dsh-subscriptions 0.2.3 → 0.3.0

package/vendor/cursor/index.js

@@ -1,3574 +0,0 @@
-/**
- * dsh-cursor-subscription — use a Cursor subscription inside DeepSeek Harness.
- *
- * Host side: PKCE browser login against Cursor's OAuth endpoints, token
- * refresh, an `LlmAdapter` that speaks the Cursor Agent protocol
- * (`agent.v1.AgentService/Run` over HTTP/2 with Connect framing), model
- * discovery via `GetUsableModels`, and a loopback RPC channel for the web
- * client.
- *
- * The Cursor Agent protocol is reverse-engineered and undocumented; wire
- * details are kept in the `proto.js` companion module and the message
- * builders/parsers below. Field numbers were verified against the live API
- * and the vendored `agent.v1` protobuf schemas published by community
- * projects (see README).
- */
-import z from "@deepseek-ai/schemastery";
-import { credentialRef } from "@deepseek-ai/dsh-credentials";
-import { ToolCallId, LlmAdapter, LlmError } from "@deepseek-ai/dsh-llm";
-import { createHash, randomUUID } from "node:crypto";
-import { appendFileSync, readFileSync } from "node:fs";
-import { homedir } from "node:os";
-import { join } from "node:path";
-import http2 from "node:http2";
-import http from "node:http";
-import https from "node:https";
-import { spawn } from "node:child_process";
-import { gunzipSync } from "node:zlib";
-import { encodeValue, decodeValue, Reader, Writer } from "./proto.js";
-
-/**
- * Read this package's own version from the manifest beside `lib/`.
- * @returns the manifest version, or undefined when it cannot be read.
- */
-function readVersion() {
-	try {
-		return JSON.parse(readFileSync(new URL("../package.json", import.meta.url), "utf8")).version;
-	} catch {
-		/* An unreadable manifest leaves the version unknown instead of failing the plugin. */
-		return undefined;
-	}
-}
-/** Build version the settings panel prints; undefined when the manifest is unreadable. */
-export const VERSION = readVersion();
-
-//#region native fetch
-/**
- * DSH web replaces the ambient `fetch` with an MCP `createFetchWithInit`
- * wrapper whose `baseFetch` is undefined in the host process, so any consumer
- * of the global `fetch` throws `baseFetch is not a function`. Cursor usage /
- * refresh / model discovery go through HTTP, so this module uses Node's own
- * `http`/`https` to talk to the upstream directly instead of trusting the
- * host-injected `fetch`. It exposes the small Fetch subset these callers use.
- */
-export async function nativeFetch(url, init = {}) {
-	const target = new URL(url);
-	const client = target.protocol === "https:" ? https : http;
-	const method = (init.method ?? "GET").toUpperCase();
-	const body = init.body === undefined ? undefined : isBytes(init.body) ? Buffer.from(init.body) : String(init.body);
-
-	const request = (u, redirect) =>
-		new Promise((resolve, reject) => {
-			const req = client.request(
-				u,
-				{
-					method,
-					headers: sanitizeHeaders(init.headers),
-					signal: init.signal,
-				},
-				(res) => {
-					const chunks = [];
-					res.on("data", (chunk) => chunks.push(chunk));
-					res.on("end", () => {
-						const data = Buffer.concat(chunks);
-						const status = res.statusCode ?? 0;
-						if ((redirect === "error" || redirect === "manual") && status >= 300 && status < 400) {
-							reject(new Error(`HTTP ${status} from ${u.origin}${u.pathname}`));
-							return;
-						}
-						const location = res.headers.location;
-						if ((redirect ?? "follow") === "follow" && location && status >= 300 && status < 400) {
-							const next = new URL(location, u);
-							resolve({ redirect: next });
-							return;
-						}
-						resolve({
-							response: makeResponse(status, res, data),
-						});
-					});
-				},
-			);
-			req.on("error", reject);
-			if (body !== undefined) req.write(body);
-			req.end();
-		});
-
-	let u = target;
-	for (let hops = 0; ; hops++) {
-		const result = await request(u, init.redirect);
-		if (result.redirect !== undefined) {
-			if (hops >= 6) throw new Error(`too many redirects from ${target.href}`);
-			u = result.redirect;
-			continue;
-		}
-		return result.response;
-	}
-}
-
-function isBytes(value) {
-	return value instanceof Uint8Array || (typeof value !== "string" && ArrayBuffer.isView(value));
-}
-
-function sanitizeHeaders(headers) {
-	if (headers === undefined) return undefined;
-	const out = {};
-	for (const [key, value] of Object.entries(headers)) {
-		if (value === undefined) continue;
-		out[key] = value;
-	}
-	return out;
-}
-
-function makeResponse(status, res, data) {
-	const text = async () => data.toString("utf8");
-	return {
-		ok: status >= 200 && status < 300,
-		status,
-		statusText: res.statusMessage ?? "",
-		headers: {
-			get: (name) => res.headers[String(name).toLowerCase()] ?? null,
-		},
-		text: async () => text(),
-		json: async () => JSON.parse(await text()),
-		arrayBuffer: async () => data.buffer.slice(data.byteOffset, data.byteOffset + data.byteLength),
-	};
-}
-//#endregion
-
-//#region constants
-export const name = "cursor-subscription";
-export const inject = ["llm", "credentials", "connection"];
-
-export const PROVIDER = "cursor-subscription";
-export const CREDENTIAL_REF = credentialRef("CURSOR_SUBSCRIPTION_OAUTH");
-export const CHANNEL = "/cursor-subscription";
-/** Connection registry options: accept the account channel from loopback Hosts only. */
-const LOOPBACK_RPC = Object.freeze({ authority: "loopback" });
-
-export const CURSOR_BASE_URL = "https://api2.cursor.sh";
-export const CURSOR_LOGIN_URL = "https://cursor.com/loginDeepControl";
-export const CURSOR_AUTH_ORIGIN = "https://cursor.com";
-export const CURSOR_POLL_PATH = "/auth/poll";
-export const CURSOR_REFRESH_PATH = "/auth/exchange_user_api_key";
-export const CURSOR_RUN_PATH = "/agent.v1.AgentService/Run";
-export const CURSOR_MODELS_PATH = "/agent.v1.AgentService/GetUsableModels";
-
-/** Cursor dashboard usage endpoints (authenticated with the session cookie). */
-export const USAGE_API_ORIGIN = "https://cursor.com";
-export const USAGE_URL = `${USAGE_API_ORIGIN}/api/usage`;
-export const USAGE_SUMMARY_URL = `${USAGE_API_ORIGIN}/api/usage-summary`;
-export const USAGE_TEAMS_URL = `${USAGE_API_ORIGIN}/api/dashboard/teams`;
-export const USAGE_AGGREGATED_URL = `${USAGE_API_ORIGIN}/api/dashboard/get-aggregated-usage-events`;
-export const USAGE_TTL_MS = 60 * 1000;
-/** Cap the per-model spend list returned to the settings panel. */
-export const MAX_USAGE_MODELS = 20;
-
-/** Client version reported to the Agent service; bump when Cursor requires it. */
-export const CURSOR_CLIENT_VERSION = "cli-2026.02.13-41ac335";
-/** Keep-alive heartbeats while an agent run is streaming. */
-export const HEARTBEAT_INTERVAL_MS = 5000;
-/** How long an agent run may stay completely silent before we abort. */
-export const STREAM_IDLE_TIMEOUT_MS = 120 * 1000;
-/** How long a run may keep sending heartbeats without any real content. */
-export const STREAM_PROGRESS_TIMEOUT_MS = 60 * 1000;
-/** Retain checkpoints/live tool bridges for inactive DSH sessions. */
-export const SESSION_STATE_TTL_MS = 30 * 60 * 1000;
-/** Stop one Cursor Run before an unconstrained agent can loop forever. */
-export const MAX_TOOL_ROUNDS = 200;
-/**
- * How long a tool-call step keeps reading after `mcpArgs` for the frames that
- * close the burst (sibling calls, blob writes, checkpoint). Sibling calls of
- * one burst arrive within ~100 ms of each other.
- */
-export const TOOL_CALL_SETTLE_MS = 500;
-export const DEFAULT_RETRY_COUNT = 0;
-export const DEFAULT_RETRY_INTERVAL_MS = 1000;
-export const DEFAULT_RETRY_HTTP_STATUS_CODES = Object.freeze([408, 425, 429, 500, 502, 503, 504]);
-/**
- * Rebuild the Cursor conversation from the DSH history on every step instead of
- * resuming the conversation Cursor keeps server-side.
- *
- * A plugin whose transport re-sends its own message history — as the standard
- * function-calling ones do — hands the model the same explicit history at every
- * step. Delegating that history to Cursor's server instead is what a Cursor turn
- * was observed to re-derive from: the step re-stated its goal, re-read its goal
- * and rewrote its plan while its tool calls kept succeeding. Off by default: a
- * rebuild re-sends the whole transcript each step, and Cursor has no
- * provider-side compaction to bound it.
- */
-export const DEFAULT_REPLAY_HISTORY_EACH_STEP = false;
-/**
- * Settings namespace. DSH derives a plugin's settings section from its own
- * profile entry, so this has to keep matching the entry id in
- * `cordis.patch.yml`: it is the namespace `describe` and `update` are keyed by.
- */
-export const SETTINGS_NAMESPACE = "cursor-subscription";
-/**
- * Editable runtime settings. Every field is `.volatile()`: the Loader parses a
- * volatile field into a live reference that is updated in place when the
- * settings page saves, which is what carries a saved value into a Cursor run
- * that is already in flight. On DSH 0.1.7 a Config without a volatile field
- * produces no settings section at all, so none of these may lose the marker.
- */
-export const Config = z.object({
-	maxToolRounds: z.number().step(1).min(1).max(1000).default(MAX_TOOL_ROUNDS).volatile(),
-	retryCount: z.number().step(1).min(0).max(10).default(DEFAULT_RETRY_COUNT).volatile(),
-	retryIntervalMs: z.number().step(1).min(0).max(300_000).default(DEFAULT_RETRY_INTERVAL_MS).volatile(),
-	retryHttpStatusCodes: z.array(z.number().step(1).min(400).max(599)).default([...DEFAULT_RETRY_HTTP_STATUS_CODES]).volatile(),
-	replayHistoryEachStep: z.boolean().default(DEFAULT_REPLAY_HISTORY_EACH_STEP).volatile(),
-});
-/** Refresh the access token this early before its JWT expiry. */
-export const REFRESH_AHEAD_MS = 5 * 60 * 1000;
-/** Default access-token lifetime when the JWT has no usable exp claim. */
-export const DEFAULT_TOKEN_LIFETIME_MS = 24 * 60 * 60 * 1000;
-
-export const DEFAULT_CONTEXT_WINDOW = 200000;
-export const DEFAULT_MAX_TOKENS = 64000;
-
-/**
- * Read one Config field. DSH 0.1.7 delivers a `.volatile()` field to the plugin
- * as a reference whose `get()` returns the value the settings page last saved —
- * the Loader keeps that one object and writes into it, so reading it per call
- * is what makes a saved edit visible. Tests and the frozen fallback pass
- * ordinary values, so both shapes are accepted here.
- * @param value Config field as delivered by the Loader, or a plain value.
- * @returns The current plain value.
- */
-function readConfigField(value) {
-	return typeof value?.get === "function" ? value.get() : value;
-}
-
-export function resolveCursorSettings(input = {}) {
-	const maxToolRounds = readConfigField(input.maxToolRounds) ?? MAX_TOOL_ROUNDS;
-	const retryCount = readConfigField(input.retryCount) ?? DEFAULT_RETRY_COUNT;
-	const retryIntervalMs = readConfigField(input.retryIntervalMs) ?? DEFAULT_RETRY_INTERVAL_MS;
-	const retryHttpStatusCodes = readConfigField(input.retryHttpStatusCodes) ?? DEFAULT_RETRY_HTTP_STATUS_CODES;
-	const replayHistoryEachStep = readConfigField(input.replayHistoryEachStep) ?? DEFAULT_REPLAY_HISTORY_EACH_STEP;
-	if (!Number.isSafeInteger(maxToolRounds) || maxToolRounds < 1 || maxToolRounds > 1000) {
-		throw new Error("cursor-subscription: maxToolRounds must be an integer between 1 and 1000");
-	}
-	if (!Number.isSafeInteger(retryCount) || retryCount < 0 || retryCount > 10) {
-		throw new Error("cursor-subscription: retryCount must be an integer between 0 and 10");
-	}
-	if (!Number.isSafeInteger(retryIntervalMs) || retryIntervalMs < 0 || retryIntervalMs > 300_000) {
-		throw new Error("cursor-subscription: retryIntervalMs must be an integer between 0 and 300000");
-	}
-	if (!Array.isArray(retryHttpStatusCodes)
-		|| retryHttpStatusCodes.some((status) => !Number.isSafeInteger(status) || status < 400 || status > 599)) {
-		throw new Error("cursor-subscription: retryHttpStatusCodes must contain only HTTP status integers from 400 to 599");
-	}
-	if (new Set(retryHttpStatusCodes).size !== retryHttpStatusCodes.length) {
-		throw new Error("cursor-subscription: retryHttpStatusCodes must not contain duplicates");
-	}
-	if (typeof replayHistoryEachStep !== "boolean") {
-		throw new Error("cursor-subscription: replayHistoryEachStep must be a boolean");
-	}
-	return Object.freeze({
-		maxToolRounds,
-		retryCount,
-		retryIntervalMs,
-		retryHttpStatusCodes: Object.freeze([...retryHttpStatusCodes]),
-		replayHistoryEachStep,
-	});
-}
-
-export function shouldRetryHttpStatus(status, retriesUsed, settings) {
-	return retriesUsed < settings.retryCount && settings.retryHttpStatusCodes.includes(status);
-}
-
-export function isSuccessfulAgentResponse(status, contentType) {
-	return status === 200 && /^application\/connect\+proto(?:\s*;|\s*$)/i.test(contentType ?? "");
-}
-
-function abortableDelay(ms, signal) {
-	if (signal.aborted) return Promise.reject(signal.reason ?? new Error("Cursor retry aborted"));
-	if (ms === 0) return Promise.resolve();
-	return new Promise((resolve, reject) => {
-		const timer = setTimeout(done, ms);
-		timer.unref?.();
-		function done() {
-			signal.removeEventListener("abort", aborted);
-			resolve();
-		}
-		function aborted() {
-			clearTimeout(timer);
-			signal.removeEventListener("abort", aborted);
-			reject(signal.reason ?? new Error("Cursor retry aborted"));
-		}
-		signal.addEventListener("abort", aborted, { once: true });
-	});
-}
-//#endregion
-
-//#region model ordering
-/**
- * Cursor's `GetUsableModels` answers in its own order (newest family first),
- * and that order moves between releases, so the picker would reshuffle on its
- * own. Sort by display name instead, with the id as a tiebreak.
- *
- * The collation is pinned to `"en"` so one build lists the same order on every
- * host locale, and `numeric` so families keep their natural sequence
- * (`GPT-5` before `GPT-10`, `Composer 2` after `Composer 1.5`).
- */
-const MODEL_NAME_COLLATOR = new Intl.Collator("en", { numeric: true, sensitivity: "base" });
-
-/** Order model entries by name then id, without mutating the input. */
-export function sortModelsByName(models) {
-	return [...models].sort((a, b) => {
-		const byName = MODEL_NAME_COLLATOR.compare(a?.name ?? "", b?.name ?? "");
-		if (byName !== 0) return byName;
-		return MODEL_NAME_COLLATOR.compare(a?.id ?? "", b?.id ?? "");
-	});
-}
-//#endregion
-
-//#region credential store
-const clone = (value) => (value === undefined ? undefined : structuredClone(value));
-
-function assertOAuthCredential(value) {
-	if (value === undefined) return undefined;
-	if (
-		value === null ||
-		typeof value !== "object" ||
-		value.type !== "oauth" ||
-		typeof value.access !== "string" ||
-		value.access.length === 0 ||
-		typeof value.refresh !== "string" ||
-		value.refresh.length === 0 ||
-		typeof value.expires !== "number" ||
-		!Number.isFinite(value.expires)
-	) {
-		throw new Error("Cursor credential store received a malformed OAuth credential");
-	}
-	return clone(value);
-}
-
-function parseOAuthCredential(raw) {
-	try {
-		return assertOAuthCredential(JSON.parse(raw));
-	} catch (error) {
-		if (error?.message === "Cursor credential store received a malformed OAuth credential") throw error;
-		throw new Error("Cursor credential store contains malformed OAuth JSON", { cause: error });
-	}
-}
-
-/**
- * Adapt DSH's managed string credential service to Cursor's typed OAuth
- * credential. Write operations are serialized so an older refresh response
- * cannot overwrite a newer rotated token.
- */
-export class CursorCredentialStore {
-	#chain = Promise.resolve();
-
-	constructor(credentials, ref) {
-		if (credentials === undefined || credentials === null) {
-			throw new Error("Cursor OAuth requires the DSH credentials service");
-		}
-		this.credentials = credentials;
-		this.ref = ref;
-	}
-
-	#enqueue(operation) {
-		const current = this.#chain.catch(() => undefined).then(operation);
-		const tail = current.catch(() => undefined);
-		this.#chain = tail;
-		return current;
-	}
-
-	async read() {
-		const hit = await this.credentials.resolve(this.ref);
-		if (hit?.value === undefined || hit.value === "") return undefined;
-		return parseOAuthCredential(hit.value);
-	}
-
-	async write(credential) {
-		const validated = assertOAuthCredential(credential);
-		if (validated === undefined) throw new Error("Cursor credential store cannot write an empty credential");
-		await this.credentials.set(this.ref, JSON.stringify(validated));
-		return clone(validated);
-	}
-
-	modify(update) {
-		return this.#enqueue(async () => {
-			const current = await this.read();
-			const next = await update(clone(current));
-			if (next === undefined) return current;
-			return this.write(next);
-		});
-	}
-
-	async clear() {
-		await this.#enqueue(async () => {
-			await this.credentials.unset(this.ref);
-		});
-	}
-}
-//#endregion
-
-//#region token helpers
-/**
- * Extract a JWT expiry (ms epoch) with a safety margin.
- * Falls back to a default lifetime when the token cannot be parsed.
- */
-export function getTokenExpiry(token, now = Date.now) {
-	try {
-		const parts = token.split(".");
-		if (parts.length !== 3 || !parts[1]) return now() + DEFAULT_TOKEN_LIFETIME_MS;
-		const decoded = JSON.parse(
-			Buffer.from(parts[1].replace(/-/g, "+").replace(/_/g, "/"), "base64").toString("utf8"),
-		);
-		if (decoded && typeof decoded === "object" && typeof decoded.exp === "number") {
-			return decoded.exp * 1000;
-		}
-	} catch {}
-	return now() + DEFAULT_TOKEN_LIFETIME_MS;
-}
-
-/**
- * Decode the `sub` claim of an access token and strip the identity-provider
- * prefix (e.g. `github|user_...` → `user_...`), which is what the dashboard
- * session cookie and the `/api/usage?user=` parameter expect.
