@eddyskywalker/dsh-chatgpt-subscription 0.12.4 → 0.13.0

package/lib/index.js

@@ -516,6 +516,17 @@ function catalogSnapshotPath(name) {
 	return path.join(dshHomeDir(), "storages", `${name}-catalog.json`);
 }
 /**
+* How many models one line's catalog holds, for the hub overview's "N/M models".
+*
+* An empty catalog is reported as `null` rather than `0`: "this line has synced
+* nothing yet" and "this line offers no models" are different statements, and
+* only the second one is a fact. The overview hides the model count for a null,
+* which is what a line that has never synced should show.
+*/
+function catalogTotal(count) {
+	return count === 0 ? null : count;
+}
+/**
 * Snapshot name for one catalog *and* one credential scope.
 *
 * A persisted snapshot answers before any credential is read — that is the whole
@@ -681,7 +692,7 @@ let snapshotRehydrated = null;
 * directly; returning it here under a different entry type would force every
 * one of them to convert it.
 */
-function getCachedCatalog() {
+function getCachedCatalog$2() {
 	return catalogCache$2?.models;
 }
 /**
@@ -740,7 +751,7 @@ async function performCatalogLoad$1(credentials, options) {
 		live = true;
 		writeCatalogSnapshot(catalogSnapshotName(SNAPSHOT_SCOPE, key), [...models], Date.now());
 	} catch {
-		models = getCachedCatalog() ?? [];
+		models = getCachedCatalog$2() ?? [];
 		live = false;
 	}
 	catalogCache$2 = {
@@ -791,7 +802,7 @@ function parseListing(payload) {
 function parseListingEntry$1(value) {
 	const record = asRecord$10(value);
 	if (record === void 0) return [];
-	const id = asString$17(record.slug) ?? asString$17(record.id);
+	const id = asString$18(record.slug) ?? asString$18(record.id);
 	if (id === void 0) return [];
 	const contextWindow = asNumber$4(record.context_window) ?? asNumber$4(record.contextWindow);
 	const modalities = parseModalities(record.input_modalities);
@@ -799,18 +810,18 @@ function parseListingEntry$1(value) {
 	const fallback = CODEX_MODEL_CATALOG.find((entry) => entry.id === id);
 	return [{
 		id,
-		name: asString$17(record.display_name) ?? asString$17(record.displayName) ?? fallback?.name ?? id,
+		name: asString$18(record.display_name) ?? asString$18(record.displayName) ?? fallback?.name ?? id,
 		contextWindow: contextWindow !== void 0 && contextWindow > 0 ? contextWindow : null,
 		inputModalities: modalities ?? (fallback ? [...fallback.inputModalities] : ["text"]),
 		...efforts !== null ? { reasoningEfforts: efforts } : {},
-		defaultReasoningEffort: asString$17(record.default_reasoning_level) ?? fallback?.defaultReasoningEffort ?? null
+		defaultReasoningEffort: asString$18(record.default_reasoning_level) ?? fallback?.defaultReasoningEffort ?? null
 	}];
 }
 function parseModalities(value) {
 	if (!Array.isArray(value)) return null;
 	const out = [];
 	for (const item of value) {
-		const name = asString$17(item);
+		const name = asString$18(item);
 		if (name === "text" || name === "image") out.push(name);
 	}
 	return out.length > 0 ? out : null;
@@ -825,7 +836,7 @@ function parseEfforts(value) {
 	if (!Array.isArray(value)) return null;
 	const out = [];
 	for (const item of value) {
-		const effort = asString$17(item) ?? asString$17(asRecord$10(item)?.effort);
+		const effort = asString$18(item) ?? asString$18(asRecord$10(item)?.effort);
 		if (effort !== void 0 && !out.includes(effort)) out.push(effort);
 	}
 	return out.length > 0 ? out : null;
@@ -843,7 +854,7 @@ function accountKeyFor$1(credentials) {
 function asRecord$10(value) {
 	return typeof value === "object" && value !== null && !Array.isArray(value) ? value : void 0;
 }
-function asString$17(value) {
+function asString$18(value) {
 	return typeof value === "string" && value !== "" ? value : void 0;
 }
 function asNumber$4(value) {
@@ -1085,6 +1096,22 @@ var WindowsDpapiTokenStore = class extends WindowsDpapiCredentialStore {
 		super(path, parseStoredCredentials);
 	}
 };
+/**
+* How long a helper may run before it is killed and the read fails.
+*
+* A healthy spawn costs ~200 ms (see the cache note on this class), so the guard
+* exists only for a child that never exits, and it was set at 10 s - fifty times
+* the healthy cost. That turned out to be too tight for one real environment: a
+* cold Windows runner whose freshly-installed dependency tree is being scanned
+* on first spawn, where the helper legitimately took longer than 10 s and the
+* kill turned a working credential read into `DPAPI helper timed out`.
+*
+* The guard is kept, and only its headroom widened: 30 s is still bounded, still
+* far above anything a warm machine needs, and a hung helper now blocks one
+* credential read for 30 s rather than 10 - a worse wait in a case that is already
+* a failure, in exchange for not failing a machine that is merely cold.
+*/
+const POWERSHELL_HELPER_TIMEOUT_MS = 3e4;
 function runPowerShell(script, path, stdin) {
 	return new Promise((resolve, reject) => {
 		const child = spawn("powershell.exe", [
@@ -1110,7 +1137,7 @@ function runPowerShell(script, path, stdin) {
 		const timer = setTimeout(() => {
 			child.kill();
 			reject(/* @__PURE__ */ new Error("DPAPI helper timed out"));
-		}, 1e4);
+		}, POWERSHELL_HELPER_TIMEOUT_MS);
 		child.stdout.setEncoding("utf8");
 		child.stdout.on("data", (chunk) => {
 			stdout += chunk;
@@ -2505,13 +2532,31 @@ var OAuthService = class {
 	*   "credentials are required" on every web search.
 	*/
 	async credentials(forceRefresh = false, access = {}) {
+		return (await this.credentialSelection(forceRefresh, access)).credentials;
+	}
+	/**
+	* The credential to use next, together with the pool row that owns it.
+	*
+	* The settings card keys each account's quota snapshot by the row the reading
+	* was made for, and the credential cannot do that job: a refresh rotates its
+	* tokens, and an account may state no id of its own. Callers that care about
+	* identity ask here; every other caller keeps using {@link credentials} and
+	* sees no difference.
+	*
+	* @param forceRefresh - refresh even a token that is not close to expiry.
+	* @param access - who is asking; see {@link credentials}.
+	*/
+	async credentialSelection(forceRefresh = false, access = {}) {
 		if (this.pool !== null) {
-			if (access.purpose === "tool") return (await this.pool.getCredentialAccount(this.fetchFn, forceRefresh)).credentials;
-			return (await this.pool.getEffectiveAccount(void 0, this.fetchFn, forceRefresh)).credentials;
+			const picked = access.purpose === "tool" ? await this.pool.getCredentialAccount(this.fetchFn, forceRefresh) : await this.pool.getEffectiveAccount(void 0, this.fetchFn, forceRefresh);
+			return {
+				accountId: picked.account.id,
+				credentials: picked.credentials
+			};
 		}
 		const stored = await this.loadAuthenticated();
-		if (forceRefresh || stored.expiresAt - this.now() <= 6e4) return this.refreshCredentials(stored);
-		return stored;
+		if (forceRefresh || stored.expiresAt - this.now() <= 6e4) return { credentials: await this.refreshCredentials(stored) };
+		return { credentials: stored };
 	}
 	dispose() {
 		if (this.disposed) return;
@@ -2807,6 +2852,49 @@ async function oauthErrorIdentifier(response) {
 	return null;
 }
 //#endregion
+//#region src/host/common/account-quota.ts
+/**
+* A reset instant as Unix milliseconds.
+*
+* Providers state this in seconds, in milliseconds, or not at all, and the two
+* numeric forms are told apart by magnitude rather than by a flag: 10^10
+* seconds is the year 2286, so anything larger is already milliseconds.
+*/
+function quotaResetMs(value) {
+	if (typeof value !== "number" || !Number.isFinite(value) || value <= 0) return null;
+	return value > 1e10 ? value : value * 1e3;
+}
+/**
+* One window, or null when its consumed share was never measured.
+*
+* `label` may be empty: the card names an unlabeled window from its length, and
+* only falls back to a generic name when the line stated neither.
+*/
+function quotaWindow$1(label, usedPercent, options = {}) {
+	if (typeof usedPercent !== "number" || !Number.isFinite(usedPercent)) return null;
+	const windowDurationMins = typeof options.windowDurationMins === "number" && Number.isFinite(options.windowDurationMins) ? options.windowDurationMins : null;
+	return {
+		label,
+		usedPercent: Math.min(100, Math.max(0, usedPercent)),
+		windowDurationMins,
+		resetsAt: quotaResetMs(options.resetsAt)
+	};
+}
+/**
+* The snapshot to publish for one account, or undefined when there is none.
+*
+* Undefined is the honest answer for "never read" and for "read before this
+* process started": both mean the card has nothing to show, and neither means
+* the account is unspent.
+*/
+function poolQuota(fetchedAt, windows) {
+	if (typeof fetchedAt !== "number" || !Number.isFinite(fetchedAt) || fetchedAt <= 0) return void 0;
+	return {
+		windows: windows.filter((window) => window !== null),
+		fetchedAt
+	};
+}
+//#endregion
 //#region src/host/codex-account-pool.ts
 /** Encrypted pool file the ChatGPT line owns; the single-credential store stays the mirror. */
 function codexPoolPath() {
@@ -2875,13 +2963,32 @@ const concurrency = new ConcurrencyGate();
 function chatGPTConcurrency() {
 	return concurrency;
 }
+/**
+* One ChatGPT usage snapshot as the account card's per-account quota.
+*
+* A bucket lists its windows either in `windows` or, on older payloads, only as
+* the `primary`/`secondary` pair — the same fallback the quota card itself
+* applies, so the account row cannot show fewer windows than the card does. The
+* label stays empty on purpose: a ChatGPT window is named by its length, and
+* the card localizes that length in the reader's own language.
+*/
+function codexAccountQuota(snapshot) {
+	return poolQuota(snapshot.fetchedAt, snapshot.usage.buckets.flatMap((bucket) => {
+		return (bucket.windows.length > 0 ? bucket.windows : [bucket.primary, bucket.secondary].filter((window) => window !== null)).map((window) => quotaWindow$1("", window.usedPercent, {
+			windowDurationMins: window.windowDurationMins,
+			resetsAt: window.resetsAt
+		}));
+	}));
+}
 var CodexAccountPool = class extends AccountPoolCore {
 	refresherRef;
 	store;
+	quotaSnapshotRef;
 	quotaBlockedUntil;
 	constructor(options = {}) {
 		const store = options.store ?? createPlatformTokenStore();
 		const ref = {};
+		const quotaSnapshotRef = {};
 		const hooks = {
 			providerId: "codex-chatgpt",
 			displayName: "ChatGPT",
@@ -2932,7 +3039,11 @@ var CodexAccountPool = class extends AccountPoolCore {
 			extendSummary: (account, base) => ({
 				...base,
 				...account.email === void 0 ? {} : { email: account.email },
-				...account.planType === void 0 ? {} : { planLabel: account.planType }
+				...account.planType === void 0 ? {} : { planLabel: account.planType },
+				...(() => {
+					const quota = quotaSnapshotRef.current?.(account);
+					return quota === void 0 ? {} : { quota };
+				})()
 			}),
 			emptyMessage: "未登录 ChatGPT 账号，请在「设置 → 订阅服务 → ChatGPT」中添加并登录账号。",
 			...options.backend === void 0 ? {} : { backend: options.backend },
@@ -2940,6 +3051,7 @@ var CodexAccountPool = class extends AccountPoolCore {
 		};
 		super(hooks);
 		this.refresherRef = ref;
+		this.quotaSnapshotRef = quotaSnapshotRef;
 		this.store = store;
 	}
 	/** Hand the pool the OAuth service's per-account refresh implementation. */
@@ -2957,6 +3069,18 @@ var CodexAccountPool = class extends AccountPoolCore {
 	setQuotaBlockedUntil(lookup) {
 		this.quotaBlockedUntil = lookup;
 	}
+	/**
+	* Wire in the account's own newest quota snapshot, for display only.
+	*
+	* The same map that already spares an exhausted account a wasted request also
+	* answers "what did this account's window look like last time we read it".
+	* The lookup must be a memory read: it runs while the settings card builds
+	* its account list, and that page must never fan out an upstream request per
+	* pooled account.
+	*/
+	setQuotaSnapshot(lookup) {
+		this.quotaSnapshotRef.current = lookup;
+	}
 	isEligible(account, now, triedAccountIds) {
 		if (!super.isEligible(account, now, triedAccountIds)) return false;
 		const blockedUntil = this.quotaBlockedUntil?.(account, now);
@@ -3309,13 +3433,13 @@ function createCodexFetchProvider(options = {}) {
 //#region src/host/common/llm-compat.ts
 /** This package's kind in the harness message-source map. */
 const PLUGIN_MESSAGE_SOURCE_KIND = "dsh-chatgpt-subscription";
-function isRecord$29(value) {
+function isRecord$30(value) {
 	return typeof value === "object" && value !== null && !Array.isArray(value);
 }
 /** Read a message body's blocks, keeping every entry that carries a block tag. */
 function readBlocks(raw) {
 	if (!Array.isArray(raw)) return [];
-	return raw.filter((block) => isRecord$29(block) && typeof block.type === "string");
+	return raw.filter((block) => isRecord$30(block) && typeof block.type === "string");
 }
 /**
 * Normalize one harness message list into {@link Message}.
@@ -3330,10 +3454,10 @@ function readBlocks(raw) {
 function normalizeMessages(messages) {
 	const normalized = [];
 	for (const raw of messages ?? []) {
-		if (!isRecord$29(raw)) continue;
+		if (!isRecord$30(raw)) continue;
 		const content = readBlocks(raw.content);
 		if (raw.role === "tool") {
-			const source = isRecord$29(raw.source) ? raw.source : void 0;
+			const source = isRecord$30(raw.source) ? raw.source : void 0;
 			const callId = String(raw.toolCallId ?? source?.callId ?? "");
 			normalized.push({
 				id: raw.id,
@@ -3355,7 +3479,7 @@ function normalizeMessages(messages) {
 			id: raw.id,
 			role: typeof raw.role === "string" ? raw.role : "user",
 			content,
-			...isRecord$29(raw.source) ? { source: raw.source } : {}
+			...isRecord$30(raw.source) ? { source: raw.source } : {}
 		});
 	}
 	return normalized;
@@ -4072,6 +4196,56 @@ function createControlledModelFetch(baseFetch, options) {
 	};
 }
 //#endregion
+//#region src/host/common/provider-route.ts
+/**
+* Claim one provider route for this plugin and keep claiming it.
+*
+* The route is taken immediately, and every `llm/adapters-updated` reports the
+* claim again — it is a no-op while the registration is held, and takes the
+* route back once another owner releases it. A rejected claim is remembered
+* rather than thrown, because the line whose route table asks through
+* {@link ProviderRouteClaim.conflict} still has to render.
+* @param ctx - The plugin context owning the `llm` service and the event seam.
+* @param options - The route id, the display label the two log lines name, and the adapter serving it.
+* @returns The claim's state and its disposer.
+*/
+function claimProviderRoute(ctx, options) {
+	let registration;
+	let conflict = null;
+	const claim = () => {
+		if (registration !== void 0) return;
+		try {
+			registration = ctx.llm.registerAdapter([options.providerId], options.adapter);
+			if (conflict !== null) ctx.logger.info(`[dsh-chatgpt-subscription] ${options.label} route "${options.providerId}" is now served by this plugin`);
+			conflict = null;
+		} catch (error) {
+			conflict = error instanceof Error ? error.message : String(error);
+			ctx.logger.warn(`[dsh-chatgpt-subscription] provider route "${options.providerId}" is already owned by another adapter; ${options.label} models keep being served by that one until its configuration is removed (${conflict})`);
+		}
+	};
+	claim();
+	const watch = options.watch === false || typeof ctx.on !== "function" ? void 0 : ctx.on("llm/adapters-updated", () => {
+		claim();
+	});
+	return {
+		serving: () => registration !== void 0,
+		conflict: () => conflict,
+		dispose: () => {
+			releaseHandle(watch);
+			registration?.();
+			registration = void 0;
+		}
+	};
+}
+/** Cordis event handles are either a disposer function or a disposable object. */
+function releaseHandle(handle) {
+	if (typeof handle === "function") {
+		handle();
+		return;
+	}
+	handle?.dispose?.();
+}
+//#endregion
 //#region src/host/common/request-diagnostics.ts
 /**
 * Reusable request diagnostics wrapper for model provider calls.
@@ -6223,6 +6397,72 @@ function acceptIdentity(current, incoming) {
 	return typeof incoming === "string" && incoming.length > 0 ? incoming : current;
 }
 //#endregion
+//#region src/shared/hub-contracts.ts
+/**
+* Wire contracts for the subscription hub's overview route.
+*
+* The overview is the settings page's opening screen: one card per provider
+* line, annotated with how many accounts it holds and whether it is enabled.
+* Everything here is a non-secret fact a settings card may render; no
+* credential, quota figure, or upstream call ever crosses this boundary.
+*/
+/** The aggregated, read-only overview the hub's opening screen renders. */
+const HUB_OVERVIEW_PATH = `${ROUTE_PREFIX$5}/hub/overview`;
+//#endregion
+//#region src/host/hub-overview.ts
+function registerHubOverviewRoutes(ctx, sources) {
+	const handler = async (request, response) => {
+		if (request.method !== "GET") {
+			response.writeHead(405, {
+				allow: "GET",
+				"content-type": "application/json; charset=utf-8"
+			});
+			response.end(JSON.stringify({
+				ok: false,
+				error: {
+					code: "bad-request",
+					message: "Method not allowed."
+				}
+			}));
+			return;
+		}
+		const envelope = {
+			ok: true,
+			value: { providers: (await Promise.allSettled(sources.map(async (source) => ({
+				id: source.id,
+				providerId: source.providerId,
+				canToggle: source.canToggle,
+				...await source.read()
+			})))).map((result, index) => {
+				if (result.status === "fulfilled") return result.value;
+				const source = sources[index];
+				return {
+					id: source.id,
+					providerId: source.providerId,
+					canToggle: source.canToggle,
+					enabled: !source.canToggle,
+					accountCount: 0,
+					authenticated: false,
+					enabledModelCount: null,
+					totalModelCount: null,
+					error: true
+				};
+			}) }
+		};
+		response.writeHead(200, {
+			"content-type": "application/json; charset=utf-8",
+			"cache-control": "no-store",
+			"x-content-type-options": "nosniff"
+		});
+		response.end(JSON.stringify(envelope));
+	};
+	return ctx.webServer.register({
+		kind: "exact",
+		path: HUB_OVERVIEW_PATH,
+		handler
+	});
+}
+//#endregion
 //#region src/host/common/same-origin.ts
 /**
 * Whether one mutation request comes from this application's own document.
@@ -6330,9 +6570,9 @@ var UsageService = class {
 		};
 		const now = this.now();
 		if (!force && !this.invalidated && this.cache !== null && now - this.cache.fetchedAt < 6e4 && this.cacheIdentityRevision === this.oauth.currentIdentityRevision()) return this.fromCache(false);
-		let credentials;
+		let selection;
 		try {
-			credentials = await this.oauth.credentials(false, USAGE_CREDENTIAL_ACCESS);
+			selection = await this.oauth.credentialSelection(false, USAGE_CREDENTIAL_ACCESS);
 		} catch (error) {
 			const reason = error instanceof Error && error.message !== "" ? ` (${error.message})` : "";
 			return this.failure({
@@ -6340,7 +6580,8 @@ var UsageService = class {
 				message: `ChatGPT credentials could not be refreshed.${reason}`
 			});
 		}
-		const accountKey = identityKey(credentials);
+		const credentials = selection.credentials;
+		const accountKey = this.snapshotKeyFor(selection.accountId, credentials);
 		if (this.cache !== null && this.cache.accountKey !== accountKey) this.clear();
 		if (!force && !this.invalidated && this.cache !== null && now - this.cache.fetchedAt < 6e4) return this.fromCache(false);
 		if (this.cache !== null && now - this.lastUpstreamAt < 15e3) return this.fromCache(this.invalidated || now - this.cache.fetchedAt >= 6e4);
@@ -6368,9 +6609,9 @@ var UsageService = class {
 	* The rejection is absorbed here: this refresh has no caller to receive it,
 	* and {@link fromCache} keeps serving the snapshot it failed to replace.
 	*/
-	startBackgroundRefresh(credentials, accountKey) {
+	startBackgroundRefresh(credentials, snapshotKey) {
 		if (this.inFlight !== null) return;
-		const task = this.refreshUpstream(credentials, accountKey).finally(() => {
+		const task = this.refreshUpstream(credentials, snapshotKey).finally(() => {
 			if (this.inFlight === task) this.inFlight = null;
 		});
 		this.inFlight = task;
@@ -6387,8 +6628,8 @@ var UsageService = class {
 	* cooldown rather than forever, so a missing upstream field cannot strand an
 	* account permanently.
 	*/
-	blockedUntilFor(credentials, now) {
-		const snapshot = this.snapshots.get(identityKey(credentials));
+	blockedUntilFor(accountId, credentials, now) {
+		const snapshot = this.snapshots.get(this.snapshotKeyFor(accountId, credentials));
 		if (snapshot === void 0) return void 0;
 		let reopen;
 		for (const bucket of snapshot.usage.buckets) {
@@ -6401,6 +6642,32 @@ var UsageService = class {
 		}
 		return reopen !== void 0 && reopen > now ? reopen : void 0;
 	}
+	/**
+	* The newest quota snapshot held for one account, if any.
+	*
+	* Quota follows the account, so the settings card shows each account its own
+	* progress. The pool asks here with the row id it addresses the account by —
+	* the identity that survives a token refresh — while its credentials answer
+	* for a composition without a pool. This reads the remembered map only, never
+	* the network, because opening the settings page must not cost one upstream
+	* request per pooled account.
+	*/
+	snapshotFor(accountId, credentials) {
+		return this.snapshots.get(this.snapshotKeyFor(accountId, credentials));
+	}
+	/**
+	* Key one account's remembered snapshot is filed under.
+	*
+	* A pool row id when the caller names one, prefixed so it can never be
+	* confused with the credential-derived key, and otherwise the credential
+	* identity — which is all a composition without a pool has. Two pooled accounts
+	* that state neither an account id nor an email would otherwise collapse onto
+	* one key and show each other's numbers, so the row id is the answer wherever
+	* a pool is in charge.
+	*/
+	snapshotKeyFor(accountId, credentials) {
+		return accountId === void 0 || accountId === "" ? identityKey(credentials) : `row:${accountId}`;
+	}
 	/** Remember the newest snapshot for one account, bounded to the accounts in use. */
 	rememberSnapshot(accountKey, usage, fetchedAt) {
 		this.snapshots.set(accountKey, {
@@ -6420,8 +6687,8 @@ var UsageService = class {
 		this.invalidated = false;
 	}
 	/** Drop the remembered snapshot of one account, after its window is known to have reset. */
-	forgetSnapshot(credentials) {
-		this.snapshots.delete(identityKey(credentials));
+	forgetSnapshot(accountId, credentials) {
+		this.snapshots.delete(this.snapshotKeyFor(accountId, credentials));
 	}
 	async consumeResetCredit() {
 		if (this.resetConsumeInFlight !== null) return this.resetConsumeInFlight;
@@ -6432,12 +6699,17 @@ var UsageService = class {
 	}
 	async consumeResetCreditUpstream() {
 		let credentials;
+		let accountId;
 		try {
-			credentials = await this.oauth.credentials(false, USAGE_CREDENTIAL_ACCESS);
+			const selection = await this.oauth.credentialSelection(false, USAGE_CREDENTIAL_ACCESS);
+			credentials = selection.credentials;
+			accountId = selection.accountId;
 			let creditsResponse = await this.fetchResetCredits(credentials);
 			if (creditsResponse.status === 401) {
 				await creditsResponse.body?.cancel().catch(() => void 0);
-				credentials = await this.oauth.credentials(true, USAGE_CREDENTIAL_ACCESS);
+				const renewed = await this.oauth.credentialSelection(true, USAGE_CREDENTIAL_ACCESS);
+				credentials = renewed.credentials;
+				accountId = renewed.accountId;
 				creditsResponse = await this.fetchResetCredits(credentials);
 			}
 			if (!creditsResponse.ok) {
@@ -6468,7 +6740,9 @@ var UsageService = class {
 			});
 			if (consumeResponse.status === 401) {
 				await consumeResponse.body?.cancel().catch(() => void 0);
-				credentials = await this.oauth.credentials(true, USAGE_CREDENTIAL_ACCESS);
+				const renewed = await this.oauth.credentialSelection(true, USAGE_CREDENTIAL_ACCESS);
+				credentials = renewed.credentials;
+				accountId = renewed.accountId;
 				consumeResponse = await this.fetchFn(CODEX_RESET_CREDITS_CONSUME_URL, {
 					method: "POST",
 					headers: {
@@ -6492,7 +6766,7 @@ var UsageService = class {
 			}
 			await consumeResponse.body?.cancel().catch(() => void 0);
 			this.clear();
-			const refreshed = await this.refreshUpstream(credentials, identityKey(credentials));
+			const refreshed = await this.refreshUpstream(credentials, this.snapshotKeyFor(accountId, credentials));
 			if (refreshed.error !== void 0) throw new UsageServiceError({
 				code: "quota-failed",
 				message: "The reset credit was used, but usage could not be refreshed."
