map[string]any
+		if err := json.NewDecoder(r.Body).Decode(&body); err != nil {
+			invalidInputs.Add(1)
+			http.Error(w, err.Error(), http.StatusBadRequest)
+			return
+		}
+		inlineData := geminiStressInlineData(body)
+		if strings.HasPrefix(inlineData, "data:") {
+			_, inlineData, _ = strings.Cut(inlineData, ",")
+		}
+		if _, ok := validInputs[inlineData]; !ok {
+			invalidInputs.Add(1)
+			http.Error(w, "missing or altered Gemini inlineData", http.StatusBadRequest)
+			return
+		}
+		time.Sleep(upstreamDelay)
+		w.Header().Set("Content-Type", "application/json")
+		_ = json.NewEncoder(w).Encode(map[string]any{
+			"candidates": []any{map[string]any{
+				"index": 0,
+				"content": map[string]any{
+					"role": "model",
+					"parts": []any{map[string]any{
+						"inlineData": map[string]any{
+							"mimeType": "image/png",
+							"data":     outputBase64,
+						},
+					}},
+				},
+				"finishReason": "STOP",
+			}},
+			"usageMetadata": map[string]any{
+				"promptTokenCount":     1,
+				"candidatesTokenCount": 1,
+				"totalTokenCount":      2,
+			},
+		})
+	}))
+	defer upstream.Close()
+
+	serverCtx, cancelServer := context.WithCancel(ctx)
+	defer cancelServer()
+	storageRoot := t.TempDir()
+	server := httptest.NewServer(NewServerWithContext(serverCtx, config.Config{
+		AppEnv:                          "test",
+		HTTPAddr:                        ":0",
+		DatabaseURL:                     databaseURL,
+		DatabaseMaxConns:                16,
+		IdentityMode:                    "hybrid",
+		JWTSecret:                       "gemini-base64-stress-secret",
+		BillingEngineMode:               "observe",
+		CORSAllowedOrigin:               "*",
+		ProcessRole:                     "api",
+		LocalUploadedStorageDir:         filepath.Join(storageRoot, "uploaded"),
+		LocalGeneratedStorageDir:        filepath.Join(storageRoot, "generated"),
+		MediaRequestConcurrency:         16,
+		MediaMaterializationConcurrency: mediaConcurrent,
+		AsyncQueueWorkerEnabled:         false,
+		AsyncWorkerHardLimit:            64,
+	}, apiDB, slog.New(slog.NewTextHandler(io.Discard, nil))))
+	defer server.Close()
+
+	adminToken := createAsyncAcceptanceAdmin(t, ctx, pool, server.URL)
+	suffix := strconv.FormatInt(time.Now().UnixNano(), 10)
+	uploadAPIKey := "local-gemini-stress-upload-key"
+	var storageChannel struct {
+		ID string `json:"id"`
+	}
+	doJSON(t, server.URL, http.MethodPost, "/api/admin/system/file-storage/channels", adminToken, map[string]any{
+		"channelKey": "gemini-base64-stress-" + suffix,
+		"name":       "Gemini Base64 Stress Shared Upload",
+		"provider":   "server_main_openapi",
+		"uploadUrl":  uploadServer.URL + "/upload",
+		"apiKey":     uploadAPIKey,
+		"scenes": []string{
+			store.FileStorageSceneUpload,
+			store.FileStorageSceneRequestAsset,
+			store.FileStorageSceneImageResult,
+		},
+		"priority": 1,
+		"status":   "enabled",
+	}, http.StatusCreated, &storageChannel)
+	uploadService.setAPIKey(uploadAPIKey)
+
+	var gatewayUserID string
+	if err := pool.QueryRow(ctx, `
+SELECT id::text
+FROM gateway_users
+WHERE username LIKE 'async_acceptance_%'
+ORDER BY created_at DESC
+LIMIT 1`).Scan(&gatewayUserID); err != nil {
+		t.Fatalf("read Gemini stress user: %v", err)
+	}
+	doJSON(t, server.URL, http.MethodPatch, "/api/admin/users/"+gatewayUserID+"/wallet", adminToken, map[string]any{
+		"currency": "resource",
+		"balance":  1000000000,
+		"reason":   "seed isolated Gemini Base64 stress wallet",
+	}, http.StatusOK, &map[string]any{})
+	model := "gemini-base64-image-edit-stress-" + suffix
+	var baseModel struct {
+		ID string `json:"id"`
+	}
+	doJSON(t, server.URL, http.MethodPost, "/api/admin/catalog/base-models", adminToken, map[string]any{
+		"providerKey":       "gemini",
+		"canonicalModelKey": "gemini:" + model,
+		"invocationName":    model,
+		"providerModelName": model,
+		"modelType":         []string{"image_edit"},
+		"displayName":       model,
+		"status":            "active",
+		"capabilities": map[string]any{
+			"image_edit": map[string]any{
+				"support_base64_input": true,
+				"support_url_input":    false,
+			},
+		},
+		"baseBillingConfig": map[string]any{},
+	}, http.StatusCreated, &baseModel)
+
+	var platform struct {
+		ID string `json:"id"`
+	}
+	doJSON(t, server.URL, http.MethodPost, "/api/admin/platforms", adminToken, map[string]any{
+		"provider":    "gemini",
+		"platformKey": "gemini-base64-stress-" + suffix,
+		"name":        "Gemini Base64 Stress",
+		"baseUrl":     upstream.URL,
+		"authType":    "bearer",
+		"credentials": map[string]any{"apiKey": "local-stress-key"},
+		"priority":    1,
+		"status":      "enabled",
+		"rateLimitPolicy": map[string]any{"rules": []any{
+			map[string]any{"metric": "concurrent", "limit": upstreamConcurrent, "leaseTtlSeconds": 120},
+			map[string]any{"metric": "queue_size", "limit": 1000, "maxWaitSeconds": 900},
+		}},
+	}, http.StatusCreated, &platform)
+	var platformModel struct {
+		ID string `json:"id"`
+	}
+	doJSON(t, server.URL, http.MethodPost, "/api/admin/platforms/"+platform.ID+"/models", adminToken, map[string]any{
+		"baseModelId":         baseModel.ID,
+		"providerModelName":   model,
+		"modelAlias":          model,
+		"modelType":           []string{"image_edit"},
+		"rateLimitPolicyMode": "override",
+		"rateLimitPolicy": map[string]any{"rules": []any{
+			map[string]any{"metric": "concurrent", "limit": upstreamConcurrent, "leaseTtlSeconds": 120},
+			map[string]any{"metric": "queue_size", "limit": 1000, "maxWaitSeconds": 900},
+		}},
+		"capabilityOverride": map[string]any{
+			"image_edit": map[string]any{
+				"support_base64_input": true,
+				"support_url_input":    false,
+			},
+		},
+	}, http.StatusCreated, &platformModel)
+
+	requestBodies := make([][]byte, inputVariants)
+	for index := range requestBodies {
+		requestBodies[index], err = json.Marshal(map[string]any{
+			"contents": []any{map[string]any{
+				"role": "user",
+				"parts": []any{
+					map[string]any{"text": "把输入图片改成用于并发验收的蓝色赛博朋克风格"},
+					map[string]any{"inlineData": map[string]any{
+						"mimeType": "image/png",
+						"data":     inputBase64[index],
+					}},
+				},
+			}},
+			"generationConfig": map[string]any{
+				"responseModalities": []any{"IMAGE"},
+			},
+		})
+		if err != nil {
+			t.Fatalf("encode stress request %d: %v", index, err)
+		}
+	}
+
+	runtime.GC()
+	var baseline runtime.MemStats
+	runtime.ReadMemStats(&baseline)
+	var peakHeap atomic.Uint64
+	peakHeap.Store(baseline.HeapAlloc)
+	var queuePeak atomic.Int64
+	var runningPeak atomic.Int64
+	samplingDone := make(chan struct{})
+	go sampleGeminiStressPressure(ctx, pool, model, &peakHeap, &queuePeak, &runningPeak, samplingDone)
+
+	startedAt := time.Now()
+	loadClient := &http.Client{
+		Timeout: testTimeout,
+		Transport: &http.Transport{
+			MaxIdleConns:        clientConnections,
+			MaxIdleConnsPerHost: clientConnections,
+			MaxConnsPerHost:     clientConnections,
+		},
+	}
+	var wg sync.WaitGroup
+	var requestErrors atomic.Int64
+	errs := make(chan error, 20)
+	recordRequestError := func(err error) {
+		requestErrors.Add(1)
+		select {
+		case errs <- err:
+		default:
+		}
+	}
+	for index := 0; index < taskCount; index++ {
+		wg.Add(1)
+		go func(index int) {
+			defer wg.Done()
+			req, err := http.NewRequestWithContext(ctx, http.MethodPost,
+				server.URL+"/v1beta/models/"+model+":generateContent",
+				bytes.NewReader(requestBodies[index%inputVariants]))
+			if err != nil {
+				recordRequestError(err)
+				return
+			}
+			req.Header.Set("Authorization", "Bearer "+adminToken)
+			req.Header.Set("Content-Type", "application/json")
+			resp, err := loadClient.Do(req)
+			if err != nil {
+				recordRequestError(fmt.Errorf("request %d: %w", index, err))
+				return
+			}
+			if resp.StatusCode != http.StatusOK {
+				body, _ := io.ReadAll(io.LimitReader(resp.Body, 512))
+				_ = resp.Body.Close()
+				recordRequestError(fmt.Errorf("request %d status=%d body=%s", index, resp.StatusCode, stressBodyPreview(body)))
+				return
+			}
+			foundOutput, readErr := stressStreamContainsAll(resp.Body,
+				[]byte(outputBase64[:128]),
+				[]byte(outputBase64[len(outputBase64)-128:]),
+			)
+			_ = resp.Body.Close()
+			if readErr != nil {
+				recordRequestError(fmt.Errorf("request %d read response: %w", index, readErr))
+				return
+			}
+			if !foundOutput {
+				recordRequestError(fmt.Errorf("request %d Gemini response did not contain Base64 image output", index))
+			}
+		}(index)
+	}
+	wg.Wait()
+	close(samplingDone)
+	close(errs)
+	for err := range errs {
+		t.Error(err)
+	}
+	if requestErrors.Load() > 20 {
+		t.Errorf("%d additional request failures omitted", requestErrors.Load()-20)
+	}
+	if t.Failed() {
+		t.FailNow()
+	}
+	elapsed := time.Since(startedAt)
+
+	var tasks, succeeded, attempts, duplicateAttemptTasks int
+	if err := pool.QueryRow(ctx, `
+SELECT
+  COUNT(*)::int,
+  COUNT(*) FILTER (WHERE task.status = 'succeeded')::int,
+  COALESCE(SUM(attempts.attempt_count), 0)::int,
+  COUNT(*) FILTER (WHERE attempts.attempt_count <> 1)::int
+FROM gateway_tasks task
+LEFT JOIN LATERAL (
+  SELECT COUNT(*)::int AS attempt_count
+  FROM gateway_task_attempts attempt
+  WHERE attempt.task_id = task.id
+) attempts ON true
+WHERE task.model = $1
+  AND task.created_at >= $2`, model, startedAt).Scan(&tasks, &succeeded, &attempts, &duplicateAttemptTasks); err != nil {
+		t.Fatalf("read Gemini stress task results: %v", err)
+	}
+	if tasks != taskCount || succeeded != taskCount || attempts != taskCount || duplicateAttemptTasks != 0 {
+		t.Fatalf("task results tasks=%d succeeded=%d attempts=%d duplicate_attempt_tasks=%d", tasks, succeeded, attempts, duplicateAttemptTasks)
+	}
+	if upstreamCalls.Load() != taskCount || invalidInputs.Load() != 0 {
+		t.Fatalf("upstream calls=%d invalid_inputs=%d, want %d/0", upstreamCalls.Load(), invalidInputs.Load(), taskCount)
+	}
+	if upstreamPeak.Load() > mediaConcurrent || upstreamPeak.Load() < mediaConcurrent/2 {
+		t.Fatalf(
+			"upstream peak=%d, want 50%%-%d after media memory admission; queue_peak=%d running_peak=%d heap_peak=%d",
+			upstreamPeak.Load(),
