@cartanova/qgrid-cli 2.10.2 → 2.10.4

package/bundle/src/application/qgrid/token-subscriber.ts

@@ -25,19 +25,15 @@ export class TokenSubscriber {
   attempt = 0;
   connectedAt: Date | null = null;
   lastReconcileAt: Date | null = null;
-  private operationChain: Promise<void> = Promise.resolve();
-  private tokenChangeHandler: (() => void) | null = null;
+  operationChain: Promise<void> = Promise.resolve();
+  tokenChangeHandler: (() => void) | null = null;
 
   constructor(
     public connConfig: ClientConfig,
     public dispatcher: QgridDispatcherClass,
   ) {}
 
-  setTokenChangeHandler(handler: (() => void) | null): void {
-    this.tokenChangeHandler = handler;
-  }
-
-  private notifyTokensChanged(): void {
+  notifyTokensChanged(): void {
     try {
       this.tokenChangeHandler?.();
     } catch (error) {
@@ -168,18 +164,18 @@ export class TokenSubscriber {
     return this.enqueueOperation(() => this.handleNotificationNow(payloadJson));
   }
 
-  private async handleNotificationNow(payloadJson: string): Promise<void> {
+  async handleNotificationNow(payloadJson: string): Promise<void> {
     const payload = JSON.parse(payloadJson) as Payload;
     const previousRow = this.tokenChangeHandler
       ? this.dispatcher.tokens.get(payload.id)
       : undefined;
     const wasKeepaliveTarget = previousRow?.active === true && previousRow.provider === "anthropic";
     if (payload.op === "DELETE") {
-      this.dispatcher.removeCache(payload.id);
+      this.dispatcher.tokens.delete(payload.id);
       await this.dispatcher.openaiDispatcher
         ?.onTokenRemoved(payload.id)
         .catch((e) => logger.warn(`openai token remove failed: ${(e as Error).message}`));
-      // anthropic 토큰 이벤트는 동기(void) — provider 를 모르므로 무해하게 항상 제거 시도.
+      // 삭제된 토큰의 provider를 모를 수 있으므로 양쪽 풀에서 제거한다.
       this.dispatcher.anthropicDispatcher?.onTokenRemoved(payload.id);
       logger.info(`NOTIFY ${payload.op} id=${payload.id} → removed from cache`);
       if (previousRow === undefined || wasKeepaliveTarget) this.notifyTokensChanged();
@@ -187,7 +183,7 @@ export class TokenSubscriber {
     }
     const row = await TokenModel.findOne("A", { id: payload.id });
     if (!row) {
-      this.dispatcher.removeCache(payload.id);
+      this.dispatcher.tokens.delete(payload.id);
       await this.dispatcher.openaiDispatcher
         ?.onTokenRemoved(payload.id)
         .catch((e) => logger.warn(`openai token remove failed: ${(e as Error).message}`));
@@ -197,7 +193,7 @@ export class TokenSubscriber {
       return;
     }
 
-    this.dispatcher.upsertCache(payload.id, row);
+    this.dispatcher.tokens.set(payload.id, row);
 
     if (row.provider === "openai") {
       const creds = row.credentials as Record<string, unknown>;
@@ -229,9 +225,6 @@ export class TokenSubscriber {
         openaiDispatcher?.onTokenDeactivated(payload.id);
       }
     } else if (row.provider === "anthropic") {
-      // anthropic 토큰 이벤트는 동기(void). worker 가 없고 풀(Map)만 관리하므로 단순:
-      //  INSERT → 추가, active 면 갱신(onTokenUpdated 가 identity 변경도 처리), inactive 면 풀에서 제거.
-      //  (별도 activate/deactivate 콜백이 없어 active 토글을 add/remove 로 매핑.)
       const creds = row.credentials as AnthropicCredentials;
       if (payload.op === "INSERT" && row.active) {
         this.dispatcher.anthropicDispatcher?.onTokenAdded(
@@ -250,7 +243,6 @@ export class TokenSubscriber {
           row.weight,
         );
       } else {
-        // inactive 면 INSERT 든 UPDATE 든 풀에 넣지 않는다.
         this.dispatcher.anthropicDispatcher?.onTokenRemoved(payload.id);
       }
     }
@@ -263,7 +255,7 @@ export class TokenSubscriber {
     return this.enqueueOperation(() => this.reconcileNow());
   }
 