@@ -6519,16 +6793,17 @@ var UsageService = class {
 			checkedAt: Math.floor(this.now() / 1e3)
 		};
 	}
-	async refreshUpstream(initialCredentials, initialAccountKey) {
+	async refreshUpstream(initialCredentials, initialSnapshotKey) {
 		this.lastUpstreamAt = this.now();
 		try {
 			let credentials = initialCredentials;
-			let accountKey = initialAccountKey;
+			let snapshotKey = initialSnapshotKey;
 			let response = await this.fetch(credentials);
 			if (response.status === 401) {
 				await response.body?.cancel().catch(() => void 0);
-				credentials = await this.oauth.credentials(true, USAGE_CREDENTIAL_ACCESS);
-				accountKey = identityKey(credentials);
+				const renewed = await this.oauth.credentialSelection(true, USAGE_CREDENTIAL_ACCESS);
+				credentials = renewed.credentials;
+				snapshotKey = this.snapshotKeyFor(renewed.accountId, credentials);
 				response = await this.fetch(credentials);
 			}
 			if (response.status === 429) {
@@ -6561,10 +6836,10 @@ var UsageService = class {
 			this.cache = {
 				usage,
 				fetchedAt: this.now(),
-				accountKey
+				accountKey: snapshotKey
 			};
 			this.cacheIdentityRevision = this.oauth.currentIdentityRevision();
-			this.rememberSnapshot(accountKey, usage, this.cache.fetchedAt);
+			this.rememberSnapshot(snapshotKey, usage, this.cache.fetchedAt);
 			this.invalidated = false;
 			return this.fromCache(false);
 		} catch (error) {
@@ -7055,7 +7330,7 @@ function field(value, name) {
 	const candidate = value[name];
 	return typeof candidate === "string" && candidate !== "" ? candidate : null;
 }
-function isRecord$28(value) {
+function isRecord$29(value) {
 	return typeof value === "object" && value !== null && !Array.isArray(value);
 }
 function readPreferencesUpdate(value, current) {
@@ -7089,7 +7364,7 @@ function readPreferencesUpdate(value, current) {
 		patch.searchProvider = value.searchProvider;
 	}
 	if ("contextWindowOverrides" in value) {
-		if (!isRecord$28(value.contextWindowOverrides)) throw new PreferenceError("contextWindowOverrides must be an object.");
+		if (!isRecord$29(value.contextWindowOverrides)) throw new PreferenceError("contextWindowOverrides must be an object.");
 		const overrides = {};
 		for (const [model, contextWindow] of Object.entries(value.contextWindowOverrides)) {
 			if (!isCodexModelId(model)) throw new PreferenceError("Unknown Codex model for a context window override.");
@@ -7903,7 +8178,7 @@ var FileModelSettingsStore$2 = class {
 */
 /** Default loopback port the Antigravity CLI's own flow uses. */
 const DEFAULT_CALLBACK_PORT$2 = 51121;
-let cachedQuota$1;
+let cachedQuota;
 let quotaFetchInFlight = null;
 /** Bumped by every cache clear so an in-flight fetch cannot publish stale state. */
 let quotaCacheEpoch$2 = 0;
@@ -8080,7 +8355,7 @@ function parseQuotaSummary(data) {
 		for (const bucket of rawBuckets) {
 			if (typeof bucket !== "object" || bucket === null) continue;
 			const bRec = bucket;
-			const remaining = typeof bRec.remainingFraction === "number" ? Math.max(0, Math.min(1, bRec.remainingFraction)) : 0;
+			const remaining = typeof bRec.remainingFraction === "number" && Number.isFinite(bRec.remainingFraction) ? Math.max(0, Math.min(1, bRec.remainingFraction)) : null;
 			buckets.push({
 				bucketId: String(bRec.bucketId || bRec.displayName || "limit"),
 				displayName: String(bRec.displayName || bRec.bucketId || "Limit"),
@@ -8143,7 +8418,7 @@ function mergeCatalogSelection(stored, storedCatalog) {
 	return stored.filter((id) => known.has(id));
 }
 async function fetchAccountQuota(store = new FileCredentialStore$2(), modelSettings, fetchFn = fetch, force = false) {
-	if (!force && cachedQuota$1 && Date.now() - (cachedQuota$1.fetchedAt || 0) < 12e4) return cachedQuota$1;
+	if (!force && cachedQuota && Date.now() - (cachedQuota.fetchedAt || 0) < 12e4) return cachedQuota;
 	if (quotaFetchInFlight) return quotaFetchInFlight;
 	const epoch = quotaCacheEpoch$2;
 	const request = (async () => {
@@ -8179,7 +8454,7 @@ async function fetchAccountQuota(store = new FileCredentialStore$2(), modelSetti
 			fetchedAt: Date.now()
 		};
 		if (epoch !== quotaCacheEpoch$2) return snapshot;
-		cachedQuota$1 = snapshot;
+		cachedQuota = snapshot;
 		if (modelSettings && catalogModels.length > 0) {
 			const current = await modelSettings.read();
 			const mergedEnabled = mergeCatalogSelection(current.enabledModelIds, current.catalogModels);
@@ -8195,11 +8470,11 @@ async function fetchAccountQuota(store = new FileCredentialStore$2(), modelSetti
 	}
 }
 function getCachedQuota() {
-	return cachedQuota$1;
+	return cachedQuota;
 }
 function clearCachedQuota() {
 	quotaCacheEpoch$2 += 1;
-	cachedQuota$1 = void 0;
+	cachedQuota = void 0;
 	quotaFetchInFlight = null;
 }
 //#endregion
@@ -8589,7 +8864,7 @@ const schemaFields = /* @__PURE__ */ new Set([
 	"required",
 	"propertyOrdering"
 ]);
-function isRecord$27(value) {
+function isRecord$28(value) {
 	return typeof value === "object" && value !== null && !Array.isArray(value);
 }
 function mergeSchemas(left, right) {
@@ -8597,7 +8872,7 @@ function mergeSchemas(left, right) {
 		...left,
 		...right
 	};
-	if (isRecord$27(left.properties) && isRecord$27(right.properties)) merged.properties = {
+	if (isRecord$28(left.properties) && isRecord$28(right.properties)) merged.properties = {
 		...left.properties,
 		...right.properties
 	};
@@ -8633,9 +8908,9 @@ const mergedTypeKeys = /* @__PURE__ */ new Set([
 * model with strictly less than the schema it started from.
 */
 function describeBranch(branch) {
-	const type = typeof branch.type === "string" ? branch.type : isRecord$27(branch.properties) ? "object" : isRecord$27(branch.items) ? "array" : "value";
+	const type = typeof branch.type === "string" ? branch.type : isRecord$28(branch.properties) ? "object" : isRecord$28(branch.items) ? "array" : "value";
 	const details = [];
-	if (isRecord$27(branch.properties)) {
+	if (isRecord$28(branch.properties)) {
 		const names = Object.keys(branch.properties);
 		if (names.length > 0) details.push(`{${names.join(",")}}`);
 	}
@@ -8649,7 +8924,7 @@ function describeBranch(branch) {
 }
 /** How well one branch can stand in for a whole union; a named object shape wins. */
 function branchRank(branch) {
-	if (isRecord$27(branch.properties)) return 3;
+	if (isRecord$28(branch.properties)) return 3;
 	if (branch.type === "object") return 2;
 	if (branch.type === "array") return 1;
 	return 0;
@@ -8678,7 +8953,7 @@ function unionShape(branches) {
 		shape: branches[0],
 		folded: false
 	};
-	if (new Set(branches.map((branch) => branch.type)).size === 1 && branches.every((branch) => !isRecord$27(branch.properties) && !isRecord$27(branch.items))) {
+	if (new Set(branches.map((branch) => branch.type)).size === 1 && branches.every((branch) => !isRecord$28(branch.properties) && !isRecord$28(branch.items))) {
 		const shape = { type: branches[0].type };
 		if (branches.every((branch) => Array.isArray(branch.enum))) {
 			const values = [...new Set(branches.flatMap((branch) => branch.enum))];
@@ -8715,7 +8990,7 @@ function foldOwnFields(node) {
 	const out = {};
 	for (const [key, value] of Object.entries(node)) {
 		if (key === "anyOf" || key === "oneOf" || key === "allOf") continue;
-		if (key === "properties" && isRecord$27(value)) {
+		if (key === "properties" && isRecord$28(value)) {
 			out.properties = Object.fromEntries(Object.entries(value).map(([name, child]) => [name, foldSchemaUnions(child)]));
 			continue;
 		}
@@ -8744,15 +9019,15 @@ function foldOwnFields(node) {
 * argument the model sends anyway fails there, naming the actual problem.
 */
 function foldSchemaUnions(node) {
-	if (!isRecord$27(node)) return node;
+	if (!isRecord$28(node)) return node;
 	let own = foldOwnFields(node);
 	if (Array.isArray(node.allOf)) for (const branch of node.allOf) {
 		const folded = foldSchemaUnions(branch);
-		if (isRecord$27(folded)) own = mergeSchemas(own, folded);
+		if (isRecord$28(folded)) own = mergeSchemas(own, folded);
 	}
 	const alternatives = Array.isArray(node.anyOf) ? node.anyOf : Array.isArray(node.oneOf) ? node.oneOf : [];
 	if (alternatives.length === 0) return own;
-	const branches = alternatives.map((branch) => foldSchemaUnions(branch)).filter(isRecord$27);
+	const branches = alternatives.map((branch) => foldSchemaUnions(branch)).filter(isRecord$28);
 	if (branches.length === 0) return own;
 	const shapes = branches.filter((branch) => branch.type !== "null");
 	const acceptsNull = branches.some((branch) => branch.type === "null");
@@ -8768,7 +9043,7 @@ function foldSchemaUnions(node) {
 	return merged;
 }
 function toAntigravityToolSchema(schema, options = {}) {
-	if (!isRecord$27(schema)) return schema;
+	if (!isRecord$28(schema)) return schema;
 	const root = schema;
 	function resolveReference(reference) {
 		let target = root;
@@ -8776,13 +9051,13 @@ function toAntigravityToolSchema(schema, options = {}) {
 		const path = reference === "#" ? [] : reference.slice(2).split("/");
 		for (const segment of path) {
 			const key = segment.replace(/~1/g, "/").replace(/~0/g, "~");
-			target = isRecord$27(target) && Object.hasOwn(target, key) ? target[key] : void 0;
+			target = isRecord$28(target) && Object.hasOwn(target, key) ? target[key] : void 0;
 		}
-		if (!isRecord$27(target)) throw new LlmError(`Antigravity tool schema reference was not found: ${reference}`, "PROVIDER_ERROR");
+		if (!isRecord$28(target)) throw new LlmError(`Antigravity tool schema reference was not found: ${reference}`, "PROVIDER_ERROR");
 		return target;
 	}
 	function convert(node, references) {
-		if (!isRecord$27(node)) return {};
+		if (!isRecord$28(node)) return {};
 		let inherited = {};
 		if (typeof node.$ref === "string") {
 			if (references.has(node.$ref)) throw new LlmError(`Antigravity tool schema contains a recursive reference: ${node.$ref}`, "PROVIDER_ERROR");
@@ -8791,8 +9066,8 @@ function toAntigravityToolSchema(schema, options = {}) {
 		if (Array.isArray(node.allOf)) for (const branch of node.allOf) inherited = mergeSchemas(inherited, convert(branch, references));
 		const out = {};
 		for (const [key, value] of Object.entries(node)) if (schemaFields.has(key)) out[key] = value;
-		if (isRecord$27(node.properties)) out.properties = Object.fromEntries(Object.entries(node.properties).map(([name, child]) => [name, convert(child, references)]));
-		if (isRecord$27(node.items)) out.items = convert(node.items, references);
+		if (isRecord$28(node.properties)) out.properties = Object.fromEntries(Object.entries(node.properties).map(([name, child]) => [name, convert(child, references)]));
+		if (isRecord$28(node.items)) out.items = convert(node.items, references);
 		const alternatives = Array.isArray(node.anyOf) ? node.anyOf : node.oneOf;
 		if (Array.isArray(alternatives)) out.anyOf = alternatives.map((child) => convert(child, references));
 		if (Array.isArray(node.type)) {
@@ -8828,10 +9103,10 @@ let toolCallCounter = 0;
 function sanitizeText$5(text) {
 	return text.replace(/\0/g, "");
 }
-function isRecord$26(value) {
+function isRecord$27(value) {
 	return typeof value === "object" && value !== null && !Array.isArray(value);
 }
-function asString$16(value) {
+function asString$17(value) {
 	return typeof value === "string" ? value : void 0;
 }
 function safeJsonParse$5(text) {
@@ -8848,25 +9123,25 @@ function toolCallIdNeeded(modelId, runtimeModel) {
 	return modelId.startsWith("claude-") || modelId.startsWith("gpt-oss-") || runtimeModel.startsWith("claude-") || runtimeModel.startsWith("gpt-oss-");
 }
 function parseArguments$1(raw) {
-	if (isRecord$26(raw)) return raw;
+	if (isRecord$27(raw)) return raw;
 	if (raw === void 0 || raw === null || raw === "") return {};
 	const parsed = typeof raw === "string" ? safeJsonParse$5(raw) : raw;
-	return isRecord$26(parsed) ? parsed : {};
+	return isRecord$27(parsed) ? parsed : {};
 }
 const NO_RESOLVED_IMAGES$5 = /* @__PURE__ */ new Map();
 function attachmentOf$5(block) {
 	const attachment = block.attachment;
-	if (!isRecord$26(attachment)) return void 0;
+	if (!isRecord$27(attachment)) return void 0;
 	return typeof attachment.attachmentId === "string" ? attachment : void 0;
 }
 function attachmentLabel$6(block) {
-	const attachment = isRecord$26(block.attachment) ? block.attachment : void 0;
-	return asString$16(attachment?.name) || asString$16(attachment?.attachmentId);
+	const attachment = isRecord$27(block.attachment) ? block.attachment : void 0;
+	return asString$17(attachment?.name) || asString$17(attachment?.attachmentId);
 }
 function collectImageRefs$4(content, refs) {
 	if (!Array.isArray(content)) return;
 	for (const block of content) {
-		if (!isRecord$26(block)) continue;
+		if (!isRecord$27(block)) continue;
 		if (block.type === "image") {
 			const attachment = attachmentOf$5(block);
 			if (attachment) refs.set(attachment.attachmentId, attachment);
@@ -8904,13 +9179,13 @@ function base64Length$4(bytes) {
 function requestImageBytes$4(block) {
 	const attachment = attachmentOf$5(block);
 	if (attachment) return base64Length$4(attachment.bytes);
-	const inline = asString$16(block.data) || asString$16(block.base64);
+	const inline = asString$17(block.data) || asString$17(block.base64);
 	return inline ? inline.length : void 0;
 }
 function collectRequestImageBytes$4(content, lengths) {
 	if (!Array.isArray(content)) return;
 	for (const block of content) {
-		if (!isRecord$26(block) || block.type !== "image") continue;
+		if (!isRecord$27(block) || block.type !== "image") continue;
 		const bytes = requestImageBytes$4(block);
 		if (bytes !== void 0) lengths.push(bytes);
 	}
@@ -8945,7 +9220,7 @@ function offloadOldestRequestImages$4(options) {
 		if (remaining.count === 0 || !Array.isArray(message.content)) return message;
 		let replaced = false;
 		const content = message.content.map((block) => {
-			if (remaining.count === 0 || !isRecord$26(block) || block.type !== "image") return block;
+			if (remaining.count === 0 || !isRecord$27(block) || block.type !== "image") return block;
 			if (requestImageBytes$4(block) === void 0) return block;
 			remaining.count -= 1;
 			replaced = true;
@@ -9008,10 +9283,10 @@ function unavailableImageText$6(block) {
 	return `[image unavailable: ${label ? `${label} could not be read` : "the image could not be read"}; ask the user to attach it again if the image is needed]`;
 }
 function imageBlockToPart(block, images) {
-	let data = asString$16(block.data) || asString$16(block.base64);
-	const source = isRecord$26(block.source) ? block.source : void 0;
-	if (!data && source) data = asString$16(source.data) || asString$16(source.base64);
-	let mimeType = asString$16(block.mimeType) || asString$16(block.mediaType) || (source ? asString$16(source.mimeType) || asString$16(source.mediaType) : void 0) || "image/png";
+	let data = asString$17(block.data) || asString$17(block.base64);
+	const source = isRecord$27(block.source) ? block.source : void 0;
+	if (!data && source) data = asString$17(source.data) || asString$17(source.base64);
+	let mimeType = asString$17(block.mimeType) || asString$17(block.mediaType) || (source ? asString$17(source.mimeType) || asString$17(source.mediaType) : void 0) || "image/png";
 	if (data?.startsWith("data:")) {
 		const match = data.match(/^data:([^;,]+);base64,(.*)$/s);
 		if (match) {
@@ -9034,8 +9309,8 @@ function contentToUserParts(content, images) {
 	if (typeof content === "string") return [{ text: sanitizeText$5(content) }];
 	if (!Array.isArray(content)) return [];
 	const parts = [];
-	for (const block of content) if (isRecord$26(block) && block.type === "text" && typeof block.text === "string") parts.push({ text: sanitizeText$5(block.text) });
-	else if (isRecord$26(block) && block.type === "image") {
+	for (const block of content) if (isRecord$27(block) && block.type === "text" && typeof block.text === "string") parts.push({ text: sanitizeText$5(block.text) });
+	else if (isRecord$27(block) && block.type === "image") {
 		const img = imageBlockToPart(block, images);
 		parts.push(img ?? { text: unavailableImageText$6(block) });
 	}
@@ -9050,7 +9325,7 @@ function toolResultImageParts(blocks, images) {
 	if (!Array.isArray(blocks)) return [];
 	const parts = [];
 	for (const block of blocks) {
-		if (!isRecord$26(block)) continue;
+		if (!isRecord$27(block)) continue;
 		if (block.type === "image") {
 			parts.push(imageBlockToPart(block, images) ?? { text: unavailableImageText$6(block) });
 			continue;
@@ -9062,7 +9337,7 @@ function toolResultImageParts(blocks, images) {
 function toolResultText$5(blocks) {
 	if (!Array.isArray(blocks)) return "";
 	return blocks.map((block) => {
-		if (!isRecord$26(block)) return "";
+		if (!isRecord$27(block)) return "";
 		if (block.type === "text" && typeof block.text === "string") return sanitizeText$5(block.text);
 		if (block.type === "tool-result") return toolResultText$5(block.content);
 		if (block.type === "image") {
@@ -9076,16 +9351,16 @@ function replayBlockFor$1(message, index) {
 	const source = message.source;
 	if (!source || source.kind !== "model" || source.provider !== "antigravity") return void 0;
 	const state = source.replayState;
-	if (!isRecord$26(state)) return void 0;
+	if (!isRecord$27(state)) return void 0;
 	if (Array.isArray(state.blocks)) return state.blocks[index];
-	const resp = isRecord$26(state.response) ? state.response : void 0;
+	const resp = isRecord$27(state.response) ? state.response : void 0;
 	if (resp) {
 		if (Array.isArray(resp.outputItems)) return resp.outputItems[index];
 		if (Array.isArray(resp.blocks)) return resp.blocks[index];
 	}
 }
 function thoughtSignature(part) {
-	return asString$16(part?.thoughtSignature) || asString$16(part?.thought_signature) || asString$16(part?.thinkingSignature) || asString$16(part?.textSignature);
+	return asString$17(part?.thoughtSignature) || asString$17(part?.thought_signature) || asString$17(part?.thinkingSignature) || asString$17(part?.textSignature);
 }
 function replayPart(part) {
 	const copy = { ...part };
@@ -9138,11 +9413,11 @@ function assistantParts(message, model, runtimeModel, toolCalls) {
 	if (!Array.isArray(message.content)) return parts;
 	for (let index = 0; index < message.content.length; index++) {
 		const block = message.content[index];
-		if (!isRecord$26(block)) continue;
+		if (!isRecord$27(block)) continue;
 		if (block.type === "reasoning" && runtimeModel.startsWith("claude-") && (message.source?.kind !== "model" || message.source.provider !== "antigravity" || !("model" in message.source) || message.source.model !== model.id)) continue;
 		const replay = replayBlockFor$1(message, index);
-		const originalParts = Array.isArray(replay?.parts) ? replay.parts.filter(isRecord$26) : [];
-		if ((block.type === "text" || block.type === "reasoning") && originalParts.length > 0 && originalParts.every((part) => !part.functionCall) && originalParts.map((part) => asString$16(part.text) || "").join("") === sanitizeText$5(String(block.text || ""))) {
+		const originalParts = Array.isArray(replay?.parts) ? replay.parts.filter(isRecord$27) : [];
+		if ((block.type === "text" || block.type === "reasoning") && originalParts.length > 0 && originalParts.every((part) => !part.functionCall) && originalParts.map((part) => asString$17(part.text) || "").join("") === sanitizeText$5(String(block.text || ""))) {
 			parts.push(...originalParts.map(replayPart));
 			continue;
 		}
@@ -9162,8 +9437,8 @@ function assistantParts(message, model, runtimeModel, toolCalls) {
 		} else if (block.type === "tool-call") {
 			const toolId = String(block.id || "");
 			const toolName = String(block.name || "");
-			const originalCall = originalParts.find((part) => isRecord$26(part.functionCall));
-			const wireId = asString$16((isRecord$26(originalCall?.functionCall) ? originalCall.functionCall : void 0)?.id) || (toolCallIdNeeded(model.id, runtimeModel) ? sanitizeToolCallId(toolId, toolName) : originalCall ? void 0 : toolId || void 0);
+			const originalCall = originalParts.find((part) => isRecord$27(part.functionCall));
+			const wireId = asString$17((isRecord$27(originalCall?.functionCall) ? originalCall.functionCall : void 0)?.id) || (toolCallIdNeeded(model.id, runtimeModel) ? sanitizeToolCallId(toolId, toolName) : originalCall ? void 0 : toolId || void 0);
 			toolCalls.set(toolId, {
 				name: toolName,
 				id: wireId
@@ -9215,7 +9490,7 @@ function convertMessages(options, model, runtimeModel, images = NO_RESOLVED_IMAG
 			continue;
 		}
 		const content = Array.isArray(message.content) ? message.content : [];
-		const userParts = contentToUserParts(content.filter((b) => !isRecord$26(b) || b.type !== "tool-result"), images);
+		const userParts = contentToUserParts(content.filter((b) => !isRecord$27(b) || b.type !== "tool-result"), images);
 		if (role === "system") {
 			if (userParts.length) contents.push({
 				role: GEMINI_ROLE.user,
@@ -9227,7 +9502,7 @@ function convertMessages(options, model, runtimeModel, images = NO_RESOLVED_IMAG
 			role: GEMINI_ROLE.user,
 			parts: userParts
 		});
-		for (const b of content) if (isRecord$26(b) && b.type === "tool-result") pushToolResult(contents, b, toolCalls, model, runtimeModel, images);
+		for (const b of content) if (isRecord$27(b) && b.type === "tool-result") pushToolResult(contents, b, toolCalls, model, runtimeModel, images);
 	}
 	return contents;
 }
@@ -9362,7 +9637,7 @@ const USAGE_FIELDS = [
 	"totalTokenCount"
 ];
 function collectUsage(value, state) {
-	if (!isRecord$26(value)) return;
+	if (!isRecord$27(value)) return;
 	for (const key of USAGE_FIELDS) {
 		const count = value[key];
 		if (typeof count !== "number" || !Number.isSafeInteger(count) || count < 0) continue;
@@ -9392,17 +9667,17 @@ function processStreamLine$2(line, state) {
 	}
 	if (!json) return [];
 	const chunk = safeJsonParse$5(json);
-	if (!isRecord$26(chunk)) return [];
-	const responseData = isRecord$26(chunk.response) ? chunk.response : chunk;
-	const error = isRecord$26(chunk.error) ? chunk.error : isRecord$26(responseData.error) ? responseData.error : void 0;
-	if (error !== void 0) throw new LlmError$1(`Antigravity stream error: ${asString$16(error.message) ?? "unknown error"}`, isContextOverflow(error) ? CONTEXT_OVERFLOW_CODE : "PROVIDER_ERROR");
+	if (!isRecord$27(chunk)) return [];
+	const responseData = isRecord$27(chunk.response) ? chunk.response : chunk;
+	const error = isRecord$27(chunk.error) ? chunk.error : isRecord$27(responseData.error) ? responseData.error : void 0;
+	if (error !== void 0) throw new LlmError$1(`Antigravity stream error: ${asString$17(error.message) ?? "unknown error"}`, isContextOverflow(error) ? CONTEXT_OVERFLOW_CODE : "PROVIDER_ERROR");
 	const candidates = Array.isArray(responseData.candidates) ? responseData.candidates : [];
-	const candidate = isRecord$26(candidates[0]) ? candidates[0] : void 0;
-	const content = isRecord$26(candidate?.content) ? candidate.content : void 0;
+	const candidate = isRecord$27(candidates[0]) ? candidates[0] : void 0;
+	const content = isRecord$27(candidate?.content) ? candidate.content : void 0;
 	const parts = Array.isArray(content?.parts) ? content.parts : [];
 	const out = [];
 	for (const part of parts) {
-		if (!isRecord$26(part)) continue;
+		if (!isRecord$27(part)) continue;
 		if (typeof part.text === "string" && part.text !== "") {
 			const isThinking = Boolean(part.thought);
 			const blockType = isThinking ? "reasoning" : "text";
@@ -9434,7 +9709,7 @@ function processStreamLine$2(line, state) {
 				index: state.currentBlock.index,
 				text: delta
 			});
-		} else if (!isRecord$26(part.functionCall) && thoughtSignature(part)) {
+		} else if (!isRecord$27(part.functionCall) && thoughtSignature(part)) {
 			if (state.replayBlocks.length === 0) {
 				const type = part.thought ? "reasoning" : "text";
 				state.blocks.push({
@@ -9458,12 +9733,12 @@ function processStreamLine$2(line, state) {
 			}
 			state.replayBlocks[state.replayBlocks.length - 1].parts.push(replayPart(part));
 		}
-		if (isRecord$26(part.functionCall)) {
+		if (isRecord$27(part.functionCall)) {
 			out.push(...closeCurrentBlock(state));
 			const fc = part.functionCall;
-			const toolName = asString$16(fc.name) || "";
-			const toolId = asString$16(fc.id) || sanitizeToolCallId("", toolName);
-			const argsText = JSON.stringify(isRecord$26(fc.args) ? fc.args : {});
+			const toolName = asString$17(fc.name) || "";
+			const toolId = asString$17(fc.id) || sanitizeToolCallId("", toolName);
+			const argsText = JSON.stringify(isRecord$27(fc.args) ? fc.args : {});
 			const index = state.blocks.length;
 			const block = {
 				type: "tool-call",
@@ -9500,7 +9775,7 @@ function processStreamLine$2(line, state) {
 	}
 	collectUsage(chunk.usageMetadata, state);
 	if (responseData !== chunk) collectUsage(responseData.usageMetadata, state);
-	const finishReason = asString$16(candidate?.finishReason) || asString$16(responseData.finishReason);
+	const finishReason = asString$17(candidate?.finishReason) || asString$17(responseData.finishReason);
 	if (finishReason) {
 		state.finishReason = finishReason;
 		out.push(...closeCurrentBlock(state));
@@ -9579,6 +9854,54 @@ function createPoolCredentialBackend(filePath) {
 	throw new Error("Antigravity pool encrypted storage requires Windows, macOS, or Linux.");
 }
 /**
+* One bucket's reset moment as Unix milliseconds, when it parses as a date.
