dsh-agy-link 0.4.67 → 0.4.74

package/dist/index.js

@@ -1,13 +1,13 @@
 import { createRequire } from "node:module";
 import { execFile, execFileSync, spawn } from "node:child_process";
-import path, { delimiter, dirname, extname, isAbsolute, join, relative, resolve } from "node:path";
+import path, { basename, delimiter, dirname, extname, isAbsolute, join, relative, resolve } from "node:path";
 import { accessSync, chmodSync, constants, copyFileSync, createReadStream, existsSync, mkdirSync, readFileSync, readdirSync, renameSync, rmSync, statSync, unlinkSync, utimesSync, writeFileSync } from "node:fs";
 import { homedir, tmpdir } from "node:os";
 import * as dshLlm from "@deepseek-ai/dsh-llm";
 import { LlmAdapter, LlmError } from "@deepseek-ai/dsh-llm";
 import { createHash, randomBytes, randomUUID } from "node:crypto";
 import { defineTool } from "@deepseek-ai/dsh-tools";
-import { chmod, copyFile, mkdir, mkdtemp, readFile, readdir, rm, stat, unlink, writeFile } from "node:fs/promises";
+import { mkdir, mkdtemp, readFile, readdir, rm, stat, unlink, writeFile } from "node:fs/promises";
 import { EventEmitter } from "node:events";
 import { createServer } from "node:http";
 import { promisify } from "node:util";
@@ -130,14 +130,19 @@ function defaultConfig() {
 		mediaMaxBytes: 10485760,
 		mediaMaxImages: 8,
 		mcpBridge: false,
+		mcpConfigScope: "global",
+		mcpConfigPath: "",
 		mcpToolAllowlist: "",
 		rateLimitPerMinute: 0,
 		autoFallbackModel: false,
 		logRetentionDays: 7,
 		disableTelemetry: true,
-		quotaPollIntervalMs: 9e5,
+		quotaPollIntervalMs: 36e5,
+		quotaViaCli: true,
 		proxyUrl: "",
-		suppressAgyAutoUpdate: process.platform === "win32"
+		suppressAgyAutoUpdate: process.platform === "win32",
+		autoWarmup: false,
+		warmupModels: ["gemini-3.8-flash", "claude-sonnet-4-6"]
 	};
 }
 /**
@@ -165,6 +170,60 @@ function looksLikeNetworkError(message) {
 	return /dial tcp|fetch failed|econn(?:refused|reset|aborted)|enotfound|ehostunreach|enetunreach|socket hang up|getaddrinfo|bad gateway|service unavailable|unexpected eof|\beof\b|tls handshake timeout|i\/o timeout|network (?:error|is unreachable)|no route to host|connection reset|context deadline exceeded|temporary failure in name resolution|proxyconnect|token validation failed|validate (?:the )?token|silent auth failed/i.test(message);
 }
 /**
+* Backoff schedule for CONSECUTIVE transport (NETWORK) failures.
+*
+* DSH's generic backoff starts at 2s and doubles, which spends the whole retry
+* budget inside a few seconds. A real egress outage does not look like that:
+* on 2026-10-21 the machine's proxy answered CONNECT 200 and then killed TLS
+* for ~14 minutes, so every agy run failed its startup eligibility probe with
+* `Post "…/loadCodeAssist": EOF` and the turn died after two retries in 4s
+* while agy itself kept working whenever a gap appeared.
+*
+* These delays are stated per failure as `providerRetryAfterMs`, which the
+* retry plugin honours verbatim — but only up to the policy's maxDelayMs (a
+* longer request makes it drop the retry entirely), so the cap must stay above
+* the last entry. The schedule is capped, never unbounded: a dead egress still
+* ends the turn, just after minutes instead of seconds.
+*/
+const NETWORK_RETRY_DELAYS_MS = [
+	5e3,
+	15e3,
+	3e4
+];
+/** Delay to request after `consecutiveFailures` transport failures in a row. */
+function networkRetryDelayMs(consecutiveFailures) {
+	const index = Math.min(Math.max(Math.trunc(consecutiveFailures) - 1, 0), NETWORK_RETRY_DELAYS_MS.length - 1);
+	return NETWORK_RETRY_DELAYS_MS[index];
+}
+/**
+* Whether every ENABLED account leaves through the same proxy.
+*
+* When it does, a transport failure is a property of the machine's egress, not
+* of the account: rotating the pool repoints the session for nothing and burns
+* the rotation budget that a real quota/eligibility hop needs. Accounts with
+* their own `proxyUrl` (or one with an override next to the default) can fail
+* independently, so rotation stays meaningful there.
+*
+* `globalProxy` is the proxy an account without an override uses (config
+* `proxyUrl`, else the detected system proxy). It matters because a newly
+* created pool account is seeded with a COPY of the system proxy while the
+* primary/system account keeps an empty override — comparing raw overrides
+* would call that "two different egresses" and rotate on every blip.
+*
+* Disabled accounts are ignored: they are never selected, so their proxy cannot
+* make the reachable egress distinct.
+*/
+function accountsShareSingleEgress(accounts, globalProxy) {
+	const fallback = (globalProxy ?? "").trim();
+	const urls = /* @__PURE__ */ new Set();
+	for (const account of accounts) {
+		if (account.enabled === false) continue;
+		urls.add((account.proxyUrl ?? "").trim() || fallback);
+		if (urls.size > 1) return false;
+	}
+	return true;
+}
+/**
 * Evidence from agy's OWN per-run log that a stored credential was present but
 * could not be VALIDATED over the network — i.e. jitter wearing an auth costume.
 *
@@ -383,6 +442,10 @@ function readOverrides(file = overridesPath()) {
 function asString(v) {
 	return typeof v === "string" ? v : void 0;
 }
+/** `mcpConfigScope` is a closed enum: unknown values fall back to the default. */
+function asMcpConfigScope(v) {
+	return v === "global" || v === "workspace" || v === "both" ? v : void 0;
+}
 function asBool(v) {
 	if (typeof v === "boolean") return v;
 	if (typeof v === "string") return v === "true" || v === "1";
@@ -451,14 +514,19 @@ function resolveConfig(entry, env = process.env, overrides = readOverrides()) {
 		mediaMaxBytes: asNum(get("mediaMaxBytes")) ?? base.mediaMaxBytes,
 		mediaMaxImages: asNum(get("mediaMaxImages")) ?? base.mediaMaxImages,
 		mcpBridge: asBool(get("mcpBridge")) ?? base.mcpBridge,
+		mcpConfigScope: asMcpConfigScope(get("mcpConfigScope")) ?? base.mcpConfigScope,
+		mcpConfigPath: asString(get("mcpConfigPath")) ?? base.mcpConfigPath,
 		mcpToolAllowlist: asString(get("mcpToolAllowlist")) ?? base.mcpToolAllowlist,
 		rateLimitPerMinute: asNum(get("rateLimitPerMinute")) ?? base.rateLimitPerMinute,
 		autoFallbackModel: asBool(get("autoFallbackModel")) ?? base.autoFallbackModel,
 		logRetentionDays: asNum(get("logRetentionDays")) ?? base.logRetentionDays,
 		disableTelemetry: asBool(get("disableTelemetry")) ?? base.disableTelemetry,
 		quotaPollIntervalMs: asNum(get("quotaPollIntervalMs")) ?? base.quotaPollIntervalMs,
+		quotaViaCli: asBool(get("quotaViaCli")) ?? base.quotaViaCli,
 		proxyUrl: asString(get("proxyUrl")) ?? base.proxyUrl,
-		suppressAgyAutoUpdate: asBool(get("suppressAgyAutoUpdate")) ?? base.suppressAgyAutoUpdate
+		suppressAgyAutoUpdate: asBool(get("suppressAgyAutoUpdate")) ?? base.suppressAgyAutoUpdate,
+		autoWarmup: asBool(get("autoWarmup")) ?? base.autoWarmup,
+		warmupModels: asStringArray(get("warmupModels")) ?? base.warmupModels
 	};
 	if (env.DSH_AGY_ENABLED !== void 0) cfg.enabled = asBool(env.DSH_AGY_ENABLED) ?? cfg.enabled;
 	if (env.DSH_AGY_BIN) cfg.agyBin = env.DSH_AGY_BIN;
@@ -496,6 +564,11 @@ function resolveConfig(entry, env = process.env, overrides = readOverrides()) {
 		if (b !== void 0) cfg.mcpBridge = b;
 	}
 	if (env.DSH_AGY_MCP_TOOL_ALLOWLIST) cfg.mcpToolAllowlist = env.DSH_AGY_MCP_TOOL_ALLOWLIST;
+	if (env.DSH_AGY_MCP_CONFIG_PATH) cfg.mcpConfigPath = env.DSH_AGY_MCP_CONFIG_PATH;
+	if (env.DSH_AGY_MCP_CONFIG_SCOPE !== void 0) {
+		const scope = asMcpConfigScope(env.DSH_AGY_MCP_CONFIG_SCOPE);
+		if (scope !== void 0) cfg.mcpConfigScope = scope;
+	}
 	if (env.DSH_AGY_RATE_LIMIT_PER_MINUTE) {
 		const r = asNum(env.DSH_AGY_RATE_LIMIT_PER_MINUTE);
 		if (r !== void 0 && r >= 0) cfg.rateLimitPerMinute = r;
@@ -516,11 +589,23 @@ function resolveConfig(entry, env = process.env, overrides = readOverrides()) {
 		const q = asNum(env.DSH_AGY_QUOTA_POLL_INTERVAL_MS);
 		if (q && q >= 6e4) cfg.quotaPollIntervalMs = q;
 	}
+	if (env.DSH_AGY_QUOTA_VIA_CLI !== void 0) {
+		const qvc = asBool(env.DSH_AGY_QUOTA_VIA_CLI);
+		if (qvc !== void 0) cfg.quotaViaCli = qvc;
+	}
 	if (env.DSH_AGY_PROXY_URL) cfg.proxyUrl = env.DSH_AGY_PROXY_URL;
 	if (env.DSH_AGY_SUPPRESS_AUTOUPDATE !== void 0) {
 		const sa = asBool(env.DSH_AGY_SUPPRESS_AUTOUPDATE);
 		if (sa !== void 0) cfg.suppressAgyAutoUpdate = sa;
 	}
+	if (env.DSH_AGY_AUTO_WARMUP !== void 0) {
+		const aw = asBool(env.DSH_AGY_AUTO_WARMUP);
+		if (aw !== void 0) cfg.autoWarmup = aw;
+	}
+	if (env.DSH_AGY_WARMUP_MODELS) {
+		const wm = asStringArray(env.DSH_AGY_WARMUP_MODELS.split(","));
+		if (wm) cfg.warmupModels = wm;
+	}
 	return cfg;
 }
 //#endregion
@@ -1989,7 +2074,7 @@ var EventMapper = class {
 	* context occupancy, so a failed turn used to reset the meter and could fire
 	* a bogus compaction a step later.
 	*/
-	*emitFailure(kind, code, message) {
+	*emitFailure(kind, code, message, retryAfterMs) {
 		if (this.finished) return;
 		const close = this.closeOpen();
 		if (close) yield close;
@@ -2001,20 +2086,19 @@ var EventMapper = class {
 				outputTokens: 0
 			}
 		};
+		const failure = {
+			message,
+			code,
+			...retryAfterMs !== void 0 && Number.isFinite(retryAfterMs) && retryAfterMs > 0 ? { providerRetryAfterMs: Math.trunc(retryAfterMs) } : {}
+		};
 		yield {
 			type: "finish",
 			reason: kind === "error" ? {
 				kind: "error",
-				failure: {
-					message,
-					code
-				}
+				failure
 			} : {
 				kind: "aborted",
-				failure: {
-					message,
-					code
-				}
+				failure
 			}
 		};
 		this.finished = true;
@@ -2022,6 +2106,24 @@ var EventMapper = class {
 };
 //#endregion
 //#region src/host/models.ts
+/** Check if a model identifier belongs to the Gemini 4 Argon family. */
+function isArgonModel(id) {
+	return id.trim().toLowerCase().includes("argon");
+}
+/**
+* Determine whether a managed account has confirmed access to the Gemini 4 Argon family.
+* Argon is exclusive to paid Pro/Ultra plans and returns a separate quota bucket
+* and/or dedicated model list from agy cli. Free/standard accounts do not have Argon.
+*/
+function accountHasArgon(acc, pool, catalog) {
+	if (!acc) return false;
+	if (acc.quotas?.argon !== void 0) return true;
+	if (pool && typeof pool.hasDedicatedGroup === "function" && pool.hasDedicatedGroup(acc.id, "argon")) return true;
+	if (acc.quotas?.google?.models?.some((m) => isArgonModel(m.modelId))) return true;
+	if (acc.quotas?.google !== void 0) return false;
+	if (catalog && catalog.source === "discovered") return catalog.models.some((m) => isArgonModel(m.id));
+	return false;
+}
 /** Parse `agy models` stdout: JSON shapes first, then two-column text. */
 function parseModelsOutput(stdout) {
 	const text = stdout.trim();
@@ -2207,6 +2309,7 @@ var ModelCatalog = class ModelCatalog {
 	ttlMs;
 	onUpdate;
 	current;
+	accountCatalogs = /* @__PURE__ */ new Map();
 	refreshing = null;
 	/**
 	* While the catalog is still on the fallback source, re-probe at most once
@@ -2226,46 +2329,59 @@ var ModelCatalog = class ModelCatalog {
 			discoveredAt: 0
 		};
 	}
-	get() {
+	get(accountId) {
+		if (accountId && this.accountCatalogs.has(accountId)) return this.accountCatalogs.get(accountId);
 		return this.current;
 	}
+	hasAccountCatalog(accountId) {
+		return this.accountCatalogs.has(accountId);
+	}
 	/** Refresh if stale; never throws — failures keep the previous catalog. */
-	async refreshIfNeeded() {
+	async refreshIfNeeded(accountId) {
 		if (this.refreshing) return this.refreshing;
 		const now = Date.now();
-		const age = now - this.current.discoveredAt;
-		if (this.current.source === "discovered" && age < this.ttlMs) return;
-		if (this.current.source === "fallback" && now - this.lastAttemptAt < ModelCatalog.FALLBACK_RETRY_MS) return;
+		const accountCat = accountId ? this.accountCatalogs.get(accountId) : void 0;
+		if (!accountId || accountCat) {
+			const targetCat = accountCat ?? this.current;
+			const age = now - targetCat.discoveredAt;
+			if (targetCat.source === "discovered" && age < this.ttlMs) return;
+			if (targetCat.source === "fallback" && now - this.lastAttemptAt < ModelCatalog.FALLBACK_RETRY_MS) return;
+		}
 		this.lastAttemptAt = now;
-		this.refreshing = this.refresh().finally(() => {
+		this.refreshing = this.refresh(accountId).finally(() => {
 			this.refreshing = null;
 		});
 		return this.refreshing;
 	}
-	async forceRefresh() {
-		await this.refresh();
-		this.onUpdate?.();
-		return this.current;
+	async forceRefresh(accountId) {
+		await this.refresh(accountId);
+		return this.get(accountId);
 	}
-	async refresh() {
+	async refresh(accountId) {
 		try {
 			const ac = new AbortController();
 			const timer = setTimeout(() => ac.abort(), 3e4);
 			try {
-				const { stdout } = await this.discover(ac.signal);
+				const { stdout } = await this.discover(ac.signal, accountId);
 				const raw = parseModelsOutput(stdout);
 				if (raw.length > 0) {
 					const newModels = foldEfforts(raw);
-					const changed = this.current.source !== "discovered" || this.current.models.length !== newModels.length || this.current.models.some((m, i) => m.id !== newModels[i]?.id);
-					this.current = {
+					const newCatalog = {
 						source: "discovered",
 						models: newModels,
 						discoveredAt: Date.now()
 					};
+					const prev = this.get(accountId);
+					const changed = prev.source !== "discovered" || prev.models.length !== newModels.length || prev.models.some((m, i) => m.id !== newModels[i]?.id);
+					if (accountId) this.accountCatalogs.set(accountId, newCatalog);
+					else this.current = newCatalog;
 					if (changed) this.onUpdate?.();
 					return;
 				}
-				this.current = {
+				if (accountId) {
+					const cat = this.accountCatalogs.get(accountId);
+					if (cat) cat.lastError = "agy models returned no entries";
+				} else this.current = {
 					...this.current,
 					lastError: "agy models returned no entries"
 				};
@@ -2273,9 +2389,13 @@ var ModelCatalog = class ModelCatalog {
 				clearTimeout(timer);
 			}
 		} catch (err) {
-			this.current = {
+			const errMsg = err instanceof Error ? err.message : String(err);
+			if (accountId) {
+				const cat = this.accountCatalogs.get(accountId);
+				if (cat) cat.lastError = errMsg;
+			} else this.current = {
 				...this.current,
-				lastError: err instanceof Error ? err.message : String(err)
+				lastError: errMsg
 			};
 		}
 	}
@@ -3042,12 +3162,6 @@ function ensureIsolatedKeychain(dir, platform = process.platform, run = (file, a
 			kc
 		], env);
 		run("security", ["set-keychain-settings", kc], env);
-		run("security", [
-			"unlock-keychain",
-			"-p",
-			"",
-			kc
-		], env);
 		return true;
 	} catch {
 		return false;
@@ -3058,8 +3172,9 @@ function ensureIsolatedKeychain(dir, platform = process.platform, run = (file, a
 * Keychains from older versions (v0.4.41-v0.4.44) were provisioned with
 * ephemeral random passwords that cannot be unlocked with an empty password,
 * causing repeated modal password dialogs from macOS SecurityAgent.
