@cartanova/qgrid-cli 2.9.1 → 2.9.4

package/bundle/src/utils/providers/common/model-cost.ts

@@ -10,6 +10,8 @@
  * @see https://platform.claude.com/docs/en/about-claude/pricing
  */
 
+import { type OpenAIEffort } from "./effort";
+
 export interface ModelCosts {
   inputTokens: number;
   outputTokens: number;
@@ -32,7 +34,7 @@ export interface ModelCosts {
 
 // ── OpenAI ─────────────────────────────────────────────────────────
 //
-// 가격 출처: https://developers.openai.com/api/docs/models (2026-07-18 확인)
+// Standard pricing verified 2026-09-07: https://developers.openai.com/api/docs/pricing
 // 신모델 출시마다 단가가 바뀌므로(5.2→5.4→5.5) 모델 추가 시 반드시 공식 페이지 재확인해야함
 
 // GPT-5.4에서 처음 도입된 long-context 할증 (5.2/5.3-codex는 해당 없음, 5.4-mini/nano는 공식 표에서 long-context 단가 없음)
@@ -45,31 +47,53 @@ const LONG_CONTEXT_272K: NonNullable<ModelCosts["longContext"]> = {
   outputMultiplier: 1.5,
 };
 
-const OPENAI_COSTS: Record<string, ModelCosts> = {
+// OpenAI 모델 카탈로그. 단가와 함께 Codex 백엔드가 모델별로 받는 최대 reasoning effort 를 한 곳에 둔다.
+// maxEffort 출처: ~/.codex/models_cache.json 의 supported_reasoning_levels (2026-09-07 확인).
+// 미기재 모델은 xhigh 까지만 받는다.
+type OpenAIModelSpec = ModelCosts & { maxEffort?: OpenAIEffort };
+
+const OPENAI_COSTS: Record<string, OpenAIModelSpec> = {
+  // GPT-6 Astra: Standard API-equivalent pricing, verified 2026-09-07.
+  // Codex subscription catalog supports ultra; the public API lists only through max.
+  // @see https://developers.openai.com/api/docs/models/gpt-6-astra
+  // @see https://developers.openai.com/api/docs/pricing
+  "gpt-6-astra": {
+    maxEffort: "ultra",
+    inputTokens: 10,
+    outputTokens: 50,
+    cachedInputTokens: 1,
+    cacheCreationInputTokens: 12.5,
+    longContext: LONG_CONTEXT_272K,
+  },
   // GPT-5.6 Sol, Terra, Luna. cache write 단가는 외부 logger/manual usage 입력을 위해
   // 유지한다. 현재 OpenAI 응답 usage 에 cache write 토큰 필드가 없으면 비용은 0으로 계산된다.
   // @see https://developers.openai.com/api/docs/models/gpt-5.6-sol
   // @see https://developers.openai.com/api/docs/models/gpt-5.6-terra
   // @see https://developers.openai.com/api/docs/models/gpt-5.6-luna
+  // Sol promotional rates are available at least through 2026-11-21.
+  // No confirmed end date/replacement rate: recheck official pricing instead of auto-reverting.
   "gpt-5.6-sol": {
-    inputTokens: 5,
-    outputTokens: 30,
-    cachedInputTokens: 0.5,
-    cacheCreationInputTokens: 6.25,
+    maxEffort: "ultra",
+    inputTokens: 4,
+    outputTokens: 20,
+    cachedInputTokens: 0.4,
+    cacheCreationInputTokens: 5,
     longContext: LONG_CONTEXT_272K,
   },
   "gpt-5.6-terra": {
-    inputTokens: 2.5,
-    outputTokens: 15,
-    cachedInputTokens: 0.25,
-    cacheCreationInputTokens: 3.125,
+    maxEffort: "ultra",
+    inputTokens: 2,
+    outputTokens: 12,
+    cachedInputTokens: 0.2,
+    cacheCreationInputTokens: 2.5,
     longContext: LONG_CONTEXT_272K,
   },
   "gpt-5.6-luna": {
-    inputTokens: 1,
-    outputTokens: 6,
-    cachedInputTokens: 0.1,
-    cacheCreationInputTokens: 1.25,
+    maxEffort: "max",
+    inputTokens: 0.2,
+    outputTokens: 1.2,
+    cachedInputTokens: 0.02,
+    cacheCreationInputTokens: 0.25,
     longContext: LONG_CONTEXT_272K,
   },
   // https://openai.com/index/introducing-gpt-5-5/ (2026-04 출시)
@@ -148,6 +172,11 @@ const ANTHROPIC_COSTS: Record<string, ModelCosts> = {
 // @see https://help.openai.com/en/articles/20001106-codex-rate-card
 const DEFAULT_COSTS: ModelCosts = { inputTokens: 3, outputTokens: 15, cachedInputTokens: 0.3 };
 