+*
+* The service states it as an ISO-ish string; anything `Date.parse` refuses is
+* dropped rather than guessed at, because a wrong instant on the card is worse
+* than none.
+*/
+function resetTimeMs(resetTime) {
+	if (resetTime === void 0 || resetTime === "") return null;
+	const parsed = Date.parse(resetTime);
+	return Number.isFinite(parsed) && parsed > 0 ? parsed : null;
+}
+/**
+* One cached Antigravity snapshot as the shared account card renders it.
+*
+* One window per bucket. The service states what is LEFT, and the card draws what
+* is SPENT, so the used share is the complement. A bucket whose fraction is
+* `null` (never measured) or outside 0-1 is dropped rather than drawn as 0% or
+* 100%: "not stated" and "nothing used" are different claims and only one of
+* them is true. A stated 0 is a real measurement — the allowance is exhausted —
+* and is published as 100% used.
+*
+* `windowDurationMins` is deliberately absent. The bucket's `window` field is
+* carried through this line's cache verbatim and nothing in the line interprets
+* it as a duration, so naming a window from it would be a guess; the card names
+* each window from the label below instead.
+*/
+function antigravityPoolQuota(quota) {
+	if (quota === void 0) return void 0;
+	const nameCounts = /* @__PURE__ */ new Map();
+	for (const group of quota.groups) for (const bucket of group.buckets) {
+		const name = bucket.displayName.trim();
+		if (name === "") continue;
+		nameCounts.set(name, (nameCounts.get(name) ?? 0) + 1);
+	}
+	const windows = [];
+	for (const group of quota.groups) {
+		const groupName = group.displayName.trim();
+		for (const bucket of group.buckets) {
+			const remaining = bucket.remainingFraction;
+			if (remaining === null || !Number.isFinite(remaining) || remaining < 0 || remaining > 1) continue;
+			const name = bucket.displayName.trim();
+			const label = (name === "" || (nameCounts.get(name) ?? 0) > 1) && groupName !== "" ? name === "" ? groupName : groupName + " · " + name : name;
+			windows.push(quotaWindow$1(label, (1 - remaining) * 100, { resetsAt: resetTimeMs(bucket.resetTime) }));
+		}
+	}
+	return poolQuota(quota.fetchedAt, windows);
+}
+/**
 * Antigravity's account pool.
 *
 * The storage, eligibility, rotation and cooldown rules live in the shared
@@ -9642,7 +9965,11 @@ var AccountPoolStore = class {
 					...base,
 					...projectId === void 0 ? {} : { projectId },
 					...email === void 0 ? {} : { email },
-					...account.planLabel === void 0 ? {} : { planLabel: account.planLabel }
+					...account.planLabel === void 0 ? {} : { planLabel: account.planLabel },
+					...account.isPrimary === true ? (() => {
+						const quota = antigravityPoolQuota(getCachedQuota());
+						return quota === void 0 ? {} : { quota };
+					})() : {}
 				};
 			},
 			emptyMessage: "未登录 Antigravity 账号，请在「设置 → Antigravity」中添加并登录账号。",
@@ -11859,13 +12186,13 @@ function commandCodeHeaders(apiKey, extra = {}) {
 function timeoutSignal$3(signal, ms) {
 	return signal ? AbortSignal.any([signal, AbortSignal.timeout(ms)]) : AbortSignal.timeout(ms);
 }
-function isRecord$25(value) {
+function isRecord$26(value) {
 	return typeof value === "object" && value !== null && !Array.isArray(value);
 }
 function asRecord$9(value) {
-	return isRecord$25(value) ? value : void 0;
+	return isRecord$26(value) ? value : void 0;
 }
-function asString$15(value) {
+function asString$16(value) {
 	if (typeof value === "string" && value.trim() !== "") return value;
 	if (typeof value === "number" && Number.isFinite(value)) return String(value);
 }
@@ -11878,7 +12205,7 @@ function asNumber$3(value) {
 }
 function firstString$1(record, keys) {
 	for (const key of keys) {
-		const value = asString$15(record[key]);
+		const value = asString$16(record[key]);
 		if (value !== void 0) return value;
 	}
 }
@@ -11907,7 +12234,7 @@ function deepValue(value, keys, depth = 0) {
 		}
 		return;
 	}
-	if (!isRecord$25(value)) return void 0;
+	if (!isRecord$26(value)) return void 0;
 	for (const key of keys) if (value[key] !== void 0 && value[key] !== null) return value[key];
 	for (const nested of Object.values(value)) {
 		const hit = deepValue(nested, keys, depth + 1);
@@ -11921,7 +12248,7 @@ function collectRecords(value, keys, depth = 0, out = []) {
 		for (const entry of value) collectRecords(entry, keys, depth + 1, out);
 		return out;
 	}
-	if (!isRecord$25(value)) return out;
+	if (!isRecord$26(value)) return out;
 	if (keys.some((key) => value[key] !== void 0)) out.push(value);
 	for (const nested of Object.values(value)) collectRecords(nested, keys, depth + 1, out);
 	return out;
@@ -12125,7 +12452,7 @@ function parseCreditBalances(payload, consumed = []) {
 /** Resolve the plan id the subscription or credits payload reports, when either does. */
 function parsePlanId(payloads) {
 	for (const payload of payloads) {
-		const id = asString$15(deepValue(payload, [
+		const id = asString$16(deepValue(payload, [
 			"planId",
 			"plan_id",
 			"priceId",
@@ -12144,7 +12471,7 @@ function parsePlanId(payloads) {
 */
 function parseSubscriptionStatus(payload) {
 	const root = asRecord$9(payload) ?? {};
-	return asString$15(deepValue(asRecord$9(root.data) ?? root, ["status"])) ?? null;
+	return asString$16(deepValue(asRecord$9(root.data) ?? root, ["status"])) ?? null;
 }
 /** End of the current billing period, in Unix milliseconds. */
 function parseSubscriptionPeriodEnd(payload) {
@@ -12553,7 +12880,41 @@ function whoami(apiKey, options = {}) {
 async function verifyApiKey(apiKey, options = {}) {
 	return parseWhoami(await whoami(apiKey, options), { authenticatedAt: Date.now() });
 }
-let cachedQuotaEntry;
+/**
+* One newest snapshot PER ACCOUNT.
+*
+* Quota follows the account: rotation decides which key spends the next
+* request, so a single slot only ever described whichever key was read last.
+* The pool's card draws every account its own reading, so each account keeps
+* its own entry; past the cap the oldest read is evicted.
+*
+* This map is a memory read only. Nothing here fetches — the settings card
+* builds its account list from it, and that must never cost one upstream
+* request per pooled key.
+*/
+const cachedQuotaEntries$1 = /* @__PURE__ */ new Map();
+/** How many accounts' snapshots are remembered; mirrors the Codex usage service. */
+const QUOTA_CACHE_LIMIT$1 = 20;
+/** Key of a snapshot whose caller named no account. */
+const UNNAMED_ACCOUNT_KEY$1 = "";
+function quotaCacheKey(accountId) {
+	return accountId ?? UNNAMED_ACCOUNT_KEY$1;
+}
+/** The entry read most recently, whichever account it belongs to. */
+function newestQuotaEntry$2() {
+	let newest;
+	for (const entry of cachedQuotaEntries$1.values()) if (newest === void 0 || entry.fetchedAt > newest.fetchedAt) newest = entry;
+	return newest;
+}
+/** Remember one account's newest snapshot, bounded to the accounts in use. */
+function rememberQuotaEntry(entry) {
+	cachedQuotaEntries$1.set(quotaCacheKey(entry.accountId), entry);
+	while (cachedQuotaEntries$1.size > QUOTA_CACHE_LIMIT$1) {
+		const oldest = [...cachedQuotaEntries$1.entries()].sort((left, right) => left[1].fetchedAt - right[1].fetchedAt)[0];
+		if (oldest === void 0) break;
+		cachedQuotaEntries$1.delete(oldest[0]);
+	}
+}
 let quotaInFlight$2 = null;
 let quotaInFlightAccountId$1;
 let quotaCacheEpoch$1 = 0;
@@ -12562,21 +12923,22 @@ let quotaCacheEpoch$1 = 0;
 * Kept with a no-argument signature for backward compatibility.
 */
 function getCachedQuota$2() {
-	return cachedQuotaEntry?.quota;
+	return newestQuotaEntry$2()?.quota;
 }
 /**
 * The cached snapshot only when it belongs to this account.
 *
 * Passing no account keeps the pre-pool behavior of returning whatever was
-* fetched last.
+* fetched last. This is the lookup the account pool calls while it builds its
+* summaries: a pure memory read, never a fetch.
 */
-function getCachedQuotaFor$1(accountId) {
-	if (accountId === void 0 || accountId === null) return cachedQuotaEntry?.quota;
-	return cachedQuotaEntry?.accountId === accountId ? cachedQuotaEntry.quota : void 0;
+function getCachedQuotaFor$3(accountId) {
+	if (accountId === void 0 || accountId === null) return newestQuotaEntry$2()?.quota;
+	return cachedQuotaEntries$1.get(accountId)?.quota;
 }
 function clearCachedQuota$2() {
 	quotaCacheEpoch$1 += 1;
-	cachedQuotaEntry = void 0;
+	cachedQuotaEntries$1.clear();
 	quotaInFlight$2 = null;
 	quotaInFlightAccountId$1 = void 0;
 }
@@ -12591,8 +12953,8 @@ async function fetchAccountQuota$2(store = new FileCredentialStore$1(), fetchFnO
 	const fetchFn = typeof fetchFnOrOptions === "function" ? fetchFnOrOptions : fetchFnOrOptions?.fetchFn ?? fetch;
 	const effectiveForce = isOptionsObj ? fetchFnOrOptions.force ?? force : force;
 	const effectiveAccountId = isOptionsObj ? fetchFnOrOptions.accountId ?? accountId : accountId;
-	const sameAccount = effectiveAccountId === void 0 || cachedQuotaEntry?.accountId === effectiveAccountId;
-	if (!effectiveForce && sameAccount && cachedQuotaEntry && Date.now() - cachedQuotaEntry.fetchedAt < 12e4) return cachedQuotaEntry.quota;
+	const cached = getCachedQuotaFor$3(effectiveAccountId);
+	if (!effectiveForce && cached !== void 0 && Date.now() - (cached.fetchedAt || 0) < 12e4) return cached;
 	if (quotaInFlight$2 !== null && (effectiveAccountId === void 0 || quotaInFlightAccountId$1 === effectiveAccountId)) return quotaInFlight$2;
 	const epoch = quotaCacheEpoch$1;
 	const request = (async () => {
@@ -12648,11 +13010,11 @@ async function fetchAccountQuota$2(store = new FileCredentialStore$1(), fetchFnO
 			sources
 		};
 		if (epoch !== quotaCacheEpoch$1) return snapshot;
-		cachedQuotaEntry = {
+		rememberQuotaEntry({
 			accountId: effectiveAccountId,
 			quota: snapshot,
 			fetchedAt: snapshot.fetchedAt
-		};
+		});
 		return snapshot;
 	})();
 	quotaInFlight$2 = request;
@@ -12696,10 +13058,10 @@ function buildModelOptions$2(catalog, enabledModelIds, overrides) {
 *
 * https://commandcode.ai/blog/command-code-provider-api
 */
-function isRecord$24(value) {
+function isRecord$25(value) {
 	return typeof value === "object" && value !== null && !Array.isArray(value);
 }
-function asString$14(value) {
+function asString$15(value) {
 	return typeof value === "string" ? value : void 0;
 }
 function safeJsonParse$4(text) {
@@ -12736,17 +13098,17 @@ const SUPPORTED_IMAGE_MEDIA_TYPES$4 = /* @__PURE__ */ new Set([
 ]);
 function attachmentOf$4(block) {
 	const attachment = block.attachment;
-	if (!isRecord$24(attachment)) return void 0;
+	if (!isRecord$25(attachment)) return void 0;
 	return typeof attachment.attachmentId === "string" ? attachment : void 0;
 }
 function attachmentLabel$5(block) {
-	const attachment = isRecord$24(block.attachment) ? block.attachment : void 0;
-	return asString$14(attachment?.name) || asString$14(attachment?.attachmentId);
+	const attachment = isRecord$25(block.attachment) ? block.attachment : void 0;
+	return asString$15(attachment?.name) || asString$15(attachment?.attachmentId);
 }
 function collectImageRefs$3(content, refs) {
 	if (!Array.isArray(content)) return;
 	for (const block of content) {
-		if (!isRecord$24(block)) continue;
+		if (!isRecord$25(block)) continue;
 		if (block.type === "image") {
 			const attachment = attachmentOf$4(block);
 			if (attachment) refs.set(attachment.attachmentId, attachment);
@@ -12761,13 +13123,13 @@ function base64Length$3(bytes) {
 function requestImageBytes$3(block) {
 	const attachment = attachmentOf$4(block);
 	if (attachment) return base64Length$3(attachment.bytes);
-	const inline = asString$14(block.data) || asString$14(block.base64);
+	const inline = asString$15(block.data) || asString$15(block.base64);
 	return inline ? inline.length : void 0;
 }
 function collectRequestImageBytes$3(content, lengths) {
 	if (!Array.isArray(content)) return;
 	for (const block of content) {
-		if (!isRecord$24(block) || block.type !== "image") continue;
+		if (!isRecord$25(block) || block.type !== "image") continue;
 		const bytes = requestImageBytes$3(block);
 		if (bytes !== void 0) lengths.push(bytes);
 	}
@@ -12794,7 +13156,7 @@ function offloadOldestRequestImages$3(options) {
 		if (remaining.count === 0 || !Array.isArray(message.content)) return message;
 		let replaced = false;
 		const content = message.content.map((block) => {
-			if (remaining.count === 0 || !isRecord$24(block) || block.type !== "image") return block;
+			if (remaining.count === 0 || !isRecord$25(block) || block.type !== "image") return block;
 			if (requestImageBytes$3(block) === void 0) return block;
 			remaining.count -= 1;
 			replaced = true;
@@ -12847,10 +13209,10 @@ function unavailableImageText$5(block) {
 	return `[image unavailable: ${label ? `${label} could not be read` : "the image could not be read"}; ask the user to attach it again if the image is needed]`;
 }
 function imageBlockToInline$3(block, images) {
-	let data = asString$14(block.data) || asString$14(block.base64);
-	const source = isRecord$24(block.source) ? block.source : void 0;
-	if (!data && source) data = asString$14(source.data) || asString$14(source.base64);
-	let mediaType = asString$14(block.mimeType) || asString$14(block.mediaType) || (source ? asString$14(source.mimeType) || asString$14(source.mediaType) : void 0) || "image/png";
+	let data = asString$15(block.data) || asString$15(block.base64);
+	const source = isRecord$25(block.source) ? block.source : void 0;
+	if (!data && source) data = asString$15(source.data) || asString$15(source.base64);
+	let mediaType = asString$15(block.mimeType) || asString$15(block.mediaType) || (source ? asString$15(source.mimeType) || asString$15(source.mediaType) : void 0) || "image/png";
 	if (data?.startsWith("data:")) {
 		const matched = data.match(/^data:([^;,]+);base64,(.*)$/s);
 		if (matched) {
@@ -12874,7 +13236,7 @@ function textOf$5(content) {
 	if (!Array.isArray(content)) return "";
 	const parts = [];
 	for (const block of content) {
-		if (!isRecord$24(block)) continue;
+		if (!isRecord$25(block)) continue;
 		if (block.type === "text" && typeof block.text === "string") parts.push(sanitizeText$4(block.text));
 		else if (block.type === "tool-result") parts.push(textOf$5(block.content));
 	}
@@ -12883,7 +13245,7 @@ function textOf$5(content) {
 function toolResultText$4(blocks) {
 	if (!Array.isArray(blocks)) return "";
 	return blocks.map((block) => {
-		if (!isRecord$24(block)) return "";
+		if (!isRecord$25(block)) return "";
 		if (block.type === "text" && typeof block.text === "string") return sanitizeText$4(block.text);
 		if (block.type === "tool-result") return toolResultText$4(block.content);
 		if (block.type === "image") return `[image: ${attachmentLabel$5(block) ?? "attached image"}]`;
@@ -12901,7 +13263,7 @@ function toolResultBlocks$2(blocks, images) {
 	const out = [];
 	let hasImage = false;
 	for (const block of blocks) {
-		if (!isRecord$24(block)) continue;
+		if (!isRecord$25(block)) continue;
 		if (block.type === "text" && typeof block.text === "string") {
 			out.push({
 				type: "text",
@@ -12950,7 +13312,7 @@ function toolResultImageBlocks$2(blocks, images) {
 	if (!Array.isArray(blocks)) return [];
 	const out = [];
 	for (const block of blocks) {
-		if (!isRecord$24(block)) continue;
+		if (!isRecord$25(block)) continue;
 		if (block.type === "image") {
 			const inline = imageBlockToInline$3(block, images);
 			if (inline && SUPPORTED_IMAGE_MEDIA_TYPES$4.has(inline.mediaType)) out.push({
@@ -12994,7 +13356,7 @@ function nonSystemMessages$4(options) {
 }
 /** Drop the JSON-Schema keywords provider gateways reject or ignore. */
 function stripMetaSchema$2(schema) {
-	if (!isRecord$24(schema)) return {
+	if (!isRecord$25(schema)) return {
 		type: "object",
 		properties: {}
 	};
@@ -13011,7 +13373,7 @@ function openAIUserContent$2(message, images) {
 	const parts = [];
 	let hasImage = false;
 	for (const block of message.content) {
-		if (!isRecord$24(block)) continue;
+		if (!isRecord$25(block)) continue;
 		if (block.type === "text" && typeof block.text === "string") {
 			const text = sanitizeText$4(block.text);
 			if (text !== "") parts.push({
@@ -13039,7 +13401,7 @@ function openAIAssistantContent$2(message) {
 	const textParts = [];
 	const toolCalls = [];
 	for (const block of message.content) {
-		if (!isRecord$24(block)) continue;
+		if (!isRecord$25(block)) continue;
 		if (block.type === "text" && typeof block.text === "string") textParts.push(sanitizeText$4(block.text));
 		else if (block.type === "tool-call" && typeof block.name === "string") toolCalls.push({
 			id: typeof block.id === "string" && block.id !== "" ? block.id : `call_${toolCalls.length}`,
@@ -13071,7 +13433,7 @@ function buildOpenAIRequest$1(options, images = NO_RESOLVED_IMAGES$4) {
 			while (index < conversation.length && isToolResultMessage$4(conversation[index])) {
 				const current = conversation[index];
 				const block = current.content[0];
-				const callId = isRecord$24(block) && typeof block.toolCallId === "string" ? block.toolCallId : "";
+				const callId = isRecord$25(block) && typeof block.toolCallId === "string" ? block.toolCallId : "";
 				messages.push({
 					role: "tool",
 					tool_call_id: callId,
@@ -13132,7 +13494,7 @@ function anthropicUserContent$2(message, images) {
 	if (!Array.isArray(message.content)) return [];
 	const blocks = [];
 	for (const block of message.content) {
-		if (!isRecord$24(block)) continue;
+		if (!isRecord$25(block)) continue;
 		if (block.type === "text" && typeof block.text === "string") {
 			const text = sanitizeText$4(block.text);
 			if (text !== "") blocks.push({
@@ -13169,7 +13531,7 @@ function anthropicUserContent$2(message, images) {
 function anthropicAssistantContent$2(message) {
 	const blocks = [];
 	for (const block of message.content) {
-		if (!isRecord$24(block)) continue;
+		if (!isRecord$25(block)) continue;
 		if (block.type === "text" && typeof block.text === "string") {
 			const text = sanitizeText$4(block.text);
 			if (text !== "") blocks.push({
@@ -13182,7 +13544,7 @@ function anthropicAssistantContent$2(message) {
 				type: "tool_use",
 				id: typeof block.id === "string" && block.id !== "" ? block.id : `toolu_${blocks.length}`,
 				name: block.name,
-				input: isRecord$24(parsed) ? parsed : {}
+				input: isRecord$25(parsed) ? parsed : {}
 			});
 		}
 	}
@@ -13276,7 +13638,7 @@ function responsesUserContent(message, images) {
 	if (!Array.isArray(message.content)) return [];
 	const parts = [];
 	for (const block of message.content) {
-		if (!isRecord$24(block)) continue;
+		if (!isRecord$25(block)) continue;
 		if (block.type === "text" && typeof block.text === "string") {
 			const text = sanitizeText$4(block.text);
 			if (text !== "") parts.push({
@@ -13301,7 +13663,7 @@ function responsesToolResultImageBlocks(blocks, images) {
 	if (!Array.isArray(blocks)) return [];
 	const out = [];
 	for (const block of blocks) {
-		if (!isRecord$24(block)) continue;
+		if (!isRecord$25(block)) continue;
 		if (block.type === "image") {
 			const inline = imageBlockToInline$3(block, images);
 			if (inline && SUPPORTED_IMAGE_MEDIA_TYPES$4.has(inline.mediaType)) out.push({
@@ -13328,8 +13690,8 @@ function buildResponsesRequest(options, images = NO_RESOLVED_IMAGES$4) {
 			const imageBlocks = [];
 			while (index < conversation.length && isToolResultMessage$4(conversation[index])) {
 				const current = conversation[index];
-				const toolResult = Array.isArray(current.content) ? current.content.find((b) => isRecord$24(b) && b.type === "tool-result") : void 0;
-				const callId = (isRecord$24(toolResult) && typeof toolResult.toolCallId === "string" ? toolResult.toolCallId : "") || (typeof current.source?.callId === "string" ? current.source.callId : "");
+				const toolResult = Array.isArray(current.content) ? current.content.find((b) => isRecord$25(b) && b.type === "tool-result") : void 0;
+				const callId = (isRecord$25(toolResult) && typeof toolResult.toolCallId === "string" ? toolResult.toolCallId : "") || (typeof current.source?.callId === "string" ? current.source.callId : "");
 				input.push({
 					type: "function_call_output",
 					call_id: callId,
@@ -13352,7 +13714,7 @@ function buildResponsesRequest(options, images = NO_RESOLVED_IMAGES$4) {
 				content
 			});
 			for (const call of toolCalls) {
-				const fn = isRecord$24(call.function) ? call.function : void 0;
+				const fn = isRecord$25(call.function) ? call.function : void 0;
 				input.push({
 					type: "function_call",
 					call_id: call.id,
@@ -13475,25 +13837,33 @@ function processOpenAIStreamLine$1(line, state) {
 	}
 	if (payload === "") return [];
 	const chunk = safeJsonParse$4(payload);
-	if (!isRecord$24(chunk)) return [];
+	if (!isRecord$25(chunk)) return [];
 	const out = [];
-	if (isRecord$24(chunk.error)) throw new LlmError$1(`Command Code stream error: ${asString$14(chunk.error.message) ?? "unknown error"}`, isContextOverflow(chunk.error) ? CONTEXT_OVERFLOW_CODE : "PROVIDER_ERROR");
-	const usage = isRecord$24(chunk.usage) ? chunk.usage : void 0;
+	if (isRecord$25(chunk.error)) {
+		const message = asString$15(chunk.error.message) ?? "unknown error";
+		state.streamError = {
+			vocabulary: "openai",
+			error: chunk.error,
+			message
+		};
+		throw new LlmError$1(`Command Code stream error: ${message}`, isContextOverflow(chunk.error) ? CONTEXT_OVERFLOW_CODE : "PROVIDER_ERROR");
+	}
+	const usage = isRecord$25(chunk.usage) ? chunk.usage : void 0;
 	if (usage) {
 		state.sawUsage = true;
 		const prompt = numberOr$5(usage.prompt_tokens, 0);
-		const details = isRecord$24(usage.prompt_tokens_details) ? usage.prompt_tokens_details : void 0;
+		const details = isRecord$25(usage.prompt_tokens_details) ? usage.prompt_tokens_details : void 0;
 		const cached = details ? numberOr$5(details.cached_tokens, 0) : 0;
 		state.inputTokens = Math.max(0, prompt - cached);
 		state.cacheReadTokens = cached;
 		state.outputTokens = numberOr$5(usage.completion_tokens, state.outputTokens);
-		if (isRecord$24(usage.completion_tokens_details)) state.reasoningTokens = numberOr$5(usage.completion_tokens_details.reasoning_tokens, state.reasoningTokens);
+		if (isRecord$25(usage.completion_tokens_details)) state.reasoningTokens = numberOr$5(usage.completion_tokens_details.reasoning_tokens, state.reasoningTokens);
 	}
 	const choices = Array.isArray(chunk.choices) ? chunk.choices : [];
-	const choice = isRecord$24(choices[0]) ? choices[0] : void 0;
-	const delta = isRecord$24(choice?.delta) ? choice.delta : void 0;
+	const choice = isRecord$25(choices[0]) ? choices[0] : void 0;
+	const delta = isRecord$25(choice?.delta) ? choice.delta : void 0;
 	if (delta) {
-		const reasoning = asString$14(delta.reasoning_content) ?? asString$14(delta.reasoning);
+		const reasoning = asString$15(delta.reasoning_content) ?? asString$15(delta.reasoning);
 		if (reasoning !== void 0 && reasoning !== "") {
 			out.push(...closeToolCalls$4(state));
 			if (state.current === null || state.current.type !== "reasoning") out.push(...openTextBlock$3(state, "reasoning"));
@@ -13505,7 +13875,7 @@ function processOpenAIStreamLine$1(line, state) {
 				text: sanitizeText$4(reasoning)
 			});
 		}
-		const content = asString$14(delta.content);
+		const content = asString$15(delta.content);
 		if (content !== void 0 && content !== "") {
 			out.push(...closeToolCalls$4(state));
 			if (state.current === null || state.current.type !== "text") out.push(...openTextBlock$3(state, "text"));
@@ -13519,11 +13889,11 @@ function processOpenAIStreamLine$1(line, state) {
 		}
 		const toolDeltas = Array.isArray(delta.tool_calls) ? delta.tool_calls : [];
 		for (const entry of toolDeltas) {
-			if (!isRecord$24(entry)) continue;
+			if (!isRecord$25(entry)) continue;
 			out.push(...applyOpenAIToolDelta$1(entry, state));
 		}
 	}