- */
-export function getTokenSub(token) {
-	try {
-		const parts = token.split(".");
-		if (parts.length !== 3 || !parts[1]) return undefined;
-		const decoded = JSON.parse(Buffer.from(parts[1], "base64url").toString("utf8"));
-		if (typeof decoded?.sub !== "string" || decoded.sub.length === 0) return undefined;
-		return decoded.sub.includes("|") ? decoded.sub.split("|").pop() : decoded.sub;
-	} catch {
-		return undefined;
-	}
-}
-//#endregion
-
-//#region pkce
-async function generatePkce() {
-	const verifierBytes = new Uint8Array(96);
-	globalThis.crypto.getRandomValues(verifierBytes);
-	const verifier = Buffer.from(verifierBytes).toString("base64url");
-	const digest = await globalThis.crypto.subtle.digest("SHA-256", new TextEncoder().encode(verifier));
-	const challenge = Buffer.from(digest).toString("base64url");
-	return { verifier, challenge };
-}
-
-/** Build the browser login URL for the PKCE flow. */
-export function buildLoginUrl({ challenge, uuid }) {
-	const params = new URLSearchParams({ challenge, uuid, mode: "login", redirectTarget: "cli" });
-	return `${CURSOR_LOGIN_URL}?${params.toString()}`;
-}
-//#endregion
-
-//#region external url
-/** Validate the only external origin this plugin may launch. */
-export function assertCursorAuthUrl(value) {
-	let url;
-	try {
-		url = new URL(value);
-	} catch {
-		throw new Error("Cursor auth URL is invalid");
-	}
-	if (url.protocol !== "https:") throw new Error("Cursor auth URL must use HTTPS");
-	if (url.origin !== CURSOR_AUTH_ORIGIN || url.username !== "" || url.password !== "") {
-		throw new Error("Cursor auth URL must use the cursor.com origin");
-	}
-	return url.href;
-}
-
-/** Return a shell-free native opener command for the current desktop. */
-export function commandForCursorAuthUrl(value, platform = process.platform) {
-	const url = assertCursorAuthUrl(value);
-	if (platform === "win32") return { file: "rundll32.exe", args: ["url.dll,FileProtocolHandler", url], shell: false };
-	if (platform === "darwin") return { file: "open", args: [url], shell: false };
-	if (platform === "linux") return { file: "xdg-open", args: [url], shell: false };
-	throw new Error(`Cursor auth URL opener is unsupported on ${platform}`);
-}
-
-export function openCursorAuthUrl(value, options = {}) {
-	const command = commandForCursorAuthUrl(value, options.platform);
-	const spawnProcess = options.spawn ?? spawn;
-	return new Promise((resolve, reject) => {
-		const child = spawnProcess(command.file, command.args, {
-			detached: true,
-			stdio: "ignore",
-			windowsHide: true,
-			shell: command.shell,
-		});
-		child.once("error", reject);
-		child.once("spawn", () => {
-			child.unref();
-			resolve();
-		});
-	});
-}
-//#endregion
-
-//#region auth service
-/**
- * Owns the Cursor OAuth lifecycle: PKCE login polling, refresh, and status.
- * Tokens live in the DSH credential store; the browser client only ever sees
- * `{ authenticated, provider, type, expiresAt }`.
- */
-export class CursorAuthService {
-	constructor(store, options = {}) {
-		this.store = store;
-		this.fetch = options.fetch ?? nativeFetch;
-		this.now = options.now ?? Date.now;
-		this.logger = options.logger;
-	}
-
-	#log(level, message, ...args) {
-		try {
-			this.logger?.[level]?.(`cursor-subscription: ${message}`, ...args);
-		} catch {}
-	}
-
-	async status({ signal } = {}) {
-		signal?.throwIfAborted();
-		const current = await this.store.read();
-		if (current === undefined) return { authenticated: false, provider: PROVIDER };
-		return {
-			authenticated: true,
-			provider: PROVIDER,
-			type: "oauth",
-			expiresAt: current.expires,
-		};
-	}
-
-	/**
-	 * Resolve a usable access token, refreshing first when the stored token is
-	 * missing, expired, or about to expire.
-	 * @returns {Promise<string>} the bearer token.
-	 */
-	async accessToken({ signal } = {}) {
-		const credential = await this.credential({ signal });
-		return credential.access;
-	}
-
-	/**
-	 * Resolve the stored credential, refreshing first when its token is
-	 * missing, expired, or about to expire.
-	 * @returns {Promise<{access: string, refresh: string, expires: number}>}
-	 */
-	async credential({ signal } = {}) {
-		signal?.throwIfAborted();
-		const current = await this.store.read();
-		if (current === undefined) {
-			throw new LlmError("Cursor subscription is not signed in", "MISSING_CREDENTIAL");
-		}
-		if (current.expires - this.now() > REFRESH_AHEAD_MS) return current;
-		const refreshed = await this.refresh(current, { signal });
-		return refreshed;
-	}
-
-	async refresh(current, { signal } = {}) {
-		const response = await this.fetch(`${CURSOR_BASE_URL}${CURSOR_REFRESH_PATH}`, {
-			method: "POST",
-			redirect: "error",
-			headers: {
-				authorization: `Bearer ${current.refresh}`,
-				"content-type": "application/json",
-				accept: "application/json",
-				"user-agent": "dsh-cursor-subscription/0.1.0",
-			},
-			body: "{}",
-			signal,
-		});
-		if (!response.ok) {
-			this.#log("warn", "token refresh failed (HTTP %s)", response.status);
-			if (response.status === 401 || response.status === 403) {
-				throw new LlmError("Cursor sign-in needs to be renewed", "INVALID_CREDENTIAL");
-			}
-			throw new LlmError(`Cursor token refresh failed (HTTP ${response.status})`, "AUTH_FAILED");
-		}
-		let data;
-		try {
-			data = await response.json();
-		} catch (error) {
-			throw new LlmError("Cursor returned an unreadable token response", "AUTH_FAILED", { cause: error });
-		}
-		if (typeof data?.accessToken !== "string" || data.accessToken.length === 0) {
-			throw new LlmError("Cursor sign-in needs to be renewed", "INVALID_CREDENTIAL");
-		}
-		const next = {
-			type: "oauth",
-			access: data.accessToken,
-			refresh: typeof data.refreshToken === "string" && data.refreshToken.length > 0 ? data.refreshToken : current.refresh,
-			expires: getTokenExpiry(data.accessToken, this.now),
-		};
-		const stored = await this.store.modify((latest) => {
-			// Only rotate when the stored credential is still the one we refreshed.
-			if (latest === undefined || latest.refresh !== current.refresh) return latest;
-			return next;
-		});
-		return stored ?? next;
-	}
-
-	/** Start a login and resolve once the browser flow completes. */
-	async login({ interaction, signal }) {
-		signal?.throwIfAborted();
-		const { verifier, challenge } = await generatePkce();
-		const uuid = randomUUID();
-		const loginUrl = buildLoginUrl({ challenge, uuid });
-
-		interaction.notify?.({
-			type: "auth_url",
-			url: assertCursorAuthUrl(loginUrl),
-			instructions: "在浏览器中登录 Cursor 并授权后，此页面会自动完成登录。",
-		});
-		interaction.prompt?.({ type: "text", message: "等待浏览器登录完成…" }).catch(() => {});
-		interaction.signal?.throwIfAborted();
-
-		let delay = 1000;
-		for (let attempt = 0; attempt < 150; attempt++) {
-			await sleep(delay);
-			signal?.throwIfAborted();
-			interaction.signal?.throwIfAborted();
-			let response;
-			try {
-				response = await this.fetch(`${CURSOR_BASE_URL}${CURSOR_POLL_PATH}?uuid=${encodeURIComponent(uuid)}&verifier=${encodeURIComponent(verifier)}`, {
-					redirect: "error",
-					headers: { accept: "application/json", "user-agent": "dsh-cursor-subscription/0.1.0" },
-					signal,
-				});
-			} catch (error) {
-				this.#log("warn", "login poll request failed on attempt %d: %s", attempt, error?.cause?.code ?? error?.message);
-				throw new LlmError("Cursor login poll request failed", "AUTH_FAILED", { cause: error });
-			}
-			if (response.status === 404) {
-				delay = Math.min(delay * 1.2, 10000);
-				continue;
-			}
-			if (!response.ok) {
-				this.#log("warn", "login poll returned HTTP %s on attempt %d", response.status, attempt);
-				throw new LlmError(`Cursor login poll failed (HTTP ${response.status})`, "AUTH_FAILED");
-			}
-			const text = await Promise.resolve(response.text()).catch(() => "");
-			this.#log("info", "login poll succeeded after %d attempts", attempt);
-			let data;
-			try {
-				data = JSON.parse(text);
-			} catch (error) {
-				this.#log("warn", "login poll returned unreadable JSON: %s", text.slice(0, 200));
-				throw new LlmError("Cursor login returned an unreadable response", "AUTH_FAILED", { cause: error });
-			}
-			if (typeof data?.accessToken !== "string" || data.accessToken.length === 0) {
-				this.#log("warn", "login poll response had no accessToken");
-				throw new LlmError("Cursor login returned no access token", "AUTH_FAILED");
-			}
-			const credential = {
-				type: "oauth",
-				access: data.accessToken,
-				refresh: typeof data.refreshToken === "string" && data.refreshToken.length > 0 ? data.refreshToken : "",
-				expires: getTokenExpiry(data.accessToken, this.now),
-			};
-			try {
-				await this.store.modify(() => credential);
-			} catch (error) {
-				this.#log("error", "failed to store Cursor credential: %s", error?.message);
-				throw new LlmError("Cursor login could not store the credential", "AUTH_FAILED", { cause: error });
-			}
-			return;
-		}
-		this.#log("warn", "login timed out after 150 poll attempts");
-		throw new LlmError("Cursor login timed out", "AUTH_FAILED");
-	}
-
-	async logout({ signal } = {}) {
-		signal?.throwIfAborted();
-		await this.store.clear();
-		return this.status({ signal });
-	}
-}
-
-function sleep(ms) {
-	return new Promise((resolve) => setTimeout(resolve, ms));
-}
-//#endregion
-
-//#region login coordinator
-const TERMINAL_PHASES = new Set(["authenticated", "failed", "cancelled"]);
-const publicClone = (value) => structuredClone(value);
-const asObject = (value) => (value !== null && typeof value === "object" ? value : {});
-const ok = (value) => ({ ok: true, value });
-const badRequest = (message) => ({ ok: false, error: { code: "bad-request", message, details: { issues: [] } } });
-
-const deferred = () => {
-	let resolve;
-	let reject;
-	return {
-		promise: new Promise((onResolve, onReject) => {
-			resolve = onResolve;
-			reject = onReject;
-		}),
-		resolve,
-		reject,
-	};
-};
-
-/** Own one host-side login without exposing tokens to the browser client. */
-export class CursorLoginCoordinator {
-	#sessions = new Map();
-	#activeId;
-
-	constructor(auth, options = {}) {
-		this.auth = auth;
-		this.createId = options.createId ?? (() => randomUUID());
-		this.logger = options.logger;
-	}
-
-	#log(level, message, ...args) {
-		try {
-			this.logger?.[level]?.(`cursor-subscription: ${message}`, ...args);
-		} catch {}
-	}
-
-	async accountStatus(options) {
-		return publicClone(await this.auth.status(options));
-	}
-
-	async start() {
-		const active = this.#activeId === undefined ? undefined : this.#sessions.get(this.#activeId);
-		if (active !== undefined && !TERMINAL_PHASES.has(active.view.phase)) {
-			throw new Error("a Cursor login is already active");
-		}
-		if (active !== undefined) this.#sessions.delete(active.view.id);
-		const id = this.createId();
-		const ready = deferred();
-		const controller = new AbortController();
-		const session = {
-			controller,
-			ready,
-			view: { id, provider: PROVIDER, method: "browser", phase: "starting", authenticated: false },
-		};
-		this.#sessions.set(id, session);
-		this.#activeId = id;
-		const publishReady = () => ready.resolve(this.read(id));
-		const interaction = {
-			signal: controller.signal,
-			prompt: async (prompt) => {
-				controller.signal.throwIfAborted();
-				const answer = deferred();
-				session.prompt = answer;
-				session.view = { ...session.view, phase: "waiting_input", prompt: publicPrompt(prompt) };
-				const abortPrompt = () => answer.reject(controller.signal.reason ?? new Error("login cancelled"));
-				controller.signal.addEventListener("abort", abortPrompt, { once: true });
-				prompt.signal?.addEventListener("abort", abortPrompt, { once: true });
-				publishReady();
-				try {
-					return await answer.promise;
-				} finally {
-					controller.signal.removeEventListener("abort", abortPrompt);
-					prompt.signal?.removeEventListener("abort", abortPrompt);
-					if (session.prompt === answer) session.prompt = undefined;
-				}
-			},
-			notify: (event) => {
-				if (controller.signal.aborted) return;
-				if (event.type === "auth_url") {
-					session.view = {
-						...session.view,
-						phase: "waiting_browser",
-						authUrl: assertCursorAuthUrl(event.url),
-						...typeof event.instructions === "string" ? { instructions: event.instructions } : {},
-					};
-				} else {
-					session.view = { ...session.view, message: String(event.message ?? "") };
-				}
-				publishReady();
-			},
-		};
-		session.run = Promise.resolve()
-			.then(() => this.auth.login({ interaction, signal: controller.signal }))
-			.then(async () => {
-				if (controller.signal.aborted) return;
-				const status = await this.auth.status();
-				session.view = {
-					id,
-					provider: PROVIDER,
-					method: "browser",
-					phase: "authenticated",
-					authenticated: status.authenticated === true,
-					...typeof status.expiresAt === "number" ? { expiresAt: status.expiresAt } : {},
-				};
-			})
-			.catch((error) => {
-				if (controller.signal.aborted) {
-					session.view = { id, provider: PROVIDER, method: "browser", phase: "cancelled", authenticated: false };
-					return;
-				}
-				const message = error instanceof Error ? error.message : String(error);
-				this.#log("warn", "login failed: %s", message);
-				session.view = {
-					id,
-					provider: PROVIDER,
-					method: "browser",
-					phase: "failed",
-					authenticated: false,
-					error: "Cursor login failed",
-					...(message ? { detail: message.slice(0, 500) } : {}),
-				};
-				session.hostError = error;
-			})
-			.finally(publishReady);
-		return ready.promise;
-	}
-
-	read(id) {
-		const session = this.#sessions.get(id);
-		if (session === undefined) throw new Error("unknown Cursor login");
-		return publicClone(session.view);
-	}
-
-	async cancel(id) {
-		const session = this.#sessions.get(id);
-		if (session === undefined) throw new Error("unknown Cursor login");
-		if (!TERMINAL_PHASES.has(session.view.phase)) {
-			session.view = { id, provider: PROVIDER, method: "browser", phase: "cancelled", authenticated: false };
-			session.controller.abort(new Error("Cursor login cancelled"));
-		}
-		await Promise.resolve(session.run).catch(() => undefined);
-		return this.read(id);
-	}
-
-	async logout(options) {
-		if (this.#activeId !== undefined) {
-			const active = this.#sessions.get(this.#activeId);
-			if (active !== undefined && !TERMINAL_PHASES.has(active.view.phase)) await this.cancel(active.view.id);
-		}
-		await this.auth.logout(options);
-		return this.accountStatus(options);
-	}
-}
-
-const publicPrompt = (prompt) => ({
-	type: prompt.type,
-	message: String(prompt.message ?? ""),
-	...typeof prompt.placeholder === "string" ? { placeholder: prompt.placeholder } : {},
-});
-
-/** Map the loopback-only DSH Connection channel onto the coordinator. */
-export function createCursorRpcHandler(coordinator, options = {}) {
-	const openExternal = options.openExternal;
-	const usageReader = options.usageReader;
-	const modelsProvider = options.modelsProvider;
-	const settingsController = options.settings;
-	const version = options.version ?? VERSION;
-	const publicError = (code, message) => ({ ok: false, error: { code, message, details: { issues: [] } } });
-	return async (endpoint, payload, signal) => {
-		try {
-			signal.throwIfAborted();
-			const input = asObject(payload);
-			// The panel labels itself with the build that is actually serving it, so the
-			// version never has to be duplicated into the browser bundle.