-* If unlocking with `-p ""` fails, the file and its registration in securityd
-* are purged so that ensureIsolatedKeychain can cleanly re-provision it.
+* We test whether the keychain settings can be updated silently without prompt.
+* If setting settings fails, the file is un-unlockable or corrupted, so it is
+* unregistered and purged so ensureIsolatedKeychain can cleanly re-provision it.
 */
 function purgeStaleIsolatedKeychain(dir, platform = process.platform, run = (file, args, env) => execFileSync(file, args, {
 	stdio: "ignore",
@@ -3075,12 +3190,7 @@ function purgeStaleIsolatedKeychain(dir, platform = process.platform, run = (fil
 		HOME: dir
 	};
 	try {
-		run("security", [
-			"unlock-keychain",
-			"-p",
-			"",
-			kc
-		], env);
+		run("security", ["set-keychain-settings", kc], env);
 		return false;
 	} catch {
 		try {
@@ -3173,6 +3283,7 @@ function repairMacSystemKeychain(userHome = homedir(), platform = process.platfo
 function openAccountTerminal(accDir, platform = process.platform, run = (file, args, opts) => execFile(file, [...args], { windowsHide: opts?.windowsHide }, () => {})) {
 	if (!accDir) return false;
 	if (platform === "darwin") {
+		ensureIsolatedKeychain(accDir, platform);
 		const script = `tell application "Terminal" to activate\ntell application "Terminal" to do script "export HOME='${accDir}'; agy"`;
 		try {
 			run("osascript", ["-e", script]);
@@ -3698,7 +3809,7 @@ function runWinCredPs(target, user, action, payload = "") {
 		payload
 	], {
 		encoding: "utf8",
-		timeout: 8e3,
+		timeout: action === "write" ? 3e4 : 8e3,
 		windowsHide: true,
 		stdio: [
 			"ignore",
@@ -4118,6 +4229,7 @@ var AccountPoolManager = class {
 			cooldowns: {},
 			quotas: {}
 		};
+		if (process.platform === "darwin" && newAccount.dir) ensureIsolatedKeychain(newAccount.dir);
 		this.data.accounts.push(newAccount);
 		this.knownIds.add(newAccount.id);
 		this.persist();
@@ -4185,6 +4297,7 @@ var AccountPoolManager = class {
 		const id = `acc_${Date.now()}_${Math.random().toString(36).slice(2, 7)}`;
 		const dir = join(this.baseDir, id);
 		mkdirSync(join(dir, ".gemini", "antigravity-cli"), { recursive: true });
+		if (process.platform === "darwin") ensureIsolatedKeychain(dir);
 		const count = this.data.accounts.length + 1;
 		const newAccount = {
 			id,
@@ -4516,6 +4629,19 @@ var AccountPoolManager = class {
 		this.persist();
 	}
 	/**
+	* Drop one family's quota row.
+	*
+	* Used when stored numbers provably cannot be attributed to this account —
+	* `updateAccountQuotas` only merges, so an untrustworthy row would otherwise
+	* survive every refresh.
+	*/
+	clearQuotaFamily(id, family) {
+		const acc = this.getAccount(id);
+		if (!acc?.quotas || acc.quotas[family] === void 0) return;
+		delete acc.quotas[family];
+		this.persist();
+	}
+	/**
 	* Records a rate-limit / 429 error and sets cooldown for the requested model family.
 	*/
 	recordFailure(id, family, reason, serverResetTime) {
@@ -4555,6 +4681,28 @@ var AccountPoolManager = class {
 		this.authStrikes.delete(id);
 		this.persist();
 	}
+	/** True when this account has reported a dedicated quota group for `family`. */
+	hasDedicatedGroup(id, family) {
+		return this.getAccount(id)?.dedicatedGroups?.includes(family) === true;
+	}
+	/**
+	* Remember that `family` has its own quota group on this account.
+	* Persists only on an actual change so a steady-state poll stays write-free.
+	*/
+	noteDedicatedGroup(id, family) {
+		const acc = this.getAccount(id);
+		if (!acc) return;
+		if (acc.dedicatedGroups?.includes(family)) return;
+		acc.dedicatedGroups = [...acc.dedicatedGroups ?? [], family];
+		this.persist();
+	}
+	/** Forget a family's dedicated group after it stopped being reported. */
+	forgetDedicatedGroup(id, family) {
+		const acc = this.getAccount(id);
+		if (!acc?.dedicatedGroups?.includes(family)) return;
+		acc.dedicatedGroups = acc.dedicatedGroups.filter((f) => f !== family);
+		this.persist();
+	}
 	clearCooldown(id, family) {
 		if (id) {
 			const acc = this.getAccount(id);
@@ -28143,12 +28291,7 @@ const OAUTH_USERINFO_URL$1 = "https://www.googleapis.com/oauth2/v1/userinfo";
 const OAUTH_CALLBACK_PORT = 51121;
 const OAUTH_REDIRECT_URI = `http://localhost:${OAUTH_CALLBACK_PORT}/oauth-callback`;
 /** Runtime endpoint fallback order (Antigravity endpoint first, then Cloud Code PA fallbacks). */
-const AGY_ENDPOINTS = [
-	"https://daily-cloudcode-pa.googleapis.com",
-	"https://cloudcode-pa.googleapis.com",
-	"https://daily-cloudcode-pa.sandbox.googleapis.com",
-	"https://autopush-cloudcode-pa.sandbox.googleapis.com"
-];
+const AGY_ENDPOINTS = ["https://daily-cloudcode-pa.googleapis.com", "https://cloudcode-pa.googleapis.com"];
 function generatePkce() {
 	const verifier = randomBytes(32).toString("base64url");
 	return {
@@ -28449,7 +28592,7 @@ function credentialFingerprint(raw) {
 * back into an account file, re-asserting the primary), so "cannot prove it"
 * must mean "not the same".
 */
-function sameCredentialGrant(a, b) {
+function sameCredentialGrant$1(a, b) {
 	const docA = parseCredentialDoc(a);
 	const docB = parseCredentialDoc(b);
 	if (!docA?.refreshToken || !docB?.refreshToken) return false;
@@ -28524,8 +28667,8 @@ var AsyncMutex = class {
 	async acquire(opts = {}) {
 		const signal = opts.signal;
 		if (signal?.aborted) throw new AbortedWaitError("aborted before mutex acquisition");
-		if (this.holders < 1) {
-			this.holders += 1;
+		if (this.holders < 1 && this.queue.length === 0) {
+			this.holders = 1;
 			return () => this.releaseOne();
 		}
 		const pollMs = opts.pollMs ?? 0;
@@ -28551,15 +28694,12 @@ var AsyncMutex = class {
 					pollTimer = null;
 				}
 			};
-			const take = () => {
-				this.holders += 1;
-				return () => this.releaseOne();
-			};
 			const grant = () => {
 				if (settled) return;
 				settled = true;
 				cleanup();
-				resolve(take());
+				entry = null;
+				resolve(() => this.releaseOne());
 			};
 			const fail = (err) => {
 				if (settled) return;
@@ -28568,6 +28708,7 @@ var AsyncMutex = class {
 				if (entry !== null) {
 					const idx = this.queue.indexOf(entry);
 					if (idx >= 0) this.queue.splice(idx, 1);
+					else this.releaseOne();
 					entry = null;
 				} else if (takenDirectly) this.releaseOne();
 				reject(err);
@@ -28590,6 +28731,7 @@ var AsyncMutex = class {
 						if (idx >= 0) this.queue.splice(idx, 1);
 						entry = null;
 						takenDirectly = true;
+						this.holders = 1;
 						grant();
 					}
 				}, pollMs);
@@ -28598,12 +28740,9 @@ var AsyncMutex = class {
 		});
 	}
 	releaseOne() {
-		this.holders -= 1;
-		this.wakeNext();
-	}
-	wakeNext() {
 		const next = this.queue.shift();
 		if (next) queueMicrotask(next);
+		else this.holders = Math.max(0, this.holders - 1);
 	}
 };
 /** The caller's abort signal fired while queued on the mutex. */
@@ -28841,6 +28980,32 @@ var SlotUnavailableError = class extends Error {
 		this.name = "SlotUnavailableError";
 	}
 };
+/**
+* Install an identity into the machine-wide slot and PROVE the write landed.
+*
+* The four install sites here used to ignore `io.writeSlot()`'s boolean. On
+* Windows that write is a PowerShell cold start that compiles a C# P/Invoke
+* shim, so on a loaded machine it can exceed its own timeout and return false
+* WITHOUT touching the slot (reproduced 2026-10-23 while probing: a delete
+* followed by a write reported false and left the slot empty). The run then
+* spawned anyway and authenticated as whichever account was already installed —
+* silently spending the wrong account's quota, which is the exact corruption
+* class this module exists to prevent. One retry, then refuse.
+*/
+function installSlotOrThrow(io, raw, log, label) {
+	if (io.writeSlot(raw)) return;
+	log("[agy-link] credential slot write failed for " + label + " — retrying once");
+	if (io.writeSlot(raw)) return;
+	throw new SlotUnavailableError("could not install the " + label + " credential into the Windows credential slot (the write was rejected) — refusing to run, because agy would authenticate as whichever account is already in the slot");
+}
+/** Restore path: never fail a teardown, but say so loudly. */
+function installSlotBestEffort(io, raw, log, label) {
+	try {
+		installSlotOrThrow(io, raw, log, label);
+	} catch (err) {
+		log("[agy-link] WARNING: could not restore the " + label + " credential into the slot (" + (err instanceof Error ? err.message : String(err)) + ") — the next run re-installs it");
+	}
+}
 let primaryFastRuns = 0;
 /** The secondary identity currently installed in the slot, if any. */
 let installedIdentity = null;
@@ -28866,6 +29031,54 @@ function releaseSwitchTarget(key) {
 function fastPathRunCount() {
 	return primaryFastRuns + (installedIdentity?.holders ?? 0);
 }
+/** Default ceiling for holding the slot while waiting for the first event. */
+const SLOT_HANDOFF_TIMEOUT_MS = 45e3;
+/**
+* Arrange for a credential-slot lease to be released after the startup read.
+*
+* Correctness rests on the caller's spawn ordering, which does not change: the
+* identity is installed UNDER the mutex immediately before the spawn, so
+* whatever a later switcher does cannot affect a run that is already past its
+* read. The residual exposure is a mid-run 401 forcing a re-auth (agy would
+* then read whatever identity is installed at that moment); the hand-off logs
+* when it fires so that case is diagnosable instead of silent.