-  private async reconcileNow(): Promise<void> {
+  async reconcileNow(): Promise<void> {
     const previousAnthropicIds = this.tokenChangeHandler
       ? new Set(
           [...this.dispatcher.tokens.values()]
@@ -274,8 +266,6 @@ export class TokenSubscriber {
     await TokenModel.deactivateExpiredTokens();
     const rows = await TokenModel.findActive("A");
     this.dispatcher.replaceCache(rows);
-    // NOTIFY 유실 대비: AnthropicDispatcher 풀도 DB active anthropic 토큰 기준으로 재동기화.
-    // (rows 는 active 만 — inactive/삭제된 토큰은 여기 없으므로 replaceTokens 가 풀에서 제거한다.)
     const anthropicRows = rows
       .filter((r) => r.provider === "anthropic")
       .map((r) => ({
@@ -308,7 +298,7 @@ export class TokenSubscriber {
     }
   }
 
-  private enqueueOperation(operation: () => Promise<void>): Promise<void> {
+  enqueueOperation(operation: () => Promise<void>): Promise<void> {
     const next = this.operationChain.then(operation, operation);
     this.operationChain = next.catch(() => {});
     return next;

package/bundle/src/application/qgrid/token-window-keepalive.test.ts

@@ -1,7 +1,6 @@
 import { afterEach, beforeEach, describe, expect, it, vi } from "vitest";
 
 import {
-  KEEPALIVE_MODEL,
   KEEPALIVE_PROJECT_NAME,
   POST_KEEPALIVE_USAGE_DELAY_MS,
   rescheduleTokenWindowKeepalive,
@@ -76,7 +75,8 @@ describe("token window keepalive", () => {
 
     expect(dispatch).toHaveBeenCalledWith({
       prompt: "Reply OK.",
-      model: KEEPALIVE_MODEL,
+      model: "anthropic/claude-haiku-5-5",
+      thinking: false,
       effort: "low",
       projectName: KEEPALIVE_PROJECT_NAME,
       preferredTokenId: 1,

package/bundle/src/application/qgrid/token-window-keepalive.ts

@@ -15,7 +15,7 @@ const RESET_GRACE_MS = 1_000;
 // usage API 성공 응답은 1분 캐시된다. keepalive 직후 즉시 재조회하면 이전 윈도우를
 // 다시 보므로 캐시가 지난 뒤 한 번 확인하고, 그때도 새 윈도우가 없으면 다음 5h까지 기다린다.
 export const POST_KEEPALIVE_USAGE_DELAY_MS = 61_000;
-export const KEEPALIVE_MODEL = "anthropic/claude-haiku-4-5";
+export const KEEPALIVE_MODEL = "anthropic/claude-haiku-5-5";
 export const KEEPALIVE_PROJECT_NAME = "qgrid-token-window-keepalive";
 export const KEEPALIVE_SETTING_KEY = "qgrid.tokenWindowKeepaliveEnabled";
 export const KEEPALIVE_ENV_KEY = TOKEN_WINDOW_KEEPALIVE_RUNNER_ENV_KEY;
@@ -163,6 +163,7 @@ async function inspectToken(
     await deps.dispatch({
       prompt: "Reply OK.",
       model: KEEPALIVE_MODEL,
+      thinking: false,
       effort: "low",
       projectName: KEEPALIVE_PROJECT_NAME,
       preferredTokenId: token.id,

package/bundle/src/application/qgrid/tool-emulation-schema.test.ts

@@ -23,16 +23,16 @@ describe("buildToolCallSchema", () => {
     },
   ];
 