-			if (endpoint === "version") return ok({ version });
-			if (endpoint === "settings") return ok(settingsController.read());
-			if (endpoint === "settings/update") {
-				if (!Number.isSafeInteger(input.revision) || input.revision < 0) throw new Error("invalid Cursor settings revision");
-				const patch = {};
-				for (const key of ["maxToolRounds", "retryCount", "retryIntervalMs", "retryHttpStatusCodes", "replayHistoryEachStep"]) {
-					if (Object.hasOwn(input, key)) patch[key] = input[key];
-				}
-				return ok(await settingsController.update(patch, input.revision));
-			}
-			if (endpoint === "status") return ok(await coordinator.accountStatus({ signal }));
-			if (endpoint === "usage") {
-				try {
-					return ok(await usageReader.read({ force: input.force === true, signal }));
-				} catch (error) {
-					const known = new Set(["Cursor subscription is not signed in", "Cursor sign-in needs to be renewed"]);
-					if (error instanceof Error && known.has(error.message)) return publicError("internal", error.message);
-					const raw = error instanceof Error ? error.message : "";
-					const http = raw.match(/^HTTP (\d{3})\b/);
-					const message = http ? `Could not read Cursor usage (HTTP ${http[1]})` : `Could not read Cursor usage: ${raw}`;
-					return publicError("internal", message);
-				}
-			}
-			if (endpoint === "models") {
-				try {
-					return ok({
-						models: await modelsProvider.listModelsForRpc({ force: input.force === true, signal }),
-						fetchedAt: Date.now(),
-					});
-				} catch (error) {
-					return publicError("internal", error instanceof Error ? error.message : "Could not list Cursor models");
-				}
-			}
-			if (endpoint === "login/start") {
-				const started = await coordinator.start();
-				if (input.openExternal !== true) return ok(started);
-				const url = started.authUrl;
-				if (typeof url !== "string" || openExternal === undefined) return ok({ ...started, externalOpened: false });
-				try {
-					await openExternal(url);
-					return ok({ ...started, externalOpened: true });
-				} catch {
-					return ok({ ...started, externalOpened: false });
-				}
-			}
-			if (endpoint === "login/status") return ok(coordinator.read(input.id));
-			if (endpoint === "login/cancel") return ok(await coordinator.cancel(input.id));
-			if (endpoint === "logout") {
-				const result = await coordinator.logout({ signal });
-				usageReader?.clear();
-				return ok(result);
-			}
-			return badRequest(`unknown Cursor auth endpoint: ${endpoint}`);
-		} catch (error) {
-			if (signal.aborted) throw error;
-			const message = error instanceof Error && /^(unknown|unsupported|a Cursor|Cursor login|Cursor auth URL)/.test(error.message)
-				? error.message
-				: "Cursor request failed";
-			return badRequest(message);
-		}
-	};
-}
-//#endregion
-
-//#region protobuf message builders (agent.v1 subset)
-const textEncoder = new TextEncoder();
-
-function bytesOf(value) {
-	return value instanceof Uint8Array ? value : new Uint8Array(value);
-}
-
-function encodeUserMessage({ text, messageId }) {
-	const writer = new Writer();
-	if (typeof text === "string" && text.length > 0) writer.string(1, text); // text
-	if (typeof messageId === "string" && messageId.length > 0) writer.string(2, messageId); // message_id
-	return writer.finish();
-}
-
-function encodeAssistantMessage(text) {
-	return new Writer().string(1, text).finish(); // text
-}
-
-/** ConversationStep { assistant_message = 1 } */
-function encodeAssistantStep(text) {
-	const inner = encodeAssistantMessage(text);
-	return new Writer().message(1, inner).finish();
-}
-
-/** AgentConversationTurnStructure { user_message = 1, steps = 2 } */
-function encodeAgentTurn(userBytes, stepBytes) {
-	const writer = new Writer();
-	if (userBytes.length > 0) writer.bytes(1, userBytes);
-	for (const step of stepBytes) writer.bytes(2, step);
-	return writer.finish();
-}
-
-/** ConversationTurnStructure { agent_conversation_turn = 1 } */
-function encodeTurnStructure(turnBytes) {
-	return new Writer().message(1, turnBytes).finish();
-}
-
-/** ModelDetails { model_id = 1, display_model_id = 3, display_name = 4 } */
-function encodeModelDetails(modelId) {
-	const writer = new Writer();
-	writer.string(1, modelId);
-	writer.string(3, modelId);
-	writer.string(4, modelId);
-	return writer.finish();
-}
-
-/**
- * ConversationHistoryImageContent { data = 1, mime_type = 2 }
- * wrapped as ConversationHistoryMessage.user.content[].image
- */
-function encodeConversationHistoryImage({ data, mimeType }) {
-	const image = new Writer().string(1, data);
-	if (mimeType) image.string(2, mimeType);
-	const content = new Writer().message(2, image.finish()).finish(); // UserContent.image
-	const user = new Writer().message(1, content).finish(); // UserMessage.content
-	return new Writer().message(1, user).finish(); // HistoryMessage.user
-}
-
-/** ConversationHistory { messages = 1 } with one history message per image. */
-function encodeConversationHistory(images) {
-	const writer = new Writer();
-	for (const image of images) writer.message(1, encodeConversationHistoryImage(image));
-	return writer.finish();
-}
-
-/** ConversationAction { user_message_action = 1 } → UserMessageAction { user_message = 1, conversation_history = 7 } */
-function encodeUserMessageAction(userBytes, images = []) {
-	const inner = new Writer().message(1, userBytes); // UserMessageAction.user_message
-	if (images.length > 0) inner.message(7, encodeConversationHistory(images));
-	return new Writer().message(1, inner.finish()).finish(); // ConversationAction
-}
-
-/** ConversationStateStructure — the durable conversation payload. */
-function encodeConversationState({ rootPromptBlobIds, turns }) {
-	const writer = new Writer();
-	for (const id of rootPromptBlobIds) writer.bytes(1, id); // root_prompt_messages_json
-	for (const turn of turns) writer.bytes(8, turn); // turns
-	return writer.finish();
-}
-
-/** AgentRunRequest { conversation_state=1, action=2, model_details=3, conversation_id=5 } */
-function encodeRunRequest({ conversationState, action, modelDetails, conversationId }) {
-	const writer = new Writer();
-	writer.message(1, conversationState);
-	writer.message(2, action);
-	writer.message(3, modelDetails);
-	if (conversationId) writer.string(5, conversationId);
-	return writer.finish();
-}
-
-/** AgentClientMessage { run_request = 1 } */
-function encodeRunMessage(runRequestBytes) {
-	return new Writer().message(1, runRequestBytes).finish();
-}
-
-/** AgentClientMessage { client_heartbeat = 7 } */
-function encodeHeartbeat() {
-	return new Writer().message(7, new Uint8Array(0)).finish();
-}
-
-/** KvClientMessage { id=1, get_blob_result=2 { blob_data=1 } } */
-export function encodeGetBlobResult(id, blobData) {
-	const inner = blobData === undefined ? new Uint8Array(0) : blobData;
-	const result = new Writer().bytes(1, inner).finish(); // GetBlobResult
-	const writer = new Writer();
-	writer.varint(1, id);
-	writer.message(2, result);
-	return writer.finish();
-}
-
-/** AgentClientMessage { kv_client_message = 3 } */
-export function encodeKvClientMessage(kvBytes) {
-	return new Writer().message(3, kvBytes).finish();
-}
-
-/** KvClientMessage { id=1, set_blob_result=3 } — acknowledge a server blob write. */
-export function encodeSetBlobResult(id) {
-	const writer = new Writer();
-	writer.varint(1, id);
-	writer.message(3, new Uint8Array(0)); // set_blob_result (empty success)
-	return writer.finish();
-}
-
-/** McpResult.success with one text content item. */
-export function encodeMcpResult({ content, isError = false }) {
-	return encodeMcpResultSuccess(String(content ?? ""), isError);
-}
-
-/** ExecClientMessage { id=1, exec_id=15, message=... } */
-export function encodeExecClientMessage(id, execId, messageField, messageBytes) {
-	const writer = new Writer();
-	writer.varint(1, id);
-	if (typeof execId === "string" && execId.length > 0) writer.string(15, execId);
-	if (messageBytes !== undefined) writer.message(messageField, messageBytes);
-	return writer.finish();
-}
-
-/** AgentClientMessage { exec_client_message = 2 } */
-export function encodeExecClientMessageEnvelope(execBytes) {
-	return new Writer().message(2, execBytes).finish();
-}
-
-/** RequestContextResult { success=1 { request_context=1 } } with tools. */
-export function encodeRequestContextResult(tools) {
-	const context = new Writer();
-	for (const tool of tools) context.message(7, tool); // RequestContext.tools
-	const success = new Writer().message(1, context.finish()).finish(); // RequestContextSuccess
-	return new Writer().message(1, success).finish(); // RequestContextResult
-}
-
-/** McpToolDefinition { name=1, description=2, input_schema=3, provider_identifier=4, tool_name=5 } */
-export function encodeMcpToolDefinition({ name, description, inputSchema, providerIdentifier, toolName }) {
-	const writer = new Writer();
-	writer.string(1, name);
-	writer.string(2, description ?? "");
-	writer.bytes(3, bytesOf(inputSchema));
-	writer.string(4, providerIdentifier ?? "dsh-cursor-subscription");
-	writer.string(5, toolName ?? name);
-	return writer.finish();
-}
-
-/** McpTextContent { text = 1 } */
-function encodeMcpTextContent(text) {
-	return new Writer().string(1, text).finish();
-}
-
-/** McpToolResultContentItem { text = 1 } */
-function encodeMcpToolResultContentItem(text) {
-	return new Writer().message(1, encodeMcpTextContent(text)).finish();
-}
-
-/** McpSuccess { content=1, is_error=2 } */
-function encodeMcpSuccess(text, isError) {
-	const writer = new Writer();
-	writer.message(1, encodeMcpToolResultContentItem(text));
-	writer.varint(2, isError ? 1 : 0);
-	return writer.finish();
-}
-
-/** McpResult { success=1 | error=2 } */
-function encodeMcpResultSuccess(text, isError) {
-	return new Writer().message(1, encodeMcpSuccess(text, isError)).finish();
-}
-
-function encodeMcpError(error) {
-	return new Writer().message(2, new Writer().string(1, error).finish()).finish();
-}
-
-/** ReadResult { rejected=3 { path=1, reason=2 } } */
-function encodeReadRejected(path, reason) {
-	const rejected = new Writer().string(1, path ?? "").string(2, reason).finish();
-	return new Writer().message(3, rejected).finish();
-}
-
-/** LsResult { rejected=3 { path=1, reason=2 } } */
-function encodeLsRejected(path, reason) {
-	const rejected = new Writer().string(1, path ?? "").string(2, reason).finish();
-	return new Writer().message(3, rejected).finish();
-}
-
-/** GrepResult { error=2 { error=1 } } */
-function encodeGrepError(error) {
-	return new Writer().message(2, new Writer().string(1, error).finish()).finish();
-}
-
-/** WriteResult { rejected = 6 } (oneof: success=1, permission_denied=3, no_space=4, error=5, rejected=6) */
-function encodeWriteRejected(path, reason) {
-	const rejected = new Writer().string(1, path ?? "").string(2, reason).finish();
-	return new Writer().message(6, rejected).finish();
-}
-
-/** DeleteResult { rejected = 6 } (oneof: success=1, file_not_found=2, not_file=3, permission_denied=4, file_busy=5, rejected=6) */
-function encodeDeleteRejected(path, reason) {
-	const rejected = new Writer().string(1, path ?? "").string(2, reason).finish();
-	return new Writer().message(6, rejected).finish();
-}
-
-/** ShellRejected { command=1, working_directory=2, reason=3, is_readonly=4 } */
-function encodeShellRejected(command, workingDirectory, reason) {
-	const writer = new Writer();
-	writer.string(1, command ?? "");
-	writer.string(2, workingDirectory ?? "");
-	writer.string(3, reason);
-	writer.varint(4, 0);
-	return writer.finish();
-}
-
-/** ShellResult { rejected = 4 } (oneof: success=1, failure=2, timeout=3, rejected=4) */
-function encodeShellRejectedResult(command, workingDirectory, reason) {
-	return new Writer().message(4, encodeShellRejected(command, workingDirectory, reason)).finish();
-}
-
-/**
- * ShellStream { rejected = 5 } — the reply type for `shellStreamArgs` execs
- * (streaming shell), carried by ExecClientMessage.shell_stream (field 14).
- */
-function encodeShellStreamRejected(command, workingDirectory, reason) {
-	return new Writer().message(5, encodeShellRejected(command, workingDirectory, reason)).finish();
-}
-
-/** BackgroundShellSpawnResult { rejected = 3 } (oneof: success=1, error=2, rejected=3, permission_denied=4) */
-function encodeBackgroundShellRejectedResult(command, workingDirectory, reason) {
-	return new Writer().message(3, encodeShellRejected(command, workingDirectory, reason)).finish();
-}
-
-/** FetchResult { error=2 { url=1, error=2 } } */
-function encodeFetchError(url, error) {
-	const inner = new Writer().string(1, url ?? "").string(2, error).finish();
-	return new Writer().message(2, inner).finish();
-}
-
-/** WriteShellStdinResult { error=2 { error=1 } } */
-function encodeWriteShellStdinError(error) {
-	return new Writer().message(2, new Writer().string(1, error).finish()).finish();
-}
-
-/** DiagnosticsResult { success = 1 } — empty success keeps the exec well-formed. */
-function encodeDiagnosticsResult() {
-	return new Writer().message(1, new Uint8Array(0)).finish();
-}
-//#endregion
-
-//#region protobuf message parsers (agent.v1 subset)
-const decoder = new TextDecoder();
-
-function readTag(reader) {
-	if (reader.done) return undefined;
-	const { field, wireType } = reader.tag();
-	return { field, wireType };
-}
-
-/** AgentServerMessage { interaction_update=1, exec_server_message=2, kv_server_message=4, ... } */
-export function decodeAgentServerMessage(bytes) {
-	const reader = new Reader(bytes);
-	while (!reader.done) {
-		const { field, wireType } = reader.tag();
-		if (wireType !== 2) {
-			reader.skip(wireType);
-			continue;
-		}
-		const payload = reader.bytes();
-		if (field === 1) return { case: "interactionUpdate", value: decodeInteractionUpdate(payload) };
-		if (field === 2) return { case: "execServerMessage", value: decodeExecServerMessage(payload) };
-		if (field === 3) return { case: "conversationCheckpointUpdate", value: payload };
-		if (field === 4) return { case: "kvServerMessage", value: decodeKvServerMessage(payload) };
-		// 5 = exec_server_control, 7 = interaction_query (ignored)
-	}
-	return { case: "unknown", value: undefined };
-}
-
-/** Field numbers inside one protobuf message; used to describe frames we do not decode. */
-function messageFields(bytes) {
-	const reader = new Reader(bytes);
-	const fields = [];
-	while (!reader.done) {
-		const { field, wireType } = reader.tag();
-		fields.push(field);
-		reader.skip(wireType);
-	}
-	return fields.join(",");
-}
-
-/** Field numbers of a nested message (`wanted` inside `bytes`). */
-function nestedMessageFields(bytes, wanted) {
-	const reader = new Reader(bytes);
-	while (!reader.done) {
-		const { field, wireType } = reader.tag();
-		if (wireType !== 2) {
-			reader.skip(wireType);
-			continue;
-		}
-		const payload = reader.bytes();
-		if (field === wanted) return messageFields(payload);
-	}
-	return "";
-}
-
-/**
- * One-line description of a server frame for the stall trace. Undecoded
- * frames report their protobuf field numbers so a server message the adapter
- * ignores is still visible when a run goes silent.
- */
-export function describeServerFrame(frame) {
-	if ((frame.flags & CONNECT_END_STREAM_FLAG) !== 0) {
-		return `endstream(${Buffer.from(frame.payload).toString("utf8").replace(/\s+/g, " ").slice(0, 60)})`;
-	}
-	try {
-		const message = decodeAgentServerMessage(frame.payload);
-		if (message.case === "interactionUpdate") {
-			const update = message.value;
-			if (update.type === "textDelta") return `text(${update.text.length})`;
-			if (update.type === "thinkingDelta") return `thinking(${update.text.length})`;
-			if (update.type === "tokenDelta") return `tokens(${update.tokens})`;
-			if (update.type === "partialToolCall") return "partialToolCall";
-			if (update.type === "unknown") return `interaction#${nestedMessageFields(frame.payload, 1)}`;
-			return update.type;
-		}
-		if (message.case === "conversationCheckpointUpdate") return `checkpoint(${message.value.length}B)`;
-		if (message.case === "kvServerMessage") return `kv:${message.value.case}`;
-		if (message.case === "execServerMessage") {
-			const exec = message.value;
-			return exec.case === "unknown"
-				? `exec:unknown#${exec.field ?? "?"}[${(exec.seen ?? []).join(",")}]`
-				: `exec:${exec.case}`;
-		}
-		return `message#${messageFields(frame.payload)}`;
-	} catch (error) {
-		return `undecodable(${(error instanceof Error ? error.message : String(error)).slice(0, 40)})`;
-	}
-}
-
-/**
- * Prompt tokens Cursor reports for the conversation it is holding.
- *
- * `ConversationStateStructure.token_details.used_tokens` rides in every
- * `conversation_checkpoint_update`, so the run that resumes that checkpoint can
- * report the real prompt size. Without it the adapter reported a placeholder
- * zero, and the context gauge showed nothing however large the model's
- * conversation had grown.
- *
- * @param checkpoint Raw checkpoint bytes from the last update, if any.
- * @returns Used tokens, or undefined when the field is absent or unreadable.
- */
-export function decodeCheckpointUsedTokens(checkpoint) {
-	if (checkpoint === undefined) return undefined;
-	try {
-		const reader = new Reader(checkpoint);
-		while (!reader.done) {
-			const { field, wireType } = reader.tag();
-			if (field === 5 && wireType === 2) {
-				const details = new Reader(reader.bytes());
-				while (!details.done) {
-					const tag = details.tag();
-					if (tag.field === 1 && tag.wireType === 0) return details.varint();
-					details.skip(tag.wireType);
-				}
-				return undefined;
-			}
-			reader.skip(wireType);
-		}
-	} catch {
-		// A checkpoint this build cannot parse still resumes the conversation;
-		// the usage figure stays unknown rather than failing the step.
-	}
-	return undefined;
-}
-
-/** InteractionUpdate { text_delta=1, thinking_delta=4, token_delta=8, turn_ended=14, ... } */
-export function decodeInteractionUpdate(bytes) {
-	const reader = new Reader(bytes);
-	while (!reader.done) {
-		const { field, wireType } = reader.tag();
-		if (wireType !== 2) {
-			reader.skip(wireType);
-			continue;
-		}
-		const payload = reader.bytes();
-		if (field === 1) return { type: "textDelta", text: decodeTextDelta(payload) }; // TextDeltaUpdate.text=1
-		if (field === 4) return { type: "thinkingDelta", text: decodeTextDelta(payload) }; // ThinkingDeltaUpdate.text=1
-		if (field === 5) return { type: "thinkingCompleted", durationMs: decodeThinkingDuration(payload) }; // ThinkingCompletedUpdate.thinking_duration_ms=1
-		if (field === 8) return { type: "tokenDelta", tokens: decodeTokenDelta(payload) }; // TokenDeltaUpdate.tokens=1
-		if (field === 14) return { type: "turnEnded" };
-		if (field === 2) return { type: "toolCallStarted" };
-		if (field === 3) return { type: "toolCallCompleted" };
-		if (field === 7) return { type: "partialToolCall", ...decodePartialToolCall(payload) };
-		if (field === 15) return { type: "toolCallDelta" };
-		if (field === 13) return { type: "heartbeat" };
-		// others (summary*, step_*, user_message_appended, shell_output_delta) ignored
-	}
-	return { type: "unknown" };
-}
-
-function decodeThinkingDuration(bytes) {
-	const reader = new Reader(bytes);
-	while (!reader.done) {
-		const { field, wireType } = reader.tag();
-		if (field === 1 && wireType === 0) return reader.varint();
-		reader.skip(wireType);
-	}
-	return 0;
-}
-
-function decodeTextDelta(bytes) {
-	const reader = new Reader(bytes);
-	while (!reader.done) {
-		const { field, wireType } = reader.tag();
-		if (field === 1 && wireType === 2) return reader.string();
-		reader.skip(wireType);
-	}
-	return "";
-}
-
-function decodeTokenDelta(bytes) {
-	const reader = new Reader(bytes);
-	while (!reader.done) {
-		const { field, wireType } = reader.tag();
-		if (field === 1 && wireType === 0) return reader.varint();
-		reader.skip(wireType);
-	}
-	return 0;
-}
-
-/** PartialToolCallUpdate { call_id=1, args_text_delta=3 } */
-function decodePartialToolCall(bytes) {
-	const reader = new Reader(bytes);
-	let callId = "";
-	let argsTextDelta = "";
-	while (!reader.done) {
-		const { field, wireType } = reader.tag();
-		if (field === 1 && wireType === 2) callId = reader.string();
-		else if (field === 3 && wireType === 2) argsTextDelta = reader.string();
-		else reader.skip(wireType);
-	}
-	return { callId, argsTextDelta };
-}
-
-/** KvServerMessage { id=1, get_blob_args=2 { blob_id=1 } | set_blob_args=3 { blob_id=1, blob_data=2 } } */
-export function decodeKvServerMessage(bytes) {
-	const reader = new Reader(bytes);
-	let id = 0;
-	let blobId;
-	while (!reader.done) {
-		const { field, wireType } = reader.tag();
-		if (field === 1 && wireType === 0) id = reader.varint();
-		else if (field === 2 && wireType === 2) blobId = decodeBlobId(reader.bytes());
-		else if (field === 3 && wireType === 2) {
-			return { id, case: "setBlobArgs", ...decodeSetBlobArgs(reader.bytes()) };
-		} else reader.skip(wireType);
-	}
-	return blobId === undefined ? { id, case: "unknown" } : { id, case: "getBlobArgs", blobId };
-}
-
-/** SetBlobArgs { blob_id=1, blob_data=2 } */
-function decodeSetBlobArgs(bytes) {
-	const reader = new Reader(bytes);
-	let blobId;
-	let blobData;
-	while (!reader.done) {
-		const { field, wireType } = reader.tag();
-		if (field === 1 && wireType === 2) blobId = reader.bytes();
-		else if (field === 2 && wireType === 2) blobData = reader.bytes();
-		else reader.skip(wireType);
-	}
-	return { blobId, blobData };
-}
-
-function decodeBlobId(bytes) {
-	const reader = new Reader(bytes);
-	while (!reader.done) {
-		const { field, wireType } = reader.tag();
-		if (field === 1 && wireType === 2) return reader.bytes();
-		reader.skip(wireType);
-	}
-	return undefined;
-}
-
-/**
- * Field number of `ExecServerMessage.span_context`. It is a plain (non-oneof)
- * field, so a message carrying only ids, a span context, and an exec variant
- * this build does not know must not mistake it for the exec itself.
