@cartanova/qgrid-cli 2.10.3 → 2.10.5

package/bundle/src/application/qgrid/qgrid-run-lifecycle.ts

@@ -1,8 +1,5 @@
 import { getLogger } from "@logtape/logtape";
-import { ServiceUnavailableException } from "sonamu";
 
-import { SD } from "../../i18n/sd.generated";
-import { isRestartPending } from "../../utils/server-restart";
 import { RequestLogModel } from "../request-log/request-log.model";
 import {
   estimateImageGenerationCostMicroUsd,
@@ -20,44 +17,6 @@ const STALE_RUN_THRESHOLD_MS = 30 * 60 * 1000;
 const STALE_CLEANUP_INTERVAL_MS = 60 * 1000;
 let staleCleanupInFlight: Promise<void> | undefined;
 let lastStaleCleanupStartedAt = 0;
-const activeNativeRunRefs = new Map<number, number>();
-const interruptedNativeRunIds = new Set<number>();
-const activeNativeFinalizers = new Map<number, Set<Promise<void>>>();
-
-function restartUnavailable(): ServiceUnavailableException {
-  return new ServiceUnavailableException(SD("qgrid.restartPending")());
-}
-
-function registerNativeRun(requestLogId: number): void {
-  activeNativeRunRefs.set(requestLogId, (activeNativeRunRefs.get(requestLogId) ?? 0) + 1);
-}
-
-function releaseNativeRun(requestLogId: number): void {
-  const refs = activeNativeRunRefs.get(requestLogId);
-  if (refs === undefined || refs <= 1) activeNativeRunRefs.delete(requestLogId);
-  else activeNativeRunRefs.set(requestLogId, refs - 1);
-}
-
-function trackNativeFinalizer<T>(requestLogId: number, finalize: () => Promise<T>): Promise<T> {
-  const operation = finalize();
-  const settled = operation.then(
-    () => undefined,
-    () => undefined,
-  );
-  const finalizers = activeNativeFinalizers.get(requestLogId) ?? new Set<Promise<void>>();
-  finalizers.add(settled);
-  activeNativeFinalizers.set(requestLogId, finalizers);
-  void settled.then(() => {
-    finalizers.delete(settled);
-    if (finalizers.size === 0) activeNativeFinalizers.delete(requestLogId);
-  });
-  return operation;
-}
-
-export function assertNativeRunAdmission(): void {
-  if (isRestartPending()) throw restartUnavailable();
-}
-
 function withProviderPrefix(
   routeModel: string | undefined,
   model: string | undefined,
@@ -110,8 +69,6 @@ export async function beforeQuery(args: QueryInput): Promise<{
   requestLogId: number;
   stepIndex: number;
 }> {
-  assertNativeRunAdmission();
-
   let requestLogId: number;
   let stepIndex = 0;
 
@@ -140,14 +97,6 @@ export async function beforeQuery(args: QueryInput): Promise<{
     });
   }
 
-  // DB 준비 중에 재시작이 시작됐다면 provider 실행으로 넘어가지 않는다.
-  // 이 요청은 재시작 snapshot에 없었으므로 방금 준비한 run을 직접 마감한다.
-  if (isRestartPending()) {
-    await finishTerminalRun(requestLogId, "error", "server restarted");
-    throw restartUnavailable();
-  }
-
-  registerNativeRun(requestLogId);
   return { requestLogId, stepIndex };
 }
 
@@ -157,102 +106,94 @@ export async function afterQuery(
   args: QueryInput,
   result: QueryOutput,
 ): Promise<RunLifecycleResult> {
-  if (interruptedNativeRunIds.has(requestLogId)) return {};
+  // provider가 canonical model을 돌려줘도 요청 route 자체([1m] 등)는 보존한다.
+  const requestedModelName = withProviderPrefix(
+    args.model,
+    args.model ?? result.requestedModel ?? result.model,
+  );
+  const servedModelName = withProviderPrefix(args.model, result.model);
+
+  // generate step 기록
+  await RequestLogModel.appendStep(requestLogId, {
+    step_index: stepIndex,
+    type: "generate",
+    model_name: servedModelName,
+    requested_model_name: requestedModelName,
+    fallback_count: result.modelFallbacks?.length ?? 0,
+    input_tokens: result.usage.input_tokens,
+    output_tokens: result.usage.output_tokens,
+    cache_read_tokens: result.usage.cache_read_input_tokens,
+    cache_creation_tokens: result.usage.cache_creation_input_tokens,
+    cache_creation_5m_tokens: result.usage.cache_creation_5m_input_tokens,
+    cache_creation_1h_tokens: result.usage.cache_creation_1h_input_tokens,
+    cost_usd: Math.round(result.costUsd * 1_000_000),
+    cost_source: result.costSource,
+    duration_ms: result.durationMs,
+    ttft_ms: result.ttftMs,
+    finish_reason: result.finishReason,
+  });
 