-  it("documents that emulated tools must be requested through StructuredOutput", () => {
+  it("requests tools in the response body and ends the turn until the client returns results", () => {
     const schema = buildToolCallSchema(tools);
 
     expect(schema).toHaveProperty(
       "description",
-      expect.stringContaining("Do not invoke listed client tools as native Claude Code tools"),
+      expect.stringContaining("then end your turn"),
     );
     expect(schema).toHaveProperty(
       ["properties", "result", "anyOf", 1, "description"],
-      expect.stringContaining("Do not call these tools as native"),
+      expect.stringContaining("wait for the next request containing tool results"),
     );
     expect(schema).toHaveProperty(
       [
@@ -49,6 +49,9 @@ describe("buildToolCallSchema", () => {
       ],
       expect.stringContaining("getWeather"),
     );
+    expect(JSON.stringify(schema)).not.toContain("StructuredOutput");
+    expect(JSON.stringify(schema)).not.toContain("Claude Code");
+    expect(schema).toHaveProperty("description", expect.stringContaining("Do not invent tool results"));
   });
 
   it("makes degenerate combinations grammatically impossible", () => {

package/bundle/src/application/qgrid/tool-emulation-schema.ts

@@ -54,7 +54,7 @@ export function buildToolCallSchema(tools: QgridTool[], answerSchema?: JsonValue
   return {
     type: "object",
     description:
-      "Use this schema only through StructuredOutput. Do not invoke listed client tools as native Claude Code tools. To request client-side tool execution, set result.action to tool_call and include result.toolCalls. Use action answer only when no further client tool result is needed.",
+      "Return exactly one JSON object matching this schema as your final response, including when requesting a tool. This object is the response body, not a call to a built-in tool. To request client-side tools, set result.action to tool_call and include result.toolCalls, then end your turn. The client executes those tools and supplies their results in a later request. Do not invent tool results or emit an answer before receiving the required results.",
     properties: {
       result: {
         anyOf: [
@@ -75,7 +75,7 @@ export function buildToolCallSchema(tools: QgridTool[], answerSchema?: JsonValue
           {
             type: "object",
             description:
-              "Request client-side tool execution through this structured output. Do not call these tools as native Claude Code tools.",
+              "Request client-side tool execution. Return this object as the complete response, then wait for the next request containing tool results.",
             properties: {
               action: { type: "string", enum: ["tool_call"] },
               answer: { type: "null" },
@@ -88,7 +88,7 @@ export function buildToolCallSchema(tools: QgridTool[], answerSchema?: JsonValue
                     toolName: {
                       type: "string",
                       enum: tools.map((tool) => tool.name),
-                      description: `Client-side tool name to request through structured output. Do not call this as a native Claude Code tool.\n${toolDescriptions}`,
+                      description: `Client-side tool name. The caller executes these tools after receiving your JSON response.\n${toolDescriptions}`,
                     },
                     args: {
                       type: "string",

package/bundle/src/application/qgrid/tool-emulation.test.ts

@@ -2,7 +2,7 @@ import { describe, expect, it } from "vitest";
 
 import { CALLER_SCHEMA_LIMITS } from "../../utils/providers/common/schema-validation";
 import { type QgridTool } from "./qgrid.types";
-import { applyToolCallEmulation, ToolCallEmulationError } from "./tool-emulation";
+import { applyToolCallEmulation, isToolCallEnvelope, ToolCallEmulationError } from "./tool-emulation";
 
 const baseResult = {
   text: "",
@@ -84,6 +84,13 @@ describe("applyToolCallEmulation image parts", () => {
 });
 
 describe("applyToolCallEmulation envelope validation", () => {
+  it("ignores intermediate prose but never repairs malformed final JSON or accepts an unknown control tool", () => {
+    const tool = JSON.stringify({result: {action: "tool_call", answer: null, toolCalls: [{toolName: "notAllowed", args: "{}"}]}});
+    const allowed = [{name: "lookup", inputSchema: {type: "object"}}];
+    expect(isToolCallEnvelope("Checking the source.", allowed, "json", "commentary")).toBe(false);
+    expect(() => isToolCallEnvelope(tool, allowed, "json", "commentary")).toThrow("unknown emulated tool");
+    expect(() => isToolCallEnvelope(tool + tool, allowed, "json", "final_answer")).toThrow("response is not JSON");
+  });
   const tools: QgridTool[] = [
     {
       name: "lookup",

