@cartanova/qgrid-cli 2.10.2 → 2.10.4

package/bundle/web-dist/server/assets/{settings-BFVuldch.js → settings-DeHNwoMv.js}

@@ -1,7 +1,7 @@
-import { A as require_jsx_runtime, D as useQuery, T as isAxiosError, Z as __toESM, a as SettingService, j as require_react, k as useQueryClient, l as isSonamuError, n as QgridService, o as TokenService } from "./services.generated-BM4X3Ecj.js";
+import { A as require_react, O as useQueryClient, X as __toESM, a as SettingService, k as require_jsx_runtime, o as TokenService } from "./services.generated-FPwsG9p3.js";
 import { l as lucideRotateCcw } from "../entry-server.generated.js";
-import { t as useUpdateTokenMutation } from "./use-update-token-mutation-DqGMopII.js";
-import { t as lucidePlus } from "./plus-CmvbWuDf.js";
+import { t as useUpdateTokenMutation } from "./use-update-token-mutation-xvXbPrCI.js";
+import { t as lucidePlus } from "./plus-Bvdj8Pwx.js";
 //#region ~icons/lucide/eye.jsx
 var import_react = /* @__PURE__ */ __toESM(require_react());
 var import_jsx_runtime = require_jsx_runtime();
@@ -62,7 +62,6 @@ var GROUP_LABELS = {
 	slack: "Slack 알림",
 	slackConnection: "Slack 연결"
 };
-var RESTART_READY_TIMEOUT_MS = 180 * 1e3;
 function maskValue(value) {
 	if (!value) return "";
 	if (value.length <= 12) return "•".repeat(value.length);
@@ -357,212 +356,6 @@ function TriggerReminderButton() {
 		})]
 	});
 }
-/**
-* 재시작은 스스로 종료하고 감독자(pm2)가 다시 띄우는 것에 맡긴다. 되돌릴 수 없고 1~2분간
-* 요청이 실패하므로 확인 단계를 둔다 — 무엇이 일어나는지 모달이 먼저 설명한다.
-*/
-function RestartPanel({ supervisor }) {
-	const [confirming, setConfirming] = (0, import_react.useState)(false);
-	const [restartRequested, setRestartRequested] = (0, import_react.useState)(false);
-	const [restartStartedAt, setRestartStartedAt] = (0, import_react.useState)(null);
-	const [sawUnavailable, setSawUnavailable] = (0, import_react.useState)(false);
-	const [completing, setCompleting] = (0, import_react.useState)(false);
-	const [error, setError] = (0, import_react.useState)(null);
-	const completingRef = (0, import_react.useRef)(false);
-	const restartAttemptRef = (0, import_react.useRef)(0);
-	const queryClient = useQueryClient();
-	const mutation = SettingService.useRestartServerMutation();
-	const resetMutation = mutation.reset;
-	const healthQuery = useQuery({
-		...QgridService.healthQueryOptions(),
-		enabled: restartRequested && !completing,
-		refetchInterval: restartRequested && !completing ? 1e3 : false,
-		refetchIntervalInBackground: true,
-		retry: false
-	});
-	(0, import_react.useEffect)(() => {
-		if (!restartRequested || restartStartedAt === null || completing) return;
-		const remaining = Math.max(0, RESTART_READY_TIMEOUT_MS - (Date.now() - restartStartedAt));
-		const timeout = window.setTimeout(() => {
-			restartAttemptRef.current += 1;
-			resetMutation();
-			setConfirming(false);
-			setRestartRequested(false);
-			setRestartStartedAt(null);
-			setSawUnavailable(false);
-			setError("재시작 후 3분 안에 서버 준비를 확인하지 못했습니다. 상태를 확인한 뒤 다시 시도하세요.");
-		}, remaining);
-		return () => window.clearTimeout(timeout);
-	}, [
-		completing,
-		resetMutation,
-		restartRequested,
-		restartStartedAt
-	]);
-	(0, import_react.useEffect)(() => {
-		if (!restartRequested || restartStartedAt === null) return;
-		const currentAttemptError = healthQuery.isError && healthQuery.errorUpdatedAt >= restartStartedAt;
-		const currentAttemptData = healthQuery.isSuccess && healthQuery.dataUpdatedAt >= restartStartedAt;
-		if (currentAttemptError || currentAttemptData && !healthQuery.data.ready) {
-			setSawUnavailable(true);
-			return;
-		}
-		if (!sawUnavailable || !currentAttemptData || !healthQuery.data.ready || completingRef.current) return;
-		completingRef.current = true;
-		setCompleting(true);
-		queryClient.invalidateQueries({ queryKey: SettingService.getSettingListQueryOptions().queryKey }).then(() => {
-			restartAttemptRef.current += 1;
-			resetMutation();
-			setConfirming(false);
-			setRestartRequested(false);
-			setRestartStartedAt(null);
-			setSawUnavailable(false);
-		}).catch((cause) => {
-			restartAttemptRef.current += 1;
-			resetMutation();
-			setConfirming(false);
-			setRestartRequested(false);
-			setRestartStartedAt(null);
-			setSawUnavailable(false);
-			setError(`서버는 준비됐지만 설정을 새로고침하지 못했습니다: ${cause.message}`);
-		}).finally(() => {
-			completingRef.current = false;
-			setCompleting(false);
-		});
-	}, [
-		healthQuery.dataUpdatedAt,
-		healthQuery.data?.ready,
-		healthQuery.errorUpdatedAt,
-		healthQuery.isError,
-		healthQuery.isSuccess,
-		queryClient,
-		resetMutation,
-		restartStartedAt,
-		restartRequested,
-		sawUnavailable
-	]);
-	const restart = async () => {
-		const attempt = restartAttemptRef.current + 1;
-		restartAttemptRef.current = attempt;
-		setError(null);
-		setSawUnavailable(false);
-		setCompleting(false);
-		completingRef.current = false;
-		setRestartRequested(true);
-		setRestartStartedAt(Date.now());
-		try {
-			await mutation.mutateAsync();
-		} catch (e) {
-			if (restartAttemptRef.current !== attempt) return;