-  return trackNativeFinalizer(requestLogId, async () => {
-    try {
-      // provider가 canonical model을 돌려줘도 요청 route 자체([1m] 등)는 보존한다.
-      const requestedModelName = withProviderPrefix(
-        args.model,
-        args.model ?? result.requestedModel ?? result.model,
-      );
-      const servedModelName = withProviderPrefix(args.model, result.model);
+  const imageParts = getImageParts(result);
+  const imageCostMicroUsd = estimateImageGenerationCostMicroUsd(
+    result,
+    args.imageGenerationOptions,
+  );
+  for (const step of buildImageGenerationToolSteps(args, imageParts, stepIndex)) {
+    await RequestLogModel.appendStep(requestLogId, step);
+  }
 
-      // generate step 기록
+  if (result.finishReason === "tool-calls") {
+    // tool-call step을 즉시 기록하고, 다음 follow-up이 같은 row에 결과를 채운다.
+    const toolCalls = result.content.filter((c) => c.type === "tool-call");
+    for (let i = 0; i < toolCalls.length; i++) {
+      const tc = toolCalls[i]!;
+      if (tc.type !== "tool-call") continue;
       await RequestLogModel.appendStep(requestLogId, {
         step_index: stepIndex,
-        type: "generate",
-        model_name: servedModelName,
-        requested_model_name: requestedModelName,
-        fallback_count: result.modelFallbacks?.length ?? 0,
-        input_tokens: result.usage.input_tokens,
-        output_tokens: result.usage.output_tokens,
-        cache_read_tokens: result.usage.cache_read_input_tokens,
-        cache_creation_tokens: result.usage.cache_creation_input_tokens,
-        cache_creation_5m_tokens: result.usage.cache_creation_5m_input_tokens,
-        cache_creation_1h_tokens: result.usage.cache_creation_1h_input_tokens,
-        cost_usd: Math.round(result.costUsd * 1_000_000),
-        cost_source: result.costSource,
-        duration_ms: result.durationMs,
-        ttft_ms: result.ttftMs,
-        finish_reason: result.finishReason,
+        type: "tool_call",
+        tool_call_index: imageParts.length + i,
+        tool_call_id: tc.toolCallId,
+        tool_name: tc.toolName,
+        tool_args: tc.input,
       });
+    }
 