package/bundle/src/application/qgrid/tool-emulation.ts

@@ -111,6 +111,35 @@ export function applyToolCallEmulation(
   );
 }
 
+export function isToolCallEnvelope(
+  text: string,
+  tools: QgridTool[],
+  answerKind: AnswerKind,
+  phase?: string,
+): boolean {
+  if (phase === "commentary") {
+    // Commentary is not the structured answer. Only an explicit tool control
+    // object is eligible, and that object must still pass the strict parser.
+    let value: unknown;
+    try {
+      value = JSON.parse(text);
+    } catch {
+      return false;
+    }
+    if (
+      !value ||
+      typeof value !== "object" ||
+      !("result" in value) ||
+      !value.result ||
+      typeof value.result !== "object" ||
+      !("action" in value.result) ||
+      value.result.action !== "tool_call"
+    )
+      return false;
+  }
+  return parseEnvelope(text, tools, answerKind).action === "tool_call";
+}
+
 function parseEnvelope(text: string, tools: QgridTool[], answerKind: AnswerKind) {
   if (Buffer.byteLength(text, "utf8") > STRUCTURED_ENVELOPE_LIMITS.maxUtf8Bytes) {
     throw new ToolCallEmulationError(

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

@@ -257,7 +257,7 @@ class RequestLogModelClass extends BaseModelClass<
    * 화면에서 필터를 걸면 행 수와 비용이 함께 좁혀져야 하는데, 두 쿼리가 각자 필터를
    * 구현하면 한쪽에 축을 추가할 때 조용히 어긋난다.
    */
-  private applyListFilters(
+  applyListFilters(
     qb: ReturnType<RequestLogModelClass["getSubsetQueries"]>["qb"],
     params: RequestLogListParams,
   ): void {
@@ -378,7 +378,7 @@ class RequestLogModelClass extends BaseModelClass<
       string,
       { requests: number; errors: number; inputTokens: number; cacheReadTokens: number }
     >();
-    for (const row of rows) {
+    rows.forEach((row) => {
       const model = row.requested_model_name ?? "";
       const provider = model.startsWith("openai/")
         ? "openai"
@@ -396,7 +396,7 @@ class RequestLogModelClass extends BaseModelClass<
       agg.inputTokens += row.input_tokens;
       agg.cacheReadTokens += row.cache_read_tokens;
       byProvider.set(provider, agg);
-    }
+    });
     return [...byProvider.entries()]
       .map(([provider, agg]) => ({ provider, ...agg }))
       .toSorted((a, b) => a.provider.localeCompare(b.provider));
@@ -584,9 +584,9 @@ class RequestLogModelClass extends BaseModelClass<
       "is_image_generation",
       "response_json_ok",
     ] as const;
-    for (const key of fields) {
+    fields.forEach((key) => {
       if (params[key] !== undefined) update[key] = params[key];
-    }
+    });
     if (params.history !== undefined) update.history = params.history;
     update.ttft_ms = (await this.firstGenerateStepTtft(requestLogId)) ?? 0;
 

package/bundle/src/application/setting/setting.model.test.ts

@@ -1,10 +1,6 @@
-import { EventEmitter } from "node:events";
 
-import { Sonamu } from "sonamu";
-import { afterEach, beforeEach, describe, expect, it, vi } from "vitest";
+import { afterEach, describe, expect, it, vi } from "vitest";
 