+/** Codex 백엔드가 이 모델에 허용하는 최대 effort. 카탈로그에 없는 모델은 xhigh 까지다. */
+export function openaiModelMaxEffort(model: string): OpenAIEffort {
+  return OPENAI_COSTS[model]?.maxEffort ?? "xhigh";
+}
+
 export function getModelCosts(model: string): ModelCosts {
   const normalizedModel = (model.split("/").pop() ?? model).replace(/\[1m\]$/i, "");
   return OPENAI_COSTS[normalizedModel] ?? ANTHROPIC_COSTS[normalizedModel] ?? DEFAULT_COSTS;

package/bundle/src/utils/providers/openai/openai-backend-protocol.ts

@@ -1,5 +1,7 @@
 import os from "node:os";
 
+import { type OpenAIEffort } from "../common/effort";
+
 export const CHATGPT_CODEX_RESPONSES_URL =
   "https://chatgpt.com/backend-api/codex/responses" as const;
 
@@ -58,7 +60,7 @@ export interface OpenAIResponsesOptions {
   toolChoice?: string | OpenAIResponseItem;
   parallelToolCalls?: boolean;
   reasoning?: {
-    effort?: "none" | "minimal" | "low" | "medium" | "high" | "xhigh";
+    effort?: OpenAIEffort;
     summary?: "auto" | "concise" | "detailed";
   };
   verbosity?: "low" | "medium" | "high";

package/bundle/src/utils/providers/openai/openai-dispatcher.test.ts

@@ -416,3 +416,24 @@ describe("OpenAIDispatcher direct runtime", () => {
     await expect(blocked.generate(request())).rejects.toThrow("NO_OPENAI_WORKERS");
   });
 });
+
+describe("effort 어휘 해석 (Codex 카탈로그 기준)", () => {
+  async function mappedEffortFor(model: string, effort: string): Promise<unknown> {
+    let mapped: OpenAIResponsesOptions | undefined;
+    const d = dispatcher((options) => {
+      mapped = options;
+      return events({ type: "completed", responseId: "r1" });
+    });
+    await d.onTokenAdded(7, "primary", credentials, null, 1);
+    await d.generate(request({ model, effort }));
+    return mapped?.reasoning?.effort;
+  }
+
+  it("모델이 지원하는 값은 그대로 보내고, 상한 초과·공개 API 어휘는 reasoning 에서 빼 백엔드 기본값을 쓴다", async () => {
+    await expect(mappedEffortFor("gpt-6-astra", "ultra")).resolves.toBe("ultra");
+    await expect(mappedEffortFor("gpt-5.6-terra", "ultra")).resolves.toBe("ultra");
+    await expect(mappedEffortFor("gpt-5.5", "xhigh")).resolves.toBe("xhigh");
+    await expect(mappedEffortFor("gpt-5.5", "max")).resolves.toBeUndefined();
+    await expect(mappedEffortFor("gpt-5.6-terra", "minimal")).resolves.toBeUndefined();
+  });
+});

package/bundle/src/utils/providers/openai/openai-dispatcher.ts

@@ -3,6 +3,7 @@ import { getLogger } from "@logtape/logtape";
 import { QuotaThresholdExceededError } from "../../../application/qgrid/qgrid.types";
 import { TokenModel } from "../../../application/token/token.model";
 import { type OpenAICredentials } from "../../../application/token/token.types";
+import { resolveOpenAIEffort } from "../common/effort";
 import {
   type GenerateRequest,
   type GenerateResult,
@@ -97,11 +98,11 @@ function inputItems(req: GenerateRequest): OpenAIResponseItem[] {
 function asReasoning(req: GenerateRequest): OpenAIResponsesOptions["reasoning"] | undefined {
   const summary =
     req.reasoningSummary && req.reasoningSummary !== "none" ? req.reasoningSummary : undefined;
-  if (!req.effort && !summary) return undefined;
+  // Codex 어휘·모델 상한 밖의 effort 는 여기서 조용히 버린다(백엔드 기본값 적용).
+  const effort = resolveOpenAIEffort(req.model ?? "", req.effort);
+  if (!effort && !summary) return undefined;
   return {
-    ...(req.effort
-      ? { effort: req.effort as NonNullable<OpenAIResponsesOptions["reasoning"]>["effort"] }
-      : {}),
+    ...(effort ? { effort } : {}),
     ...(summary
       ? {
           summary: summary as NonNullable<OpenAIResponsesOptions["reasoning"]>["summary"],