-      const imageParts = getImageParts(result);
-      const imageCostMicroUsd = estimateImageGenerationCostMicroUsd(
-        result,
-        args.imageGenerationOptions,
-      );
-      for (const step of buildImageGenerationToolSteps(args, imageParts, stepIndex)) {
-        await RequestLogModel.appendStep(requestLogId, step);
-      }
-
-      if (result.finishReason === "tool-calls") {
-        // tool-call step을 즉시 기록하고, 다음 follow-up이 같은 row에 결과를 채운다.
-        const toolCalls = result.content.filter((c) => c.type === "tool-call");
-        for (let i = 0; i < toolCalls.length; i++) {
-          const tc = toolCalls[i]!;
-          if (tc.type !== "tool-call") continue;
-          await RequestLogModel.appendStep(requestLogId, {
-            step_index: stepIndex,
-            type: "tool_call",
-            tool_call_index: imageParts.length + i,
-            tool_call_id: tc.toolCallId,
-            tool_name: tc.toolName,
-            tool_args: tc.input,
-          });
-        }
-
-        return { runContext: { requestLogId } };
-      }
-
-      // finishReason === "stop" → run 종료 (전체 step usage 합산)
-      const agg = await RequestLogModel.aggregateStepUsage(requestLogId);
-      const responseText = formatResponseForLog(result);
-      await RequestLogModel.finishRun(requestLogId, {
-        status: "succeeded",
-        response: responseText,
-        // structured 요청만 판정한다. deti 배치의 broken JSON 류를 목록에서 바로
-        // 구별하기 위한 기록이며, 비structured 요청은 null 로 남는다.
-        response_json_ok: args.jsonSchema ? isParseableJson(responseText) : null,
-        token_name: result.tokenName,
-        input_tokens: agg.input_tokens,
-        output_tokens: agg.output_tokens,
-        cache_read_tokens: agg.cache_read_tokens,
-        cache_creation_tokens: agg.cache_creation_tokens,
-        cache_creation_5m_tokens: agg.cache_creation_5m_tokens,
-        cache_creation_1h_tokens: agg.cache_creation_1h_tokens,
-        duration_ms: agg.duration_ms,
-        // multi-step run 부모의 모델은 최종 turn을 서빙한 모델이다.
-        requested_model_name: requestedModelName,
-        model_name: servedModelName,
-        fallback_count: agg.fallback_count,
-        cost_usd: agg.cost_usd,
-        cost_source: agg.cost_source,
-        image_cost_usd: imageCostMicroUsd,
-        ...(imageParts.length > 0 ? { is_image_generation: true } : {}),
-        image_cost_method:
-          imageCostMicroUsd !== null
-            ? imageGenerationCostMethod(args.imageGenerationOptions, result)
-            : null,
-        history: filterHistoryForStorage(args.history),
-      });
+    return { runContext: { requestLogId } };
+  }
 
-      return {};
-    } finally {
-      releaseNativeRun(requestLogId);
-    }
+  // finishReason === "stop" → run 종료 (전체 step usage 합산)
+  const agg = await RequestLogModel.aggregateStepUsage(requestLogId);
+  const responseText = formatResponseForLog(result);
+  await RequestLogModel.finishRun(requestLogId, {
+    status: "succeeded",
+    response: responseText,
+    // structured 요청만 판정한다. deti 배치의 broken JSON 류를 목록에서 바로
+    // 구별하기 위한 기록이며, 비structured 요청은 null 로 남는다.
+    response_json_ok: args.jsonSchema ? isParseableJson(responseText) : null,
+    token_name: result.tokenName,
+    input_tokens: agg.input_tokens,
+    output_tokens: agg.output_tokens,
+    cache_read_tokens: agg.cache_read_tokens,
+    cache_creation_tokens: agg.cache_creation_tokens,
+    cache_creation_5m_tokens: agg.cache_creation_5m_tokens,
+    cache_creation_1h_tokens: agg.cache_creation_1h_tokens,
+    duration_ms: agg.duration_ms,
+    // multi-step run 부모의 모델은 최종 turn을 서빙한 모델이다.
+    requested_model_name: requestedModelName,
+    model_name: servedModelName,
+    fallback_count: agg.fallback_count,
+    cost_usd: agg.cost_usd,
+    cost_source: agg.cost_source,
+    image_cost_usd: imageCostMicroUsd,
+    ...(imageParts.length > 0 ? { is_image_generation: true } : {}),
+    image_cost_method:
+      imageCostMicroUsd !== null
+        ? imageGenerationCostMethod(args.imageGenerationOptions, result)
+        : null,
+    history: filterHistoryForStorage(args.history),
   });
+
+  return {};
 }
 
 export async function finishRunWithError(
@@ -260,48 +201,11 @@ export async function finishRunWithError(
   errorMessage: string,
   args?: QueryInput,
 ): Promise<void> {
-  if (interruptedNativeRunIds.has(requestLogId)) return;
-  return trackNativeFinalizer(requestLogId, async () => {
-    try {
-      await finishTerminalRun(requestLogId, "error", errorMessage, args);
-    } finally {
-      releaseNativeRun(requestLogId);
-    }
-  });
+  await finishTerminalRun(requestLogId, "error", errorMessage, args);
 }
 