-import { resetNativeRunRegistryForTests } from "../qgrid/qgrid-run-lifecycle";
-import { resetServerRestartForTests } from "../../utils/server-restart";
 import { SettingModel } from "./setting.model";
 import * as reminders from "../qgrid/expired-token-reminder";
 import { SlackNotificationError } from "../../utils/slack-notify";
@@ -26,99 +22,3 @@ describe("SettingModel.triggerExpiryReminder", () => {
     });
   });
 });
-
-function installContext(headers: Record<string, string>) {
-  const raw = Object.assign(new EventEmitter(), {
-    destroyed: false,
-    writableEnded: false,
-  });
-  vi.spyOn(Sonamu, "getContext").mockReturnValue({
-    transport: "http",
-    headers,
-    request: { protocol: "http" },
-    reply: { raw },
-  } as never);
-  return raw;
-}
-
-describe("SettingModel.restartServer", () => {
-  beforeEach(() => {
-    vi.spyOn(process, "exit").mockImplementation((() => undefined) as never);
-  });
-
-  afterEach(() => {
-    resetServerRestartForTests();
-    resetNativeRunRegistryForTests();
-    vi.unstubAllEnvs();
-    vi.restoreAllMocks();
-  });
-
-  it("rejects restart when pm2 will not respawn the process", async () => {
-    const raw = installContext({ host: "localhost:44900" });
-    vi.stubEnv("pm_id", undefined);
-
-    await expect(SettingModel.restartServer()).rejects.toMatchObject({ statusCode: 400 });
-    expect(raw.listenerCount("finish")).toBe(0);
-    expect(process.exit).not.toHaveBeenCalled();
-  });
-
-  it("rejects a browser request whose Origin host does not match Host", async () => {
-    installContext({ origin: "https://evil.example", host: "qgrid.example.com" });
-    vi.stubEnv("pm_id", "0");
-
-    await expect(SettingModel.restartServer()).rejects.toMatchObject({ statusCode: 403 });
-    expect(process.exit).not.toHaveBeenCalled();
-  });
-
-  it("rejects a malformed present Origin instead of treating it as an operational client", async () => {
-    installContext({ origin: "null", host: "qgrid.example.com" });
-    vi.stubEnv("pm_id", "0");
-
-    await expect(SettingModel.restartServer()).rejects.toMatchObject({ statusCode: 403 });
-    expect(process.exit).not.toHaveBeenCalled();
-  });
-
-  it("accepts a matching Origin using the Caddy-forwarded host", async () => {
-    const raw = installContext({
-      origin: "https://qgrid.example.com",
-      host: "127.0.0.1:44900",
-      "x-forwarded-host": "qgrid.example.com",
-    });
-    vi.stubEnv("pm_id", "0");
-
-    await expect(SettingModel.restartServer()).resolves.toEqual({ supervisor: "pm2" });
-    expect(process.exit).not.toHaveBeenCalled();
-
-    raw.emit("finish");
-    expect(process.exit).toHaveBeenCalledOnce();
-    expect(process.exit).toHaveBeenCalledWith(0);
-  });
-
-  it("accepts an Origin-less operational request", async () => {
-    const raw = installContext({ host: "qgrid.example.com" });
-    vi.stubEnv("pm_id", "0");
-
-    await expect(SettingModel.restartServer()).resolves.toEqual({ supervisor: "pm2" });
-    expect(process.exit).not.toHaveBeenCalled();
-
-    raw.emit("finish");
-    expect(process.exit).toHaveBeenCalledWith(0);
-  });
-
-  it("coordinates simultaneous calls through one process exit", async () => {
-    const firstRaw = installContext({ host: "qgrid.example.com" });
-    vi.stubEnv("pm_id", "0");
-
-    const first = SettingModel.restartServer();
-    const second = SettingModel.restartServer();
-    await expect(Promise.all([first, second])).resolves.toEqual([
-      { supervisor: "pm2" },
-      { supervisor: "pm2" },
-    ]);
-
-    firstRaw.emit("finish");
-    firstRaw.emit("close");
-    expect(process.exit).toHaveBeenCalledOnce();
-    expect(process.exit).toHaveBeenCalledWith(0);
-  });
-});