-	const finish = asString$14(choice?.finish_reason);
+	const finish = asString$15(choice?.finish_reason);
 	if (finish !== void 0 && finish !== "") {
 		state.finishReason = finish;
 		out.push(...closeCurrent$3(state));
@@ -13533,15 +13903,15 @@ function processOpenAIStreamLine$1(line, state) {
 }
 function applyOpenAIToolDelta$1(entry, state) {
 	const wireIndex = typeof entry.index === "number" ? entry.index : 0;
-	const fn = isRecord$24(entry.function) ? entry.function : {};
+	const fn = isRecord$25(entry.function) ? entry.function : {};
 	const out = [];
 	let call = state.toolCalls.get(wireIndex);
 	if (call === void 0) {
 		out.push(...closeCurrent$3(state));
 		call = {
 			blockIndex: state.blocks.length,
-			id: asString$14(entry.id) ?? `call_${wireIndex}`,
-			name: asString$14(fn.name) ?? "",
+			id: asString$15(entry.id) ?? `call_${wireIndex}`,
+			name: asString$15(fn.name) ?? "",
 			arguments: "",
 			started: false
 		};
@@ -13554,13 +13924,13 @@ function applyOpenAIToolDelta$1(entry, state) {
 		state.toolCalls.set(wireIndex, call);
 	} else {
 		if (call.id === `call_${wireIndex}`) {
-			const id = asString$14(entry.id);
+			const id = asString$15(entry.id);
 			if (id !== void 0) call.id = id;
 		}
-		const name = asString$14(fn.name);
+		const name = asString$15(fn.name);
 		if (name !== void 0 && name !== "") call.name = name;
 	}
-	const argsDelta = asString$14(fn.arguments) ?? "";
+	const argsDelta = asString$15(fn.arguments) ?? "";
 	if (argsDelta !== "") call.arguments += argsDelta;
 	if (!call.started) {
 		call.started = true;
@@ -13591,12 +13961,12 @@ function processAnthropicStreamLine$1(line, state) {
 	const payload = trimmed.slice(5).trim();
 	if (payload === "" || payload === "[DONE]") return [];
 	const event = safeJsonParse$4(payload);
-	if (!isRecord$24(event)) return [];
-	const type = asString$14(event.type);
+	if (!isRecord$25(event)) return [];
+	const type = asString$15(event.type);
 	const out = [];
 	if (type === "message_start") {
-		const message = isRecord$24(event.message) ? event.message : void 0;
-		const usage = message && isRecord$24(message.usage) ? message.usage : void 0;
+		const message = isRecord$25(event.message) ? event.message : void 0;
+		const usage = message && isRecord$25(message.usage) ? message.usage : void 0;
 		if (usage) {
 			state.sawUsage = true;
 			state.inputTokens = numberOr$5(usage.input_tokens, 0);
@@ -13604,21 +13974,21 @@ function processAnthropicStreamLine$1(line, state) {
 			state.cacheWriteTokens = numberOr$5(usage.cache_creation_input_tokens, 0);
 			state.outputTokens = numberOr$5(usage.output_tokens, 0);
 		}
-		const stop = message ? asString$14(message.stop_reason) : void 0;
+		const stop = message ? asString$15(message.stop_reason) : void 0;
 		if (stop !== void 0 && stop !== null) state.finishReason = stop;
 		return out;
 	}
 	if (type === "content_block_start") {
 		const contentIndex = numberOr$5(event.index, 0);
-		const block = isRecord$24(event.content_block) ? event.content_block : {};
-		const blockType = asString$14(block.type);
+		const block = isRecord$25(event.content_block) ? event.content_block : {};
+		const blockType = asString$15(block.type);
 		out.push(...closeCurrent$3(state));
 		if (blockType === "tool_use") {
 			const index = state.blocks.length;
 			const pending = {
 				blockIndex: index,
-				id: asString$14(block.id) ?? `toolu_${contentIndex}`,
-				name: asString$14(block.name) ?? "",
+				id: asString$15(block.id) ?? `toolu_${contentIndex}`,
+				name: asString$15(block.name) ?? "",
 				arguments: "",
 				started: true
 			};
@@ -13660,11 +14030,11 @@ function processAnthropicStreamLine$1(line, state) {
 	}
 	if (type === "content_block_delta") {
 		const contentIndex = numberOr$5(event.index, 0);
-		const delta = isRecord$24(event.delta) ? event.delta : {};
-		const deltaType = asString$14(delta.type);
+		const delta = isRecord$25(event.delta) ? event.delta : {};
+		const deltaType = asString$15(delta.type);
 		if (deltaType === "input_json_delta") {
 			const pending = state.toolCalls.get(contentIndex);
-			const partial = asString$14(delta.partial_json) ?? "";
+			const partial = asString$15(delta.partial_json) ?? "";
 			if (pending !== void 0) {
 				pending.arguments += partial;
 				out.push({
@@ -13677,7 +14047,7 @@ function processAnthropicStreamLine$1(line, state) {
 			}
 			return out;
 		}
-		const text = deltaType === "thinking_delta" ? asString$14(delta.thinking) : asString$14(delta.text);
+		const text = deltaType === "thinking_delta" ? asString$15(delta.thinking) : asString$15(delta.text);
 		if (text !== void 0 && text !== "") {
 			const index = state.contentIndexes.get(contentIndex) ?? state.current?.index;
 			const kind = deltaType === "thinking_delta" ? "reasoning" : "text";
@@ -13733,10 +14103,10 @@ function processAnthropicStreamLine$1(line, state) {
 		return out;
 	}
 	if (type === "message_delta") {
-		const delta = isRecord$24(event.delta) ? event.delta : void 0;
-		const stop = delta ? asString$14(delta.stop_reason) : void 0;
+		const delta = isRecord$25(event.delta) ? event.delta : void 0;
+		const stop = delta ? asString$15(delta.stop_reason) : void 0;
 		if (stop !== void 0 && stop !== "") state.finishReason = stop;
-		const usage = isRecord$24(event.usage) ? event.usage : void 0;
+		const usage = isRecord$25(event.usage) ? event.usage : void 0;
 		if (usage) {
 			state.sawUsage = true;
 			state.outputTokens = numberOr$5(usage.output_tokens, state.outputTokens);
@@ -13748,8 +14118,14 @@ function processAnthropicStreamLine$1(line, state) {
 		return closeStream$4(state);
 	}
 	if (type === "error") {
-		const error = isRecord$24(event.error) ? event.error : {};
-		throw new LlmError$1(`Command Code stream error: ${asString$14(error.message) ?? "unknown error"}`, isContextOverflow(error) ? CONTEXT_OVERFLOW_CODE : "PROVIDER_ERROR");
+		const error = isRecord$25(event.error) ? event.error : {};
+		const message = asString$15(error.message) ?? "unknown error";
+		state.streamError = {
+			vocabulary: "anthropic",
+			error,
+			message
+		};
+		throw new LlmError$1(`Command Code stream error: ${message}`, isContextOverflow(error) ? CONTEXT_OVERFLOW_CODE : "PROVIDER_ERROR");
 	}
 	return out;
 }
@@ -13788,7 +14164,7 @@ function processResponsesStreamLine(line, state) {
 	} catch {
 		return [];
 	}
-	if (!isRecord$24(parsed)) return [];
+	if (!isRecord$25(parsed)) return [];
 	const type = typeof parsed.type === "string" ? parsed.type : "";
 	if (type === "response.output_text.delta") {
 		const delta = typeof parsed.delta === "string" ? parsed.delta : "";
@@ -13815,7 +14191,7 @@ function processResponsesStreamLine(line, state) {
 		}];
 	}
 	if (type === "response.output_item.added") {
-		const item = isRecord$24(parsed.item) ? parsed.item : null;
+		const item = isRecord$25(parsed.item) ? parsed.item : null;
 		if (item === null || item.type !== "function_call") return [];
 		return startResponsesToolCall(state, parsed, item);
 	}
@@ -13839,19 +14215,24 @@ function processResponsesStreamLine(line, state) {
 		return [];
 	}
 	if (type === "response.completed" || type === "response.incomplete") {
-		const response = isRecord$24(parsed.response) ? parsed.response : null;
+		const response = isRecord$25(parsed.response) ? parsed.response : null;
 		if (response !== null) {
 			if (response.status === "failed") {
 				state.done = true;
-				const errorObj = isRecord$24(response.error) ? response.error : null;
+				const errorObj = isRecord$25(response.error) ? response.error : null;
 				const detail = errorObj && typeof errorObj.message === "string" ? errorObj.message : typeof response.error === "string" ? response.error : "";
+				state.streamError = {
+					vocabulary: "responses",
+					error: errorObj,
+					message: detail
+				};
 				throw new LlmError$1(detail === "" ? "Command Code responses stream failed" : `Command Code responses stream failed: ${detail}`, "PROVIDER_ERROR");
 			}
 			readResponsesUsage(response, state);
 		}
 		state.done = true;
 		if (type === "response.incomplete") {
-			const details = response && isRecord$24(response.incomplete_details) ? response.incomplete_details : null;
+			const details = response && isRecord$25(response.incomplete_details) ? response.incomplete_details : null;
 			const reason = details && typeof details.reason === "string" ? details.reason : "";
 			state.finishReason = reason === "max_output_tokens" || reason === "length" || reason === "max_tokens" ? "length" : reason || "length";
 		} else state.finishReason = "stop";
@@ -13859,8 +14240,13 @@ function processResponsesStreamLine(line, state) {
 	}
 	if (type === "response.failed" || type === "error") {
 		state.done = true;
-		const errorObj = isRecord$24(parsed.error) ? parsed.error : isRecord$24(parsed.response) && isRecord$24(parsed.response.error) ? parsed.response.error : null;
-		const detail = (typeof parsed.message === "string" ? parsed.message : "") || (errorObj && typeof errorObj.message === "string" ? errorObj.message : "") || (isRecord$24(parsed.response) && typeof parsed.response.error === "string" ? parsed.response.error : typeof parsed.error === "string" ? parsed.error : "");
+		const errorObj = isRecord$25(parsed.error) ? parsed.error : isRecord$25(parsed.response) && isRecord$25(parsed.response.error) ? parsed.response.error : null;
+		const detail = (typeof parsed.message === "string" ? parsed.message : "") || (errorObj && typeof errorObj.message === "string" ? errorObj.message : "") || (isRecord$25(parsed.response) && typeof parsed.response.error === "string" ? parsed.response.error : typeof parsed.error === "string" ? parsed.error : "");
+		state.streamError = {
+			vocabulary: "responses",
+			error: errorObj,
+			message: detail
+		};
 		throw new LlmError$1(detail === "" ? "Command Code responses stream failed" : `Command Code responses stream failed: ${detail}`, isContextOverflow(errorObj ?? detail) ? CONTEXT_OVERFLOW_CODE : "PROVIDER_ERROR");
 	}
 	return [];
@@ -13943,15 +14329,15 @@ function responsesToolCall(state, key) {
 	for (const call of state.toolCalls.values()) if (call.id === toToolCallId(key)) return call;
 }
 function readResponsesUsage(response, state) {
-	const usage = isRecord$24(response.usage) ? response.usage : null;
+	const usage = isRecord$25(response.usage) ? response.usage : null;
 	if (usage === null) return;
-	const cached = isRecord$24(usage.input_tokens_details) ? numberOrZero(usage.input_tokens_details.cached_tokens) : 0;
-	const written = isRecord$24(usage.input_tokens_details) ? numberOrZero(usage.input_tokens_details.cache_write_tokens) : 0;
+	const cached = isRecord$25(usage.input_tokens_details) ? numberOrZero(usage.input_tokens_details.cached_tokens) : 0;
+	const written = isRecord$25(usage.input_tokens_details) ? numberOrZero(usage.input_tokens_details.cache_write_tokens) : 0;
 	state.inputTokens = Math.max(0, numberOrZero(usage.input_tokens) - cached - written);
 	state.cacheReadTokens = cached;
 	state.cacheWriteTokens = written;
 	state.outputTokens = numberOrZero(usage.output_tokens);
-	const outputDetails = isRecord$24(usage.output_tokens_details) ? usage.output_tokens_details : null;
+	const outputDetails = isRecord$25(usage.output_tokens_details) ? usage.output_tokens_details : null;
 	if (outputDetails !== null) state.reasoningTokens = numberOrZero(outputDetails.reasoning_tokens);
 	state.sawUsage = true;
 }
@@ -13978,6 +14364,136 @@ function assertStreamComplete$4(state) {
 	if (!state.done && state.finishReason === null) throw new LlmError$1("Command Code stream ended before its terminal event", "PROVIDER_ERROR");
 }
 //#endregion
+//#region src/host/common/stream-error.ts
+/**
+* A failure the provider reported INSIDE a 200 stream.
+*
+* A service can answer 200 and then report the failure in the stream, in the
+* same `{ error: { type, message } }` envelope a non-2xx body carries. A mapper
+* has to type that event with one code, and the code that is right in general is
+* PROVIDER_ERROR - it cannot know whether output has started - which is outside
+* every retry set. So a transient failure delivered this way ends the turn on
+* the first try while the identical failure delivered as a status is retried.
+*
+* The reclassification itself is per line, because only the adapter knows
+* whether anything has reached the caller yet, and because every line owns rules
+* its own HTTP path already owns. What is shared is the question each line asks
+* its own classifier: what status WOULD this failure have carried?
+*/
+/**
+* The status an in-band `error` event stands for, or null for a type this
+* vocabulary does not name.
+*
+* The Anthropic-compatible shims (MiniMax Code, Kimi Code, and Command Code's
+* messages stream) speak one vocabulary, and each of these types already has a
+* status: 529 is Anthropic's documented overload, 429 the rate limit, and a 5xx
+* the server-side failure. Naming the STATUS rather than the verdict is what lets
+* each line hand the answer to its own classifier, so a rule that line already
+* owns - a 429 that names an exhausted plan, a credential refusal reported as
+* 403 - keeps owning it here.
+*
+* Null for everything else: a request the provider rejected, a context
+* overflow, a credential refusal, a type from the future. Null is what keeps
+* the mapper's own verdict, and that is the safe direction to fail in - a
+* fallback to some made-up status would file an unknown type as a retryable
+* server error, which is worse than not retrying at all.
+*/
+function inBandAnthropicStatus(error) {
+	const type = readErrorType(error);
+	if (type === null) return null;
+	return ANTHROPIC_IN_BAND_STATUS[type] ?? null;
+}
+/** The status each transient Anthropic-compatible in-band type stands for. */
+const ANTHROPIC_IN_BAND_STATUS = {
+	overloaded_error: 529,
+	rate_limit_error: 429,
+	api_error: 500
+};
+/**
+* The DSH code an in-band Responses-vocabulary failure becomes, or null when it
+* names nothing transient.
+*
+* The Responses stream reports failure as `response.failed` or `error`, whose
+* message is free text and whose structured evidence is split across TWO fields:
+* this plugin's own routes document `{"error":{"message":"Upstream model
+* provider is temporarily unavailable. Please try again in a moment.","type":
+* "server_error"}}` - `type`, not `code` - so reading only `code` would miss
+* the exact body this exists for, and its message names nothing the heuristic
+* would catch. Both fields are read, and the message is read too, because a
+* deployment that says "overloaded" in prose is the whole signal there.
+*
+* Anything not recognized keeps the mapper's verdict.
+*/
+function inBandResponsesCode(error, message) {
+	const fields = isRecord$24(error) ? error : {};
+	const rawCode = (asString$14(fields.code) ?? asString$14(fields.type) ?? "").toLowerCase();
+	const text = message.toLowerCase();
+	if (text.includes("rate limit") || rawCode === "rate_limit" || rawCode === "rate_limit_exceeded") return "RATE_LIMIT";
+	if (text.includes("overload") || text.includes("server error") || rawCode === "server_error" || rawCode === "service_unavailable" || rawCode === "internal_error") return "SERVER";
+	return null;
+}
+/**
+* The error one in-band stream failure becomes, or `thrown` itself when the
+* line's own verdict stands.
+*
+* Call this only while nothing has reached the caller: the whole reason a fresh
+* request is still free is what makes the retry safe (see module note 1 in the
+* Claude adapter, which this generalizes). After the first chunk the mapper's
+* verdict stands and this must not be called at all.
+*
+* Only the CODE changes. The message stays the mapper's, which names the wire
+* type the user saw, and no `status` is attached: the response really was a
+* 200, and a synthetic one would misreport what the provider said.
+*
+* @param thrown - the mapper's own verdict, passed through when it stands.
+* @param rawError - the event's wire `error` object, from the stream state.
+* @param classify - this line's own failure classifier, so every rule its HTTP
+*   path owns still owns the verdict reached here.
+* @param statusFor - the vocabulary's status table; Responses lines pass
+*   {@link inBandResponsesCode} through {@link reclassifyInBandResponsesError}.
+*/
+function reclassifyInBandError(thrown, rawError, classify, statusFor = inBandAnthropicStatus) {
+	const status = statusFor(rawError);
+	if (status === null) return thrown;
+	const failure = classify(status, JSON.stringify({ error: rawError }));
+	if (!failure.retryable) return thrown;
+	return new LlmError(thrown.message, failure.code, { cause: thrown });
+}
+/**
+* {@link reclassifyInBandError} for a line whose classifier reads a status, fed
+* by the Responses vocabulary instead of the Anthropic one.
+*
+* The transient verdict found here is a QUESTION, not the answer: a
+* `rate_limit` code says only that the failure is worth retrying, and whether
+* it actually is - a 429 naming a spent balance is this route's PROVIDER_ERROR,
+* not a RATE_LIMIT - is a rule the line's own classifier owns. So the answer is
+* asked of `classify` at the status the verdict stands for, exactly as the
+* Anthropic sibling does, and the CODE that comes back is the one the HTTP path
+* would have produced for the same body. That agreement is the whole point: the
+* in-band delivery must not be the one delivery with its own opinion.
+*/
+function reclassifyInBandResponsesError(thrown, rawError, message, classify) {
+	const code = inBandResponsesCode(rawError, message);
+	if (code === null) return thrown;
+	const failure = classify(code === "RATE_LIMIT" ? 429 : 500, JSON.stringify({
+		error: rawError,
+		message
+	}));
+	if (!failure.retryable) return thrown;
+	return new LlmError(thrown.message, failure.code, { cause: thrown });
+}
+/** The wire `error` object of an in-band event, read defensively. */
+function readErrorType(error) {
+	if (!isRecord$24(error)) return null;
+	return typeof error.type === "string" ? error.type : null;
+}
+function asString$14(value) {
+	return typeof value === "string" ? value : void 0;
+}
+function isRecord$24(value) {
+	return typeof value === "object" && value !== null && !Array.isArray(value);
+}
+//#endregion
 //#region src/host/common/capabilities.ts
 /** Error thrown by requireCapability when a capability contract fails. */
 var CapabilityError = class extends Error {
@@ -14253,6 +14769,35 @@ function isCommandCodeCredentialInvalid(status, detail) {
 	}
 	return false;
 }
+/**
+* The verdict one failed response becomes on this route.
+*
+* An in-band `error` event is a status the provider chose not to send as a
+* status, so the chain that answers a non-2xx body has to answer it too rather
+* than a second, looser one: every rule this route already owns - a model the
+* plan does not include, a rejected key, a spent quota - keeps owning the
+* answer, and the codes below are the ONLY place the HTTP path reads them from.
+*
+* `retryable` is membership of this route's own policy rather than a second
+* judgement, because a code is worth repeating exactly when RETRY_POLICY repeats
+* it; a verdict the policy would not act on must not be filed as transient.
+*
+* @param status - the response status, or the status an in-band type stands for.
+* @param detail - the response body, or the serialized in-band error envelope.
+*/
+function classifyCommandCodeFailure(status, detail) {
+	let code;
+	if (status === 422) code = "PROVIDER_ERROR";
+	else if (isCommandCodeModelAccessDenied(status, detail)) code = "PROVIDER_ERROR";
+	else if (isCommandCodeCredentialInvalid(status, detail) || status === 401) code = "INVALID_CREDENTIAL";
+	else if (status === 429) code = "RATE_LIMIT";
+	else if (status >= 500) code = "SERVER";
+	else code = isHttpContextOverflow(status, detail) ? CONTEXT_OVERFLOW_CODE : "PROVIDER_ERROR";
+	return {
+		code,
+		retryable: RETRY_POLICY$5.mode === "normal" && RETRY_POLICY$5.retryableCodes.includes(code)
+	};
+}
 /** Cooldown one rate-limited key takes when the provider states no delay. */
 const POOL_COOLDOWN_MS$4 = 15 * 6e4;
 var CommandCodeAdapter = class extends LlmAdapter {
@@ -14447,49 +14992,105 @@ var CommandCodeAdapter = class extends LlmAdapter {
 		}
 		if (response === void 0 || !response.ok) {
 			const status = response?.status ?? 500;
+			const { code } = classifyCommandCodeFailure(status, detail);
 			if (status === 422) {
-				if (detail.includes("cmd_zdr_no_providers")) throw new LlmError(`${PROVIDER_NAME$5} rejected request under Zero Data Retention: no ZDR-capable upstream is available for this model (${detail || "cmd_zdr_no_providers"}).`, "PROVIDER_ERROR", { status: 422 });
-				throw new LlmError(`${PROVIDER_NAME$5} validation error (422): ${detail || "Unprocessable Entity"}`, "PROVIDER_ERROR", { status: 422 });
+				if (detail.includes("cmd_zdr_no_providers")) throw new LlmError(`${PROVIDER_NAME$5} rejected request under Zero Data Retention: no ZDR-capable upstream is available for this model (${detail || "cmd_zdr_no_providers"}).`, code, { status: 422 });
+				throw new LlmError(`${PROVIDER_NAME$5} validation error (422): ${detail || "Unprocessable Entity"}`, code, { status: 422 });
 			}
-			if (isCommandCodeModelAccessDenied(status, detail)) throw new LlmError(`${PROVIDER_NAME$5} access denied for model ${options.model}: this model is not included in the plan or requires higher entitlement (${status}).${detail ? ` ${detail}` : ""}`, "PROVIDER_ERROR", { status });
-			if (isCommandCodeCredentialInvalid(status, detail) || status === 401) throw new LlmError(`${PROVIDER_NAME$5} rejected the stored API key (${status}). Sign in again from Settings > Command Code.${detail ? ` ${detail}` : ""}`, "INVALID_CREDENTIAL", { status });
+			if (isCommandCodeModelAccessDenied(status, detail)) throw new LlmError(`${PROVIDER_NAME$5} access denied for model ${options.model}: this model is not included in the plan or requires higher entitlement (${status}).${detail ? ` ${detail}` : ""}`, code, { status });
+			if (isCommandCodeCredentialInvalid(status, detail) || status === 401) throw new LlmError(`${PROVIDER_NAME$5} rejected the stored API key (${status}). Sign in again from Settings > Command Code.${detail ? ` ${detail}` : ""}`, code, { status });
 			if (status === 429) {
 				const after = response === void 0 ? void 0 : retryAfterMs$1(response.headers);
-				throw new LlmError(`${PROVIDER_NAME$5} rate limit or plan quota reached (429). Check the quota card in Settings > Command Code.${detail ? ` ${detail}` : ""}`, "RATE_LIMIT", {
+				throw new LlmError(`${PROVIDER_NAME$5} rate limit or plan quota reached (429). Check the quota card in Settings > Command Code.${detail ? ` ${detail}` : ""}`, code, {
 					status: 429,
 					...after === void 0 ? {} : { providerRetryAfterMs: after }
 				});
 			}
-			if (status >= 500) throw new LlmError(`${PROVIDER_NAME$5} upstream server error (${status}): ${detail || "No response"}`, "SERVER", { status });
-			throw new LlmError(`${PROVIDER_NAME$5} API error (${status}): ${detail || "No response"}`, isHttpContextOverflow(status, detail) ? CONTEXT_OVERFLOW_CODE : "PROVIDER_ERROR", { status });
+			if (status >= 500) throw new LlmError(`${PROVIDER_NAME$5} upstream server error (${status}): ${detail || "No response"}`, code, { status });
+			throw new LlmError(`${PROVIDER_NAME$5} API error (${status}): ${detail || "No response"}`, code, { status });
 		}
 		if (response.body === null) throw new LlmError("Command Code returned an empty response body", "PROVIDER_ERROR");
 		const reader = response.body.getReader();
 		const decoder = new TextDecoder();
 		const state = createStreamState$5(wire);
 		let buffer = "";
+		/**
+		* Set the moment ANY chunk reaches the caller, and never cleared.