 export async function finishRunAborted(requestLogId: number, args?: QueryInput): Promise<void> {
-  if (interruptedNativeRunIds.has(requestLogId)) return;
-  return trackNativeFinalizer(requestLogId, async () => {
-    try {
-      await finishTerminalRun(requestLogId, "aborted", "client disconnected", args);
-    } finally {
-      releaseNativeRun(requestLogId);
-    }
-  });
-}
-
-/** 현재 이 프로세스가 provider를 실행 중인 native run만 재시작 오류로 마감한다. */
-export async function finishActiveNativeRunsForRestart(): Promise<void> {
-  const requestLogIds = [...activeNativeRunRefs.keys()];
-  for (const requestLogId of requestLogIds) {
-    interruptedNativeRunIds.add(requestLogId);
-    activeNativeRunRefs.delete(requestLogId);
-  }
-
-  await Promise.all(
-    requestLogIds.map(async (requestLogId) => {
-      await Promise.all(activeNativeFinalizers.get(requestLogId) ?? []);
-      await finishTerminalRun(requestLogId, "error", "server restarted");
-    }),
-  );
-}
-
-/** 프로세스 전역 registry를 격리해야 하는 단위 테스트 전용. */
-export function resetNativeRunRegistryForTests(): void {
-  activeNativeRunRefs.clear();
-  interruptedNativeRunIds.clear();
-  activeNativeFinalizers.clear();
+  await finishTerminalRun(requestLogId, "aborted", "client disconnected", args);
 }
 
 async function finishTerminalRun(

package/bundle/src/application/qgrid/qgrid.frame.test.ts

@@ -21,7 +21,6 @@ const {
   dispatcherQueryStreamMock,
   beforeQueryMock,
   afterQueryMock,
-  assertNativeRunAdmissionMock,
   finishRunWithErrorMock,
   finishRunAbortedMock,
   getRateLimitsByTokenIdMock,
@@ -45,7 +44,6 @@ const {
     dispatcherQueryStreamMock: vi.fn(),
     beforeQueryMock: vi.fn(),
     afterQueryMock: vi.fn(),
-    assertNativeRunAdmissionMock: vi.fn(),
     finishRunWithErrorMock: vi.fn(),
     finishRunAbortedMock: vi.fn(),
     getRateLimitsByTokenIdMock: vi.fn(),
@@ -95,7 +93,6 @@ vi.mock("./qgrid.dispatcher", async (importOriginal) => {
 });
 
 vi.mock("./qgrid-run-lifecycle", () => ({
-  assertNativeRunAdmission: assertNativeRunAdmissionMock,
   beforeQuery: beforeQueryMock,
   afterQuery: afterQueryMock,
   finishRunWithError: finishRunWithErrorMock,
@@ -215,7 +212,6 @@ describe("QgridFrame.query request logging", () => {
   beforeEach(() => {
     beforeQueryMock.mockReset().mockResolvedValue({ requestLogId: 41, stepIndex: 0 });
     afterQueryMock.mockReset().mockResolvedValue({});
-    assertNativeRunAdmissionMock.mockReset();
     finishRunWithErrorMock.mockReset();
     dispatcherQueryMock.mockReset();
     findActiveByProviderAndNameMock.mockReset();
@@ -368,18 +364,6 @@ describe("QgridFrame.query request logging", () => {
     expect(finishRunWithErrorMock).not.toHaveBeenCalled();
   });
 
-  it("rejects logger-disabled native dispatch while restart admission is closed", async () => {
-    assertNativeRunAdmissionMock.mockImplementationOnce(() => {
-      throw Object.assign(new Error("server restarting"), { statusCode: 503 });
-    });
-
-    await expect(
-      QgridFrame.query({ prompt: "hi", model: "openai/gpt-5-codex", logger: false }),
-    ).rejects.toMatchObject({ statusCode: 503 });
-    expect(dispatcherQueryMock).not.toHaveBeenCalled();
-    expect(beforeQueryMock).not.toHaveBeenCalled();
-  });
-
   it("merges a tool run id with the provider thread coordinate", async () => {
     dispatcherQueryMock.mockResolvedValueOnce(queryOutput());
     afterQueryMock.mockResolvedValueOnce({ runContext: { requestLogId: 41 } });
@@ -952,7 +936,6 @@ describe("QgridFrame.queryStream request logging", () => {
   beforeEach(() => {
     beforeQueryMock.mockReset().mockResolvedValue({ requestLogId: 52, stepIndex: 0 });
     afterQueryMock.mockReset().mockResolvedValue({});
-    assertNativeRunAdmissionMock.mockReset();
     finishRunWithErrorMock.mockReset();
     finishRunAbortedMock.mockReset();
     dispatcherQueryStreamMock.mockReset();
@@ -986,27 +969,6 @@ describe("QgridFrame.queryStream request logging", () => {
     };
   }
 
