@cartanova/qgrid-cli 2.4.2 → 2.4.4

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

@@ -95,6 +95,15 @@ function firstRefreshTokenArg() {
   return call;
 }
 
+async function selectedTokenNames(
+  dispatcher: AnthropicDispatcher,
+  count: number,
+): Promise<string[]> {
+  return Promise.all(
+    Array.from({ length: count }, async () => (await dispatcher.generate(baseReq())).tokenName),
+  );
+}
+
 describe("AnthropicDispatcher", () => {
   beforeEach(() => {
     runClaudeSessionMock.mockReset();
@@ -113,7 +122,7 @@ describe("AnthropicDispatcher", () => {
 
   it("happy: generate → GenerateResult (threadCoord 조립, systemHash 없음)", async () => {
     const d = new AnthropicDispatcher();
-    d.onTokenAdded(1, "tok-A", creds());
+    d.onTokenAdded(1, "tok-A", creds(), null, 1);
     runClaudeSessionMock.mockResolvedValueOnce(sessionResult({ ttftMs: 42 }));
 
     const result = await d.generate(baseReq());
@@ -127,7 +136,7 @@ describe("AnthropicDispatcher", () => {
 
   it("cold 호출: coldInput/coldHistory 를 그대로 전달하고 continuation session id 는 없다", async () => {
     const d = new AnthropicDispatcher();
-    d.onTokenAdded(1, "tok-A", creds());
+    d.onTokenAdded(1, "tok-A", creds(), null, 1);
 
     await d.generate(
       baseReq({ coldHistory: [{ type: "message", role: "assistant", content: [] }] }),
@@ -141,7 +150,7 @@ describe("AnthropicDispatcher", () => {
 
   it("reuse/reuseInput 이 실려 와도 무시하고 coldInput/coldHistory 로 실행한다", async () => {
     const d = new AnthropicDispatcher();
-    d.onTokenAdded(1, "tok-A", creds());
+    d.onTokenAdded(1, "tok-A", creds(), null, 1);
 
     await d.generate(
       baseReq({
@@ -159,7 +168,7 @@ describe("AnthropicDispatcher", () => {
 
   it("structured(outputSchema): jsonSchema 로 직렬화되어 전달", async () => {
     const d = new AnthropicDispatcher();
-    d.onTokenAdded(1, "tok-A", creds());
+    d.onTokenAdded(1, "tok-A", creds(), null, 1);
     await d.generate(baseReq({ outputSchema: { type: "object", properties: {} } }));
 
     const call = firstRunRequest();
@@ -168,7 +177,7 @@ describe("AnthropicDispatcher", () => {
 
   it("quota 소진 → 에러", async () => {
     const d = new AnthropicDispatcher();
-    d.onTokenAdded(1, "tok-A", creds());
+    d.onTokenAdded(1, "tok-A", creds(), null, 1);
     runClaudeSessionMock.mockResolvedValueOnce(sessionResult({ quotaExhausted: true }));
 
     await expect(d.generate(baseReq())).rejects.toThrow(/quota exhausted/);
@@ -176,7 +185,7 @@ describe("AnthropicDispatcher", () => {
 
   it("claude 에러 → 에러", async () => {
     const d = new AnthropicDispatcher();
-    d.onTokenAdded(1, "tok-A", creds());
+    d.onTokenAdded(1, "tok-A", creds(), null, 1);
     runClaudeSessionMock.mockResolvedValueOnce(sessionResult({ isError: true, text: "boom" }));
 
     await expect(d.generate(baseReq())).rejects.toThrow(/claude error/);
@@ -184,7 +193,7 @@ describe("AnthropicDispatcher", () => {
 
   it("cold 실패는 resume retry 없이 그대로 전파한다", async () => {
     const d = new AnthropicDispatcher();
-    d.onTokenAdded(1, "tok-A", creds());
+    d.onTokenAdded(1, "tok-A", creds(), null, 1);
     runClaudeSessionMock.mockRejectedValueOnce(new Error("closed without result"));
 
     await expect(
@@ -198,25 +207,46 @@ describe("AnthropicDispatcher", () => {
     expect(runClaudeSessionMock).toHaveBeenCalledTimes(1);
   });
 