+			mediaConcurrent,
+			queuePeak.Load(),
+			runningPeak.Load(),
+			peakHeap.Load()-baseline.HeapAlloc,
+		)
+	}
+	if queuePeak.Load() == 0 {
+		t.Fatal("the synchronous distributed admission queue was never observed under the 1000-request burst")
+	}
+	requestAssetUploads, generatedUploads, retainedObjects, invalidUploads := uploadService.stats()
+	if invalidUploads != 0 {
+		t.Fatalf("shared upload validation failures=%d", invalidUploads)
+	}
+	if requestAssetUploads != inputVariants {
+		t.Fatalf("request asset uploads=%d, want %d deduplicated input variants", requestAssetUploads, inputVariants)
+	}
+	if generatedUploads != taskCount {
+		t.Fatalf("generated result uploads=%d, want %d", generatedUploads, taskCount)
+	}
+	if retainedObjects != inputVariants {
+		t.Fatalf("retained shared input objects=%d, want %d", retainedObjects, inputVariants)
+	}
+	uploadedFiles := stressDirectoryFileCount(t, filepath.Join(storageRoot, "uploaded"))
+	generatedFiles := stressDirectoryFileCount(t, filepath.Join(storageRoot, "generated"))
+	if uploadedFiles != 0 || generatedFiles != 0 {
+		t.Fatalf("local stress files uploaded=%d generated=%d, want shared upload only", uploadedFiles, generatedFiles)
+	}
+	heapGrowth := peakHeap.Load() - baseline.HeapAlloc
+	if heapGrowth > maxHeapGrowth {
+		t.Fatalf("heap growth=%d bytes exceeded configured ceiling=%d bytes", heapGrowth, maxHeapGrowth)
+	}
+	t.Logf(
+		"Gemini Base64 同步千任务模拟上游证据: tasks=%d image_bytes=%d input_variants=%d client_connections=%d platform_concurrency=%d media_concurrency=%d elapsed=%s queue_peak=%d running_peak=%d upstream_peak=%d heap_growth_peak=%d request_asset_uploads=%d generated_uploads=%d local_files=%d",
+		taskCount,
+		imageBytes,
+		inputVariants,
+		clientConnections,
+		upstreamConcurrent,
+		mediaConcurrent,
+		elapsed.Round(time.Millisecond),
+		queuePeak.Load(),
+		runningPeak.Load(),
+		upstreamPeak.Load(),
+		heapGrowth,
+		requestAssetUploads,
+		generatedUploads,
+		uploadedFiles+generatedFiles,
+	)
+}
+
+type geminiStressUploadService struct {
+	expectedBytes int
+	outputHash    string
+	inputHashes   map[string]struct{}
+
+	mu              sync.RWMutex
+	apiKey          string
+	retainedObjects map[string][]byte
+
+	requestAssetUploads atomic.Int64
+	generatedUploads    atomic.Int64
+	invalidUploads      atomic.Int64
+}
+
+func newGeminiStressUploadService(expectedBytes int, inputHashes map[string]struct{}, outputHash string) *geminiStressUploadService {
+	return &geminiStressUploadService{
+		expectedBytes:   expectedBytes,
+		outputHash:      outputHash,
+		inputHashes:     inputHashes,
+		retainedObjects: map[string][]byte{},
+	}
+}
+
+func (s *geminiStressUploadService) setAPIKey(apiKey string) {
+	s.mu.Lock()
+	s.apiKey = apiKey
+	s.mu.Unlock()
+}
+
+func (s *geminiStressUploadService) ServeHTTP(w http.ResponseWriter, r *http.Request) {
+	switch {
+	case r.Method == http.MethodPost && r.URL.Path == "/upload":
+		s.handleUpload(w, r)
+	case r.Method == http.MethodGet && strings.HasPrefix(r.URL.Path, "/objects/"):
+		s.handleObject(w, r)
+	default:
+		http.NotFound(w, r)
+	}
+}
+
+func (s *geminiStressUploadService) handleUpload(w http.ResponseWriter, r *http.Request) {
+	s.mu.RLock()
+	apiKey := s.apiKey
+	s.mu.RUnlock()
+	if apiKey == "" || r.Header.Get("Authorization") != "Bearer "+apiKey {
+		s.invalidUploads.Add(1)
+		http.Error(w, "invalid upload authorization", http.StatusUnauthorized)
+		return
+	}
+	reader, err := r.MultipartReader()
+	if err != nil {
+		s.invalidUploads.Add(1)
+		http.Error(w, "invalid multipart body", http.StatusBadRequest)
+		return
+	}
+	var (
+		filePayload []byte
+		scene       string
+	)
+	for {
+		part, nextErr := reader.NextPart()
+		if nextErr == io.EOF {
+			break
+		}
+		if nextErr != nil {
+			s.invalidUploads.Add(1)
+			http.Error(w, "read multipart body failed", http.StatusBadRequest)
+			return
+		}
+		switch part.FormName() {
+		case "file":
+			filePayload, err = io.ReadAll(io.LimitReader(part, int64(s.expectedBytes)+1))
+		case "scene":
+			var raw []byte
+			raw, err = io.ReadAll(io.LimitReader(part, 128))
+			scene = strings.TrimSpace(string(raw))
+		default:
+			_, err = io.Copy(io.Discard, part)
+		}
+		_ = part.Close()
+		if err != nil {
+			s.invalidUploads.Add(1)
+			http.Error(w, "read upload field failed", http.StatusBadRequest)
+			return
+		}
+	}
+	if len(filePayload) != s.expectedBytes {
+		s.invalidUploads.Add(1)
+		http.Error(w, "unexpected upload size", http.StatusBadRequest)
+		return
+	}
+	hash := stressPayloadSHA256(filePayload)
+	switch scene {
+	case store.FileStorageSceneRequestAsset:
+		if _, ok := s.inputHashes[hash]; !ok {
+			s.invalidUploads.Add(1)
+			http.Error(w, "unexpected request asset hash", http.StatusBadRequest)
+			return
+		}
+		s.mu.Lock()
+		if _, exists := s.retainedObjects[hash]; !exists {
+			s.retainedObjects[hash] = filePayload
+		}
+		s.mu.Unlock()
+		s.requestAssetUploads.Add(1)
+	case store.FileStorageSceneImageResult:
+		if hash != s.outputHash {
+			s.invalidUploads.Add(1)
+			http.Error(w, "unexpected generated result hash", http.StatusBadRequest)
+			return
+		}
+		s.generatedUploads.Add(1)
+	default:
+		s.invalidUploads.Add(1)
+		http.Error(w, "unexpected upload scene", http.StatusBadRequest)
+		return
+	}
+	w.Header().Set("Content-Type", "application/json")
+	_ = json.NewEncoder(w).Encode(map[string]any{
+		"url": "http://" + r.Host + "/objects/" + hash,
+	})
+}
+
+func (s *geminiStressUploadService) handleObject(w http.ResponseWriter, r *http.Request) {
+	hash := strings.TrimPrefix(r.URL.Path, "/objects/")
+	s.mu.RLock()
+	payload, ok := s.retainedObjects[hash]
+	s.mu.RUnlock()
+	if !ok {
+		http.NotFound(w, r)
+		return
+	}
+	w.Header().Set("Content-Type", "image/png")
+	w.WriteHeader(http.StatusOK)
+	_, _ = w.Write(payload)
+}
+
+func (s *geminiStressUploadService) stats() (requestAssetUploads int, generatedUploads int, retainedObjects int, invalidUploads int) {
+	s.mu.RLock()
+	retainedObjects = len(s.retainedObjects)
+	s.mu.RUnlock()
+	return int(s.requestAssetUploads.Load()), int(s.generatedUploads.Load()), retainedObjects, int(s.invalidUploads.Load())
+}
+
+func stressPayloadSHA256(payload []byte) string {
+	sum := sha256.Sum256(payload)
+	return hex.EncodeToString(sum[:])
+}
+
+func stressStreamContainsAll(reader io.Reader, needles ...[]byte) (bool, error) {
+	maxNeedle := 0
+	for _, needle := range needles {
+		if len(needle) > maxNeedle {
+			maxNeedle = len(needle)
+		}
+	}
+	if maxNeedle == 0 {
+		return true, nil
+	}
+	found := make([]bool, len(needles))
+	buffer := make([]byte, 64<<10+maxNeedle-1)
+	carried := 0
+	for {
+		n, err := reader.Read(buffer[carried:])
+		total := carried + n
+		window := buffer[:total]
+		for index, needle := range needles {
+			if !found[index] && bytes.Contains(window, needle) {
+				found[index] = true
+			}
+		}
+		if err != nil && err != io.EOF {
+			return false, err
+		}
+		if err == io.EOF {
+			for _, matched := range found {
+				if !matched {
+					return false, nil
+				}
+			}
+			return true, nil
+		}
+		carried = min(maxNeedle-1, total)
+		copy(buffer[:carried], buffer[total-carried:total])
+	}
+}
+
+func stressEnvInt(t *testing.T, key string, fallback int) int {
+	t.Helper()
+	raw := strings.TrimSpace(os.Getenv(key))
+	if raw == "" {
+		return fallback
+	}
+	value, err := strconv.Atoi(raw)
+	if err != nil || value <= 0 {
+		t.Fatalf("%s must be a positive integer", key)
+	}
+	return value
+}
+
+func stressImagePayload(size int, marker byte) []byte {
+	payload := make([]byte, size)
+	copy(payload, []byte{0x89, 'P', 'N', 'G', '\r', '\n', 0x1a, '\n'})
+	for index := 8; index < len(payload); index++ {
+		payload[index] = byte((index*31 + int(marker)) % 251)
+	}
+	if len(payload) > 8 {
+		payload[8] = marker
+	}
+	return payload
+}
+
+func geminiStressInlineData(body map[string]any) string {
+	contents, _ := body["contents"].([]any)
+	for _, rawContent := range contents {
+		content, _ := rawContent.(map[string]any)
+		parts, _ := content["parts"].([]any)
+		for _, rawPart := range parts {
+			part, _ := rawPart.(map[string]any)
+			inline, _ := part["inlineData"].(map[string]any)
+			if inline == nil {
+				inline, _ = part["inline_data"].(map[string]any)
+			}
+			if data, _ := inline["data"].(string); data != "" {
+				return data
+			}
+		}
+	}
+	return ""
+}
+
+func updateAtomicPeak(peak *atomic.Int64, value int64) {
+	for {
+		current := peak.Load()
+		if value <= current || peak.CompareAndSwap(current, value) {
+			return
+		}
+	}
+}
+
+func sampleGeminiStressPressure(
+	ctx context.Context,
+	pool *pgxpool.Pool,
+	model string,
+	peakHeap *atomic.Uint64,
+	queuePeak *atomic.Int64,
+	runningPeak *atomic.Int64,
+	done <-chan struct{},
+) {
+	ticker := time.NewTicker(100 * time.Millisecond)
+	defer ticker.Stop()
+	for {
+		select {
+		case <-ctx.Done():
+			return
+		case <-done:
+			return
+		case <-ticker.C:
+			var stats runtime.MemStats
+			runtime.ReadMemStats(&stats)
+			for {
+				current := peakHeap.Load()
+				if stats.HeapAlloc <= current || peakHeap.CompareAndSwap(current, stats.HeapAlloc) {
+					break
+				}
+			}
+			var waiting, running int64
+			err := pool.QueryRow(ctx, `
+SELECT
+  COUNT(*) FILTER (WHERE admission.status = 'waiting')::bigint,
+  COUNT(*) FILTER (WHERE task.status = 'running')::bigint
+FROM gateway_tasks task
+LEFT JOIN gateway_task_admissions admission ON admission.task_id = task.id
+WHERE task.model = $1`, model).Scan(&waiting, &running)
+			if err == nil {
+				updateAtomicPeak(queuePeak, waiting)
+				updateAtomicPeak(runningPeak, running)
+			}
+		}
+	}
+}
+
+func stressDirectoryFileCount(t *testing.T, path string) int {
+	t.Helper()
+	entries, err := os.ReadDir(path)
+	if err != nil {
+		if os.IsNotExist(err) {
+			return 0
+		}
+		t.Fatalf("read stress directory %s: %v", path, err)
+	}
+	count := 0
+	for _, entry := range entries {
+		if !entry.IsDir() {
+			count++
+		}
+	}
+	return count
+}
+
+func stressBodyPreview(body []byte) string {
+	const limit = 512
+	if len(body) <= limit {
+		return string(body)
+	}
+	return string(body[:limit]) + "..."
+}