-			if (isSonamuError(e) ? e.code >= 400 && e.code < 500 : isAxiosError(e) && e.response !== void 0 && e.response.status >= 400 && e.response.status < 500) {
-				restartAttemptRef.current += 1;
-				resetMutation();
-				setConfirming(false);
-				setRestartRequested(false);
-				setRestartStartedAt(null);
-				setError(e.message || "재시작 요청이 거부되었습니다.");
-			}
-		}
-	};
-	const closeConfirmation = () => {
-		if (restartRequested) return;
-		setConfirming(false);
-		setError(null);
-	};
-	return /* @__PURE__ */ (0, import_jsx_runtime.jsxs)("div", {
-		className: "panel overflow-hidden",
-		children: [
-			/* @__PURE__ */ (0, import_jsx_runtime.jsx)("div", {
-				className: "panel-header px-4 py-2.5 sm:px-5",
-				children: /* @__PURE__ */ (0, import_jsx_runtime.jsx)("span", {
-					className: "text-[13px] font-medium text-sand-700",
-					children: "서버 제어"
-				})
-			}),
-			/* @__PURE__ */ (0, import_jsx_runtime.jsxs)("div", {
-				className: "px-4 py-3 sm:px-5 flex items-start justify-between gap-4",
-				children: [/* @__PURE__ */ (0, import_jsx_runtime.jsxs)("div", {
-					className: "min-w-0",
-					children: [/* @__PURE__ */ (0, import_jsx_runtime.jsx)("p", {
-						className: "text-[13px] text-sand-800",
-						children: "qgrid 재시작"
-					}), /* @__PURE__ */ (0, import_jsx_runtime.jsx)("p", {
-						className: "mt-0.5 text-[11px] text-sand-500",
-						children: supervisor ? "재시작 필요 설정을 반영하고 qgrid·Codex·Claude CLI를 확인합니다" : "로컬 실행 환경에서는 재시작할 수 없습니다"
-					})]
-				}), /* @__PURE__ */ (0, import_jsx_runtime.jsx)("button", {
-					type: "button",
-					disabled: !supervisor || restartRequested || mutation.isPending,
-					onClick: () => setConfirming(true),
-					className: "shrink-0 px-2.5 py-1.5 text-[11px] font-medium rounded-lg border border-red-200 text-red-600 hover:bg-red-50 disabled:opacity-40 disabled:hover:bg-transparent transition-colors duration-150",
-					children: "재시작"
-				})]
-			}),
-			error && /* @__PURE__ */ (0, import_jsx_runtime.jsx)("p", {
-				className: "px-4 pb-3 sm:px-5 text-[11px] text-red-500",
-				children: error
-			}),
-			confirming && /* @__PURE__ */ (0, import_jsx_runtime.jsxs)("div", {
-				className: "fixed inset-0 z-50 flex items-start justify-center pt-[15vh]",
-				children: [/* @__PURE__ */ (0, import_jsx_runtime.jsx)("div", {
-					className: "absolute inset-0 bg-sand-900/8 backdrop-blur-sm",
-					onClick: closeConfirmation,
-					onKeyDown: () => {}
-				}), /* @__PURE__ */ (0, import_jsx_runtime.jsxs)("div", {
-					className: "relative panel shadow-xl w-full max-w-sm mx-4",
-					children: [
-						/* @__PURE__ */ (0, import_jsx_runtime.jsx)("div", {
-							className: "px-5 py-4 border-b border-sand-100/60",
-							children: /* @__PURE__ */ (0, import_jsx_runtime.jsx)("h3", {
-								className: "text-base font-medium text-sand-900",
-								children: restartRequested ? mutation.isPending ? "재시작 요청 중" : "서버 재연결 중" : "qgrid 재시작"
-							})
-						}),
-						/* @__PURE__ */ (0, import_jsx_runtime.jsxs)("div", {
-							className: "px-5 py-4 space-y-2",
-							children: [
-								/* @__PURE__ */ (0, import_jsx_runtime.jsxs)("p", {
-									className: "text-[13px] text-sand-700",
-									children: [
-										"워커가 다시 뜰 때까지 ",
-										/* @__PURE__ */ (0, import_jsx_runtime.jsx)("strong", {
-											className: "text-sand-900",
-											children: "1~2분간 요청이 실패"
-										}),
-										"합니다."
-									]
-								}),
-								/* @__PURE__ */ (0, import_jsx_runtime.jsx)("p", {
-									className: "text-[12px] text-sand-500",
-									children: "서버를 종료하면 모니터링 도구가 다시 띄웁니다. 이때 재시작 필요 설정을 반영하고, qgrid와 Codex CLI·Claude Code의 최신 버전을 확인합니다."
-								}),
-								restartRequested && /* @__PURE__ */ (0, import_jsx_runtime.jsx)("p", {
-									className: "text-[12px] font-medium text-sienna-600",
-									children: mutation.isPending ? "종료 요청을 처리하고 있습니다…" : sawUnavailable ? "새 서버가 준비될 때까지 연결을 확인하고 있습니다…" : "서버 종료를 기다리고 있습니다…"
-								})
-							]
-						}),
-						/* @__PURE__ */ (0, import_jsx_runtime.jsxs)("div", {
-							className: "px-5 py-3 border-t border-sand-100/60 flex justify-end gap-2",
-							children: [!restartRequested && /* @__PURE__ */ (0, import_jsx_runtime.jsx)("button", {
-								type: "button",
-								onClick: closeConfirmation,
-								className: "px-3 py-1.5 text-[12px] rounded-lg text-sand-600 hover:bg-sand-100 transition-colors duration-150",
-								children: "취소"
-							}), /* @__PURE__ */ (0, import_jsx_runtime.jsx)("button", {
-								type: "button",
-								disabled: restartRequested || mutation.isPending,
-								onClick: () => void restart(),
-								className: "px-3 py-1.5 text-[12px] font-medium rounded-lg bg-red-500 text-white hover:bg-red-600 disabled:opacity-50 transition-colors duration-150",
-								children: restartRequested ? mutation.isPending ? "요청 중" : "재연결 중" : "재시작"
-							})]
-						})
-					]
-				})]
-			})
-		]
-	});
-}
 function SettingsPanel() {
 	const settingsQuery = SettingService.useSettingList();