-  it("least-used RR: 두 토큰 번갈아 분배", async () => {
+  it("routes Anthropic requests in a 3:1 ratio", async () => {
     const d = new AnthropicDispatcher();
-    d.onTokenAdded(1, "tok-A", creds());
-    d.onTokenAdded(2, "tok-B", creds());
+    d.onTokenAdded(1, "tok-A", creds(), null, 3);
+    d.onTokenAdded(2, "tok-B", creds(), null, 1);
 
-    const names = new Set<string>();
-    const r1 = await d.generate(baseReq());
-    const r2 = await d.generate(baseReq());
-    names.add(r1.tokenName);
-    names.add(r2.tokenName);
+    const names = await selectedTokenNames(d, 4);
 
-    expect(names.size).toBe(2);
+    expect(names.filter((name) => name === "tok-A")).toHaveLength(3);
+    expect(names.filter((name) => name === "tok-B")).toHaveLength(1);
     expect(readAnthropicQuotaUsageMock).not.toHaveBeenCalled();
   });
 
+  it("recomputes weighted selection from quota-eligible tokens", async () => {
+    const d = new AnthropicDispatcher();
+    d.onTokenAdded(1, "tok-A", creds(), 80, 10);
+    d.onTokenAdded(2, "tok-B", creds(), 80, 1);
+    readAnthropicQuotaUsageMock
+      .mockResolvedValueOnce(quotaOk(90))
+      .mockResolvedValueOnce(quotaOk(10));
+
+    expect((await d.generate(baseReq())).tokenName).toBe("tok-B");
+  });
+
+  it("resets the schedule when a token weight changes", async () => {
+    const d = new AnthropicDispatcher();
+    d.onTokenAdded(1, "tok-A", creds(), null, 1);
+    d.onTokenAdded(2, "tok-B", creds(), null, 1);
+    await d.generate(baseReq());
+
+    d.onTokenUpdated(1, "tok-A", creds(), null, 1);
+    d.onTokenUpdated(2, "tok-B", creds(), null, 3);
+    const names = await selectedTokenNames(d, 4);
+
+    expect(names.filter((name) => name === "tok-B")).toHaveLength(3);
+  });
+
   it("threshold 초과 토큰은 후보에서 제외하고 미설정 토큰을 선택한다", async () => {
     const d = new AnthropicDispatcher();
-    d.onTokenAdded(1, "tok-A", creds(), 80);
-    d.onTokenAdded(2, "tok-B", creds());
+    d.onTokenAdded(1, "tok-A", creds(), 80, 1);
+    d.onTokenAdded(2, "tok-B", creds(), null, 1);
     readAnthropicQuotaUsageMock.mockResolvedValueOnce(quotaOk(85, 1_000));
 
     const result = await d.generate(baseReq());
@@ -239,8 +269,8 @@ describe("AnthropicDispatcher", () => {
 
   it("threshold 경계는 utilization >= threshold 일 때 제외한다", async () => {
     const d = new AnthropicDispatcher();
-    d.onTokenAdded(1, "tok-A", creds(), 80);
-    d.onTokenAdded(2, "tok-B", creds(), 80);
+    d.onTokenAdded(1, "tok-A", creds(), 80, 1);
+    d.onTokenAdded(2, "tok-B", creds(), 80, 1);
     readAnthropicQuotaUsageMock
       .mockResolvedValueOnce(quotaOk(80))
       .mockResolvedValueOnce(quotaOk(79));
@@ -252,7 +282,7 @@ describe("AnthropicDispatcher", () => {
 
   it("quota 조회 실패는 fail-open 으로 통과시키고 lookup_fail_open 로그를 남긴다", async () => {
     const d = new AnthropicDispatcher();
-    d.onTokenAdded(1, "tok-A", creds(), 80);
+    d.onTokenAdded(1, "tok-A", creds(), 80, 1);
     readAnthropicQuotaUsageMock.mockResolvedValueOnce(quotaFail("timeout"));
 
     const result = await d.generate(baseReq());
@@ -272,7 +302,7 @@ describe("AnthropicDispatcher", () => {
 