apps/api/internal/httpapi/gemini_compat.go

@@ -117,6 +117,15 @@ func (s *Server) geminiGenerateContent(w http.ResponseWriter, r *http.Request) {
 		writeGeminiTaskError(http.StatusUnauthorized, "unauthorized", nil, "unauthorized")
 		return
 	}
+	releaseRequestBody, err := s.acquireMediaRequestBodySlot(r.Context())
+	if err != nil {
+		return
+	}
+	defer func() {
+		if releaseRequestBody != nil {
+			releaseRequestBody()
+		}
+	}()
 	var native map[string]any
 	if err := json.NewDecoder(r.Body).Decode(&native); err != nil {
 		writeGeminiTaskError(http.StatusBadRequest, "invalid json body", nil, "invalid_json_body")
@@ -150,6 +159,13 @@ func (s *Server) geminiGenerateContent(w http.ResponseWriter, r *http.Request) {
 		writeGeminiTaskError(status, err.Error(), nil, clients.ErrorCode(err))
 		return
 	}
+	// The queued task only needs the materialized assetRef. Release the native
+	// Base64 request body before it waits for distributed admission or an
+	// upstream response.
+	native = nil
+	mapping.Body = nil
+	releaseRequestBody()
+	releaseRequestBody = nil
 	createInput := store.CreateTaskInput{
 		Kind:            mapping.Kind,
 		Model:           mapping.Model,

apps/api/internal/httpapi/handlers.go

@@ -1171,7 +1171,7 @@ func (s *Server) createTask(kind string, compatible bool) http.Handler {
 				writeTaskError(http.StatusBadRequest, err.Error(), nil, clients.ErrorCode(err))
 				return
 			}
-			s.logger.Error("create task failed", "kind", kind, "error_category", "task_create_failed")
+			s.logger.Error("create task failed", "kind", kind, "error_category", "task_create_failed", "error", err)
 			writeTaskError(http.StatusInternalServerError, "create task failed", nil, "task_create_failed")
 			return
 		}

apps/api/internal/httpapi/request_preparation.go

@@ -9,9 +9,12 @@ import (
 	"fmt"
 	"mime"
 	"net/http"
+	"os"
 	"path/filepath"
 	"sort"
+	"strconv"
 	"strings"
+	"sync"
 	"time"
 
 	"github.com/easyai/easyai-ai-gateway/apps/api/internal/auth"
@@ -41,6 +44,11 @@ type requestAssetOptions struct {
 	Source           string
 }
 
+type requestAssetLock struct {
+	mu   sync.Mutex
+	refs int
+}
+
 func (s *Server) prepareTaskRequest(ctx context.Context, r *http.Request, user *auth.User, body map[string]any) (preparedTaskRequest, error) {
 	preparedBody, err := s.prepareRequestAssetRefs(ctx, body)
 	if err != nil {
@@ -120,13 +128,9 @@ func (s *Server) prepareRequestAssetValue(ctx context.Context, value any, path [
 		sort.Strings(keys)
 		for _, key := range keys {
 			item := typed[key]
-			if decoded, ok, err := requestAssetFromValue(key, path, item, typed); err != nil {
+			if ref, ok, err := s.prepareRequestAssetField(ctx, key, path, item, typed); err != nil {
 				return nil, err
 			} else if ok {
-				ref, err := s.ensureRequestAsset(ctx, decoded)
-				if err != nil {
-					return nil, err
-				}
 				next[key] = requestAssetWrapper(ref)
 				continue
 			}
@@ -152,6 +156,67 @@ func (s *Server) prepareRequestAssetValue(ctx context.Context, value any, path [
 	}
 }
 
+func (s *Server) prepareRequestAssetField(ctx context.Context, key string, path []string, value any, siblings map[string]any) (map[string]any, bool, error) {
+	text, ok := value.(string)
+	if !ok {
+		return nil, false, nil
+	}
+	raw := strings.TrimSpace(text)
+	if raw == "" || mediaURLString(raw) {
+		return nil, false, nil
+	}
+	if !strings.HasPrefix(strings.ToLower(raw), "data:") &&
+		!strictRequestBase64Field(key, path) &&
+		!likelyRequestBase64MediaField(key, path, raw) {
+		return nil, false, nil
+	}
+	if err := s.acquireMediaRequestSlot(ctx); err != nil {
+		return nil, false, err
+	}
+	defer s.releaseMediaRequestSlot()
+	decoded, ok, err := requestAssetFromValue(key, path, value, siblings)
+	if err != nil || !ok {
+		return nil, ok, err
+	}
+	ref, err := s.ensureRequestAsset(ctx, decoded)
+	return ref, true, err
+}
+
+func (s *Server) acquireMediaRequestSlot(ctx context.Context) error {
+	if s.mediaRequestSlots == nil {
+		return nil
+	}
+	select {
+	case s.mediaRequestSlots <- struct{}{}:
+		return nil
+	case <-ctx.Done():
+		return ctx.Err()
+	}
+}
+
+func (s *Server) releaseMediaRequestSlot() {
+	if s.mediaRequestSlots != nil {
+		<-s.mediaRequestSlots
+	}
+}
+
+func (s *Server) acquireMediaRequestBodySlot(ctx context.Context) (func(), error) {
+	if s.mediaRequestBodySlots == nil {
+		return func() {}, nil
+	}
+	select {
+	case s.mediaRequestBodySlots <- struct{}{}:
+		var once sync.Once
+		return func() {
+			once.Do(func() {
+				<-s.mediaRequestBodySlots
+			})
+		}, nil
+	case <-ctx.Done():
+		return nil, ctx.Err()
+	}
+}
+
 func requestAssetFromValue(key string, path []string, value any, siblings map[string]any) (decodedRequestAsset, bool, error) {
 	text, ok := value.(string)
 	if !ok {
@@ -215,6 +280,8 @@ func (s *Server) ensureRequestAssetWithOptions(ctx context.Context, decoded deco
 	if contentType == "" {
 		contentType = "application/octet-stream"
 	}
+	release := s.acquireRequestAssetLock(sha + "\x00" + contentType + "\x00" + options.UploadScene + "\x00" + strconv.FormatBool(options.RequirePublicURL))
+	defer release()
 	now := time.Now()
 	if existing, ok, err := s.store.FindRequestAsset(ctx, sha, contentType); err != nil && !store.IsUndefinedDatabaseObject(err) {
 		return nil, err
@@ -286,6 +353,31 @@ func (s *Server) ensureRequestAssetWithOptions(ctx context.Context, decoded deco
 	return requestAssetRef(asset), nil
 }
 
+func (s *Server) acquireRequestAssetLock(key string) func() {
+	s.requestAssetLocksMu.Lock()
+	if s.requestAssetLocks == nil {
+		s.requestAssetLocks = map[string]*requestAssetLock{}
+	}
+	entry := s.requestAssetLocks[key]
+	if entry == nil {
+		entry = &requestAssetLock{}
+		s.requestAssetLocks[key] = entry
+	}
+	entry.refs++
+	s.requestAssetLocksMu.Unlock()
+
+	entry.mu.Lock()
+	return func() {
+		entry.mu.Unlock()
+		s.requestAssetLocksMu.Lock()
+		entry.refs--
+		if entry.refs == 0 {
+			delete(s.requestAssetLocks, key)
+		}
+		s.requestAssetLocksMu.Unlock()
+	}
+}
+
 func requestConversationKey(r *http.Request, body map[string]any) string {
 	if r != nil {
 		if value := strings.TrimSpace(r.Header.Get("X-EasyAI-Conversation-ID")); value != "" {
@@ -397,6 +489,16 @@ func requestAssetStillUsable(asset store.RequestAsset, now time.Time) bool {
 	if asset.ExpiresAt != nil && !asset.ExpiresAt.After(now) {
 		return false
 	}
+	if strings.EqualFold(strings.TrimSpace(asset.StorageProvider), "local_static") {
+		localPath := strings.TrimSpace(asset.LocalPath)
+		info, err := os.Stat(localPath)
+		if localPath == "" || err != nil || !info.Mode().IsRegular() {
+			return false
+		}
+		if asset.ByteSize > 0 && info.Size() != asset.ByteSize {
+			return false
+		}
+	}
 	return strings.TrimSpace(asset.URL) != ""
 }
 

apps/api/internal/httpapi/request_preparation_test.go

@@ -16,6 +16,7 @@ import (
 	"time"
 
 	"github.com/easyai/easyai-ai-gateway/apps/api/internal/config"
+	"github.com/easyai/easyai-ai-gateway/apps/api/internal/store"
 )
 
 func TestRequestAssetFromValueDetectsDataURLAndRawBase64(t *testing.T) {
@@ -66,6 +67,33 @@ func TestRequestAssetFromValueDetectsGeminiInlineData(t *testing.T) {
 	}
 }
 