- */
-const EXEC_SPAN_CONTEXT_FIELD = 19;
-
-/** ExecServerMessage { id=1, exec_id=15, request_context_args=10, mcp_args=11, ... } */
-export function decodeExecServerMessage(bytes) {
-	const reader = new Reader(bytes);
-	let id = 0;
-	let execId = "";
-	// Kept for the trace and for rejectionFor(): a server exec this build does
-	// not decode must still be answered, so its field number has to survive.
-	// `seen` lists every top-level field of the message, because an unknown
-	// exec can also carry non-oneof fields (ids, span context, tracing).
-	let unknownField;
-	const seen = [];
-	while (!reader.done) {
-		const { field, wireType } = reader.tag();
-		seen.push(field);
-		if (field === 1 && wireType === 0) {
-			id = reader.varint();
-		} else if (field === 15 && wireType === 2) {
-			execId = reader.string();
-		} else if (wireType === 2) {
-			const payload = reader.bytes();
-			if (field === 10) return { id, execId, case: "requestContextArgs" };
-			if (field === 11) return { id, execId, case: "mcpArgs", args: decodeMcpArgs(payload) };
-			if (field === 2) return { id, execId, case: "shellArgs", path: decodeSinglePathArg(payload) };
-			if (field === 3) return { id, execId, case: "writeArgs", path: decodeSinglePathArg(payload) };
-			if (field === 4) return { id, execId, case: "deleteArgs", path: decodeSinglePathArg(payload) };
-			if (field === 5) return { id, execId, case: "grepArgs" };
-			if (field === 7) return { id, execId, case: "readArgs", path: decodeReadArgsPath(payload) };
-			if (field === 8) return { id, execId, case: "lsArgs", path: decodeSinglePathArg(payload) };
-			if (field === 9) return { id, execId, case: "diagnosticsArgs" };
-			if (field === 14) return { id, execId, case: "shellStreamArgs", path: decodeSinglePathArg(payload) };
-			if (field === 16) return { id, execId, case: "backgroundShellSpawnArgs" };
-			if (field === 17) return { id, execId, case: "listMcpResourcesExecArgs" };
-			if (field === 18) return { id, execId, case: "readMcpResourceExecArgs" };
-			if (field === 20) return { id, execId, case: "fetchArgs", url: decodeFetchUrl(payload) };
-			if (field === 21) return { id, execId, case: "recordScreenArgs" };
-			if (field === 22) return { id, execId, case: "computerUseArgs" };
-			if (field === 23) return { id, execId, case: "writeShellStdinArgs" };
-			if (field !== EXEC_SPAN_CONTEXT_FIELD) unknownField ??= field;
-		} else {
-			reader.skip(wireType);
-		}
-	}
-	return { id, execId, case: "unknown", field: unknownField, seen };
-}
-
-function decodeSinglePathArg(bytes) {
-	const reader = new Reader(bytes);
-	while (!reader.done) {
-		const { field, wireType } = reader.tag();
-		if (field === 1 && wireType === 2) return reader.string();
-		reader.skip(wireType);
-	}
-	return "";
-}
-
-function decodeReadArgsPath(bytes) {
-	return decodeSinglePathArg(bytes);
-}
-
-function decodeFetchUrl(bytes) {
-	const reader = new Reader(bytes);
-	while (!reader.done) {
-		const { field, wireType } = reader.tag();
-		if (field === 1 && wireType === 2) return reader.string();
-		reader.skip(wireType);
-	}
-	return "";
-}
-
-/** McpArgs { name=1, args=2 (map<string,bytes>), tool_call_id=3, provider_identifier=4, tool_name=5 } */
-export function decodeMcpArgs(bytes) {
-	const reader = new Reader(bytes);
-	let name = "";
-	let toolCallId = "";
-	let providerIdentifier = "";
-	let toolName = "";
-	const args = {};
-	while (!reader.done) {
-		const { field, wireType } = reader.tag();
-		if (field === 1 && wireType === 2) name = reader.string();
-		else if (field === 2 && wireType === 2) {
-			const entry = new Reader(reader.bytes());
-			let key = "";
-			let value;
-			while (!entry.done) {
-				const tag = entry.tag();
-				if (tag.field === 1 && tag.wireType === 2) key = entry.string();
-				else if (tag.field === 2 && tag.wireType === 2) value = entry.bytes();
-				else entry.skip(tag.wireType);
-			}
-			if (key) args[key] = value;
-		} else if (field === 3 && wireType === 2) toolCallId = reader.string();
-		else if (field === 4 && wireType === 2) providerIdentifier = reader.string();
-		else if (field === 5 && wireType === 2) toolName = reader.string();
-		else reader.skip(wireType);
-	}
-	return { name, toolCallId, providerIdentifier, toolName, args };
-}
-
-/** GetUsableModelsResponse { models = 1: repeated ModelDetails } */
-export function decodeUsableModels(bytes) {
-	const reader = new Reader(bytes);
-	const models = [];
-	while (!reader.done) {
-		const { field, wireType } = reader.tag();
-		if (field === 1 && wireType === 2) {
-			const model = decodeModelDetails(reader.bytes());
-			if (model?.id) models.push(model);
-		} else {
-			reader.skip(wireType);
-		}
-	}
-	return models;
-}
-
-/** ModelDetails { model_id=1, display_model_id=3, display_name=4, display_name_short=5, aliases=6 } */
-function decodeModelDetails(bytes) {
-	const reader = new Reader(bytes);
-	const model = { id: "", name: "" };
-	while (!reader.done) {
-		const { field, wireType } = reader.tag();
-		if (wireType === 2) {
-			if (field === 1) model.id = reader.string();
-			else if (field === 3) model.displayModelId = reader.string();
-			else if (field === 4) model.name = reader.string();
-			else if (field === 5) model.displayNameShort = reader.string();
-			else if (field === 6) reader.bytes(); // aliases
-			else reader.bytes();
-		} else {
-			reader.skip(wireType);
-		}
-	}
-	return model;
-}
-//#endregion
-
-//#region Connect framing
-/** Connect frame: [1-byte flags][4-byte big-endian length][payload]. */
-export function frameEncode(payload, flags = 0) {
-	const out = new Uint8Array(5 + payload.length);
-	const view = new DataView(out.buffer);
-	view.setUint8(0, flags);
-	view.setUint32(1, payload.length, false);
-	out.set(payload, 5);
-	return out;
-}
-
-export const CONNECT_END_STREAM_FLAG = 0b00000010;
-export const CONNECT_COMPRESSED_FLAG = 0b00000001;
-
-/** Upper bound for a decompressed Connect frame; guards against zip bombs. */
-export const MAX_CONNECT_FRAME_BYTES = 64 * 1024 * 1024;
-
-/**
- * Incremental Connect frame parser fed by response chunks.
- * Yields { flags, payload } objects. Frames flagged compressed (bit 0) are
- * gzip-decompressed here so consumers only ever see plain payloads.
- */
-export class ConnectFrameReader {
-	constructor() {
-		this.buffer = new Uint8Array(0);
-		this.frames = [];
-		this.waiters = [];
-		this.ended = false;
-		this.paused = false;
-		this.error = undefined;
-	}
-
-	push(chunk) {
-		const combined = new Uint8Array(this.buffer.length + chunk.length);
-		combined.set(this.buffer);
-		combined.set(chunk, this.buffer.length);
-		this.buffer = combined;
-		while (this.buffer.length >= 5) {
-			const view = new DataView(this.buffer.buffer, this.buffer.byteOffset, 5);
-			const flags = view.getUint8(0);
-			const length = view.getUint32(1, false);
-			if (this.buffer.length < 5 + length) break;
-			const payload = this.buffer.slice(5, 5 + length);
-			this.buffer = this.buffer.slice(5 + length);
-			let data = payload;
-			if ((flags & CONNECT_COMPRESSED_FLAG) !== 0) {
-				try {
-					data = gunzipSync(Buffer.from(payload), { maxOutputLength: MAX_CONNECT_FRAME_BYTES });
-				} catch (error) {
-					this.fail(new Error("Cursor sent an unreadable compressed frame", { cause: error }));
-					return;
-				}
-			}
-			this.#enqueue({ flags: flags & ~CONNECT_COMPRESSED_FLAG, payload: data });
-		}
-	}
-
-	#enqueue(frame) {
-		if (this.waiters.length > 0) {
-			const waiter = this.waiters.shift();
-			waiter.resolve(frame);
-			return;
-		}
-		this.frames.push(frame);
-	}
-
-	finish() {
-		this.ended = true;
-		for (const waiter of this.waiters.splice(0)) waiter.resolve(undefined);
-	}
-
-	/**
-	 * Stop blocking for new frames without ending the stream: buffered frames
-	 * are still returned, and `resume()` restores normal waiting. A Cursor run
-	 * stays alive across DSH steps, so a reader that is paused to end one step
-	 * must accept frames again when the next step continues the same run.
-	 */
-	pause() {
-		this.paused = true;
-		for (const waiter of this.waiters.splice(0)) waiter.resolve(undefined);
-	}
-
-	resume() {
-		this.paused = false;
-	}
-
-	fail(error) {
-		this.error = error;
-		this.ended = true;
-		for (const waiter of this.waiters.splice(0)) waiter.reject(error);
-	}
-
-	async next() {
-		if (this.frames.length > 0) return this.frames.shift();
-		if (this.paused) return undefined;
-		if (this.ended) {
-			if (this.error !== undefined) throw this.error;
-			return undefined;
-		}
-		return new Promise((resolve, reject) => this.waiters.push({ resolve, reject }));
-	}
-}
-//#endregion
-
-//#region agent transport
-const AGENT_HEADERS = {
-	"content-type": "application/connect+proto",
-	"connect-protocol-version": "1",
-	"connect-accept-encoding": "gzip",
-	te: "trailers",
-	"x-ghost-mode": "true",
-	"x-cursor-client-version": CURSOR_CLIENT_VERSION,
-	"x-cursor-client-type": "cli",
-};
-
-/**
- * One bidirectional agent run over HTTP/2. Writes Connect-framed
- * `AgentClientMessage`s and surfaces `AgentServerMessage`s through the frame
- * reader. Heartbeats keep the run alive.
- */
-export class AgentRun {	constructor(accessToken, options = {}) {
-		this.accessToken = accessToken;
-		this.baseUrl = options.baseUrl ?? CURSOR_BASE_URL;
-		this.session = undefined;
-		this.frames = new ConnectFrameReader();
-		this.responseStatus = undefined;
-		this.responseContentType = undefined;
-		this.trailers = {};
-		this.finished = false;
-	}
-
-	async start() {
-		const url = new URL(CURSOR_RUN_PATH, this.baseUrl);
-		const headers = {
-			":method": "POST",
-			":path": url.pathname,
-			authorization: `Bearer ${this.accessToken}`,
-			...AGENT_HEADERS,
-		};
-		this.responsePromise = new Promise((resolve, reject) => {
-			this.resolveResponse = resolve;
-			this.rejectResponse = reject;
-		});
-		// The original promise remains rejectable for waitForResponse(); this
-		// observer only prevents a close-before-wait race from becoming unhandled.
-		void this.responsePromise.catch(() => {});
-		this.session = http2.connect(this.baseUrl);
-		this.stream = this.session.request(headers);
-		// NOTE: never call stream.setEncoding() here — Node's setEncoding(null)
-		// still decodes binary frames as UTF-8, corrupting any non-UTF-8 byte
-		// (e.g. blob ids) into U+FFFD replacement characters.
-		this.stream.on("response", (headers) => {
-			this.responseStatus = Number(headers[":status"]);
-			this.responseContentType = typeof headers["content-type"] === "string" ? headers["content-type"] : undefined;
-			this.#resolveResponse(this.responseStatus);
-		});
-		this.stream.on("data", (chunk) => this.frames.push(Buffer.from(chunk)));
-		this.stream.on("trailers", (trailers) => {
-			this.trailers = trailers;
-		});
-		this.stream.on("end", () => this.frames.finish());
-		this.stream.on("close", () => {
-			if (this.responseStatus === undefined) this.#rejectResponse(new Error("Cursor HTTP stream closed before response"));
-		});
-		this.stream.on("error", (error) => {
-			this.#rejectResponse(error);
-			this.frames.fail(error);
-		});
-		this.session.on("error", (error) => {
-			this.#rejectResponse(error);
-			this.frames.fail(error);
-		});
-	}
-
-	#resolveResponse(status) {
-		const resolve = this.resolveResponse;
-		this.resolveResponse = undefined;
-		this.rejectResponse = undefined;
-		resolve?.(status);
-	}
-
-	#rejectResponse(error) {
-		const reject = this.rejectResponse;
-		this.resolveResponse = undefined;
-		this.rejectResponse = undefined;
-		reject?.(error);
-	}
-
-	waitForResponse(timeoutMs) {
-		if (this.responsePromise === undefined) throw new Error("Cursor AgentRun has not started");
-		if (timeoutMs === undefined) return this.responsePromise;
-		return new Promise((resolve, reject) => {
-			const timer = setTimeout(() => reject(new Error("Cursor HTTP response timeout")), timeoutMs);
-			timer.unref?.();
-			this.responsePromise.then(resolve, reject).finally(() => clearTimeout(timer));
-		});
-	}
-
-	write(payload) {
-		if (this.finished || this.stream === undefined || this.stream.destroyed) return false;
-		try {
-			this.stream.write(Buffer.from(payload));
-			return true;
-		} catch {
-			return false;
-		}
-	}
-
-	writeMessage(bytes) {
-		return this.write(frameEncode(bytes));
-	}
-
-	startHeartbeat() {
-		this.heartbeat = setInterval(() => {
-			this.writeMessage(encodeHeartbeat());
-		}, HEARTBEAT_INTERVAL_MS);
-		this.heartbeat.unref?.();
-	}
-
-	/** Fail the run with a terminal error and tear down the connection. */
-	abort(error) {
-		if (this.finished) return;
-		this.#rejectResponse(error);
-		this.frames.fail(error);
-		this.close();
-	}
-
-	close() {
-		if (this.finished) return;
-		this.#rejectResponse(new Error("Cursor AgentRun closed before response"));
-		this.finished = true;
-		if (this.heartbeat !== undefined) clearInterval(this.heartbeat);
-		if (!this.frames.ended) this.frames.finish();
-		try {
-			this.stream?.close();
-		} catch {}
-		// Force-destroy the session so no socket keeps the process alive after a
-		// completed run; the agent response is fully consumed by then.
-		try {
-			this.session.destroy();
-		} catch {}
-	}
-}
-//#endregion
-
-//#region conversation building
-function sha256(bytes) {
-	return createHash("sha256").update(bytes).digest();
-}
-
-function collectImageAttachments(options) {
-	const refs = new Map();
-	const visit = (blocks) => {
-		for (const block of blocks ?? []) {
-			if (block.type === "image") refs.set(block.attachment.attachmentId, block.attachment);
-			else if (block.type === "tool-result") visit(block.content);
-		}
-	};
-	for (const message of options.messages ?? []) if (message.role === "user") visit(message.content);
-	return [...refs.values()];
-}
-
-export async function prepareCursorImages(options, attachments) {
-	const refs = collectImageAttachments(options);
-	if (refs.length === 0) return [];
-	if (attachments === undefined) throw new LlmError("Cursor image input requires the durable attachment service", "UNSUPPORTED_CONTENT");
-	const stored = await Promise.all(refs.map((ref) => attachments.readImage(ref, options.signal)));
-	return stored.map(({ ref, data }) => ({ data: Buffer.from(data).toString("base64"), mimeType: ref.mediaType }));
-}
-
-function flattenBlocks(content) {
-	let text = "";
-	for (const block of content ?? []) {
-		if (block.type === "text") text += block.text;
-		else if (block.type === "tool-result") {
-			const callId = block.toolCallId ? ` for ${block.toolCallId}` : "";
-			text += `\n[TOOL RESULT${callId}]\n${flattenBlocks(block.content)}`;
-		}
-	}
-	return text;
-}
-
-function renderAssistant(content) {
-	let text = "";
-	for (const block of content ?? []) {
-		if (block.type === "text") text += block.text;
-		else if (block.type === "tool-call") {
-			text += `\n<tool_call id="${block.id}" name="${block.name}">${block.arguments}</tool_call>`;
-		} else if (block.type === "reasoning") {
-			text += `\n<thinking>${block.text}</thinking>`;
-		}
-	}
-	return text.trim();
-}
-
-function sanitizeReplayedAssistantText(text) {
-	let omitted = false;
-	const sanitized = String(text ?? "").replace(/<tool_call\b[\s\S]*?<\/tool_call>/gi, () => {
-		omitted = true;
-		return "";
-	}).replace(/<\/?invoke>|<\|eos\|>/gi, "").trim();
-	if (!omitted) return sanitized;
-	return `${sanitized}${sanitized ? "\n" : ""}[Previous textual tool-call markup omitted from replay.]`;
-}
-
-/**
- * How much of a replayed tool request to keep. Enough for the arguments that are
- * state — a plan (`todo_write`) is its list, and a repeated `grep` is avoided
- * only if the model can see what it already asked for — while a `write` of a
- * whole file stays bounded.