package/bundle/src/application/setting/setting.model.ts

@@ -1,19 +1,14 @@
-import { getLogger } from "@logtape/logtape";
 import {
   BaseModelClass,
   type ListResult,
   asArray,
   NotFoundException,
   BadRequestException,
-  SoException,
-  Sonamu,
   api,
   exhaustive,
 } from "sonamu";
 
 import { SD, type LocalizedString } from "../../i18n/sd.generated";
-import { detectSupervisor } from "../../utils/process-supervisor";
-import { armServerRestartExit, beginServerRestart } from "../../utils/server-restart";
 import { SlackNotificationError } from "../../utils/slack-notify";
 import { type SettingSubsetKey, type SettingSubsetMapping } from "../sonamu.generated";
 import { settingSubsetQueries, settingLoaderQueries } from "../sonamu.generated.sso";
@@ -37,37 +32,8 @@ import {
   type SettingListParams,
   type SettingSaveParams,
   type SettingsResponse,
-  type SupervisorKind,
 } from "./setting.types";
 
-const logger = getLogger(["qgrid", "setting"]);
-
-class ForbiddenException extends SoException {
-  constructor() {
-    super(403, SD("error.forbidden"));
-  }
-}
-
-function firstHeaderValue(value: string | string[] | undefined): string | undefined {
-  return (Array.isArray(value) ? value[0] : value)?.split(",")[0]?.trim();
-}
-
-/** Origin 없는 운영 CLI 요청은 통과시키되, 브라우저 교차 출처 POST 는 막는다. */
-function assertRestartOrigin(): void {
-  const { headers } = Sonamu.getContext();
-  const origin = firstHeaderValue(headers.origin);
-  if (headers.origin === undefined) return;
-  if (!origin) throw new ForbiddenException();
-
-  const host = firstHeaderValue(headers["x-forwarded-host"]) ?? firstHeaderValue(headers.host);
-  try {
-    if (!host || new URL(origin).host !== host) throw new ForbiddenException();
-  } catch (error) {
-    if (error instanceof ForbiddenException) throw error;
-    throw new ForbiddenException();
-  }
-}
-
 /*
   Setting Model
 */
@@ -246,12 +212,11 @@ class SettingModelClass extends BaseModelClass<
     return {
       settings,
       runtime: this.runtimeInfo(),
-      supervisor: detectSupervisor(),
       keepaliveRunnerEnabled: process.env[TOKEN_WINDOW_KEEPALIVE_RUNNER_ENV_KEY] === "true",
     };
   }
 
-  private runtimeInfo(): RuntimeInfoItem[] {
+  runtimeInfo(): RuntimeInfoItem[] {
     const env = process.env;
     const password = env.QGRID_DB_PASSWORD ?? env.SONAMU_DB_PASSWORD ?? "";
     return [
@@ -309,44 +274,6 @@ class SettingModelClass extends BaseModelClass<
       throw error;
     }
   }
-
-  /**
-   * 서버를 재시작한다 — 정확히는 스스로 종료하고, pm2 가 다시 띄우는 것에 맡긴다.
-   *
-   * `restart` 로 표시된 워커 설정은 dispatcher 생성자에서 한 번만 읽히므로 이 경로 말고는
-   * 반영할 방법이 없었다. 지금까지는 SSH 로 `pm2 restart` 를 쳐야 했다.
-   *
-   * 부수 효과가 하나 더 있다: pm2 는 `ecosystem.config.cjs` 의 원래 명령으로 다시 띄우고,
-   * 그 명령에 `--skip-update` 가 없으면 CLI 가 npm 최신 버전을 받아 올린다. 즉 재시작은
-   * 배포이기도 하다 — 화면 문구가 이 사실을 함께 알린다.
-   */
-  @api({ httpMethod: "POST", clients: ["axios", "tanstack-mutation"] })
-  async restartServer(): Promise<{ supervisor: SupervisorKind }> {
-    assertRestartOrigin();
-    const supervisor = detectSupervisor();
-    if (!supervisor) {
-      // 다시 띄워줄 주체가 없으면 종료가 곧 정지다. 살릴 수 없는데 죽이지 않는다.
-      throw new BadRequestException(SD("setting.noSupervisor")());
-    }
-
-    const { reply } = Sonamu.getContext();
-    const started = beginServerRestart();
-    if (started) {
-      logger.warn(`restart requested via dashboard, exiting for ${supervisor} to respawn`);
-
-      try {
-        // 정적 import로 qgrid.frame/token 그래프를 당겨오면 모듈 목 테스트가 오염된다.
-        // 재시작이 실제로 확정된 시점에 active native run 정합성 처리만 로드한다.
-        const { finishActiveNativeRunsForRestart } = await import("../qgrid/qgrid-run-lifecycle");
-        await finishActiveNativeRunsForRestart();
-      } finally {
-        // 정합성 처리가 실패해도 응답이 flush 된 뒤에는 재시작을 끝까지 진행한다.
-        armServerRestartExit(reply.raw);
-      }
-    }
-
-    return { supervisor };
-  }
 }
 