+*/
+function armSlotHandoff(opts) {
+	const timeoutMs = opts.timeoutMs !== void 0 && opts.timeoutMs > 0 ? opts.timeoutMs : SLOT_HANDOFF_TIMEOUT_MS;
+	const log = opts.log ?? (() => {});
+	let done = false;
+	let byTimeout = false;
+	let timer = null;
+	const clear = () => {
+		if (timer !== null) {
+			clearTimeout(timer);
+			timer = null;
+		}
+	};
+	const fire = (reason) => {
+		if (done) return;
+		done = true;
+		clear();
+		try {
+			opts.release();
+		} catch {}
+		if (reason !== "first-event") log("[agy-link] credential slot handed back at " + reason + " (no first event yet)");
+	};
+	timer = setTimeout(() => {
+		timer = null;
+		if (done) return;
+		byTimeout = true;
+		log("[agy-link] credential slot held for " + Math.round(timeoutMs / 1e3) + "s without a first event — handing it back so other sessions are not blocked");
+		fire("timeout");
+	}, timeoutMs);
+	timer.unref?.();
+	return {
+		onFirstEvent: () => fire("first-event"),
+		dispose: () => fire("run-end"),
+		released: () => done,
+		timedOut: () => byTimeout
+	};
+}
 /** True while a run is waiting for (or performing) a slot identity switch. */
 function isSlotSwitchPending() {
 	return pendingSwitchTargets.size > 0;
@@ -28920,7 +29133,7 @@ async function acquireCredentialSlot(account, pool, options = {}) {
 	const isPrimarySlot = Boolean(account.systemHome) || !account.dir;
 	if (parallel && isPrimarySlot && !isSlotSwitchPending() && !options.signal?.aborted) {
 		const stored = readPrimaryCredential(baseDir);
-		if (stored && sameCredentialGrant(io.readSlot(), stored.raw)) {
+		if (stored && sameCredentialGrant$1(io.readSlot(), stored.raw)) {
 			primaryFastRuns += 1;
 			log("[agy-link] same-account fast path: slot already holds the primary credential — running without exclusivity");
 			return {
@@ -28929,7 +29142,7 @@ async function acquireCredentialSlot(account, pool, options = {}) {
 					primaryFastRuns = Math.max(0, primaryFastRuns - 1);
 					try {
 						const afterRaw = io.readSlot();
-						if (afterRaw && afterRaw !== stored.raw && sameCredentialGrant(afterRaw, stored.raw)) writePrimaryCredential({
+						if (afterRaw && afterRaw !== stored.raw && sameCredentialGrant$1(afterRaw, stored.raw)) writePrimaryCredential({
 							raw: afterRaw,
 							email: stored.email,
 							savedAt: Date.now()
@@ -28949,7 +29162,7 @@ async function acquireCredentialSlot(account, pool, options = {}) {
 		const state = installedIdentity;
 		if (state === null || state.fingerprint !== requestedFingerprint) return null;
 		for (const key of pendingSwitchTargets.keys()) if (key !== switchKey) return null;
-		if (!sameCredentialGrant(io.readSlot(), requestedRaw)) return null;
+		if (!sameCredentialGrant$1(io.readSlot(), requestedRaw)) return null;
 		state.holders += 1;
 		return {
 			identity: account.email,
@@ -29070,9 +29283,9 @@ async function acquireCredentialSlot(account, pool, options = {}) {
 			throw new SlotUnavailableError("Cannot verify the primary Antigravity credential, so the Windows credential slot cannot be handed over safely. Sign the primary account in again, then retry.");
 		}
 		if (isPrimarySlot) {
-			if (!sameCredentialGrant(io.readSlot(), primary.raw)) {
+			if (!sameCredentialGrant$1(io.readSlot(), primary.raw)) {
 				log("[agy-link] re-asserting the primary credential into the Windows credential slot");
-				io.writeSlot(primary.raw);
+				installSlotOrThrow(io, primary.raw, log, "primary");
 			}
 			releaseLockFileNow();
 			releaseClaim();
@@ -29082,7 +29295,7 @@ async function acquireCredentialSlot(account, pool, options = {}) {
 				release: () => {
 					try {
 						const afterRaw = io.readSlot();
-						if (sameCredentialGrant(afterRaw, primary.raw)) {
+						if (sameCredentialGrant$1(afterRaw, primary.raw)) {
 							if (afterRaw && afterRaw !== primary.raw) writePrimaryCredential({
 								raw: afterRaw,
 								email: primary.email,
@@ -29090,7 +29303,7 @@ async function acquireCredentialSlot(account, pool, options = {}) {
 							}, baseDir);
 						} else if (afterRaw) {
 							log("[agy-link] credential slot changed during a primary run — restoring the primary credential");
-							io.writeSlot(primary.raw);
+							installSlotBestEffort(io, primary.raw, log, "primary");
 							writePrimaryCredential(primary, baseDir);
 						}
 					} finally {
@@ -29117,11 +29330,11 @@ async function acquireCredentialSlot(account, pool, options = {}) {
 			done();
 			throw new SlotUnavailableError(`Account ${account.alias || account.id} has an unreadable credential file. Sign it in again.`);
 		}
-		if (sameCredentialGrant(accountRaw, primary.raw)) {
+		if (sameCredentialGrant$1(accountRaw, primary.raw)) {
 			done();
 			throw new SlotUnavailableError(`Account ${account.alias || account.id} holds the SAME credential as the primary account — its login was overwritten. Sign this account in again to restore its own token.`);
 		}
-		io.writeSlot(accountRaw);
+		installSlotOrThrow(io, accountRaw, log, account.alias || account.id);
 		releaseLockFileNow();
 		const state = {
 			fingerprint: accountFingerprint,
@@ -29130,12 +29343,12 @@ async function acquireCredentialSlot(account, pool, options = {}) {
 				try {
 					const afterRaw = io.readSlot();
 					const afterFingerprint = credentialFingerprint(afterRaw);
-					if (sameCredentialGrant(afterRaw, accountRaw)) {
+					if (sameCredentialGrant$1(afterRaw, accountRaw)) {
 						if (afterRaw && afterRaw !== accountRaw) writeFileAtomic(accountFile, afterRaw);
 					} else if (afterFingerprint !== null) log(`[agy-link] credential slot was replaced during ${account.alias || account.id}'s run — refusing to write a foreign credential into the account file`);
 				} finally {
 					try {
-						io.writeSlot(primary.raw);
+						installSlotBestEffort(io, primary.raw, log, "primary");
 					} finally {
 						done();
 					}
@@ -29543,8 +29756,19 @@ async function queryViaNodeSqlite(dbPath, sql) {
 		if (typeof mod.DatabaseSync !== "function") throw new Error("DatabaseSync missing");
 		const Database = mod.DatabaseSync;
 		nodeSqliteQuery = async (p, q) => {
-			const db = new Database(p);
+			let db;
+			try {
+				db = new Database(p, { readOnly: true });
+			} catch {
+				db = new Database(p);
+			}
 			try {
+				try {
+					db.exec("PRAGMA busy_timeout = 2000;");
+				} catch {}
+				try {
+					db.exec("PRAGMA query_only = ON;");
+				} catch {}
 				return db.prepare(q).all().map((r) => `${String(r.idx)}|${String(r.step_type)}|${typeof r.payload_hex === "string" ? r.payload_hex : ""}`);
 			} finally {
 				db.close();
@@ -29557,9 +29781,13 @@ async function queryViaNodeSqlite(dbPath, sql) {
 	return nodeSqliteQuery(dbPath, sql);
 }
 /**
-* Copy the agy SQLite DB to a temp path (avoiding WAL lock issues), then
-* query tool and thought step payloads — node:sqlite first, sqlite3 CLI as
-* the fallback reader.
+* Query tool and thought step payloads directly from the agy SQLite DB —
+* node:sqlite first, sqlite3 CLI as the fallback reader.
+*
+* Direct read avoids the file-copy busy loop on macOS (issue #52) where
+* libuv uv_fs_copyfile spins at 100% CPU if an active WAL file is truncated
+* during copy. SQLite WAL mode provides concurrent read MVCC snapshots
+* natively without copying to temp.
 */
 async function loadAgyDbData(conversationId, accountHome) {
 	const steps = /* @__PURE__ */ new Map();
@@ -29576,26 +29804,34 @@ async function loadAgyDbData(conversationId, accountHome) {
 		thoughts,
 		seenSteps
 	};
-	let tmpDir = null;
-	const tmpDb = join(tmpdir(), `agy-db-${conversationId.slice(0, 8)}-${Date.now()}.db`);
 	try {
-		tmpDir = await mkdtemp(join(tmpdir(), "agy-db-"));
-		const target = join(tmpDir, "conversation.db");
-		await copyFile(dbPath, target);
-		await chmod(target, 384).catch(() => void 0);
-		try {
-			await copyFile(dbPath + "-wal", target + "-wal");
-		} catch {}
-		try {
-			await copyFile(dbPath + "-shm", target + "-shm");
-		} catch {}
+		if ((await stat(dbPath)).size === 0) return {
+			steps,
+			thoughts,
+			seenSteps
+		};
+	} catch {
+		return {
+			steps,
+			thoughts,
+			seenSteps
+		};
+	}
+	try {
 		const sql = `SELECT idx, step_type, hex(step_payload) AS payload_hex FROM steps WHERE (step_type IN (5,7,8,9,14,15,17,21,33,101,132,138,139)) AND length(step_payload) > 20`;
 		let lines = null;
 		try {
-			lines = await queryViaNodeSqlite(target, sql);
+			lines = await queryViaNodeSqlite(dbPath, sql);
 		} catch {
 			try {
-				const { stdout } = await execFileAsync("sqlite3", [target, sql], {
+				const { stdout } = await execFileAsync("sqlite3", [
+					"-cmd",
+					".timeout 2000",
+					"-cmd",
+					"PRAGMA query_only = ON;",
+					dbPath,
+					sql
+				], {
 					timeout: 5e3,
 					encoding: "utf-8",
 					maxBuffer: 26214400
@@ -29640,24 +29876,6 @@ async function loadAgyDbData(conversationId, accountHome) {
 				console.error("[dsh-agy-link] agy conversation DB unavailable (" + String(err) + ") — falling back to stream-carried metadata only" + (code === "ENOENT" ? "; install the sqlite3 CLI for full tool args and thoughts" : ""));
 			} catch {}
 		}
-	} finally {
-		if (tmpDir !== null) try {
-			await rm(tmpDir, {
-				recursive: true,
-				force: true
-			});
-		} catch {}
-		else {
-			try {
-				await unlink(tmpDb);
-			} catch {}
-			try {
-				await unlink(tmpDb + "-wal");
-			} catch {}
-			try {
-				await unlink(tmpDb + "-shm");
-			} catch {}
-		}
 	}
 	return {
 		steps,
@@ -30056,13 +30274,13 @@ function buildDigest(messages, fromIdx, maxChars) {
 		otherLines.push(digestLabel(m) + text);
 	}
 	let budget = maxChars;
-	const kept = [];
+	let taskPart = "";
 	if (taskLine !== "") {
 		if (taskLine.length > budget) {
-			kept.push(taskLine.slice(0, Math.max(0, budget)));
+			taskPart = taskLine.slice(0, Math.max(0, budget));
 			budget = 0;
 		} else {
-			kept.push(taskLine);
+			taskPart = taskLine;
 			budget -= taskLine.length;
 		}
 	}
@@ -30076,7 +30294,8 @@ function buildDigest(messages, fromIdx, maxChars) {
 		keptOther.unshift(line);
 		budget -= line.length;
 	}
-	kept.push(...keptOther);
+	const kept = [...keptOther];
+	if (taskPart !== "") kept.push(taskPart);
 	if (kept.length === 0) return "";
 	return "[conversation so far]\n" + kept.join("\n\n") + "\n[end of conversation so far]\n\n";
 }
@@ -30111,6 +30330,92 @@ function brief(s) {
 	const flat = s.trim().replace(/\s+/g, " ");
 	return flat.length > 300 ? flat.slice(0, 300) + "..." : flat;
 }
+/**
+* Await `work` while periodically yielding a SESSION-VISIBLE notice.
+*
+* Why this exists: Windows has ONE machine-global credential slot, so every
+* session's agy run is serialized behind whichever session got there first.
+* A healthy long run (a 15-minute build, a 240-minute long tool) made the
+* OTHER sessions sit with a completely empty transcript for up to
+* `dispatchTimeoutMinutes` — indistinguishable from a freeze. The console
+* heartbeat was invisible to the user, so the only honest signal was "nothing
+* is happening" (2026-10-22: two sessions looked hung for 15 minutes while
+* another project's agy run held the slot).
+*
+* Resolves with `work`'s value; rejects with `work`'s rejection untouched, so
+* the caller's existing error classification is unchanged.
+*/
+async function* awaitWithNotices(work, opts) {
+	const baseInterval = opts.intervalMs > 0 ? opts.intervalMs : 3e4;
+	const escalate = opts.escalate !== void 0 && opts.escalate > 1 ? opts.escalate : 1;
+	const maxInterval = opts.maxIntervalMs !== void 0 && opts.maxIntervalMs > baseInterval ? opts.maxIntervalMs : baseInterval * 100;
+	let intervalMs = baseInterval;
+	const startedAt = Date.now();
+	let settled = null;
+	let wake = null;
+	const ping = () => {
+		const resume = wake;
+		wake = null;
+		resume?.();
+	};
+	work.then((value) => {
+		if (settled === null) {
+			settled = {
+				ok: true,
+				value
+			};
+			ping();
+		}
+	}, (error) => {
+		if (settled === null) {
+			settled = {
+				ok: false,
+				error
+			};
+			ping();
+		}
+	});
+	let pendingNotices = 0;
+	let timer = null;
+	const arm = () => {
+		if (timer !== null || settled !== null) return;
+		timer = setTimeout(() => {
+			timer = null;
+			pendingNotices++;
+			intervalMs = Math.min(Math.round(intervalMs * escalate), maxInterval);
+			ping();
+		}, intervalMs);
+		timer.unref?.();
+	};
+	try {
+		arm();
+		for (;;) {
+			if (settled !== null) break;
+			if (pendingNotices > 0) {
+				pendingNotices--;
+				yield {
+					type: "reasoning-delta",
+					index: -1,
+					text: opts.notice(Date.now() - startedAt) + "\n"
+				};
+				arm();
+				continue;
+			}
+			await new Promise((resolve) => {
+				wake = resolve;
+			});
+		}
+	} finally {
+		if (timer !== null) {
+			clearTimeout(timer);
+			timer = null;
+		}
+	}
+	const done = settled;
+	if (done === null) throw new Error("awaitWithNotices lost its completion signal");
+	if (done.ok) return done.value;
+	throw done.error;
+}
 function sawAuthFailure(parser, outcome) {
 	if (parser.stats.sawAuthFailure) return true;
 	if (looksLikeAuthFailure(outcome.stderrTail)) return true;
@@ -30172,6 +30477,16 @@ var AgyAdapter = class AgyAdapter extends LlmAdapter {
 	*/
 	rateLimitFailover = /* @__PURE__ */ new Map();
 	static RATE_LIMIT_FAILOVER_WINDOW_MS = 6e5;
+	/**
+	* sessionKey -> consecutive transport failures and when the streak started.
+	*
+	* The streak drives the 5s/15s/30s retry delay handed to DSH. A success (or
+	* any non-transport failure) clears it, and a gap longer than the window
+	* treats the next failure as a fresh streak: a user who retries a turn ten
+	* minutes after an outage must not start by waiting 30 seconds.
+	*/
+	networkStreak = /* @__PURE__ */ new Map();
+	static NETWORK_STREAK_WINDOW_MS = 6e5;
 	constructor(deps) {
 		super();
 		this.deps = deps;
@@ -30181,6 +30496,39 @@ var AgyAdapter = class AgyAdapter extends LlmAdapter {
 		this.warnedKeys.add(key);
 		this.deps.log?.("WARNING: " + msg);
 	}
+	/**
+	* Record one transport failure and return the streak position (1-based).
+	* Position drives the delay requested from DSH.
+	*/
+	noteNetworkFailure(sessionKey) {
+		if (sessionKey === "") return 1;
+		const now = Date.now();
+		const prev = this.networkStreak.get(sessionKey);
+		const count = (prev === void 0 || now - prev.at > AgyAdapter.NETWORK_STREAK_WINDOW_MS ? 0 : prev.count) + 1;
+		this.networkStreak.set(sessionKey, {
+			count,
+			at: now
+		});
+		return count;
+	}
+	/** Drop the transport streak: the session reached Google (or failed another way). */
+	clearNetworkStreak(sessionKey) {
+		if (sessionKey === "") return;
+		this.networkStreak.delete(sessionKey);
+	}
+	/**
+	* Whether every enabled account leaves through the same proxy.
+	*
+	* A shared egress makes a transport failure a property of the machine's
+	* network, not of the account: rotating the sticky pointer repoints the
+	* session for nothing and spends the rotation budget a real quota or
+	* eligibility hop needs (live 2026-10-21: one flapping proxy produced EOFs
+	* on every account and the turn still rotated the pool).