   it("utilization 0 은 정상 조회로 보고 fail-open 로그를 남기지 않는다", async () => {
     const d = new AnthropicDispatcher();
-    d.onTokenAdded(1, "tok-A", creds(), 80);
+    d.onTokenAdded(1, "tok-A", creds(), 80, 1);
     readAnthropicQuotaUsageMock.mockResolvedValueOnce(quotaOk(0));
 
     const result = await d.generate(baseReq());
@@ -286,8 +316,8 @@ describe("AnthropicDispatcher", () => {
 
   it("모든 threshold 설정 토큰이 초과되면 typed error 로 실패한다", async () => {
     const d = new AnthropicDispatcher();
-    d.onTokenAdded(1, "tok-A", creds(), 80);
-    d.onTokenAdded(2, "tok-B", creds(), 90);
+    d.onTokenAdded(1, "tok-A", creds(), 80, 1);
+    d.onTokenAdded(2, "tok-B", creds(), 90, 1);
     readAnthropicQuotaUsageMock
       .mockResolvedValueOnce(quotaOk(85, 100))
       .mockResolvedValueOnce(quotaOk(95, 200));
@@ -301,11 +331,11 @@ describe("AnthropicDispatcher", () => {
     );
   });
 
-  it("minCount 는 초과 토큰을 제외한 eligible 집합에서 계산한다", async () => {
+  it("가중 선택은 quota 초과 토큰을 제외한 eligible 집합에서 계산한다", async () => {
     const d = new AnthropicDispatcher();
-    d.onTokenAdded(1, "tok-A", creds(), 80);
-    d.onTokenAdded(2, "tok-B", creds(), 80);
-    d.onTokenAdded(3, "tok-C", creds(), 80);
+    d.onTokenAdded(1, "tok-A", creds(), 80, 1);
+    d.onTokenAdded(2, "tok-B", creds(), 80, 1);
+    d.onTokenAdded(3, "tok-C", creds(), 80, 1);
 
     readAnthropicQuotaUsageMock
       .mockResolvedValueOnce(quotaOk(10))
@@ -325,10 +355,10 @@ describe("AnthropicDispatcher", () => {
     expect(new Set(["tok-B", "tok-C"]).has(result.tokenName)).toBe(true);
   });
 
-  it("동시 요청은 quota await 이후 charge 선반영으로 서로 다른 eligible 토큰을 고른다", async () => {
+  it("동시 요청은 quota await 이후 동기 선택으로 서로 다른 eligible 토큰을 고른다", async () => {
     const d = new AnthropicDispatcher();
-    d.onTokenAdded(1, "tok-A", creds(), 80);
-    d.onTokenAdded(2, "tok-B", creds(), 80);
+    d.onTokenAdded(1, "tok-A", creds(), 80, 1);
+    d.onTokenAdded(2, "tok-B", creds(), 80, 1);
     readAnthropicQuotaUsageMock.mockResolvedValue(quotaOk(0));
 
     const results = await Promise.all([d.generate(baseReq()), d.generate(baseReq())]);
@@ -338,7 +368,7 @@ describe("AnthropicDispatcher", () => {
 
   it("onTokenRemoved 후 그 토큰 미선택", async () => {
     const d = new AnthropicDispatcher();
-    d.onTokenAdded(1, "tok-A", creds());
+    d.onTokenAdded(1, "tok-A", creds(), null, 1);
     d.onTokenRemoved(1);
 
     await expect(d.generate(baseReq())).rejects.toThrow(/No anthropic tokens/);
@@ -346,8 +376,8 @@ describe("AnthropicDispatcher", () => {
 
   it("onTokenUpdated 는 토큰 credentials 를 in-place 갱신한다", async () => {
     const d = new AnthropicDispatcher();
-    d.onTokenAdded(1, "tok-A", creds(3_600_000, "acc-1"));
-    d.onTokenUpdated(1, "tok-A", creds(7_200_000, "acc-2"));
+    d.onTokenAdded(1, "tok-A", creds(3_600_000, "acc-1"), null, 1);
+    d.onTokenUpdated(1, "tok-A", creds(7_200_000, "acc-2"), null, 1);
 
     await d.generate(baseReq());
 