- */
-const REPLAYED_CALL_CHARS = 600;
-
-/** Replay assistant history without teaching models to imitate XML tool tags. */
-function renderColdStartAssistant(content) {
-	const parts = [];
-	for (const block of content ?? []) {
-		if (block.type === "text") {
-			const text = sanitizeReplayedAssistantText(block.text);
-			if (text) parts.push(text);
-		} else if (block.type === "tool-call") {
-			// The request's arguments are part of the history, not decoration: a
-			// replay that keeps only the tool name leaves the model unable to
-			// recall the plan it wrote or the search it already ran, and it
-			// re-derives both — reworded, which is what plan churn looks like.
-			const args = typeof block.arguments === "string" ? block.arguments : JSON.stringify(block.arguments ?? "");
-			const shown = args.length > REPLAYED_CALL_CHARS ? `${args.slice(0, REPLAYED_CALL_CHARS)}…` : args;
-			parts.push(`[Previous tool request: ${block.name}${shown.length > 0 ? ` ${shown}` : ""}]`);
-		}
-	}
-	return parts.join("\n").trim();
-}
-
-function actionBoundary(messages) {
-	for (let index = messages.length - 1; index >= 0; index--) {
-		if (messages[index].role === "assistant") return index + 1;
-	}
-	return 0;
-}
-
-/** Combine the current DSH turn's adjacent user/context messages into one action. */
-function currentActionText(options) {
-	const messages = options.messages ?? [];
-	const boundary = actionBoundary(messages);
-	const parts = [];
-	for (let index = boundary; index < messages.length; index++) {
-		const message = messages[index];
-		if (message.role !== "user") continue;
-		const text = flattenBlocks(message.content).trim();
-		if (text.length > 0) parts.push(text);
-	}
-	if (parts.length > 0) return parts.join("\n\n");
-	for (let index = messages.length - 1; index >= 0; index--) {
-		if (messages[index].role !== "user") continue;
-		const text = flattenBlocks(messages[index].content).trim();
-		if (text.length > 0) return text;
-	}
-	return "";
-}
-
-/**
- * Markers this harness leaves in a result when the sandbox refused the work.
- *
- * They call for different answers, and telling them apart matters: the
- * `SetNamedSecurityInfoW` failure is the ACL grant the confinement itself needs
- * and cannot get, so *no* narrower mode can start and only a mode without that
- * grant can run the call. `[sandbox: ... denied under <mode> mode]` is a
- * refusal from a sandbox that did start, where one step wider may well suffice.
- * Naming the narrowest mode first for the ACL failure is what turned a single
- * refusal into a retry loop: the model asked for `workspace-write` — the mode it
- * was already in — and failed the same way 101 times in one session.
- */
-const SANDBOX_ACL_FAILURE = /SetNamedSecurityInfoW failed/i;
-const SANDBOX_POLICY_DENIAL = /\[sandbox:[^\]]*denied/i;
-/**
- * DSH's escalation rule, restated where the refusal lands.
- *
- * Every provider gets this rule — retry the denied call once with
- * `sandbox_permissions` and a justification, which raises an approval prompt —
- * but only the `pwsh` description spells it out, and the refusal result carries
- * nothing at all. A model that does not recall the rule from a tool description
- * reads the refusal as "this tool does not work here" and stops using it, which
- * is how a Cursor turn ends up saying local tools are unavailable instead of
- * asking the user for permission.
- */
-const SANDBOX_ACL_HINT =
-	"\n\n[cursor] The DSH sandbox cannot start in this workspace — its own ACL grant fails "
-	+ "(`SetNamedSecurityInfoW`) — so no narrower mode can succeed. Retry this exact call once with "
-	+ "sandbox_permissions=\"danger-full-access\" and a one-sentence justification; DSH then asks the user to "
-	+ "approve it. Do not retry it in a narrower mode, and if the retry is refused, stop and report the refusal "
-	+ "instead of trying again.";
-const SANDBOX_POLICY_HINT =
-	"\n\n[cursor] The DSH sandbox refused this call, and that is recoverable: retry this exact call once with "
-	+ "sandbox_permissions set to the narrowest wider mode that suffices (workspace-write, or danger-full-access "
-	+ "when that is not enough) plus a one-sentence justification. DSH then asks the user to approve it. If the "
-	+ "retry is refused as well, stop and report it instead of trying another mode.";
-
-/** Whether one result is the harness refusing, rather than the tool failing. */
-function isSandboxRefusal(content) {
-	return SANDBOX_ACL_FAILURE.test(content) || SANDBOX_POLICY_DENIAL.test(content);
-}
-
-/**
- * Result text as the model should read it: a sandbox refusal gains the rule that
- * turns it back into a request the user can approve.
- * @param content Result text on its way to the model.
- * @returns The text, with the escalation rule a refusal calls for appended.
- */
-function withSandboxHint(content) {
-	if (content.includes("sandbox_permissions")) return content;
-	if (SANDBOX_ACL_FAILURE.test(content)) return `${content}${SANDBOX_ACL_HINT}`;
-	if (SANDBOX_POLICY_DENIAL.test(content)) return `${content}${SANDBOX_POLICY_HINT}`;
-	return content;
-}
-
-/** Extract DSH tool-result blocks for resuming a live Cursor exec bridge. */
-function collectToolResults(options) {
-	const results = [];
-	for (const message of options.messages ?? []) {
-		for (const block of message.content ?? []) {
-			if (block.type !== "tool-result") continue;
-			const content = flattenBlocks(block.content).trim();
-			const sandboxDenied = isSandboxRefusal(content);
-			results.push({
-				toolCallId: block.toolCallId ?? message.source?.callId,
-				content: withSandboxHint(content),
-				isError: block.isError === true,
-				sandboxDenied,
-			});
-		}
-	}
-	return results;
-}
-
-/**
- * The compaction DSH applied to the history carried by one request, if any.
- *
- * DSH frees context by replacing a span of the conversation with a summary and
- * marking that message `source.kind = "compact-checkpoint"` with the id of the
- * compaction. The adapter compares that id with the one its persisted Cursor
- * conversation was built from: equal means the checkpoint still describes the
- * history DSH is sending, a new id means DSH has rewritten it.
- *
- * @param options The DSH request whose message history is being sent.
- * @returns The newest compaction id in the history, or undefined when it has none.
- */
-function compactionIdOf(options) {
-	let id;
-	for (const message of options.messages ?? []) {
-		if (message.source?.kind !== "compact-checkpoint") continue;
-		id = message.source.compactionId ?? "compact-checkpoint";
-	}
-	return id;
-}
-
-function decodeXmlText(value) {
-	return String(value ?? "")
-		.replace(/&quot;/gi, '"')
-		.replace(/&apos;/gi, "'")
-		.replace(/&lt;/gi, "<")
-		.replace(/&gt;/gi, ">")
-		.replace(/&amp;/gi, "&");
-}
-
-function parseTagAttributes(source) {
-	const attributes = {};
-	const pattern = /([A-Za-z_][\w-]*)\s*=\s*"([\s\S]*?)"/g;
-	for (let match; (match = pattern.exec(source)) !== null;) attributes[match[1]] = decodeXmlText(match[2]);
-	return attributes;
-}
-
-function resolveTextTool(candidate, tools) {
-	const available = tools ?? [];
-	const normalized = String(candidate ?? "").toLowerCase();
-	let tool = available.find((entry) => {
-		const name = String(entry.name).toLowerCase();
-		return normalized === name || normalized.endsWith(`_${name}`);
-	});
-	if (tool !== undefined) return tool;
-	const aliases = {
-		shell: ["pwsh", "bash"],
-		read: ["read"],
-		glob: ["glob"],
-		grep: ["grep"],
-		write: ["write"],
-		edit: ["edit"],
-	};
-	for (const name of aliases[normalized] ?? []) {
-		tool = available.find((entry) => String(entry.name).toLowerCase() === name);
-		if (tool !== undefined) return tool;
-	}
-	return undefined;
-}
-
-function normalizeTextToolArguments(tool, raw) {
-	const args = { ...raw };
-	if (tool.name === "read" && args.file_path === undefined && args.path !== undefined) {
-		args.file_path = args.path;
-		delete args.path;
-	}
-	if (tool.name === "glob") {
-		if (args.pattern === undefined && args.glob_pattern !== undefined) args.pattern = args.glob_pattern;
-		if (args.path === undefined && args.target_directory !== undefined) args.path = args.target_directory;
-		delete args.glob_pattern;
-		delete args.target_directory;
-	}
-	for (const [key, value] of Object.entries(args)) {
-		const type = tool.parameters?.properties?.[key]?.type;
-		if ((type === "number" || type === "integer") && typeof value === "string" && value.trim() !== "") {
-			const number = Number(value);
-			if (Number.isFinite(number)) args[key] = number;
-		} else if (type === "boolean" && typeof value === "string") {
-			if (value.toLowerCase() === "true") args[key] = true;
-			else if (value.toLowerCase() === "false") args[key] = false;
-		} else if ((type === "object" || type === "array") && typeof value === "string") {
-			try { args[key] = JSON.parse(value); } catch {}
-		}
-	}
-	return args;
-}
-
-/**
- * Recover Cursor models that print tool-call XML as text instead of emitting
- * an execServerMessage.mcpArgs frame. This is a fallback for contaminated or
- * compacted conversations; protocol-native MCP calls remain the primary path.
- */
-export function parseTextToolCalls(text, tools) {
-	const calls = [];
-	const pattern = /<tool_call\b([^>]*)>([\s\S]*?)<\/tool_call>/gi;
-	for (let match; (match = pattern.exec(String(text ?? ""))) !== null;) {
-		const attributes = parseTagAttributes(match[1]);
-		const tool = resolveTextTool(attributes.name ?? attributes.id, tools);
-		if (tool === undefined) continue;
-		const rawArgs = {};
-		const parameterPattern = /<parameter\s+name="([^"]+)"\s*>([\s\S]*?)<\/parameter>/gi;
-		for (let parameter; (parameter = parameterPattern.exec(match[2])) !== null;) {
-			rawArgs[parameter[1]] = decodeXmlText(parameter[2].trim());
-		}
-		for (const [key, value] of Object.entries(attributes)) {
-			if (key !== "id" && key !== "name") rawArgs[key] = value;
-		}
-		if (Object.keys(rawArgs).length === 0) {
-			const body = decodeXmlText(match[2]).trim();
-			if (body.startsWith("{")) {
-				try { Object.assign(rawArgs, JSON.parse(body)); } catch {}
-			}
-		}
-		calls.push({
-			id: `text-tool-${crypto.randomUUID()}`,
-			name: tool.name,
-			arguments: JSON.stringify(normalizeTextToolArguments(tool, rawArgs)),
-		});
-	}
-	return calls;
-}
-
-function coldStartLabel(message) {
-	if (message.role === "assistant") return "ASSISTANT";
-	if ((message.content ?? []).some((block) => block.type === "tool-result")) return "TOOL RESULT";
-	if (message.source?.kind === "plugin") return "RUNTIME CONTEXT";
-	return "USER";
-}
-
-/**
- * Rehydrate a DSH conversation after process restart without using Cursor's
- * field-8 turns (the current server interprets those entries as blob ids).
- */
-function coldStartActionText(options) {
-	const entries = [];
-	for (const message of options.messages ?? []) {
-		if (message.role === "system") continue;
-		const text = message.role === "assistant" ? renderColdStartAssistant(message.content) : withSandboxHint(flattenBlocks(message.content).trim());
-		if (text.length === 0) continue;
-		entries.push({ label: coldStartLabel(message), text });
-	}
-	if (entries.length === 0) return "";
-	if (entries.length === 1 && entries[0].label === "USER") return entries[0].text;
-	return [
-		"Continue the DSH conversation below. Treat entries according to their labels. Respond to the final USER request; RUNTIME CONTEXT and TOOL RESULT entries provide context only.",
-		...entries.map((entry) => `[${entry.label}]\n${entry.text}`),
-	].join("\n\n");
-}
-
-/**
- * Build a fresh state with root prompt blobs only. Current Cursor servers reject
- * hand-encoded field-8 turns, so cold starts carry textual history in the action.
- */
-export function buildInitialConversationState(options) {
-	const systemParts = [options.system ?? ""];
-	for (const message of options.messages ?? []) {
-		if (message.role === "system") systemParts.push(flattenBlocks(message.content));
-	}
-	const systemText = systemParts.filter((part) => part.length > 0).join("\n").trim();
-	const blobStore = new Map();
-	const rootPromptBlobIds = [];
-	if (systemText.length > 0) {
-		const payload = textEncoder.encode(JSON.stringify({ role: "system", content: systemText }));
-		const id = sha256(payload);
-		blobStore.set(Buffer.from(id).toString("hex"), payload);
-		rootPromptBlobIds.push(id);
-	}
-	return {
-		conversationState: encodeConversationState({ rootPromptBlobIds, turns: [], systemText }),
-		blobStore,
-		systemText,
-	};
-}
-
-/**
- * Fold DSH history into Cursor's turn-based conversation state.
- *
- * The last user message becomes the new `ConversationAction`; everything
- * before it is the durable turn history. Tool results are rendered as user
- * messages so a follow-up run after DSH executes a tool carries the result
- * back to the model. The system prompt travels as a blob referenced by
- * `rootPromptMessagesJson`; the server fetches it through the KV handshake.
- */
-export function buildConversationState(options) {
-	const systemParts = [options.system ?? ""];
-	const turns = [];
-	let currentUser = null;
-	let currentSteps = [];
-
-	const flushTurn = () => {
-		if (currentUser !== null) {
-			const userBytes = encodeUserMessage({ text: currentUser, messageId: randomUUID() });
-			const turnBytes = encodeAgentTurn(userBytes, currentSteps);
-			turns.push(encodeTurnStructure(turnBytes));
-		}
-		currentUser = null;
-		currentSteps = [];
-	};
-
-	for (const message of options.messages ?? []) {
-		if (message.role === "system") {
-			systemParts.push(flattenBlocks(message.content));
-			continue;
-		}
-		if (message.role === "user") {
-			flushTurn();
-			currentUser = flattenBlocks(message.content);
-			continue;
-		}
-		if (message.role === "assistant") {
-			const text = renderAssistant(message.content);
-			if (currentUser === null) currentUser = ""; // assistant-first history: implicit turn
-			if (text.length > 0) currentSteps.push(encodeAssistantStep(text));
-		}
-	}
-
-	// The final pending user message is the action, not part of history.
-	let actionText = "";
-	if (currentUser !== null) {
-		actionText = currentUser;
-	}
-	currentUser = null;
-	currentSteps = [];
-
-	const systemText = systemParts.filter((part) => part.length > 0).join("\n").trim();
-	const blobStore = new Map();
-	const rootPromptBlobIds = [];
-	if (systemText.length > 0) {
-		const payload = textEncoder.encode(JSON.stringify({ role: "system", content: systemText }));
-		const id = sha256(payload);
-		blobStore.set(Buffer.from(id).toString("hex"), payload);
-		rootPromptBlobIds.push(id);
-	}
-
-	const conversationState = encodeConversationState({ rootPromptBlobIds, turns, systemText });
-	const userBytes = encodeUserMessage({ text: actionText, messageId: randomUUID() });
-	const action = encodeUserMessageAction(userBytes);
-	return { conversationState, action, blobStore, systemText };
-}
-
-/**
- * Build the full `AgentClientMessage` payload for one run.
- *
- * When a persisted checkpoint for this conversation exists (captured from a
- * previous run's `conversation_checkpoint_update`), it becomes the
- * conversation state and its referenced blobs are served from the persisted
- * blob store. Otherwise the state is built from the DSH message history.
- */
-export function buildRunPayload(options, modelId, persisted, images = []) {
-	let conversationState;
-	let blobStore;
-	let actionText;
-	if (persisted?.checkpoint !== undefined) {
-		// The checkpoint already encodes prior turns. DSH may append both the
-		// human request and a runtime-context user message, so combine the whole
-		// trailing user group rather than selecting only the final message.
-		conversationState = persisted.checkpoint;
-		blobStore = persisted.blobs ?? new Map();
-		actionText = currentActionText(options);
-	} else {
-		// On first use or after a DSH restart there is no server checkpoint. Never
-		// hand-encode field-8 turns: current Cursor servers treat them as blob ids.
-		const built = buildInitialConversationState(options);
-		conversationState = built.conversationState;
-		blobStore = built.blobStore;
-		actionText = coldStartActionText(options);
-	}
-	const userBytes = encodeUserMessage({ text: actionText, messageId: randomUUID() });
-	const action = encodeUserMessageAction(userBytes, images);
-	const modelDetails = encodeModelDetails(modelId);
-	const runRequest = encodeRunRequest({
-		conversationState,
-		action,
-		modelDetails,
-		conversationId: randomUUID(),
-	});
-	return { payload: encodeRunMessage(runRequest), blobStore };
-}
-
-//#endregion
-
-//#region models discovery
-/**
- * Fetch usable models from Cursor's unary `GetUsableModels` endpoint.
- * The request body is the raw (unframed) empty protobuf message.
- */
-export async function fetchUsableModels(accessToken, options = {}) {
-	const fetchImpl = options.fetch ?? nativeFetch;
-	const response = await fetchImpl(`${CURSOR_BASE_URL}${CURSOR_MODELS_PATH}`, {
-		method: "POST",
-		redirect: "error",
-		headers: {
-			"content-type": "application/proto",
-			authorization: `Bearer ${accessToken}`,
-			"x-ghost-mode": "true",
-			"x-cursor-client-version": CURSOR_CLIENT_VERSION,
-			"x-cursor-client-type": "cli",
-			"user-agent": "dsh-cursor-subscription/0.1.0",
-		},
-		body: new Uint8Array(0),
-		signal: options.signal,
-	});
-	if (!response.ok) throw new Error(`Cursor model discovery failed (HTTP ${response.status})`);
-	const bytes = new Uint8Array(await response.arrayBuffer());
-	if (bytes.length === 0) throw new Error("Cursor model discovery returned an empty response");
-	// The body may be raw protobuf or Connect-framed; try both.
-	const models = decodeUsableModels(bytes);
-	if (models.length > 0) return models;
-	const frameReader = new ConnectFrameReader();
-	frameReader.push(bytes);
-	frameReader.finish();
-	const framed = [];
-	for (;;) {
-		const frame = await frameReader.next();
-		if (frame === undefined) break;
-		if ((frame.flags & CONNECT_END_STREAM_FLAG) !== 0) break;
-		framed.push(...decodeUsableModels(frame.payload));
-	}
-	return framed;
-}
-//#endregion
-
-//#region usage
-const record = (value) => value !== null && typeof value === "object" && !Array.isArray(value);
-
-/**
- * Extract the legacy request bucket from the `/api/usage` model map. Cursor's
- * dashboard hardcodes the `"gpt-4"` entry for the legacy quota; when absent we
- * fall back to the quota-bearing bucket with the largest limit.