+	*/
+	sharesSingleEgress() {
+		const fallback = (this.deps.getConfig().proxyUrl ?? "").trim() || (detectSystemProxy() ?? "");
+		return accountsShareSingleEgress(this.deps.pool?.getAccounts() ?? [], fallback);
+	}
 	/** Mark a session prompt as no longer in flight so retries are not debounced. */
 	clearInFlightPrompt(sessionKey, prompt) {
 		if (sessionKey === "") return;
@@ -30199,6 +30547,7 @@ var AgyAdapter = class AgyAdapter extends LlmAdapter {
 	*/
 	claimFailover(sessionKey, code) {
 		if (!isAccountScopedFailure(code)) return false;
+		if (code === Err.NETWORK && this.sharesSingleEgress()) return true;
 		const enabled = (this.deps.pool?.getAccounts() ?? []).filter((a) => a.enabled).length;
 		const max = Math.max(1, enabled - 1);
 		const now = Date.now();
@@ -30253,10 +30602,10 @@ var AgyAdapter = class AgyAdapter extends LlmAdapter {
 		if (enabledAccounts >= 2) retryableCodes.push(Err.AUTH);
 		return {
 			mode: "normal",
-			maxRetries: enabledAccounts >= 3 ? 2 : 1,
+			maxRetries: Math.max(enabledAccounts >= 3 ? 2 : 1, NETWORK_RETRY_DELAYS_MS.length),
 			retryableCodes,
 			initialDelayMs: 2e3,
-			maxDelayMs: 1e4,
+			maxDelayMs: 45e3,
 			jitterRatio: .1
 		};
 	}
@@ -30266,10 +30615,20 @@ var AgyAdapter = class AgyAdapter extends LlmAdapter {
 			name: "Antigravity (agy CLI)"
 		};
 	}
+	getActiveAccount() {
+		if (!this.deps.pool) return void 0;
+		const pool = this.deps.pool;
+		const activeId = pool.getActiveAccountId("google") ?? pool.getPoolData().lastActiveAccountId ?? pool.getPoolData().primaryAccountId;
+		return activeId ? pool.getAccount(activeId) : void 0;
+	}
 	async listModels(_provider) {
-		this.deps.catalog.refreshIfNeeded();
+		const activeAcc = this.getActiveAccount();
+		this.deps.catalog.refreshIfNeeded(activeAcc?.id);
 		const cfg = this.deps.getConfig();
-		const { entries, dropped } = sanitizeCatalogEntries(this.deps.catalog.get().models);
+		const cat = this.deps.catalog.get(activeAcc?.id);
+		let models = cat.models;
+		if (!activeAcc || !accountHasArgon(activeAcc, this.deps.pool, cat)) models = models.filter((m) => !isArgonModel(m.id));
+		const { entries, dropped } = sanitizeCatalogEntries(models);
 		if (dropped.length > 0) this.warnOnce("catalog-dupes", "model catalog contained duplicate ids [" + dropped.join(", ") + "] — kept first occurrence so DSH does not drop the whole provider group (INVALID_CATALOG)");
 		const hidden = new Set(cfg.hiddenModels.map((s) => s.trim()));
 		return entries.filter((m) => !hidden.has(m.id)).map((m) => ({
@@ -30280,8 +30639,9 @@ var AgyAdapter = class AgyAdapter extends LlmAdapter {
 		}));
 	}
 	async resolveModel(_provider, model, _signal) {
+		const activeAcc = this.getActiveAccount();
 		const cfg = this.deps.getConfig();
-		const cat = this.deps.catalog.get();
+		const cat = this.deps.catalog.get(activeAcc?.id);
 		const rawModel = typeof model === "string" ? model.trim() : "";
 		const cleanId = rawModel !== "" ? rawModel : cfg.defaultModel;
 		const entry = findEntry(cat, cleanId);
@@ -30411,7 +30771,7 @@ var AgyAdapter = class AgyAdapter extends LlmAdapter {
 			}
 			if (rec !== void 0) {
 				this.deps.runs.forget(continuation.runId);
-				this.warnOnce("dead-continuation:" + continuation.runId, "agy run " + continuation.runId + " settled in " + (deadFailure?.code ?? "") + " — retrying this step on another account instead of replaying it");
+				this.warnOnce("dead-continuation:" + continuation.runId, "agy run " + continuation.runId + " settled in " + (deadFailure?.code ?? "") + " — retrying this step instead of replaying it" + (deadFailure?.code === Err.NETWORK && this.sharesSingleEgress() ? " on the same account (shared egress)" : " on another account"));
 			} else this.warnOnce("stale-continuation:" + continuation.runId, "agy run " + continuation.runId + " is no longer available in memory (server restarted?) — falling back to fresh turn prompt assembly");
 		}
 		const catalog = this.deps.catalog.get();
@@ -30467,18 +30827,24 @@ var AgyAdapter = class AgyAdapter extends LlmAdapter {
 		if (model === "" && activeModel !== "" && !isAux && sessionKey !== "") this.deps.syncSessionModel?.(sessionKey, activeModel, options.reasoningEffort ? String(options.reasoningEffort) : void 0);
 		let account;
 		if (this.deps.pool && !isAux && sessionKey !== "") {
-			const prefix = `${sessionKey}:`;
-			let mostRecentAccount;
-			let latestUpdated = -1;
-			for (const [k, b] of Object.entries(this.deps.store.all())) if (k.startsWith(prefix) && b.updatedAt > latestUpdated) {
-				const boundId = k.slice(prefix.length);
-				const boundAcc = this.deps.pool.getAccount(boundId);
-				if (boundAcc) {
-					mostRecentAccount = boundAcc;
-					latestUpdated = b.updatedAt;
-				} else this.deps.store.delete(k);
-			}
-			if (mostRecentAccount && this.deps.pool.isAccountEligible(mostRecentAccount, family)) account = mostRecentAccount;
+			if (process.platform === "win32" && cfg.parallelSameAccount) {
+				const poolActive = this.deps.pool.selectAccount(family);
+				if (poolActive) account = poolActive;
+			}
+			if (!account) {
+				const prefix = `${sessionKey}:`;
+				let mostRecentAccount;
+				let latestUpdated = -1;
+				for (const [k, b] of Object.entries(this.deps.store.all())) if (k.startsWith(prefix) && b.updatedAt > latestUpdated) {
+					const boundId = k.slice(prefix.length);
+					const boundAcc = this.deps.pool.getAccount(boundId);
+					if (boundAcc) {
+						mostRecentAccount = boundAcc;
+						latestUpdated = b.updatedAt;
+					} else this.deps.store.delete(k);
+				}
+				if (mostRecentAccount && this.deps.pool.isAccountEligible(mostRecentAccount, family)) account = mostRecentAccount;
+			}
 		}
 		if (!account && this.deps.pool) account = this.deps.pool.selectAccount(family) ?? void 0;
 		if (this.deps.pool && this.deps.pool.getAccounts().length > 0 && !account) {
@@ -30560,9 +30926,13 @@ var AgyAdapter = class AgyAdapter extends LlmAdapter {
 			if (sessionAccountKey !== "") this.deps.store.delete(sessionAccountKey);
 			binding = void 0;
 		}
-		if (!isAux && binding !== void 0 && messages.length < binding.lastMessageCount) {
+		const isCompacted = binding !== void 0 && (messages.length <= 4 && binding.lastMessageCount >= 5 || binding.lastMessageCount > 10 && messages.length < Math.floor(binding.lastMessageCount * .8) || binding.lastMessageCount <= 10 && messages.length < binding.lastMessageCount - 2);
+		if (!isAux && binding !== void 0 && isCompacted) {
 			if (sessionAccountKey !== "") this.deps.store.delete(sessionAccountKey);
 			binding = void 0;
+		} else if (binding !== void 0 && messages.length < binding.lastMessageCount) {
+			binding.lastMessageCount = messages.length;
+			if (sessionAccountKey !== "") this.deps.store.set(sessionAccountKey, binding);
 		}
 		let prompt = "";
 		if (isAux && options.purpose === "compaction") {
@@ -30773,6 +31143,8 @@ var AgyAdapter = class AgyAdapter extends LlmAdapter {
 		const dispatchHardCapMs = cfg.dispatchTimeoutMinutes > 0 ? cfg.dispatchTimeoutMinutes * 6e4 : 0;
 		const dispatchStartedAt = Date.now();
 		const DISPATCH_HEARTBEAT_MS = 12e4;
+		/** First session-visible "waiting for the slot" notice (then ×3 each time). */
+		const DISPATCH_NOTICE_MS = 2e4;
 		let dispatchHeartbeat = null;
 		const stopDispatchHeartbeat = () => {
 			if (dispatchHeartbeat !== null) {
@@ -30809,9 +31181,22 @@ var AgyAdapter = class AgyAdapter extends LlmAdapter {
 		let granted = false;
 		let dispatchBackoffMs = 2e3;
 		const maxDispatchAttempts = 1 + Math.max(0, cfg.dispatchRetries);
+		let abandonedDuringDispatch = false;
+		let dispatchSettled = false;
 		try {
 			for (let attempt = 1; attempt <= maxDispatchAttempts && !granted; attempt++) try {
-				release = await this.deps.acquire(options.signal);
+				const pendingConcurrencySlot = this.deps.acquire(options.signal);
+				pendingConcurrencySlot.then((late) => {
+					if (abandonedDuringDispatch) try {
+						late();
+					} catch {}
+				}, () => {});
+				release = yield* awaitWithNotices(pendingConcurrencySlot, {
+					intervalMs: DISPATCH_NOTICE_MS,
+					escalate: 3,
+					maxIntervalMs: 54e4,
+					notice: (waitedMs) => "[agy] 正在等待 agy 凭据槽：另一个会话正在运行（已等待 " + Math.max(1, Math.round(waitedMs / 6e4)) + " 分钟）。同一台机器同一时刻只能有一个 agy 身份占用系统凭据槽；稍候会自动开始，按 Stop 可取消。\n[agy] waiting for the agy credential slot: another session is running (" + Math.max(1, Math.round(waitedMs / 6e4)) + "m elapsed)."
+				});
 				startDispatchHeartbeat();
 				slotLease = process.env.NODE_TEST_CONTEXT === void 0 ? await acquireCredentialSlot(account ?? void 0, this.deps.pool, {
 					baseDir: this.deps.pool?.dir,
@@ -30829,6 +31214,11 @@ var AgyAdapter = class AgyAdapter extends LlmAdapter {
 				if (e instanceof WaitTimeoutError) {
 					if (attempt < maxDispatchAttempts) {
 						this.deps.log?.("agy dispatch attempt " + attempt + "/" + maxDispatchAttempts + " timed out after " + cfg.dispatchTimeoutMinutes + "m waiting for the credential slot — releasing and retrying in " + Math.round(dispatchBackoffMs / 1e3) + "s");
+						yield {
+							type: "reasoning-delta",
+							index: -1,
+							text: "[agy] 等待 agy 凭据槽已超过 " + cfg.dispatchTimeoutMinutes + " 分钟，正在重新排队（" + Math.round(dispatchBackoffMs / 1e3) + " 秒后重试 " + (attempt + 1) + "/" + maxDispatchAttempts + "）。\n[agy] credential-slot wait exceeded " + cfg.dispatchTimeoutMinutes + "m — re-queueing (attempt " + (attempt + 1) + "/" + maxDispatchAttempts + ").\n"
+						};
 						stopDispatchHeartbeat();
 						await dispatchSleep(dispatchBackoffMs);
 						dispatchBackoffMs = Math.min(dispatchBackoffMs * 2, 15e3);
@@ -30839,18 +31229,30 @@ var AgyAdapter = class AgyAdapter extends LlmAdapter {
 				const detail = e instanceof SlotUnavailableError ? e.message : brief(String(e));
 				throw new LlmError(detail, Err.AGY_ERROR);
 			}
+			dispatchSettled = true;
 		} catch (e) {
+			dispatchSettled = true;
 			stopDispatchHeartbeat();
 			this.clearInFlightPrompt(sessionKey, prompt);
 			throw e;
+		} finally {
+			if (!dispatchSettled) abandonedDuringDispatch = true;
 		}
 		stopDispatchHeartbeat();
-		const releaseOnce = () => {
-			if (released) return;
-			released = true;
+		let slotHandoff = null;
+		const releaseSlotLease = () => {
+			if (slotHandoff !== null) {
+				slotHandoff.dispose();
+				return;
+			}
 			try {
 				slotLease?.release();
 			} catch {}
+		};
+		const releaseOnce = () => {
+			if (released) return;
+			released = true;
+			releaseSlotLease();
 			release?.();
 			release = null;
 		};
@@ -30864,6 +31266,7 @@ var AgyAdapter = class AgyAdapter extends LlmAdapter {
 		const effectiveProxy = account?.proxyUrl || cfg.proxyUrl || detectSystemProxy() || void 0;
 		const env = {
 			...process.env,
+			...sessionKey === "" ? {} : { DSH_MCP_SESSION: sessionKey },
 			...cfg.disableTelemetry ? {
 				DO_NOT_TRACK: "1",
 				DISABLE_TELEMETRY: "1",
@@ -30897,6 +31300,14 @@ var AgyAdapter = class AgyAdapter extends LlmAdapter {
 				}
 				this.lastAccountSpawnTime.set(account.id, Date.now());
 			}
+			if (slotLease !== void 0) slotHandoff = armSlotHandoff({
+				release: () => {
+					try {
+						slotLease?.release();
+					} catch {}
+				},
+				log: (msg) => this.deps.log?.(msg)
+			});
 			proc = startAgyProcess({
 				bin,
 				args,
@@ -30909,6 +31320,7 @@ var AgyAdapter = class AgyAdapter extends LlmAdapter {
 				stderrKeepalive: true,
 				onLine: (line) => {
 					sawAnyEvent = true;
+					slotHandoff?.onFirstEvent();
 					lastOutputAt = Date.now();
 					startToolHeartbeat();
 					applyToolIdleBudget();
@@ -31048,12 +31460,17 @@ var AgyAdapter = class AgyAdapter extends LlmAdapter {
 					message: "agy produced no result event (" + parser.stats.garbage + " unparseable lines)"
 				};
 			}
+			if (failure !== null) {
+				if (failure.code === Err.NETWORK) failure.retryAfterMs = networkRetryDelayMs(this.noteNetworkFailure(sessionKey));
+				else this.clearNetworkStreak(sessionKey);
+			}
 			rec.settle(failure);
 			const superseded = this.activeRuns.get(sessionKey) !== void 0 && this.activeRuns.get(sessionKey) !== rec;
 			if (failure === null) {
 				if (account) this.deps.pool?.recordSuccess(account.id, family);
 				this.failoverBudget.delete(sessionKey);
 				this.rateLimitFailover.delete(sessionKey);
+				this.clearNetworkStreak(sessionKey);
 				if (!isAux && sessionAccountKey !== "" && !superseded) {
 					const finalId = binding !== void 0 ? binding.conversationId : conversationId;
 					if (finalId) this.deps.store.set(sessionAccountKey, {
@@ -31068,7 +31485,8 @@ var AgyAdapter = class AgyAdapter extends LlmAdapter {
 				if (account && !outcome.aborted && looksLikeHardRateLimit(rawErrText)) this.deps.pool?.recordFailure(account.id, family, failure.message);
 				if (account && silentTimeout) this.deps.pool?.advanceAccount(account.id, family);
 				if (account && failure.code === Err.ELIGIBILITY) this.deps.pool?.advanceAccount(account.id, family);
-				if (!outcome.aborted && failure.code !== Err.RATE_LIMIT && failure.code !== Err.OVERLOADED && (failure.code === Err.NETWORK || failure.code === Err.AUTH) && account) this.deps.pool?.advanceAccount(account.id, family);
+				const sharedEgressNetwork = failure.code === Err.NETWORK && this.sharesSingleEgress();
+				if (!outcome.aborted && failure.code !== Err.RATE_LIMIT && failure.code !== Err.OVERLOADED && !sharedEgressNetwork && (failure.code === Err.NETWORK || failure.code === Err.AUTH) && account) this.deps.pool?.advanceAccount(account.id, family);
 				if (account && shouldMarkAuthRequired(failure.code, failure.message)) {
 					const authoritative = isAuthoritativeAuthError(rawErrText + " " + runLogText);
 					if (failure.code === Err.AUTH && !authoritative) this.deps.pool?.noteAuthFailure(account.id, failure.message);
@@ -31244,7 +31662,7 @@ var AgyAdapter = class AgyAdapter extends LlmAdapter {
 				}
 				if (!mapper.isFinished && !abortedLocally && !signal?.aborted) {
 					const f = rec.failureInfo;
-					if (f !== null) for (const ch of mapper.emitFailure(f.kind, f.code, f.message)) queue.push(ch);
+					if (f !== null) for (const ch of mapper.emitFailure(f.kind, f.code, f.message, f.retryAfterMs)) queue.push(ch);
 					else for (const ch of mapper.emitFailure("error", Err.INVALID_OUTPUT, "agy stream ended without a result event")) queue.push(ch);
 				}
 			} catch (err) {
@@ -32584,6 +33002,174 @@ var PoolAuthFlow = class {
 	}
 };
 //#endregion
+//#region src/host/agy-usage.ts
+function num(v) {
+	return typeof v === "number" && Number.isFinite(v) ? Math.min(1, Math.max(0, v)) : void 0;
+}
+function str(v) {
+	return typeof v === "string" && v.trim() !== "" ? v.trim() : void 0;
+}
+function normalizeBucket(raw) {
+	const id = str(raw.id);
+	if (id === void 0) return null;
+	return {
+		id,
+		name: str(raw.name),
+		window: (str(raw.window) ?? "").toLowerCase(),
+		remainingFraction: num(raw.remaining_fraction ?? raw.remainingFraction),
+		resetTime: str(raw.reset_time ?? raw.resetTime)
+	};
+}
+/** Pull groups out of one `command_result` envelope. */
+function groupsFromCommandResult(parsed) {
+	const command = parsed.command;
+	if (command === void 0 || command.name !== "usage") return null;
+	const rawGroups = command.data?.groups;
+	if (!Array.isArray(rawGroups) || rawGroups.length === 0) return null;
+	const groups = [];
+	for (const g of rawGroups) {
+		if (g === null || typeof g !== "object") continue;
+		const group = g;
+		const name = str(group.name);
+		if (name === void 0) continue;
+		const buckets = [];
+		if (Array.isArray(group.buckets)) for (const b of group.buckets) {
+			if (b === null || typeof b !== "object") continue;
+			const bucket = normalizeBucket(b);
+			if (bucket !== null) buckets.push(bucket);
+		}
+		groups.push({
+			name,
+			description: str(group.description),
+			buckets
+		});
+	}
+	return groups.length > 0 ? groups : null;
+}
+/**
+* Parse `agy -p /usage --output-format stream-json` stdout.