 export const SettingModel = new SettingModelClass();

package/bundle/src/application/setting/setting.store.ts

@@ -1,9 +1,8 @@
 /**
  * 런타임 설정 조회 캐시.
  *
- * 값의 출처는 DB(`settings`) → env 순이다. env 를 남겨두는 이유는 전환 때문이다 — dev0 는
- * `ecosystem.config.json` 으로 값을 주입하는데, DB 만 보게 만들면 배포 직후 설정이 전부
- * 기본값으로 떨어진다.
+ * 값의 출처는 DB(`settings`) → env 순이다. 저장된 설정이 없으면 기동 시 전달된
+ * 환경변수를 사용하므로, 설정을 DB에 저장하기 전에도 기존 값이 유지된다.
  *
  * 설정은 요청 경로(permit 해석, 알림 전송)에서 읽히므로 매번 DB 를 때리지 않고 메모리에
  * 둔다. `immediate` 변경도 현재 API 프로세스에만 즉시 반영된다. 인스턴스가 여러 개면 다른

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

@@ -21,16 +21,6 @@ export type SettingApplies = z.infer<typeof SettingApplies>;
 export const SettingSource = z.enum(["db", "env", "default"]);
 export type SettingSource = z.infer<typeof SettingSource>;
 
-/**
- * 프로세스를 다시 띄워주는 도구. 재시작은 스스로 종료하고 이 도구에 맡기므로,
- * 없으면(null) 화면에서 재시작을 시도조차 하지 않는다.
- *
- * API 반환 타입에 쓰이므로 zod 로 둔다 — sonamu syncer 가 클라이언트 코드를 만들 때
- * 타입만 있는 선언은 모듈을 찾지 못한다.
- */
-export const SupervisorKind = z.enum(["pm2"]);
-export type SupervisorKind = z.infer<typeof SupervisorKind>;
-
 export const SettingItem = z.object({
   key: z.string(),
   group: z.string(),
@@ -59,8 +49,6 @@ export type RuntimeInfoItem = z.infer<typeof RuntimeInfoItem>;
 export const SettingsResponse = z.object({
   settings: SettingItem.array(),
   runtime: RuntimeInfoItem.array(),
-  /** null 이면 재시작 버튼을 쓸 수 없다 — 종료해도 되살릴 주체가 없다. */
-  supervisor: SupervisorKind.nullable(),
   /** 이 인스턴스가 keepalive 스케줄러를 소유하도록 명시적으로 허용됐는지. */
   keepaliveRunnerEnabled: z.boolean(),
 });

package/bundle/src/application/sonamu.generated.http

@@ -98,13 +98,6 @@ Content-Type: application/json
 
 ###
 
-POST {{baseUrl}}/api/setting/restartServer
-Content-Type: application/json
-
-{}
-
-###
-
 GET {{baseUrl}}/api/requestLogStep/findById
 	?subset=A
 	&id=0
@@ -297,6 +290,7 @@ Content-Type: application/json
         "inputSchema": "unknown"
       }
     ],
+    "thinking": true,
     "effort": "EFFORT",
     "verbosity": "low",
     "reasoningSummary": "auto",
@@ -361,6 +355,7 @@ Content-Type: application/json
         "inputSchema": "unknown"
       }
     ],
+    "thinking": true,
     "effort": "EFFORT",
     "verbosity": "low",
     "reasoningSummary": "auto",