+func TestRequestAssetStillUsableRequiresExistingLocalFile(t *testing.T) {
+	path := filepath.Join(t.TempDir(), "asset.png")
+	if err := os.WriteFile(path, []byte("image"), 0o600); err != nil {
+		t.Fatalf("write local request asset: %v", err)
+	}
+	asset := store.RequestAsset{
+		URL:             "http://127.0.0.1/static/uploaded/asset.png",
+		StorageProvider: "local_static",
+		LocalPath:       path,
+		ByteSize:        5,
+	}
+	if !requestAssetStillUsable(asset, time.Now()) {
+		t.Fatal("existing local request asset was rejected")
+	}
+	asset.ByteSize = 6
+	if requestAssetStillUsable(asset, time.Now()) {
+		t.Fatal("truncated local request asset was treated as reusable")
+	}
+	asset.ByteSize = 5
+	if err := os.Remove(path); err != nil {
+		t.Fatalf("remove local request asset: %v", err)
+	}
+	if requestAssetStillUsable(asset, time.Now()) {
+		t.Fatal("missing local request asset was treated as reusable")
+	}
+}
+
 func TestCanonicalConversationMessageHashUsesTextAndAssetRefs(t *testing.T) {
 	message := map[string]any{
 		"role": "user",

apps/api/internal/httpapi/server.go

@@ -32,6 +32,10 @@ type Server struct {
 	runner                     *runner.Service
 	logger                     *slog.Logger
 	geminiUploadSessions       sync.Map
+	requestAssetLocksMu        sync.Mutex
+	requestAssetLocks          map[string]*requestAssetLock
+	mediaRequestSlots          chan struct{}
+	mediaRequestBodySlots      chan struct{}
 	securityEventReceiver      http.Handler
 	securityEventManager       *ssfreceiver.ConnectionManager
 	identityRuntime            *identityruntime.Manager
@@ -68,16 +72,25 @@ func NewServer(cfg config.Config, db *store.Store, logger *slog.Logger) http.Han
 }
 
 func NewServerWithContext(ctx context.Context, cfg config.Config, db *store.Store, logger *slog.Logger) http.Handler {
+	if cfg.MediaMaterializationConcurrency == 0 {
+		cfg.MediaMaterializationConcurrency = 8
+	}
+	if cfg.MediaRequestConcurrency == 0 {
+		cfg.MediaRequestConcurrency = 16
+	}
 	securityEventMetrics := &ssfreceiver.Metrics{}
 	server := &Server{
-		ctx:              ctx,
-		cfg:              cfg,
-		store:            db,
-		oidcUserResolver: db,
-		auth:             auth.New(cfg.JWTSecret, cfg.ServerMainBaseURL, cfg.ServerMainInternalToken),
-		runner:           runner.New(cfg, db, logger, securityEventMetrics),
-		logger:           logger,
-		billingMetrics:   securityEventMetrics,
+		ctx:                   ctx,
+		cfg:                   cfg,
+		store:                 db,
+		oidcUserResolver:      db,
+		auth:                  auth.New(cfg.JWTSecret, cfg.ServerMainBaseURL, cfg.ServerMainInternalToken),
+		runner:                runner.New(cfg, db, logger, securityEventMetrics),
+		logger:                logger,
+		billingMetrics:        securityEventMetrics,
+		requestAssetLocks:     map[string]*requestAssetLock{},
+		mediaRequestSlots:     make(chan struct{}, cfg.MediaMaterializationConcurrency),
+		mediaRequestBodySlots: make(chan struct{}, cfg.MediaRequestConcurrency),
 	}
 	server.auth.ServerMainInternalKey = cfg.ServerMainInternalKey
 	server.auth.ServerMainInternalSecret = cfg.ServerMainInternalSecret

apps/api/internal/runner/admission.go

@@ -3,6 +3,7 @@ package runner
 import (
 	"context"
 	"errors"
+	"hash/fnv"
 	"strings"
 	"time"
 
@@ -22,6 +23,17 @@ type taskAdmissionPlan struct {
 	Eligible  bool
 }
 
+type admissionTaskWaiter struct {
+	wake     chan struct{}
+	waiterID string
+}
+
+const (
+	asyncWorkerCapacityScopeKey = "global"
+	asyncWorkerQueueLimit       = 10000
+	asyncWorkerMaxWaitSeconds   = 24 * 60 * 60
+)
+
 func distributedAdmissionModelType(modelType string) bool {
 	switch strings.ToLower(strings.TrimSpace(modelType)) {
 	case "image_generate",
@@ -99,6 +111,12 @@ func (s *Service) buildTaskAdmissionPlan(ctx context.Context, task store.Gateway
 			continue
 		}
 		scopes, groupID := s.admissionScopes(ctx, user, candidate)
+		if task.AsyncMode {
+			scopes, err = s.withAsyncWorkerCapacityScope(ctx, scopes)
+			if err != nil {
+				return taskAdmissionPlan{}, err
+			}
+		}
 		hasConcurrentLimit := false
 		for _, scope := range scopes {
 			if scope.ConcurrentLimit > 0 {
@@ -128,10 +146,67 @@ func (s *Service) buildTaskAdmissionPlan(ctx context.Context, task store.Gateway
 	return taskAdmissionPlan{}, store.ErrNoModelCandidate
 }
 
+func (s *Service) withAsyncWorkerCapacityScope(ctx context.Context, scopes []store.AdmissionScope) ([]store.AdmissionScope, error) {
+	capacity, err := s.store.ActiveWorkerCapacity(ctx)
+	if err != nil {
+		return nil, err
+	}
+	if capacity <= 0 {
+		return scopes, nil
+	}
+	out := append([]store.AdmissionScope{}, scopes...)
+	out = append(out, store.AdmissionScope{
+		ScopeType:       "worker_capacity",
+		ScopeKey:        asyncWorkerCapacityScopeKey,
+		ScopeName:       "async worker capacity",
+		ConcurrentLimit: float64(capacity),
+		Amount:          1,
+		LeaseTTLSeconds: 120,
+		QueueLimit:      asyncWorkerQueueLimit,
+		MaxWaitSeconds:  asyncWorkerMaxWaitSeconds,
+	})
+	return out, nil
+}
+
 func (s *Service) tryTaskAdmission(ctx context.Context, task store.GatewayTask, plan taskAdmissionPlan, waiterID string) (store.TaskAdmissionResult, error) {
 	return s.tryTaskAdmissionWithAdmittedHook(ctx, task, plan, waiterID, nil)
 }
 
+func (s *Service) activeAsyncTaskAdmission(
+	ctx context.Context,
+	task store.GatewayTask,
+	plan taskAdmissionPlan,
+) (store.TaskAdmissionResult, bool, error) {
+	if !task.AsyncMode {
+		return store.TaskAdmissionResult{}, false, nil
+	}
+	admission, err := s.store.GetTaskAdmission(ctx, task.ID)
+	if errors.Is(err, pgx.ErrNoRows) {
+		return store.TaskAdmissionResult{}, false, nil
+	}
+	if err != nil {
+		return store.TaskAdmissionResult{}, false, err
+	}
+	if admission.Status != "admitted" ||
+		admission.PlatformID != plan.Candidate.PlatformID ||
+		admission.PlatformModelID != plan.Candidate.PlatformModelID ||
+		admission.UserGroupID != plan.GroupID {
+		return store.TaskAdmissionResult{}, false, nil
+	}
+	leases, err := s.store.ActiveTaskAdmissionLeases(ctx, task.ID)
+	if err != nil {
+		return store.TaskAdmissionResult{}, false, err
+	}
+	if len(leases) == 0 {
+		return store.TaskAdmissionResult{}, false, nil
+	}
+	return store.TaskAdmissionResult{
+		Admission: admission,
+		Admitted:  true,
+		Leases:    leases,
+	}, true, nil
+}
+
 func (s *Service) tryTaskAdmissionWithAdmittedHook(
 	ctx context.Context,
 	task store.GatewayTask,
@@ -139,21 +214,7 @@ func (s *Service) tryTaskAdmissionWithAdmittedHook(
 	waiterID string,
 	onAdmitted func(pgx.Tx) error,
 ) (store.TaskAdmissionResult, error) {
-	mode := "sync"
-	if task.AsyncMode {
-		mode = "async"
-	}
-	input := store.TaskAdmissionInput{
-		TaskID:          task.ID,
-		PlatformID:      plan.Candidate.PlatformID,
-		PlatformModelID: plan.Candidate.PlatformModelID,
-		UserGroupID:     plan.GroupID,
-		QueueKey:        plan.Candidate.QueueKey,
-		Mode:            mode,
-		Priority:        task.Priority,
-		WaiterID:        waiterID,
-		Scopes:          plan.Scopes,
-	}
+	input := taskAdmissionInput(task, plan, waiterID)
 	current, currentErr := s.store.GetTaskAdmission(ctx, task.ID)
 	if currentErr != nil && !errors.Is(currentErr, pgx.ErrNoRows) {
 		return store.TaskAdmissionResult{}, currentErr
@@ -186,14 +247,34 @@ func (s *Service) tryTaskAdmissionWithAdmittedHook(
 	return result, err
 }
 
+func taskAdmissionInput(task store.GatewayTask, plan taskAdmissionPlan, waiterID string) store.TaskAdmissionInput {
+	mode := "sync"
+	if task.AsyncMode {
+		mode = "async"
+	}
+	return store.TaskAdmissionInput{
+		TaskID:          task.ID,
+		PlatformID:      plan.Candidate.PlatformID,
+		PlatformModelID: plan.Candidate.PlatformModelID,
+		UserGroupID:     plan.GroupID,
+		QueueKey:        plan.Candidate.QueueKey,
+		Mode:            mode,
+		Priority:        task.Priority,
+		WaiterID:        waiterID,
+		Scopes:          plan.Scopes,
+	}
+}
+
 func (s *Service) waitForSynchronousAdmission(ctx context.Context, task store.GatewayTask, plan taskAdmissionPlan) (store.TaskAdmissionResult, error) {
 	waiterID := uuid.NewString()
 	return s.waitForTaskAdmission(ctx, task, plan, waiterID)
 }
 
 func (s *Service) waitForTaskAdmission(ctx context.Context, task store.GatewayTask, plan taskAdmissionPlan, waiterID string) (store.TaskAdmissionResult, error) {
-	ticker := time.NewTicker(time.Second)
-	defer ticker.Stop()
+	taskWake, unregister := s.registerAdmissionTaskWaiter(task.ID, waiterID)
+	defer unregister()
+	pollTimer := time.NewTimer(admissionPollInterval(task.ID))
+	defer pollTimer.Stop()
 	for {
 		result, err := s.tryTaskAdmission(ctx, task, plan, waiterID)
 		if err != nil || result.Admitted {
@@ -204,12 +285,19 @@ func (s *Service) waitForTaskAdmission(ctx context.Context, task store.GatewayTa
 			_ = s.store.DeleteTaskAdmission(context.WithoutCancel(ctx), task.ID)
 			s.observeTaskAdmission("cancelled")
 			return store.TaskAdmissionResult{}, ctx.Err()
-		case <-s.currentAdmissionWake():
-		case <-ticker.C:
+		case <-taskWake:
+		case <-pollTimer.C:
+			pollTimer.Reset(admissionPollInterval(task.ID))
 		}
 	}
 }
 
+func admissionPollInterval(taskID string) time.Duration {
+	hasher := fnv.New32a()
+	_, _ = hasher.Write([]byte(taskID))
+	return 30*time.Second + time.Duration(hasher.Sum32()%5000)*time.Millisecond
+}
+
 func (s *Service) ensureCandidateAdmission(
 	ctx context.Context,
 	task store.GatewayTask,
@@ -218,6 +306,13 @@ func (s *Service) ensureCandidateAdmission(
 	candidate store.RuntimeModelCandidate,
 ) (store.TaskAdmissionResult, bool, error) {
 	scopes, groupID := s.admissionScopes(ctx, user, candidate)
+	if task.AsyncMode {
+		var err error
+		scopes, err = s.withAsyncWorkerCapacityScope(ctx, scopes)
+		if err != nil {
+			return store.TaskAdmissionResult{}, true, err
+		}
+	}
 	hasConcurrentLimit := false
 	for _, scope := range scopes {
 		if scope.ConcurrentLimit > 0 {
@@ -270,10 +365,12 @@ func (s *Service) observeTaskAdmissionWait(wait time.Duration) {
 
 func (s *Service) StartAdmissionNotifier(ctx context.Context) {
 	s.admissionListener.Do(func() {
+		go s.dispatchWaitingSynchronousAdmissions(ctx)
+		go s.renewSynchronousAdmissionWaiters(ctx)
 		go func() {
 			for ctx.Err() == nil {
-				err := s.store.ListenTaskAdmissionNotifications(ctx, func(string) {
-					s.broadcastAdmissionWake()
+				err := s.store.ListenTaskAdmissionNotifications(ctx, func(taskID string) {
+					s.signalAdmissionWake(taskID)
 				})
 				if ctx.Err() != nil {
 					return
@@ -293,17 +390,99 @@ func (s *Service) StartAdmissionNotifier(ctx context.Context) {
 	})
 }
 