+		*
+		* Deliberately broader than `a text delta or a tool call`: a retry would repeat
+		* a block-start the caller has already seen just as surely, and could re-issue
+		* a tool call the agent has already run, so the conservative reading is the
+		* one that cannot be wrong. It is what keeps the in-band reclassification
+		* below pre-output only, and before that point a fresh request is still free.
+		*/
+		let outputStarted = false;
 		try {
-			while (true) {
-				const { done, value } = await reader.read();
-				if (done) break;
-				buffer += decoder.decode(value, { stream: true });
-				const lines = buffer.split("\n");
-				buffer = lines.pop() ?? "";
-				for (const line of lines) {
-					for (const chunk of processLine$1(line, state, wire)) yield chunk;
-					if (state.finished) return;
+			try {
+				while (true) {
+					const { done, value } = await reader.read();
+					if (done) break;
+					buffer += decoder.decode(value, { stream: true });
+					const lines = buffer.split("\n");
+					buffer = lines.pop() ?? "";
+					for (const line of lines) {
+						for (const chunk of processLine$1(line, state, wire)) {
+							outputStarted = true;
+							yield chunk;
+						}
+						if (state.finished) return;
+					}
+				}
+				buffer += decoder.decode();
+				if (buffer.trim() !== "") for (const line of buffer.split("\n")) for (const chunk of processLine$1(line, state, wire)) {
+					outputStarted = true;
+					yield chunk;
+				}
+				if (state.finished) return;
+				assertStreamComplete$4(state);
+				for (const chunk of closeStream$4(state)) {
+					outputStarted = true;
+					yield chunk;
 				}
+			} catch (error) {
+				const inBand = state.streamError;
+				if (error instanceof LlmError && !outputStarted && inBand !== void 0 && error.code === "PROVIDER_ERROR") {
+					const reclassified = reclassifyInBandStreamError(error, inBand);
+					if (reclassified.code === "RATE_LIMIT" && pool !== null && accountId !== void 0) {
+						const after = retryAfterMs$1(response.headers);
+						await pool.markCooldown(accountId, after ?? POOL_COOLDOWN_MS$4, `${PROVIDER_NAME$5} in-stream 429`).catch(() => void 0);
+					}
+					throw reclassified;
+				}
+				throw error;
 			}
-			buffer += decoder.decode();
-			if (buffer.trim() !== "") for (const line of buffer.split("\n")) for (const chunk of processLine$1(line, state, wire)) yield chunk;
-			if (state.finished) return;
-			assertStreamComplete$4(state);
-			for (const chunk of closeStream$4(state)) yield chunk;
 		} finally {
 			reader.cancel().catch(() => void 0);
 		}
 	}
 };
+/**
+* The error one recorded in-band failure becomes, or the mapper's own verdict.
+*
+* Two rules, because this line speaks three in-band vocabularies and they do
+* not carry the same evidence. The messages wire names a transient failure with
+* a `type` that STANDS FOR A STATUS - an `overloaded_error` is the 529 it would
+* have sent - so its envelope is answered by the status table. The
+* chat-completions and Responses wires have no such type: their evidence is a
+* `code`, a `type` in the looser `server_error` sense, or plain prose, and the
+* status-reading helper is what turns that into the question this line's own
+* classifier answers.
+*
+* The switch is exhaustive on purpose. A fourth vocabulary must be a compile
+* error here rather than a silent fall into the rule above, which is the one
+* shape of this bug that cannot be caught by a test.
+*/
+function reclassifyInBandStreamError(thrown, inBand) {
+	switch (inBand.vocabulary) {
+		case "anthropic": return reclassifyInBandError(thrown, inBand.error, classifyCommandCodeFailure);
+		case "openai":
+		case "responses": return reclassifyInBandResponsesError(thrown, inBand.error, inBand.message, classifyCommandCodeFailure);
+	}
+}
 /** Path suffix the chosen route answers on. */
 function endpointPathFor(wire) {
 	if (wire === "anthropic") return "/messages";
@@ -14503,6 +15104,34 @@ function processLine$1(line, state, wire) {
 }
 //#endregion
 //#region src/host/command-code/account-pool.ts
+/**
+* One cached snapshot as the shared account card renders it.
+*
+* A usage window states its own consumed share, so it maps straight across; a
+* window whose share nobody measured is dropped rather than drawn at 0, which
+* would claim an unspent window no reading supports.
+*
+* Exported for the pool's regression tests; the card only reaches it through
+* the summary hook.
+*/
+function commandCodePoolQuota(cached) {
+	return poolQuota(cached.fetchedAt, cached.windows.map((window) => quotaWindow$1(window.label, window.usedPercent, {
+		windowDurationMins: window.windowDurationMins,
+		resetsAt: window.resetsAt
+	})));
+}
+/**
+* This key's own newest snapshot, ready for the account card.
+*
+* Quota follows the account, so the pool reports what the line already
+* remembered for THIS account id. The lookup is a pure memory read: the card
+* builds its account list here, and that must never cost one upstream request
+* per pooled key.
+*/
+function quotaFor$1(accountId) {
+	const cached = getCachedQuotaFor$3(accountId);
+	return cached === void 0 ? void 0 : commandCodePoolQuota(cached);
+}
 /** Encrypted pool file this line owns; the pre-pool credential file stays the mirror. */
 function commandCodePoolPath() {
 	return path.join(dshHomeDir(), "storages", "command-code-pool.json");
@@ -14640,7 +15269,11 @@ var CommandCodeAccountPool = class extends AccountPoolCore {
 					...account.userName === void 0 ? {} : { userName: account.userName },
 					...account.organizationName === void 0 ? {} : { organizationName: account.organizationName },
 					...account.planId === void 0 ? {} : { planId: account.planId },
-					...account.userId === void 0 ? {} : { userId: account.userId }
+					...account.userId === void 0 ? {} : { userId: account.userId },
+					...(() => {
+						const quota = quotaFor$1(account.id);
+						return quota === void 0 ? {} : { quota };
+					})()
 				};
 			},
 			emptyMessage: "Not signed in to Command Code. 请在「设置 → 订阅服务 → Command Code」中添加账号或粘贴 API Key。",
@@ -15195,7 +15828,7 @@ async function getCommandCodeWebStatus(store, modelSettings, preferences, option
 	const catalog = live.length > 0 ? live : fallbackCatalog$1();
 	const enabled = settings.enabled !== false;
 	const models = buildModelOptions$2(catalog, resolveEnabledModelIds$4(settings.enabledModelIds, catalog, enabled), settings.contextWindowOverrides);
-	const quota = getCachedQuotaFor$1(activeAccount?.id ?? null);
+	const quota = getCachedQuotaFor$3(activeAccount?.id ?? null);
 	const activeAccountView = activePoolAccount === void 0 ? null : {
 		userId: activePoolAccount.userId ?? activePoolAccount.credentials.userId ?? null,
 		userName: activePoolAccount.userName ?? activePoolAccount.credentials.userName ?? null,
@@ -15268,7 +15901,7 @@ function registerCommandCodeRoutes(ctx, store, modelSettings, preferences, optio
 					if (method !== "GET") return sendMethodNotAllowed$5(response);
 					const credentials = await store.read();
 					const targetId = await activeAccountId();
-					const cached = getCachedQuotaFor$1(targetId);
+					const cached = getCachedQuotaFor$3(targetId);
 					if ((credentials !== null || targetId !== void 0) && (cached === void 0 || Date.now() - (cached.fetchedAt || 0) > 12e4)) {
 						const refresh = () => fetchAccountQuota$2(quotaStore, fetchFn, false, targetId);
 						if (cached === void 0) await quotaRefresh.run(refresh);
@@ -15541,6 +16174,24 @@ const DEFAULT_CONTEXT_WINDOW$4 = 128e3;
 * ceiling rather than a claim about what the model can do.
 */
 const DEFAULT_MAX_OUTPUT_TOKENS = 8192;
+/**
+* Raise a cap that is too small for this model's forced thinking.
+*
+* Left alone: a model whose thinking CAN be disabled - its caller may have asked
+* for thinking off on purpose, and raising the cap would not enable it while the
+* number read back would be misleading - and an absent cap, which
+* `DEFAULT_MAX_OUTPUT_TOKENS` already owns at a size nothing is short of.
+*
+* @param modelId - the routed model id.
+* @param requested - the caller's cap, undefined when it stated none.
+* @returns the cap to request, raised only when that is the difference between an
+* empty truncated answer and a usable one.
+*/
+function floorForcedThinkingTokens$1(modelId, requested) {
+	if (requested === void 0) return void 0;
+	if (thinkingModeForModel(modelId) !== "levels") return requested;
+	return Math.max(requested, 512);
+}
 /** How long a catalog sync may take before the cached list is used instead. */
 const CATALOG_TIMEOUT_MS = 15e3;
 /**
@@ -16586,6 +17237,23 @@ var OllamaAdapter = class extends LlmAdapter {
 		let lastStatus;
 		let lastDetail = "";
 		let accountId;
+		/**
+		* Set the moment ANY chunk reaches the caller, and never cleared.
+		*
+		* Deliberately broader than "some text arrived": a block start or a tool
+		* call the caller has already seen is exactly as visible as the text, so the
+		* conservative reading is the one that cannot be wrong. It also survives the
+		* pool rotation below, because output an account produced before failing is
+		* still output.
+		*/
+		let outputStarted = false;
+		/**
+		* The code a transient in-band failure becomes, or null for this attempt.
+		*
+		* Assigned per failure rather than accumulated, so a later attempt that
+		* failed for a reason of its own is never judged by an earlier one's wording.
+		*/
+		let lastInBandCode = null;
 		for (;;) {
 			let credentials;
 			if (pool === null) {
@@ -16613,6 +17281,7 @@ var OllamaAdapter = class extends LlmAdapter {
 				if (event.type === "error") {
 					lastStatus = openedStatus;
 					lastDetail = event.message;
+					lastInBandCode = outputStarted || !(openedStatus !== void 0 && openedStatus >= 200 && openedStatus < 300) ? null : inBandTransientCode(event.message);
 					failed = true;
 					break;
 				}
@@ -16621,12 +17290,18 @@ var OllamaAdapter = class extends LlmAdapter {
 					continue;
 				}
 				if (event.type === "text" || event.type === "thinking" || event.type === "tool_call") {
-					for (const chunk of applyEvent(state, event)) yield chunk;
+					for (const chunk of applyEvent(state, event)) {
+						outputStarted = true;
+						yield chunk;
+					}
 					continue;
 				}
 			}
 			if (!failed) {
-				for (const chunk of closeStream$3(state)) yield chunk;
+				for (const chunk of closeStream$3(state)) {
+					outputStarted = true;
+					yield chunk;
+				}
 				if (pool !== null && accountId !== void 0 && spent !== null) await pool.recordUsage(accountId, spent.inputTokens, spent.outputTokens);
 				return;
 			}
@@ -16646,9 +17321,76 @@ var OllamaAdapter = class extends LlmAdapter {
 		const status = lastStatus;
 		if (status === 401 || status === 403) throw new LlmError(`${PROVIDER_NAME$4} rejected the API key (${status}). Replace it from Settings > ${PROVIDER_NAME$4}.${lastDetail ? ` ${lastDetail}` : ""}`, "INVALID_CREDENTIAL", { status });
 		if (status === 429) throw new LlmError(`${PROVIDER_NAME$4} rate limit or plan quota reached (429). Add another key, or wait for the cooldown shown in Settings > ${PROVIDER_NAME$4}.${lastDetail ? ` ${lastDetail}` : ""}`, "RATE_LIMIT", { status: 429 });
-		throw new LlmError(`${PROVIDER_NAME$4} request failed${status === void 0 ? "" : ` (${status})`}.${lastDetail}`, status !== void 0 && status >= 500 ? "SERVER" : "INVALID_REQUEST", status === void 0 ? {} : { status });
+		throw new LlmError(`${PROVIDER_NAME$4} request failed${status === void 0 ? "" : ` (${status})`}.${lastDetail}`, lastInBandCode ?? (status !== void 0 && status >= 500 ? "SERVER" : "INVALID_REQUEST"), status === void 0 ? {} : { status });
 	}
 };
+/**
+* The DSH code an in-band Ollama failure becomes, or null when the message names
+* nothing transient.
+*
+* Ollama reports a failure inside a 200 as `{"error": "<message>"}` - one string,
+* no type - so the shared helper's Anthropic table cannot apply and there is no
+* status to name either. The wording is the whole signal, which is the same
+* message heuristic the Codex line already reads on `response.failed` in
+* `responses-client.ts` (generalized for the Responses vocabulary by
+* `inBandResponsesCode` in `common/stream-error.ts`); it is kept local here
+* because these are Ollama's own phrasings, not a vocabulary lines share.
+*
+* Every phrase below names a condition Ollama reports that clears on its own: a
+* model still being pulled into memory, a runner that would not start, a server
+* with no slot for this request, an overloaded or throttled front door. That is
+* what makes a retry worth taking - the request did not change, so a later one
+* is genuinely a different bet.
+*
+* Deliberately narrow, because the cost of guessing is paid by the user in
+* backoff. A refusal the wording does not name is usually the service objecting
+* to the caller's message, and retrying repeats it verbatim: three identical
+* requests ending in the same verdict, with the real fault unreported. Falling
+* back to some invented status instead would be worse on both counts, filing an
+* unrecognized refusal as a retryable server error and inviting the pool chain
+* to cool an account down for something nothing observed.
+*/
+function inBandTransientCode(message) {
+	const text = message.toLowerCase();
+	if (IN_BAND_RATE_LIMIT.some((phrase) => text.includes(phrase))) return "RATE_LIMIT";
+	if (IN_BAND_SERVER.some((phrase) => text.includes(phrase))) return "SERVER";
+	return null;
+}
+/**
+* Throttling wording, read first because it is the narrower of the two verdicts.
+*
+* The same conditions this line already types RATE_LIMIT for when they arrive as
+* a 429; a gateway that reports one inside the stream is describing the same
+* state, and the harness backoff that 429 earns is what the turn should get.
+*/
+const IN_BAND_RATE_LIMIT = [
+	"rate limit",
+	"rate-limit",
+	"rate_limit",
+	"too many requests",
+	"quota"
+];
+/**
+* Server-side conditions, in no particular order.
+*
+* The model-loading and runner wording covers the two failures Ollama raises while
+* it is still getting ready to serve - the request was well formed and the
+* service was not ready for it - and the rest is what its front door says when it
+* is saturated.
+*/
+const IN_BAND_SERVER = [
+	"overload",
+	"server busy",
+	"model is loading",
+	"loading model",
+	"couldn't load",
+	"could not load",
+	"unable to load",
+	"failed to load",
+	"runner",
+	"service unavailable",
+	"internal server error"
+];
 function isAbort$3(error, signal) {
 	return signal?.aborted === true || error instanceof Error && error.name === "AbortError";
 }
@@ -16812,7 +17554,8 @@ async function toOllamaRequest(options, attachments, signal) {
 		model: options.model,
 		messages
 	};
-	if (options.maxTokens !== void 0) request.maxOutputTokens = options.maxTokens;
+	const flooredMax = floorForcedThinkingTokens$1(options.model, options.maxTokens);
+	if (flooredMax !== void 0) request.maxOutputTokens = flooredMax;
 	if (options.temperature !== void 0) request.temperature = options.temperature;
 	const think = thinkForModel(options.model, options.reasoningEffort);
 	if (think !== void 0) request.think = think;
@@ -18971,19 +19714,89 @@ function buildModelOptions$1(catalog, enabledModelIds, contextWindowOverrides) {
 		};
 	});
 }
-let quotaCache$1 = null;
-let quotaAccountId;
+/**
+* How many accounts' snapshots are remembered.
+*
+* Mirrors the bound the ChatGPT line's usage service applies to its own
+* per-account map: room for every account a user is likely to pool, and no way
+* for a long-lived process to accumulate one snapshot per account it has seen.
+*/
+const QUOTA_ACCOUNT_LIMIT$1 = 20;
+/** Key of a snapshot whose caller named no account: the single-credential read. */
+const UNNAMED_ACCOUNT_KEY = "";
+/**
+* One newest snapshot PER ACCOUNT, keyed by the account id the caller stated.
+*
+* Quota follows the account: rotation decides which account spends the next
+* request, so a single slot could only ever describe whichever account was read
+* last. The pool's card draws every account its own reading, so each account
+* keeps its own entry; past the cap the oldest read is evicted.
+*
+* This map is a memory read only. Nothing here fetches — the settings card
+* builds its account list from it, and that must never cost one upstream
+* request per pooled account.
+*/
+const quotaCache$1 = /* @__PURE__ */ new Map();
+/**
+* Key of the entry written most recently.
+*
+* The pre-pool card asks "whatever was fetched last"; the last WRITE answers
+* that exactly, where a scan by `fetchedAt` would break a tie between two
+* accounts read inside the same millisecond by map order.
+*/
+let quotaLastKey$1 = UNNAMED_ACCOUNT_KEY;
+function quotaCacheKeyFor(accountId) {
+	return accountId ?? UNNAMED_ACCOUNT_KEY;
+}
+/** The entry read most recently, whichever account it belongs to. */
+function newestQuotaEntry$1() {
+	return quotaCache$1.get(quotaLastKey$1);
+}
+/** Remember one account's newest snapshot, bounded to the accounts in use. */
+function rememberQuota$1(accountId, snapshot) {
+	const key = quotaCacheKeyFor(accountId);
+	quotaCache$1.set(key, snapshot);
+	quotaLastKey$1 = key;
+	while (quotaCache$1.size > QUOTA_ACCOUNT_LIMIT$1) {
+		const oldest = [...quotaCache$1.entries()].sort((left, right) => left[1].fetchedAt - right[1].fetchedAt)[0];
+		if (oldest === void 0) break;
+		quotaCache$1.delete(oldest[0]);
+		if (quotaLastKey$1 === oldest[0]) quotaLastKey$1 = UNNAMED_ACCOUNT_KEY;
+	}
+}
+/**
+* The snapshot read most recently, whichever account it belongs to.
+*
+* Kept for the pre-pool single-credential card, which has no account id to ask
+* with. A pooled caller asks {@link getCachedQuotaFor} instead.
+*/
 function getCachedQuota$3() {
-	return quotaCache$1;
+	return newestQuotaEntry$1() ?? null;
 }
-/** The cached snapshot only when it belongs to this account. */
-function getCachedQuotaFor(accountId) {
-	if (accountId === void 0 || accountId === null) return quotaCache$1;
-	return quotaAccountId === accountId ? quotaCache$1 : null;
+/**
+* The cached snapshot only when it belongs to this account.
+*
+* Passing no account keeps the pre-pool behavior of returning whatever was
+* fetched last. Passing one answers that account's own newest snapshot, or null
+* when this line has never read it — null is "never read", which is a different
+* statement from a snapshot whose windows are all at zero.
+*/
+function getCachedQuotaFor$2(accountId) {
+	if (accountId === void 0 || accountId === null) return newestQuotaEntry$1() ?? null;
+	return cachedQuotaFor$1(accountId);
+}
+/**
+* The newest snapshot held for one account, for display only.
+*
+* The pool asks here while it builds the account list for the settings card, so
+* this is a pure memory read: it never fetches.
+*/
+function cachedQuotaFor$1(accountId) {
+	return quotaCache$1.get(accountId) ?? null;
 }
 function clearCachedQuota$3() {
-	quotaCache$1 = null;
-	quotaAccountId = void 0;
+	quotaCache$1.clear();
+	quotaLastKey$1 = UNNAMED_ACCOUNT_KEY;
 }
 /**
 * Display name and nominal length for each window key the service reports.
@@ -19230,8 +20043,8 @@ function parsePlanLevel(payload) {
 * account's real usage.
 */
 async function fetchAccountQuota$3(store, options = {}) {
-	const sameAccount = options.accountId === void 0 || quotaAccountId === options.accountId;
-	if (options.force !== true && sameAccount && quotaCache$1 !== null && Date.now() - quotaCache$1.fetchedAt < 12e4) return quotaCache$1;
+	const cached = quotaCache$1.get(quotaCacheKeyFor(options.accountId));
+	if (options.force !== true && cached !== void 0 && Date.now() - cached.fetchedAt < 12e4) return cached;
 	const fetchFn = options.fetchFn ?? fetch;
 	const pooled = options.credentials !== void 0;
 	let credentials;
@@ -19280,8 +20093,7 @@ async function fetchAccountQuota$3(store, options = {}) {
 		fetchedAt: Date.now(),
 		sources: [openAIUrl(USAGES_PATH, credentials.region)]
 	};
-	quotaCache$1 = snapshot;
-	quotaAccountId = options.accountId;
+	rememberQuota$1(options.accountId, snapshot);
 	return snapshot;
 }
 /**
@@ -21104,7 +21916,15 @@ function processOpenAIStreamLine(line, state) {
 	if (!isRecord$17(chunk)) return [];
 	const out = [];
 	const errorPayload = isRecord$17(chunk.error) ? chunk.error : void 0;
-	if (errorPayload !== void 0) throw new LlmError$1(`Kimi Code stream error: ${asString$9(errorPayload.message) ?? "unknown error"}`, isContextOverflow(errorPayload) ? CONTEXT_OVERFLOW_CODE : "PROVIDER_ERROR");
+	if (errorPayload !== void 0) {
+		const message = asString$9(errorPayload.message) ?? "unknown error";
+		state.streamError = {
+			vocabulary: "openai",
+			error: errorPayload,
+			message
+		};
+		throw new LlmError$1(`Kimi Code stream error: ${message}`, isContextOverflow(errorPayload) ? CONTEXT_OVERFLOW_CODE : "PROVIDER_ERROR");
+	}
 	const usage = isRecord$17(chunk.usage) ? chunk.usage : void 0;
 	if (usage) {
 		state.sawUsage = true;
@@ -21373,7 +22193,13 @@ function processAnthropicStreamLine(line, state) {
 	}
 	if (type === "error") {
 		const error = isRecord$17(event.error) ? event.error : {};
-		throw new LlmError$1(`Kimi Code stream error: ${asString$9(error.message) ?? "unknown error"}`, isContextOverflow(error) ? CONTEXT_OVERFLOW_CODE : "PROVIDER_ERROR");
+		const message = asString$9(error.message) ?? "unknown error";
+		state.streamError = {
+			vocabulary: "anthropic",
+			error,
+			message
+		};
+		throw new LlmError$1(`Kimi Code stream error: ${message}`, isContextOverflow(error) ? CONTEXT_OVERFLOW_CODE : "PROVIDER_ERROR");
 	}
 	return out;
 }
@@ -22014,27 +22840,98 @@ var KimiCodeAdapter = class extends LlmAdapter {
 		const state = createStreamState$3(wire, requestOptions.sessionId);
 		markSessionActive(requestOptions.sessionId);
 		let buffer = "";
+		/**
+		* Set the moment ANY chunk reaches the caller, and never cleared.
+		*
+		* Deliberately broader than "text arrived": a block-start the caller has
+		* already seen repeats just as badly as text it has already read, so the
+		* conservative reading is the one that cannot be wrong. It is what keeps the
+		* in-band reclassification below pre-output only.