- */
-export function parseLegacyBucket(json) {
-	if (!record(json)) return undefined;
-	const gpt4 = json["gpt-4"];
-	if (record(gpt4) && typeof gpt4.numRequests === "number") {
-		return { numRequests: gpt4.numRequests, maxRequestUsage: gpt4.maxRequestUsage };
-	}
-	let best;
-	let bestLimit = 0;
-	for (const value of Object.values(json)) {
-		if (record(value) && typeof value.numRequests === "number") {
-			const limit = typeof value.maxRequestUsage === "number" ? value.maxRequestUsage : 0;
-			if (limit > bestLimit) {
-				bestLimit = limit;
-				best = { numRequests: value.numRequests, maxRequestUsage: value.maxRequestUsage };
-			}
-		}
-	}
-	return best;
-}
-
-/** Cursor converts included on-plan spend (cents) to requests at ~4 cents each. */
-export function getRequestCountFromSpendCents(cents) {
-	return typeof cents === "number" && cents > 0 ? Math.ceil(cents / 4) : 0;
-}
-
-/**
- * Compute included-request usage the way Cursor's dashboard does:
- * `used = team ? ceil(spendCents/4) : legacy.numRequests`,
- * `limit = team ? 500 * requestQuotaPerSeat : legacy.maxRequestUsage`.
- */
-export function computeIncludedRequests({ legacy, isTeam, planUsedCents, requestQuotaPerSeat }) {
-	const usedFromSpend =
-		typeof planUsedCents === "number" && planUsedCents > 0 ? getRequestCountFromSpendCents(planUsedCents) : undefined;
-	const used = isTeam ? usedFromSpend ?? legacy?.numRequests : legacy?.numRequests;
-	const limit = isTeam && typeof requestQuotaPerSeat === "number" ? 500 * requestQuotaPerSeat : legacy?.maxRequestUsage;
-	if (typeof used !== "number" || typeof limit !== "number" || limit <= 0) return undefined;
-	return {
-		used,
-		limit,
-		remaining: Math.max(0, limit - used),
-		pct: Math.round((used / limit) * 1000) / 10,
-	};
-}
-
-function centsToDollars(cents) {
-	return typeof cents === "number" && Number.isFinite(cents) ? Math.round((cents / 100) * 100) / 100 : undefined;
-}
-
-function asFiniteNumber(value) {
-	if (typeof value === "number" && Number.isFinite(value)) return value;
-	if (typeof value === "string" && value.trim() !== "") {
-		const n = Number(value);
-		if (Number.isFinite(n)) return n;
-	}
-	return undefined;
-}
-
-function firstFinite(...values) {
-	for (const value of values) {
-		const n = asFiniteNumber(value);
-		if (n !== undefined) return n;
-	}
-	return undefined;
-}
-
-function percentFromUsageMessage(message) {
-	if (typeof message !== "string") return undefined;
-	const match = message.match(/(\d+(?:\.\d+)?)\s*%/);
-	return match ? asFiniteNumber(match[1]) : undefined;
-}
-
-/** Project `/api/usage-summary` into browser-safe usage facts. */
-export function parseUsageSummary(json) {
-	if (!record(json)) return {};
-	const out = {};
-	if (typeof json.membershipType === "string") out.membershipType = json.membershipType;
-	if (typeof json.isUnlimited === "boolean") out.isUnlimited = json.isUnlimited;
-	if (typeof json.billingCycleStart === "string") out.billingCycleStart = json.billingCycleStart;
-	if (typeof json.billingCycleEnd === "string") out.billingCycleEnd = json.billingCycleEnd;
-	if (json.limitType === "team") out.isTeam = true;
-	const individualPlan = record(json.individualUsage?.plan) ? json.individualUsage.plan : undefined;
-	const teamPlan = record(json.teamUsage?.plan) ? json.teamUsage.plan : undefined;
-	if (individualPlan !== undefined || teamPlan !== undefined) {
-		const used = firstFinite(individualPlan?.used, teamPlan?.used);
-		const limit = firstFinite(individualPlan?.limit, teamPlan?.limit);
-		const totalPercentUsed = firstFinite(individualPlan?.totalPercentUsed, teamPlan?.totalPercentUsed);
-		const autoPercentUsed = firstFinite(individualPlan?.autoPercentUsed, teamPlan?.autoPercentUsed);
-		let apiPercentUsed = firstFinite(individualPlan?.apiPercentUsed, teamPlan?.apiPercentUsed);
-		if (apiPercentUsed === undefined) {
-			apiPercentUsed = percentFromUsageMessage(json.namedModelSelectedDisplayMessage);
-		}
-		if (used !== undefined) out.planUsedCents = used;
-		if (limit !== undefined) out.planLimitCents = limit;
-		if (totalPercentUsed !== undefined) out.totalPercentUsed = totalPercentUsed;
-		if (autoPercentUsed !== undefined) out.autoPercentUsed = autoPercentUsed;
-		if (apiPercentUsed !== undefined) out.apiPercentUsed = apiPercentUsed;
-	}
-	const onDemand = json.individualUsage?.onDemand;
-	if (record(onDemand)) {
-		const usedDollars = centsToDollars(onDemand.used);
-		const limitDollars = centsToDollars(onDemand.limit);
-		const remainingDollars = centsToDollars(onDemand.remaining);
-		if (usedDollars !== undefined || limitDollars !== undefined) {
-			out.individualOnDemand = {
-				...usedDollars === undefined ? {} : { usedDollars },
-				...limitDollars === undefined ? {} : { limitDollars },
-				...remainingDollars === undefined ? {} : { remainingDollars },
-			};
-		}
-	}
-	const teamOnDemand = json.teamUsage?.onDemand;
-	if (record(teamOnDemand)) {
-		const usedDollars = centsToDollars(teamOnDemand.used);
-		const limitDollars = centsToDollars(teamOnDemand.limit);
-		const remainingDollars = centsToDollars(teamOnDemand.remaining);
-		if (usedDollars !== undefined || limitDollars !== undefined) {
-			out.teamOnDemand = {
-				...usedDollars === undefined ? {} : { usedDollars },
-				...limitDollars === undefined ? {} : { limitDollars },
-				...remainingDollars === undefined ? {} : { remainingDollars },
-			};
-		}
-	}
-	return out;
-}
-
-/**
- * Classify one aggregation row into the included-usage pool whose percentage
- * it draws on. Cursor's own usage table splits rows with the same `tier`
- * field: tier 2 is the Cursor Models (Auto + Composer) pool, and every other
- * row — named models, plus `default` rows at tier 1 that the dashboard labels
- * "auto (overflow)" — is billed to the Other Models (API) pool.
- */
-export function usageModelPool(entry) {
-	const intent = typeof entry?.modelIntent === "string" ? entry.modelIntent.trim() : "";
-	const tier = asFiniteNumber(entry?.tier);
-	if (tier === 2) return "auto";
-	if (tier === undefined && intent === "default") return "auto";
-	return "api";
-}
-
-/** Token counts one aggregation row reports; a row with none stays unknown. */
-function aggregationTokens(entry) {
-	let total;
-	for (const field of ["inputTokens", "outputTokens", "cacheWriteTokens", "cacheReadTokens"]) {
-		const value = asFiniteNumber(entry?.[field]);
-		if (value === undefined) continue;
-		total = (total ?? 0) + value;
-	}
-	return total;
-}
-
-/**
- * Read the aggregation rows into the pool subtotals both projections need.
- * Rows are classified exactly like the dashboard's usage table, and a
- * `default` row at tier 1 is Auto usage that spilled out of the Cursor Models
- * pool and is billed to Other Models.
- */
-function collectModelUsage(json) {
-	const rows = [];
-	const poolCents = { auto: 0, api: 0 };
-	const poolTokens = { auto: 0, api: 0 };
-	if (!record(json) || !Array.isArray(json.aggregations)) return { rows, poolCents, poolTokens };
-	for (const entry of json.aggregations) {
-		if (!record(entry) || typeof entry.modelIntent !== "string") continue;
-		const id = entry.modelIntent.trim();
-		if (!id) continue;
-		const cents = asFiniteNumber(entry.totalCents);
-		if (cents === undefined || cents <= 0) continue;
-		const pool = usageModelPool(entry);
-		const tokens = aggregationTokens(entry);
-		rows.push({
-			id,
-			cents,
-			pool,
-			overflow: id === "default" && pool === "api" && asFiniteNumber(entry.tier) === 1,
-			...tokens === undefined ? {} : { tokens },
-		});
-		poolCents[pool] += cents;
-		poolTokens[pool] += tokens ?? 0;
-	}
-	rows.sort((a, b) => b.cents - a.cents);
-	return { rows, poolCents, poolTokens };
-}
-
-/**
- * Project one aggregation payload into the per-model rows plus the pool
- * token totals.
- *
- * `pct` is what Cursor's own usage table prints for a model row: the model's
- * share of its pool's spend times that pool's reported percentage, so one
- * Other Models model reads the same percentage as the pool row above it. A
- * model whose pool percentage was not reported carries no `pct` at all.
- */
-function projectModelUsage(json, options = {}) {
-	const limit = options.limit ?? MAX_USAGE_MODELS;
-	const { rows, poolCents, poolTokens } = collectModelUsage(json);
-	const poolPercent = {
-		auto: asFiniteNumber(options.autoPercentUsed),
-		api: asFiniteNumber(options.apiPercentUsed),
-	};
-	const cap = Number.isFinite(limit) && limit > 0 ? Math.min(Math.floor(limit), rows.length) : rows.length;
-	const models = [];
-	for (let index = 0; index < cap; index++) {
-		const row = rows[index];
-		const model = {
-			id: row.id,
-			pool: row.pool,
-			spentDollars: Math.round((row.cents / 100) * 100) / 100,
-		};
-		if (row.tokens !== undefined) model.tokens = row.tokens;
-		if (row.overflow === true) model.overflow = true;
-		const percent = poolPercent[row.pool];
-		if (percent !== undefined && poolCents[row.pool] > 0) {
-			model.pct = Math.round((row.cents / poolCents[row.pool]) * percent * 10) / 10;
-		}
-		models.push(model);
-	}
-	return { models, pools: { auto: { tokens: poolTokens.auto }, api: { tokens: poolTokens.api } } };
-}
-
-/**
- * Project `/api/dashboard/get-aggregated-usage-events` into a browser-safe
- * per-model list. Cursor's named/API models (Claude, Grok, GPT, …) live here;
- * `/api/usage` only exposes the legacy `gpt-4` quota bucket.
- */
-export function parseModelAggregations(json, options = {}) {
-	return projectModelUsage(json, options).models;
-}
-
-/** The full projection the settings panel's included-usage table renders. */
-export function parseModelUsage(json, options = {}) {
-	return projectModelUsage(json, options);
-}
-
-/**
- * Pool facts for the panel's included-usage table: Cursor's reported pool
- * percentage, plus the pool's token total when the aggregated read succeeded.
- * A pool with neither is left out, so the panel never prints an empty row.
- */
-export function usagePools(parsed, modelUsage) {
-	const pools = {};
-	for (const [pool, percent] of [["auto", parsed?.autoPercentUsed], ["api", parsed?.apiPercentUsed]]) {
-		const facts = {};
-		const pct = asFiniteNumber(percent);
-		if (pct !== undefined) facts.pct = pct;
-		if (modelUsage !== undefined) facts.tokens = modelUsage.pools[pool].tokens;
-		if (Object.keys(facts).length > 0) pools[pool] = facts;
-	}
-	return pools;
-}
-
-/** Pull `requestQuotaPerSeat` for the active team from `/api/dashboard/teams`. */
-export function parseRequestQuotaPerSeat(json, teamId) {
-	if (!record(json) || !Array.isArray(json.teams)) return undefined;
-	const match = (teamId ? json.teams.find((t) => String(t?.id) === String(teamId)) : undefined) ?? json.teams[0];
-	const quota = match?.requestQuotaPerSeat ?? match?.request_quota_per_seat;
-	return typeof quota === "number" && Number.isFinite(quota) ? quota : undefined;
-}
-
-function usageDaysLeft(billingCycleEnd, now) {
-	const end = new Date(billingCycleEnd).getTime();
-	if (!Number.isFinite(end)) return undefined;
-	return Math.max(0, Math.ceil((end - now) / (24 * 60 * 60 * 1000)));
-}
-
-/**
- * Read usage from Cursor's dashboard endpoints using the same session cookie
- * the web dashboard sends. The browser receives only the parsed projection;
- * the access token and cookie are request-local host values. Concurrent polls
- * share one in-flight request; successful reads are cached for a short TTL.
- */
-export class CursorUsageReader {
-	constructor(auth, options = {}) {
-		this.auth = auth;
-		this.fetch = options.fetch ?? nativeFetch;
-		this.now = options.now ?? Date.now;
-		this.ttlMs = options.ttlMs ?? USAGE_TTL_MS;
-		this.logger = options.logger;
-		this.#cache = { at: 0, value: undefined };
-	}
-
-	#cache;
-
-	#log(level, message, ...args) {
-		try {
-			this.logger?.[level]?.(`cursor-subscription: ${message}`, ...args);
-		} catch {}
-	}
-
-	async #request(url, cookie, options = {}) {
-		const response = await this.fetch(url, {
-			method: options.method ?? "GET",
-			redirect: "follow",
-			headers: {
-				accept: "application/json",
-				origin: USAGE_API_ORIGIN,
-				referer: "https://cursor.com/dashboard",
-				cookie,
-				"user-agent": "dsh-cursor-subscription/0.2.0",
-				...(options.body === undefined ? {} : { "content-type": "application/json" }),
-			},
-			...options.body === undefined ? {} : { body: JSON.stringify(options.body) },
-			signal: options.signal,
-		});
-		if (!response.ok) throw new Error(`HTTP ${response.status} from ${url}`);
-		const text = await response.text();
-		try {
-			return JSON.parse(text);
-		} catch (error) {
-			throw new Error(`unreadable JSON from ${url}`, { cause: error });
-		}
-	}
-
-	async #tryRequest(url, cookie, options = {}) {
-		try {
-			return { ok: true, value: await this.#request(url, cookie, options) };
-		} catch (error) {
-			this.#log("warn", "usage request failed: %s", error instanceof Error ? error.message : error);
-			return { ok: false, error };
-		}
-	}
-
-	#mergeSignals(signal, timeoutMs) {
-		const timeout = Number.isFinite(timeoutMs) && timeoutMs > 0 && typeof AbortSignal.timeout === "function"
-			? AbortSignal.timeout(timeoutMs)
-			: undefined;
-		if (signal === undefined) return timeout;
-		if (timeout === undefined) return signal;
-		return typeof AbortSignal.any === "function" ? AbortSignal.any([signal, timeout]) : signal;
-	}
-
-	async read({ force = false, signal } = {}) {
-		const now = this.now();
-		if (!force && this.#cache.value !== undefined && now - this.#cache.at < this.ttlMs) {
-			return structuredClone(this.#cache.value);
-		}
-		const credential = await this.auth.credential({ signal });
-		const userSub = getTokenSub(credential.access);
-		if (userSub === undefined) throw new Error("Cursor session token has no usable user identity");
-		const cookie = `WorkosCursorSessionToken=${userSub}::${credential.access}`;
-
-		const [usageResult, summaryResult, teamsResult] = await Promise.all([
-			this.#tryRequest(`${USAGE_URL}?user=${encodeURIComponent(userSub)}`, cookie, { signal }),
-			this.#tryRequest(USAGE_SUMMARY_URL, cookie, { signal }),
-			this.#tryRequest(USAGE_TEAMS_URL, cookie, { method: "POST", body: { activeOnly: false }, signal }),
-		]);
-		if (!usageResult.ok && !summaryResult.ok) {
-			const failed = summaryResult.error ?? usageResult.error;
-			throw failed instanceof Error ? failed : new Error("Could not read Cursor usage");
-		}
-		const usage = usageResult.ok ? usageResult.value : undefined;
-		const summary = summaryResult.ok ? summaryResult.value : undefined;
-		const teams = teamsResult.ok ? teamsResult.value : undefined;
-
-		const legacy = parseLegacyBucket(usage);
-		const parsed = parseUsageSummary(summary);
-		const requestQuotaPerSeat = parseRequestQuotaPerSeat(teams);
-		const includedRequests =
-			computeIncludedRequests({
-				legacy,
-				isTeam: parsed.isTeam === true,
-				planUsedCents: parsed.planUsedCents,
-				requestQuotaPerSeat,
-			}) ?? (legacy !== undefined ? computeIncludedRequests({ legacy, isTeam: false }) : undefined);
-
-		const modelUsage = await this.#readModelUsage(cookie, parsed, now, signal);
-		const hasPlan =
-			parsed.totalPercentUsed !== undefined || parsed.autoPercentUsed !== undefined || parsed.apiPercentUsed !== undefined;
-
-		const value = {
-			fetchedAt: now,
-			membershipType: parsed.membershipType,
-			isUnlimited: parsed.isUnlimited,
-			isTeam: parsed.isTeam,
-			billingCycle: parsed.billingCycleEnd === undefined ? undefined : {
-				start: parsed.billingCycleStart,
-				end: parsed.billingCycleEnd,
-				daysLeft: usageDaysLeft(parsed.billingCycleEnd, now),
-			},
-			plan: hasPlan ? {
-				...parsed.totalPercentUsed === undefined ? {} : { totalPercentUsed: parsed.totalPercentUsed },
-				...parsed.autoPercentUsed === undefined ? {} : { autoPercentUsed: parsed.autoPercentUsed },
-				...parsed.apiPercentUsed === undefined ? {} : { apiPercentUsed: parsed.apiPercentUsed },
-			} : undefined,
-			includedRequests,
-			individualOnDemand: parsed.individualOnDemand,
-			teamOnDemand: parsed.teamOnDemand,
-			pools: usagePools(parsed, modelUsage),
-			models: modelUsage?.models ?? [],
-		};
-		this.#cache = { at: now, value };
-		this.#log("info", "usage read: %s", JSON.stringify({ membershipType: value.membershipType, includedRequests: value.includedRequests, models: value.models.length, apiPercentUsed: value.plan?.apiPercentUsed }));
-		return structuredClone(value);
-	}
-
-	/**
-	 * Read the per-model aggregation for the billing window. Returns the model
-	 * rows plus the pool token totals, or undefined when the read failed — the
-	 * panel then keeps the pool percentages from `/api/usage-summary` and shows
-	 * the token cells as unavailable.