+*
+* Primary shape: one NDJSON line carrying a `command_result` event with
+* `command.name === "usage"`. Fallback: the plain-text table mode
+* (`<group>\t<window label>\t<pct>%\t<reset ISO>`), which the CLI prints when
+* run without `--output-format stream-json`.
+*
+* Returns null when nothing usable was found — callers keep their fallback data.
+*/
+function parseAgyUsageOutput(stdout) {
+	const lines = stdout.split(/\r?\n/);
+	for (const line of lines) {
+		const trimmed = line.trim();
+		if (!trimmed.startsWith("{")) continue;
+		let parsed;
+		try {
+			parsed = JSON.parse(trimmed);
+		} catch {
+			continue;
+		}
+		if (parsed === null || typeof parsed !== "object") continue;
+		const obj = parsed;
+		if (obj.event !== "command_result") continue;
+		const groups = groupsFromCommandResult(obj);
+		if (groups !== null) return groups;
+	}
+	return parseUsageTextTable(stdout);
+}
+/** `<group>\t<label>\t<pct>%\t<reset>` rows; weekly labels carry "week". */
+function parseUsageTextTable(text) {
+	const byName = /* @__PURE__ */ new Map();
+	for (const raw of text.split(/\r?\n/)) {
+		const line = raw.trim();
+		if (line === "") continue;
+		const cells = line.split("	");
+		if (cells.length < 3) continue;
+		const [groupName, label, pctCell] = cells;
+		const pct = Number.parseFloat(pctCell.replace("%", "").trim());
+		if (!Number.isFinite(pct)) continue;
+		const window = /week/i.test(label) ? "weekly" : "5h";
+		let group = byName.get(groupName);
+		if (group === void 0) {
+			group = {
+				name: groupName,
+				buckets: []
+			};
+			byName.set(groupName, group);
+		}
+		group.buckets.push({
+			id: `${groupName.toLowerCase().replace(/\s+/g, "-")}-${window}`,
+			name: label,
+			window,
+			remainingFraction: Math.min(1, Math.max(0, pct / 100)),
+			resetTime: cells[3]?.trim() || void 0
+		});
+	}
+	return byName.size > 0 ? [...byName.values()] : null;
+}
+/**
+* Map CLI groups onto pool families.
+*
+* Order matters: the Argon group's description mentions "Gemini quota bucket",
+* so the Argon gate must run before the generic Gemini one or the two groups
+* collapse into `google` and overwrite each other.
+*/
+function usageGroupsToFamilies(groups, now = Date.now()) {
+	const out = {};
+	if (groups === null) return out;
+	for (const group of groups) {
+		const name = group.name.toLowerCase();
+		const desc = (group.description ?? "").toLowerCase();
+		const hasArgonBucket = group.buckets.some((b) => b.id.toLowerCase().includes("argon"));
+		const isArgon = name.includes("argon") || desc.includes("argon") || hasArgonBucket;
+		const isGoogle = !isArgon && (name.includes("gemini") || desc.includes("gemini"));
+		const is3P = name.includes("claude") || name.includes("gpt") || desc.includes("claude") || desc.includes("gpt");
+		const families = isArgon ? ["argon"] : isGoogle ? ["google"] : is3P ? ["anthropic", "openai"] : [];
+		if (families.length === 0) continue;
+		let fiveHourFrac;
+		let fiveHourReset;
+		let weeklyFrac;
+		let weeklyReset;
+		for (const bucket of group.buckets) {
+			const w = (bucket.window || bucket.id).toLowerCase();
+			if (w.includes("week")) {
+				weeklyFrac = bucket.remainingFraction;
+				weeklyReset = bucket.resetTime;
+			} else if (w.includes("5h") || w.includes("five")) {
+				fiveHourFrac = bucket.remainingFraction;
+				fiveHourReset = bucket.resetTime;
+			}
+		}
+		for (const family of families) out[family] = {
+			remainingFraction: fiveHourFrac,
+			resetTime: fiveHourReset,
+			weeklyFraction: weeklyFrac,
+			weeklyResetTime: weeklyReset,
+			description: group.description,
+			updatedAt: now
+		};
+	}
+	return out;
+}
+/** Spawn one `agy -p /usage` and parse its groups. Never throws, never rejects. */
+async function readAgyUsage(opts) {
+	try {
+		return parseAgyUsageOutput((await startAgyProcess({
+			bin: opts.bin,
+			args: [
+				"-p",
+				"/usage",
+				"--output-format",
+				"stream-json"
+			],
+			timeoutMs: opts.timeoutMs ?? 2e4,
+			signal: opts.signal,
+			env: opts.env,
+			preemptAutoUpdate: opts.preemptAutoUpdate
+		}).outcome).stdout);
+	} catch {
+		return null;
+	}
+}
+//#endregion
 //#region src/host/quota.ts
 /**
 * When the quota-summary endpoint transiently fails, per-model fallback
@@ -32597,6 +33183,56 @@ function mergeFallbackFamilyQuota(prev, fallback) {
 	if (prev && typeof prev.remainingFraction === "number") return prev;
 	return fallback;
 }
+/**
+* True when any returned group is the dedicated Argon bucket.
+*
+* The backend stops emitting that group once the plan folds Argon into the
+* larger Gemini bucket (Ultra). From then on the family is governed by the
+* shared bucket, so a cooldown recorded against the retired pool — e.g. a 429
+* taken minutes before an upgrade — must not keep blocking the family for days.
+*/
+function summaryHasArgonGroup(groups) {
+	return groups.some((group) => {
+		const fields = [group.displayName, group.description];
+		for (const bucket of group.buckets ?? []) fields.push(bucket.bucketId, bucket.displayName);
+		return fields.some((value) => typeof value === "string" && value.toLowerCase().includes("argon"));
+	});
+}
+/**
+* Decide whether a family's cooldown belongs to a quota bucket that no longer exists.
+*
+* A cooldown is recorded against the pool that produced the 429. A plan change can
+* retire that pool — a Pro → Ultra upgrade folds Argon into the larger Gemini bucket
+* and the backend stops emitting the `Gemini 4 Argon` group — leaving a cooldown that
+* blocks the family until a reset that no longer applies:
+*
+*   All Antigravity accounts are unavailable: no account can serve argon right now
+*   (1 in cooldown) (earliest reset in 345338s)
+*
+* `groupSeenBefore` is what keeps this from firing on accounts that never had the
+* group at all (free tier, or any plan without an Argon bucket): their argon
+* cooldowns come from genuine 429s and must survive.
+*/
+function shouldReleaseOrphanedCooldown(state) {
+	return state.summaryAuthoritative && !state.groupPresent && state.groupSeenBefore && state.hasCooldown;
+}
+/**
+* Pick the family fraction when this poll's summary carried no group for it.
+*
+* An *authoritative* summary (at least one group returned) means the absence is
+* real: the family lost its own bucket — a Pro → Ultra upgrade folds the
+* separate Argon pool into the larger Gemini bucket, and the backend then stops
+* reporting a `Gemini 4 Argon` group entirely while `fetchAvailableModels`
+* keeps listing the argon models with fresh per-model fractions. Keeping the
+* last-known-good row there would freeze a dead window on screen (0%, a reset
+* time from the previous tier).
+*
+* A failed or empty summary is the case `mergeFallbackFamilyQuota` protects:
+* nothing authoritative arrived, so the previous complete value stays.
+*/
+function pickFamilyQuota(prev, fresh, summaryAuthoritative) {
+	return summaryAuthoritative ? fresh : mergeFallbackFamilyQuota(prev, fresh);
+}
 function detectEmailFromAgyLogs(homeDir) {
 	const logDir = join(homeDir, ".gemini", "antigravity-cli", "log");
 	if (!existsSync(logDir)) return void 0;
@@ -32615,9 +33251,25 @@ function detectEmailFromAgyLogs(homeDir) {
 	} catch {}
 }
 const OAUTH_USERINFO_URL = "https://www.googleapis.com/oauth2/v1/userinfo";
+/** Used when the installed agy version is unknown or older than AGY_UA_MIN_VERSION. */
+const AGY_UA_FALLBACK_VERSION = "1.4.5";
+function semverAtLeast(candidate, floor) {
+	if (!candidate) return false;
+	const a = candidate.match(/^(\d+)\.(\d+)\.(\d+)$/);
+	const b = floor.match(/^(\d+)\.(\d+)\.(\d+)$/);
+	if (!a || !b) return false;
+	for (let i = 1; i <= 3; i += 1) {
+		const x = Number(a[i]);
+		const y = Number(b[i]);
+		if (x !== y) return x > y;
+	}
+	return true;
+}
 /** Platform-faithful agy User-Agent (hardcoding darwin/arm64 on Windows is a bad fingerprint). */
-function agyUserAgent(version = "2.24.0") {
-	return `antigravity/${version} ${process.platform === "win32" ? "windows" : process.platform}/${process.arch === "x64" ? "amd64" : process.arch}`;
+function agyUserAgent(version) {
+	const os = process.platform === "win32" ? "windows" : process.platform;
+	const arch = process.arch === "x64" ? "amd64" : process.arch;
+	return `antigravity/cli/${semverAtLeast(version, "1.3.0") ? version : AGY_UA_FALLBACK_VERSION} (aidev_client; os_type=${os}; arch=${arch}; auth_method=consumer)`;
 }
 /** Coerce an expiry value (ISO string, epoch millis, or epoch seconds) to ms. */
 function parseExpiryMs(value) {
@@ -32663,6 +33315,49 @@ function isStale5hBucket(remainingFraction, resetTime, now = Date.now()) {
 	return resetMs > 162e5 && resetMs < 1872e4;
 }
 /**
+* True only when two token snapshots belong to the SAME OAuth grant.
+*
+* Used to decide whether the machine credential slot really holds the account
+* we are about to poll through the `agy` CLI. A refresh token identifies a
+* grant; an access token rotates, so it is only consulted when neither side
+* carries a refresh token. Mismatched shapes count as "not proven" — the
+* caller then takes the safe path instead of misattributing a quota row.
+*/
+function sameCredentialGrant(a, b) {
+	const fingerprint = (token) => {
+		if (token?.refreshToken) return "r:" + token.refreshToken;
+		if (token?.accessToken) return "a:" + token.accessToken;
+		return null;
+	};
+	const fa = fingerprint(a);
+	return fa !== null && fa === fingerprint(b);
+}
+/** How close two ISO reset stamps must be to name the SAME quota window. */
+const QUOTA_WINDOW_MATCH_TOLERANCE_MS = 3e5;
+/**
+* Organic stagger delay with randomized jitter for multi-account sequential polling.
+* Prevents rigid periodic cadence against Google Cloud Code endpoints (default: 2000~3500ms).
+*/
+function staggerJitterMs(baseMs = 2e3, maxJitterMs = 1500) {
+	return baseMs + Math.floor(Math.random() * maxJitterMs);
+}
+/**
+* True when two reset stamps point at the same quota window.
+*
+* `fetchAvailableModels` returns a per-model `quotaInfo` with NO window label:
+* for some accounts it carries the 5-hour bucket, for others the WEEKLY bucket
+* (live: daoxzeex@gmail.com reports the weekly fraction + weekly reset on every
+* Gemini model). The only reliable discriminator is the reset instant — it
+* matches exactly one of the `retrieveUserQuotaSummary` buckets.
+*/
+function sameQuotaWindow(a, b, toleranceMs = QUOTA_WINDOW_MATCH_TOLERANCE_MS) {
+	if (!a || !b) return false;
+	const ta = Date.parse(a);
+	const tb = Date.parse(b);
+	if (!Number.isFinite(ta) || !Number.isFinite(tb)) return false;
+	return Math.abs(ta - tb) <= toleranceMs;
+}
+/**
 * Parse a go-keyring payload (optionally "go-keyring-base64:" prefixed) into
 * a normalized StoredToken. Shared by the macOS and Linux keyring readers.