+		*/
+		let outputStarted = false;
 		try {
-			while (true) {
-				const { done, value } = await reader.read();
-				if (done) break;
-				buffer += decoder.decode(value, { stream: true });
-				const lines = buffer.split("\n");
-				buffer = lines.pop() ?? "";
-				for (const line of lines) {
-					for (const chunk of processLine(line, state, wire)) yield chunk;
-					if (state.finished) return;
+			try {
+				while (true) {
+					const { done, value } = await reader.read();
+					if (done) break;
+					buffer += decoder.decode(value, { stream: true });
+					const lines = buffer.split("\n");
+					buffer = lines.pop() ?? "";
+					for (const line of lines) {
+						for (const chunk of processLine(line, state, wire)) {
+							outputStarted = true;
+							yield chunk;
+						}
+						if (state.finished) return;
+					}
+				}
+				buffer += decoder.decode();
+				if (buffer.trim() !== "") for (const line of buffer.split("\n")) for (const chunk of processLine(line, state, wire)) {
+					outputStarted = true;
+					yield chunk;
 				}
+				if (state.finished) return;
+				assertStreamComplete$3(state);
+				for (const chunk of closeStream$2(state, servedByAccountId)) {
+					outputStarted = true;
+					yield chunk;
+				}
+			} catch (error) {
+				const inBand = state.streamError;
+				if (error instanceof LlmError && !outputStarted && error.code === "PROVIDER_ERROR" && inBand !== void 0) throw await this.inBandStreamFailure(error, inBand, servedByAccountId);
+				throw error;
 			}
-			buffer += decoder.decode();
-			if (buffer.trim() !== "") for (const line of buffer.split("\n")) for (const chunk of processLine(line, state, wire)) yield chunk;
-			if (state.finished) return;
-			assertStreamComplete$3(state);
-			for (const chunk of closeStream$2(state, servedByAccountId)) yield chunk;
 		} finally {
 			reader.cancel().catch(() => void 0);
 		}
 	}
+	/**
+	* What one in-band failure becomes once the pool has been told what the
+	* verdict means for it.
+	*
+	* Reclassification and the pool reaction are separate because they answer
+	* different questions about the same classification: the code decides whether
+	* the harness retries, `accountScoped` decides whether a retry could land
+	* somewhere that can answer. The mapper can supply only the first, which is
+	* why the wire object is recorded there and read here.
+	*
+	* The body is the one a non-2xx response would have carried, so every pattern
+	* below reads exactly what the HTTP branch reads and a limit cannot mean one
+	* thing as a status and another as an event. The request itself is NOT
+	* re-issued: its body is already open and part-read, and the harness retry
+	* policy owns the repeat.
+	*/
+	async inBandStreamFailure(thrown, inBand, accountId) {
+		const bodyText = JSON.stringify({ error: inBand.error });
+		const reclassified = inBand.vocabulary === "anthropic" ? reclassifyInBandError(thrown, inBand.error, classifyKimiFailure) : reclassifyInBandResponsesError(thrown, inBand.error, inBand.message, classifyKimiFailure);
+		const pool = this.accountPool;
+		if (pool === null || accountId === void 0) return reclassified;
+		if (inBand.vocabulary === "anthropic") {
+			const status = inBandAnthropicStatus(inBand.error);
+			if (status === null) return reclassified;
+			const failure = classifyKimiFailure(status, bodyText);
+			const planScoped = status === 429 && matchesAny$1(bodyText, ENTITLEMENT_PATTERNS);
+			if (status === 429 && !planScoped) await this.coolInBandRateLimit(pool, accountId, failure.accountScoped, bodyText);
+			else if (failure.code === "INVALID_CREDENTIAL") await pool.markAuthFailed(accountId, failure.message).catch(() => void 0);
+			else if (failure.accountScoped) await pool.markCooldown(accountId, accountLimitCooldownMs(bodyText), `${PROVIDER_NAME$3} in-stream 403`).catch(() => void 0);
+			return reclassified;
+		}
+		if (reclassified.code === "RATE_LIMIT" && !matchesAny$1(bodyText, ENTITLEMENT_PATTERNS)) await this.coolInBandRateLimit(pool, accountId, classifyKimiFailure(429, bodyText).accountScoped, bodyText);
+		return reclassified;
+	}
+	/**
+	* Cool the account whose own rate limit the stream reported.
+	*
+	* The window this route already defaults to, never a Retry-After: there is no
+	* 429 response to read one from, and a delay stated on the 200 belongs to a
+	* different message. An account-scoped verdict takes its own window instead,
+	* because a spent balance is not back-pressure and 15 minutes would put the
+	* account back into rotation four times inside the window that refused it.
+	*/
+	async coolInBandRateLimit(pool, accountId, accountScoped, bodyText) {
+		const cooldownMs = accountScoped ? accountLimitCooldownMs(bodyText) : POOL_COOLDOWN_MS$2;
+		await pool.markCooldown(accountId, cooldownMs, `${PROVIDER_NAME$3} in-stream 429`).catch(() => void 0);
+	}
 };
 function processLine(line, state, wire) {
 	return wire === "anthropic" ? processAnthropicStreamLine(line, state) : processOpenAIStreamLine(line, state);
@@ -22097,6 +22994,34 @@ function parseKimiCodePoolData(value) {
 	return result;
 }
 /**
+* One account's own newest quota snapshot, in the shape the account card draws.
+*
+* The snapshot comes from 'client.ts''s per-account cache keyed by the pool
+* account id, and this is where this line's own numbers are translated into the
+* shared shape. Every judgement below trades a possibly-nice-looking bar for an
+* honest one:
+*
+* - a window whose consumed share nobody measured is DROPPED (`quotaWindow`
+*   returns null), because "0% used" and "not stated" are different claims. This
+*   parser only mints a window once it has a ratio, so a window that states none
+*   never reaches here at all — the drop is the second line of defence;
+* - `resetsAt` is already Unix MILLISECONDS on this line, so it is passed
+*   through and normalized rather than multiplied;
+* - the read time is the snapshot's own `fetchedAt`, which this line always
+*   sets. A snapshot with no read time is not a snapshot, and `poolQuota`
+*   reports that as undefined.
+*
+* A null snapshot means "this account was never read", which the card renders as
+* exactly that instead of drawing a zero.
+*/
+function kimiCodePoolQuota(snapshot) {
+	if (snapshot === null) return void 0;
+	return poolQuota(snapshot.fetchedAt, snapshot.windows.map((window) => quotaWindow$1(window.label, window.usedPercent, {
+		windowDurationMins: window.windowDurationMins,
+		resetsAt: window.resetsAt
+	})));
+}
+/**
 * Kimi Code's account pool.
 *
 * The credential is an OAuth pair whose refresh token rotates, so the pool owns
@@ -22175,6 +23100,7 @@ var KimiCodeAccountPool = class extends AccountPoolCore {
 				const nickname = account.nickname ?? account.credentials.nickname;
 				const planName = account.planName ?? account.credentials.planName;
 				const region = account.region ?? account.credentials.region;
+				const quota = kimiCodePoolQuota(cachedQuotaFor$1(account.id));
 				return {
 					...base,
 					...email === void 0 ? {} : { email },
@@ -22184,7 +23110,8 @@ var KimiCodeAccountPool = class extends AccountPoolCore {
 						planLabel: planName
 					},
 					...region === void 0 ? {} : { region },
-					...account.userId === void 0 ? {} : { userId: account.userId }
+					...account.userId === void 0 ? {} : { userId: account.userId },
+					...quota === void 0 ? {} : { quota }
 				};
 			},
 			emptyMessage: "Not signed in to Kimi Code. Sign in from Settings > Kimi Code.",
@@ -22298,7 +23225,7 @@ async function getKimiCodeWebStatus(store, modelSettings, preferences, options =
 	const catalog = live.length > 0 ? live : fallbackCatalog();
 	const enabled = settings.enabled !== false;
 	const models = buildModelOptions$1(catalog, resolveEnabledModelIds$3(settings.enabledModelIds, catalog, enabled), settings.contextWindowOverrides);
-	const quota = getCachedQuotaFor(activeAccount?.id ?? null);
+	const quota = getCachedQuotaFor$2(activeAccount?.id ?? null);
 	const account = quota?.account ?? (active === null || active === void 0 ? null : accountFromCredentials$2(active));
 	return {
 		enabled,
@@ -22432,7 +23359,7 @@ function registerKimiCodeRoutes(ctx, store, modelSettings, preferences, options
 				if (path === "" || path === "status") {
 					if (method !== "GET") return sendMethodNotAllowed$3(response);
 					const poolActive = await activeAccount();
-					const cached = getCachedQuotaFor(poolActive.id ?? null);
+					const cached = getCachedQuotaFor$2(poolActive.id ?? null);
 					if (poolActive.credentials !== void 0 && (cached === null || Date.now() - (cached.fetchedAt || 0) > 12e4)) {
 						const refresh = async () => {
 							const credentials = await effectiveCredentials(poolActive);
@@ -23997,6 +24924,26 @@ function outputConfigFor(modelId, requestedEffort) {
 	const effort = effortForModel(modelId, requestedEffort ?? null);
 	return effort === "default" ? void 0 : { effort };
 }
+/**
+* Raise a cap that is too small for this line's forced thinking.
+*
+* Left alone: a model that can disable thinking (its caller may have asked for
+* thinking off on purpose, and raising the cap would not enable it but a caller
+* reading the number back would be misled), and an absent cap (the caller stated
+* none, so \`maxOutputTokensFor\` owns the ceiling).
+*
+* @param modelId - the routed model id.
+* @param requested - the caller's cap, undefined when it stated none.
+* @returns the cap to request, raised only when that is the difference between an
+* empty truncated answer and a usable one.
+*/
+function floorForcedThinkingTokens(modelId, requested) {
+	if (requested === void 0) return void 0;
+	const model = catalogEntry(modelId);
+	const thinking = model?.thinking;
+	if (thinking !== "always-on" && thinking !== "forced-effort") return requested;
+	return Math.min(Math.max(requested, 512), Math.max(requested, model?.maxTokens ?? requested));
+}
 /** Output cap one request asks for, tracked against the model's declared ceiling. */
 function maxOutputTokensFor$1(modelId, contextWindow) {
 	const model = catalogEntry(modelId);
@@ -24515,6 +25462,7 @@ function processMinimaxStreamLine(line, state) {
 	}
 	if (type === "error") {
 		const error = isRecord$16(event.error) ? event.error : {};
+		state.streamError = error;
 		throw new LlmError$1("MiniMax Code（编程订阅） stream error: " + (asString$8(error.message) ?? "unknown error"), isContextOverflow(error) ? CONTEXT_OVERFLOW_CODE : "PROVIDER_ERROR");
 	}
 	return out;
@@ -26254,29 +27202,75 @@ const QUOTA_FAILURE_CACHE_MS = 10 * 6e4;
 const QUOTA_TIMEOUT_MS = 8e3;
 let quotaHostInForce = null;
 /**
-* The last usage snapshot, plus the last failure to have one.
+* How many accounts' snapshots are remembered.
 *
-* `value` and `reason` are tracked SEPARATELY from the last GOOD value on
-* purpose. A read that fails carries no information about the previous one, so
-* replacing a good snapshot with `null` turned every transient refusal into a
-* visible disappearance of the quota box; `quotaLastGood` lets the card keep
-* showing the last known numbers while saying they are old.
+* Bounded like the ChatGPT line's own per-account map: the pool accepts a
+* handful of accounts, and a long-lived process must not accumulate one entry
+* per id it has ever seen. The oldest WRITE is evicted, which is also the read
+* least likely to be looked at again.
+*/
+const QUOTA_SNAPSHOT_ACCOUNT_LIMIT = 20;
+/**
+* The newest read, whichever account it belonged to.
+*
+* Kept beside the per-account map because the page-level block and the
+* "is a read even worth starting" test are about the line's latest reading,
+* not about one account.
 */
 let quotaSnapshot = null;
-/** The last snapshot that actually parsed, kept across failures. */
-let quotaLastGood = null;
+/** Per-account snapshots, so an account is never shown the numbers read for another. */
+const quotaByAccount = /* @__PURE__ */ new Map();
 /** How long a refused credential is remembered before asking again. */
 const QUOTA_REJECTED_CACHE_MS = 30 * 6e4;
 /** Forget the cached usage snapshot (a sign-out, a sign-in, a test). */
 function clearCachedQuota$4() {
 	quotaSnapshot = null;
-	quotaLastGood = null;
+	quotaByAccount.clear();
 	quotaHostInForce = null;
 }
 /** The last usage snapshot without touching the network. */
 function getCachedQuota$4() {
 	return quotaSnapshot?.value ?? null;
 }
+/**
+* The cached snapshot only when it belongs to this account, without touching the
+* network.
+*
+* Quota follows the account: a rotation decides whose credential is spent next,
+* so the snapshot a read produced describes the row it was made for. The pool
+* asks here while it builds the settings card's account list, which must never
+* wait on — or trigger — an upstream read.
+*
+* An absent, empty or null id keeps the pre-pool behaviour and answers with the
+* newest read, which is what the page-level block wants.
+*/
+function getCachedQuotaFor$1(accountId) {
+	if (accountId === void 0 || accountId === null || accountId === "") return getCachedQuota$4();
+	return quotaByAccount.get(accountId)?.value ?? null;
+}
+/** The entry a read for this account should consult, when there is one. */
+function quotaEntryFor(accountId) {
+	if (accountId === void 0 || accountId === "") return quotaSnapshot;
+	return quotaByAccount.get(accountId) ?? null;
+}
+/**
+* Remember one read against the account it was made for.
+*
+* Re-inserting an id moves it to the newest position, so the map's first key is
+* always the least recently written entry and eviction is a single delete.
+*/
+function rememberQuotaSnapshot(entry) {
+	quotaSnapshot = entry;
+	const accountId = entry.accountId;
+	if (accountId === void 0 || accountId === "") return;
+	quotaByAccount.delete(accountId);
+	quotaByAccount.set(accountId, entry);
+	while (quotaByAccount.size > QUOTA_SNAPSHOT_ACCOUNT_LIMIT) {
+		const oldest = quotaByAccount.keys().next().value;
+		if (oldest === void 0) break;
+		quotaByAccount.delete(oldest);
+	}
+}
 /** Why the last read produced no snapshot, or null when it produced one. */
 function getQuotaUnavailable() {
 	return quotaSnapshot?.reason ?? null;
@@ -26398,9 +27392,10 @@ function parseTokenPlanQuota(payload, now = Date.now()) {
 async function fetchTokenPlanQuota(credentials, options = {}) {
 	const fetchFn = options.fetchFn ?? fetch;
 	const now = Date.now();
-	if (options.force !== true && quotaSnapshot !== null) {
-		const ttl = quotaSnapshot.reason === "credential-not-accepted" ? QUOTA_REJECTED_CACHE_MS : quotaSnapshot.reason === "token-expired" && quotaSnapshot.value === null ? QUOTA_FAILURE_CACHE_MS : quotaSnapshot.value === null ? QUOTA_FAILURE_CACHE_MS : QUOTA_CACHE_MS;
-		if (now - quotaSnapshot.at < ttl) return quotaSnapshot.value;
+	const cached = quotaEntryFor(options.accountId);
+	if (options.force !== true && cached !== null) {
+		const ttl = cached.reason === "credential-not-accepted" ? QUOTA_REJECTED_CACHE_MS : cached.reason === "token-expired" && cached.value === null ? QUOTA_FAILURE_CACHE_MS : cached.value === null ? QUOTA_FAILURE_CACHE_MS : QUOTA_CACHE_MS;
+		if (now - cached.at < ttl) return cached.value;
 	}
 	const candidates = quotaHostCandidates(credentials.region);
 	const ordered = quotaHostInForce === null ? candidates : [quotaHostInForce, ...candidates.filter((host) => host !== quotaHostInForce)];
@@ -26440,19 +27435,19 @@ async function fetchTokenPlanQuota(credentials, options = {}) {
 		});
 	}
 	if (value === null && reason === null) reason = authRefused ? "token-expired" : "unreachable";
-	if (value !== null) quotaLastGood = {
-		value,
-		at: Date.now()
-	};
-	else if (quotaLastGood !== null) {
-		value = quotaLastGood.value;
-		reason = "stale";
+	if (value === null) {
+		const previous = quotaEntryFor(options.accountId);
+		if (previous !== null && previous.value !== null) {
+			value = previous.value;
+			reason = "stale";
+		}
 	}
-	quotaSnapshot = {
+	rememberQuotaSnapshot({
+		accountId: options.accountId,
 		at: Date.now(),
 		value,
 		reason
-	};
+	});
 	return value;
 }
 /**
@@ -27175,7 +28170,7 @@ var MinimaxCodeAdapter = class extends LlmAdapter {
 		}
 		let authOwner = computeMinimaxAuthOwner(credentials, accountId);
 		const withUploads = await uploadOversizedMedia(requestOptions, images, videos, credentials, accountId, this.options.fetchFn ?? fetch, signal);
-		const effectiveMax = clampOutputToContext(requestOptions.maxTokens ?? maxOutputTokensFor$1(options.model, contextWindow), contextWindow, estimatedInputTokens(requestOptions));
+		const effectiveMax = clampOutputToContext(floorForcedThinkingTokens(options.model, requestOptions.maxTokens) ?? maxOutputTokensFor$1(options.model, contextWindow), contextWindow, estimatedInputTokens(requestOptions));
 		const buildBodyForOwner = (owner) => {
 			return assertRequestBodyFits(buildMinimaxRequest({
 				...withUploads,
@@ -27235,23 +28230,46 @@ var MinimaxCodeAdapter = class extends LlmAdapter {
 			provider: PROVIDER_ID$2
 		});
 		let buffer = "";
+		/**
+		* Set the moment ANY chunk reaches the caller, and never cleared.
+		*
+		* Deliberately broader than "a content delta or a tool block": a retry would
+		* duplicate a block-start the caller has already seen just as surely, so the
+		* conservative reading is the one that cannot be wrong. It is what keeps the
+		* in-band reclassification below pre-output only.
+		*/
+		let outputStarted = false;
 		try {
-			while (true) {
-				const { done, value } = await reader.read();
-				if (done) break;
-				buffer += decoder.decode(value, { stream: true });
-				const lines = buffer.split("\n");
-				buffer = lines.pop() ?? "";
-				for (const line of lines) {
-					for (const chunk of processMinimaxStreamLine(line, state)) yield chunk;
-					if (state.finished) return;
+			try {
+				while (true) {
+					const { done, value } = await reader.read();
+					if (done) break;
+					buffer += decoder.decode(value, { stream: true });
+					const lines = buffer.split("\n");
+					buffer = lines.pop() ?? "";
+					for (const line of lines) {
+						for (const chunk of processMinimaxStreamLine(line, state)) {
+							outputStarted = true;
+							yield chunk;
+						}
+						if (state.finished) return;
+					}
 				}
+				buffer += decoder.decode();
+				if (buffer.trim() !== "") for (const line of buffer.split("\n")) for (const chunk of processMinimaxStreamLine(line, state)) {
+					outputStarted = true;
+					yield chunk;
+				}
+				if (state.finished) return;
+				assertStreamComplete(state);
+				for (const chunk of closeMinimaxStream(state)) {
+					outputStarted = true;
+					yield chunk;
+				}
+			} catch (error) {
+				if (error instanceof LlmError && !outputStarted && error.code === "PROVIDER_ERROR" && state.streamError !== void 0) throw reclassifyInBandError(error, state.streamError, classifyMinimaxFailure);
+				throw error;
 			}
-			buffer += decoder.decode();
-			if (buffer.trim() !== "") for (const line of buffer.split("\n")) for (const chunk of processMinimaxStreamLine(line, state)) yield chunk;
-			if (state.finished) return;
-			assertStreamComplete(state);
-			for (const chunk of closeMinimaxStream(state)) yield chunk;
 		} finally {
 			reader.cancel().catch(() => void 0);
 		}
@@ -27467,6 +28485,38 @@ function parseMinimaxCodePoolData(value) {
 	return result;
 }
 /**
+* Nominal length of the two windows the Token Plan reports, in minutes.
+*
+* The line's own card names these two keys (`quotaWindowInterval` = 5 小时窗口,
+* `quotaWindowWeekly` = 每周窗口), and 5 hours / 7 days is what those words mean,
+* so the shared account card can name an unlabeled window from its length
+* instead of showing a generic one. The same two numbers label the equivalent
+* windows on the Command Code and Kimi Code lines.
+*/
+const WINDOW_MINUTES = {
+	interval: 300,
+	weekly: 10080
+};
+/**
+* One cached read as the shared account card renders it.
+*
+* Every window is labeled through its LENGTH rather than with a string: the
+* canonical window names live in the client's dictionary, and the host has no
+* business shipping display text. A window whose consumed share nobody measured
+* is dropped by {@link quotaWindow} — a null percentage is not 0%, it is
+* "not stated", and a bar would lie about it. An `unlimited` window states no
+* consumed share at all and the per-account shape has no way to say
+* "unlimited", so it is dropped too rather than drawn as an empty bar.
+*/
+function minimaxCodePoolQuota(quota) {
+	if (quota === null || quota === void 0) return void 0;
+	const windows = quota.windows === void 0 || quota.windows.length === 0 ? [quotaWindow$1("", quota.usedPercent, { resetsAt: quota.resetsAtMs })] : quota.windows.map((window) => quotaWindow$1("", window.usedPercent, {
+		windowDurationMins: WINDOW_MINUTES[window.key],
+		resetsAt: window.resetsAtMs
+	}));
+	return poolQuota(quota.fetchedAtMs, windows);
+}
+/**
 * MiniMax Code's account pool.
 *
 * It owns rotation and the write-back, exactly like the sibling lines, with one
@@ -27547,7 +28597,11 @@ var MinimaxCodeAccountPool = class extends AccountPoolCore {
 					...planName === void 0 ? {} : {
 						planName,
 						planLabel: planName
-					}
+					},
+					...(() => {
+						const quota = minimaxCodePoolQuota(getCachedQuotaFor$1(account.id));
+						return quota === void 0 ? {} : { quota };
+					})()
 				};
 			},
 			liveCredentialsFor: (account) => this.nativeAuthority(account),
@@ -27634,6 +28688,34 @@ var MinimaxCodeAccountPool = class extends AccountPoolCore {
 		});
 	}
 	/**
+	* The account a usage read belongs to, together with the credential to read it
+	* with.
+	*
+	* Authored here rather than in the routes because it is provider-shaped: the
+	* routes need a credential and the ACCOUNT it was read for, so a snapshot can
+	* be remembered against the row whose numbers it describes. The selection is
+	* the core's own rule for a caller that names no account — the active one, then
+	* the primary, then the first — repeated in these three lines because the core
+	* exposes it only through `getFreshCredential`, which returns the credential
+	* alone; the credential (and its single-flight refresh, write-back and
+	* bookkeeping) still comes from the core, so no rotation rule lives here.
+	*
+	* No upstream request is made for the selection itself, and none of this moves
+	* the rotation: reading quota is not serving a request.
+	*/
+	async getQuotaReadTarget(fetchFn = fetch) {
+		const data = await this.read();
+		const target = data.accounts.find((account) => account.id === data.activeAccountId) ?? data.accounts.find((account) => account.isPrimary) ?? data.accounts[0];
+		if (target === void 0) {
+			await this.getFreshCredential(void 0, fetchFn);
+			throw new Error("MiniMax Code 号池中没有账号。");
+		}
+		return {
+			accountId: target.id,
+			credentials: await this.getFreshCredential(target.id, fetchFn)
+		};
+	}
+	/**
 	* Pick the account for the next request.
 	*
 	* The returned credential carries its own region and record key, so a caller
@@ -27967,8 +29049,12 @@ function registerMinimaxCodeRoutes(ctx, store = new MinimaxCodeCredentialStore()
 	* network, and the status route reads the snapshot through the cache instead.
 	*/
 	const refreshQuota = async (force) => {
-		const credentials = await (options.accountPool === void 0 ? ensureAccessToken$2(store, { fetchFn }).catch(() => null) : options.accountPool.getFreshCredential(void 0, fetchFn).catch(() => null));
-		if (credentials === null) return null;
+		const target = await (options.accountPool === void 0 ? ensureAccessToken$2(store, { fetchFn }).then((credentials) => ({
+			accountId: void 0,
+			credentials
+		})).catch(() => null) : options.accountPool.getQuotaReadTarget(fetchFn).catch(() => null));
+		if (target === null) return null;
+		const credentials = target.credentials;
 		const readToken = credentials.refreshToken;
 		const isCredentialStale = async () => {
 			const live = await store.readFresh().catch(() => null);
@@ -27978,10 +29064,11 @@ function registerMinimaxCodeRoutes(ctx, store = new MinimaxCodeCredentialStore()
 			fetchFn,
 			force,
 			isCredentialStale,
+			...target.accountId === void 0 ? {} : { accountId: target.accountId },
 			renewCredential: async () => await (options.accountPool === void 0 ? ensureAccessToken$2(store, {
 				fetchFn,
 				force: true
-			}) : options.accountPool.renewCredential(void 0, fetchFn)).catch(() => null)
+			}) : options.accountPool.renewCredential(target.accountId, fetchFn)).catch(() => null)
 		}).catch(() => null);
 	};
 	/**
@@ -30046,6 +31133,9 @@ const CATALOG_CACHE_TTL_MS$1 = 1800 * 1e3;
 * every cache miss.
 */
 const catalogSnapshotScopesLoaded = /* @__PURE__ */ new Set();
+function getCachedCatalog$1() {
+	return cachedCatalog?.models ?? [];
+}
 function clearCachedCatalog$1() {
 	cachedCatalog = void 0;
 	catalogCacheEpoch += 1;
@@ -30237,16 +31327,64 @@ function billingMeters(billing) {
 	}
 	return meters;
 }
-let cachedQuota;
+/**
+* One newest snapshot PER ACCOUNT, keyed by the snapshot's own `account.id` —
+* the same public key the pool addresses an account with.
+*
+* Quota follows the account: rotation decides which account spends the next
+* request, so a single slot only ever described whichever account was read
+* last. The pool's card draws every account its own reading, so each account
+* keeps its own entry; past the cap the oldest read is evicted.
+*
+* This map is a memory read only. Nothing here fetches — the settings card
+* builds its account list from it, and that must never cost one upstream
+* request per pooled account.