package/bundle/src/i18n/en.ts

@@ -49,12 +49,9 @@ export default {
   "error.serviceUnavailable": "Service unavailable",
   "setting.invalidValue": (reason: string) => `Invalid setting value: ${reason}`,
   "setting.unknownKey": (key: string) => `Unknown setting key ${key}`,
-  "setting.noSupervisor": () =>
-    "Restart is unavailable when running locally — nothing would bring the server back up",
   "qgrid.dispatcherFailed": (provider: string) => `${provider} dispatcher failed to initialize`,
   "qgrid.dispatcherStarting": (provider: string) =>
     `${provider} dispatcher is starting up. Please retry shortly`,
-  "qgrid.restartPending": () => "Server restart is in progress. Please retry shortly",
   "error.unauthorized": "Unauthorized",
   notFound: (name: string, id: number) => `${name} ID ${id} not found`,
   "validation.email": "Invalid email format",

package/bundle/src/i18n/ko.ts

@@ -48,12 +48,9 @@ export default {
   "error.serviceUnavailable": "서비스를 사용할 수 없습니다",
   "setting.invalidValue": (reason: string) => `설정 값이 올바르지 않습니다: ${reason}`,
   "setting.unknownKey": (key: string) => `알 수 없는 설정 키 ${key}`,
-  "setting.noSupervisor": () =>
-    "로컬 실행 환경에서는 재시작할 수 없습니다 — 종료하면 다시 켜줄 모니터링 도구가 없습니다",
   "qgrid.dispatcherFailed": (provider: string) => `${provider} dispatcher 초기화에 실패했습니다`,
   "qgrid.dispatcherStarting": (provider: string) =>
     `${provider} dispatcher 기동 중입니다. 잠시 후 다시 시도해주세요`,
-  "qgrid.restartPending": () => "서버 재시작 중입니다. 잠시 후 다시 시도해주세요",
   "error.unauthorized": "인증이 필요합니다",
   "error.unknownSearchField": (field: string) => `구현되지 않은 검색 필드 ${field}`,
   notFound: (name: string, id: number) => `존재하지 않는 ${name} ID ${id}`,

package/bundle/src/sonamu.config.ts

@@ -195,14 +195,14 @@ export default defineConfig({
           log.warn(`anthropic dispatcher failed: ${(e as Error).message}`);
         }
 
-        subscriber.setTokenChangeHandler(() => {
+        subscriber.tokenChangeHandler = () => {
           void rescheduleTokenWindowKeepalive();
-        });
+        };
         startExpiredTokenReminder();
         void startTokenWindowKeepalive();
 
-        const anthropicCount = QgridDispatcher.anthropicDispatcher?.tokenCount ?? 0;
-        const openaiCount = QgridDispatcher.openaiDispatcher?.tokenCount ?? 0;
+        const anthropicCount = QgridDispatcher.anthropicDispatcher?.tokenPool.size ?? 0;
+        const openaiCount = QgridDispatcher.openaiDispatcher?.countActiveTokens() ?? 0;
 
         log.info(`listening on http://${host}:${port}`);
         log.info(`anthropic: ${anthropicCount} tokens ready`);
@@ -217,7 +217,7 @@ export default defineConfig({
         stopTokenWindowKeepalive();
         stopOpenAICallbackRelay();
         setSettingChangeHandler(null);
-        QgridDispatcher.subscriber?.setTokenChangeHandler(null);
+        if (QgridDispatcher.subscriber) QgridDispatcher.subscriber.tokenChangeHandler = null;
         if (QgridDispatcher.openaiDispatcher) {
           await QgridDispatcher.openaiDispatcher.stop();
         }