 */
@@ -32828,6 +33523,8 @@ async function readWindowsCredentialToken() {
 var QuotaService = class {
 	pool;
 	globalProxy;
+	agyVersion;
+	cliDeps;
 	preferredEndpointIndex = 0;
 	/** accountId -> in-flight getValidAccessToken attempt (dedup). */
 	tokenRefreshInFlight = /* @__PURE__ */ new Map();
@@ -32835,9 +33532,58 @@ var QuotaService = class {
 	* @param globalProxy resolves the plugin-wide proxy URL; used when the
 	* account has no dedicated proxyUrl. Injected so tests stay network-free.
 	*/
-	constructor(pool, globalProxy = () => "") {
+	constructor(pool, globalProxy = () => "", agyVersion = () => null, cliDeps = {}) {
 		this.pool = pool;
 		this.globalProxy = globalProxy;
+		this.agyVersion = agyVersion;
+		this.cliDeps = cliDeps;
+	}
+	/**
+	* One `agy -p /usage` spawn, mapped onto pool families.
+	*
+	* SCOPE: only the account that owns the OS credential store. `agy` resolves
+	* its credential from the platform keystore (Windows Credential Manager,
+	* macOS Keychain, Linux Secret Service) BEFORE the per-HOME token file, and
+	* that store is a single per-user entry — so the CLI can only ever answer for
+	* one identity no matter which HOME we hand it. Isolated pool accounts never
+	* own it, so polling them through the CLI filed the owner's numbers under
+	* their rows (2026-10-26: four pool rows showing one account's quota with
+	* identical reset instants).
+	*
+	* That single-owner rule keeps the CLI usage check strictly attributed to the
+	* slot owner: isolated accounts fall back to the REST path, which signs with
+	* each account's own token and reports authentic quotas (including the dedicated
+	* Gemini 4 Argon group for Pro/Ultra plans via the agy User-Agent).
+	*/
+	async readCliUsageFamilies(account) {
+		const bin = this.cliDeps.bin?.();
+		if (!bin) return null;
+		if (account.dir && !account.systemHome) return null;
+		if (!await this.cliSlotHoldsAccount(account)) return null;
+		const groups = await readAgyUsage({
+			bin,
+			env: {
+				...process.env,
+				...account.dir ? isolatedHomeEnv(account.dir) : {},
+				...proxyEnvFor(this.proxyFor(account))
+			},
+			timeoutMs: 2e4,
+			preemptAutoUpdate: this.cliDeps.preemptAutoUpdate?.()
+		});
+		return groups === null ? null : usageGroupsToFamilies(groups);
+	}
+	/**
+	* True when the machine credential slot provably holds this account's grant.
+	*
+	* This gates the CLI quota read, which can only ever see the slot owner.
+	* Conservative by design: when the two snapshots cannot be shown to be the
+	* same grant, the caller falls back to REST rather than risk attributing one
+	* account's quota to another.
+	*/
+	async cliSlotHoldsAccount(account) {
+		const slot = await this.readSystemKeychainToken();
+		if (slot === null) return false;
+		return sameCredentialGrant(slot, await this.getStoredToken(account));
 	}
 	/** Dedicated account proxy wins, then the global config proxy, then detected system proxy. */
 	proxyFor(account) {
@@ -33002,7 +33748,7 @@ var QuotaService = class {
 			headers: {
 				Authorization: `Bearer ${accessToken}`,
 				"Content-Type": "application/json",
-				"User-Agent": agyUserAgent()
+				"User-Agent": agyUserAgent(this.agyVersion())
 			},
 			body: JSON.stringify({})
 		}, proxyUrl);
@@ -33052,7 +33798,7 @@ var QuotaService = class {
 			headers: {
 				Authorization: `Bearer ${accessToken}`,
 				"Content-Type": "application/json",
-				"User-Agent": agyUserAgent()
+				"User-Agent": agyUserAgent(this.agyVersion())
 			},
 			body: JSON.stringify({})
 		}, proxyUrl);
@@ -33085,7 +33831,7 @@ var QuotaService = class {
 					headers: {
 						Authorization: `Bearer ${accessToken}`,
 						"Content-Type": "application/json",
-						"User-Agent": agyUserAgent()
+						"User-Agent": agyUserAgent(this.agyVersion())
 					},
 					body: JSON.stringify({})
 				}, proxyUrl);
@@ -33130,6 +33876,20 @@ var QuotaService = class {
 			const siblingEmails = new Set(this.pool.getAccounts().filter((a) => a.id !== account.id && a.email).map((a) => a.email));
 			if (detected && !siblingEmails.has(detected)) email = detected;
 		}
+		if (this.cliDeps.viaCli?.() !== false) {
+			const cliFamilies = await this.readCliUsageFamilies(account);
+			if (cliFamilies !== null && Object.keys(cliFamilies).length > 0) {
+				if (account.authRequired) this.pool.clearAuthRequired(account.id);
+				if (cliFamilies.argon) this.pool.noteDedicatedGroup(account.id, "argon");
+				else {
+					this.pool.clearQuotaFamily(account.id, "argon");
+					this.pool.forgetDedicatedGroup(account.id, "argon");
+					this.pool.clearCooldown(account.id, "argon");
+				}
+				this.pool.updateAccountQuotas(account.id, cliFamilies, email);
+				return cliFamilies;
+			}
+		}
 		const accessToken = await this.getValidAccessToken(account);
 		if (!accessToken) return null;
 		if (force) {
@@ -33241,24 +34001,32 @@ var QuotaService = class {
 			});
 			const isStale = stale5hFamilies.has(fam);
 			if (!familyQuotas[fam]) {
-				familyQuotas[fam] = mergeFallbackFamilyQuota(account.quotas[fam], {
+				familyQuotas[fam] = pickFamilyQuota(account.quotas[fam], {
 					remainingFraction: remaining,
 					resetTime,
 					updatedAt: now
-				});
+				}, summaryGroups.length > 0);
 				if (fam === "argon" && familyQuotas[fam].weeklyFraction === void 0) {
 					familyQuotas[fam].weeklyFraction = remaining;
 					familyQuotas[fam].weeklyResetTime = resetTime;
 				}
 			} else {
-				if (remaining < (familyQuotas[fam].remainingFraction ?? 1) || isStale) {
-					familyQuotas[fam].remainingFraction = remaining;
-					familyQuotas[fam].resetTime = resetTime;
+				const familyQuota = familyQuotas[fam];
+				if (sameQuotaWindow(resetTime, familyQuota.weeklyResetTime) && !sameQuotaWindow(resetTime, familyQuota.resetTime)) {
+					if (remaining < (familyQuota.weeklyFraction ?? 1)) {
+						familyQuota.weeklyFraction = remaining;
+						familyQuota.weeklyResetTime = resetTime;
+					}
+					continue;
+				}
+				if (remaining < (familyQuota.remainingFraction ?? 1) || isStale) {
+					familyQuota.remainingFraction = remaining;
+					familyQuota.resetTime = resetTime;
 					stale5hFamilies.delete(fam);
 				}
-				if (fam === "argon" && familyQuotas[fam].weeklyFraction === void 0) {
-					familyQuotas[fam].weeklyFraction = remaining;
-					familyQuotas[fam].weeklyResetTime = resetTime;
+				if (fam === "argon" && familyQuota.weeklyFraction === void 0) {
+					familyQuota.weeklyFraction = remaining;
+					familyQuota.weeklyResetTime = resetTime;
 				}
 			}
 		}
@@ -33266,6 +34034,21 @@ var QuotaService = class {
 			const fam = famKey;
 			if (familyQuotas[fam]) familyQuotas[fam].models = list;
 		}
+		if (summaryGroups.length > 0) {
+			if (summaryHasArgonGroup(summaryGroups) || familyQuotas.argon) this.pool.noteDedicatedGroup(account.id, "argon");
+			else {
+				this.pool.clearQuotaFamily(account.id, "argon");
+				if (shouldReleaseOrphanedCooldown({
+					summaryAuthoritative: true,
+					groupPresent: false,
+					groupSeenBefore: this.pool.hasDedicatedGroup(account.id, "argon"),
+					hasCooldown: account.cooldowns?.argon !== void 0
+				})) {
+					this.pool.forgetDedicatedGroup(account.id, "argon");
+					this.pool.clearCooldown(account.id, "argon");
+				}
+			}
+		}
 		this.pool.updateAccountQuotas(account.id, familyQuotas, email);
 		return familyQuotas;
 	}
@@ -33303,11 +34086,454 @@ var QuotaService = class {
 				if (detected && detected !== acc.email) this.pool.resetAccountIdentity(acc.id, detected);
 			}
 			accounts = accounts.filter((acc) => shouldPollAccount(acc) || needsAuthReverify(acc));
+			for (let i = 0; i < accounts.length; i++) {
+				const acc = accounts[i];
+				await this.refreshAccountQuota(acc, false).catch(() => null);
+				if (i < accounts.length - 1) await new Promise((resolve) => setTimeout(resolve, staggerJitterMs()));
+			}
+			return;
 		}
 		await Promise.allSettled(accounts.map((acc) => this.refreshAccountQuota(acc, force)));
 	}
 };
 //#endregion
+//#region src/host/warmup.ts
+/** Canonical model route resolution for upstream Google Cloud Code PA endpoints. */
+function toCanonicalWarmupModel(model) {
+	const clean = model.trim().toLowerCase();
+	if (clean === "gemini-3-flash") return "gemini-3.8-flash-tiered";
+	if (/^gemini-3\.\d+-flash$/.test(clean)) return clean + "-tiered";
+	if (clean.includes("argon")) return "gemini-4-argon";
+	return model.trim();
+}
+/** Select target model for a given family honoring user monitored_models preferences. */
+function pickWarmupModelForFamily(family, monitoredModels = []) {
+	if (family === "anthropic") return monitoredModels.find((m) => m.toLowerCase().includes("claude")) ?? "claude-sonnet-4-6";
+	if (family === "openai") return monitoredModels.find((m) => m.toLowerCase().includes("gpt")) ?? "gpt-oss-120b-medium";
+	if (family === "argon") return "gemini-4-argon";
+	return monitoredModels.find((m) => m.toLowerCase().includes("gemini") && !m.toLowerCase().includes("argon")) ?? "gemini-3.8-flash";
+}
+/**
+* `loadCodeAssist` returns `cloudaicompanionProject` either as a bare string or
+* as an object `{ id }` (the official clients and this repo's own OAuth fixture
+* both show the object form). Normalize both; an unresolved project must stay
+* empty instead of falling back to a shared project id — the backend rejects
+* foreign project ids with 403 "Caller does not have required permission".
+*/
+function extractWarmupProjectId(data) {
+	const project = data?.cloudaicompanionProject;
+	if (typeof project === "string" && project.trim() !== "") return project.trim();
+	const id = project?.id;
+	if (typeof id === "string" && id.trim() !== "") return id.trim();
+	return "";
+}
+const SIX_DAYS_MS = 5184e5;
+const FOUR_HOURS_MS = 144e5;
+const THIRTY_DAYS_MS = 2592e6;
+/**
+* A failed probe (429/403/network) must not be retried by the 5-minute
+* scheduler; otherwise one restricted account turns into an endless WAF
+* hammer. Back off for 30 minutes before the next scheduled attempt.
+*/
+const FAILURE_BACKOFF_MS = 18e5;
+/** 1.5–3s randomized gap: a metronomic interval is itself a WAF signal. */
+function warmupDelayMs() {
+	return 1500 + Math.floor(Math.random() * 1500);
+}
+/** Project ID resolver endpoints */
+const LOAD_PROJECT_ENDPOINTS = [
+	"https://daily-cloudcode-pa.googleapis.com/v1internal:loadCodeAssist",
+	"https://daily-cloudcode-pa.sandbox.googleapis.com/v1internal:loadCodeAssist",
+	"https://cloudcode-pa.googleapis.com/v1internal:loadCodeAssist"
+];
+var WarmupHistoryStore = class {
+	history = /* @__PURE__ */ new Map();
+	filePath;
+	constructor(filePath) {
+		this.filePath = filePath ?? join(stateDir(), "warmup-history.json");
+		this.load();
+	}
+	load() {
+		try {
+			if (existsSync(this.filePath)) {
+				const raw = readFileSync(this.filePath, "utf8");
+				const data = JSON.parse(raw);
+				if (data && typeof data === "object") {
+					const next = /* @__PURE__ */ new Map();
+					for (const [key, value] of Object.entries(data)) if (typeof value === "number" && Number.isFinite(value)) next.set(key, value);
+					this.history = next;
+				}
+			}
+		} catch {}
+	}
+	save() {
+		try {
+			this.pruneOldEntries();
+			const obj = Object.fromEntries(this.history.entries());
+			const tmp = `${this.filePath}.${process.pid}.tmp`;
+			writeFileSync(tmp, JSON.stringify(obj, null, 2), "utf8");
+			renameSync(tmp, this.filePath);
+		} catch {}
+	}
+	checkCooldown(key, cooldownMs, now = Date.now()) {
+		this.load();
+		const last = this.history.get(key);
+		if (last === void 0) return false;
+		return now - last < cooldownMs;
+	}
+	record(key, now = Date.now()) {
+		this.history.set(key, now);
+		this.save();
+	}
+	pruneOldEntries(now = Date.now()) {
+		const cutoff = now - THIRTY_DAYS_MS;
+		for (const [k, ts] of this.history.entries()) if (ts < cutoff) this.history.delete(k);
+	}
+	clear() {
+		this.history.clear();
+		this.save();
+	}
+};
+var WarmupService = class {
+	pool;
+	quota;
+	globalProxy;
+	agyVersion;
+	config;
+	log;
+	history;
+	constructor(pool, quota, globalProxy = () => "", agyVersion = () => null, config = () => ({}), log = () => void 0, historyStore) {
+		this.pool = pool;
+		this.quota = quota;
+		this.globalProxy = globalProxy;
+		this.agyVersion = agyVersion;
+		this.config = config;
+		this.log = log;
+		this.history = historyStore ?? new WarmupHistoryStore();
+	}
+	proxyFor(account) {
+		return account.proxyUrl?.trim() || this.globalProxy() || void 0;
+	}
+	/**
+	* Fetch cloudaicompanionProject via loadCodeAssist (official Antigravity project resolution).
+	*/
+	async fetchProjectId(accessToken, proxyUrl) {
+		const reqBody = JSON.stringify({ metadata: { ideType: "ANTIGRAVITY" } });
+		const userAgent = agyUserAgent(this.agyVersion());
+		for (const url of LOAD_PROJECT_ENDPOINTS) try {
+			const res = await agyFetch(url, {
+				method: "POST",
+				headers: {
+					Authorization: `Bearer ${accessToken}`,
+					"Content-Type": "application/json",
+					"User-Agent": userAgent,
+					"Client-Metadata": "{\"ideType\":\"ANTIGRAVITY\"}"
+				},
+				body: reqBody
+			}, proxyUrl);
+			if (res.ok) {
+				const projectId = extractWarmupProjectId(await res.json());
+				if (projectId !== "") return projectId;
+			}
+		} catch {}
+		return "";
+	}
+	/**
+	* Send a minimal probe to upstream v1internal:streamGenerateContent (fallback: generateContent).