+*/
+const cachedQuotaEntries = /* @__PURE__ */ new Map();
+/** How many accounts' snapshots are remembered; mirrors the Codex usage service. */
+const QUOTA_CACHE_LIMIT = 20;
+/** The entry read most recently, whichever account it belongs to. */
+function newestCachedQuota() {
+	let newest;
+	for (const snapshot of cachedQuotaEntries.values()) if (newest === void 0 || (snapshot.fetchedAt || 0) > (newest.fetchedAt || 0)) newest = snapshot;
+	return newest;
+}
+/** Remember one account's newest snapshot, bounded to the accounts in use. */
+function rememberCachedQuota(snapshot) {
+	cachedQuotaEntries.set(snapshot.account.id, snapshot);
+	while (cachedQuotaEntries.size > QUOTA_CACHE_LIMIT) {
+		const oldest = [...cachedQuotaEntries.entries()].sort((left, right) => (left[1].fetchedAt || 0) - (right[1].fetchedAt || 0))[0];
+		if (oldest === void 0) break;
+		cachedQuotaEntries.delete(oldest[0]);
+	}
+}
 let quotaInFlight$1 = null;
 let quotaInFlightAccountId = null;
 let quotaCacheEpoch = 0;
+/**
+* The snapshot read most recently, whichever account it belongs to.
+*
+* Kept with a no-argument signature for backward compatibility; the page-level
+* facts use {@link getCachedQuotaFor} instead, because they describe one
+* account and must never be another account's numbers.
+*/
 function getCachedQuota$5() {
-	return cachedQuota;
+	return newestCachedQuota();
+}
+/**
+* The snapshot remembered for one account, if it was read before.
+*
+* This is the lookup the account pool calls while it builds its summaries: a
+* pure memory read, never a fetch. Passing no account keeps the pre-pool
+* behavior of returning whatever was fetched last.
+*/
+function getCachedQuotaFor(accountId) {
+	if (accountId === void 0 || accountId === null) return newestCachedQuota();
+	return cachedQuotaEntries.get(accountId);
 }
 function clearCachedQuota$5() {
 	quotaCacheEpoch += 1;
-	cachedQuota = void 0;
+	cachedQuotaEntries.clear();
 	quotaInFlight$1 = null;
 	quotaInFlightAccountId = null;
 }
@@ -30263,7 +31401,8 @@ async function fetchAccountQuota$4(store = new FileCredentialStore$4(), fetchFn
 	});
 	if (credentials === null) throw new Error(`Not signed in to ${PROVIDER_NAME$1}.`);
 	const requestedAccountId = workBuddyAccountId(credentials);
-	if (!force && cachedQuota && cachedQuota.account.id === requestedAccountId && Date.now() - (cachedQuota.fetchedAt || 0) < 12e4) return cachedQuota;
+	const cached = cachedQuotaEntries.get(requestedAccountId);
+	if (!force && cached !== void 0 && Date.now() - (cached.fetchedAt || 0) < 12e4) return cached;
 	if (quotaInFlight$1 && quotaInFlightAccountId === requestedAccountId) return quotaInFlight$1;
 	const epoch = quotaCacheEpoch;
 	const request = (async () => {
@@ -30299,7 +31438,7 @@ async function fetchAccountQuota$4(store = new FileCredentialStore$4(), fetchFn
 			sources: ["billing/meter/get-user-resource"]
 		};
 		if (epoch !== quotaCacheEpoch) return snapshot;
-		cachedQuota = snapshot;
+		rememberCachedQuota(snapshot);
 		return snapshot;
 	})();
 	quotaInFlight$1 = request;
@@ -30614,7 +31753,7 @@ async function getWorkBuddyWebStatus(store, modelSettings, preferences, options
 	const catalog = live.length > 0 ? live : FALLBACK_MODELS$1;
 	const available = modelsForRegion(credentials?.region ?? "intl", catalog).map((model) => model.id);
 	const models = buildModelOptions(catalog, available, resolveEnabledModelIds$2(settings.enabledModelIds, available, enabled), settings.contextWindowOverrides);
-	const quota = getCachedQuota$5();
+	const quota = getCachedQuotaFor(credentials === null ? null : accountFromCredentials$1(credentials).id);
 	const poolData = options.accountPool === void 0 ? null : await options.accountPool.read().catch(() => null);
 	const poolAccounts = options.accountPool === void 0 ? [] : await options.accountPool.listAccounts().catch(() => []);
 	const hiddenIds = new Set(settings.hiddenAccountIds);
@@ -30673,12 +31812,10 @@ function registerWorkBuddyRoutes(ctx, store, modelSettings, preferences, options
 						accountId: settings.selectedAccountId,
 						hiddenAccountIds: settings.hiddenAccountIds
 					});
-					const cached = getCachedQuota$5();
-					const quotaMatches = cached?.account.id === (credentials === null ? void 0 : accountFromCredentials$1(credentials).id);
-					if (credentials !== null && (!quotaMatches || cached === void 0 || Date.now() - (cached.fetchedAt || 0) > 12e4)) {
-						if (!quotaMatches) clearCachedQuota$5();
+					const cached = getCachedQuotaFor(credentials === null ? null : accountFromCredentials$1(credentials).id);
+					if (credentials !== null && (cached === void 0 || Date.now() - (cached.fetchedAt || 0) > 12e4)) {
 						const refresh = () => fetchAccountQuota$4(store, fetchFn, false, settings.selectedAccountId, settings.hiddenAccountIds);
-						if (cached === void 0 || !quotaMatches) await quotaRefresh.run(refresh);
+						if (cached === void 0) await quotaRefresh.run(refresh);
 						else quotaRefresh.start(refresh);
 					}
 					return sendJson$1(response, 200, {
@@ -32164,6 +33301,47 @@ function defaultAliasFor(credentials, position) {
 	return credentials.nickname || credentials.uin || credentials.uid || `账号 ${position}`;
 }
 /**
+* A meter's consumed share, from whichever fraction this line measured.
+*
+* `usedFraction` is the DTO's own "share consumed"; `remainingFraction` is the
+* same statement read from the other end, and is the one the rest of the card
+* renders. A meter that states neither is dropped (null) rather than drawn at
+* 0, which would claim a full window no reading supports.
+*/
+function meterUsedPercent(meter) {
+	if (typeof meter.usedFraction === "number" && Number.isFinite(meter.usedFraction)) return meter.usedFraction * 100;
+	if (typeof meter.remainingFraction === "number" && Number.isFinite(meter.remainingFraction)) return (1 - meter.remainingFraction) * 100;
+	return null;
+}
+/**
+* One cached snapshot as the shared account card renders it.
+*
+* Only what the DTO states is published: every meter that measured a share
+* becomes a window, and a meter that measured none is dropped rather than drawn
+* at 0 — which would claim an unspent allowance no reading supports.
+*
+* Exported for the pool's regression tests; the card only reaches it through
+* the summary hook.
+*/
+function workBuddyPoolQuota(snapshot) {
+	const windows = snapshot.meters.map((meter) => quotaWindow$1(meter.label, meterUsedPercent(meter), { resetsAt: meter.resetsAt }));
+	if (!snapshot.meters.some((meter) => meter.id === "cycle") && snapshot.cycleCredits !== null && snapshot.cycleCredits > 0 && snapshot.cycleUsedCredits !== null) windows.push(quotaWindow$1(snapshot.packageName ?? "Billing cycle", snapshot.cycleUsedCredits / snapshot.cycleCredits * 100, { resetsAt: snapshot.cycleEndsAt }));
+	return poolQuota(snapshot.fetchedAt, windows);
+}
+/**
+* This account's own newest snapshot, ready for the account card.
+*
+* Quota follows the account, so the pool reports the snapshot the line already
+* remembered for THIS account id. The lookup is a pure memory read: the card
+* builds its account list here, and that must never cost one upstream request
+* per pooled account. Undefined means exactly that — never read — which is not
+* the same statement as "nothing used".
+*/
+function quotaFor(accountId) {
+	const snapshot = getCachedQuotaFor(accountId);
+	return snapshot === void 0 ? void 0 : workBuddyPoolQuota(snapshot);
+}
+/**
 * WorkBuddy's account pool.
 *
 * Desktop accounts are adopted into the pool so they can be scheduled and
@@ -32224,7 +33402,11 @@ var WorkBuddyAccountPool = class extends AccountPoolCore {
 				...account.credentials.domain === "" ? {} : { domain: account.credentials.domain },
 				...account.credentials.backend === "" ? {} : { backend: account.credentials.backend },
 				...account.credentials.accountType === void 0 ? {} : { accountType: account.credentials.accountType },
-				...account.credentials.expiresAt > 0 ? {} : { expiresAt: void 0 }
+				...account.credentials.expiresAt > 0 ? {} : { expiresAt: void 0 },
+				...(() => {
+					const quota = quotaFor(account.id);
+					return quota === void 0 ? {} : { quota };
+				})()
 			}),
 			preferAccountId: () => this.resolveSelectionId(options.selection?.().selectedAccountId ?? null),
 			emptyMessage: `Not signed in to ${PROVIDER_NAME$1}. Sign in with the CodeBuddy desktop client, or add an account from Settings > WorkBuddy.`,
@@ -34233,6 +35415,10 @@ const catalogInFlight = /* @__PURE__ */ new Map();
 function clearCachedCatalog() {
 	catalogCache = null;
 }
+/** The listing currently held, or an empty array before the first load. */
+function getCachedCatalog() {
+	return catalogCache?.models ?? [];
+}
 /**
 * Read one entry of the live listing.
 *
@@ -34528,22 +35714,108 @@ function parseQuotaHeaders(headers) {
 		}
 	};
 }
-let quotaCache = null;
+/**
+* Cache key of one account's snapshot.
+*
+* A pool row id when the caller names one — that is the identity that survives a
+* token refresh and the one the account card asks with — and otherwise the
+* credential's token tail, which is all a single-credential process has. The
+* `acc:` prefix keeps the two apart, so no row id can ever be answered with an
+* entry written under a token tail.
+*/
+function quotaEntryKey(accountId, credentials) {
+	return accountId === void 0 || accountId === "" ? accountKeyFor(credentials) : `acc:${accountId}`;
+}
+/**
+* How many accounts' snapshots are remembered.
+*
+* Mirrors the bound the ChatGPT line's usage service applies to its own
+* per-account map: room for every account a user is likely to pool, and no way
+* for a long-lived process to accumulate one snapshot per token it has seen.
+*/
+const QUOTA_ACCOUNT_LIMIT = 20;
+/**
+* The newest snapshot of each account, keyed by {@link quotaEntryKey}.
+*
+* ONE ENTRY PER ACCOUNT rather than one slot for the line: quota follows the
+* account, and rotation decides which account spends the next request, so a
+* single slot could only ever describe whichever account was read last. This is
+* a memory read only — the settings card builds its account list from it, and
+* that must never cost one upstream request per pooled account.
+*/
+const quotaCache = /* @__PURE__ */ new Map();
+/**
+* Key of the entry written most recently.
+*
+* The pre-pool card asks "whatever was read last", and a header reading that
+* names no account may only merge into the snapshot it was read beside. Both
+* need the last WRITE rather than a scan: two accounts can be written inside one
+* millisecond, and a tie would otherwise be broken by map order.
+*/
+let quotaLastKey = null;
 const quotaInFlight = /* @__PURE__ */ new Map();
-/** Drop the cached snapshot. Called on sign-out and after a credential change. */
+/** Drop every cached snapshot. Called on sign-out and after a credential change. */
 function clearCachedQuota$1() {
-	quotaCache = null;
+	quotaCache.clear();
+	quotaLastKey = null;
+}
+/** The entry written most recently, whichever account it belongs to. */
+function newestQuotaEntry() {
+	return quotaLastKey === null ? null : quotaCache.get(quotaLastKey) ?? null;
+}
+/**
+* Remember one account's newest snapshot, bounded to the accounts in use.
+*
+* The oldest entry is measured by {@link QuotaCacheEntry.observedAt} rather than
+* by `fetchedAt`: a snapshot kept warm by response headers has no full read
+* behind it at all, and it is still the newest thing known about that account.
+*/
+function rememberQuota(entry) {
+	quotaCache.set(entry.key, entry);
+	quotaLastKey = entry.key;
+	while (quotaCache.size > QUOTA_ACCOUNT_LIMIT) {
+		const oldest = [...quotaCache.entries()].sort((left, right) => left[1].observedAt - right[1].observedAt)[0];
+		if (oldest === void 0) break;
+		quotaCache.delete(oldest[0]);
+		if (quotaLastKey === oldest[0]) quotaLastKey = null;
+	}
 }
 /**
 * The cached snapshot, optionally only when it belongs to this credential.
 *
 * With one stored account the argument is unnecessary; it exists so a later
 * pool chunk cannot render one account's usage under another account's name.
+* Given a credential the answer is THAT account's own newest snapshot, which is
+* what keeps two pooled accounts apart: a later read of a different account
+* cannot displace it. Without one the pre-pool answer is kept — whatever was
+* written last — because the single-account card has no identity to ask with.
 */
 function getCachedQuota$1(credentials) {
-	if (quotaCache === null) return null;
-	if (credentials === void 0) return quotaCache.quota;
-	return quotaCache.key === accountKeyFor(credentials) ? quotaCache.quota : null;
+	if (credentials === void 0) return newestQuotaEntry()?.quota ?? null;
+	return cachedQuotaFor(credentials);
+}
+/**
+* The newest snapshot held for one credential's account, for display only.
+*
+* The token-tail form, kept for a composition with a single stored credential.
+* A pooled caller asks {@link cachedQuotaForPool} with the row id instead: a
+* token is rotated by every refresh, so an entry filed under a tail is lost the
+* moment the account renews.
+*/
+function cachedQuotaFor(credentials) {
+	return quotaCache.get(accountKeyFor(credentials))?.quota ?? null;
+}
+/**
+* The newest snapshot held for one pool row, for display only.
+*
+* The pool asks here while it builds the account list for the settings card, so
+* this is a pure memory read: it never fetches. Null means nothing was ever read
+* for this account, which is a different statement from "nothing used". A row id
+* that is undefined falls back to the credential's own tail, which is what a
+* composition without a pool has.
+*/
+function cachedQuotaForPool(accountId, credentials) {
+	return quotaCache.get(quotaEntryKey(accountId, credentials))?.quota ?? null;
 }
 /**
 * Read the account's usage meters.
@@ -34561,12 +35833,13 @@ function getCachedQuota$1(credentials) {
 * up staring at a card that claims they are fine.
 */
 function fetchAccountQuota$1(credentials, options = {}) {
-	const key = accountKeyFor(credentials);
+	const key = quotaEntryKey(options.accountId, credentials);
 	const at = now();
-	if (options.force !== true && quotaCache !== null && quotaCache.key === key) {
-		const warm = at - quotaCache.observedAt < QUOTA_CACHE_TTL_MS;
-		const freshFullRead = quotaCache.fullReadAt !== null && at - quotaCache.fullReadAt < 3e5;
-		if (warm && freshFullRead) return Promise.resolve(quotaCache.quota);
+	const cached = quotaCache.get(key);
+	if (options.force !== true && cached !== void 0) {
+		const warm = at - cached.observedAt < QUOTA_CACHE_TTL_MS;
+		const freshFullRead = cached.fullReadAt !== null && at - cached.fullReadAt < 3e5;
+		if (warm && freshFullRead) return Promise.resolve(cached.quota);
 	}
 	const pending = quotaInFlight.get(key);
 	if (pending !== void 0) return pending;
@@ -34581,9 +35854,9 @@ function fetchAccountQuota$1(credentials, options = {}) {
 	return run;
 }
 async function performQuotaFetch(credentials, options) {
-	const key = accountKeyFor(credentials);
+	const key = quotaEntryKey(options.accountId, credentials);
 	const fetchFn = options.fetchFn ?? fetch;
-	const previous = quotaCache !== null && quotaCache.key === key ? quotaCache.quota : null;
+	const previous = quotaCache.get(key)?.quota ?? null;
 	let response;
 	try {
 		response = await fetchFn("https://api.anthropic.com/api/oauth/usage", {
@@ -34621,12 +35894,12 @@ async function performQuotaFetch(credentials, options) {
 		status: null,
 		representativeClaim: null
 	};
-	quotaCache = {
+	rememberQuota({
 		quota,
 		key,
 		fullReadAt: at,
 		observedAt: at
-	};
+	});
 	return quota;
 }
 /**
@@ -36641,6 +37914,7 @@ function processStreamLine(line, state) {
 		const error = isRecord$5(event.error) ? event.error : {};
 		const message = asString$2(error.message) ?? "unknown error";
 		const kind = asString$2(error.type);
+		state.streamError = error;
 		throw new LlmError$1("Claude stream error" + (kind === void 0 ? "" : " (" + kind + ")") + ": " + message, isContextOverflow(error) ? CONTEXT_OVERFLOW_CODE : "PROVIDER_ERROR");
 	}
 	return out;
@@ -37814,6 +39088,13 @@ function cancelLogin() {
 * - `request` -> PROVIDER_ERROR;
 * - `network` -> TRANSPORT.
 *
+* An in-band `error` event inside a 200 stream carries the same envelope and is
+* classified the same way while nothing has reached the caller; after the first
+* chunk it is surfaced as PROVIDER_ERROR, for the reason in note 1
+* ({@link inBandStreamVerdict}). A reclassified in-band failure is weighed by
+* {@link shouldRotateAccount} exactly as a non-2xx one is, so an account-scoped
+* limit reported inside the stream still takes that account out of rotation.
+*
 * A reported-client-version rejection is a REQUEST problem even though it
 * arrives as a 400 that mentions the client: it must NOT sign the user out, and
 * it has a real remedy (raise the reported version), so it is handled
@@ -38213,8 +39494,19 @@ var ClaudeAdapter = class ClaudeAdapter extends LlmAdapter {
 		* a later reordering of this method cannot silently drop it.
 		*/
 		let outputStarted = false;
+		/**
+		* The account the response now being streamed was issued with.
+		*
+		* Hoisted out of the loop because the in-stream failure path needs it: a
+		* rate limit reported inside a 200 stream belongs to the account that was
+		* asked, and that account is what the failure takes out of rotation — the
+		* harness is about to repeat this request, and repeating it against the same
+		* exhausted account fails identically.
+		*/
+		let usedAccountId;
 		while (true) {
 			const { credentials, accountId } = await this.resolveCredential(pool, tried, fetchFn, signal);
+			usedAccountId = accountId;
 			response = await this.attemptRequest(credentials, requestOptions, images, toolNames, thinking, settings, signal, fetchFn);
 			if (response.ok) break;
 			const detail = (await response.text().catch(() => "")).slice(0, 2e3);
@@ -38260,7 +39552,12 @@ var ClaudeAdapter = class ClaudeAdapter extends LlmAdapter {
 					yield chunk;
 				}
 			} catch (error) {
-				if (error instanceof LlmError) throw error;
+				if (error instanceof LlmError) {
+					if (outputStarted || error.code !== "PROVIDER_ERROR" || state.streamError === void 0) throw error;
+					const verdict = inBandStreamVerdict(error, state.streamError, response.headers);
+					if (verdict.failure !== null && shouldRotateAccount(verdict.failure, outputStarted) && usedAccountId !== void 0) await pool?.markCooldown(usedAccountId, cooldownMsFor(verdict.failure), "Claude（订阅） in-stream 429").catch(() => void 0);
+					throw verdict.error;
+				}
 				if (signal.aborted) throw new LlmError("Claude request aborted", "ABORTED", { cause: error });
 				throw new LlmError("Claude stream failed: " + (error instanceof Error ? error.message : String(error)), outputStarted ? "PROVIDER_ERROR" : "TRANSPORT", { cause: error });
 			}
@@ -38350,6 +39647,49 @@ function codeForFailure(failure) {
 	return "PROVIDER_ERROR";
 }
 /**
+* The DSH error an in-band `error` event becomes when it arrived BEFORE any
+* chunk reached the caller.
+*
+* Anthropic can answer 200 and then report the failure inside the stream, in
+* the same `{ error: { type, message } }` envelope a non-2xx body carries. The
+* mapper types every such event PROVIDER_ERROR, because it cannot know whether
+* output has started — and PROVIDER_ERROR is outside the retry set, so an
+* `overloaded_error` delivered this way ended the turn on the first try while
+* the identical overload delivered as a 529 is retried. Before any output a
+* fresh request is still free (module note 1), so the envelope is classified
+* exactly as a response body would be, and a transient verdict (overloaded,
+* server, rate limit) takes its retryable code.
+*
+* Everything else keeps the mapper's verdict: a request problem, a context
+* overflow, a credential refusal, and a type this line does not recognize. The
+* last matters: classifyFailure falls back to the status line for an unknown
+* type, and no status here describes the failure — the response itself was a
+* 200 — so the status passed is 0 and never consulted for a recognized type.
+*
+* The message stays the mapper's, which names the wire type the user saw.
+*
+* The classified failure rides along beside the error so the caller can weigh it
+* with {@link shouldRotateAccount} instead of re-deriving what it says. It is
+* null whenever the mapper's verdict stands, so a caller never has to guess
+* whether `accountScoped` is meaningful.
+*
+* @param thrown - the mapper's PROVIDER_ERROR (or context-overflow) verdict.
+* @param streamError - the event's wire `error` object, from the stream state.
+* @param headers - the 200 response's headers, read for the rate-limit verdict.
+*/
+function inBandStreamVerdict(thrown, streamError, headers) {
+	const failure = classifyFailure(0, JSON.stringify({ error: streamError }), headers);
+	if (failure.type === null || !failure.retryable) return {
+		error: thrown,
+		failure: null
+	};
+	const retryAfter = failure.retryAfterMs !== null && failure.retryAfterMs > 0 ? failure.retryAfterMs : void 0;
+	return {
+		error: new LlmError(thrown.message, codeForFailure(failure), retryAfter === void 0 ? {} : { providerRetryAfterMs: retryAfter }),
+		failure
+	};
+}
+/**
 * Turn one classified failure into the DSH error the caller sees.
 *
 * The facts the harness needs ride along: the HTTP status, and the delay the
@@ -39186,6 +40526,32 @@ async function refreshClaudeCredential(credentials, fetchFn) {
 	};
 }
 /**
+* One account's own newest quota snapshot, in the shape the account card draws.
+*
+* The snapshot comes from 'client.ts''s per-account cache, and this is where
+* this line's own numbers are translated into the shared shape. Every judgement
+* below trades a possibly-nice-looking bar for an honest one:
+*
+* - a window whose consumed share nobody measured is DROPPED (`quotaWindow`
+*   returns null), because "0% used" and "not stated" are different claims;
+* - the share falls back to `100 - remainingPercent` only when that field was
+*   stated; when the payload stated NEITHER share there is nothing to draw;
+* - the read time is the full read's own `fetchedAt`, or — for a snapshot that
+*   only response headers ever kept warm, which performed no full read — the
+*   `observedAt` instant those numbers arrived. Either way it is when this line
+*   saw the numbers, never when upstream served them.
+*
+* A null snapshot means "this account was never read", which `poolQuota` reports
+* as undefined so the card can say exactly that instead of drawing a zero.
+*/
+function claudePoolQuota(snapshot) {
+	if (snapshot === null) return void 0;
+	return poolQuota(snapshot.fetchedAt ?? snapshot.observedAt, snapshot.windows.map((window) => quotaWindow$1(window.label, window.usedPercent ?? (window.remainingPercent === null ? null : 100 - window.remainingPercent), {
+		windowDurationMins: window.windowMinutes,
+		resetsAt: typeof window.resetsAt === "string" ? Date.parse(window.resetsAt) : null
+	})));
+}
+/**
 * The Claude subscription account pool.