@@ -357,7 +387,7 @@ describe("AnthropicDispatcher", () => {
 
   it("model prefix 정규화: 'anthropic/claude-opus-4-8' → canonical 로 세션/result.model", async () => {
     const d = new AnthropicDispatcher();
-    d.onTokenAdded(1, "tok-A", creds());
+    d.onTokenAdded(1, "tok-A", creds(), null, 1);
 
     const result = await d.generate(baseReq({ model: "anthropic/claude-opus-4-8" }));
 
@@ -368,7 +398,7 @@ describe("AnthropicDispatcher", () => {
 
   it("[1m] suffix 정규화: result/runClaudeSession 에는 base canonical 만 전달", async () => {
     const d = new AnthropicDispatcher();
-    d.onTokenAdded(1, "tok-A", creds());
+    d.onTokenAdded(1, "tok-A", creds(), null, 1);
 
     const result = await d.generate(baseReq({ model: "anthropic/claude-sonnet-4-6[1m]" }));
 
@@ -379,7 +409,7 @@ describe("AnthropicDispatcher", () => {
 
   it("unsupported alias + [1m] 은 조용히 다운그레이드하지 않는다", async () => {
     const d = new AnthropicDispatcher();
-    d.onTokenAdded(1, "tok-A", creds());
+    d.onTokenAdded(1, "tok-A", creds(), null, 1);
 
     await expect(d.generate(baseReq({ model: "sonnet[1m]" }))).rejects.toThrow(
       /Unsupported Anthropic 1M model suffix/,
@@ -388,7 +418,7 @@ describe("AnthropicDispatcher", () => {
 
   it("model 미지정 → ANTHROPIC_DEFAULT_MODEL 적용", async () => {
     const d = new AnthropicDispatcher();
-    d.onTokenAdded(1, "tok-A", creds());
+    d.onTokenAdded(1, "tok-A", creds(), null, 1);
 
     const result = await d.generate(baseReq({ model: undefined }));
 
@@ -399,9 +429,9 @@ describe("AnthropicDispatcher", () => {
 
   it("replaceTokens: DB 기준 풀 재동기화 — 없는 토큰 제거, 새 토큰 추가", async () => {
     const d = new AnthropicDispatcher();
-    d.onTokenAdded(1, "tok-A", creds());
+    d.onTokenAdded(1, "tok-A", creds(), null, 1);
 
-    d.replaceTokens([{ id: 2, name: "tok-B", credentials: creds() }]);
+    d.replaceTokens([{ id: 2, name: "tok-B", credentials: creds(), weight: 1 }]);
     const result = await d.generate(baseReq());
 
     expect(result.tokenName).toBe("tok-B");
@@ -409,7 +439,7 @@ describe("AnthropicDispatcher", () => {
 
   it("generateStream: onDelta/onComplete 호출", async () => {
     const d = new AnthropicDispatcher();
-    d.onTokenAdded(1, "tok-A", creds());
+    d.onTokenAdded(1, "tok-A", creds(), null, 1);
     runClaudeSessionMock.mockImplementationOnce(async (_req, onDelta: (t: string) => void) => {
       onDelta("부분");
       return sessionResult({ text: "부분완성" });
@@ -430,7 +460,7 @@ describe("AnthropicDispatcher", () => {
 
   it("generateStream: Claude server error 는 onError 로 전달하고 complete 하지 않는다", async () => {
     const d = new AnthropicDispatcher();
-    d.onTokenAdded(1, "tok-A", creds());
+    d.onTokenAdded(1, "tok-A", creds(), null, 1);
     const serverError = new Error(
       "claude error (anthropic/yds): API Error: 529 Overloaded. This is a server-side issue",
     );
@@ -449,7 +479,7 @@ describe("AnthropicDispatcher", () => {
 
   it("refresh: 만료 임박 토큰은 provider 포함해 refreshToken 호출, 새 access token 으로 세션 진행", async () => {
     const d = new AnthropicDispatcher();
-    d.onTokenAdded(1, "tok-A", creds(30_000));
+    d.onTokenAdded(1, "tok-A", creds(30_000), null, 1);
     refreshTokenMock.mockResolvedValueOnce("sk-ant-oat01-refreshed");
 