+	*/
+	async warmupModelDirectly(account, modelName) {
+		const start = Date.now();
+		const accessToken = await this.quota.getValidAccessToken(account);
+		if (!accessToken) return {
+			ok: false,
+			error: "Could not obtain valid access token",
+			durationMs: Date.now() - start
+		};
+		const proxy = this.proxyFor(account);
+		const projectId = await this.fetchProjectId(accessToken, proxy);
+		if (projectId === "") return {
+			ok: false,
+			error: "Could not resolve cloudaicompanionProject (loadCodeAssist failed); refusing to probe a foreign project",
+			durationMs: Date.now() - start
+		};
+		const mappedModel = toCanonicalWarmupModel(modelName);
+		const isClaude = mappedModel.toLowerCase().includes("claude");
+		const sessionId = `warmup_${Date.now()}_${randomBytes(4).toString("hex")}`;
+		const reqBody = JSON.stringify({
+			...projectId !== "" ? { project: projectId } : {},
+			model: mappedModel,
+			requestId: `agent/${randomUUID()}`,
+			request: {
+				contents: [{
+					role: "user",
+					parts: [{ text: isClaude ? "ping" : "Say hi" }]
+				}],
+				generationConfig: {
+					temperature: 0,
+					maxOutputTokens: 1,
+					...isClaude ? {} : { responseModalities: ["TEXT"] }
+				},
+				sessionId
+			}
+		});
+		const headers = {
+			Authorization: `Bearer ${accessToken}`,
+			"Content-Type": "application/json",
+			"User-Agent": agyUserAgent(this.agyVersion()),
+			"Client-Metadata": "{\"ideType\":\"ANTIGRAVITY\"}"
+		};
+		if (isClaude) headers["anthropic-beta"] = "claude-code-20250219";
+		const endpoints = AGY_ENDPOINTS;
+		const attempts = [{
+			method: "streamGenerateContent",
+			query: "?alt=sse"
+		}, {
+			method: "generateContent",
+			query: ""
+		}];
+		let lastStatus = 0;
+		let lastError = "";
+		for (const attempt of attempts) for (const ep of endpoints) try {
+			const res = await agyFetch(`${ep}/v1internal:${attempt.method}${attempt.query}`, {
+				method: "POST",
+				headers,
+				body: reqBody
+			}, proxy);
+			lastStatus = res.status;
+			if (res.ok) {
+				const bodyText = await res.text().catch(() => "");
+				if (bodyText.includes("no longer available") || bodyText.includes("is not available") || bodyText.includes("is retired")) return {
+					ok: false,
+					status: res.status,
+					error: `Model is retired upstream: ${mappedModel}`,
+					durationMs: Date.now() - start
+				};
+				return {
+					ok: true,
+					status: res.status,
+					durationMs: Date.now() - start
+				};
+			}
+			const errBody = await res.text().catch(() => "");
+			if (errBody.toLowerCase().includes("validation_required")) return {
+				ok: false,
+				status: res.status,
+				error: "Account requires browser validation (VALIDATION_REQUIRED)",
+				durationMs: Date.now() - start
+			};
+			lastError = `HTTP ${res.status}: ${errBody.slice(0, 160)}`;
+			if (res.status === 401 || res.status === 403) return {
+				ok: false,
+				status: res.status,
+				error: lastError,
+				durationMs: Date.now() - start
+			};
+		} catch (err) {
+			lastError = String(err);
+		}
+		return {
+			ok: false,
+			status: lastStatus,
+			error: lastError,
+			durationMs: Date.now() - start
+		};
+	}
+	/**
+	* Warm up one account.
+	* If options.model is specified, warms up that model directly.
+	* Otherwise screens quota families for eligible weekly reset windows or cold-starts.
+	*/
+	async warmupAccount(account, options = {}) {
+		const results = [];
+		const identity = account.email || account.id;
+		const now = Date.now();
+		if (!account.enabled) return [{
+			accountId: account.id,
+			alias: account.alias,
+			email: account.email,
+			model: options.model ?? "default",
+			success: false,
+			durationMs: 0,
+			skipped: true,
+			skipReason: "Account is disabled"
+		}];
+		if (account.authRequired) return [{
+			accountId: account.id,
+			alias: account.alias,
+			email: account.email,
+			model: options.model ?? "default",
+			success: false,
+			durationMs: 0,
+			skipped: true,
+			skipReason: "Account requires authentication"
+		}];
+		if (!options.force && !shouldPollAccount(account)) return [{
+			accountId: account.id,
+			alias: account.alias,
+			email: account.email,
+			model: options.model ?? "default",
+			success: false,
+			durationMs: 0,
+			skipped: true,
+			skipReason: "Account is in cooldown or auth quarantine"
+		}];
+		if (options.model) {
+			const model = options.model;
+			const manualKey = `${identity}:${model}:100`;
+			const manualFailureKey = `${manualKey}:failed`;
+			if (!options.force && this.history.checkCooldown(manualKey, FOUR_HOURS_MS, now)) return [{
+				accountId: account.id,
+				alias: account.alias,
+				email: account.email,
+				model,
+				success: false,
+				durationMs: 0,
+				skipped: true,
+				skipReason: "Model in 4-hour warmup cooldown"
+			}];
+			if (!options.force && this.history.checkCooldown(manualFailureKey, FAILURE_BACKOFF_MS, now)) return [{
+				accountId: account.id,
+				alias: account.alias,
+				email: account.email,
+				model,
+				success: false,
+				durationMs: 0,
+				skipped: true,
+				skipReason: "Recent warmup attempt failed; retrying after backoff"
+			}];
+			this.log(`[Warmup] pinging ${model} for account ${account.id}...`);
+			const res = await this.warmupModelDirectly(account, model);
+			if (res.ok) this.history.record(manualKey, now);
+			else {
+				this.history.record(manualFailureKey, now);
+				if (res.status === 401) this.pool.markAuthRequired(account.id, "Warmup probe rejected with HTTP 401");
+			}
+			return [{
+				accountId: account.id,
+				alias: account.alias,
+				email: account.email,
+				model,
+				success: res.ok,
+				status: res.status,
+				error: res.error,
+				durationMs: res.durationMs
+			}];
+		}
+		const configuredModels = this.config().warmupModels;
+		if (configuredModels !== void 0 && configuredModels.length === 0) return [{
+			accountId: account.id,
+			alias: account.alias,
+			email: account.email,
+			model: "none",
+			success: true,
+			durationMs: 0,
+			skipped: true,
+			skipReason: "No warmup models configured"
+		}];
+		const monitored = configuredModels ?? ["gemini-3.8-flash", "claude-sonnet-4-6"];
+		const quotas = account.quotas ?? {};
+		const families = Object.keys(quotas);
+		const modelsToPing = [];
+		for (const fam of families) {
+			const q = quotas[fam];
+			if (!q) continue;
+			const familyCooldown = account.cooldowns?.[fam];
+			if (!options.force && familyCooldown && familyCooldown.cooldownUntil > now) continue;
+			const weeklyFrac = q.weeklyFraction;
+			if (weeklyFrac === void 0 || weeklyFrac < .999) continue;
+			let shouldWarm = false;
+			let historyKey = "";
+			if (q.weeklyResetTime) {
+				const resetTs = Date.parse(q.weeklyResetTime);
+				if (!isNaN(resetTs)) {
+					if (now >= resetTs - 6e4) {
+						historyKey = `${identity}:${fam}:weekly:${resetTs}`;
+						shouldWarm = options.force || !this.history.checkCooldown(historyKey, SIX_DAYS_MS, now);
+					}
+				}
+			} else {
+				historyKey = `${identity}:${fam}:weekly:initial`;
+				shouldWarm = options.force || !this.history.checkCooldown(historyKey, SIX_DAYS_MS, now);
+			}
+			const failureKey = `${historyKey}:failed`;
+			if (shouldWarm && !options.force && this.history.checkCooldown(failureKey, FAILURE_BACKOFF_MS, now)) shouldWarm = false;
+			if (shouldWarm) {
+				const model = pickWarmupModelForFamily(fam, monitored);
+				modelsToPing.push({
+					family: fam,
+					model,
+					historyKey,
+					failureKey
+				});
+			}
+		}
+		if (modelsToPing.length === 0 && options.force) {
+			const fallbackModel = pickWarmupModelForFamily("google", monitored);
+			modelsToPing.push({
+				family: "google",
+				model: fallbackModel,
+				historyKey: `${identity}:google:forced:${Math.floor(now / SIX_DAYS_MS)}`,
+				failureKey: `${identity}:google:forced:${Math.floor(now / SIX_DAYS_MS)}:failed`
+			});
+		}
+		if (modelsToPing.length === 0) return [{
+			accountId: account.id,
+			alias: account.alias,
+			email: account.email,
+			model: "none",
+			success: true,
+			durationMs: 0,
+			skipped: true,
+			skipReason: "All weekly quota windows in cooldown or not yet due for reset"
+		}];
+		for (let i = 0; i < modelsToPing.length; i++) {
+			const target = modelsToPing[i];
+			this.log(`[Warmup] weekly warmup for ${target.model} @ ${account.id} (${target.family})...`);
+			const res = await this.warmupModelDirectly(account, target.model);
+			if (res.ok) {
+				this.history.record(target.historyKey, now);
+				this.history.record(`${identity}:${target.model}:100`, now);
+			} else {
+				this.history.record(target.failureKey, now);
+				if (res.status === 401) this.pool.markAuthRequired(account.id, "Warmup probe rejected with HTTP 401");
+			}
+			results.push({
+				accountId: account.id,
+				alias: account.alias,
+				email: account.email,
+				model: target.model,
+				success: res.ok,
+				status: res.status,
+				error: res.error,
+				durationMs: res.durationMs
+			});
+			if (i < modelsToPing.length - 1) await new Promise((r) => setTimeout(r, warmupDelayMs()));
+		}
+		return results;
+	}
+	/**
+	* Warm up all accounts sequentially.
+	*/
+	async warmupAllAccounts(options = {}) {
+		const accounts = this.pool.getAccounts().filter((a) => options.force ? a.enabled && !a.authRequired : shouldPollAccount(a));
+		const allResults = [];
+		for (let i = 0; i < accounts.length; i++) {
+			const acc = accounts[i];
+			const res = await this.warmupAccount(acc, options);
+			allResults.push(...res);
+			if (i < accounts.length - 1) await new Promise((r) => setTimeout(r, warmupDelayMs()));
+		}
+		return allResults;
+	}
+	/**
+	* Scheduled periodic run: checks weekly quotas and triggers reset activation when due.
+	*/
+	async runScheduledWarmup() {
+		if (!this.config().autoWarmup) return [];
+		return this.warmupAllAccounts({ force: false });
+	}
+};
+//#endregion
 //#region src/host/mcp-bridge.ts
 const MCP_SERVER_KEY = "dsh-tools";
 /** MCP tool names are [a-zA-Z0-9_-]; DSH names may contain dots. */
@@ -33410,16 +34636,39 @@ function startMcpBridge(opts) {
 					return;
 				}
 				callSeq++;
+				const startedAt = Date.now();
+				const audit = (outcome) => {
+					try {
+						opts.log?.("mcp call " + dshName + " [" + outcome + "] " + (Date.now() - startedAt) + "ms");
+					} catch {}
+				};
 				try {
+					const sessionHeader = String(req.headers["x-dsh-session"] ?? "");
+					const resolveSession = async (id) => id === "" ? void 0 : await Promise.resolve(opts.resolveAgent?.(id));
+					let agent = await resolveSession(sessionHeader);
+					if (agent === void 0 && sessionHeader !== "" && !sessionHeader.startsWith("session-")) agent = await resolveSession("session-" + sessionHeader);
+					const controller = new AbortController();
+					const abortIfOpen = () => {
+						if (!res.writableEnded) controller.abort();
+					};
+					req.once("aborted", abortIfOpen);
+					res.once("close", () => {
+						if (!res.writableEnded) controller.abort();
+					});
+					const result = await svc.execute({
+						callId: "agy-mcp-" + callSeq,
+						name: dshName,
+						arguments: parsed.arguments ?? {},
+						signal: controller.signal,
+						...agent === void 0 ? {} : { agent }
+					});
+					audit("ok caller=" + (agent === void 0 ? "none" : "agent"));
 					sendJson(res, 200, {
 						ok: true,
-						text: resultText(await svc.execute({
-							callId: "agy-mcp-" + callSeq,
-							name: dshName,
-							arguments: parsed.arguments ?? {}
-						}))
+						text: resultText(result)
 					});
 				} catch (e) {
+					audit("failed: " + String(e).slice(0, 200));
 					sendJson(res, 200, {
 						ok: false,
 						error: String(e)
@@ -33452,12 +34701,30 @@ function startMcpBridge(opts) {
 	});
 }
 /**
-* Merge our server entry into the workspace .mcp.json. Returns a restore
-* function that puts the previous content back (or deletes the file we
-* created). Never throws.
+* Absolute path of the MCP server registry `agy` 1.4+ actually loads.
+* `override` wins so callers can target a specific file.
+* Correctly distinguishes between a home dir (`/path/to/home`) and a
+* direct .gemini dir (`/path/to/home/.gemini`).
+*/
+function agyMcpConfigPath(override = "", env = process.env) {
+	if (override !== "") return override;
+	const gch = env.GEMINI_CLI_HOME;
+	if (typeof gch === "string" && gch !== "") return basename(gch) === ".gemini" ? join(gch, "config", "mcp_config.json") : join(gch, ".gemini", "config", "mcp_config.json");
+	const home = typeof env.HOME === "string" && env.HOME !== "" ? env.HOME : homedir();
+	return join(home, ".gemini", "config", "mcp_config.json");
+}
+/** Registry key for one bridge instance: unique per port so two DSH hosts sharing a global file never clobber each other. */
+function mcpServerKey(port) {
+	return "dsh-tools-" + String(port);
+}
+/**
+* Merge our server entry into the MCP registry file at `file` (`agy` 1.4+
+* reads the global one; a workspace `.mcp.json` is ignored, see issue #55).
+* Returns a restore function that puts the previous content back (or deletes
+* the file we created). Never throws.