-	 */
-	async #readModelUsage(cookie, parsed, now, signal) {
-		const start = parsed.billingCycleStart !== undefined ? new Date(parsed.billingCycleStart).getTime() : now - 32 * 24 * 60 * 60 * 1000;
-		const end = parsed.billingCycleEnd !== undefined ? new Date(parsed.billingCycleEnd).getTime() : now;
-		if (!Number.isFinite(start) || !Number.isFinite(end)) return undefined;
-		try {
-			const aggregated = await this.#request(USAGE_AGGREGATED_URL, cookie, {
-				method: "POST",
-				body: { teamId: 0, startDate: start, endDate: end },
-				signal: this.#mergeSignals(signal, 8_000),
-			});
-			return parseModelUsage(aggregated, {
-				autoPercentUsed: parsed.autoPercentUsed,
-				apiPercentUsed: parsed.apiPercentUsed,
-			});
-		} catch (error) {
-			this.#log("warn", "aggregated model usage unavailable: %s", error instanceof Error ? error.message : error);
-			return undefined;
-		}
-	}
-
-	clear() {
-		this.#cache = { at: 0, value: undefined };
-	}
-}
-//#endregion
-
-//#region adapter
-export const TOOL_REJECT_REASON =
-	"Tool not available in this environment. Use the MCP tools provided instead.";
-
-/** Block indices are fixed per type so the assembler keeps blocks distinct. */
-const TEXT_BLOCK_INDEX = 0;
-const REASONING_BLOCK_INDEX = 1;
-const TOOL_BLOCK_INDEX = 2;
-/**
- * Progress notes ride their own reasoning block, so the model's own thinking
- * stays exactly what the model produced. Reasoning is the only channel an
- * adapter has for showing work that is not answer text, and a Cursor step
- * otherwise reduces to one sentence plus a tool card: the protocol round trips
- * that cost the turn most of its wall clock — a declined native tool call, a
- * tool-schema fetch — are invisible, and a turn of them reads as the same
- * request being submitted again and again.
- */
-export const PROGRESS_BLOCK_INDEX = 3;
-
-/**
- * One reader-visible progress line.
- * @param text What just happened, in one clause.
- * @returns A reasoning delta the conversation view renders as process.
- */
-function cursorProgress(text) {
-	return { type: "reasoning-delta", index: PROGRESS_BLOCK_INDEX, text: `[cursor] ${text}\n` };
-}
-
-/** Reader-facing name of one decoded exec. */
-function nativeToolName(exec) {
-	if (exec.case === "unknown") return `tool (field ${exec.field ?? "?"})`;
-	return exec.case.replace(/ExecArgs$|Args$/, "");
-}
-
-/**
- * DSH `LlmAdapter` for the Cursor subscription. Normal turns use Cursor
- * checkpoints; MCP tool calls keep their HTTP/2 Run alive across DSH steps so
- * the following tool-result can resume the exact pending exec with mcpResult.
- */
-export class CursorAdapter extends LlmAdapter {
-	constructor(options) {
-		super();
-		this.auth = options.auth;
-		this.fetchModels = options.fetchModels ?? fetchUsableModels;
-		this.idleTimeoutMs = options.idleTimeoutMs ?? STREAM_IDLE_TIMEOUT_MS;
-		this.progressTimeoutMs = options.progressTimeoutMs ?? STREAM_PROGRESS_TIMEOUT_MS;
-		this.idleCheckIntervalMs = options.idleCheckIntervalMs ?? 15000;
-		this.modelsCacheTtlMs = options.modelsCacheTtlMs ?? 5 * 60 * 1000;
-		this.sessionStateTtlMs = options.sessionStateTtlMs ?? SESSION_STATE_TTL_MS;
-		const fallbackSettings = resolveCursorSettings({
-			maxToolRounds: options.maxToolRounds,
-			retryCount: options.retryCount,
-			retryIntervalMs: options.retryIntervalMs,
-			retryHttpStatusCodes: options.retryHttpStatusCodes,
-		});
-		this.settings = options.settings ?? (() => fallbackSettings);
-		this.createAgentRun = options.createAgentRun ?? ((access) => new AgentRun(access));
-		this.resolveAttachments = options.resolveAttachments;
-		this.sleep = options.sleep ?? abortableDelay;
-		this.now = options.now ?? Date.now;
-		this.logger = options.logger;
-		this.hangTracePath = options.hangTracePath ?? join(homedir(), ".dsh", "cursor-hang-trace.log");
-		this.toolCallSettleMs = options.toolCallSettleMs ?? TOOL_CALL_SETTLE_MS;
-		this.#modelsCache = { at: 0, models: undefined };
-	}
-
-	#modelsCache;
-	/** Persisted conversation checkpoints + referenced blobs per DSH session id. */
-	#sessions = new Map();
-
-	#evictStaleSessions() {
-		const now = this.now();
-		for (const [key, state] of this.#sessions) {
-			if (now - (state.lastAccessMs ?? now) <= this.sessionStateTtlMs) continue;
-			state.bridge?.run.close();
-			this.#sessions.delete(key);
-		}
-	}
-
-	async #discoverModels(force = false) {
-		const now = this.now();
-		if (!force && this.#modelsCache.models !== undefined && now - this.#modelsCache.at < this.modelsCacheTtlMs) {
-			return this.#modelsCache.models;
-		}
-		const access = await this.auth.accessToken();
-		const source = await this.fetchModels(access);
-		if (source.length === 0) throw new Error("Cursor returned no usable models");
-		// Sorted here so the picker and the settings list agree, and so a change
-		// in Cursor's response order cannot reshuffle either surface.
-		const result = sortModelsByName(
-			source.map((model) => ({
-				id: model.id,
-				name: model.name || model.id,
-				inputModalities: ["text", "image"],
-			})),
-		);
-		this.#modelsCache = { at: now, models: result };
-		return result;
-	}
-
-	invalidateModels() {
-		this.#modelsCache = { at: 0, models: undefined };
-	}
-
-	providerInfo(provider) {
-		return { id: provider, name: "Cursor subscription" };
-	}
-
-	async listModels(provider) {
-		const models = await this.#discoverModels();
-		return models.map((model) => ({ provider, ...model }));
-	}
-
-	/** RPC-facing model listing; `force` bypasses the discovery cache. */
-	async listModelsForRpc({ force = false, signal } = {}) {
-		const models = await this.#discoverModels(force);
-		if (signal?.aborted) throw new LlmError("Cursor model listing aborted", "ABORTED");
-		return models;
-	}
-
-	async resolveModel(provider, model, _signal) {
-		let name = model;
-		try {
-			const models = await this.#discoverModels();
-			const found = models.find((entry) => entry.id === model);
-			if (found !== undefined) name = found.name;
-		} catch {}
-		return {
-			provider,
-			id: model,
-			name,
-			inputModalities: ["text", "image"],
-			context: { contextWindow: DEFAULT_CONTEXT_WINDOW },
-			defaultMaxTokens: DEFAULT_MAX_TOKENS,
-		};
-	}
-
-	async *stream(options) {
-		const consumer = new AbortController();
-		const upstream = options.signal === undefined ? consumer.signal : AbortSignal.any([options.signal, consumer.signal]);
-		let run;
-		let abortRun;
-		let bridgeStored = false;
-		// Persisted conversation state per session: the server's checkpoint,
-		// referenced blobs, and (while an MCP tool runs) the live HTTP/2 bridge.
-		const sessionKey = options.sessionId === undefined ? undefined : String(options.sessionId);
-		this.#evictStaleSessions();
-		const prior = sessionKey === undefined ? undefined : this.#sessions.get(sessionKey);
-		const persisted = {
-			checkpoint: prior?.checkpoint === undefined ? undefined : Uint8Array.from(prior.checkpoint),
-			blobs: new Map(prior?.blobs ?? []),
-		};
-		// DSH condenses the history mid-session. A Cursor conversation checkpoint
-		// captured before that condensing still carries the pre-compaction
-		// transcript, so reusing it hands the model exactly the context DSH just
-		// freed: the turn then runs on a stale, oversized conversation and
-		// re-plans the same work step after step. Drop the checkpoint and its
-		// blobs once per new compaction, and never resume a live bridge across
-		// one, so the next run is rebuilt from what DSH kept.
-		const compactionId = compactionIdOf(options);
-		const compacted = compactionId !== undefined && compactionId !== prior?.compactionId;
-		if (compacted) {
-			persisted.checkpoint = undefined;
-			persisted.blobs.clear();
-		}
-		// The opt-in rebuild: drop the server-side conversation as well, so the run
-		// below is assembled from the DSH history alone.
-		const replayHistoryEachStep = readConfigField(this.settings().replayHistoryEachStep) === true;
-		if (replayHistoryEachStep) {
-			persisted.checkpoint = undefined;
-			persisted.blobs.clear();
-		}
-		persisted.compactionId = compactionId ?? prior?.compactionId;
-		try {
-			if (options.stop !== undefined) throw new LlmError("cursor-subscription does not support GenerateOptions.stop", "UNSUPPORTED_OPTION");
-			const requestSettings = resolveCursorSettings(this.settings());
-			const mcpTools = (options.tools ?? []).map((tool) =>
-				encodeMcpToolDefinition({
-					name: tool.name,
-					description: tool.description,
-					inputSchema: encodeValue(tool.parameters ?? {}),
-				}),
-			);
-			const toolResults = collectToolResults(options);
-			const liveBridge = prior?.bridge;
-			const canResume = !replayHistoryEachStep && !compacted && liveBridge !== undefined && !liveBridge.run.finished && !liveBridge.run.stream?.destroyed && toolResults.length > 0;
-			let toolRoundCount = canResume ? (liveBridge.toolRoundCount ?? 0) : 0;
-			let blobStore;
-			if (canResume) {
-				run = liveBridge.run;
-				// A previous step paused this reader to stop waiting for frames it
-				// no longer needed; continue the same Run and read again.
-				run.frames?.resume?.();
-				blobStore = persisted.blobs;
-				abortRun = () => run.abort(new Error("Cursor request aborted by caller"));
-				upstream.addEventListener("abort", abortRun, { once: true });
-				if (upstream.aborted) abortRun();
-				for (const pending of liveBridge.pendingExecs) {
-					const result = toolResults.find((entry) => entry.toolCallId === pending.toolCallId);
-					if (result?.sandboxDenied === true) {
-						yield cursorProgress("the sandbox refused a tool call; told the model how to ask for approval");
-					}
-					const payload = result === undefined
-						? encodeMcpError("Tool result not provided")
-						: encodeMcpResult(result);
-					if (!run.writeMessage(encodeExecClientMessageEnvelope(encodeExecClientMessage(pending.id, pending.execId, 11, payload)))) {
-						throw new Error("Cursor tool continuation bridge closed before accepting the result");
-					}
-				}
-			} else {
-				if (liveBridge !== undefined) {
-					liveBridge.run.close();
-					// A checkpoint paused on an MCP call cannot safely accept a normal
-					// user action. Fall back to textual cold-start reconstruction.
-					persisted.checkpoint = undefined;
-					persisted.blobs.clear();
-				}
-				const access = await this.auth.accessToken({ signal: upstream });
-				const images = await prepareCursorImages(options, this.resolveAttachments?.());
-				const built = buildRunPayload(options, options.model, persisted, images);
-				blobStore = built.blobStore;
-				let retriesUsed = 0;
-				for (;;) {
-					run = this.createAgentRun(access);
-					await run.start();
-					abortRun = () => run.abort(new Error("Cursor request aborted by caller"));
-					upstream.addEventListener("abort", abortRun, { once: true });
-					if (upstream.aborted) abortRun();
-					if (!run.writeMessage(built.payload)) throw new Error("Cursor agent bridge closed before accepting the request");
-					const status = await run.waitForResponse(this.idleTimeoutMs);
-					if (isSuccessfulAgentResponse(status, run.responseContentType)) {
-						run.startHeartbeat();
-						break;
-					}
-					upstream.removeEventListener("abort", abortRun);
-					abortRun = undefined;
-					run.close();
-					if (status === 200) {
-						throw new LlmError(
-							`Cursor agent returned an unexpected content type: ${run.responseContentType ?? "missing"}`,
-							"CURSOR_PROTOCOL",
-						);
-					}
-					if (!shouldRetryHttpStatus(status, retriesUsed, requestSettings)) {
-						const message = `Cursor agent returned HTTP ${status}`;
-						throw new LlmError(message, classifyCursorError(message));
-					}
-					retriesUsed++;
-					await this.sleep(requestSettings.retryIntervalMs, upstream);
-				}
-			}
-
-			const started = this.now();
-			let lastActivity = started;
-			let lastProgress = started;
-			let emittedToolCall = false;
-			let toolCallBuffer = undefined;
-			let textOutput = "";
-			let outputTokens = 0;
-			let streamClosed = false;
-			let terminalErrorEmitted = false;
-			let toolCallPending = false;
-			let toolRoundCounted = false;
-			let unknownExecWarned = false;
-			const pendingExecs = [];
-			let toolCallSettle;
-			/**
-			 * Stop blocking for more frames without ending the reader, so the same
-			 * Run can be continued by the next DSH step.
-			 */
-			const stopReading = () => {
-				if (typeof run.frames?.pause === "function") run.frames.pause();
-				else if (typeof run.frames?.finish === "function") run.frames.finish();
-			};
-			/**
-			 * Cap how long the loop waits for the frame that closes a tool-call
-			 * burst. Cursor usually follows mcpArgs with blob writes and a
-			 * checkpoint, but it may instead send the checkpoint *before* the burst
-			 * and then nothing at all; waiting for a second checkpoint in that case
-			 * hung the step until the progress watchdog aborted it.
-			 */
-			const armToolCallSettle = () => {
-				toolCallSettle ??= setTimeout(stopReading, this.toolCallSettleMs);
-				toolCallSettle.unref?.();
-			};
-			// Compact trace of the frames this run received, kept for the stall
-			// report so a silent server can be diagnosed from the log alone.
-			const frameTrace = [];
-			const traceFrame = (label) => {
-				const last = frameTrace.at(-1);
-				if (label === "heartbeat" && last?.label === "heartbeat") {
-					last.count++;
-					last.at = this.now();
-					return;
-				}
-				frameTrace.push({ label, at: this.now(), count: 1 });
-				if (frameTrace.length > 120) frameTrace.shift();
-			};
-			const stallReport = (message) => {
-				const lines = [
-					`=== ${new Date().toISOString()} ${message} ===`,
-					`session=${sessionKey ?? "-"} model=${options.model ?? "-"} textChars=${textOutput.length}`
-						+ ` toolCallPending=${toolCallPending} toolRounds=${toolRoundCount}`
-						+ ` checkpoint=${persisted.checkpoint?.length ?? 0}B blobs=${blobStore?.size ?? 0}`,
-					`lastFrame=${Math.round((this.now() - lastActivity) / 1000)}s ago`
-						+ ` lastContent=${Math.round((this.now() - lastProgress) / 1000)}s ago`,
-					...frameTrace.map((entry) => `  ${new Date(entry.at).toISOString()} ${entry.label}${entry.count > 1 ? ` x${entry.count}` : ""}`),
-					"",
-				];
-				try {
-					appendFileSync(this.hangTracePath, lines.join("\n"));
-				} catch {}
-				this.logger?.warn?.(`cursor-subscription: ${message}; trace in ${this.hangTracePath}`);
-			};
-			const handleStall = (message) => {
-				stallReport(message);
-				// The step already delivered something the user can see — answer
-				// text or a tool call — and Cursor stopped sending content without
-				// ending the turn. Stop reading instead of raising a retryable
-				// error: DSH would discard the delivered answer or tool call and
-				// re-run the whole step.
-				if ((textOutput.length > 0 || toolCallPending) && typeof run.frames?.pause === "function") {
-					stopReading();
-					return;
-				}
-				run.abort(new Error(message));
-			};
-			const idleCheck = setInterval(() => {
-				const now = this.now();
-				if (now - lastActivity > this.idleTimeoutMs) {
-					handleStall("Cursor stream idle timeout");
-				} else if (now - lastProgress > this.progressTimeoutMs) {
-					handleStall(`Cursor stream progress timeout: no content for ${this.progressTimeoutMs}ms`);
-				}
-			}, this.idleCheckIntervalMs);
-			idleCheck.unref?.();
-
-			try {
-				for (;;) {
-					const frame = await run.frames.next();
-					if (frame === undefined) break;
-					lastActivity = this.now();
-					traceFrame(describeServerFrame(frame));
-					if ((frame.flags & CONNECT_END_STREAM_FLAG) !== 0) {
-						const end = parseEndStream(frame.payload);
-						if (end !== undefined) {
-							terminalErrorEmitted = true;
-							yield {
-								type: "finish",
-								reason: {
-									kind: "error",
-									failure: {
-										message: end.message,
-										code: classifyCursorError(`${end.code} ${end.debugCode ?? ""} ${end.message}`),
-									},
-								},
-							};
-							streamClosed = true;
-							break;
-						}
-						break;
-					}
-					const message = decodeAgentServerMessage(frame.payload);
-					// Heartbeats keep the connection alive but are not content
-					// progress; a server that only pings us (e.g. after an empty
-					// partialToolCall placeholder) must still hit the stall timeout
-					// instead of leaving the UI spinning forever.
-					if (!(message.case === "interactionUpdate" && message.value?.type === "heartbeat")) {
-						lastProgress = this.now();
-					}
-					if (message.case === "conversationCheckpointUpdate") {
-						// Reader.bytes() returns a view into the frame buffer; copy it before
-						// retaining it for a later request.
-						persisted.checkpoint = Uint8Array.from(message.value);
-						if (toolCallPending) {
-							// A checkpoint after mcpArgs closes the tool-call step. Keep the
-							// same Run alive; the next DSH step resumes it with McpResult
-							// field 11. Cursor does not always send this checkpoint after
-							// the burst — the traced hangs have it *before* mcpArgs — so the
-							// wait for it is capped by the settle timer armed below.
-							streamClosed = true;
-							break;
-						}
-					} else if (message.case === "interactionUpdate") {
-						const update = message.value;
-						if (update.type === "textDelta") {
-							if (update.text.length > 0) {
-								textOutput += update.text;
-								yield { type: "text-delta", index: TEXT_BLOCK_INDEX, text: update.text };
-							}
-						} else if (update.type === "thinkingDelta") {
-							if (update.text.length > 0) {
-								yield { type: "reasoning-delta", index: REASONING_BLOCK_INDEX, text: update.text };
-							}
-						} else if (update.type === "thinkingCompleted") {
-							if (update.durationMs > 0) {
-								yield cursorProgress(`thinking finished after ${(update.durationMs / 1000).toFixed(1)}s`);
-							}
-						} else if (update.type === "tokenDelta") {
-							// Each frame carries the tokens that delta produced, not a
-							// running total, so the step's output count is their sum;
-							// keeping the last frame reported "1 token" for a paragraph.
-							if (Number.isFinite(update.tokens) && update.tokens > 0) outputTokens += update.tokens;
-						} else if (update.type === "partialToolCall") {
-							// Buffer only. Emitting incomplete tool-call chunks here created
-							// nameless DSH tool blocks that competed with real mcpArgs frames.