-func (s *Service) currentAdmissionWake() <-chan struct{} {
-	s.admissionWakeMu.Lock()
-	defer s.admissionWakeMu.Unlock()
-	return s.admissionWake
+func (s *Service) dispatchWaitingSynchronousAdmissions(ctx context.Context) {
+	for {
+		select {
+		case <-ctx.Done():
+			return
+		case <-s.admissionWake:
+		}
+		taskIDs, err := s.store.ListWaitingTaskAdmissionIDs(ctx, 1)
+		if err != nil {
+			if s.logger != nil {
+				s.logger.Warn("list waiting synchronous admissions failed", "error", err)
+			}
+			continue
+		}
+		s.admissionWakeMu.Lock()
+		waiters := make([]chan struct{}, 0, len(taskIDs))
+		for _, taskID := range taskIDs {
+			if waiter := s.admissionTaskWaiters[taskID]; waiter != nil {
+				waiters = append(waiters, waiter.wake)
+			}
+		}
+		s.admissionWakeMu.Unlock()
+		for _, taskWake := range waiters {
+			select {
+			case taskWake <- struct{}{}:
+			default:
+			}
+		}
+	}
 }
 
-func (s *Service) broadcastAdmissionWake() {
+func (s *Service) renewSynchronousAdmissionWaiters(ctx context.Context) {
+	ticker := time.NewTicker(5 * time.Second)
+	defer ticker.Stop()
+	for {
+		select {
+		case <-ctx.Done():
+			return
+		case <-ticker.C:
+		}
+		s.admissionWakeMu.Lock()
+		taskIDs := make([]string, 0, len(s.admissionTaskWaiters))
+		waiterIDs := make([]string, 0, len(s.admissionTaskWaiters))
+		for taskID, waiter := range s.admissionTaskWaiters {
+			if waiter == nil || strings.TrimSpace(waiter.waiterID) == "" {
+				continue
+			}
+			taskIDs = append(taskIDs, taskID)
+			waiterIDs = append(waiterIDs, waiter.waiterID)
+		}
+		s.admissionWakeMu.Unlock()
+		if err := s.store.RenewTaskAdmissionWaiters(ctx, taskIDs, waiterIDs); err != nil && s.logger != nil {
+			s.logger.Warn("renew synchronous admission waiters failed", "waiterCount", len(taskIDs), "error", err)
+		}
+	}
+}
+
+func (s *Service) registerAdmissionTaskWaiter(taskID string, waiterID string) (<-chan struct{}, func()) {
 	s.admissionWakeMu.Lock()
-	close(s.admissionWake)
-	s.admissionWake = make(chan struct{})
+	waiter := &admissionTaskWaiter{wake: make(chan struct{}, 1), waiterID: waiterID}
+	s.admissionTaskWaiters[taskID] = waiter
 	s.admissionWakeMu.Unlock()
+	return waiter.wake, func() {
+		s.admissionWakeMu.Lock()
+		if s.admissionTaskWaiters[taskID] == waiter {
+			delete(s.admissionTaskWaiters, taskID)
+		}
+		s.admissionWakeMu.Unlock()
+	}
+}
+
+func (s *Service) signalAdmissionWake(taskID string) {
+	taskID = strings.TrimSpace(taskID)
+	if taskID == "" || taskID == "*" {
+		select {
+		case s.admissionWake <- struct{}{}:
+		default:
+		}
+	} else {
+		s.admissionWakeMu.Lock()
+		waiter := s.admissionTaskWaiters[taskID]
+		s.admissionWakeMu.Unlock()
+		if waiter != nil {
+			select {
+			case waiter.wake <- struct{}{}:
+			default:
+			}
+		}
+	}
+	select {
+	case s.asyncAdmissionWake <- struct{}{}:
+	default:
+	}
 }
 
 func (s *Service) SubmitAsyncTask(ctx context.Context, task store.GatewayTask) error {
@@ -335,28 +514,49 @@ func (s *Service) SubmitAsyncTask(ctx context.Context, task store.GatewayTask) e
 		}
 		return nil
 	}
-	result, err := s.tryTaskAdmissionWithAdmittedHook(ctx, task, plan, "", func(tx pgx.Tx) error {
-		return s.enqueueAsyncTaskTx(ctx, tx, task.ID, asyncTaskInsertOpts(task))
-	})
-	if err != nil {
+	input := taskAdmissionInput(task, plan, "")
+	current, currentErr := s.store.GetTaskAdmission(ctx, task.ID)
+	if currentErr != nil && !errors.Is(currentErr, pgx.ErrNoRows) {
+		return currentErr
+	}
+	if currentErr == nil && current.Status == "waiting" &&
+		(current.PlatformID != input.PlatformID || current.PlatformModelID != input.PlatformModelID || current.UserGroupID != input.UserGroupID) {
+		if _, err := s.store.RebindWaitingTaskAdmission(ctx, input); err != nil {
+			return err
+		}
+		s.observeTaskAdmission("candidate_migrated")
+	}
+	// Register FIFO order without creating a River job. A Worker dispatcher
+	// first reserves both business concurrency and a live execution slot, then
+	// creates the unique River job in the same transaction.
+	if _, err := s.store.QueueTaskAdmissionWithHook(ctx, input, nil); err != nil {
 		if s.cancelAsyncSubmissionIfDisconnected(ctx, task.ID) {
 			return ctx.Err()
 		}
 		_, _ = s.store.FailQueuedTask(context.WithoutCancel(ctx), task.ID, clients.ErrorCode(err), err.Error())
 		return err
 	}
-	if !result.Admitted {
-		if s.cancelAsyncSubmissionIfDisconnected(ctx, task.ID) {
-			return ctx.Err()
-		}
-		return nil
-	}
 	if s.cancelAsyncSubmissionIfDisconnected(ctx, task.ID) {
 		return ctx.Err()
 	}
 	return nil
 }
 
+func (s *Service) dispatchWaitingAsyncTask(ctx context.Context, task store.GatewayTask) error {
+	user := authUserFromTask(task)
+	plan, err := s.buildTaskAdmissionPlan(ctx, task, user)
+	if err != nil {
+		return err
+	}
+	if !plan.Eligible {
+		return s.EnqueueAsyncTask(ctx, task)
+	}
+	_, err = s.tryTaskAdmissionWithAdmittedHook(ctx, task, plan, "", func(tx pgx.Tx) error {
+		return s.enqueueAsyncTaskTx(ctx, tx, task.ID, asyncTaskInsertOpts(task))
+	})
+	return err
+}
+
 func (s *Service) cancelAsyncSubmissionIfDisconnected(ctx context.Context, taskID string) bool {
 	if ctx.Err() == nil {
 		return false
@@ -380,10 +580,10 @@ func (s *Service) dispatchWaitingAsyncAdmissions(ctx context.Context) {
 		select {
 		case <-ctx.Done():
 			return
-		case <-s.currentAdmissionWake():
+		case <-s.asyncAdmissionWake:
 		case <-ticker.C:
 		}
-		taskIDs, err := s.store.ListWaitingAsyncAdmissionTaskIDs(ctx, 100)
+		taskIDs, err := s.store.ListWaitingAsyncAdmissionTaskIDs(ctx, 1000)
 		if err != nil {
 			s.logger.Warn("list waiting async admissions failed", "error", err)
 			continue
@@ -399,7 +599,7 @@ func (s *Service) dispatchWaitingAsyncAdmissions(ctx context.Context) {
 			if task.Status != "queued" || task.RiverJobID > 0 {
 				continue
 			}
-			if err := s.SubmitAsyncTask(ctx, task); err != nil {
+			if err := s.dispatchWaitingAsyncTask(ctx, task); err != nil {
 				s.logger.Warn("dispatch waiting async admission failed", "taskID", taskID, "error", err)
 			}
 		}

apps/api/internal/runner/proxy.go

@@ -21,6 +21,11 @@ type httpClientCache struct {
 
 const providerHTTPClientTimeout = 10 * time.Minute
 
+const (
+	providerHTTPMaxIdleConnections        = 2048
+	providerHTTPMaxIdleConnectionsPerHost = 1024
+)
+
 func newHTTPClientCache() *httpClientCache {
 	return &httpClientCache{
 		none:   newHTTPClient(nil),
@@ -68,6 +73,12 @@ func (c *httpClientCache) customClient(rawProxy string) (*http.Client, error) {
 func newHTTPClient(proxy func(*http.Request) (*url.URL, error)) *http.Client {
 	transport := http.DefaultTransport.(*http.Transport).Clone()
 	transport.Proxy = proxy
+	// Media workers may keep hundreds of slow upstream tasks in flight and
+	// poll the same provider repeatedly. The standard library only retains two
+	// idle connections per host, which turns a high-concurrency polling load
+	// into avoidable TCP/TLS handshakes.
+	transport.MaxIdleConns = providerHTTPMaxIdleConnections
+	transport.MaxIdleConnsPerHost = providerHTTPMaxIdleConnectionsPerHost
 	return &http.Client{
 		Timeout:   providerHTTPClientTimeout,
 		Transport: transport,

apps/api/internal/runner/proxy_test.go

@@ -16,6 +16,20 @@ func TestProviderHTTPClientTimeoutAllowsLongRunningMediaRequests(t *testing.T) {
 	if client.Timeout != 10*time.Minute {
 		t.Fatalf("unexpected provider HTTP timeout: got %s want %s", client.Timeout, 10*time.Minute)
 	}
+	transport, ok := client.Transport.(*http.Transport)
+	if !ok {
+		t.Fatalf("provider transport type = %T, want *http.Transport", client.Transport)
+	}
+	if transport.MaxIdleConns != providerHTTPMaxIdleConnections ||
+		transport.MaxIdleConnsPerHost != providerHTTPMaxIdleConnectionsPerHost {
+		t.Fatalf(
+			"provider idle connection pool = %d/%d, want %d/%d",
+			transport.MaxIdleConns,
+			transport.MaxIdleConnsPerHost,
+			providerHTTPMaxIdleConnections,
+			providerHTTPMaxIdleConnectionsPerHost,
+		)
+	}
 }
 
 func TestPlatformProxyModeNoneIgnoresEnvironmentProxy(t *testing.T) {

apps/api/internal/runner/service.go

@@ -40,8 +40,11 @@ type Service struct {
 	asyncCapacityLoader  func(context.Context, int) (store.AsyncWorkerCapacitySnapshot, error)
 	admissionWakeMu      sync.Mutex
 	admissionWake        chan struct{}
+	asyncAdmissionWake   chan struct{}
+	admissionTaskWaiters map[string]*admissionTaskWaiter
 	admissionListener    sync.Once
 	asyncClientFactory   func(int) (asyncExecutionClient, error)
+	mediaResultSlots     chan struct{}
 	billingMetrics       billingMetricsObserver
 }
 