 */
-function writeMcpConfig(workspaceRoot, bridge) {
-	const file = join(workspaceRoot, ".mcp.json");
+function writeMcpConfig(file, bridge) {
+	const key = mcpServerKey(bridge.port);
 	let previous = null;
 	try {
 		if (existsSync(file)) previous = readFileSync(file, "utf8");
@@ -33472,7 +34739,8 @@ function writeMcpConfig(workspaceRoot, bridge) {
 		root = {};
 	}
 	const servers = root.mcpServers && typeof root.mcpServers === "object" ? root.mcpServers : {};
-	servers[MCP_SERVER_KEY] = {
+	if (MCP_SERVER_KEY in servers) delete servers[MCP_SERVER_KEY];
+	servers[key] = {
 		type: "stdio",
 		command: process.execPath,
 		args: [bridge.bridgeScript],
@@ -33483,8 +34751,11 @@ function writeMcpConfig(workspaceRoot, bridge) {
 	};
 	root.mcpServers = servers;
 	let written = false;
+	let writtenBytes = "";
 	try {
-		writeFileSync(file, JSON.stringify(root, null, 2) + "\n", "utf8");
+		mkdirSync(dirname(file), { recursive: true });
+		writtenBytes = JSON.stringify(root, null, 2) + "\n";
+		writeFileSync(file, writtenBytes, "utf8");
 		written = true;
 	} catch {
 		return () => {};
@@ -33499,6 +34770,31 @@ function writeMcpConfig(workspaceRoot, bridge) {
 				current = null;
 			}
 			if (current === null) return;
+			if (current !== writtenBytes) {
+				let live = {};
+				try {
+					const v = JSON.parse(current);
+					if (v && typeof v === "object") live = v;
+				} catch {
+					return;
+				}
+				const liveServers = live.mcpServers && typeof live.mcpServers === "object" ? live.mcpServers : {};
+				if (!(key in liveServers)) return;
+				delete liveServers[key];
+				const emptyRegistry = Object.keys(liveServers).length === 0;
+				const nothingElse = Object.keys(live).every((k) => k === "mcpServers");
+				if (emptyRegistry && nothingElse) {
+					try {
+						unlinkSync(file);
+					} catch {}
+					return;
+				}
+				writeFileSync(file, JSON.stringify({
+					...live,
+					mcpServers: liveServers
+				}, null, 2) + "\n", "utf8");
+				return;
+			}
 			let parsed = {};
 			try {
 				const v = JSON.parse(current);
@@ -33507,9 +34803,9 @@ function writeMcpConfig(workspaceRoot, bridge) {
 				return;
 			}
 			const curServers = parsed.mcpServers && typeof parsed.mcpServers === "object" ? parsed.mcpServers : {};
-			if (!(MCP_SERVER_KEY in curServers)) return;
+			if (!(key in curServers)) return;
 			if (previous === null) {
-				delete curServers[MCP_SERVER_KEY];
+				delete curServers[key];
 				if (Object.keys(curServers).length === 0) unlinkSync(file);
 				else writeFileSync(file, JSON.stringify({
 					...parsed,
@@ -33747,7 +35043,12 @@ function apply(ctx, entryConfig = {}) {
 	const version = () => versionCache;
 	const store = new SessionStore(join(stateDir(), "sessions.json"));
 	const pool = new AccountPoolManager();
-	const quota = new QuotaService(pool, () => getConfig().proxyUrl || detectSystemProxy() || "");
+	const quota = new QuotaService(pool, () => getConfig().proxyUrl || detectSystemProxy() || "", version, {
+		bin: () => bin(),
+		viaCli: () => getConfig().quotaViaCli,
+		preemptAutoUpdate: () => getConfig().suppressAgyAutoUpdate
+	});
+	const warmup = new WarmupService(pool, quota, () => getConfig().proxyUrl || detectSystemProxy() || "", version, getConfig, log);
 	const semaphore = new Semaphore(() => getConfig().maxConcurrent);
 	const getDiscoveryEnv = (account) => {
 		const cfg = getConfig();
@@ -33763,12 +35064,13 @@ function apply(ctx, entryConfig = {}) {
 			...proxyEnvFor(account?.proxyUrl || cfg.proxyUrl || detectSystemProxy() || void 0)
 		};
 	};
-	const catalog = new ModelCatalog(async (signal) => {
+	const catalog = new ModelCatalog(async (signal, accountId) => {
 		const b = bin();
 		if (!b) throw new Error("agy binary not found");
 		const poolData = pool.getPoolData();
-		const primaryAcc = poolData.primaryAccountId ? pool.getAccount(poolData.primaryAccountId) : void 0;
-		const candidateAccount = primaryAcc && primaryAcc.enabled && !primaryAcc.authRequired ? primaryAcc : pool.getAccounts().find((a) => a.enabled && !a.authRequired);
+		const targetAcc = accountId ? pool.getAccount(accountId) : void 0;
+		const activeId = pool.getActiveAccountId("google") ?? poolData.lastActiveAccountId ?? poolData.primaryAccountId;
+		const candidateAccount = targetAcc && targetAcc.enabled && !targetAcc.authRequired ? targetAcc : (activeId ? pool.getAccount(activeId) : void 0) ?? (poolData.primaryAccountId ? pool.getAccount(poolData.primaryAccountId) : void 0) ?? pool.getAccounts().find((a) => a.enabled && !a.authRequired);
 		let out = await probeProcess(b, ["models"], 3e4, signal, getDiscoveryEnv(candidateAccount), getConfig().suppressAgyAutoUpdate);
 		if ((out.code !== 0 || out.stdout.includes("Please sign in") || /sign in|auth|not logged in/i.test(out.stderrTail)) && (!candidateAccount || !candidateAccount.dir)) {
 			const secondary = pool.getAccounts().find((a) => a.enabled && !a.authRequired && a.dir);
@@ -34034,7 +35336,7 @@ function apply(ctx, entryConfig = {}) {
 			res.writeHead(status, { "Content-Type": "application/json; charset=utf-8" });
 			res.end(JSON.stringify(body));
 		};
-		const readBody = (req) => {
+		const readBody = (req, strict = false) => {
 			const r = req;
 			return new Promise((resolve) => {
 				const chunks = [];
@@ -34044,7 +35346,7 @@ function apply(ctx, entryConfig = {}) {
 						const v = JSON.parse(Buffer.concat(chunks).toString("utf8"));
 						resolve(v && typeof v === "object" ? v : {});
 					} catch {
-						resolve({});
+						resolve(strict ? { __invalidJson: true } : {});
 					}
 				});
 			});
@@ -34081,6 +35383,8 @@ function apply(ctx, entryConfig = {}) {
 						},
 						catalogModels: sanitizeCatalogEntries(cat.models).entries,
 						hiddenModels: cfg.hiddenModels,
+						autoWarmup: cfg.autoWarmup,
+						warmupModels: cfg.warmupModels,
 						bindings: Object.keys(store.all()).length,
 						lastRun
 					});
@@ -34127,7 +35431,7 @@ function apply(ctx, entryConfig = {}) {
 							if (acc) quota.refreshAccountQuota(acc).catch(() => void 0);
 						} else quota.refreshAllQuotas().catch(() => void 0);
 						store.clear();
-						catalog.forceRefresh().then(() => notifyPicker()).catch(() => void 0);
+						catalog.forceRefresh().catch(() => void 0);
 					}
 					sendJson(res, 200, st);
 				})();
@@ -34256,6 +35560,13 @@ function apply(ctx, entryConfig = {}) {
 					const body = await readBody(req);
 					const alias = typeof body.alias === "string" ? body.alias : void 0;
 					const acc = pool.createAccountSlot(alias);
+					if (bridgeState.bridge !== null) {
+						const cfg = getConfig();
+						if (cfg.mcpConfigScope === "global" || cfg.mcpConfigScope === "both") {
+							const accTarget = agyMcpConfigPath("", isolatedHomeEnv(acc.dir));
+							bridgeState.restores.push(writeMcpConfig(accTarget, bridgeState.bridge));
+						}
+					}
 					pool.markAuthRequired(acc.id, "not signed in yet — complete the terminal login");
 					if (body.autoOpenTerminal !== false && acc.dir) openAccountTerminal(acc.dir);
 					sendJson(res, 200, {
@@ -34401,7 +35712,7 @@ function apply(ctx, entryConfig = {}) {
 						const acc = pool.getAccount(id);
 						if (acc) await quota.refreshAccountQuota(acc, true);
 					} else await quota.refreshAllQuotas(true);
-					catalog.forceRefresh().then(() => notifyPicker()).catch(() => void 0);
+					catalog.forceRefresh().catch(() => void 0);
 					sendJson(res, 200, {
 						ok: true,
 						pool: pool.getPoolData()
@@ -34409,6 +35720,60 @@ function apply(ctx, entryConfig = {}) {
 				})();
 			}
 		});
+		reg({
+			kind: "exact",
+			path: "/plugins/agy-link/pool/warmup",
+			handler: (req, res) => {
+				(async () => {
+					try {
+						if (methodOf(req) !== "POST") {
+							sendJson(res, 405, { error: "POST only" });
+							return;
+						}
+						const body = await readBody(req, true);
+						if (body.__invalidJson === true) {
+							sendJson(res, 400, { error: "invalid JSON body" });
+							return;
+						}
+						const id = typeof body.id === "string" ? body.id : "";
+						const force = body.force === true;
+						const rawModel = body.model;
+						if (rawModel !== void 0 && (typeof rawModel !== "string" || !/^[A-Za-z0-9._/-]{1,80}$/.test(rawModel))) {
+							sendJson(res, 400, { error: "invalid model" });
+							return;
+						}
+						const model = typeof rawModel === "string" ? rawModel : void 0;
+						if (id === "" && body.all !== true) {
+							sendJson(res, 400, { error: "id or { all: true } required" });
+							return;
+						}
+						let results = [];
+						if (id) {
+							const acc = pool.getAccount(id);
+							if (!acc) {
+								sendJson(res, 404, { error: "Account not found" });
+								return;
+							}
+							results = await warmup.warmupAccount(acc, {
+								force,
+								model
+							});
+							await quota.refreshAccountQuota(acc, true).catch(() => void 0);
+						} else {
+							results = await warmup.warmupAllAccounts({ force });
+							await quota.refreshAllQuotas(false).catch(() => void 0);
+						}
+						sendJson(res, 200, {
+							ok: true,
+							results,
+							pool: pool.getPoolData()
+						});
+					} catch (e) {
+						sendJson(res, 500, { error: String(e) });
+					}
+				})();
+			}
+		});
 		reg({
 			kind: "exact",
 			path: "/plugins/agy-link/pool/clear-cooldown",
@@ -34441,6 +35806,7 @@ function apply(ctx, entryConfig = {}) {
 					const body = await readBody(req);
 					const id = typeof body.id === "string" ? body.id : "";
 					const ok = id !== "" && pool.setActiveAccount(id);
+					if (ok) catalog.forceRefresh(id).catch(() => void 0);
 					sendJson(res, ok ? 200 : 400, {
 						ok,
 						pool: pool.getPoolData()
@@ -34484,7 +35850,9 @@ function apply(ctx, entryConfig = {}) {
 						"defaultEffort",
 						"askTool",
 						"workspaceRoot",
-						"hiddenModels"
+						"hiddenModels",
+						"autoWarmup",
+						"warmupModels"
 					].includes(key)) {
 						sendJson(res, 400, { error: "key not settable" });
 						return;
@@ -34493,6 +35861,10 @@ function apply(ctx, entryConfig = {}) {
 						sendJson(res, 400, { error: "hiddenModels must be an array of model ids" });
 						return;
 					}
+					if (key === "warmupModels" && (!Array.isArray(body.value) || !body.value.every((x) => typeof x === "string"))) {
+						sendJson(res, 400, { error: "warmupModels must be an array of model names" });
+						return;
+					}
 					setOverride(key, body.value);
 					syncAskTool();
 					syncMirrorTool();
@@ -34577,7 +35949,11 @@ function apply(ctx, entryConfig = {}) {
 	ctx.effect(() => {
 		if (!getConfig().enabled) return () => void 0;
 		const refresh = () => {
-			quota.refreshAllQuotas().catch(() => void 0);
+			const prevArgon = pool.getAccounts().map((a) => Boolean(a.quotas?.argon)).join(",");
+			quota.refreshAllQuotas().then(() => {
+				const nextArgon = pool.getAccounts().map((a) => Boolean(a.quotas?.argon)).join(",");
+				if (prevArgon !== nextArgon) notifyPicker();
+			}).catch(() => void 0);
 		};
 		const boot = setTimeout(refresh, 5e3);
 		boot.unref?.();
@@ -34588,9 +35964,25 @@ function apply(ctx, entryConfig = {}) {
 			clearInterval(timer);
 		};
 	});
+	ctx.effect(() => {
+		if (!getConfig().enabled) return () => void 0;
+		let running = false;
+		const checkWarmup = () => {
+			if (running) return;
+			const cfg = getConfig();
+			if (!cfg.enabled || !cfg.autoWarmup) return;
+			running = true;
+			warmup.runScheduledWarmup().catch(() => void 0).finally(() => {
+				running = false;
+			});
+		};
+		const timer = setInterval(checkWarmup, 3e5);
+		timer.unref?.();
+		return () => clearInterval(timer);
+	});
 	const bridgeState = {
 		bridge: null,
-		restore: null
+		restores: []
 	};
 	let bridgeDisposed = false;
 	const syncMcpBridge = () => {
@@ -34600,10 +35992,14 @@ function apply(ctx, entryConfig = {}) {
 			try {
 				const script = fileURLToPath(new URL("./bridge.mjs", import.meta.url));
 				const toolsSvc = ctx.get("tools");
+				const agents = () => {
+					return ctx.get("agents") ?? ctx.agents;
+				};
 				const bridge = await startMcpBridge({
 					bridgeScript: script,
 					tools: () => ctx.get("tools"),
 					allowlist: () => getConfig().mcpToolAllowlist,
+					resolveAgent: (sessionId) => agents()?.get(sessionId),
 					log
 				});
 				if (bridgeDisposed) {
@@ -34612,21 +36008,28 @@ function apply(ctx, entryConfig = {}) {
 				}
 				bridgeState.bridge = bridge;
 				const root = cfg.workspaceRoot !== "" ? cfg.workspaceRoot : process.cwd();
-				bridgeState.restore = writeMcpConfig(root, bridge);
-				log("mcp bridge ready at " + bridge.url + (toolsSvc ? "" : " (tools service not yet available)"));
+				const targets = [];
+				if (cfg.mcpConfigScope === "global" || cfg.mcpConfigScope === "both") {
+					targets.push(agyMcpConfigPath(cfg.mcpConfigPath));
+					for (const acc of pool.getAccounts()) if (acc.dir) targets.push(agyMcpConfigPath("", isolatedHomeEnv(acc.dir)));
+				}
+				if (cfg.mcpConfigScope === "workspace" || cfg.mcpConfigScope === "both") targets.push(join(root, ".mcp.json"));
+				const uniqueTargets = Array.from(new Set(targets));
+				bridgeState.restores = uniqueTargets.map((file) => writeMcpConfig(file, bridge));
+				log("mcp bridge ready at " + bridge.url + " (registries: " + uniqueTargets.join(", ") + ")" + (toolsSvc ? "" : " (tools service not yet available)"));
 			} catch (e) {
-				bridgeState.restore?.();
+				for (const undo of bridgeState.restores) undo();
 				await bridgeState.bridge?.close();
 				bridgeState.bridge = null;
-				bridgeState.restore = null;
+				bridgeState.restores = [];
 				log("mcp bridge failed to start: " + String(e));
 			}
 		})();
 		else if (!want && bridgeState.bridge !== null) {
-			bridgeState.restore?.();
+			for (const undo of bridgeState.restores) undo();
 			bridgeState.bridge.close();
 			bridgeState.bridge = null;
-			bridgeState.restore = null;
+			bridgeState.restores = [];
 		}
 	};
 	syncMcpBridge();
@@ -34637,7 +36040,7 @@ function apply(ctx, entryConfig = {}) {
 			poolAuth.cancel();
 			if (askToolDispose.current !== null) askToolDispose.current();
 			if (mirrorToolDispose.current !== null) mirrorToolDispose.current();
-			bridgeState.restore?.();
+			for (const undo of bridgeState.restores) undo();
 			bridgeState.bridge?.close();
 		};
 	});