-							if (toolCallBuffer === undefined || toolCallBuffer.id !== update.callId) {
-								toolCallBuffer = { id: update.callId, name: "", arguments: "" };
-							}
-							if (update.argsTextDelta.length > 0) {
-								toolCallBuffer.arguments += update.argsTextDelta;
-							}
-						} else if (update.type === "toolCallStarted" || update.type === "toolCallCompleted" || update.type === "toolCallDelta") {
-							emittedToolCall = true;
-						} else if (update.type === "turnEnded") {
-							// Cursor ends every assistant turn with this update, but it does
-							// not reliably close the HTTP/2 stream afterwards: the server can
-							// keep the run alive with heartbeats only. Ignoring the update
-							// therefore left a finished answer waiting until the progress
-							// watchdog aborted the run, and the resulting retryable TIMEOUT
-							// made DSH discard the answer the user had already read and
-							// re-run the whole step. The trailing conversation checkpoint
-							// always arrives before turnEnded, so stopping here keeps
-							// everything a later run needs. A pending MCP tool call still
-							// waits for its checkpoint so the live bridge stays resumable.
-							if (!toolCallPending) break;
-						}
-					} else if (message.case === "kvServerMessage") {
-						const kv = message.value;
-						if (kv.case === "getBlobArgs") {
-							const key = Buffer.from(kv.blobId).toString("hex");
-							const blob = blobStore.get(key);
-							run.writeMessage(encodeKvClientMessage(encodeGetBlobResult(kv.id, blob)));
-						} else if (kv.case === "setBlobArgs") {
-							if (kv.blobId !== undefined && kv.blobData !== undefined) {
-								const key = Buffer.from(kv.blobId).toString("hex");
-								// Reader.bytes() returns views; retain independent copies for both
-								// this run's GET handshake and the next run's checkpoint.
-								const data = Uint8Array.from(kv.blobData);
-								blobStore.set(key, data);
-								persisted.blobs.set(key, data);
-							}
-							// Acknowledge server blob writes; without the reply the
-							// run can stall waiting for the handshake to finish.
-							run.writeMessage(encodeKvClientMessage(encodeSetBlobResult(kv.id)));
-						}
-					} else if (message.case === "execServerMessage") {
-						const exec = message.value;
-						if (exec.case === "requestContextArgs") {
-							run.writeMessage(encodeExecClientMessageEnvelope(encodeExecClientMessage(exec.id, exec.execId, 10, encodeRequestContextResult(mcpTools))));
-							yield cursorProgress(`loaded the DSH tool schemas (${mcpTools.length} tools)`);
-						} else if (exec.case === "mcpArgs") {
-							if (!toolRoundCounted) {
-								toolRoundCount++;
-								toolRoundCounted = true;
-							}
-							if (toolRoundCount > requestSettings.maxToolRounds) {
-								throw new LlmError(
-									`Cursor tool-call safety limit reached after ${requestSettings.maxToolRounds} rounds. The run was stopped to prevent an infinite loop; send a new message to continue if needed.`,
-									"TOOL_LIMIT",
-								);
-							}
-							// decodeExecServerMessage nests the parsed McpArgs under
-							// `exec.args`: { name, toolCallId, providerIdentifier, toolName, args }.
-							const mcp = exec.args ?? {};
-							const args = {};
-							for (const [key, value] of Object.entries(mcp.args ?? {})) {
-								try {
-									args[key] = decodeValue(value);
-								} catch {
-									args[key] = null;
-								}
-							}
-							const candidateName = mcp.toolName || mcp.name;
-							const tool = resolveTextTool(candidateName, options.tools ?? []);
-							if (tool === undefined) {
-								// Cursor occasionally emits empty/placeholder MCP execs around
-								// real calls. Reject them immediately so they never become
-								// pending bridge entries that later poison the resume path with
-								// "Tool result not provided".
-								run.writeMessage(encodeExecClientMessageEnvelope(encodeExecClientMessage(
-									exec.id,
-									exec.execId,
-									11,
-									encodeMcpError(`Unknown or unsupported MCP tool: ${candidateName || "(empty)"}`),
-								)));
-								continue;
-							}
-							emittedToolCall = true;
-							const id = mcp.toolCallId || exec.execId || crypto.randomUUID();
-							const name = tool.name;
-							const normalizedArgs = normalizeTextToolArguments(tool, args);
-							const argumentsJson = JSON.stringify(normalizedArgs);
-							// Parallel MCP calls must use distinct DSH block indexes. Reusing
-							// TOOL_BLOCK_INDEX causes the agent loop to keep only one call while
-							// pendingExecs still waits for every sibling.
-							const blockIndex = TOOL_BLOCK_INDEX + pendingExecs.length;
-							yield { type: "block-start", index: blockIndex, blockType: "tool-call" };
-							yield {
-								type: "tool-call-delta",
-								index: blockIndex,
-								id: ToolCallId(id),
-								name,
-								argumentsDelta: argumentsJson,
-							};
-							yield {
-								type: "block-end",
-								index: blockIndex,
-								block: {
-									type: "tool-call",
-									id: ToolCallId(id),
-									name,
-									arguments: argumentsJson,
-								},
-							};
-							pendingExecs.push({ id: exec.id, execId: exec.execId, toolCallId: id });
-							toolCallPending = true;
-							yield cursorProgress(`calling the DSH tool ${name}`);
-							// Do not close the Run. Cursor usually sends blob writes + a
-							// checkpoint right after mcpArgs, which this loop still wants to
-							// capture for the persisted session, but the ordering is not
-							// guaranteed: traced hangs end with `checkpoint | mcpArgs` and no
-							// further frame, which left the step waiting until the progress
-							// watchdog aborted it and DSH re-ran the whole step instead of
-							// running the tool. Cap the wait; sibling calls of one burst
-							// arrive within ~100 ms of each other.
-							streamClosed = true;
-							armToolCallSettle();
-						} else {
-							const reply = rejectionFor(exec);
-							if (reply !== undefined) {
-								if (exec.case === "unknown" && !unknownExecWarned) {
-									unknownExecWarned = true;
-									this.logger?.warn?.(
-										`cursor-subscription: Cursor requested an exec this build does not know (field ${exec.field}); rejected it so the run can continue with MCP tools`,
-									);
-								}
-								run.writeMessage(encodeExecClientMessageEnvelope(encodeExecClientMessage(exec.id, exec.execId, reply.field, reply.payload)));
-								yield cursorProgress(`native ${nativeToolName(exec)} call declined; using the DSH tools instead`);
-							} else if (exec.case === "unknown") {
-								this.logger?.warn?.("cursor-subscription: Cursor requested an exec with no addressable field; the run may stall");
-							}
-						}
-					}
-				}
-			} finally {
-				clearInterval(idleCheck);
-				clearTimeout(toolCallSettle);
-			}
-
-			if (!terminalErrorEmitted && !toolCallPending) {
-				const textToolCalls = parseTextToolCalls(textOutput, options.tools ?? []);
-				for (let index = 0; index < textToolCalls.length; index++) {
-					const call = textToolCalls[index];
-					const blockIndex = TOOL_BLOCK_INDEX + index;
-					yield { type: "block-start", index: blockIndex, blockType: "tool-call" };
-					yield {
-						type: "tool-call-delta",
-						index: blockIndex,
-						id: ToolCallId(call.id),
-						name: call.name,
-						argumentsDelta: call.arguments,
-					};
-					yield {
-						type: "block-end",
-						index: blockIndex,
-						block: { type: "tool-call", id: ToolCallId(call.id), name: call.name, arguments: call.arguments },
-					};
-				}
-				if (textToolCalls.length > 0) {
-					emittedToolCall = true;
-					streamClosed = true;
-				}
-			}
-
-			const liveToolBridge = toolCallPending && pendingExecs.length > 0 && !run.finished;
-			if (sessionKey !== undefined && (persisted.checkpoint !== undefined || liveToolBridge)) {
-				persisted.lastAccessMs = this.now();
-				if (liveToolBridge) {
-					persisted.bridge = { run, pendingExecs, toolRoundCount };
-					bridgeStored = true;
-				}
-				// Refresh insertion order for simple LRU-style eviction. A checkpoint
-				// may reference every retained blob, so prune whole sessions rather than
-				// individual blobs. Close an evicted live bridge to avoid socket leaks.
-				this.#sessions.delete(sessionKey);
-				this.#sessions.set(sessionKey, persisted);
-				while (this.#sessions.size > 100) {
-					const oldest = this.#sessions.keys().next().value;
-					if (oldest === undefined) break;
-					this.#sessions.get(oldest)?.bridge?.run.close();
-					this.#sessions.delete(oldest);
-				}
-			}
-
-			if (!terminalErrorEmitted) {
-				yield {
-					type: "usage",
-					usage: {
-						// Cursor reports the conversation's prompt size in the
-						// checkpoint it last sent; zero only while none has arrived.
-						inputTokens: decodeCheckpointUsedTokens(persisted.checkpoint) ?? 0,
-						outputTokens,
-					},
-				};
-				if (streamClosed && emittedToolCall) {
-					yield { type: "finish", reason: { kind: "tool-calls" } };
-				} else if (run.trailers && Number(run.trailers["grpc-status"]) > 0) {
-					const message = run.trailers["grpc-message"] ?? "Cursor agent run failed";
-					yield { type: "finish", reason: { kind: "error", failure: { message, code: classifyCursorError(message) } } };
-				} else {
-					yield { type: "finish", reason: { kind: "stop" } };
-				}
-			}
-		} catch (error) {
-			if (upstream.aborted) {
-				yield {
-					type: "finish",
-					reason: { kind: "aborted", failure: { message: "Cursor request aborted by caller", code: "ABORTED" } },
-				};
-			} else if (error instanceof LlmError) {
-				yield { type: "finish", reason: { kind: "error", failure: { message: error.message, code: error.code } } };
-			} else {
-				const message = error instanceof Error ? error.message : String(error);
-				yield { type: "finish", reason: { kind: "error", failure: { message, code: classifyCursorError(message) } } };
-			}
-		} finally {
-			if (abortRun !== undefined) upstream.removeEventListener("abort", abortRun);
-			consumer.abort("cursor stream consumer stopped");
-			if (!bridgeStored) run?.close();
-		}
-	}
-}
-
-/**
- * Parse a Connect end-stream frame into a structured error. Cursor's agent
- * errors carry a `debug` payload with a stable code and human-readable
- * `title`/`detail`; extracting those gives the user the real reason instead of
- * a bare `resource_exhausted`.
- * @returns {undefined | {code: string, debugCode?: string, message: string}}
- */
-export function parseEndStream(payload) {
-	try {
-		const json = JSON.parse(new TextDecoder().decode(payload));
-		const error = json?.error;
-		if (!error) return undefined;
-		const code = typeof error.code === "string" ? error.code : "unknown";
-		let debugCode;
-		let title;
-		let detail;
-		if (Array.isArray(error.details)) {
-			for (const entry of error.details) {
-				if (!record(entry) || !record(entry.debug)) continue;
-				if (typeof entry.debug.error === "string") debugCode = entry.debug.error;
-				const details = record(entry.debug.details) ? entry.debug.details : undefined;
-				if (details !== undefined) {
-					if (typeof details.title === "string") title = details.title;
-					if (typeof details.detail === "string") detail = details.detail;
-				}
-			}
-		}
-		const fallback = typeof error.message === "string" && error.message.length > 0 ? error.message : undefined;
-		const message = [title, detail].filter((part) => typeof part === "string" && part.length > 0).join(" ")
-			|| (debugCode !== undefined ? `Cursor: ${debugCode}` : undefined)
-			|| (fallback !== undefined ? `Cursor agent error ${code}: ${fallback}` : `Cursor agent error ${code}`);
-		return { code, debugCode, message };
-	} catch {
-		return undefined;
-	}
-}
-
-export function classifyCursorError(message) {
-	if (/\b(?:401|403)\b|unauth|invalid.*(?:key|token|credential)/i.test(message)) return "AUTH";
-	if (/rate.?limit|\b429\b|quota|resource_exhausted|spend.?limit|usage limit|exceeded/i.test(message)) return "RATE_LIMIT";
-	if (/\b400\b|invalid.?request/i.test(message)) return "INVALID_REQUEST";
-	if (/\b408\b|timeout|timed out/i.test(message)) return "TIMEOUT";
-	if (/\b5\d\d\b|internal/i.test(message)) return "SERVER";
-	if (/\b(?:network|connection|socket|fetch|ECONN|http2)/i.test(message)) return "TRANSPORT";
-	return "CURSOR_ERROR";
-}
-
-/**
- * Build the exec-client reply for a native Cursor tool we reject.
- *
- * `ExecClientMessage` numbers its result variants exactly like
- * `ExecServerMessage` numbers its exec variants, which is what lets an exec
- * this build cannot name still be answered: reply on the exec's own field.
- * Every unanswered exec leaves the server waiting for its result, and that
- * wait is what the user sees as a turn that ends right after the model says
- * what it is about to do — the progress watchdog aborts the step and only the
- * preamble text survives.
- */
-export function rejectionFor(exec) {
-	switch (exec.case) {
-		case "readArgs":
-			return { field: 7, payload: encodeReadRejected(exec.path, TOOL_REJECT_REASON) };
-		case "lsArgs":
-			return { field: 8, payload: encodeLsRejected(exec.path, TOOL_REJECT_REASON) };
-		case "grepArgs":
-			return { field: 5, payload: encodeGrepError(TOOL_REJECT_REASON) };
-		case "writeArgs":
-			return { field: 3, payload: encodeWriteRejected(exec.path, TOOL_REJECT_REASON) };
-		case "deleteArgs":
-			return { field: 4, payload: encodeDeleteRejected(exec.path, TOOL_REJECT_REASON) };
-		case "shellArgs":
-			// Non-streaming shell exec replies with ShellResult (field 2).
-			return { field: 2, payload: encodeShellRejectedResult(undefined, undefined, TOOL_REJECT_REASON) };
-		case "shellStreamArgs":
-			// Streaming shell exec replies with ShellStream (field 14); a
-			// ShellResult here leaves the server waiting for the stream.
-			return { field: 14, payload: encodeShellStreamRejected(undefined, undefined, TOOL_REJECT_REASON) };
-		case "backgroundShellSpawnArgs":
-			return { field: 16, payload: encodeBackgroundShellRejectedResult(undefined, undefined, TOOL_REJECT_REASON) };
-		case "fetchArgs":
-			return { field: 20, payload: encodeFetchError(exec.url, TOOL_REJECT_REASON) };
-		case "writeShellStdinArgs":
-			return { field: 23, payload: encodeWriteShellStdinError(TOOL_REJECT_REASON) };
-		case "diagnosticsArgs":
-			return { field: 9, payload: encodeDiagnosticsResult() };
-		case "recordScreenArgs":
-			return { field: 21, payload: encodeMcpError("Screen recording is not available") };
-		case "computerUseArgs":
-			return { field: 22, payload: encodeMcpError("Computer use is not available") };
-		case "listMcpResourcesExecArgs":
-			return { field: 17, payload: encodeMcpError("MCP resources are not available") };
-		case "readMcpResourceExecArgs":
-			return { field: 18, payload: encodeMcpError("MCP resources are not available") };
-		case "unknown": {
-			// Cursor adds exec variants without notice (field 36 appeared this
-			// cycle). Answer on the exec's own field with the generic error
-			// shape, so the model learns the tool is unavailable and falls back
-			// to the MCP tools; live probing shows every coherent reply resumes
-			// the run while silence stalls it.
-			if (typeof exec.field !== "number") return undefined;
-			return { field: exec.field, payload: encodeMcpError(TOOL_REJECT_REASON) };
-		}
-		default:
-			return undefined;
-	}
-}
-//#endregion
-
-//#region apply
-export function apply(ctx, config = {}) {
-	// DSH 0.1.7 composes a plugin's settings section from its own Config, so
-	// there is no section to register: the entry id is the namespace, the
-	// `.volatile()` fields above are the editable surface, and `config` holds
-	// live references rather than a snapshot. All that is left to state is the
-	// page policy — this plugin ships its own Settings → Cursor panel, so the
-	// generated form stays off.
-	//
-	// `configure` replaced `installSection`, which took a namespace, a schema,
-	// and a `setSource` callback to track edits. A DSH build without it (the
-	// 0.1.5 stable line) keeps the Cursor provider and loses only the settings
-	// page, rather than failing the whole entry at boot the way an unguarded
-	// call to a removed method does.
-	ctx.inject(["settings"], (settingsCtx) => {
-		settingsCtx.effect(() => {
-			const settings = settingsCtx.settings;
-			if (typeof settings.configure !== "function") return undefined;
-			return settings.configure({ auto: false }, ctx.fiber);
-		});
-	});
-	const readSettings = () => resolveCursorSettings(config);
-	const readSettingsView = () => {
-		const value = readSettings();
-		const descriptor = ctx.get("settings")?.describe({ redactSecrets: true })
-			.find((entry) => entry.ns === SETTINGS_NAMESPACE);
-		return { ...value, revision: descriptor?.revision ?? 0 };
-	};
-	const settingsController = {
-		read: readSettingsView,
-		update: async (patch, expectedRevision) => {
-			resolveCursorSettings({ ...readSettings(), ...patch });
-			const service = ctx.get("settings");
-			if (service === undefined) throw new Error("Cursor settings storage is unavailable");
-			await service.update(SETTINGS_NAMESPACE, patch, expectedRevision);
-			return readSettingsView();
-		},
-	};
-	const store = new CursorCredentialStore(ctx.credentials, CREDENTIAL_REF);
-	const auth = new CursorAuthService(store, { logger: ctx.logger });
-	const adapter = new CursorAdapter({ auth, settings: readSettings, resolveAttachments: () => ctx.get("attachments"), logger: ctx.logger });
-	ctx.llm.registerAdapter([PROVIDER], adapter);
-	const usageReader = new CursorUsageReader(auth, { logger: ctx.logger });
-
-	const coordinator = new CursorLoginCoordinator(auth, { logger: ctx.logger });
-	const handler = createCursorRpcHandler(coordinator, {
-		openExternal: openCursorAuthUrl,
-		usageReader,
-		modelsProvider: adapter,
-		settings: settingsController,
-	});
-	// Mount the loopback account channel through Connection. `rpc.handle`
-	// registers the channel on the Context that reads the Connection service,
-	// but resolves that Context's `webServer` through the providing plugin's own
-	// scope, so it only works while the Connection plugin itself declares
-	// `webServer`. @deepseek-ai/dsh-client-connection 0.1.5-rc.1 declares
-	// `credentials` alone and mounts `/api` from a nested `webServer` Context, so
-	// `rpc.handle` fails the entire plugin tree with
-	// `cannot get property "webServer" without inject`. Calling the same registry
-	// method Connection uses internally with an explicit owner Context that
-	// declares `webServer` reuses Connection's envelope, trust fence, and browser
-	// authentication, and keeps the route on this plugin's fiber. Fall back to
-	// the public `rpc.handle` if that registry method ever disappears.
-	ctx.inject(["connection", "webServer"], (webCtx) => {
-		const connection = webCtx.connection;
-		webCtx.effect(() => {
-			if (typeof connection.register === "function") {
-				return connection.register(webCtx, CHANNEL, handler, LOOPBACK_RPC);
-			}
-			return connection.rpc.handle(CHANNEL, handler, LOOPBACK_RPC);
-		}, "cursor-subscription: loopback account RPC");
-	});
-}
-//#endregion