@@ -93,6 +96,9 @@ func New(cfg config.Config, db *store.Store, logger *slog.Logger, observers ...b
 	if cfg.AsyncWorkerRefreshIntervalSeconds == 0 {
 		cfg.AsyncWorkerRefreshIntervalSeconds = 5
 	}
+	if cfg.MediaMaterializationConcurrency == 0 {
+		cfg.MediaMaterializationConcurrency = 8
+	}
 	if cfg.TaskProgressCallbackTimeoutMS == 0 {
 		cfg.TaskProgressCallbackTimeoutMS = 5000
 	}
@@ -139,9 +145,12 @@ func New(cfg config.Config, db *store.Store, logger *slog.Logger, observers ...b
 			"universal":             clients.UniversalClient{HTTPClient: httpClients.none, ScriptExecutor: scriptExecutor},
 			"simulation":            clients.SimulationClient{},
 		},
-		httpClients:      httpClients,
-		workerInstanceID: asyncWorkerID(),
-		admissionWake:    make(chan struct{}),
+		httpClients:          httpClients,
+		workerInstanceID:     asyncWorkerID(),
+		admissionWake:        make(chan struct{}, 4096),
+		asyncAdmissionWake:   make(chan struct{}, 1),
+		admissionTaskWaiters: map[string]*admissionTaskWaiter{},
+		mediaResultSlots:     make(chan struct{}, cfg.MediaMaterializationConcurrency),
 	}
 	if len(observers) > 0 {
 		service.billingMetrics = observers[0]
@@ -539,7 +548,11 @@ func (s *Service) executeWithToken(ctx context.Context, task store.GatewayTask,
 			var admissionResult store.TaskAdmissionResult
 			var admissionErr error
 			if task.AsyncMode {
-				admissionResult, admissionErr = s.tryTaskAdmission(ctx, task, plan, "")
+				var alreadyAdmitted bool
+				admissionResult, alreadyAdmitted, admissionErr = s.activeAsyncTaskAdmission(ctx, task, plan)
+				if admissionErr == nil && !alreadyAdmitted {
+					admissionResult, admissionErr = s.tryTaskAdmission(ctx, task, plan, "")
+				}
 				if admissionErr == nil && !admissionResult.Admitted {
 					_ = s.store.ReleaseTaskPreparation(context.WithoutCancel(ctx), task.ID, task.ExecutionToken)
 					return Result{Task: task, Output: task.Result}, &TaskQueuedError{Delay: time.Second}
@@ -1272,6 +1285,15 @@ func (s *Service) runCandidate(
 		})
 		return clients.Response{}, err
 	}
+	mediaSlotPreacquired := mediaTaskNeedsPreUpstreamMaterializationSlot(candidate.ModelType)
+	releaseMediaSlot := func() {}
+	if mediaSlotPreacquired {
+		releaseMediaSlot, err = s.acquireMediaMaterializationSlot(ctx)
+		if err != nil {
+			return clients.Response{}, err
+		}
+		defer releaseMediaSlot()
+	}
 	providerBody, err = s.hydrateProviderRequestAssets(ctx, providerBody, candidate)
 	if err != nil {
 		_ = s.store.FinishTaskAttempt(ctx, store.FinishTaskAttemptInput{
@@ -1465,6 +1487,21 @@ func (s *Service) runCandidate(
 		}
 		return clients.Response{}, err
 	}
+	if task.AsyncMode {
+		// Async callers consume the canonical task result through the task API,
+		// never the original provider wire response. Release the raw JSON before
+		// decoding/uploading inline media so a large Base64 response is not held
+		// twice during the most memory-intensive phase.
+		response.Wire = nil
+		submissionWire = nil
+	}
+	if !mediaSlotPreacquired {
+		releaseMediaSlot, err = s.acquireMediaResultSlot(ctx, response.Result)
+		if err != nil {
+			return clients.Response{}, err
+		}
+		defer releaseMediaSlot()
+	}
 	uploadedResult, err := s.uploadGeneratedAssets(ctx, task.ID, task.Kind, response.Result)
 	if err != nil {
 		metrics := mergeMetrics(taskMetrics(task, user, body, candidate, response, simulated), parameterPreprocessingMetrics(preprocessing), map[string]any{

apps/api/internal/runner/upload.go

@@ -20,6 +20,7 @@ import (
 	"os"
 	"path/filepath"
 	"strings"
+	"sync"
 	"time"
 
 	"github.com/easyai/easyai-ai-gateway/apps/api/internal/clients"
@@ -77,6 +78,40 @@ func defaultGeneratedAssetUploadPolicy() generatedAssetUploadPolicy {
 	}
 }
 
+func (s *Service) acquireMediaMaterializationSlot(ctx context.Context) (func(), error) {
+	if s.mediaResultSlots == nil {
+		return func() {}, nil
+	}
+	select {
+	case s.mediaResultSlots <- struct{}{}:
+		var once sync.Once
+		return func() {
+			once.Do(func() {
+				<-s.mediaResultSlots
+			})
+		}, nil
+	case <-ctx.Done():
+		return nil, ctx.Err()
+	}
+}
+
+func (s *Service) acquireMediaResultSlot(ctx context.Context, result map[string]any) (func(), error) {
+	if s.mediaResultSlots == nil ||
+		(!TaskResultHasInlineBinary(result) && !generatedRawValueHasInlineMedia(result["raw"], "", nil)) {
+		return func() {}, nil
+	}
+	return s.acquireMediaMaterializationSlot(ctx)
+}
+
+func mediaTaskNeedsPreUpstreamMaterializationSlot(modelType string) bool {
+	switch canonicalModelType(modelType) {
+	case "image_generate", "image_edit", "audio_generate", "text_to_speech", "voice_clone", "omni":
+		return true
+	default:
+		return false
+	}
+}
+
 func (s *Service) uploadGeneratedAssets(ctx context.Context, taskID string, taskKind string, result map[string]any) (map[string]any, error) {
 	data, _ := result["data"].([]any)
 	rawNeedsUpload := generatedRawValueHasInlineMedia(result["raw"], "", nil)

apps/api/internal/runner/upload_test.go

@@ -53,6 +53,60 @@ func TestGeneratedAssetDecisionUploadsInlineImageBase64(t *testing.T) {
 	}
 }
 
+func TestMediaResultMaterializationConcurrencyIsBounded(t *testing.T) {
+	service := New(config.Config{MediaMaterializationConcurrency: 1}, nil, nil)
+	result := map[string]any{
+		"data": []any{map[string]any{
+			"b64_json":  base64.StdEncoding.EncodeToString([]byte("inline image")),
+			"mime_type": "image/png",
+		}},
+	}
+	release, err := service.acquireMediaResultSlot(context.Background(), result)
+	if err != nil {
+		t.Fatalf("acquire first media result slot: %v", err)
+	}
+
+	waitCtx, cancel := context.WithTimeout(context.Background(), 25*time.Millisecond)
+	defer cancel()
+	if _, err := service.acquireMediaResultSlot(waitCtx, result); err == nil {
+		t.Fatal("second media result materialization bypassed configured concurrency")
+	}
+	release()
+
+	release, err = service.acquireMediaResultSlot(context.Background(), result)
+	if err != nil {
+		t.Fatalf("acquire released media result slot: %v", err)
+	}
+	release()
+}
+
+func TestMediaTaskPreacquiresMaterializationSlotForInlineResultTypes(t *testing.T) {
+	t.Parallel()
+	for _, modelType := range []string{
+		"image_generate",
+		"image_edit",
+		"audio_generate",
+		"text_to_speech",
+		"voice_clone",
+		"omni",
+	} {
+		if !mediaTaskNeedsPreUpstreamMaterializationSlot(modelType) {
+			t.Fatalf("%s should preacquire the media materialization slot", modelType)
+		}
+	}
+	for _, modelType := range []string{
+		"text_generate",
+		"text_embedding",
+		"text_rerank",
+		"video_generate",
+		"image_to_video",
+	} {
+		if mediaTaskNeedsPreUpstreamMaterializationSlot(modelType) {
+			t.Fatalf("%s should not hold a media materialization slot across the upstream call", modelType)
+		}
+	}
+}
+
 func TestGeneratedAssetDecisionUploadsVectorDocumentWithoutChangingType(t *testing.T) {
 	item := map[string]any{
 		"type":      "file",

apps/api/internal/store/admission_queue.go

@@ -5,6 +5,7 @@ import (
 	"database/sql"
 	"errors"
 	"fmt"
+	"math"
 	"sort"
 	"strings"
 	"time"
@@ -15,6 +16,7 @@ import (
 const (
 	admissionWaiterLeaseTTL = 15 * time.Second
 	admissionNotifyChannel  = "gateway_task_admission"
+	executionSlotWaitMax    = 24 * time.Hour
 )
 
 var ErrQueueTimeout = errors.New("task admission queue wait timed out")
@@ -113,6 +115,111 @@ func (s *Store) TryTaskAdmissionWithAdmittedHook(
 	return s.tryTaskAdmission(ctx, input, onAdmitted)
 }
 