 	const { data, isLoading } = settingsQuery;
@@ -611,53 +404,49 @@ function SettingsPanel() {
 	const groups = [...new Set(settings.map((s) => s.group))];
 	return /* @__PURE__ */ (0, import_jsx_runtime.jsxs)("div", {
 		className: "space-y-4",
-		children: [
-			/* @__PURE__ */ (0, import_jsx_runtime.jsxs)("div", {
+		children: [/* @__PURE__ */ (0, import_jsx_runtime.jsxs)("div", {
+			className: "panel overflow-hidden",
+			children: [/* @__PURE__ */ (0, import_jsx_runtime.jsxs)("div", {
+				className: "panel-header flex items-center justify-between px-4 py-2.5 sm:px-5",
+				children: [/* @__PURE__ */ (0, import_jsx_runtime.jsx)("span", {
+					className: "text-[13px] font-medium text-sand-700",
+					children: "환경"
+				}), /* @__PURE__ */ (0, import_jsx_runtime.jsx)("span", {
+					className: "text-[11px] text-sand-400",
+					children: "기동 시점 값"
+				})]
+			}), /* @__PURE__ */ (0, import_jsx_runtime.jsx)("dl", {
+				className: "px-4 py-3 sm:px-5 grid grid-cols-1 sm:grid-cols-2 gap-x-8 gap-y-1.5",
+				children: runtime.map((info) => /* @__PURE__ */ (0, import_jsx_runtime.jsxs)("div", {
+					className: "flex items-baseline justify-between gap-3 min-w-0",
+					children: [/* @__PURE__ */ (0, import_jsx_runtime.jsx)("dt", {
+						className: "text-[11px] text-sand-500 shrink-0",
+						children: info.label
+					}), /* @__PURE__ */ (0, import_jsx_runtime.jsx)("dd", {
+						className: "text-[12px] font-mono text-sand-700 truncate",
+						children: info.value
+					})]
+				}, info.label))
+			})]
+		}), groups.map((group) => {
+			const items = settings.filter((s) => s.group === group);
+			return /* @__PURE__ */ (0, import_jsx_runtime.jsxs)("div", {
 				className: "panel overflow-hidden",
 				children: [/* @__PURE__ */ (0, import_jsx_runtime.jsxs)("div", {
 					className: "panel-header flex items-center justify-between px-4 py-2.5 sm:px-5",
 					children: [/* @__PURE__ */ (0, import_jsx_runtime.jsx)("span", {
 						className: "text-[13px] font-medium text-sand-700",
-						children: "환경"
-					}), /* @__PURE__ */ (0, import_jsx_runtime.jsx)("span", {
-						className: "text-[11px] text-sand-400",
-						children: "기동 시점 값"
-					})]
-				}), /* @__PURE__ */ (0, import_jsx_runtime.jsx)("dl", {
-					className: "px-4 py-3 sm:px-5 grid grid-cols-1 sm:grid-cols-2 gap-x-8 gap-y-1.5",
-					children: runtime.map((info) => /* @__PURE__ */ (0, import_jsx_runtime.jsxs)("div", {
-						className: "flex items-baseline justify-between gap-3 min-w-0",
-						children: [/* @__PURE__ */ (0, import_jsx_runtime.jsx)("dt", {
-							className: "text-[11px] text-sand-500 shrink-0",
-							children: info.label
-						}), /* @__PURE__ */ (0, import_jsx_runtime.jsx)("dd", {
-							className: "text-[12px] font-mono text-sand-700 truncate",
-							children: info.value
-						})]
-					}, info.label))
+						children: GROUP_LABELS[group] ?? group
+					}), group === "slack" && /* @__PURE__ */ (0, import_jsx_runtime.jsx)(TriggerReminderButton, {})]
+				}), /* @__PURE__ */ (0, import_jsx_runtime.jsx)("div", {
+					className: "divide-y divide-sand-100/80",
+					children: items.map((item) => /* @__PURE__ */ (0, import_jsx_runtime.jsxs)("div", { children: [/* @__PURE__ */ (0, import_jsx_runtime.jsx)(SettingRow, {
+						item,
+						disabled: item.key === "qgrid.tokenWindowKeepaliveEnabled" && !data?.keepaliveRunnerEnabled
+					}), item.key === "qgrid.tokenWindowKeepaliveEnabled" && /* @__PURE__ */ (0, import_jsx_runtime.jsx)(KeepaliveTokenSettings, { runnerEnabled: data?.keepaliveRunnerEnabled ?? false })] }, item.key))
 				})]
-			}),
-			groups.map((group) => {
-				const items = settings.filter((s) => s.group === group);
-				return /* @__PURE__ */ (0, import_jsx_runtime.jsxs)("div", {
-					className: "panel overflow-hidden",
-					children: [/* @__PURE__ */ (0, import_jsx_runtime.jsxs)("div", {
-						className: "panel-header flex items-center justify-between px-4 py-2.5 sm:px-5",
-						children: [/* @__PURE__ */ (0, import_jsx_runtime.jsx)("span", {
-							className: "text-[13px] font-medium text-sand-700",
-							children: GROUP_LABELS[group] ?? group
-						}), group === "slack" && /* @__PURE__ */ (0, import_jsx_runtime.jsx)(TriggerReminderButton, {})]
-					}), /* @__PURE__ */ (0, import_jsx_runtime.jsx)("div", {
-						className: "divide-y divide-sand-100/80",
-						children: items.map((item) => /* @__PURE__ */ (0, import_jsx_runtime.jsxs)("div", { children: [/* @__PURE__ */ (0, import_jsx_runtime.jsx)(SettingRow, {
-							item,
-							disabled: item.key === "qgrid.tokenWindowKeepaliveEnabled" && !data?.keepaliveRunnerEnabled
-						}), item.key === "qgrid.tokenWindowKeepaliveEnabled" && /* @__PURE__ */ (0, import_jsx_runtime.jsx)(KeepaliveTokenSettings, { runnerEnabled: data?.keepaliveRunnerEnabled ?? false })] }, item.key))
-					})]
-				}, group);
-			}),
-			/* @__PURE__ */ (0, import_jsx_runtime.jsx)(RestartPanel, { supervisor: data?.supervisor ?? null })
-		]
+			}, group);
+		})]
 	});
 }
 //#endregion

package/bundle/web-dist/server/assets/{show-Cswh_tqs.js → show-C4jk1Hrp.js}

@@ -1,5 +1,5 @@
-import { A as require_jsx_runtime, D as useQuery, J as __commonJSMin, Z as __toESM, i as RequestLogStepService, j as require_react, l as isSonamuError, r as RequestLogService } from "./services.generated-BM4X3Ecj.js";
-import { a as Link } from "./lazyRouteComponent-Bj1cPq6x.js";
+import { A as require_react, E as useQuery, X as __toESM, i as RequestLogStepService, k as require_jsx_runtime, l as isSonamuError, q as __commonJSMin, r as RequestLogService } from "./services.generated-FPwsG9p3.js";
+import { a as Link } from "./lazyRouteComponent-CUR-kWzI.js";
 import { c as lucideX, d as Markdown, t as Route, u as remarkGfm } from "../entry-server.generated.js";
 import { n as formatMicroUsd, t as cacheHitRate } from "./cost-CSH_t17I.js";
 import { t as requestModelDisplay } from "./request-model-CnCeaNLE.js";

package/bundle/web-dist/server/assets/{tokens-DiTUwUwr.js → tokens-odgS2-9q.js}

@@ -1,8 +1,8 @@
-import { A as require_jsx_runtime, Z as __toESM, j as require_react, k as useQueryClient, n as QgridService, o as TokenService } from "./services.generated-BM4X3Ecj.js";
+import { A as require_react, O as useQueryClient, X as __toESM, k as require_jsx_runtime, n as QgridService, o as TokenService } from "./services.generated-FPwsG9p3.js";
 import { f as lucideKeyRound, o as Q } from "../entry-server.generated.js";
-import { t as useUpdateTokenMutation } from "./use-update-token-mutation-DqGMopII.js";
-import { n as OAuthCodeEntry, t as useOAuthLoginFlow } from "./use-oauth-login-flow-DYKnG4A2.js";
-import { t as lucidePlus } from "./plus-CmvbWuDf.js";
+import { t as useUpdateTokenMutation } from "./use-update-token-mutation-xvXbPrCI.js";
+import { n as OAuthCodeEntry, t as useOAuthLoginFlow } from "./use-oauth-login-flow-DTQKdYBO.js";
+import { t as lucidePlus } from "./plus-Bvdj8Pwx.js";
 //#region src/components/qgrid/AddTokenModal.tsx
 var import_react = /* @__PURE__ */ __toESM(require_react(), 1);
 var import_jsx_runtime = require_jsx_runtime();

package/bundle/web-dist/server/assets/{use-oauth-login-flow-DYKnG4A2.js → use-oauth-login-flow-DTQKdYBO.js}

@@ -1,4 +1,4 @@
-import { A as require_jsx_runtime, Z as __toESM, j as require_react, k as useQueryClient, n as QgridService, o as TokenService } from "./services.generated-BM4X3Ecj.js";
+import { A as require_react, O as useQueryClient, X as __toESM, k as require_jsx_runtime, n as QgridService, o as TokenService } from "./services.generated-FPwsG9p3.js";
 import { o as Q } from "../entry-server.generated.js";
 //#region src/components/qgrid/OAuthCodeEntry.tsx
 var import_react = /* @__PURE__ */ __toESM(require_react(), 1);

package/bundle/web-dist/server/assets/{use-update-token-mutation-DqGMopII.js → use-update-token-mutation-xvXbPrCI.js}

@@ -1,4 +1,4 @@
-import { E as useMutation, n as QgridService } from "./services.generated-BM4X3Ecj.js";
+import { T as useMutation, n as QgridService } from "./services.generated-FPwsG9p3.js";
 //#region src/services/token/use-update-token-mutation.ts
 function useUpdateTokenMutation() {
 	return useMutation({ mutationFn: ({ id, name, quotaThreshold, weight, keepaliveEnabled }) => QgridService.updateToken(id, name, quotaThreshold, weight, keepaliveEnabled) });

package/bundle/web-dist/server/entry-server.generated.js

@@ -1,6 +1,6 @@
-import { A as require_jsx_runtime, B as matchMutation, C as union, F as onlineManager, G as skipToken, H as noop$1, I as notifyManager, J as __commonJSMin, K as focusManager, L as functionalUpdate, M as Mutation, N as infiniteQueryBehavior, O as QueryClientProvider, P as Query, R as hashKey, S as string$2, U as partialMatchKey, V as matchQuery, W as resolveStaleTime, X as __require, Y as __exportAll, Z as __toESM, _ as custom, b as number$1, c as fetch$1, d as SD, f as SUPPORTED_LOCALES, g as boolean$1, h as array, j as require_react, m as _enum, n as QgridService, o as TokenService, p as setLocale, q as Subscribable, s as SonamuQueryMode, u as zArrayable, v as date, w as unknown, x as object, y as int, z as hashQueryKeyByOptions } from "./assets/services.generated-BM4X3Ecj.js";
-import { A as decodePath, C as findRouteMatch, D as createLRUCache, E as processRouteTree, F as last, I as replaceEqualDeep, M as findLast, N as functionalUpdate$1, O as invariant, P as isPromise, S as findFlatMatch, T as processRouteMasks, _ as interpolatePath, a as Link$1, b as trimPath, c as usePrevious$1, d as useRouter, f as getRouterContext, g as cleanPath, h as isNotFound, i as createRootRouteWithContext, j as deepEqual$2, k as createControlledPromise, l as matchContext, m as rootRouteId, n as createFileRoute, p as warning, r as createLazyFileRoute, s as useLayoutEffect$3, t as lazyRouteComponent, u as useRouterState, v as joinPaths, w as findSingleMatch, x as trimPathRight, y as resolvePath } from "./assets/lazyRouteComponent-Bj1cPq6x.js";
-import { n as require_shim, t as require_react_dom } from "./assets/react-dom-BjR1Iaz8.js";
+import { A as require_react, B as matchQuery, C as union, D as QueryClientProvider, F as notifyManager, G as focusManager, H as partialMatchKey, I as functionalUpdate, J as __exportAll, K as Subscribable, L as hashKey, M as infiniteQueryBehavior, N as Query, P as onlineManager, R as hashQueryKeyByOptions, S as string$2, U as resolveStaleTime, V as noop$1, W as skipToken, X as __toESM, Y as __require, _ as custom, b as number$1, c as fetch$1, d as SD, f as SUPPORTED_LOCALES, g as boolean$1, h as array, j as Mutation, k as require_jsx_runtime, m as _enum, n as QgridService, o as TokenService, p as setLocale, q as __commonJSMin, s as SonamuQueryMode, u as zArrayable, v as date, w as unknown, x as object, y as int, z as matchMutation } from "./assets/services.generated-FPwsG9p3.js";
+import { A as decodePath, C as findRouteMatch, D as createLRUCache, E as processRouteTree, F as last, I as replaceEqualDeep, M as findLast, N as functionalUpdate$1, O as invariant, P as isPromise, S as findFlatMatch, T as processRouteMasks, _ as interpolatePath, a as Link$1, b as trimPath, c as usePrevious$1, d as useRouter, f as getRouterContext, g as cleanPath, h as isNotFound, i as createRootRouteWithContext, j as deepEqual$2, k as createControlledPromise, l as matchContext, m as rootRouteId, n as createFileRoute, p as warning, r as createLazyFileRoute, s as useLayoutEffect$3, t as lazyRouteComponent, u as useRouterState, v as joinPaths, w as findSingleMatch, x as trimPathRight, y as resolvePath } from "./assets/lazyRouteComponent-CUR-kWzI.js";
+import { n as require_shim, t as require_react_dom } from "./assets/react-dom-CyuShs9K.js";
 import minpath from "node:path";
 import minproc from "node:process";
 import { fileURLToPath as urlToPath } from "node:url";
@@ -46874,7 +46874,7 @@ function ChatWidget() {
 }
 //#endregion
 //#region src/components/Sidebar.tsx
-var qgridVersion = "2.10.2";
+var qgridVersion = "2.10.4";
 function Sidebar({ className }) {
 	const pathname = useRouterState({ select: (s) => s.location.pathname });
 	return /* @__PURE__ */ (0, import_jsx_runtime.jsxs)("aside", {
@@ -68752,15 +68752,15 @@ function RootComponent() {
 }
 //#endregion
 //#region src/routes/tokens.tsx
-var $$splitComponentImporter$5 = () => import("./assets/tokens-DiTUwUwr.js");
+var $$splitComponentImporter$5 = () => import("./assets/tokens-odgS2-9q.js");
 var Route$5 = createFileRoute("/tokens")({ component: lazyRouteComponent($$splitComponentImporter$5, "component") });
 //#endregion
 //#region src/routes/settings.tsx
-var $$splitComponentImporter$4 = () => import("./assets/settings-BFVuldch.js");
+var $$splitComponentImporter$4 = () => import("./assets/settings-DeHNwoMv.js");
 var Route$4 = createFileRoute("/settings")({ component: lazyRouteComponent($$splitComponentImporter$4, "component") });
 //#endregion
 //#region src/routes/monit.tsx
-var $$splitComponentImporter$3 = () => import("./assets/monit-D9W4ioUi.js");
+var $$splitComponentImporter$3 = () => import("./assets/monit-DSxs4Jnl.js");
 var monitSearchSchema = object({ level: _enum([
 	"all",
 	"warn",
@@ -69137,7 +69137,7 @@ object({
 _enum(["A"]);
 //#endregion
 //#region src/routes/logs.tsx
-var $$splitComponentImporter$2 = () => import("./assets/logs-BSk9lOHe.js");
+var $$splitComponentImporter$2 = () => import("./assets/logs-CagNGTBj.js");
 var logsSearchSchema = object({
 	token: string$2().optional(),
 	project: string$2().optional(),
@@ -69151,11 +69151,11 @@ var Route$2 = createFileRoute("/logs")({
 });
 //#endregion
 //#region src/routes/index.tsx
-var $$splitComponentImporter$1 = () => import("./assets/routes-C7lB2Q85.js");
+var $$splitComponentImporter$1 = () => import("./assets/routes-i-sTqXMX.js");
 var Route$1 = createFileRoute("/")({ component: lazyRouteComponent($$splitComponentImporter$1, "component") });
 //#endregion
 //#region src/routes/requests/show.tsx
-var $$splitComponentImporter = () => import("./assets/show-Cswh_tqs.js");
+var $$splitComponentImporter = () => import("./assets/show-C4jk1Hrp.js");
 var showSearchSchema = object({ id: number$1() });
 var Route = createFileRoute("/requests/show")({
 	validateSearch: showSearchSchema,

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@cartanova/qgrid-cli",
-  "version": "2.10.2",
+  "version": "2.10.4",
   "repository": {
     "type": "git",
     "url": "https://github.com/cartanova-ai/qgrid"

package/skills/qgrid/SKILL.md

@@ -45,6 +45,7 @@ Request logging is enabled by default. Use `providerOptions.qgrid.logger: false`
 ## Non-Negotiable Boundaries
 
 - Use `packages/ai-sdk` as the active public SDK surface.
+- Expose thinking only as `providerOptions.qgrid.thinking?: boolean`: omission preserves the model policy, false overrides effort, and unsupported/unverified off requests fail explicitly. See the SDK and provider runtime references for the dated support snapshot. Keep Anthropic thinking controls separate from mandatory configuration and memory isolation. Haiku/Sonnet 5.5 off uses a computed thinking-only native body override with Claude Code >=2.1.295; never inherit or expose arbitrary extra body.
 - The old v1 SDK package (`packages/sdk`, `@cartanova/qgrid-sdk`) has been removed from the repository. It survives only as a deprecated npm artifact (1.9.0); do not resurrect it or write new code against it.
 - Let the server infer single-turn versus tool-run request-log lifecycle. Do not recreate caller-selected logging modes in the SDK or API examples.
 - Route provider models by prefix: `openai/*` goes to the OpenAI Codex runtime, `anthropic/*` goes to the Anthropic Claude Code runtime. Prefix-less model fallback is not implemented.

package/skills/qgrid/references/ai-sdk-provider-contract.md

@@ -63,7 +63,8 @@ Do not write `providerOptions: { ... } satisfies QgridProviderOptions`: the expo
   rejected by the SDK before transport; omission keeps weighted round-robin routing.
 - `logger`: request-log switch. Omitted or `true` is the default and enables qgrid logging; `false` guarantees that generation creates zero qgrid request-log rows.
 - `sessionKey`: source for model-scoped opaque OpenAI prompt-cache affinity. Disabled for Anthropic models.
-- `effort`: provider-specific vocabulary. OpenAI (ChatGPT-subscription Codex route): `low | medium | high | xhigh | max | ultra`, where GPT-6 Astra, GPT-6 Sol, GPT-5.6 Sol, and Terra support `ultra`, and GPT-6 Luna and GPT-5.6 Luna support through `max`; the public OpenAI API's `none`/`minimal` do not exist on this route. Anthropic (Claude Code `--effort`): `low | medium | high | xhigh | max`. The server resolves the value per provider and model (`providers/common/effort.ts`): anything outside the provider set, or above the Codex model's supported maximum, is silently dropped and treated as unspecified — Anthropic falls back to qgrid's default `low`, OpenAI omits `reasoning` so the backend default applies. No error, no metadata; a debug log line only. Use `QgridOpenAIProviderOptions` / `QgridAnthropicProviderOptions` for provider-typed `satisfies`; `QgridProviderOptions` is their union.
+- `thinking`: optional boolean for both providers. Omission preserves existing model behavior. True enables thinking with the normal request effort → config `defaultEffort` → default resolution; it does not guarantee thinking tokens on every response. False ignores both efforts, including in request logging, and rejects unsupported or unverified model/runtime combinations. Generate, stream, tool continuation, and structured output carry the same boolean. Update SDK and server together: older servers can ignore the field. No `defaultThinking` or public `effort: "none"` is exposed. Haiku/Sonnet 5.5 off requires Claude Code >=2.1.295; Sonnet uses `between_tools` and may retain thinking blocks between tools.
+- `effort`: provider-specific vocabulary. OpenAI (ChatGPT-subscription Codex route): `low | medium | high | xhigh | max | ultra`, where GPT-6 Astra, GPT-6 Sol, GPT-5.6 Sol, and Terra support `ultra`, and GPT-6 Luna and GPT-5.6 Luna support through `max`; `none`/`minimal` are not public effort values; use `thinking: false` for verified off-capable models. Anthropic (Claude Code `--effort`): `low | medium | high | xhigh | max`. The server resolves the value per provider and model (`providers/common/effort.ts`): anything outside the provider set, or above the Codex model's supported maximum, is silently dropped and treated as unspecified. With thinking omitted, Anthropic falls back to qgrid's default `low`; OpenAI omits `reasoning.effort` so the backend default applies. With `thinking: true`, OpenAI explicitly falls back to `low` for absent or unsupported effort, including raw API requests. With `thinking: false`, both providers ignore caller/default effort and apply their off policy instead. Invalid effort itself produces no error or metadata, only a debug log; unsupported thinking-off requests still fail explicitly. Use `QgridOpenAIProviderOptions` / `QgridAnthropicProviderOptions` for provider-typed `satisfies`; `QgridProviderOptions` is their union.
 - `verbosity`: OpenAI/Codex route only.
 - `reasoningSummary`: OpenAI/Codex route only.
 - `serviceTier`: OpenAI/Codex route only.

package/skills/qgrid/references/anthropic-claude-code-runtime.md

@@ -77,7 +77,7 @@ claude -p
   --model <canonical-model-or-1m-suffix>
   --system-prompt <text>                  # small system prompt
   --system-prompt-file <path>             # large system prompt
-  --thinking disabled                       # omitted for Fable 5/5.1, Opus 5, Opus 5.5, Sonnet 5.5 and Haiku 5.5
+  --thinking <disabled|adaptive>             # resolved from thinking and the existing model default
   --effort <effort-or-low>
   --disable-slash-commands
   --session-id <uuid>
@@ -97,7 +97,11 @@ Important details:
 - Large system prompts over 64 KiB are written to a temporary file to avoid argv `E2BIG` — this
   same branch absorbs large injected schema contracts, so the old 64 KiB schema argv limit no
   longer applies (only the global 512 KiB caller-schema complexity limit remains).
-- `--thinking disabled`, `MAX_THINKING_TOKENS=0`, and adaptive thinking env suppression keep thinking off for existing models. Fable 5, Fable 5.1, Opus 5.5, Sonnet 5.5 and Haiku 5.5 require always-on adaptive thinking (Fable: CLI catalog `rejects_disabled_thinking`; Opus 5.5: official spec, thinking cannot be turned off; Sonnet 5.5: API `{type: "disabled"}` returns 400 and the CLI catalog marks `rejects_disabled_thinking`; Haiku 5.5 (released 2026-10-07): the API accepts disabled thinking only at `high` effort or below and returns 400 at `xhigh`/`max`, and the Claude Code 2.1.293 catalog marks `rejects_disabled_thinking`). Opus 5 defaults to adaptive thinking and rejects disabled thinking at `xhigh`/`max` effort. qgrid omits all three suppressors for these models (`usesAdaptiveThinking`) and uses `effort` to control depth.
+- One policy resolves the CLI thinking flag and effort before spawning. Omission preserves the existing model policy: Fable 5/5.1, Opus 5/5.5, Sonnet 5.5 and Haiku 5.5 keep adaptive defaults; other models use `--thinking disabled`. Explicit `true` uses `--thinking adaptive` and the normal effort resolution. Explicit `false` uses `--thinking disabled --effort low` only on supported models, ignoring caller/default effort. `CLAUDE_CODE_DISABLE_ADAPTIVE_THINKING` and `MAX_THINKING_TOKENS` are no longer sent. Do not reintroduce overlapping suppressors or inherit `CLAUDE_CODE_EFFORT_LEVEL`.
+- The 2026-10-10 off policy includes Haiku 4.5/5.5, Sonnet 4.5/4.6/5/5.5, and Opus 4.5/4.6/4.7/4.8/5. Fable 5/5.1 and Opus 5.5 reject disabled thinking at the backend. Off requests for unverified models fail before spawning instead of changing models or silently enabling thinking.
+- Haiku 5.5 and Sonnet 5.5 require Claude Code 2.1.295 or newer for explicit false. Cache the result of `claude --version`; unknown or older versions return an input error. The policy computes `CLAUDE_CODE_EXTRA_BODY` containing only `thinking: {type: "disabled"}` for Haiku or `thinking: {type: "between_tools"}` for Sonnet, alongside `--thinking disabled --effort low`. Native delivery was verified on 2.1.295; invalid-type and Sonnet-disabled negative controls confirmed that the CLI applied the body field. `between_tools` disables pre-answer thinking but may emit thinking blocks between tools. Never inherit this env variable or expose arbitrary body overrides to callers.
+- Legacy environment suppression versus CLI-only disabled was compared on Sonnet 4.6 and Opus 4.8: both completed with zero thinking tokens in the short-text baseline. The comparison supports removing the duplicate suppressors without changing the omitted-option policy.
+
 
 ## Spawn env
 
@@ -109,14 +113,13 @@ Included env:
 - `TMPDIR`
 - `CLAUDE_CODE_OAUTH_TOKEN`
 - `CLAUDE_CONFIG_DIR`
+- `CLAUDE_CODE_EXTRA_BODY` only when computed by the thinking policy for explicit Haiku/Sonnet 5.5 off; never inherited
 - `CLAUDE_CODE_DISABLE_AUTO_MEMORY=1`
-- `CLAUDE_CODE_DISABLE_ADAPTIVE_THINKING=1` except for Fable 5/5.1 and Opus 5
 - `CLAUDE_CODE_DISABLE_CLAUDE_MDS=1`
 - `CLAUDE_CODE_DISABLE_TERMINAL_TITLE=1`
 - `CLAUDE_CODE_DISABLE_BUNDLED_SKILLS=1`
 - `CLAUDE_CODE_DISABLE_WORKFLOWS=1`
 - `CLAUDE_CODE_ATTRIBUTION_HEADER=0`
-- `MAX_THINKING_TOKENS=0` except for Fable 5/5.1 and Opus 5
 - `CLAUDE_CODE_DISABLE_1M_CONTEXT=1` when model does not support qgrid's 1M path
 - `MAX_STRUCTURED_OUTPUT_RETRIES` for streaming structured output only (not non-streaming `generate`)
 

package/skills/qgrid/references/cli-env-and-server-boot.md

@@ -92,17 +92,20 @@ Some env values are editable at runtime through the dashboard's Settings page, b
 - Editable: Slack delivery control (master switch, reminder on/off, reminder interval, quiet-hours window, weekend behaviour), and the Slack channel/user-map/bot-token. `SETTING_DEFS` in `setting.constant.ts` is the single definition — key, env name, type, bounds, and whether the change applies immediately or needs a restart. The `settings` table itself (`setting.entity.json`) only stores `(key, value)` strings; everything the UI needs to render and validate a key lives in that constant, not in the schema.
 - Not editable: anything needed before the server can read its own database (`QGRID_DB_*`, `HOST`, `PORT`, `NODE_ENV`). These are exposed read-only on the same page so an operator can confirm which environment they are looking at; the DB password is masked.
 
-### Restart from the dashboard
+### Process lifecycle
 
-The Settings page can restart the server. There is no `pm2 restart` call — the process exits and the supervisor respawns it, which is what "restart" means here.
+qgrid does not detect a process supervisor or expose a restart API or dashboard
+control. Restarting, replacing, and recovering the process belong to the
+operator's deployment tooling. Settings marked `restart` still take effect on
+the next process start; the label does not imply an in-app restart action.
+
+There is no SIGTERM drain handler. Replacing the process during traffic can
+interrupt active responses, so deployment tooling must account for startup
+readiness and interruption. Ordinary request success/error/abort accounting and
+stale tool-wait cleanup remain independent of the deployment mechanism.
+
+### Live notification settings
 
-- `detectSupervisor` (`utils/process-supervisor.ts`) accepts only `pm_id` (pm2). `INVOCATION_ID` proves systemd launched a process but does not prove the unit has a restart policy, so it is not sufficient. Without pm2, `restartServer` returns 400 and the button is disabled: exiting with nothing to relaunch the process is a shutdown, not a restart. Presence is the signal, not the value — pm2 sets `pm_id=0` for the first process.
-- Browser requests must be same-host: when `Origin` is present and its host differs from `X-Forwarded-Host`/`Host`, `restartServer` returns 403. Origin-less operational clients such as `curl` remain supported. This is CSRF protection, not a substitute for authenticating the management API at Caddy.
-- The first restart request latches `restartPending`, blocks new native query/stream dispatch with 503, and marks only process-local active native request-log IDs as `error: server restarted`. Tool-result-waiting runs and externally owned logger runs are untouched. This is state reconciliation, not request draining.
-- Exit is process-wide and one-shot. It runs after the restart response emits `finish`; a response `close` or a five-second fallback covers disconnects. Concurrent restart calls do not schedule multiple exits.
-- The supervisor relaunches from `ecosystem.config.cjs`, not from the dying process's argv. That file carries no `--skip-update`, so each restart also runs the CLI self-update — a restart is a deploy. The confirmation dialog says so; dropping the update means adding `--skip-update` there.
-- The relaunch reuses the environment pm2 captured when the app was first started. Neither the dashboard restart nor `pm2 restart qgrid` re-reads `ecosystem.config.cjs`, so an env edit in that file (for example `QGRID_TOKEN_WINDOW_KEEPALIVE_ENABLED`) stays invisible until `pm2 restart ecosystem.config.cjs --update-env` (or `pm2 delete` + `pm2 start`) followed by `pm2 save`. Verify with `GET /api/setting/listSettings`, which reports the env the process actually sees. dev0 ran for a week with the keepalive env in the file but not in the process because of this (2026-09-02).
-- In-flight provider responses are not drained. There is no SIGTERM handler, so a restart during traffic cuts active responses, and the OpenAI pool needs 1–2 minutes before requests succeed again (`QgridDispatcher.startupState` answers 503 in that window). The dashboard stays locked after confirmation, observes an unavailable/not-ready health state, then waits for a later `ready: true` before refreshing settings.
 - Slack channel, bot-token, and user-map changes are read on the next notification. Changing the expiry-reminder interval or toggling the reminder replaces the current timer immediately without sending a notification merely because the setting changed; process boot still performs the intended immediate reminder run.
 - `slack.enabled` is the holiday switch: an operator turns it off for periods no rule can express. It suppresses ordinary notifications but not `urgent` ones, so a provider losing its last token still reports during a long break. `slack.remindersEnabled` is narrower — it stops only the repeat timer and preserves the chosen interval so re-enabling restores it.
 - Quiet hours are `slack.quietFromHour`/`slack.quietUntilHour` (0–23, Asia/Seoul) plus `slack.notifyOnWeekends`. From > until reads as an overnight window (20→8); from < until reads as a same-day window, which night-shift teams may want. Equal values mean no quiet window at all. Out-of-range values fall back to 20/8 at runtime, since env can bypass the stored-value validation.

package/skills/qgrid/references/decision-rationale.md

@@ -37,6 +37,15 @@ Key decisions:
 - The SDK should stay a thin adapter. Server APIs own request-log lifecycle, provider dispatch, structured-output emulation, and provider runtime details.
 - Request logging defaults to enabled. `providerOptions.qgrid.logger: false` is the single per-generation opt-out for both native qgrid logs and `createQgridLogger` telemetry logs; it must not alter generation or tool behavior.
 
+## Thinking Control
+
+Source: `docs/plans/2026-10-10-001-feat-thinking-control-plan.md`.
+
+- Use the independent `providerOptions.qgrid.thinking?: boolean` switch. False makes effort irrelevant; do not reject the combination or expose a cross-provider `effort: "none"` value.
+- Preserve omitted-option behavior. Reject unsupported and runtime-unverified off requests before token acquisition/spawn, without fallback. Backend capability and Claude Code delivery are separate evidence.
+- Use one Anthropic CLI thinking policy instead of overlapping environment suppressors. Preserve the OAuth env allowlist, token-specific config, empty settings, isolated cwd, memory/CLAUDE.md/skills/workflow disabling, tools, attribution, and 1M policy. These isolation controls prevent previously observed cross-request context injection and account metadata leakage.
+- Haiku/Sonnet 5.5 use a narrowly computed native `CLAUDE_CODE_EXTRA_BODY` thinking field, guarded by Claude Code >=2.1.295. Do not inherit that variable, expose arbitrary body overrides, add a transport, or add a relay. Sonnet's `between_tools` exception remains visible: it turns off pre-answer thinking but can retain thinking blocks between tools.
+
 ## Tool Calling And Request-Log Lifecycle
 
 Sources:

package/skills/qgrid/references/openai-codex-runtime.md

@@ -12,6 +12,12 @@ Use this reference before changing OpenAI transport, concurrency, routing, promp
 - Direct quota lookup: `openai-quota.ts`.
 - Provider integration and history: `packages/api/src/application/qgrid/qgrid.dispatcher.ts`, `conv-routing.ts`.
 
+## Thinking control
+
+`providerOptions.qgrid.thinking` reaches the Responses builder as a boolean. False overrides request/default effort and sends `reasoning: { effort: "none" }` without a reasoning summary. True uses a valid resolved effort, or explicitly sends `low` when effort is absent or unsupported, so a raw API request cannot fall back to a non-thinking backend default. Omission preserves the existing behavior: valid effort is forwarded; absent or unsupported effort leaves the backend default in effect. True and omission retain the requested reasoning summary policy. `none` belongs only to the internal wire type, not the public effort enum.
+
+The 2026-10-10 subscription probe accepted off for `gpt-6-sol`, `gpt-6-luna`, `gpt-5.6-sol`, `gpt-5.6-terra`, `gpt-5.6-luna`, and `gpt-5.5`. `gpt-6-astra` and `gpt-6.1-sol` explicitly rejected it. Other model IDs are unverified for off and are rejected locally only when false is requested. Model-access failures are not evidence of mandatory thinking. Validate terminal success and response settings, not HTTP 200 alone: errors can arrive within SSE. The same policy applies to text-model requests driving image generation, not to the image model itself.
+
 ## Adding an OpenAI model
 
 Use this checklist whenever a model is added. A public API release does not by itself prove

package/skills/qgrid/references/token-auth-quota-lifecycle.md

@@ -253,6 +253,8 @@ Anthropic's 5-hour usage window anchors on the first request, so an unused token
 
 Keepalive breaks that coupling by burning the window clock during idle time. Every keepalive-enabled active Anthropic token gets one cheap request as soon as its window expires, so windows always run and a workday lands mid-window rather than on a boundary. Across an eight-hour day this raises the number of windows the day can touch from a fixed 2 to an expected 2.6.
 
+The keepalive request uses `anthropic/claude-haiku-5-5`, `thinking: false`, `effort: "low"`, and the prompt `Reply OK.`. Thinking is disabled because this request only opens the usage window; the false policy overrides the supplied effort and uses the provider's low-effort off path. Haiku 5.5 off requires Claude Code >=2.1.295. The model constant lives in `token-window-keepalive.ts`; changing it does not change the scheduler or opt-in gates below.
+
 - Since 2.8.0 (`ea7feb3`) keepalive is opt-in at two levels. The runner instance must have env `QGRID_TOKEN_WINDOW_KEEPALIVE_ENABLED=true` (default off; set it on one instance per shared DB), and a token is a target only when `tokens.keepalive_enabled` is `true` (default `false`, toggled in the dashboard settings panel). The global setting `qgrid.tokenWindowKeepaliveEnabled` is a kill switch (`false` disables everything), not a substitute for either gate. In 2.7.6 every instance fired for every active token.
 - Operational trap: the 2.8.0 migration adds `keepalive_enabled` as default `false` without backfilling existing tokens. After upgrading, both the runner env and the per-token toggles must be set, or keepalive silently schedules zero tokens. This happened on dev0 after the 2026-08-31 self-update; a lingering 2.7.6 instance masked it until 2026-09-01. Diagnose with `GET /api/setting/listSettings` (`keepaliveRunnerEnabled`), `GET /api/token/findMany?subset=A` (`keepalive_enabled`), and the latest request log under project `qgrid-token-window-keepalive`.
 - The next fire is scheduled from the token's real `resetsAt` rather than a fixed five-hour timer, so a provider-side change to the window length is followed automatically.

package/skills/qgrid/references/tool-calling-and-multiturn.md

@@ -24,6 +24,16 @@ qgrid supports the AI SDK `tools` interface, but it does not use native provider
 - qgrid maps that structured object back into AI SDK `tool-call` content.
 - The AI SDK executes the actual tool functions and calls qgrid again with tool results.
 
+OpenAI can emit separate commentary and final-answer message items in one response.
+Plain commentary is an intermediate update, not an emulation payload: do not join
+it to the final JSON. Known final-answer message deltas still stream incrementally.
+For tool emulation, a complete message that passes the strict envelope and tool-name
+validation with `action: tool_call` is the client handoff boundary. Stop aggregating
+and streaming later text from that response: it cannot incorporate a tool result
+that the client has not produced yet. Drain the upstream response to preserve usage,
+errors, and cancellation. Never split concatenated JSON inside a single message or
+recover malformed envelopes. Ordinary text requests keep their existing aggregation.
+
 `generateText` and `streamText` default to `stepCountIs(1)`. When an executable
 tool can run before the final answer, callers must set a bounded `stopWhen`
 (for example `stepCountIs(3)`) or the AI SDK stops after the first tool-call
@@ -99,9 +109,16 @@ The schema explicitly tells the model:
 
 - use `tool_call` when client-side tool execution is needed;
 - use `answer` only for a final answer;
-- do not invoke listed tools as native Claude Code tools;
+- return the tool request as the JSON response body, then end the turn and wait for client results;
+- do not invent tool results or emit an answer before the required results arrive;
 - put tool arguments in `args` as a JSON string.
 
+Do not instruct OpenAI to use a `StructuredOutput` or Claude Code tool: neither is
+provided on this route. The legacy schema descriptions contained that stale
+instruction; schema descriptions now explain the actual client handoff protocol.
+The discriminated union, required fields, argument encoding and user-schema refs
+are unchanged. Prompt wording improves guidance but does not guarantee model compliance.
+
 Delivery differs by provider (SON-532). OpenAI: `qgrid.dispatcher.ts` strictifies the composed
 envelope through `buildStrictOutputSchema` and passes it as the response format — constrained
 decoding enforces it. Anthropic: no schema is passed to CC at all; `schema-prompt.ts` renders the