 *
 * Read the module comment first: it carries the two rules this class exists to
@@ -39283,6 +40649,7 @@ var ClaudeAccountPool = class extends AccountPoolCore {
 				const subscriptionType = account.subscriptionType ?? credentials.subscriptionType;
 				const adopted = account.adopted === true;
 				const expired = adopted && adoptedCredentialExpired(credentials);
+				const quota = claudePoolQuota(cachedQuotaForPool(account.id, credentials));
 				return {
 					...base,
 					...email === void 0 ? {} : { email },
@@ -39294,6 +40661,7 @@ var ClaudeAccountPool = class extends AccountPoolCore {
 					source: account.source,
 					removable: !adopted,
 					...account.sourcePath === void 0 ? {} : { sourcePath: account.sourcePath },
+					...quota === void 0 ? {} : { quota },
 					...expired ? {
 						authStatus: "expired",
 						authFailedReason: ADOPTED_CREDENTIAL_EXPIRED_HINT
@@ -39860,7 +41228,7 @@ async function getClaudeWebStatus(store, modelSettings, preferences, options = {
 	const catalog = live.length > 0 ? live : FALLBACK_MODELS;
 	const available = catalog.map((model) => model.id);
 	const models = buildClaudeModelOptions(catalog, available, resolveEnabledModelIds(settings.enabledModelIds, available, enabled), settings.contextWindowOverrides);
-	const quota = credential === void 0 ? getCachedQuota$1() : getCachedQuota$1(credential);
+	const quota = credential === void 0 ? getCachedQuota$1() : cachedQuotaForPool(active.id, credential);
 	const poolData = options.accountPool === void 0 ? null : await options.accountPool.read().catch(() => null);
 	const poolAccounts = options.accountPool === void 0 ? [] : await options.accountPool.listAccounts().catch(() => []);
 	const presence = await claudeCodeCredentialPresence(options.adoptPaths ?? claudeCodeCredentialPaths()).catch(() => ({
@@ -39994,13 +41362,14 @@ function registerClaudeRoutes(ctx, store, modelSettings, preferences, options =
 				if (path === "" || path === "status") {
 					if (method !== "GET") return sendMethodNotAllowed(response);
 					const active = await activeAccount();
-					const cached = active.credentials === void 0 ? null : getCachedQuota$1(active.credentials);
+					const cached = active.credentials === void 0 ? null : cachedQuotaForPool(active.id, active.credentials);
 					const stale = cached === null || Date.now() - (cached.fetchedAt || 0) > 12e4;
 					if (active.credentials !== void 0 && stale) {
 						const credentials = active.credentials;
 						const refresh = () => fetchAccountQuota$1(credentials, {
 							fetchFn,
-							signal: quotaAbort.signal
+							signal: quotaAbort.signal,
+							accountId: active.id
 						});
 						if (cached === null) await quotaRefresh.run(refresh);
 						else quotaRefresh.start(refresh);
@@ -40122,7 +41491,8 @@ function registerClaudeRoutes(ctx, store, modelSettings, preferences, options =
 						await fetchAccountQuota$1(credentials, {
 							fetchFn,
 							force: method === "POST",
-							signal: quotaAbort.signal
+							signal: quotaAbort.signal,
+							accountId: active.id
 						});
 					} catch (error) {
 						quotaError = error instanceof Error ? error.message : String(error);
@@ -43738,39 +45108,24 @@ function apply(ctx, pluginConfig = {}) {
 			provider,
 			limits: providerLimits
 		})]));
-		const antigravityAdapter = new AntigravityAdapter(antigravityStore, antigravityModelSettings, antigravityPreferences, {
-			fetchFn: modelFetch.antigravity,
-			attachments: ctx.attachments
-		}, antigravityAccountPool);
-		let antigravityRegistration;
-		let antigravityConflict = null;
-		const claimAntigravityRoute = () => {
-			if (antigravityRegistration !== void 0) return;
-			try {
-				antigravityRegistration = ctx.llm.registerAdapter([PROVIDER_ID$6], antigravityAdapter);
-				if (antigravityConflict !== null) ctx.logger.info(`[dsh-chatgpt-subscription] Antigravity route "${PROVIDER_ID$6}" is now served by this plugin`);
-				antigravityConflict = null;
-			} catch (error) {
-				antigravityConflict = error instanceof Error ? error.message : String(error);
-				ctx.logger.warn(`[dsh-chatgpt-subscription] provider route "${PROVIDER_ID$6}" is already owned by another adapter; Antigravity models keep being served by that one until its configuration is removed (${antigravityConflict})`);
-			}
-		};
-		claimAntigravityRoute();
-		const antigravityRouteWatch = typeof ctx.on === "function" ? ctx.on("llm/adapters-updated", () => {
-			claimAntigravityRoute();
-		}) : void 0;
+		const antigravityClaim = claimProviderRoute(ctx, {
+			providerId: PROVIDER_ID$6,
+			label: PROVIDER_NAME$6,
+			adapter: new AntigravityAdapter(antigravityStore, antigravityModelSettings, antigravityPreferences, {
+				fetchFn: modelFetch.antigravity,
+				attachments: ctx.attachments
+			}, antigravityAccountPool)
+		});
 		const disposeAntigravityRoutes = registerAntigravityRoutes(ctx, antigravityStore, antigravityModelSettings, antigravityPreferences, proxyFetch, antigravityAccountPool);
 		const disposeCommandCodeRoutes = registerCommandCodeRoutes(ctx, commandCodeStore, commandCodeModelSettings, commandCodePreferences, {
 			fetchFn: proxyFetch,
-			serving: () => commandCodeRegistration !== void 0,
-			conflict: () => commandCodeConflict
+			serving: () => commandCodeClaim.serving(),
+			conflict: () => commandCodeClaim.conflict()
 		}, commandCodeAccountPool);
 		const commandCodeAdapter = new CommandCodeAdapter(commandCodeStore, commandCodeModelSettings, commandCodePreferences, {
 			fetchFn: modelFetch["command-code"],
 			attachments: ctx.attachments
 		}, commandCodeAccountPool);
-		let commandCodeRegistration;
-		let commandCodeConflict = null;
 		const ollamaStore = new FileCredentialStore$5();
 		const ollamaModelSettings = new FileModelSettingsStore$5();
 		const ollamaAccountPool = new OllamaAccountPool({ store: ollamaStore });
@@ -43778,92 +45133,48 @@ function apply(ctx, pluginConfig = {}) {
 			fetchFn: modelFetch.ollama,
 			attachments: ctx.attachments
 		}, ollamaAccountPool);
-		let ollamaRegistration;
-		let ollamaConflict = null;
-		const kimiCodeAdapter = new KimiCodeAdapter(kimiCodeStore, kimiCodeModelSettings, kimiCodePreferences, {
-			fetchFn: modelFetch["kimi-code"],
-			attachments: ctx.attachments,
-			videos: { readVideo: async (ref) => ({
-				data: await readVideoBytes(ref),
-				mediaType: ref.mediaType
-			}) }
-		}, kimiCodeAccountPool);
-		let kimiCodeRegistration;
-		let kimiCodeConflict = null;
-		const claimKimiCodeRoute = () => {
-			if (kimiCodeRegistration !== void 0) return;
-			try {
-				kimiCodeRegistration = ctx.llm.registerAdapter([PROVIDER_ID$3], kimiCodeAdapter);
-				if (kimiCodeConflict !== null) ctx.logger.info(`[dsh-chatgpt-subscription] ${PROVIDER_NAME$3} route "${PROVIDER_ID$3}" is now served by this plugin`);
-				kimiCodeConflict = null;
-			} catch (error) {
-				kimiCodeConflict = error instanceof Error ? error.message : String(error);
-				ctx.logger.warn(`[dsh-chatgpt-subscription] provider route "${PROVIDER_ID$3}" is already owned by another adapter; ${PROVIDER_NAME$3} models keep being served by that one until its configuration is removed (${kimiCodeConflict})`);
-			}
-		};
-		claimKimiCodeRoute();
-		const kimiCodeRouteWatch = typeof ctx.on === "function" ? ctx.on("llm/adapters-updated", () => {
-			claimKimiCodeRoute();
-		}) : void 0;
-		const claimOllamaRoute = () => {
-			if (ollamaRegistration !== void 0) return;
-			try {
-				ollamaRegistration = ctx.llm.registerAdapter([PROVIDER_ID$4], ollamaAdapter);
-				if (ollamaConflict !== null) ctx.logger.info(`[dsh-chatgpt-subscription] ${PROVIDER_NAME$4} route "${PROVIDER_ID$4}" is now served by this plugin`);
-				ollamaConflict = null;
-			} catch (error) {
-				ollamaConflict = error instanceof Error ? error.message : String(error);
-				ctx.logger.warn(`[dsh-chatgpt-subscription] provider route "${PROVIDER_ID$4}" is already owned by another adapter; ${PROVIDER_NAME$4} models keep being served by that one until its configuration is removed (${ollamaConflict})`);
-			}
-		};
-		claimOllamaRoute();
+		const kimiCodeClaim = claimProviderRoute(ctx, {
+			providerId: PROVIDER_ID$3,
+			label: PROVIDER_NAME$3,
+			adapter: new KimiCodeAdapter(kimiCodeStore, kimiCodeModelSettings, kimiCodePreferences, {
+				fetchFn: modelFetch["kimi-code"],
+				attachments: ctx.attachments,
+				videos: { readVideo: async (ref) => ({
+					data: await readVideoBytes(ref),
+					mediaType: ref.mediaType
+				}) }
+			}, kimiCodeAccountPool)
+		});
+		const ollamaClaim = claimProviderRoute(ctx, {
+			providerId: PROVIDER_ID$4,
+			label: PROVIDER_NAME$4,
+			adapter: ollamaAdapter,
+			watch: false
+		});
 		const disposeOllamaRoutes = registerOllamaRoutes(ctx, {
 			accountPool: ollamaAccountPool,
 			modelSettings: ollamaModelSettings,
 			fetchFn: proxyFetch
 		});
-		const claimCommandCodeRoute = () => {
-			if (commandCodeRegistration !== void 0) return;
-			try {
-				commandCodeRegistration = ctx.llm.registerAdapter([PROVIDER_ID$5], commandCodeAdapter);
-				if (commandCodeConflict !== null) ctx.logger.info(`[dsh-chatgpt-subscription] ${PROVIDER_NAME$5} route "${PROVIDER_ID$5}" is now served by this plugin`);
-				commandCodeConflict = null;
-			} catch (error) {
-				commandCodeConflict = error instanceof Error ? error.message : String(error);
-				ctx.logger.warn(`[dsh-chatgpt-subscription] provider route "${PROVIDER_ID$5}" is already owned by another adapter; ${PROVIDER_NAME$5} models keep being served by that one until its configuration is removed (${commandCodeConflict})`);
-			}
-		};
-		claimCommandCodeRoute();
-		const commandCodeRouteWatch = typeof ctx.on === "function" ? ctx.on("llm/adapters-updated", () => {
-			claimCommandCodeRoute();
-		}) : void 0;
+		const commandCodeClaim = claimProviderRoute(ctx, {
+			providerId: PROVIDER_ID$5,
+			label: PROVIDER_NAME$5,
+			adapter: commandCodeAdapter
+		});
 		const disposeKimiCodeRoutes = registerKimiCodeRoutes(ctx, kimiCodeStore, kimiCodeModelSettings, kimiCodePreferences, {
 			fetchFn: proxyFetch,
-			serving: () => kimiCodeRegistration !== void 0,
-			conflict: () => kimiCodeConflict
+			serving: () => kimiCodeClaim.serving(),
+			conflict: () => kimiCodeClaim.conflict()
 		}, kimiCodeAccountPool);
-		const minimaxCodeAdapter = new MinimaxCodeAdapter(minimaxCodeStore, {
-			fetchFn: modelFetch["minimax-code"],
-			attachments: ctx.attachments,
-			accountPool: minimaxCodeAccountPool
-		}, minimaxCodeModelSettings, minimaxCodePreferences);
-		let minimaxCodeRegistration;
-		let minimaxCodeConflict = null;
-		const claimMinimaxCodeRoute = () => {
-			if (minimaxCodeRegistration !== void 0) return;
-			try {
-				minimaxCodeRegistration = ctx.llm.registerAdapter([PROVIDER_ID$2], minimaxCodeAdapter);
-				if (minimaxCodeConflict !== null) ctx.logger.info(`[dsh-chatgpt-subscription] ${PROVIDER_NAME$2} route "${PROVIDER_ID$2}" is now served by this plugin`);
-				minimaxCodeConflict = null;
-			} catch (error) {
-				minimaxCodeConflict = error instanceof Error ? error.message : String(error);
-				ctx.logger.warn(`[dsh-chatgpt-subscription] provider route "${PROVIDER_ID$2}" is already owned by another adapter; ${PROVIDER_NAME$2} models keep being served by that one until its configuration is removed (${minimaxCodeConflict})`);
-			}
-		};
-		claimMinimaxCodeRoute();
-		const minimaxCodeRouteWatch = typeof ctx.on === "function" ? ctx.on("llm/adapters-updated", () => {
-			claimMinimaxCodeRoute();
-		}) : void 0;
+		const minimaxCodeClaim = claimProviderRoute(ctx, {
+			providerId: PROVIDER_ID$2,
+			label: PROVIDER_NAME$2,
+			adapter: new MinimaxCodeAdapter(minimaxCodeStore, {
+				fetchFn: modelFetch["minimax-code"],
+				attachments: ctx.attachments,
+				accountPool: minimaxCodeAccountPool
+			}, minimaxCodeModelSettings, minimaxCodePreferences)
+		});
 		let minimaxCodeAppVersion;
 		try {
 			const manifest = createRequire(import.meta.url)("../package.json");
@@ -43887,33 +45198,20 @@ function apply(ctx, pluginConfig = {}) {
 		minimaxCodeCheckinTick();
 		const disposeMinimaxCodeRoutes = registerMinimaxCodeRoutes(ctx, minimaxCodeStore, {
 			fetchFn: proxyFetch,
-			serving: () => minimaxCodeRegistration !== void 0,
-			conflict: () => minimaxCodeConflict,
+			serving: () => minimaxCodeClaim.serving(),
+			conflict: () => minimaxCodeClaim.conflict(),
 			accountPool: minimaxCodeAccountPool,
 			checkin: minimaxCodeCheckin
 		}, minimaxCodeModelSettings, minimaxCodePreferences);
-		const workBuddyAdapter = new WorkBuddyAdapter(workBuddyStore, workBuddyModelSettings, workBuddyPreferences, {
-			fetchFn: modelFetch.workbuddy,
-			attachments: ctx.attachments,
-			accountPool: workBuddyAccountPool
+		const workBuddyClaim = claimProviderRoute(ctx, {
+			providerId: PROVIDER_ID$1,
+			label: PROVIDER_NAME$1,
+			adapter: new WorkBuddyAdapter(workBuddyStore, workBuddyModelSettings, workBuddyPreferences, {
+				fetchFn: modelFetch.workbuddy,
+				attachments: ctx.attachments,
+				accountPool: workBuddyAccountPool
+			})
 		});
-		let workBuddyRegistration;
-		let workBuddyConflict = null;
-		const claimWorkBuddyRoute = () => {
-			if (workBuddyRegistration !== void 0) return;
-			try {
-				workBuddyRegistration = ctx.llm.registerAdapter([PROVIDER_ID$1], workBuddyAdapter);
-				if (workBuddyConflict !== null) ctx.logger.info(`[dsh-chatgpt-subscription] ${PROVIDER_NAME$1} route "${PROVIDER_ID$1}" is now served by this plugin`);
-				workBuddyConflict = null;
-			} catch (error) {
-				workBuddyConflict = error instanceof Error ? error.message : String(error);
-				ctx.logger.warn(`[dsh-chatgpt-subscription] provider route "${PROVIDER_ID$1}" is already owned by another adapter; ${PROVIDER_NAME$1} models keep being served by that one until its configuration is removed (${workBuddyConflict})`);
-			}
-		};
-		claimWorkBuddyRoute();
-		const workBuddyRouteWatch = typeof ctx.on === "function" ? ctx.on("llm/adapters-updated", () => {
-			claimWorkBuddyRoute();
-		}) : void 0;
 		const workBuddyCheckin = new WorkBuddyCheckinService(workBuddyStore, {
 			fetchFn: proxyFetch,
 			settings: () => workBuddyPreferences.status().checkin,
@@ -43929,37 +45227,24 @@ function apply(ctx, pluginConfig = {}) {
 		checkinTick();
 		const disposeWorkBuddyRoutes = registerWorkBuddyRoutes(ctx, workBuddyStore, workBuddyModelSettings, workBuddyPreferences, {
 			fetchFn: proxyFetch,
-			serving: () => workBuddyRegistration !== void 0,
-			conflict: () => workBuddyConflict,
+			serving: () => workBuddyClaim.serving(),
+			conflict: () => workBuddyClaim.conflict(),
 			accountPool: workBuddyAccountPool,
 			checkin: workBuddyCheckin
 		});
-		const claudeAdapter = new ClaudeAdapter(claudeStore, claudeModelSettings, claudePreferences, {
-			fetchFn: modelFetch.claude,
-			attachments: ctx.attachments,
-			accountPool: claudeAccountPool
+		const claudeClaim = claimProviderRoute(ctx, {
+			providerId: PROVIDER_ID,
+			label: PROVIDER_NAME,
+			adapter: new ClaudeAdapter(claudeStore, claudeModelSettings, claudePreferences, {
+				fetchFn: modelFetch.claude,
+				attachments: ctx.attachments,
+				accountPool: claudeAccountPool
+			})
 		});
-		let claudeRegistration;
-		let claudeConflict = null;
-		const claimClaudeRoute = () => {
-			if (claudeRegistration !== void 0) return;
-			try {
-				claudeRegistration = ctx.llm.registerAdapter([PROVIDER_ID], claudeAdapter);
-				if (claudeConflict !== null) ctx.logger.info(`[dsh-chatgpt-subscription] ${PROVIDER_NAME} route "${PROVIDER_ID}" is now served by this plugin`);
-				claudeConflict = null;
-			} catch (error) {
-				claudeConflict = error instanceof Error ? error.message : String(error);
-				ctx.logger.warn(`[dsh-chatgpt-subscription] provider route "${PROVIDER_ID}" is already owned by another adapter; ${PROVIDER_NAME} models keep being served by that one until its configuration is removed (${claudeConflict})`);
-			}
-		};
-		claimClaudeRoute();
-		const claudeRouteWatch = typeof ctx.on === "function" ? ctx.on("llm/adapters-updated", () => {
-			claimClaudeRoute();
-		}) : void 0;
 		const disposeClaudeRoutes = registerClaudeRoutes(ctx, claudeStore, claudeModelSettings, claudePreferences, {
 			fetchFn: proxyFetch,
-			serving: () => claudeRegistration !== void 0,
-			conflict: () => claudeConflict,
+			serving: () => claudeClaim.serving(),
+			conflict: () => claudeClaim.conflict(),
 			accountPool: claudeAccountPool
 		});
 		workBuddyAccountPool.syncDesktopAccounts().catch(() => void 0);
@@ -43970,7 +45255,11 @@ function apply(ctx, pluginConfig = {}) {
 			pool: codexAccountPool
 		});
 		const usage = new UsageService(oauth, { fetchFn: proxyFetch });
-		codexAccountPool.setQuotaBlockedUntil((account, now) => usage.blockedUntilFor(account.credentials, now));
+		codexAccountPool.setQuotaBlockedUntil((account, now) => usage.blockedUntilFor(account.id, account.credentials, now));
+		codexAccountPool.setQuotaSnapshot((account) => {
+			const snapshot = usage.snapshotFor(account.id, account.credentials);
+			return snapshot === void 0 ? void 0 : codexAccountQuota(snapshot);
+		});
 		const adapter = new CodexChatGptAdapter(new ResponsesClient(oauth, ctx.attachments, {
 			fetchFn: modelFetch["codex-chatgpt"],
 			accountPool: codexAccountPool,
@@ -44026,6 +45315,145 @@ function apply(ctx, pluginConfig = {}) {
 			});
 		};
 		const disposeRoutes = registerRoutes(ctx, oauth, usage, preferences, proxyManager, searchSwitcher, readRouteAudit, codexAccountPool, fetchConfiguration);
+		const routable = (accounts) => accounts.some((account) => account.authStatus === void 0 || account.authStatus === "ok");
+		const disposeHubOverview = registerHubOverviewRoutes(ctx, [
+			{
+				id: "chatgpt",
+				providerId: PROVIDER_ID$7,
+				canToggle: true,
+				read: async () => {
+					const current = preferences.status();
+					const accounts = await codexAccountPool.listAccounts().catch(() => []);
+					return {
+						enabled: current.enabled !== false,
+						accountCount: accounts.length,
+						authenticated: routable(accounts),
+						enabledModelCount: current.visibleModelIds.length,
+						totalModelCount: CODEX_MODEL_CATALOG.length
+					};
+				}
+			},
+			{
+				id: "antigravity",
+				providerId: PROVIDER_ID$6,
+				canToggle: true,
+				read: async () => {
+					const current = antigravityPreferences.status();
+					const accounts = await antigravityAccountPool.listAccounts().catch(() => []);
+					const total = catalogTotal((await antigravityModelSettings.read()).catalogModels.length);
+					return {
+						enabled: current.enabled !== false,
+						accountCount: accounts.length,
+						authenticated: routable(accounts),
+						enabledModelCount: total === null ? null : Math.min(current.enabledModelIds?.length ?? 0, total),
+						totalModelCount: total
+					};
+				}
+			},
+			{
+				id: "command-code",
+				providerId: PROVIDER_ID$5,
+				canToggle: true,
+				read: async () => {
+					const current = commandCodePreferences.status();
+					const accounts = await commandCodeAccountPool.listAccounts().catch(() => []);
+					const total = catalogTotal((await commandCodeModelSettings.read()).catalogModels.length);
+					return {
+						enabled: current.enabled !== false,
+						accountCount: accounts.length,
+						authenticated: routable(accounts),
+						enabledModelCount: total === null ? null : Math.min(current.enabledModelIds?.length ?? 0, total),
+						totalModelCount: total
+					};
+				}
+			},
+			{
+				id: "kimi-code",
+				providerId: PROVIDER_ID$3,
+				canToggle: true,
+				read: async () => {
+					const current = kimiCodePreferences.status();
+					const accounts = await kimiCodeAccountPool.listAccounts().catch(() => []);
+					const total = catalogTotal((await kimiCodeModelSettings.read()).catalogModels.length);
+					return {
+						enabled: current.enabled !== false,
+						accountCount: accounts.length,
+						authenticated: routable(accounts),
+						enabledModelCount: total === null ? null : Math.min(current.enabledModelIds?.length ?? 0, total),
+						totalModelCount: total
+					};
+				}
+			},
+			{
+				id: "workbuddy",
+				providerId: PROVIDER_ID$1,
+				canToggle: true,
+				read: async () => {
+					const current = workBuddyPreferences.status();
+					const accounts = await workBuddyAccountPool.listAccounts().catch(() => []);
+					const total = catalogTotal(getCachedCatalog$1().length);
+					return {
+						enabled: current.enabled !== false,
+						accountCount: accounts.length,
+						authenticated: routable(accounts),
+						enabledModelCount: total === null ? null : Math.min(current.enabledModelIds?.length ?? 0, total),
+						totalModelCount: total
+					};
+				}
+			},
+			{
+				id: "minimax-code",
+				providerId: PROVIDER_ID$2,
+				canToggle: true,
+				read: async () => {
+					const current = minimaxCodePreferences.status();
+					const accounts = await minimaxCodeAccountPool.listAccounts().catch(() => []);
+					const total = catalogTotal(catalogSize());
+					return {
+						enabled: current.enabled !== false,
+						accountCount: accounts.length,
+						authenticated: routable(accounts),
+						enabledModelCount: total === null ? null : Math.min(current.enabledModelIds?.length ?? 0, total),
+						totalModelCount: total
+					};
+				}
+			},
+			{
+				id: "claude",
+				providerId: PROVIDER_ID,
+				canToggle: true,
+				read: async () => {
+					const current = claudePreferences.status();
+					const accounts = await claudeAccountPool.listAccounts().catch(() => []);
+					const total = catalogTotal(getCachedCatalog().length);
+					return {
+						enabled: current.enabled !== false,
+						accountCount: accounts.length,
+						authenticated: routable(accounts),
+						enabledModelCount: total === null ? null : Math.min(current.enabledModelIds?.length ?? 0, total),
+						totalModelCount: total
+					};
+				}
+			},
+			{
+				id: "ollama",
+				providerId: PROVIDER_ID$4,
+				canToggle: false,
+				read: async () => {
+					const accounts = await ollamaAccountPool.listAccounts().catch(() => []);
+					const current = ollamaModelSettings.status();
+					const total = catalogTotal(current.catalogModels.length);
+					const enabled = current.enabledModelIds.length === 0 ? total : current.enabledModelIds.length;
+					return {
+						enabled: true,
+						accountCount: accounts.length,
+						authenticated: routable(accounts),
+						enabledModelCount: total === null ? null : Math.min(enabled ?? 0, total),
+						totalModelCount: total
+					};
+				}
+			}
+		]);
 		const disposeAdapter = ctx.llm.registerAdapter([PROVIDER_ID$7], adapter);
 		const disposeImageTool = ctx.tools.register(createCodexImageTool(oauth, ctx.attachments, { fetchFn: proxyFetch }));
 		const disposeVideoTool = ctx.tools.register(createKimiVideoTool(ctx, { fetchFn: proxyFetch }));
@@ -44052,48 +45480,28 @@ function apply(ctx, pluginConfig = {}) {
 			disposeVideoTool();
 			disposeAdapter();
 			disposeRoutes();
+			disposeHubOverview();
 			disposeAntigravityRoutes();
-			releaseHandle(antigravityRouteWatch);
-			antigravityRegistration?.();
-			antigravityRegistration = void 0;
+			antigravityClaim.dispose();
 			disposeOllamaRoutes();
-			ollamaRegistration?.();
-			ollamaRegistration = void 0;
+			ollamaClaim.dispose();
 			disposeCommandCodeRoutes();
-			releaseHandle(commandCodeRouteWatch);
-			commandCodeRegistration?.();
-			commandCodeRegistration = void 0;
+			commandCodeClaim.dispose();
 			disposeKimiCodeRoutes();
-			releaseHandle(kimiCodeRouteWatch);
-			kimiCodeRegistration?.();
-			kimiCodeRegistration = void 0;
+			kimiCodeClaim.dispose();
 			disposeMinimaxCodeRoutes();
-			releaseHandle(minimaxCodeRouteWatch);
-			minimaxCodeRegistration?.();
-			minimaxCodeRegistration = void 0;
+			minimaxCodeClaim.dispose();
 			clearInterval(minimaxCodeCheckinTimer);
 			clearInterval(checkinTimer);
 			disposeWorkBuddyRoutes();
-			releaseHandle(workBuddyRouteWatch);
-			workBuddyRegistration?.();
-			workBuddyRegistration = void 0;
+			workBuddyClaim.dispose();
 			disposeClaudeRoutes();
-			releaseHandle(claudeRouteWatch);
-			claudeRegistration?.();
-			claudeRegistration = void 0;
+			claudeClaim.dispose();
 			oauth.dispose();
 			proxyManager.dispose();
 		};
 	}, "dsh-chatgpt-subscription: adapter, routes, and lifecycle");
 }
-/** Cordis event handles are either a disposer function or a disposable object. */
-function releaseHandle(handle) {
-	if (typeof handle === "function") {
-		handle();
-		return;
-	}
-	handle?.dispose?.();
-}
 function localWebServerBaseUrl(host, port) {
 	return `http://${host === "0.0.0.0" ? "127.0.0.1" : host}:${port}`;
 }