     await d.generate(baseReq());
@@ -463,7 +493,7 @@ describe("AnthropicDispatcher", () => {
 
   it("refresh: 만료 여유 있으면 refresh 안 함, 기존 access token 사용", async () => {
     const d = new AnthropicDispatcher();
-    d.onTokenAdded(1, "tok-A", creds(3_600_000));
+    d.onTokenAdded(1, "tok-A", creds(3_600_000), null, 1);
 
     await d.generate(baseReq());
 
@@ -473,7 +503,7 @@ describe("AnthropicDispatcher", () => {
 
   it("refresh 실패해도 기존 access token 으로 진행 — 요청을 죽이지 않음", async () => {
     const d = new AnthropicDispatcher();
-    d.onTokenAdded(1, "tok-A", creds(30_000));
+    d.onTokenAdded(1, "tok-A", creds(30_000), null, 1);
     refreshTokenMock.mockRejectedValueOnce(new Error("refresh boom"));
 
     const result = await d.generate(baseReq());

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

@@ -2,7 +2,7 @@
  * AnthropicDispatcher
  *
  * OpenAIDispatcher와 달리, worker pool을 두는 대신 요청별 fresh spawn
- * 토큰은 인메모리(MAP)으로 관리하고 least-used, round-robin 으로 고른다
+ * 토큰은 인메모리(MAP)으로 관리하고 smooth weighted round-robin 으로 고른다
  *
  * - 모든 요청은 fresh `--session-id` 로 실행
  * - 멀티턴 문맥은 클라이언트가 매 호출 보내는 full history 를 평탄화해 전달한다.
@@ -23,6 +23,7 @@ import {
   type GenerateStreamCallbacks,
   type ProviderDispatcher,
 } from "../common/provider-dispatcher";
+import { SmoothWeightedRoundRobin } from "../common/smooth-weighted-round-robin";
 import { assertSupportedOneMillionSuffix, canonicalAnthropicModel } from "./anthropic-constants";
 import { readAnthropicQuotaUsage, type AnthropicQuotaUsageResult } from "./anthropic-quota";
 import { makeAnthropicWorkerId, runClaudeSession } from "./claude-session";
@@ -40,6 +41,7 @@ interface PooledToken {
   name: string;
   credentials: AnthropicCredentials;
   quotaThreshold?: number | null;
+  weight: number;
 }
 
 function quotaThresholdExceededMessage(
@@ -57,9 +59,7 @@ function quotaThresholdExceededMessage(
 export class AnthropicDispatcher implements ProviderDispatcher {
   // tokenId → 풀 항목. start()/token 이벤트로 채움
   private tokenPool = new Map<number, PooledToken>();
-  // tokenId 별 사용 카운터(least-used RR). name 이 아니라 id 기준 — 이름 충돌 안전
-  private requestCounts = new Map<number, number>();
-  private rrIndex = 0;
+  private readonly weightedSelector = new SmoothWeightedRoundRobin();
 
   async start(): Promise<void> {
     // OpenAIDispatcher.start() 와 동일하게 DB 에서 기존 anthropic 토큰을 self-bootstrap 한다.
@@ -73,16 +73,18 @@ export class AnthropicDispatcher implements ProviderDispatcher {
         name: t.name,
         credentials: t.credentials as AnthropicCredentials,
         quotaThreshold: t.quota_threshold,
+        weight: t.weight,
       });
+      this.weightedSelector.setToken(t.id, t.weight);
       logger.info(`worker spawned: ${t.name}`);
     }
   }
 
   async stop(): Promise<void> {
+    const tokenIds = [...this.tokenPool.keys()];
+    for (const id of tokenIds) this.weightedSelector.removeToken(id);
     this.tokenPool.clear();
-    // lifecycle 정합: RR 카운터/인덱스도 비워 stop→start 재등록 시 이전 카운트가 안 남게 한다.
-    this.requestCounts.clear();
-    this.rrIndex = 0;
+    this.weightedSelector.resetScores();
   }
 