+// QueueTaskAdmissionWithHook durably registers an asynchronous FIFO waiter and
+// runs onQueued in the same transaction. It intentionally does not reserve
+// concurrency leases: the Worker dispatcher obtains business and execution
+// capacity only when the task reaches the head of every applicable queue.
+func (s *Store) QueueTaskAdmissionWithHook(
+	ctx context.Context,
+	input TaskAdmissionInput,
+	onQueued func(pgx.Tx) error,
+) (TaskAdmission, error) {
+	if err := validateTaskAdmissionInput(input); err != nil {
+		return TaskAdmission{}, err
+	}
+	if input.Mode != "async" {
+		return TaskAdmission{}, errors.New("queued admission registration requires asynchronous mode")
+	}
+	tx, err := s.pool.Begin(ctx)
+	if err != nil {
+		return TaskAdmission{}, err
+	}
+	defer tx.Rollback(ctx)
+
+	if _, err := tx.Exec(ctx, `SELECT pg_advisory_xact_lock(hashtextextended($1, 0))`, "task-admission:"+input.TaskID); err != nil {
+		return TaskAdmission{}, err
+	}
+	var taskActive bool
+	if err := tx.QueryRow(ctx, `
+SELECT status IN ('queued', 'running')
+FROM gateway_tasks
+WHERE id = $1::uuid`, input.TaskID).Scan(&taskActive); err != nil {
+		return TaskAdmission{}, err
+	}
+	if !taskActive {
+		return TaskAdmission{}, ErrTaskExecutionFinished
+	}
+
+	admission, found, err := loadTaskAdmissionTx(ctx, tx, input.TaskID)
+	if err != nil {
+		return TaskAdmission{}, err
+	}
+	scopes := normalizedAdmissionScopes(input.Scopes)
+	lockScopes := append([]AdmissionScope{}, scopes...)
+	if found {
+		lockScopes = append(lockScopes,
+			AdmissionScope{ScopeType: "platform_model", ScopeKey: admission.PlatformModelID},
+			AdmissionScope{ScopeType: "user_group", ScopeKey: admission.UserGroupID},
+		)
+	}
+	for _, scope := range normalizedAdmissionScopes(lockScopes) {
+		if _, err := tx.Exec(ctx, `SELECT pg_advisory_xact_lock(hashtextextended($1, 0))`, admissionLockKey(scope)); err != nil {
+			return TaskAdmission{}, err
+		}
+	}
+
+	if found && admission.Status == "admitted" {
+		activeLeases, leaseErr := activeTaskAdmissionLeaseCountTx(ctx, tx, input.TaskID)
+		if leaseErr != nil {
+			return TaskAdmission{}, leaseErr
+		}
+		if activeLeases == 0 {
+			admission, err = resetTaskAdmissionToWaitingTx(ctx, tx, input)
+			if err != nil {
+				return TaskAdmission{}, err
+			}
+		}
+	}
+	if found && admission.Status == "waiting" && !admission.WaitDeadlineAt.After(time.Now()) {
+		if err := expireTaskAdmissionTx(ctx, tx, input.TaskID); err != nil {
+			return TaskAdmission{}, err
+		}
+		if err := notifyTaskAdmissionTx(ctx, tx, "*"); err != nil {
+			return TaskAdmission{}, err
+		}
+		if err := tx.Commit(ctx); err != nil {
+			return TaskAdmission{}, err
+		}
+		return TaskAdmission{}, &QueueTimeoutError{TaskID: input.TaskID}
+	}
+	if !found {
+		scopeStates, stateErr := admissionScopeStatesTx(ctx, tx, scopes)
+		if stateErr != nil {
+			return TaskAdmission{}, stateErr
+		}
+		if admissionErr := validateQueuedAdmissionCapacity(scopeStates); admissionErr != nil {
+			return TaskAdmission{}, admissionErr
+		}
+		admission, err = insertTaskAdmissionTx(ctx, tx, input, admissionDeadline(scopes))
+		if err != nil {
+			return TaskAdmission{}, err
+		}
+		found = true
+	}
+	if onQueued != nil {
+		if err := onQueued(tx); err != nil {
+			return TaskAdmission{}, err
+		}
+	}
+	if err := notifyTaskAdmissionTx(ctx, tx, input.TaskID); err != nil {
+		return TaskAdmission{}, err
+	}
+	if err := tx.Commit(ctx); err != nil {
+		return TaskAdmission{}, err
+	}
+	return admission, nil
+}
+
 func (s *Store) tryTaskAdmission(
 	ctx context.Context,
 	input TaskAdmissionInput,
@@ -196,7 +303,7 @@ WHERE id = $1::uuid`, input.TaskID).Scan(&taskActive); err != nil {
 		if err := expireTaskAdmissionTx(ctx, tx, input.TaskID); err != nil {
 			return TaskAdmissionResult{}, err
 		}
-		if err := notifyTaskAdmissionTx(ctx, tx, input.TaskID); err != nil {
+		if err := notifyTaskAdmissionTx(ctx, tx, "*"); err != nil {
 			return TaskAdmissionResult{}, err
 		}
 		if err := tx.Commit(ctx); err != nil {
@@ -276,7 +383,10 @@ WHERE id = $1::uuid`, input.TaskID).Scan(&taskActive); err != nil {
 	if err != nil {
 		return TaskAdmissionResult{}, err
 	}
-	if err := notifyTaskAdmissionTx(ctx, tx, input.TaskID); err != nil {
+	// Continue the FIFO chain after one task is admitted. Every API process
+	// only probes the global head it owns, so free capacity is filled without
+	// broadcasting an advisory-lock attempt to every waiter.
+	if err := notifyTaskAdmissionTx(ctx, tx, "*"); err != nil {
 		return TaskAdmissionResult{}, err
 	}
 	if onAdmitted != nil {
@@ -312,6 +422,29 @@ func validateNewAdmissionCapacity(states []admissionScopeState) error {
 	return nil
 }
 
+func validateQueuedAdmissionCapacity(states []admissionScopeState) error {
+	// Waiting asynchronous tasks have not acquired leases yet. Count them as
+	// virtual capacity reservations so a burst cannot bypass concurrent and
+	// queue limits before the dispatcher creates River jobs.
+	for _, state := range states {
+		if state.Scope.ConcurrentLimit <= 0 {
+			continue
+		}
+		virtualUsage := state.Active + float64(state.Waiting)*state.Scope.Amount + state.Scope.Amount
+		if virtualUsage <= state.Scope.ConcurrentLimit {
+			continue
+		}
+		if state.Scope.QueueLimit <= 0 {
+			return concurrencyAdmissionError(state)
+		}
+		virtualQueueDepth := int(math.Ceil(virtualUsage - state.Scope.ConcurrentLimit))
+		if virtualQueueDepth > state.Scope.QueueLimit {
+			return queueFullAdmissionError(state)
+		}
+	}
+	return nil
+}
+
 // RebindWaitingTaskAdmission migrates a queued task to a newly selected
 // candidate without changing its FIFO age or priority. Capacity checks happen
 // while both the old and new queue locks are held.
@@ -469,6 +602,32 @@ WHERE task_id = $1::uuid
 	return nil
 }
 
+// RenewTaskAdmissionWaiters refreshes all synchronous waiters owned by one API
+// process in a single statement. This keeps the 15-second crash-recovery lease
+// without running a full advisory-lock admission transaction per waiter.
+func (s *Store) RenewTaskAdmissionWaiters(ctx context.Context, taskIDs []string, waiterIDs []string) error {
+	if len(taskIDs) == 0 {
+		return nil
+	}
+	if len(taskIDs) != len(waiterIDs) {
+		return errors.New("task admission waiter renewal requires matching task and waiter IDs")
+	}
+	_, err := s.pool.Exec(ctx, `
+UPDATE gateway_task_admissions admission
+SET waiter_lease_expires_at = now() + $3::interval,
+    updated_at = now()
+FROM unnest($1::uuid[], $2::text[]) AS owned(task_id, waiter_id)
+WHERE admission.task_id = owned.task_id
+  AND admission.waiter_id = owned.waiter_id
+  AND admission.mode = 'sync'
+  AND admission.status = 'waiting'`,
+		taskIDs,
+		waiterIDs,
+		admissionWaiterLeaseTTL.String(),
+	)
+	return err
+}
+
 func (s *Store) DeleteTaskAdmission(ctx context.Context, taskID string) error {
 	tx, err := s.pool.Begin(ctx)
 	if err != nil {
@@ -488,7 +647,7 @@ WHERE task_id = $1::uuid
 	if _, err := tx.Exec(ctx, `DELETE FROM gateway_task_admissions WHERE task_id = $1::uuid`, taskID); err != nil {
 		return err
 	}
-	if err := notifyTaskAdmissionTx(ctx, tx, taskID); err != nil {
+	if err := notifyTaskAdmissionTx(ctx, tx, "*"); err != nil {
 		return err
 	}
 	return tx.Commit(ctx)
@@ -546,6 +705,52 @@ LIMIT $1`, limit)
 	return taskIDs, rows.Err()
 }
 
+// ListWaitingTaskAdmissionIDs returns a bounded set of FIFO leaders across
+// platform-model and user-group queues. It is used after capacity is released
+// so API processes wake only plausible queue heads instead of every
+// synchronous waiter.
+func (s *Store) ListWaitingTaskAdmissionIDs(ctx context.Context, limit int) ([]string, error) {
+	if limit <= 0 {
+		limit = 256
+	}
+	rows, err := s.pool.Query(ctx, `
+WITH ranked AS (
+  SELECT task_id,
+         priority,
+         enqueued_at,
+         row_number() OVER (
+           PARTITION BY platform_model_id
+           ORDER BY priority ASC, enqueued_at ASC, task_id ASC
+         ) AS platform_rank,
+         row_number() OVER (
+           PARTITION BY user_group_id
+           ORDER BY priority ASC, enqueued_at ASC, task_id ASC
+         ) AS group_rank
+  FROM gateway_task_admissions
+  WHERE status = 'waiting'
+    AND mode = 'sync'
+)
+SELECT task_id::text
+FROM ranked
+WHERE platform_rank <= $1
+  AND group_rank <= $1
+ORDER BY priority ASC, enqueued_at ASC, task_id ASC
+LIMIT $1`, limit)
+	if err != nil {
+		return nil, err
+	}
+	defer rows.Close()
+	taskIDs := make([]string, 0, limit)
+	for rows.Next() {
+		var taskID string
+		if err := rows.Scan(&taskID); err != nil {
+			return nil, err
+		}
+		taskIDs = append(taskIDs, taskID)
+	}
+	return taskIDs, rows.Err()
+}
+
 func (s *Store) GetTaskAdmission(ctx context.Context, taskID string) (TaskAdmission, error) {
 	admission, found, err := loadTaskAdmission(ctx, s.pool, taskID)
 	if err != nil {
@@ -793,6 +998,12 @@ SELECT COUNT(*)
 FROM gateway_task_admissions
 WHERE status = 'waiting' AND user_group_id::text = $1`, scope.ScopeKey).Scan(&waiting)
 		return waiting, err
+	case "worker_capacity":
+		err := tx.QueryRow(ctx, `
+SELECT COUNT(*)
+FROM gateway_task_admissions
+WHERE status = 'waiting' AND mode = 'async'`).Scan(&waiting)
+		return waiting, err
 	default:
 		return 0, nil
 	}
@@ -813,7 +1024,10 @@ func admissionDeadline(scopes []AdmissionScope) time.Time {
 		}
 	}
 	if maxWait == 0 {
-		maxWait = DefaultQueueMaxWaitSeconds
+		// A task may be registered before a River execution slot is available
+		// even when business queueing is disabled. Keep that infrastructure wait
+		// separate from the policy's default 600-second queue timeout.
+		return time.Now().Add(executionSlotWaitMax)
 	}
 	return time.Now().Add(time.Duration(maxWait) * time.Second)
 }

apps/api/internal/store/admission_queue_integration_test.go

@@ -402,6 +402,59 @@ VALUES ($1::uuid, 1, $2::uuid, $3::uuid, $4, 'running', 'submitting')`,
 		t.Fatalf("post-submission cancellation changed=%v err=%v, want unchanged", changed, err)
 	}
 
+	queuedAtomicTask := createTask(true)
+	queuedHookFailure := errors.New("synthetic queued hook failure")
+	_, err = first.QueueTaskAdmissionWithHook(ctx, inputFor(queuedAtomicTask, 100, ""), func(pgx.Tx) error {
+		return queuedHookFailure
+	})
+	if !errors.Is(err, queuedHookFailure) {
+		t.Fatalf("queued hook failure = %v, want synthetic failure", err)
+	}
+	var queuedAdmissions, queuedLeases int
+	if err := first.pool.QueryRow(ctx, `
+SELECT
+  (SELECT COUNT(*) FROM gateway_task_admissions WHERE task_id = $1::uuid),
+  (SELECT COUNT(*) FROM gateway_concurrency_leases WHERE task_id = $1::uuid AND released_at IS NULL)`,
+		queuedAtomicTask.ID,
+	).Scan(&queuedAdmissions, &queuedLeases); err != nil {
+		t.Fatalf("read rolled back queued hook state: %v", err)
+	}
+	if queuedAdmissions != 0 || queuedLeases != 0 {
+		t.Fatalf("failed queued hook left admissions=%d leases=%d", queuedAdmissions, queuedLeases)
+	}
+	const queuedSyntheticRiverJobID int64 = 987654320
+	queuedAdmission, err := first.QueueTaskAdmissionWithHook(ctx, inputFor(queuedAtomicTask, 100, ""), func(tx pgx.Tx) error {
+		_, hookErr := tx.Exec(ctx, `
+UPDATE gateway_tasks
+SET river_job_id = $2
+WHERE id = $1::uuid`, queuedAtomicTask.ID, queuedSyntheticRiverJobID)
+		return hookErr
+	})
+	if err != nil || queuedAdmission.Status != "waiting" {
+		t.Fatalf("atomic queued admission=%+v err=%v, want waiting", queuedAdmission, err)
+	}
+	var riverJobID int64
+	if err := first.pool.QueryRow(ctx, `
+SELECT
+  river_job_id,
+  (SELECT COUNT(*)
+   FROM gateway_concurrency_leases
+   WHERE task_id = $1::uuid AND released_at IS NULL)
+FROM gateway_tasks
+WHERE id = $1::uuid`, queuedAtomicTask.ID).Scan(&riverJobID, &queuedLeases); err != nil {
+		t.Fatalf("read atomic queued River marker: %v", err)
+	}
+	if riverJobID != queuedSyntheticRiverJobID || queuedLeases != 0 {
+		t.Fatalf("queued River marker=%d leases=%d, want %d/0", riverJobID, queuedLeases, queuedSyntheticRiverJobID)
+	}
+	result, err = second.TryTaskAdmission(ctx, inputFor(queuedAtomicTask, 100, ""))
+	if err != nil || !result.Admitted || len(result.Leases) != 1 {
+		t.Fatalf("worker-time queued admission result=%+v err=%v", result, err)
+	}
+	if err := first.DeleteTaskAdmission(ctx, queuedAtomicTask.ID); err != nil {
+		t.Fatalf("release queued atomic task: %v", err)
+	}
+
 	atomicTask := createTask(true)
 	hookFailure := errors.New("synthetic admitted hook failure")
 	_, err = first.TryTaskAdmissionWithAdmittedHook(ctx, inputFor(atomicTask, 100, ""), func(pgx.Tx) error {
@@ -433,7 +486,6 @@ WHERE id = $1::uuid`, atomicTask.ID, syntheticRiverJobID)
 	if err != nil || !result.Admitted {
 		t.Fatalf("atomic admitted hook result=%+v err=%v", result, err)
 	}
-	var riverJobID int64
 	if err := first.pool.QueryRow(ctx, `
 SELECT river_job_id
 FROM gateway_tasks