-  it("rejects stream preparation while restart admission is closed", async () => {
-    assertNativeRunAdmissionMock.mockImplementationOnce(() => {
-      throw Object.assign(new Error("server restarting"), { statusCode: 503 });
-    });
-
-    await expect(QgridFrame.prepareStream({ prompt: "hi" })).rejects.toMatchObject({
-      statusCode: 503,
-    });
-    expect(dispatcherQueryStreamMock).not.toHaveBeenCalled();
-  });
-
-  it("rechecks admission before dispatching a previously prepared stream", async () => {
-    const { streamId } = await QgridFrame.prepareStream({ prompt: "hi", logger: false });
-    assertNativeRunAdmissionMock.mockImplementationOnce(() => {
-      throw Object.assign(new Error("server restarting"), { statusCode: 503 });
-    });
-
-    await expect(QgridFrame.queryStream(streamId)).rejects.toMatchObject({ statusCode: 503 });
-    expect(dispatcherQueryStreamMock).not.toHaveBeenCalled();
-  });
-
   function installSseContext() {
     let closeHandler: (() => void) | undefined;
     const sse = {

package/bundle/src/application/qgrid/qgrid.frame.ts

@@ -38,7 +38,6 @@ import {
   RefreshFailedError,
 } from "./oauth";
 import {
-  assertNativeRunAdmission,
   afterQuery,
   beforeQuery,
   finishRunAborted,
@@ -248,7 +247,6 @@ class QgridFrameClass extends BaseFrameClass {
 
   @api({ httpMethod: "POST", clients: ["axios", "tanstack-mutation"] })
   async query(args: QueryInput): Promise<QueryOutput> {
-    assertNativeRunAdmission();
     rejectInvalidCallerSchemas(args);
     const resolvedArgs = await resolveTokenName(args);
     const disconnect = createHttpDisconnectHandle();
@@ -299,7 +297,6 @@ class QgridFrameClass extends BaseFrameClass {
 
   @api({ httpMethod: "POST", clients: ["axios", "tanstack-mutation"] })
   async prepareStream(args: QueryInput): Promise<{ streamId: string }> {
-    assertNativeRunAdmission();
     rejectImageGenerationStream(args);
     rejectInvalidCallerSchemas(args);
     const resolvedArgs = await resolveTokenName(args);
@@ -315,7 +312,6 @@ class QgridFrameClass extends BaseFrameClass {
     const args = pendingStreams.get(streamId);
     pendingStreams.delete(streamId);
     if (!args) throw new Error("invalid or expired streamId");
-    assertNativeRunAdmission();
     rejectImageGenerationStream(args);
 
     const ctx = Sonamu.getContext();
@@ -504,8 +500,10 @@ class QgridFrameClass extends BaseFrameClass {
   // 화면의 목록 필터를 그대로 받아 같은 조건의 합계를 돌려준다 — 목록과 비용이
   // 서로 다른 모수를 보면 사용자에게는 숫자가 어긋난 것으로 읽힌다.
   @api({ httpMethod: "GET", clients: ["axios", "tanstack-query"] })
-  async totalCost(params?: RequestLogListParams): Promise<{ usd: number }> {
-    return { usd: await RequestLogModel.totalCost(params) };
+  async totalCost(
+    params?: RequestLogListParams,
+  ): Promise<{ usd: number; unknownImageCostCount: number }> {
+    return RequestLogModel.totalCost(params);
   }
 
   @api({ httpMethod: "GET", clients: ["axios", "tanstack-query"] })

package/bundle/src/application/qgrid/qgrid.types.ts

@@ -12,10 +12,10 @@ export const QgridTool = z.object({
 });
 export type QgridTool = z.infer<typeof QgridTool>;
 
-// Observed standalone Images response values, distinct from requested generation hints.
+// Observed image response values, distinct from requested generation hints.
 export const ImageGenerationMetadata = z.object({
-  route: z.literal("codex-images"),
-  model: z.literal("gpt-image-2"),
+  route: z.enum(["codex-images", "codex-responses"]),
+  model: z.string().optional(),
   background: z.string().optional(),
   quality: z.string().optional(),
   size: z.string().optional(),
@@ -79,7 +79,17 @@ export type ServiceTier = z.infer<typeof ServiceTier>;
 export const ImageGenerationQuality = z.enum(["low", "medium", "high"]);
 export type ImageGenerationQuality = z.infer<typeof ImageGenerationQuality>;
 
-export const ImageGenerationSize = z.enum(["1024x1024", "1024x1536", "1536x1024"]);
+// Subscription-backend hints, not public Images API resolution guarantees.
+export const ImageGenerationSize = z.union([
+  z.literal("auto"),
+  z
+    .string()
+    .regex(/^[1-9]\d*x[1-9]\d*$/)
+    .refine(
+      (size) => size.split("x").every((edge) => Number.isSafeInteger(Number(edge))),
+      "Image dimensions must be positive safe integers",
+    ),
+]);
 export type ImageGenerationSize = z.infer<typeof ImageGenerationSize>;
 
 export const ImageGenerationOptions = z

package/bundle/src/application/request-log/request-log.model.test.ts

@@ -564,7 +564,7 @@ describe("RequestLogModel cost provenance", () => {
    * @param confirmedTotal SUM 쿼리가 돌려줄 micro-USD 합계
    * @param legacyRows 재계산 대상 row 들
    */
-  function mockTotalCostQueries(confirmedTotal: number, legacyRows: unknown[]) {
+  function mockTotalCostQueries(confirmedTotal: number, legacyRows: unknown[], unknownImageCostCount = 0) {
     // subset qb 는 SELECT 를 비우고(clear) 집계/축소된 컬럼만 다시 얹는다 — 구현과 같은 체인.
     const sumChain = {
       where: vi.fn(),
@@ -572,7 +572,7 @@ describe("RequestLogModel cost provenance", () => {
       whereRaw: vi.fn(),
       clear: vi.fn(),
       select: vi.fn(),
-      first: vi.fn(async () => ({ total: confirmedTotal })),
+      first: vi.fn(async () => ({ total: confirmedTotal, unknownImageCostCount })),
     };
     sumChain.where.mockReturnValue(sumChain);
     sumChain.whereIn.mockReturnValue(sumChain);
@@ -606,7 +606,7 @@ describe("RequestLogModel cost provenance", () => {
   it("uses exact stored cost for rows with a cost source", async () => {
     mockTotalCostQueries(3_000, []);
 
-    await expect(RequestLogModel.totalCost()).resolves.toBe(0.003);
+    await expect(RequestLogModel.totalCost()).resolves.toEqual({ usd: 0.003, unknownImageCostCount: 0 });
   });
 
   it("sums every row in SQL without a cost_source predicate", async () => {
@@ -637,7 +637,7 @@ describe("RequestLogModel cost provenance", () => {
       },
     ]);
 
-    await expect(RequestLogModel.totalCost()).resolves.toBeCloseTo(0.4725, 10);
+    await expect(RequestLogModel.totalCost().then((result) => result.usd)).resolves.toBeCloseTo(0.4725, 10);
   });
 
   it("adds only the delta for legacy rows already counted in the SQL sum", async () => {
@@ -656,7 +656,7 @@ describe("RequestLogModel cost provenance", () => {
       },
     ]);
 
-    await expect(RequestLogModel.totalCost()).resolves.toBeCloseTo(0.4725, 10);
+    await expect(RequestLogModel.totalCost().then((result) => result.usd)).resolves.toBeCloseTo(0.4725, 10);
   });
 
   it("leaves the stored value untouched for legacy rows without a model name", async () => {
@@ -675,7 +675,7 @@ describe("RequestLogModel cost provenance", () => {
       },
     ]);
 
-    await expect(RequestLogModel.totalCost()).resolves.toBeCloseTo(0.005, 10);
+    await expect(RequestLogModel.totalCost().then((result) => result.usd)).resolves.toBeCloseTo(0.005, 10);
   });
 
   it("applies list filters to both cost queries so totals match the filtered list", async () => {
@@ -721,7 +721,7 @@ describe("RequestLogModel cost provenance", () => {
       },
     ]);
 
-    await expect(RequestLogModel.totalCost()).resolves.toBeCloseTo(0.4755, 10);
+    await expect(RequestLogModel.totalCost().then((result) => result.usd)).resolves.toBeCloseTo(0.4755, 10);
   });
 });
 
@@ -738,10 +738,13 @@ describe("RequestLogModel PostgreSQL", () => {
         { cost_usd: 5_000, image_cost_usd: 10_000, cost_source: "provider" },
         { cost_usd: 1_000, image_cost_usd: 20_000, cost_source: null },
         { cost_usd: 900_000, image_cost_usd: 700_000, cost_source: null },
+        { cost_usd: 1_000, image_cost_usd: null, cost_source: "provider", is_image_generation: true },
+        { cost_usd: 2_000, image_cost_usd: null, cost_source: "provider", is_image_generation: false },
+        { cost_usd: 500_000, image_cost_usd: null, cost_source: "provider", is_image_generation: true },
       ].entries()) {
         const id = await RequestLogModel.createRun({
           user_prompt: "cost regression",
-          project_name: index < 2 ? projectName : `${projectName}-other`,
+          project_name: index === 2 || index === 5 ? `${projectName}-other` : projectName,
         });
         ids.push(id);
         await RequestLogModel.getPuri("w").table("request_logs").where("id", id).update({
@@ -754,10 +757,12 @@ describe("RequestLogModel PostgreSQL", () => {
 
       const params = { sonamuFilter: { project_name: projectName }, num: 0, page: 1 };
       const { rows } = await RequestLogModel.findMany("C", params);
-      expect(rows.map((row) => row.id)).toEqual(expect.arrayContaining(ids.slice(0, 2)));
-      expect(rows).toHaveLength(2);
-      // Confirmed: .005 + .010; legacy driver repriced to .003, plus .020 image.
-      await expect(RequestLogModel.totalCost(params)).resolves.toBeCloseTo(0.038, 10);
+      expect(rows.map((row) => row.id)).toEqual(expect.arrayContaining([ids[0], ids[1], ids[3], ids[4]]));
+      expect(rows).toHaveLength(4);
+      // .038 known images + .001 hosted driver + .002 text; other project excluded.
+      const total = await RequestLogModel.totalCost(params);
+      expect(total.usd).toBeCloseTo(0.041, 10);
+      expect(total.unknownImageCostCount).toBe(1);
     } finally {
       await RequestLogModel.del(ids);
     }

package/bundle/src/application/request-log/request-log.model.ts

@@ -820,7 +820,9 @@ class RequestLogModelClass extends BaseModelClass<
    * 필터 조건을 여기서 다시 구현하지 않고 `applyListFilters` 를 공유하는 것이 핵심이다 —
    * 두 쿼리가 각자 필터를 가지면 화면에서 필터를 걸었을 때 행 수와 비용이 조용히 어긋난다.
    */
-  async totalCost(rawParams: RequestLogListParams = { num: 0, page: 1 }): Promise<number> {
+  async totalCost(
+    rawParams: RequestLogListParams = { num: 0, page: 1 },
+  ): Promise<{ usd: number; unknownImageCostCount: number }> {
     const params = {
       num: 0,
       page: 1,
@@ -843,6 +845,9 @@ class RequestLogModelClass extends BaseModelClass<
         total: Puri.rawNumber(
           "COALESCE(SUM(GREATEST(request_logs.cost_usd, 0) + COALESCE(request_logs.image_cost_usd, 0)), 0)",
         ),
+        unknownImageCostCount: Puri.rawNumber(
+          "COUNT(*) FILTER (WHERE request_logs.is_image_generation = true AND request_logs.image_cost_usd IS NULL)",
+        ),
       })
       .first();
 
@@ -867,7 +872,7 @@ class RequestLogModelClass extends BaseModelClass<
     const storedTotal = Number(stored?.total ?? 0) / MICRO_USD;
 
     // legacy row 는 위 SUM 에 저장값으로 이미 포함되어 있다. 재계산값과의 차액만 더한다.
-    return legacyRows.reduce((sum, row) => {
+    const usd = legacyRows.reduce((sum, row) => {
       const alreadySummed = Math.max(row.cost_usd ?? 0, 0) / MICRO_USD;
       const usage = normalizedUsageForCost(row);
       if (!usage.model) return sum;
@@ -884,6 +889,7 @@ class RequestLogModelClass extends BaseModelClass<
         alreadySummed
       );
     }, storedTotal);
+    return { usd, unknownImageCostCount: Number(stored?.unknownImageCostCount ?? 0) };
   }
 
   async distinctProjectNames(): Promise<string[]> {