   // 현재 풀의 토큰 수(startup 로그/health 용).
@@ -101,6 +103,7 @@ export class AnthropicDispatcher implements ProviderDispatcher {
       name: string;
       credentials: AnthropicCredentials;
       quotaThreshold?: number | null;
+      weight: number;
     }>,
   ): void {
     const next = new Set(rows.map((r) => r.id));
@@ -111,9 +114,9 @@ export class AnthropicDispatcher implements ProviderDispatcher {
     }
     for (const r of rows) {
       if (this.tokenPool.has(r.id)) {
-        this.onTokenUpdated(r.id, r.name, r.credentials, r.quotaThreshold);
+        this.onTokenUpdated(r.id, r.name, r.credentials, r.quotaThreshold, r.weight);
       } else {
-        this.onTokenAdded(r.id, r.name, r.credentials, r.quotaThreshold);
+        this.onTokenAdded(r.id, r.name, r.credentials, r.quotaThreshold, r.weight);
       }
     }
   }
@@ -124,8 +127,10 @@ export class AnthropicDispatcher implements ProviderDispatcher {
     name: string,
     credentials: AnthropicCredentials,
     quotaThreshold?: number | null,
+    weight = 1,
   ): void {
-    this.tokenPool.set(id, { id, name, credentials, quotaThreshold });
+    this.tokenPool.set(id, { id, name, credentials, quotaThreshold, weight });
+    this.weightedSelector.setToken(id, weight);
   }
 
   onTokenUpdated(
@@ -133,20 +138,18 @@ export class AnthropicDispatcher implements ProviderDispatcher {
     name: string,
     credentials: AnthropicCredentials,
     quotaThreshold?: number | null,
+    weight = 1,
   ): void {
-    this.tokenPool.set(id, { id, name, credentials, quotaThreshold });
+    this.tokenPool.set(id, { id, name, credentials, quotaThreshold, weight });
+    this.weightedSelector.setToken(id, weight);
   }
 
   onTokenRemoved(id: number): void {
     this.tokenPool.delete(id);
-    this.requestCounts.delete(id);
+    this.weightedSelector.removeToken(id);
   }
 
-  private countOf(id: number): number {
-    return this.requestCounts.get(id) ?? 0;
-  }
-
-  // 요청마다 least-used RR 로 토큰을 고른다(동점이면 rrIndex 로 회전).
+  // quota 통과 후보만으로 요청마다 smooth weighted round-robin 선택한다.
   private async selectToken(): Promise<PooledToken | null> {
     const rows = [...this.tokenPool.values()];
     if (rows.length === 0) return null;
@@ -164,12 +167,9 @@ export class AnthropicDispatcher implements ProviderDispatcher {
       );
     }
 
-    const minCount = Math.min(...eligible.map((r) => this.countOf(r.id)));
-    const idle = eligible.filter((r) => this.countOf(r.id) === minCount);
-    const picked = idle[this.rrIndex % idle.length];
-    if (!picked) return null;
-    this.rrIndex++;
-    return this.charge(picked);
+    const selectedId = this.weightedSelector.select(new Set(eligible.map((token) => token.id)));
+    if (selectedId === null) return null;
+    return this.tokenPool.get(selectedId) ?? null;
   }
 
   private async filterEligibleTokens(rows: PooledToken[]): Promise<{
@@ -252,12 +252,6 @@ export class AnthropicDispatcher implements ProviderDispatcher {
     });
   }
 
-  // 선택된 토큰의 사용 카운트를 await 전에 선반영(동시 요청이 다른 토큰을 고르도록).
-  private charge(token: PooledToken): PooledToken {
-    this.requestCounts.set(token.id, this.countOf(token.id) + 1);
-    return token;
-  }
-
   async generate(req: GenerateRequest): Promise<GenerateResult> {
     return this.run(req, () => {});
   }

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

@@ -3,6 +3,60 @@ import { describe, expect, it } from "vitest";
 import { calculateCostUsd, getModelCosts } from "./model-cost";
 
 describe("calculateCostUsd", () => {
+  it.each([
+    ["gpt-5.6-sol", 5, 30, 0.5, 6.25, 3.30625],
+    ["gpt-5.6-terra", 2.5, 15, 0.25, 3.125, 1.653125],
+    ["gpt-5.6-luna", 1, 6, 0.1, 1.25, 0.66125],
+  ])(
+    "%s applies published input, output, cache-read, and cache-write pricing",
+    (model, inputTokens, outputTokens, cachedInputTokens, cacheCreationInputTokens, expectedCost) => {
+      expect(getModelCosts(model)).toMatchObject({
+        inputTokens,
+        outputTokens,
+        cachedInputTokens,
+        cacheCreationInputTokens,
+      });
+
+      expect(
+        calculateCostUsd(model, {
+          inputTokens: 100_000,
+          outputTokens: 100_000,
+          cachedInputTokens: 50_000,
+          cacheCreationInputTokens: 25_000,
+        }),
+      ).toBeCloseTo(expectedCost, 10);
+    },
+  );
+
+  it.each([
+    ["gpt-5.6-sol", 1.245],
+    ["gpt-5.6-terra", 0.6225],
+    ["gpt-5.6-luna", 0.249],
+  ])("%s applies the published long-context surcharge", (model, expectedCost) => {
+    expect(
+      calculateCostUsd(model, {
+        inputTokens: 300_000,
+        outputTokens: 1_000,
+        cachedInputTokens: 200_000,
+      }),
+    ).toBeCloseTo(expectedCost, 10);
+  });
+
+  it.each([
+    ["gpt-5.6-sol", 1.37],
+    ["gpt-5.6-terra", 0.685],
+    ["gpt-5.6-luna", 0.274],
+  ])("%s applies the long-context input multiplier to cache writes", (model, expectedCost) => {
+    expect(
+      calculateCostUsd(model, {
+        inputTokens: 300_000,
+        outputTokens: 1_000,
+        cachedInputTokens: 200_000,
+        cacheCreationInputTokens: 50_000,
+      }),
+    ).toBeCloseTo(expectedCost, 10);
+  });
+
   it("Anthropic cache read/write 를 전체 입력에서 분리해 각각 단가를 적용한다", () => {
     const cost = calculateCostUsd("claude-sonnet-4-6", {
       inputTokens: 1_917,

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

@@ -43,6 +43,37 @@ const LONG_CONTEXT_272K: NonNullable<ModelCosts["longContext"]> = {
 };
 
 const OPENAI_COSTS: Record<string, ModelCosts> = {
+  // GPT-5.6 Sol, Terra, Luna. OpenAI native API 는 1.05M context / 128K max output 이지만,
+  // qgrid 의 codex app-server 경로는 context_window=372K (max_context_window 도 372K,
+  // effective 95% ≈ 353K) 로 제한된다 — codex 0.144.1 models_cache 실측 (2026-07-10).
+  // cache write 는 uncached input 의 1.25x 로 과금되지만, codex app-server usage
+  // (TokenUsageBreakdown) 에 cache write 토큰 필드가 없어 현재 OpenAI 실행 경로의
+  // cost 계산에서는 항상 0 으로 잡힌다 → 그만큼 과소집계. 단가는 외부 logger/manual
+  // usage 입력과 향후 프로토콜 확장을 위해 유지한다.
+  // @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
+  "gpt-5.6-sol": {
+    inputTokens: 5,
+    outputTokens: 30,
+    cachedInputTokens: 0.5,
+    cacheCreationInputTokens: 6.25,
+    longContext: LONG_CONTEXT_272K,
+  },
+  "gpt-5.6-terra": {
+    inputTokens: 2.5,
+    outputTokens: 15,
+    cachedInputTokens: 0.25,
+    cacheCreationInputTokens: 3.125,
+    longContext: LONG_CONTEXT_272K,
+  },
+  "gpt-5.6-luna": {
+    inputTokens: 1,
+    outputTokens: 6,
+    cachedInputTokens: 0.1,
+    cacheCreationInputTokens: 1.25,
+    longContext: LONG_CONTEXT_272K,
+  },
   // https://openai.com/index/introducing-gpt-5-5/ (2026-04 출시)
   "gpt-5.5": {
     inputTokens: 5,

package/bundle/src/utils/providers/common/provider-dispatcher.ts

@@ -4,10 +4,10 @@
  * MVP 에서는 stream() 만. getRateLimits(), listModels() 는 future.
  */
 
+import { type ImageGenerationOptions } from "../../../application/qgrid/qgrid.types";
 import { type JsonValue } from "../../../codex-protocol/serde_json/JsonValue";
 import { type TokenUsageBreakdown } from "../../../codex-protocol/v2/TokenUsageBreakdown";
 import { type UserInput } from "../../../codex-protocol/v2/UserInput";
-import { type ImageGenerationOptions } from "../../../application/qgrid/qgrid.types";
 
 // qgrid provider 내부 표준 usage. codex-protocol 의 TokenUsageBreakdown 은 생성 파일이므로
 // provider 전용 cache write 세부 필드는 여기서 확장해 보존한다.

package/bundle/src/utils/providers/common/smooth-weighted-round-robin.test.ts

@@ -0,0 +1,73 @@
+import { describe, expect, it } from "vitest";
+
+import { SmoothWeightedRoundRobin } from "./smooth-weighted-round-robin";
+
+function pickMany(selector: SmoothWeightedRoundRobin, ids: number[], count: number): number[] {
+  const eligible = new Set(ids);
+  return Array.from({ length: count }, () => selector.select(eligible)!);
+}
+
+describe("SmoothWeightedRoundRobin", () => {
+  it("uses token-id order as the deterministic tie break for equal weights", () => {
+    const selector = new SmoothWeightedRoundRobin();
+    selector.setToken(2, 1);
+    selector.setToken(1, 1);
+
+    expect(pickMany(selector, [1, 2], 4)).toEqual([1, 2, 1, 2]);
+  });
+
+  it("produces a smooth 3:1 distribution", () => {
+    const selector = new SmoothWeightedRoundRobin();
+    selector.setToken(1, 3);
+    selector.setToken(2, 1);
+
+    expect(pickMany(selector, [1, 2], 4)).toEqual([1, 1, 2, 1]);
+  });
+
+  it("produces a complete 5:2:1 weighted cycle", () => {
+    const selector = new SmoothWeightedRoundRobin();
+    selector.setToken(1, 5);
+    selector.setToken(2, 2);
+    selector.setToken(3, 1);
+
+    const picked = pickMany(selector, [1, 2, 3], 8);
+    expect(picked.filter((id) => id === 1)).toHaveLength(5);
+    expect(picked.filter((id) => id === 2)).toHaveLength(2);
+    expect(picked.filter((id) => id === 3)).toHaveLength(1);
+  });
+
+  it("selects only from the current eligible set", () => {
+    const selector = new SmoothWeightedRoundRobin();
+    selector.setToken(1, 10);
+    selector.setToken(2, 1);
+
+    expect(selector.select(new Set([2]))).toBe(2);
+    expect(selector.select(new Set([1, 2]))).toBe(1);
+  });
+
+  it("resets scores when a weight changes or a token is removed", () => {
+    const selector = new SmoothWeightedRoundRobin();
+    selector.setToken(1, 3);
+    selector.setToken(2, 1);
+    selector.select(new Set([1, 2]));
+
+    selector.setToken(1, 1);
+    expect(pickMany(selector, [1, 2], 2)).toEqual([1, 2]);
+
+    selector.removeToken(1);
+    expect(selector.select(new Set([1, 2]))).toBe(2);
+  });
+
+  it("returns null for empty or unregistered candidates", () => {
+    const selector = new SmoothWeightedRoundRobin();
+    expect(selector.select(new Set())).toBeNull();
+    expect(selector.select(new Set([99]))).toBeNull();
+  });
+
+  it("rejects weights outside 1 through 100", () => {
+    const selector = new SmoothWeightedRoundRobin();
+    expect(() => selector.setToken(1, 0)).toThrow(RangeError);
+    expect(() => selector.setToken(1, 101)).toThrow(RangeError);
+    expect(() => selector.setToken(1, 1.5)).toThrow(RangeError);
+  });
+});