apps/api/internal/store/postgres.go

@@ -26,6 +26,7 @@ type Store struct {
 const (
 	postgresApplicationName = "easyai-ai-gateway"
 	postgresConnectTimeout  = 5 * time.Second
+	postgresPoolWarmLimit   = 64
 )
 
 func defaultAPIKeyScopes() []string {
@@ -88,7 +89,7 @@ func ConnectWithMaxConns(ctx context.Context, databaseURL string, maxConns int)
 	if err != nil {
 		return nil, err
 	}
-	if err := pool.Ping(ctx); err != nil {
+	if err := warmPostgresPool(ctx, pool, int(config.MinIdleConns)); err != nil {
 		pool.Close()
 		return nil, err
 	}
@@ -104,10 +105,31 @@ func postgresPoolConfig(databaseURL string, maxConns ...int) (*pgxpool.Config, e
 	config.ConnConfig.RuntimeParams["application_name"] = postgresApplicationName
 	if len(maxConns) > 0 && maxConns[0] > 0 {
 		config.MaxConns = int32(maxConns[0])
+		config.MinIdleConns = int32(min(maxConns[0], postgresPoolWarmLimit))
 	}
 	return config, nil
 }
 
+func warmPostgresPool(ctx context.Context, pool *pgxpool.Pool, count int) error {
+	if count <= 0 {
+		return pool.Ping(ctx)
+	}
+	connections := make([]*pgxpool.Conn, 0, count)
+	defer func() {
+		for _, connection := range connections {
+			connection.Release()
+		}
+	}()
+	for len(connections) < count {
+		connection, err := pool.Acquire(ctx)
+		if err != nil {
+			return err
+		}
+		connections = append(connections, connection)
+	}
+	return nil
+}
+
 func IsPostgresUnavailable(err error) bool {
 	if err == nil {
 		return false

apps/api/internal/store/postgres_config_test.go

@@ -27,6 +27,34 @@ func TestPostgresPoolConfigRejectsMalformedURL(t *testing.T) {
 	}
 }
 
+func TestPostgresPoolConfigWarmsBoundedIdleConnections(t *testing.T) {
+	config, err := postgresPoolConfig(
+		"postgresql://gateway:password@localhost:5432/gateway?sslmode=disable",
+		24,
+	)
+	if err != nil {
+		t.Fatalf("parse PostgreSQL pool config: %v", err)
+	}
+	if config.MaxConns != 24 || config.MinIdleConns != 24 {
+		t.Fatalf(
+			"pool bounds max=%d minIdle=%d, want 24/24",
+			config.MaxConns,
+			config.MinIdleConns,
+		)
+	}
+
+	small, err := postgresPoolConfig(
+		"postgresql://gateway:password@localhost:5432/gateway?sslmode=disable",
+		4,
+	)
+	if err != nil {
+		t.Fatalf("parse small PostgreSQL pool config: %v", err)
+	}
+	if small.MinIdleConns != 4 {
+		t.Fatalf("small pool minIdle=%d, want 4", small.MinIdleConns)
+	}
+}
+
 func TestIsPostgresUnavailableClassifiesConnectivityFailures(t *testing.T) {
 	for _, testCase := range []struct {
 		name string

apps/api/internal/store/rate_limit_policy_test.go

@@ -1,6 +1,10 @@
 package store
 
-import "testing"
+import (
+	"errors"
+	"testing"
+	"time"
+)
 
 func TestNormalizeAndValidateQueuePolicy(t *testing.T) {
 	t.Run("defaults maximum wait and preserves extensions", func(t *testing.T) {
@@ -227,3 +231,64 @@ func TestAsyncWorkerCapacityRespectsUserGroupCeiling(t *testing.T) {
 		})
 	}
 }
+
+func TestAdmissionDeadlineSeparatesExecutionSlotWaitFromPolicyQueueTimeout(t *testing.T) {
+	now := time.Now()
+	infrastructureDeadline := admissionDeadline([]AdmissionScope{{
+		ConcurrentLimit: 10,
+	}})
+	if wait := infrastructureDeadline.Sub(now); wait < 23*time.Hour || wait > executionSlotWaitMax+time.Second {
+		t.Fatalf("execution-slot deadline wait=%s, want about %s", wait, executionSlotWaitMax)
+	}
+
+	defaultPolicyDeadline := admissionDeadline([]AdmissionScope{{
+		ConcurrentLimit: 10,
+		QueueLimit:      20,
+	}})
+	if wait := defaultPolicyDeadline.Sub(now); wait < 599*time.Second || wait > 601*time.Second {
+		t.Fatalf("default policy deadline wait=%s, want about 600s", wait)
+	}
+
+	explicitPolicyDeadline := admissionDeadline([]AdmissionScope{{
+		ConcurrentLimit: 10,
+		QueueLimit:      20,
+		MaxWaitSeconds:  90,
+	}})
+	if wait := explicitPolicyDeadline.Sub(now); wait < 89*time.Second || wait > 91*time.Second {
+		t.Fatalf("explicit policy deadline wait=%s, want about 90s", wait)
+	}
+}
+
+func TestQueuedAdmissionCountsFIFOWaitersAsVirtualCapacity(t *testing.T) {
+	scope := AdmissionScope{
+		ScopeType:       "platform_model",
+		ScopeKey:        "model-1",
+		ConcurrentLimit: 1,
+		Amount:          1,
+		QueueLimit:      2,
+	}
+	for waiting := 0; waiting <= 2; waiting++ {
+		if err := validateQueuedAdmissionCapacity([]admissionScopeState{{
+			Scope:   scope,
+			Waiting: waiting,
+		}}); err != nil {
+			t.Fatalf("waiting=%d rejected within concurrent=1 queue=2: %v", waiting, err)
+		}
+	}
+	err := validateQueuedAdmissionCapacity([]admissionScopeState{{
+		Scope:   scope,
+		Waiting: 3,
+	}})
+	var limitErr *RateLimitExceededError
+	if !errors.As(err, &limitErr) || limitErr.Reason != "queue_full" {
+		t.Fatalf("fourth task error=%v, want queue_full", err)
+	}
+
+	scope.QueueLimit = 0
+	if err := validateQueuedAdmissionCapacity([]admissionScopeState{{
+		Scope:   scope,
+		Waiting: 1,
+	}}); !errors.As(err, &limitErr) || limitErr.Metric != "concurrent" {
+		t.Fatalf("queue-disabled second task error=%v, want concurrent rate limit", err)
+	}
+}

apps/api/internal/store/tasks_runtime.go

@@ -684,7 +684,7 @@ WHERE task_id = $1::uuid
 		if _, err := tx.Exec(ctx, `DELETE FROM gateway_task_admissions WHERE task_id = $1::uuid`, taskID); err != nil {
 			return err
 		}
-		return notifyTaskAdmissionTx(ctx, tx, taskID)
+		return notifyTaskAdmissionTx(ctx, tx, "*")
 	})
 	if err != nil {
 		return GatewayTask{}, false, err
@@ -770,7 +770,7 @@ WHERE task_id = $1::uuid
 		if _, err := tx.Exec(ctx, `DELETE FROM gateway_task_admissions WHERE task_id = $1::uuid`, taskID); err != nil {
 			return err
 		}
-		return notifyTaskAdmissionTx(ctx, tx, taskID)
+		return notifyTaskAdmissionTx(ctx, tx, "*")
 	})
 	if err != nil {
 		return GatewayTask{}, false, err

apps/api/internal/store/worker_registry.go

@@ -29,6 +29,18 @@ type WorkerAllocation struct {
 	HeartbeatAt     time.Time
 }
 
+// ActiveWorkerCapacity returns the cluster-wide execution capacity currently
+// owned by live Worker instances.
+func (s *Store) ActiveWorkerCapacity(ctx context.Context) (int, error) {
+	var capacity int
+	err := s.pool.QueryRow(ctx, `
+SELECT COALESCE(SUM(allocated_capacity), 0)::int
+FROM gateway_worker_instances
+WHERE status = 'active'
+  AND heartbeat_at > now() - $1::interval`, workerHeartbeatStaleAfter.String()).Scan(&capacity)
+	return capacity, err
+}
+
 func (s *Store) RegisterWorkerInstance(ctx context.Context, input WorkerRegistrationInput) (WorkerAllocation, error) {
 	input.InstanceID = strings.TrimSpace(input.InstanceID)
 	if input.InstanceID == "" {

deploy/kubernetes/production/application.yaml

@@ -54,6 +54,10 @@ spec:
               value: "false"
             - name: AI_GATEWAY_DATABASE_MAX_CONNS
               value: "16"
+            - name: AI_GATEWAY_MEDIA_REQUEST_CONCURRENCY
+              value: "16"
+            - name: AI_GATEWAY_MEDIA_MATERIALIZATION_CONCURRENCY
+              value: "8"
             - name: POD_NAMESPACE
               valueFrom:
                 fieldRef:
@@ -179,6 +183,10 @@ spec:
               value: "false"
             - name: AI_GATEWAY_DATABASE_MAX_CONNS
               value: "16"
+            - name: AI_GATEWAY_MEDIA_REQUEST_CONCURRENCY
+              value: "16"
+            - name: AI_GATEWAY_MEDIA_MATERIALIZATION_CONCURRENCY
+              value: "8"
             - name: POD_NAMESPACE
               valueFrom:
                 fieldRef:
@@ -304,6 +312,8 @@ spec:
               value: "true"
             - name: AI_GATEWAY_DATABASE_MAX_CONNS
               value: "24"
+            - name: AI_GATEWAY_MEDIA_MATERIALIZATION_CONCURRENCY
+              value: "8"
             - name: POD_NAMESPACE
               valueFrom:
                 fieldRef:
@@ -429,6 +439,8 @@ spec:
               value: "true"
             - name: AI_GATEWAY_DATABASE_MAX_CONNS
               value: "24"
+            - name: AI_GATEWAY_MEDIA_MATERIALIZATION_CONCURRENCY
+              value: "8"
             - name: POD_NAMESPACE
               valueFrom:
                 fieldRef: