feat: add verified desktop ChatGPT subscription connection

app/api/models/route.ts

@@ -1,4 +1,4 @@
-import { modelCatalog } from "@/src/server";
+import { modelCatalog, refreshConnectionCatalog } from "@/src/server";
 
 /**
  * The model picker's data (model-control wave): pi-ai's configured-provider
@@ -7,5 +7,6 @@ import { modelCatalog } from "@/src/server";
  * catalog: no network, stable within a pi-ai version.
  */
 export async function GET() {
-	return Response.json(modelCatalog());
+	await refreshConnectionCatalog();
+	return Response.json(modelCatalog(), { headers: { "Cache-Control": "no-store" } });
 }

components/activity-dock.tsx

@@ -5,7 +5,7 @@ import { ActivityIcon, BotIcon, FileTextIcon, PenToolIcon, PlayIcon, SearchIcon,
 import { cn } from "@/lib/utils";
 import type { CallUsage, JobKind, JobRecord } from "@/src/core";
 import { isModelsRefusal, isProviderKeyRefusal, jobCost } from "@/src/core";
-import { formatTokens, formatUsd } from "@/components/usage-format";
+import { formatTokens, planCostLabel } from "@/components/usage-format";
 import { useServerEvent, useServerResync } from "@/components/server-events";
 
 /**
@@ -348,7 +348,7 @@ export function ActivitySection({ feed: feedProp }: { feed?: ActivityFeed }) {
 				Background agents
 				{blobs.length > 0 && (
 					<span className="ml-auto tabular-nums" data-testid="activity-total-cost" title="Total running cost (inferred)">
-						{totalCost === null ? "—" : formatUsd(totalCost)}
+						{planCostLabel(totalCost, blobs.some((blob) => blob.model?.startsWith("chatgpt:")))}
 				</span>
 				)}
 			</div>
@@ -423,7 +423,7 @@ function ActivityBlobRow({
 					)}
 				</span>
 				<span className="flex shrink-0 items-center gap-2 tabular-nums text-muted-foreground">
-					<span data-testid="activity-blob-cost">{blob.cost === null ? "—" : formatUsd(blob.cost)}</span>
+					<span data-testid="activity-blob-cost">{planCostLabel(blob.cost, blob.model?.startsWith("chatgpt:") === true)}</span>
 					<span data-testid="activity-blob-elapsed">{formatElapsed(now - blob.startedAt)}</span>
 				</span>
 			</button>
@@ -462,7 +462,7 @@ function ActivityBlobRow({
 						<div className="flex justify-between gap-2">
 							<span>Cost</span>
 							<span className="tabular-nums text-card-foreground">
-								{blob.cost === null ? "—" : formatUsd(blob.cost)}
+								{planCostLabel(blob.cost, blob.model?.startsWith("chatgpt:") === true)}
 							</span>
 						</div>
 						<div className="flex justify-between gap-2">

components/graph-view.tsx

@@ -294,9 +294,17 @@ function timeAgo(startedAt: number): string {
 
 /** Run/stage cost at 3 decimals (pipeline pennies are real money at this
  *  scale); headline totals use 2. Null (unpriced model) renders "—". */
-function fmtRunCost(dollars: number | null): string {
+function fmtRunCost(dollars: number | null, subscription = false): string {
+	if (subscription) return dollars === null ? "ChatGPT plan" : `$${dollars.toFixed(3)} API + ChatGPT plan`;
 	return dollars === null ? "—" : `$${dollars.toFixed(3)}`;
 }
+function graphUsageLabel(costs: Pick<GraphCostSummary, "stages">): string {
+	const stages = [costs.stages.extraction, costs.stages.audit];
+	const plan = stages.some((stage) => (stage.subscriptionRuns ?? 0) > 0);
+	const priced = stages.some((stage) => stage.runs - stage.unpricedRuns - (stage.subscriptionRuns ?? 0) > 0);
+	const dollars = displayCostTotals(costs).totalDollars;
+	return plan ? (priced ? `$${dollars.toFixed(2)} API + ChatGPT plan` : "ChatGPT plan") : `$${dollars.toFixed(2)}`;
+}
 
 /**
  * The cost figures the UI DISPLAYS (owner refinement round, 2026-09-24):
@@ -349,7 +357,7 @@ export function GraphCostControl({ costs }: { costs: GraphCostSummary }) {
 					open ? "text-foreground" : "text-muted-foreground hover:text-foreground"
 				}`}
 			>
-				${displayCostTotals(costs).totalDollars.toFixed(2)}
+				{graphUsageLabel(costs)}
 			</button>
 			</Hint>
 			{open && (
@@ -378,7 +386,7 @@ function GraphCostPanel({ costs }: { costs: GraphCostSummary }) {
 	];
 	// The displayed figures come from displayCostTotals (naming excluded —
 	// the pill and this panel must agree).
-	const { totalDollars, unpricedRuns } = displayCostTotals(costs);
+	const { unpricedRuns } = displayCostTotals(costs);
 	return (
 		<div className="w-80 max-w-[calc(100vw-1.5rem)] rounded-xl border bg-card/95 p-3 text-xs shadow-lg backdrop-blur">
 			<div className="mb-2 flex items-baseline justify-between">
@@ -386,7 +394,7 @@ function GraphCostPanel({ costs }: { costs: GraphCostSummary }) {
 					Graph pipeline costs
 				</span>
 				<span className="text-foreground tabular-nums">
-					${totalDollars.toFixed(2)}
+					{graphUsageLabel(costs)}
 					{unpricedRuns > 0 && (
 						<span className="text-muted-foreground"> · {unpricedRuns} unpriced</span>
 					)}
@@ -397,8 +405,8 @@ function GraphCostPanel({ costs }: { costs: GraphCostSummary }) {
 					<div key={label} className="flex items-baseline justify-between gap-2 text-[11px]">
 						<span className="shrink-0 text-muted-foreground">{label}</span>
 						<span className="min-w-0 text-right tabular-nums">
-							{fmtRunCost(stage.runs === 0 || (stage.unpricedRuns > 0 && stage.dollars === 0) ? null : stage.dollars)}
-							{stage.unpricedRuns > 0 && stage.dollars > 0 && (
+							{fmtRunCost(stage.runs - stage.unpricedRuns - (stage.subscriptionRuns ?? 0) <= 0 ? null : stage.dollars, (stage.subscriptionRuns ?? 0) > 0)}
+							{stage.unpricedRuns > 0 && (stage.runs - stage.unpricedRuns - (stage.subscriptionRuns ?? 0) > 0 || (stage.subscriptionRuns ?? 0) > 0) && (
 								<span className="text-muted-foreground"> +{stage.unpricedRuns} unpriced</span>
 							)}
 							<span className="text-muted-foreground">
@@ -422,7 +430,7 @@ function GraphCostPanel({ costs }: { costs: GraphCostSummary }) {
 									</span>
 								)}
 							</span>
-							<span className="shrink-0 tabular-nums">{fmtRunCost(run.dollars)}</span>
+							<span className="shrink-0 tabular-nums">{fmtRunCost(run.dollars, (run.subscriptionRuns ?? 0) > 0)}{run.dollars === null && run.subscriptionRuns === 1 ? " +1 unpriced" : ""}</span>
 						</div>
 					))}
 			</div>

components/model-picker.tsx

@@ -8,7 +8,7 @@ import { ModelMenuList, ThinkingMenuList } from "@/components/settings/model-men
 import { composeSlotModel } from "@/components/settings/agent-section";
 import { MenuShell } from "@/components/menu-shell";
 import { Hint } from "@/components/ui/hint";
-import { notifySettingsSaved, useSettingsSaved } from "@/components/settings-signal";
+import { notifySettingsSaved, onSettingsSaved, useSettingsSaved } from "@/components/settings-signal";
 import { defaultOrderLevel, populateLevel, THINKING_LABEL } from "@/components/thinking-default";
 
 /** GET /api/models is cached per page load for instant menu opens, but
@@ -17,51 +17,40 @@ import { defaultOrderLevel, populateLevel, THINKING_LABEL } from "@/components/t
  *  every model-menu open (stale-while-revalidate: the cached list shows
  *  instantly, the fresh one swaps in while the menu is open). */
 let catalogCache: ModelCatalogResponse | undefined;
-let catalogInflight: Promise<ModelCatalogResponse> | undefined;
-
-function fetchCatalog(force = false): Promise<ModelCatalogResponse> {
-	if (force) {
-		catalogCache = undefined;
-		catalogInflight = undefined;
-	}
-	catalogInflight ??= fetch("/api/models")
-		.then((r) => {
-			if (!r.ok) throw new Error(`catalog request failed: ${r.status}`);
-			return r.json() as Promise<ModelCatalogResponse>;
-		})
-		.then((c) => {
-			catalogCache = c;
-			return c;
-		})
-		.catch((err) => {
-			catalogInflight = undefined; // retry on next mount/open
-			throw err;
-		});
-	return catalogInflight;
+let catalogInflight: Promise<void> | undefined;
+let catalogGeneration = 0;
+let stopCatalogSignals: (() => void) | undefined;
+const catalogReaders = new Set<() => void>();
+const publishCatalog = () => { for (const reader of catalogReaders) reader(); };
+function fetchCatalog(revalidate = false): void {
+	if (revalidate) { catalogGeneration++; catalogInflight = undefined; }
+	if (catalogInflight) return;
+	const generation = catalogGeneration;
+	catalogInflight = fetch("/api/models")
+		.then((r) => { if (!r.ok) throw new Error(`catalog request failed: ${r.status}`); return r.json() as Promise<ModelCatalogResponse>; })
+		.then((c) => { if (generation === catalogGeneration) { catalogCache = c; publishCatalog(); } })
+		.catch(() => { /* existing unavailable/usage-only posture, no retry timer */ })
+		.finally(() => { if (generation === catalogGeneration) catalogInflight = undefined; });
 }
-
 export function useModelCatalog(): { catalog: ModelCatalogResponse | undefined; refresh: () => void } {
 	const [catalog, setCatalog] = useState(catalogCache);
 	useEffect(() => {
-		if (catalogCache) return;
-		let alive = true;
-		fetchCatalog()
-			.then((c) => alive && setCatalog(c))
-			.catch(() => {
-				/* the picker stays closed on failure; the buttons still
-				   display the strong slot from the record fetch */
-			});
-		return () => {
-			alive = false;
-		};
-	}, []);
-	const refresh = useCallback(() => {
-		fetchCatalog(true)
-			.then(setCatalog)
-			.catch(() => {
-				/* refresh is best-effort: the cached list stays on failure */
+		const first = catalogReaders.size === 0;
+		const read = () => setCatalog(catalogCache);
+		catalogReaders.add(read);
+		if (first) {
+			stopCatalogSignals = onSettingsSaved((section) => {
+				if (section === "keys") {
+					catalogGeneration++; catalogInflight = undefined; catalogCache = undefined;
+					publishCatalog(); fetchCatalog();
+				} else if (section === "models") fetchCatalog(true);
 			});
+			fetchCatalog(true);
+		} else fetchCatalog();
+		read();
+		return () => { catalogReaders.delete(read); if (!catalogReaders.size) { stopCatalogSignals?.(); stopCatalogSignals = undefined; } };
 	}, []);
+	const refresh = useCallback(() => fetchCatalog(true), []);
 	return { catalog, refresh };
 }
 
@@ -322,6 +311,7 @@ export function ModelControl({ chatId }: { chatId?: string }) {
 
 	return (
 		<div ref={containerRef} className="relative flex items-center gap-1.5" data-testid="model-control">
+			{effectiveProvider === "chatgpt" && <a href="https://chatgpt.com/settings/usage" target="_blank" rel="noreferrer" className="text-xs text-muted-foreground underline-offset-2 hover:underline">Using ChatGPT plan</a>}
 			<Hint
 				// Trimmed (owner, subtlety round 2026-09-25): "Model selector"
 				// — the long sentence was way too much. ⌘K opens this same

components/settings/agent-section.tsx

@@ -344,18 +344,23 @@ export function AgentSection({
 	// fetch each per section mount. The models read ALSO re-runs on the
 	// cross-surface signal (fix wave: another surface's write re-reads
 	// into the slot rows live).
-	const readModels = useCallback(() => {
+	const modelsRequest = useRef(0);
+	const modelsMounted = useRef(false);
+	const readModels = useCallback((connectionChanged = false) => {
+		const request = ++modelsRequest.current;
+		if (connectionChanged) setCatalog(null);
 		fetch("/api/models")
-			.then((r) => r.json())
+			.then((r) => { if (!r.ok) throw new Error("Catalog unavailable"); return r.json(); })
 			.then((body: ModelsBody) => {
+				if (!modelsMounted.current || request !== modelsRequest.current) return;
 				setCatalog(body.models);
 				setSlots(body.slots);
 			})
-			.catch(() => setCatalog(null));
+			.catch(() => { if (modelsMounted.current && request === modelsRequest.current) setCatalog(null); });
 	}, []);
 
 	useSettingsSaved((section) => {
-		if (section === "models") readModels();
+		if (section === "models" || section === "keys") readModels(section === "keys");
 		// THE state-echo root cause: the prompt switches write through the
 		// agent section, but this view (the toggles' truth) never re-read —
 		// the settings-saved law (every reader re-reads) was missing this
@@ -364,6 +369,7 @@ export function AgentSection({
 	});
 
 	useEffect(() => {
+		modelsMounted.current = true;
 		readModels();
 		fetch("/api/settings/prompts")
 			.then((r) => r.json())
@@ -380,6 +386,7 @@ export function AgentSection({
 				setEngineKeys(present);
 			})
 			.catch(() => {});
+		return () => { modelsMounted.current = false; modelsRequest.current++; };
 	}, []);
 
 	// The reason line re-reads /api/search AFTER an engine write lands —

components/settings/chatgpt-connection.tsx

@@ -0,0 +1,86 @@
+"use client";
+
+import { useEffect, useState } from "react";
+import type { ChatGPTStatus } from "@/src/core";
+
+export const CHATGPT_NOTICE = "LearnHarness stores your ChatGPT account details and encrypted credentials locally on this device. When this connection is used, relevant conversation context, learning material, and tool results are sent directly to OpenAI. ChatGPT plan limits apply; any additional credits are authorized and managed in ChatGPT, not LearnHarness.";
+const empty: ChatGPTStatus = { available: false, pending: false, activeId: null, accounts: [], error: null };
+
+/** Status/control surface only; OAuth URLs, codes, and tokens never reach React. */
+export function ChatGPTConnection() {
+	const [status, setStatus] = useState<ChatGPTStatus>(empty);
+	const [selected, setSelected] = useState("");
+	const [busy, setBusy] = useState(false);
+	const [error, setError] = useState<string | null>(null);
+	useEffect(() => {
+		const bridge = window.lh?.chatgpt;
+		if (!bridge) return;
+		let mounted = true;
+		const refresh = async () => {
+			try {
+				const next = await bridge.status();
+				if (!mounted) return;
+				setStatus(next);
+				setSelected((old) => old || next.activeId || "");
+			} catch { if (mounted) setError("The desktop connection is unavailable. Retry when the app is ready."); }
+		};
+		void refresh();
+		const off = bridge.onChange(() => void refresh());
+		return () => { mounted = false; off(); };
+	}, []);
+	const account = status.accounts.find((a) => a.id === selected);
+	async function control(action: () => Promise<ChatGPTStatus>) {
+		if (busy) return;
+		setBusy(true);
+		setError(null);
+		try {
+			const next = await action();
+			if (typeof next.available === "boolean") setStatus(next);
+			setError(next.error ?? null);
+		} catch { setError("The ChatGPT connection could not be changed. Try again."); }
+		finally { setBusy(false); }
+	}
+	const bridge = typeof window !== "undefined" ? window.lh?.chatgpt : undefined;
+	return (
+		<div className="flex flex-col gap-3" data-testid="chatgpt-connection">
+			<div className="flex items-center justify-between gap-3">
+				<span className="font-ui text-xs text-foreground">ChatGPT</span>
+				<span className="text-xs text-muted-foreground" role="status">{status.pending ? "Waiting for browser sign-in…" : status.accounts.some((a) => a.id === status.activeId && a.connected) ? "ChatGPT plan connection" : "Not connected"}</span>
+			</div>
+			<p className="text-xs leading-relaxed text-muted-foreground" data-testid="chatgpt-privacy-notice">{CHATGPT_NOTICE}</p>
+			<div className="flex flex-wrap gap-3 text-xs text-muted-foreground">
+				<a href="https://openai.com/policies/privacy-policy/" target="_blank" rel="noreferrer" className="underline underline-offset-2">OpenAI privacy</a>
+				<a href="https://chatgpt.com/settings/usage" target="_blank" rel="noreferrer" className="underline underline-offset-2">Manage usage</a>
+			</div>
+			{status.accounts.length > 0 && (
+				<label className="flex flex-col gap-1 text-xs text-muted-foreground">
+					Saved ChatGPT account / registration
+					<select aria-label="Saved ChatGPT account" value={selected} onChange={(event) => setSelected(event.target.value)} disabled={busy || status.pending} className="rounded-sm border border-border bg-background px-2 py-1.5 text-foreground">
+						<option value="">Add another account or workspace</option>
+						{status.accounts.map((item) => <option key={item.id} value={item.id}>{item.email} · registration {item.registration}{item.id === status.activeId ? " · active" : ""}</option>)}
+					</select>
+					<span>Registrations can belong to different workspaces. An email does not identify a workspace.</span>
+				</label>
+			)}
+			<div className="flex flex-wrap items-center gap-3">
+				{/* Official black/white format and unmodified OpenAI-provided mark. */}
+				<button type="button" data-testid="chatgpt-login" disabled={!status.available || busy || status.pending} onClick={() => bridge && void control(() => bridge.login(selected || undefined))} className="inline-flex min-h-10 items-center justify-center gap-2 rounded-md border border-white/20 bg-black px-4 py-2 text-sm font-medium text-white disabled:opacity-40">
+					{/* eslint-disable-next-line @next/next/no-img-element */}
+					<img src="/siwc/chatgpt-logo-white.svg" width="21" height="21" alt="" aria-hidden="true" />
+					Continue with ChatGPT
+				</button>
+				{status.pending && <button type="button" onClick={() => bridge && void control(() => bridge.cancel())} className="text-xs underline underline-offset-2">Cancel sign-in</button>}
+				{account?.connected && selected !== status.activeId && <button type="button" disabled={busy || status.pending} onClick={() => bridge && void control(() => bridge.select(selected))} className="text-xs underline underline-offset-2">Use this account</button>}
+				{account && <button type="button" disabled={busy} onClick={() => bridge && void control(() => bridge.logout(selected))} className="text-xs underline underline-offset-2">Disconnect</button>}
+			</div>
+			{!status.available && <p className="text-xs text-muted-foreground">ChatGPT sign-in is available only in the desktop app with secure OS credential storage. API-key connections remain available here.</p>}
+			{account && (
+				<label className="flex items-start gap-2 text-xs text-foreground">
+					<input type="checkbox" checked={account.background} disabled={busy || !status.available || status.pending} onChange={(event) => bridge && void control(() => bridge.consent(account.id, event.target.checked))} className="mt-0.5" />
+					<span>Allow background use<span className="mt-1 block leading-relaxed text-muted-foreground">Allow extraction, compaction, delegated jobs, and automatic follow-up work to use this ChatGPT connection.</span></span>
+				</label>
+			)}
+			{(error || status.error || status.catalogError) && <p role="alert" className="text-xs text-red-400">{error || status.error || status.catalogError}</p>}
+		</div>
+	);
+}

components/settings/keys-section.tsx

@@ -26,6 +26,7 @@
 import { useEffect, useRef, useState } from "react";
 import { AUTH_KEYS, type AuthKeyStatus } from "@/src/core";
 import { SettingsGroup } from "./controls.tsx";
+import { ChatGPTConnection } from "./chatgpt-connection.tsx";
 import { ProviderMenuList } from "./model-menu.tsx";
 import { MenuShell } from "@/components/menu-shell";
 import { notifySettingsSaved } from "@/components/settings-signal";
@@ -412,6 +413,7 @@ export function KeysSection({
 
 	return (
 		<>
+			<SettingsGroup label="ChatGPT" testid="group-chatgpt"><ChatGPTConnection /></SettingsGroup>
 			<SettingsGroup label="Model providers" testid="group-provider-keys">
 				{providerRows.map((status) => (
 					<KeyTableRow

components/settings/model-menu.tsx

@@ -126,7 +126,7 @@ export function ModelMenuList({
 							<span className="flex items-center justify-between gap-2">
 								<span className="truncate text-card-foreground">{m.name}</span>
 								<span className="shrink-0 tabular-nums text-muted-foreground" title="USD per 1M tokens in/out">
-									{priceLabel(m.cost)}
+									{m.provider === "chatgpt" ? "ChatGPT plan" : priceLabel(m.cost)}
 								</span>
 							</span>
 							<span className="flex items-center justify-between gap-2 text-muted-foreground">

components/settings/settings-dialog.tsx

@@ -55,7 +55,7 @@ import { KeyboardSection } from "./keyboard-section.tsx";
  *  real question users ask). */
 const SECTIONS = [
 	{ id: "general", label: "General" },
-	{ id: "keys", label: "Keys" },
+	{ id: "keys", label: "Connections" },
 	{ id: "agent", label: "Agent" },
 	{ id: "appearance", label: "Appearance" },
 	{ id: "graph", label: "Graph" },
@@ -376,7 +376,7 @@ export function SettingsDialogBody({
 								</SettingsSection>
 							)}
 							{active === "keys" && (
-								<SettingsSection title="Keys" testid="section-keys">
+								<SettingsSection title="Connections" testid="section-keys">
 									<KeysSection initialRow={initialRow ?? undefined} />
 								</SettingsSection>
 							)}

components/setup-cta.tsx

@@ -15,12 +15,10 @@
  * its own save), the same re-read pattern chat.tsx uses for "models".
  */
 
-import { useCallback, useEffect, useState } from "react";
 import type { AuthKeyStatus } from "@/src/core";
 import { classifyKeys, type ProviderKeyInfo } from "@/components/settings/keys-section";
 import { useModelCatalog } from "@/components/model-picker";
 import type { ModelCatalogResponse } from "@/src/core";
-import { useSettingsSaved } from "@/components/settings-signal";
 
 /** Pure: does a MODEL-PROVIDER key exist among the statuses? Reuses
  *  classifyKeys (the keys section's own classification — no drift):
@@ -40,22 +38,8 @@ export function hasProviderKey(
  *  failed read claiming "zero keys" would be a lie the CTA flow then
  *  leans on). Re-reads on the keys signal. */
 export function useZeroKeys(): boolean {
-	const [zeroKeys, setZeroKeys] = useState(false);
-	const read = useCallback(() => {
-		fetch("/api/keys")
-			.then((r) => r.json())
-			.then((body: { keys?: AuthKeyStatus[]; providers?: ProviderKeyInfo[] }) => {
-				setZeroKeys(!hasProviderKey(body.keys ?? [], body.providers));
-			})
-			.catch(() => {});
-	}, []);
-	useEffect(() => {
-		read();
-	}, [read]);
-	useSettingsSaved((section) => {
-		if (section === "keys") read();
-	});
-	return zeroKeys;
+	const { catalog } = useModelCatalog();
+	return catalog ? !catalog.providers.some((provider) => provider.keyPresent) : false;
 }
 
 /** The pill itself — the models CTA's chrome and placement, the keys
@@ -93,9 +77,6 @@ export function strongKeyless(catalog: ModelCatalogResponse | undefined): boolea
  *  (a provider key saved or deleted) and "models" (the strong slot
  *  re-pinned). Live-updates with the page, never mid-send. */
 export function useStrongKeyless(): boolean {
-	const { catalog, refresh } = useModelCatalog();
-	useSettingsSaved((section) => {
-		if (section === "keys" || section === "models") refresh();
-	});
+	const { catalog } = useModelCatalog();
 	return strongKeyless(catalog);
 }

components/setup-dialog.tsx

@@ -44,6 +44,7 @@ import { notifySettingsSaved, useSettingsSaved } from "@/components/settings-sig
 import { Hint } from "@/components/ui/hint";
 import { type ProviderKeyInfo } from "@/components/settings/keys-section";
 import { ProviderMenuList } from "@/components/settings/model-menu";
+import { ChatGPTConnection } from "@/components/settings/chatgpt-connection";
 import { defaultOrderLevel, populateLevel, THINKING_LABEL } from "@/components/thinking-default";
 import { ThinkingMenuList } from "@/components/settings/model-menu";
 import type { ThinkingLevel } from "@earendil-works/pi-agent-core";
@@ -147,7 +148,8 @@ export function SetupKeysBody({
 }) {
 	return (
 		<div className="flex flex-col gap-3 px-6 py-5" data-testid="setup-stage-keys">
-			<p className="text-sm text-foreground">Add a provider key.</p>
+			<ChatGPTConnection />
+			<p className="border-t border-border pt-3 text-sm text-foreground">Or add an API key.</p>
 			<p className="text-xs text-muted-foreground">
 				Choose your AI provider and paste its API key. You’ll pick your models next.
 			</p>
@@ -265,16 +267,19 @@ export function SetupDialog({
 	// moment ago (stage 1's own screen) widens the pickers immediately.
 	useEffect(() => {
 		if (!open) return;
+		let cancelled = false;
+		setBody(null);
 		setPicked({});
 		setLevels({});
 		setError(null);
 		fetch("/api/keys")
 			.then((r) => r.json())
-			.then((b: { providers?: ProviderKeyInfo[] }) => setKeyProviders(b.providers ?? []))
-			.catch(() => setKeyProviders([]));
+			.then((b: { providers?: ProviderKeyInfo[] }) => { if (!cancelled) setKeyProviders(b.providers ?? []); })
+			.catch(() => { if (!cancelled) setKeyProviders([]); });
 		fetch("/api/models")
-			.then((r) => r.json())
+			.then((r) => { if (!r.ok) throw new Error("Catalog unavailable"); return r.json(); })
 			.then((b: ModelsBody) => {
+				if (cancelled) return;
 				setBody(b);
 				setFailed(false);
 				// Prefill (owner, 2026-09-22): each pinned tier opens with its
@@ -292,7 +297,8 @@ export function SetupDialog({
 				}
 				setLevels(preLevels);
 			})
-			.catch(() => setFailed(true));
+			.catch(() => { if (!cancelled) setFailed(true); });
+		return () => { cancelled = true; };
 	}, [open, keysTick]);
 
 	// Esc + outside-click — layered dismissal, closable WITHOUT

components/sidebar.tsx

@@ -21,6 +21,7 @@ import { fuzzyScore } from "@/src/core";
 import { useServerEvent, useServerResync } from "@/components/server-events";
 import type { ChatListItem } from "@/src/core";
 import { kbdEventForAction } from "@/src/core";
+import { notifySettingsSaved } from "@/components/settings-signal";
 
 /**
  * Toggleable sidebar — NOT a modal drawer: no backdrop, no focus theft. It
@@ -47,6 +48,7 @@ const STORAGE_KEY = "lh:sidebar-open";
  *
  * Mirrors SIDEBAR_AUTOCLOSE in electron/main.cjs — keep in sync.
  */
+import type { ChatGPTControlBridge } from "@/src/core";
 const SIDEBAR_AUTOCLOSE = 928;
 
 // Module scope, guarded for SSR (the server has no navigator).
@@ -58,6 +60,7 @@ declare global {
 		lh?: {
 			/** Read-only host OS; determines the native-control gutter. */
 			platform?: string;
+			chatgpt?: ChatGPTControlBridge;
 			expandWindow: () => void;
 			/** The zoom truth (owner round 7): subscribes to "lh:zoom" pushes; returns unsubscribe. */
 			onZoom?: (cb: (factor: number) => void) => () => void;
@@ -79,6 +82,13 @@ export function useSidebar() {
 }
 
 export function SidebarProvider({ children }: { children: React.ReactNode }) {
+	useEffect(() => {
+		const bridge = window.lh?.chatgpt;
+		if (!bridge) return;
+		const off = bridge.onChange(() => notifySettingsSaved("keys"));
+		notifySettingsSaved("keys"); // cover changes missed before this root-lifetime subscription
+		return off;
+	}, []);
 	// Start closed so server and client first-render agree (no hydration
 	// mismatch); the stored value is applied after mount.
 	const [open, setOpen] = useState(false);

components/stats-bar.tsx

@@ -83,11 +83,11 @@ export function StatsBar({
 	const { catalog } = useModelCatalog();
 	const limits = useMemo(() => (catalog ? catalogLimits(catalog) : undefined), [catalog]);
 	const ctx = contextWindow({ turns }, limits);
-	const { cost, unpriced } = sessionCost({ turns });
+	const { cost, unpriced, subscription = 0 } = sessionCost({ turns });
 	// Honest cost: show dollars only when at least one turn priced;
 	// otherwise "—" (never a fake $0.0000).
-	const pricedCount = turns.length - unpriced;
-	const costLabel = pricedCount > 0 ? formatUsd(cost) : "—";
+	const pricedCount = turns.length - unpriced - subscription;
+	const costLabel = subscription ? (pricedCount > 0 ? `${formatUsd(cost)} API + ChatGPT plan` : "ChatGPT plan") : pricedCount > 0 ? formatUsd(cost) : "—";
 	const lastModel = turns.at(-1)?.model;
 
 	return (
@@ -143,9 +143,10 @@ export function StatsBar({
 								</span>
 							</div>
 							<div className="flex justify-between gap-2">
-								<span>Total cost</span>
+								<span>{subscription ? "Usage" : "Total cost"}</span>
 								<span className="tabular-nums text-card-foreground">{costLabel}</span>
 							</div>
+							{subscription > 0 && <a href="https://chatgpt.com/settings/usage" target="_blank" rel="noreferrer" className="underline underline-offset-2">Manage ChatGPT usage</a>}
 							{unpriced > 0 && (
 								<div className="flex justify-between gap-2">
 									<span>Unpriced turns</span>
@@ -168,7 +169,7 @@ export function StatsBar({
 												</span>
 												<span className="flex shrink-0 items-center gap-2 tabular-nums text-card-foreground">
 													<span className="text-muted-foreground">{turn.steps.length} steps</span>
-													<span>{c === null ? "—" : formatUsd(c)}</span>
+													<span>{turn.model.startsWith("chatgpt:") ? "ChatGPT plan" : c === null ? "—" : formatUsd(c)}</span>
 												</span>
 											</li>
 										);

components/usage-format.ts

@@ -8,6 +8,11 @@ export function formatUsd(c: number): string {
 	return c >= 1 ? `$${c.toFixed(2)}` : `$${c.toFixed(4)}`;
 }
 
+export function planCostLabel(dollars: number | null, subscription: boolean, format = formatUsd): string {
+	if (subscription) return dollars === null ? "ChatGPT plan" : `${format(dollars)} API + ChatGPT plan`;
+	return dollars === null ? "—" : format(dollars);
+}
+
 /** 38000 → "38k", 1234567 → "1.2M" — trimmed, never "38.0k". */
 export function formatTokens(n: number): string {
 	if (n >= 1e6) return `${+(n / 1e6).toFixed(1)}M`;

electron-builder.yml

@@ -81,6 +81,8 @@ files:
   - node_modules/universalify/**/*
   - node_modules/ms/**/*
   - node_modules/argparse/**/*
+  # Main-process ID-token signature verification (no bundled OAuth implementation).
+  - node_modules/jose/**/*
 extraResources:
   # Copy from the PARENT dir (.next), not .next/standalone itself:
   # electron-builder hard-excludes a copy's root node_modules and the

electron/chatgpt-ipc.cjs

@@ -0,0 +1,42 @@
+"use strict";
+const { publicError } = require("./chatgpt.cjs");
+
+/** Inherited Node child IPC, not a listening HTTP/socket token service. */
+function attachChatGPTChild(child, connection) {
+	const send = (value) => {
+		if (child.connected) child.send({ channel: "lh:chatgpt", ...value }, () => {});
+	};
+	const listener = async (message) => {
+		if (!message || message.channel !== "lh:chatgpt" || typeof message.id !== "string" || message.id.length > 100) return;
+		try {
+			let value;
+			if (message.op === "status") value = { status: await connection.status(), ...connection.stamp() };
+			else if (message.op === "models") value = await connection.models();
+			else if (message.op === "lease" && typeof message.background === "boolean") value = await connection.lease(message.background, message);
+			else throw new Error("Invalid private operation");
+			send({ id: message.id, op: message.op, ok: true, value });
+		} catch (err) { send({ id: message.id, op: message.op, ok: false, stamp: connection.stamp(), error: publicError(err) }); }
+	};
+	child.on("message", listener);
+	return { send, dispose: () => child.removeListener("message", listener) };
+}
+
+/** Sender must be the trusted window's main frame at its exact app origin. */
+function trustedSender(event, origin, BrowserWindow) {
+	const win = BrowserWindow.fromWebContents(event.sender);
+	try {
+		return !!win && event.senderFrame === event.sender.mainFrame && new URL(event.senderFrame.url).origin === origin;
+	} catch { return false; }
+}
+async function chatGPTControl(command, connection) {
+	try {
+		if (command?.op === "status") return await connection.status();
+		if (command?.op === "login" && (command.id === undefined || typeof command.id === "string")) return await connection.login(command.id);
+		if (command?.op === "cancel") return connection.cancelLogin();
+		if (command?.op === "select" && typeof command.id === "string") return await connection.select(command.id);
+		if (command?.op === "logout" && typeof command.id === "string") return await connection.logout(command.id);
+		if (command?.op === "consent" && typeof command.id === "string" && typeof command.enabled === "boolean") return await connection.consent(command.id, command.enabled);
+		return { error: "Unsupported ChatGPT control." };
+	} catch (err) { return { error: publicError(err) }; }
+}
+module.exports = { attachChatGPTChild, trustedSender, chatGPTControl };

electron/chatgpt.cjs

@@ -0,0 +1,513 @@
+"use strict";
+
+// LearnHarness-owned public-client OAuth. Never import pi's bundled login.
+// This module is main-process-only; dependencies are injectable for scratch tests.
+const crypto = require("node:crypto");
+const http = require("node:http");
+const fs = require("node:fs/promises");
+const path = require("node:path");
+
+const ISSUER = "https://auth.openai.com";
+const RESOURCE = "https://api.openai.com/v1";
+const AUTHORIZE = `${ISSUER}/api/accounts/authorize`;
+const TOKEN = `${ISSUER}/api/accounts/oauth/token`;
+const JWKS = `${ISSUER}/.well-known/jwks.json`;
+const SCOPES = ["openid", "profile", "email", "offline_access", "resource.invoke", "chatgpt.tokens.use.direct"];
+const random = () => crypto.randomBytes(32).toString("base64url");
+const fail = (code) => Object.assign(new Error(code), { code });
+const publicError = (err) => ({
+	cancelled: "ChatGPT sign-in cancelled.",
+	timeout: "ChatGPT sign-in timed out. Try again.",
+	secure_storage: "Secure credential storage is unavailable. ChatGPT sign-in requires the desktop app and an OS secure-storage backend.",
+	storage: "ChatGPT credentials could not be saved securely. Reconnect before making requests.",
+	identity: "ChatGPT sign-in could not be verified. Select the intended saved account and try again.",
+	scope: "ChatGPT plan permission is missing. Reconnect and authorize ChatGPT plan use.",
+	disconnected: "This ChatGPT connection is disconnected. Reconnect in Connections.",
+	background: "Background use is off for this ChatGPT account. Enable Allow background use in Connections, or explicitly select an API-key model for that workload.",
+	catalog: "The ChatGPT account model catalog is unavailable. Retry the connection; no API-key fallback was used.",
+	expired: "The ChatGPT session expired. Reconnect in Connections.",
+	network: "ChatGPT could not be reached. Check your connection and try again.",
+	pending: "A ChatGPT sign-in is already pending.",
+	unavailable: "ChatGPT sign-in is unavailable in browser-only mode.",
+}[err?.code] || "ChatGPT sign-in failed. Try again; no API-key fallback was used.");
+
+const nonempty = (v) => typeof v === "string" && v.length > 0 && !/[\x00-\x20\x7f]/.test(v);
+const planScopes = (tokens) => !!tokens?.accessToken && SCOPES.every((scope) => tokens.scopes.includes(scope));
+/** Reject the entire encrypted record. Encryption is not schema validation. */
+function validateState(state) {
+	if (!state || state.version !== 1 || !nonempty(state.hostId) || !Array.isArray(state.registrations) || !(state.activeId === null || nonempty(state.activeId))) throw fail("storage");
+	const ids = new Set(); const clients = new Set();
+	for (const r of state.registrations) {
+		if (!r || !nonempty(r.id) || ids.has(r.id) || !nonempty(r.clientId) || r.clientId === "dynamic_agent_client" || clients.has(r.clientId) || r.issuer !== ISSUER || !nonempty(r.subject) || typeof r.background !== "boolean" || (r.email !== undefined && typeof r.email !== "string") || (Object.hasOwn(r, "reconnectRequired") && r.reconnectRequired !== true)) throw fail("storage");
+		ids.add(r.id); clients.add(r.clientId);
+		if (r.tokens !== undefined) {
+			const t = r.tokens;
+			if (!t || !nonempty(t.idToken) || !Array.isArray(t.scopes) || t.scopes.some((s) => !nonempty(s)) || !Number.isFinite(t.expiresAt) || t.expiresAt < 0 || (t.accessToken !== undefined && (!nonempty(t.accessToken) || t.accessToken.startsWith("sk-"))) || (t.refreshToken !== undefined && !nonempty(t.refreshToken))) throw fail("storage");
+		}
+	}
+	if (state.activeId !== null && !ids.has(state.activeId)) throw fail("storage");
+	return state;
+}
+
+/** Encrypted envelope only, outside learning data. No plaintext fallback. */
+function encryptedStore(file, safeStorage) {
+	function secure() {
+		if (!safeStorage.isEncryptionAvailable() || safeStorage.getSelectedStorageBackend?.() === "basic_text") throw fail("secure_storage");
+	}
+	return {
+		secure,
+		async read() {
+			secure();
+			try {
+				const envelope = JSON.parse(await fs.readFile(file, "utf8"));
+				if (envelope.version !== 1 || typeof envelope.encrypted !== "string" || !/^(?:[A-Za-z0-9+/]{4})*(?:[A-Za-z0-9+/]{2}==|[A-Za-z0-9+/]{3}=)?$/.test(envelope.encrypted) || !envelope.encrypted) throw fail("storage");
+				const state = JSON.parse(safeStorage.decryptString(Buffer.from(envelope.encrypted, "base64")));
+				return validateState(state);
+			} catch (err) {
+				if (err.code === "ENOENT") return null;
+				throw fail("storage");
+			}
+		},
+		async write(state) {
+			secure();
+			validateState(state);
+			const temp = `${file}.${crypto.randomUUID()}.tmp`;
+			let handle;
+			try {
+				const encrypted = safeStorage.encryptString(JSON.stringify(state)).toString("base64");
+				await fs.mkdir(path.dirname(file), { recursive: true, mode: 0o700 });
+				handle = await fs.open(temp, "wx", 0o600);
+				await handle.writeFile(JSON.stringify({ version: 1, encrypted }));
+				await handle.sync();
+				await handle.close();
+				handle = null;
+				await fs.rename(temp, file);
+				// Directory fsync is supported on Unix; the file fsync above is mandatory.
+				if (process.platform !== "win32") {
+					const dir = await fs.open(path.dirname(file), "r");
+					try { await dir.sync(); } finally { await dir.close(); }
+				}
+			} catch {
+				throw fail("storage");
+			} finally {
+				await handle?.close().catch(() => {});
+				await fs.unlink(temp).catch(() => {});
+			}
+		},
+	};
+}
+
+function allowedAuthEndpoint(value) {
+	try {
+		const url = new URL(value);
+		return url.origin === ISSUER && !url.username && !url.password && !url.search && !url.hash;
+	} catch { return false; }
+}
+
+function createChatGPT({ store, openExternal, fetch: request = globalThis.fetch, onChange = () => {}, timeoutMs = 180000 }) {
+	let state;
+	let init;
+	let queue = Promise.resolve();
+	let epoch = 0;
+	let revision = 0;
+	let authorityRevision = 0;
+	let catalogSequence = 0;
+	let storageDenied = false;
+	let secureAvailable = true;
+	const stamp = () => ({ epoch, authorityRevision, revision });
+	const backgroundIntent = new Map();
+	let pending;
+	let catalog = [];
+	let catalogError = null;
+	let lastError = null;
+	let closed = false;
+	let paused = false;
+	let suspended = false;
+	let retryRegistration;
+	let transportAbort = new AbortController();
+	let backgroundAbort = new AbortController();
+	const serial = (fn) => {
+		const next = queue.then(fn, fn);
+		queue = next.catch(() => {});
+		return next;
+	};
+	const active = () => !storageDenied && !suspended && !closed && !paused ? state?.registrations.find((r) => r.id === state.activeId && r.tokens && !r.reconnectRequired) : undefined;
+	const backgroundAllowed = (account) => account.background === true && backgroundIntent.get(account.id)?.enabled !== false;
+	const notify = (invalidate = "none") => { revision++; onChange({ channel: "lh:chatgpt", kind: "state", stamp: stamp(), status: snapshot(), invalidate }); };
+	function stop(backgroundOnly = false) {
+		if (backgroundOnly) {
+			backgroundAbort.abort();
+			backgroundAbort = new AbortController();
+			notify("background");
+			return;
+		}
+		suspended = true;
+		epoch++;
+		authorityRevision++;
+		catalogSequence++;
+		transportAbort.abort();
+		transportAbort = new AbortController();
+		catalog = [];
+		catalogError = null;
+		notify("all");
+	}
+	async function ready() {
+		if (closed) throw fail("disconnected");
+		if (!init) init = (async () => {
+			const loaded = await store.read();
+			state = loaded ? validateState(loaded) : null;
+			if (!state) {
+				const first = { version: 1, hostId: `urn:uuid:${crypto.randomUUID()}`, activeId: null, registrations: [] };
+				await store.write(first);
+				state = first;
+			}
+		})();
+		try { await init; checkSecure(); } catch (err) { denyStorage(err); throw err; }
+	}
+	function denyStorage(err) {
+		secureAvailable = false;
+		if (!storageDenied) { storageDenied = true; lastError = publicError(err); stop(); }
+	}
+	function checkSecure() {
+		try { store.secure(); secureAvailable = true; } catch (err) { denyStorage(err); throw err; }
+	}
+	async function save(next) {
+		try { checkSecure(); validateState(next); await store.write(next); checkSecure(); } catch (err) {
+			// A rotated token may already have invalidated its predecessor. Fail closed.
+			denyStorage(err);
+			lastError = publicError(err);
+			notify();
+			throw err;
+		}
+		state = next;
+	}
+	function snapshot() {
+		return {
+			available: !!state && secureAvailable && !closed && !paused,
+			pending: !!pending,
+			activeId: state?.activeId ?? null,
+			accounts: (state?.registrations || []).map((r) => ({
+				id: r.id,
+				email: r.email || "ChatGPT account",
+				registration: r.clientId,
+				connected: !storageDenied && !closed && !paused && !suspended && !r.reconnectRequired && planScopes(r.tokens),
+				planEnabled: !storageDenied && !r.reconnectRequired && planScopes(r.tokens),
+				background: backgroundAllowed(r),
+			})),
+			error: lastError,
+			catalogError,
+		};
+	}
+	async function status() {
+		try { await ready(); } catch (err) { return { ...snapshot(), available: false, error: publicError(err) }; }
+		return snapshot();
+	}
+	async function json(url, options = {}) {
+		try {
+			const response = await request(url, { ...options, redirect: "error", signal: options.signal ?? AbortSignal.timeout(15000) });
+			if (!response.ok) {
+				const body = await response.json().catch(() => ({}));
+				const oauthCode = typeof body.error === "string" ? body.error : body.error?.code;
+				throw Object.assign(fail(oauthCode === "invalid_grant" || response.status === 401 || response.status === 403 ? "expired" : "network"), { oauthCode });
+			}
+			return await response.json();
+		} catch (err) { throw err.code ? err : fail("network"); }
+	}
+	async function identity(idToken, clientId, nonce, expectedSubject) {
+		try {
+			const { jwtVerify, createLocalJWKSet } = await import("jose");
+			// Fixed, documented JWKS origin; never follow a discovery-supplied URL.
+			const keys = await json(JWKS);
+			const { payload } = await jwtVerify(idToken, createLocalJWKSet(keys), {
+				issuer: ISSUER, audience: clientId, algorithms: ["RS256"],
+				requiredClaims: ["sub", "exp", "iat"], clockTolerance: 5,
+			});
+			if (typeof payload.sub !== "string" || !payload.sub || (nonce !== undefined && payload.nonce !== nonce) || (expectedSubject && payload.sub !== expectedSubject)) throw fail("identity");
+			if ((payload.azp !== undefined && payload.azp !== clientId) || (Array.isArray(payload.aud) && payload.aud.length > 1 && payload.azp !== clientId)) throw fail("identity");
+			return { subject: payload.sub, email: typeof payload.email === "string" ? payload.email : undefined };
+		} catch { throw fail("identity"); }
+	}
+	function tokenSet(body, previous) {
+		const scopes = typeof body.scope === "string" ? body.scope.split(/\s+/) : previous?.scopes;
+		if (!scopes || SCOPES.some((scope) => !scopes.includes(scope))) {
+			// Lack of entitlement is not lack of revocable material.
+			return { idToken: body.id_token ?? previous?.idToken, scopes: scopes ?? [], expiresAt: 0,
+				...(nonempty(body.refresh_token) ? { refreshToken: body.refresh_token } : previous?.refreshToken ? { refreshToken: previous.refreshToken } : {}) };
+		}
+		if (typeof body.access_token !== "string" || !body.access_token || body.access_token.startsWith("sk-") || body.token_type?.toLowerCase() !== "bearer" || !Number.isFinite(body.expires_in) || body.expires_in <= 0) throw fail("identity");
+		const refreshToken = body.refresh_token ?? previous?.refreshToken;
+		if (!nonempty(refreshToken)) throw fail("identity");
+		return { accessToken: body.access_token, refreshToken, idToken: body.id_token ?? previous?.idToken, scopes, expiresAt: Date.now() + body.expires_in * 1000 };
+	}
+	async function grant(fields, signal) {
+		return json(TOKEN, { method: "POST", headers: { "content-type": "application/x-www-form-urlencoded", accept: "application/json" }, body: new URLSearchParams({ ...fields, resource: RESOURCE }), signal });
+	}
+
+	async function login(accountId) {
+		await ready();
+		if (pending) throw fail("pending");
+		if (paused) throw fail("disconnected");
+		const selected = accountId ? state.registrations.find((r) => r.id === accountId) : retryRegistration;
+		if (accountId && !selected) throw fail("identity");
+		const attemptEpoch = epoch;
+		const verifier = random();
+		const nonce = random();
+		const expectedState = random();
+		const abort = new AbortController();
+		let finish;
+		const completion = new Promise((resolve) => { finish = resolve; });
+		const listener = http.createServer();
+		let timer;
+		let consumed = false;
+		const cleanup = () => { clearTimeout(timer); listener.close(); listener.closeAllConnections(); };
+		const cancel = (code = "cancelled") => {
+			abort.abort();
+			cleanup();
+			if (pending?.abort === abort) pending = null;
+			lastError = publicError(fail(code));
+			finish();
+			notify();
+		};
+		pending = { abort, cancel };
+		lastError = null;
+		notify();
+		try {
+			await new Promise((resolve, reject) => {
+				listener.once("error", reject);
+				listener.listen(0, "127.0.0.1", resolve);
+			});
+			const redirectUri = `http://127.0.0.1:${listener.address().port}/auth/callback`;
+			listener.on("request", async (req, res) => {
+				let callback;
+				try { callback = new URL(req.url || "/", redirectUri); }
+				catch { res.writeHead(400).end(); return; }
+				res.setHeader("Cache-Control", "no-store");
+				res.setHeader("Content-Security-Policy", "default-src 'none'");
+				res.setHeader("Referrer-Policy", "no-referrer");
+				if (req.method !== "GET" || callback.origin !== new URL(redirectUri).origin || callback.pathname !== "/auth/callback" || req.headers.host !== new URL(redirectUri).host || consumed) { res.writeHead(404).end(); return; }
+				const params = callback.searchParams;
+				// Reject duplicate parameters and unsolicited callbacks without consuming a valid pending attempt.
+				if (["state", "code", "client_id", "error"].some((key) => params.getAll(key).length > 1) || params.get("state") !== expectedState) { res.writeHead(400).end("Sign-in could not be verified."); return; }
+				consumed = true;
+				res.writeHead(200, { "content-type": "text/plain; charset=utf-8" }).end("Return to LearnHarness to view the sign-in result. You may close this tab.");
+				cleanup();
+				try {
+					if (params.has("error")) throw fail(params.get("error") === "access_denied" ? "cancelled" : "identity");
+					const clientId = params.get("client_id") || selected?.clientId;
+					if (!clientId || clientId === "dynamic_agent_client" || clientId.length > 512 || /[\s\x00-\x1f]/.test(clientId) || (selected && clientId !== selected.clientId) || !params.get("code")) throw fail("identity");
+					let body;
+					try { body = await grant({ grant_type: "authorization_code", client_id: clientId, code: params.get("code"), code_verifier: verifier, redirect_uri: redirectUri }, AbortSignal.any([abort.signal, AbortSignal.timeout(15000)])); }
+					catch (err) {
+						if (err.oauthCode === "invalid_grant" && !selected?.subject) retryRegistration = { clientId };
+						throw err;
+					}
+					if (typeof body.id_token !== "string") throw fail("identity");
+					const verified = await identity(body.id_token, clientId, nonce, selected?.subject);
+					const tokens = tokenSet(body);
+					await serial(async () => {
+						if (closed || abort.signal.aborted || epoch !== attemptEpoch) throw fail("cancelled");
+						const existing = state.registrations.find((r) => r.clientId === clientId);
+						if (existing && existing.subject !== verified.subject) throw fail("identity");
+						const record = { id: existing?.id ?? crypto.randomUUID(), clientId, issuer: ISSUER, ...verified, background: existing?.background === true, tokens };
+						const next = structuredClone(state);
+						next.registrations = next.registrations.filter((r) => r.id !== record.id).concat(record);
+						next.activeId = record.id;
+						const previous = structuredClone(state);
+						await save(next); // Active identity changes only AFTER validation and durable encrypted persistence.
+						if (closed || abort.signal.aborted || epoch !== attemptEpoch) {
+							await save(previous);
+							throw fail("cancelled");
+						}
+						stop();
+						storageDenied = false;
+						suspended = false;
+						retryRegistration = undefined;
+					});
+					lastError = planScopes(tokens) ? null : publicError(fail("scope"));
+				} catch (err) { if (!abort.signal.aborted) lastError = publicError(err); }
+				finally { if (pending?.abort === abort) pending = null; finish(); notify(); }
+			});
+			timer = setTimeout(() => cancel("timeout"), timeoutMs);
+			const url = new URL(AUTHORIZE);
+			url.search = new URLSearchParams({
+				client_id: selected?.clientId ?? "dynamic_agent_client",
+				...(selected ? {} : { agent_name_hint: "LearnHarness" }),
+				ext_agent_host_id: state.hostId,
+				...(selected?.tokens?.idToken ? { id_token_hint: selected.tokens.idToken } : {}),
+				...(selected?.email ? { login_hint: selected.email } : {}),
+				response_type: "code", redirect_uri: redirectUri, scope: SCOPES.join(" "), resource: RESOURCE,
+				state: expectedState, nonce, code_challenge_method: "S256",
+				code_challenge: crypto.createHash("sha256").update(verifier).digest("base64url"),
+			}).toString();
+			await openExternal(url.toString()); // Trusted main only. Never return this URL to the renderer.
+		} catch (err) { cancel(err.code || "network"); }
+		void completion;
+		return snapshot();
+	}
+	function cancelLogin() { pending?.cancel(); return snapshot(); }
+
+	function authorized(background, expected, capturedEpoch) {
+		checkSecure();
+		const account = active();
+		if (!account || (capturedEpoch !== undefined && capturedEpoch !== epoch)) throw fail("disconnected");
+		if (!planScopes(account.tokens)) throw fail("scope");
+		if (expected && (expected.expectedAccountId !== account.id || expected.expectedAuthorityRevision !== authorityRevision || !catalog.some((m) => m.id === expected.selectedModelId))) throw fail("catalog");
+		if (background && !backgroundAllowed(account)) throw fail("background");
+		return account;
+	}
+	async function credential(background, expected) {
+		await ready();
+		return serial(async () => {
+			const requestEpoch = epoch;
+			const account = authorized(background, expected);
+			if (account.tokens.expiresAt <= Date.now() + 60000) {
+				const refreshEpoch = epoch;
+				const signal = AbortSignal.any([transportAbort.signal, ...(background ? [backgroundAbort.signal] : []), AbortSignal.timeout(15000)]);
+				let body;
+				try { body = await grant({ grant_type: "refresh_token", client_id: account.clientId, refresh_token: account.tokens.refreshToken }, signal); }
+				catch (err) {
+					if (err.oauthCode === "invalid_grant") {
+						const next = structuredClone(state);
+						delete next.registrations.find((r) => r.id === account.id).tokens;
+						next.activeId = null;
+						await save(next);
+						stop();
+						notify();
+					}
+					throw err;
+				}
+				let tokens;
+				try {
+					if (body.id_token !== undefined) await identity(body.id_token, account.clientId, undefined, account.subject);
+					tokens = tokenSet(body, account.tokens);
+				} catch (err) {
+					// The fixed token endpoint may have rotated already. Preserve only
+					// verified identity and a single likely-latest revocation handle.
+					stop();
+					const restricted = structuredClone(state);
+					const record = restricted.registrations.find((r) => r.id === account.id);
+					record.reconnectRequired = true;
+					if (nonempty(body.refresh_token)) record.tokens.refreshToken = body.refresh_token;
+					await save(restricted);
+					lastError = publicError(fail("identity")); notify();
+					throw err;
+				}
+				// A successful remote rotation must be durably retained even when
+				// logout, account selection, consent revocation, or pause denies use.
+				// Those serialized operations will revoke/clear the LATEST token.
+				const next = structuredClone(state);
+				next.registrations.find((r) => r.id === account.id).tokens = tokens;
+				await save(next);
+				if (closed || refreshEpoch !== epoch || signal.aborted) throw fail("disconnected");
+			}
+			const current = authorized(background, expected, requestEpoch);
+			return { accessToken: current.tokens.accessToken, accountId: current.id, ...stamp() };
+		});
+	}
+	async function lease(background, expected) {
+		if (typeof background !== "boolean" || !expected || !nonempty(expected.expectedAccountId) || !Number.isSafeInteger(expected.expectedAuthorityRevision) || !nonempty(expected.selectedModelId)) throw fail("catalog");
+		return credential(background, expected);
+	}
+	const modelValue = () => ({ ...stamp(), status: snapshot(), accountId: active()?.id ?? null, models: catalog });
+	async function models() {
+		const sequence = ++catalogSequence;
+		try {
+			await ready();
+			if (!active()) return modelValue();
+			// Discovery is identity/scope authorization, not an invented model lease.
+			const token = await credential(false);
+			const body = await json(`${RESOURCE}/models`, { headers: { Authorization: `Bearer ${token.accessToken}` }, signal: AbortSignal.any([transportAbort.signal, AbortSignal.timeout(15000)]) });
+			authorized(false, undefined, token.epoch);
+			if (sequence !== catalogSequence || active()?.id !== token.accountId || !Array.isArray(body.models)) throw fail("catalog");
+			const visible = body.models.filter((m) => m && m.visibility === "list" && nonempty(m.slug) && typeof m.display_name === "string").map((m) => ({ id: m.slug, name: m.display_name }));
+			const changed = JSON.stringify(visible) !== JSON.stringify(catalog);
+			catalog = visible;
+			if (changed) authorityRevision++;
+			const hadError = catalogError !== null; catalogError = null;
+			if (changed || hadError) notify();
+			return modelValue();
+		} catch {
+			if (sequence === catalogSequence) {
+				if (catalog.length) { catalog = []; authorityRevision++; }
+				catalogError = publicError(fail("catalog")); notify();
+			}
+			return modelValue();
+		}
+	}
+	async function select(id) {
+		pending?.cancel();
+		stop();
+		const selectionIntent = epoch;
+		await ready();
+		return serial(async () => {
+			const account = state.registrations.find((r) => r.id === id);
+			if (!account?.tokens) throw fail("disconnected");
+			await save({ ...state, activeId: id });
+			if (epoch !== selectionIntent || closed || paused || storageDenied) throw fail("disconnected");
+			suspended = false;
+			lastError = null;
+			notify();
+			return snapshot();
+		});
+	}
+	async function consent(id, enabled) {
+		if (typeof enabled !== "boolean") throw fail("identity");
+		const intent = { enabled: false, version: (backgroundIntent.get(id)?.version ?? 0) + 1 };
+		backgroundIntent.set(id, intent);
+		if (!enabled) {
+			const account = state?.registrations.find((r) => r.id === id);
+			if (account) account.background = false;
+			stop(true); // Revoke background calls immediately; foreground stays usable.
+		}
+		await ready();
+		return serial(async () => {
+			const next = structuredClone(state);
+			const account = next.registrations.find((r) => r.id === id);
+			if (!account) throw fail("identity");
+			account.background = enabled;
+			await save(next);
+			if (backgroundIntent.get(id) === intent && enabled) backgroundIntent.delete(id);
+			notify();
+			return snapshot();
+		});
+	}
+	async function logout(id) {
+		pending?.cancel();
+		stop();
+		await ready();
+		return serial(async () => {
+			const account = state.registrations.find((r) => r.id === id);
+			if (!account) throw fail("identity");
+			let revoked = false;
+			try {
+				if (account.tokens?.refreshToken) {
+					const config = await json(`${ISSUER}/.well-known/openid-configuration`);
+					if (config.issuer !== ISSUER || !allowedAuthEndpoint(config.revocation_endpoint)) throw fail("identity");
+					for (let attempt = 0; attempt < 2; attempt++) {
+						if (attempt) await new Promise((resolve) => setTimeout(resolve, 200));
+						try {
+							const response = await request(config.revocation_endpoint, { method: "POST", redirect: "error", signal: AbortSignal.timeout(5000), headers: { "content-type": "application/x-www-form-urlencoded" }, body: new URLSearchParams({ token: account.tokens.refreshToken, token_type_hint: "refresh_token", client_id: account.clientId }) });
+							revoked = response.status === 200;
+							if (revoked || response.status < 500) break;
+						} catch { /* retry only while the refresh token is still in memory */ }
+					}
+				}
+			} catch { /* local sign-out is mandatory even when revocation is unavailable */ }
+			const next = structuredClone(state);
+			delete next.registrations.find((r) => r.id === id).tokens;
+			if (next.activeId === id) next.activeId = null;
+			const logoutEpoch = epoch;
+			await save(next);
+			if (!closed && epoch === logoutEpoch) suspended = false;
+			lastError = revoked ? null : "Signed out locally. Remote revocation was not confirmed; disconnect LearnHarness in ChatGPT Settings.";
+			notify();
+			return snapshot();
+		});
+	}
+	function close() { closed = true; pending?.cancel(); stop(); }
+	function pause() { paused = true; pending?.cancel(); stop(); notify(); }
+	function resume() { paused = false; if (!storageDenied) suspended = false; notify(); }
+	return { status, login, cancelLogin, lease, models, select, consent, logout, close, pause, resume, stamp };
+}
+
+module.exports = { createChatGPT, encryptedStore, publicError, allowedAuthEndpoint, SCOPES, validateState };

electron/desktop-api.cjs

@@ -0,0 +1,25 @@
+"use strict";
+
+const HEADER = "x-learnharness-desktop";
+
+/** Called for EVERY outgoing request, so redirects cannot carry the secret
+ * to another origin. Never expose this function or its secret to preload. */
+function desktopAPIHeaders(details, origin, secret, windows) {
+	const headers = { ...details.requestHeaders };
+	for (const name of Object.keys(headers)) if (name.toLowerCase() === HEADER) delete headers[name];
+	if (!secret) return headers;
+	try {
+		const target = new URL(details.url);
+		if (target.origin !== origin || target.username || target.password || !(target.pathname === "/api" || target.pathname.startsWith("/api/"))) return headers;
+		const win = windows.find((candidate) => !candidate.isDestroyed() && candidate.webContents.id === details.webContentsId);
+		if (!win || win.webContents.isDestroyed()) return headers;
+		const contents = win.webContents;
+		if (details.webContents && details.webContents !== contents) return headers;
+		if (!details.frame || details.frame !== contents.mainFrame) return headers;
+		if (new URL(contents.mainFrame.url).origin !== origin) return headers;
+		if (details.initiatorOrigin !== undefined && details.initiatorOrigin !== origin) return headers;
+		headers[HEADER] = secret;
+	} catch { /* Missing/destroyed/navigated frames fail closed. */ }
+	return headers;
+}
+module.exports = { desktopAPIHeaders, HEADER };

electron/main.cjs

@@ -24,7 +24,11 @@
  *     would shadow every settings-page key change until restart.
  */
 
-const { app, BrowserWindow, Menu, ipcMain, session, shell, dialog } = require("electron");
+const { app, BrowserWindow, Menu, ipcMain, session, shell, dialog, safeStorage } = require("electron");
+const { createChatGPT, encryptedStore } = require("./chatgpt.cjs");
+const { attachChatGPTChild, trustedSender, chatGPTControl } = require("./chatgpt-ipc.cjs");
+const { desktopAPIHeaders } = require("./desktop-api.cjs");
+const { randomBytes } = require("node:crypto");
 const { spawn } = require("node:child_process");
 const fs = require("node:fs");
 const http = require("node:http");
@@ -128,6 +132,11 @@ if (process.argv.includes("--browser-service")) {
 }
 
 let serverProcess = null;
+let chatgpt = null;
+let chatgptChild = null;
+// Per-launch request capability, never OAuth tokens and never renderer state.
+let desktopAPISecret = null;
+const trustedAppWindows = new Set();
 
 // A GUI shell must never die to a stray exception (EPIPE dialogs, a race
 // in an event handler): the window + child server are the product, and
@@ -201,17 +210,23 @@ function startServer(port) {
 					// (no env injection anymore: a spawn-time copy would shadow every
 					// settings-page key change until restart; env wins over file).
 					LH_AUTH_FILE: AUTH_FILE(),
+					LH_DESKTOP_SERVER: "1",
+					LH_DESKTOP_API_SECRET: desktopAPISecret,
 				},
-				stdio: ["ignore", "pipe", "pipe"],
+				stdio: ["ignore", "pipe", "pipe", "ipc"],
 			},
 		);
 		serverProcess = child;
+		chatgptChild = attachChatGPTChild(child, chatgpt);
 		const fwd = (buf) => buf.toString().trimEnd().split("\n").forEach((l) => l && log(`server: ${l}`));
 		child.stdout.on("data", fwd);
 		child.stderr.on("data", fwd);
 		child.on("exit", (code, signal) => {
 			log(`server exited (code=${code} signal=${signal})`);
 			serverProcess = null;
+			chatgptChild?.dispose();
+			chatgptChild = null;
+			chatgpt?.close();
 		});
 		resolve(child);
 	});
@@ -235,6 +250,7 @@ function waitUntilUp(url, tries = 150) {
 }
 
 function killServer() {
+	chatgpt?.close();
 	if (!serverProcess) return;
 	const child = serverProcess;
 	serverProcess = null;
@@ -306,6 +322,8 @@ function createWindow(url) {
 		minHeight: 600,
 	});
 
+	trustedAppWindows.add(win);
+	win.once("closed", () => trustedAppWindows.delete(win));
 	const allowedOrigin = () => new URL(url).origin;
 	// Deny-by-default; the ONLY grants are media (mic/STT groundwork) and
 	// clipboard-sanitized-write — the copy affordances in the transcript
@@ -546,6 +564,7 @@ function installMenu() {
 
 async function main() {
 	app.on("window-all-closed", () => {
+		chatgpt?.pause();
 		// macOS convention: the app stays resident, reactivated from the Dock.
 		if (process.platform !== "darwin") app.quit();
 	});
@@ -563,6 +582,20 @@ async function main() {
 	const devUrl = process.env.LH_DEV_URL;
 	const url = devUrl ?? `http://127.0.0.1:${await freePort()}`;
 	if (!devUrl) {
+		desktopAPISecret = randomBytes(32).toString("hex");
+		const ownedOrigin = new URL(url).origin;
+		// All URLs, not only /api: remove the header again on external redirects.
+		session.defaultSession.webRequest.onBeforeSendHeaders({ urls: ["<all_urls>"] }, (details, callback) => {
+			callback({ requestHeaders: desktopAPIHeaders(details, ownedOrigin, desktopAPISecret, [...trustedAppWindows]) });
+		});
+		chatgpt = createChatGPT({
+			store: encryptedStore(path.join(app.getPath("userData"), "chatgpt-credentials.enc.json"), safeStorage),
+			openExternal: (target) => shell.openExternal(target),
+			onChange: (value) => {
+				chatgptChild?.send(value);
+				for (const win of BrowserWindow.getAllWindows()) win.webContents.send("lh:chatgpt:changed");
+			},
+		});
 		log(`starting server: ${serverEntry()}`);
 		log(`data dir: ${DATA_DIR()}${fs.existsSync(AUTH_FILE()) ? " · auth.json: present" : " · auth.json: ABSENT (no BYOK keys yet)"}`);
 		await startServer(new URL(url).port);
@@ -572,6 +605,14 @@ async function main() {
 		log(`dev mode: ${devUrl}`);
 	}
 
+	// Renderer controls contain no token/URL/code operation. The separate server
+	// child is the only recipient of credential leases over inherited IPC.
+	ipcMain.handle("lh:chatgpt:control", async (event, command) => {
+		if (!trustedAppWindows.has(BrowserWindow.fromWebContents(event.sender)) || !trustedSender(event, new URL(url).origin, BrowserWindow)) return { error: "Untrusted desktop request." };
+		if (!chatgpt || !serverProcess?.connected) return { available: false, pending: false, activeId: null, accounts: [], error: "ChatGPT sign-in requires the desktop app's private server connection. Browser-only and externally launched development servers keep API-key connections." };
+		return chatGPTControl(command, chatgpt);
+	});
+
 	installMenu();
 
 	// Auto-update (Wave C): wire the updater before the window so the
@@ -597,7 +638,7 @@ async function main() {
 		if (width < target) win.setBounds({ x, y, width: target, height });
 	});
 	app.on("activate", () => {
-		if (BrowserWindow.getAllWindows().length === 0) createWindow(url);
+		if (BrowserWindow.getAllWindows().length === 0) { chatgpt?.resume(); createWindow(url); }
 	});
 	createWindow(url);
 	log(`window open: ${url}`);

electron/preload.cjs

@@ -29,6 +29,19 @@ const { contextBridge, ipcRenderer } = require("electron");
 
 contextBridge.exposeInMainWorld("lh", {
 	platform: process.platform,
+	chatgpt: {
+		status: () => ipcRenderer.invoke("lh:chatgpt:control", { op: "status" }),
+		login: (id) => ipcRenderer.invoke("lh:chatgpt:control", { op: "login", id }),
+		cancel: () => ipcRenderer.invoke("lh:chatgpt:control", { op: "cancel" }),
+		select: (id) => ipcRenderer.invoke("lh:chatgpt:control", { op: "select", id }),
+		logout: (id) => ipcRenderer.invoke("lh:chatgpt:control", { op: "logout", id }),
+		consent: (id, enabled) => ipcRenderer.invoke("lh:chatgpt:control", { op: "consent", id, enabled }),
+		onChange: (cb) => {
+			const listener = () => cb();
+			ipcRenderer.on("lh:chatgpt:changed", listener);
+			return () => ipcRenderer.removeListener("lh:chatgpt:changed", listener);
+		},
+	},
 	expandWindow: () => ipcRenderer.send("lh:expand-window"),
 	onZoom: (cb) => {
 		const listener = (_event, factor) => cb(factor);

package-lock.json

@@ -31,6 +31,7 @@
 				"dagre-d3-es": "7.0.14",
 				"electron-updater": "^6.8.9",
 				"embla-carousel-react": "^8.6.0",
+				"jose": "^6.2.12",
 				"katex": "0.18.2",
 				"linkedom": "^0.18.13",
 				"lucide-react": "^1.33.0",
@@ -11655,6 +11656,13 @@
 				"jiti": "lib/jiti-cli.mjs"
 			}
 		},
+		"node_modules/jose": {
+			"version": "6.2.12",
+			"resolved": "https://registry.npmjs.org/jose/-/jose-6.2.12.tgz",
+			"integrity": "sha512-9NiFmJEex0sy2Dk58j2UGBSHgUs2ypF9eZSu4L6vjOX3Dp96Sw1F3uL+H+D1sx02jZZdzUT0HgvCy59CuvXcWw==",
+			"license": "MIT",
+			"funding": { "url": "https://github.com/sponsors/panva" }
+		},
 		"node_modules/js-yaml": {
 			"version": "4.3.2",
 			"resolved": "https://registry.npmjs.org/js-yaml/-/js-yaml-4.3.2.tgz",

package.json

@@ -46,6 +46,7 @@
 		"dagre-d3-es": "7.0.14",
 		"electron-updater": "^6.8.9",
 		"embla-carousel-react": "^8.6.0",
+		"jose": "^6.2.12",
 		"katex": "0.18.2",
 		"linkedom": "^0.18.13",
 		"lucide-react": "^1.33.0",

proxy.ts

@@ -0,0 +1,17 @@
+import { NextResponse, type NextRequest } from "next/server";
+import { desktopAPIAuthorized } from "./src/server/desktop-access.ts";
+
+/** Applies to every API method, including SSE and spending/settings routes.
+ * No public callback or worker bypass: internal jobs call server functions,
+ * and any legitimate self-HTTP request must carry the private capability. */
+export function proxy(request: NextRequest) {
+	if (!desktopAPIAuthorized(request.headers)) {
+		return NextResponse.json(
+			{ error: "Desktop API access requires the trusted LearnHarness window." },
+			{ status: 403, headers: { "Cache-Control": "no-store" } },
+		);
+	}
+	return NextResponse.next();
+}
+
+export const config = { matcher: "/api/:path*" };

public/siwc/chatgpt-logo-white.svg

@@ -0,0 +1,5 @@
+<svg preserveAspectRatio="none" overflow="visible" style="display: block;" width="21" height="21" viewBox="0 0 21 21" fill="none" xmlns="http://www.w3.org/2000/svg">
+<g id="ChatGPT logo &#226;&#128;&#148; original vector">
+<path id="Vector" d="M11.8102 19.1625C11.2327 19.1625 10.6841 19.0528 10.1643 18.8334C9.64459 18.6139 9.18259 18.3078 8.77834 17.9151C8.33944 18.0653 7.88322 18.1404 7.40967 18.1404C6.63582 18.1404 5.91972 17.9498 5.26137 17.5686C4.60302 17.1875 4.07172 16.6677 3.66747 16.0094C3.27477 15.351 3.07842 14.6176 3.07842 13.8091C3.07842 13.4742 3.12462 13.1103 3.21702 12.7176C2.75502 12.2903 2.39697 11.7994 2.14287 11.245C1.88877 10.679 1.76172 10.09 1.76172 9.47785C1.76172 8.85415 1.89454 8.25355 2.16019 7.67605C2.42584 7.09855 2.79544 6.6019 3.26899 6.1861C3.75409 5.75875 4.31427 5.46422 4.94952 5.30252C5.07657 4.64417 5.34222 4.05512 5.74647 3.53537C6.16227 3.00407 6.67047 2.58827 7.27107 2.28797C7.87167 1.98767 8.51269 1.83752 9.19414 1.83752C9.77164 1.83752 10.3203 1.94725 10.84 2.1667C11.3598 2.38615 11.8218 2.69222 12.226 3.08492C12.6649 2.93477 13.1211 2.8597 13.5947 2.8597C14.3685 2.8597 15.0846 3.05027 15.743 3.43142C16.4013 3.81257 16.9269 4.33232 17.3196 4.99067C17.7238 5.64902 17.9259 6.38245 17.9259 7.19095C17.9259 7.5259 17.8797 7.88973 17.7873 8.28242C18.2493 8.70977 18.6074 9.20643 18.8615 9.77238C19.1156 10.3268 19.2426 10.91 19.2426 11.5222C19.2426 12.1459 19.1098 12.7465 18.8442 13.324C18.5785 13.9015 18.2031 14.4039 17.718 14.8313C17.2445 15.2471 16.6901 15.5358 16.0548 15.6975C15.9278 16.3559 15.6564 16.9449 15.2406 17.4647C14.8363 17.996 14.3339 18.4118 13.7333 18.7121C13.1327 19.0124 12.4917 19.1625 11.8102 19.1625ZM7.53094 16.9969C8.10844 16.9969 8.61087 16.8756 9.03822 16.6331L12.2953 14.762C12.4108 14.6811 12.4686 14.5714 12.4686 14.4328V12.9429L8.27592 15.351C8.02182 15.5012 7.76772 15.5012 7.51362 15.351L4.23919 13.4626C4.23919 13.4973 4.23342 13.5377 4.22187 13.5839C4.22187 13.6301 4.22187 13.6994 4.22187 13.7918C4.22187 14.3808 4.36047 14.9237 4.63767 15.4203C4.92642 15.9054 5.32489 16.2866 5.83309 16.5638C6.34129 16.8525 6.90724 16.9969 7.53094 16.9969ZM7.70419 14.1729C7.77349 14.2076 7.83702 14.2249 7.89477 14.2249C7.95252 14.2249 8.01027 14.2076 8.06802 14.1729L9.36739 13.4279L5.19207 11.0024C4.93797 10.8523 4.81092 10.6271 4.81092 10.3268V6.56725C4.23342 6.82135 3.77142 7.21405 3.42492 7.74535C3.07842 8.2651 2.90517 8.8426 2.90517 9.47785C2.90517 10.0438 3.04954 10.5866 3.33829 11.1064C3.62704 11.6261 4.00242 12.0189 4.46442 12.2845L7.70419 14.1729ZM11.8102 18.0191C12.4224 18.0191 12.9768 17.8805 13.4734 17.6033C13.9701 17.3261 14.3628 16.9449 14.6515 16.4598C14.9403 15.9747 15.0846 15.4319 15.0846 14.8313V11.0891C15.0846 10.9505 15.0269 10.8465 14.9114 10.7772L13.5947 10.0149V14.8486C13.5947 15.1489 13.4676 15.3741 13.2135 15.5243L9.93912 17.4127C10.5051 17.817 11.1288 18.0191 11.8102 18.0191ZM12.4686 11.6781V9.32192L10.5108 8.21312L8.53579 9.32192V11.6781L10.5108 12.7869L12.4686 11.6781ZM7.40967 6.15145C7.40967 5.85115 7.53672 5.62592 7.79082 5.47577L11.0652 3.58735C10.4993 3.1831 9.87559 2.98097 9.19414 2.98097C8.58199 2.98097 8.02759 3.11957 7.53094 3.39677C7.03429 3.67397 6.64159 4.05512 6.35284 4.54022C6.07564 5.02532 5.93704 5.56817 5.93704 6.16878V9.89365C5.93704 10.0323 5.99479 10.142 6.11029 10.2228L7.40967 10.9851V6.15145ZM16.2108 14.4328C16.7883 14.1787 17.2445 13.786 17.5794 13.2547C17.9259 12.7234 18.0992 12.1459 18.0992 11.5222C18.0992 10.9562 17.9548 10.4134 17.6661 9.89365C17.3773 9.3739 17.0019 8.9812 16.5399 8.71555L13.3002 6.84445C13.2309 6.79825 13.1673 6.78092 13.1096 6.79248C13.0518 6.79248 12.9941 6.8098 12.9363 6.84445L11.637 7.5721L15.8296 10.0149C15.9567 10.0842 16.0491 10.1766 16.1068 10.2921C16.1761 10.3961 16.2108 10.5231 16.2108 10.6733V14.4328ZM12.7284 5.6317C12.9825 5.47 13.2366 5.47 13.4907 5.6317L16.7825 7.55477C16.7825 7.47393 16.7825 7.36997 16.7825 7.24292C16.7825 6.68852 16.6439 6.163 16.3667 5.66635C16.101 5.15815 15.7141 4.7539 15.2059 4.4536C14.7093 4.1533 14.1318 4.00315 13.4734 4.00315C12.8959 4.00315 12.3935 4.12442 11.9661 4.36697L8.70904 6.23808C8.59354 6.31892 8.53579 6.42865 8.53579 6.56725V8.0572L12.7284 5.6317Z" fill="white"/>
+</g>
+</svg>

src/agent/chatgpt-bridge.ts

@@ -0,0 +1,122 @@
+import type { Model } from "@earendil-works/pi-ai";
+import type { ChatGPTStatus } from "../core/contracts/chatgpt.ts";
+import type { AdmissionScope, NoDispatchKind } from "./chatgpt-request-cycle.ts";
+export type { ChatGPTStatus } from "../core/contracts/chatgpt.ts";
+export interface Stamp { epoch: number; authorityRevision: number; revision: number }
+interface StatusValue extends Stamp { status: ChatGPTStatus }
+interface ModelsValue extends StatusValue { accountId: string | null; models: { id: string; name: string }[] }
+interface LeaseValue extends Stamp { accessToken: string; accountId: string }
+interface Pending { op: string; resolve: (v: unknown) => void; reject: (e: Error) => void; timer: ReturnType<typeof setTimeout> }
+interface BridgeState {
+ status: ChatGPTStatus; stamp: Stamp; models: Model<"openai-responses">[]; catalogAccount: string | null;
+ pending: Map<string, Pending>; controllers: Map<AbortController, boolean>; listeners: Set<() => void>; counter: number; catalogRequest: number;
+}
+const key = Symbol.for("lh.chatgpt.bridge");
+const unavailable = (): ChatGPTStatus => ({ available: false, pending: false, activeId: null, accounts: [], error: null });
+const validStamp = (s: Stamp) => s && [s.epoch, s.authorityRevision, s.revision].every((v) => Number.isSafeInteger(v) && v >= 0);
+function bridge(): BridgeState {
+ const global = globalThis as typeof globalThis & { [key]?: BridgeState };
+ if (!global[key]) {
+  const state: BridgeState = { status: unavailable(), stamp: { epoch: 0, authorityRevision: 0, revision: 0 }, models: [], catalogAccount: null, pending: new Map(), controllers: new Map(), listeners: new Set(), counter: 0, catalogRequest: 0 };
+  global[key] = state;
+  process.on("message", (raw: unknown) => {
+   const m = raw as { channel?: string; kind?: string; id?: string; op?: string; ok?: boolean; value?: unknown; stamp: Stamp; status: ChatGPTStatus; invalidate?: string; error?: string };
+   if (m?.channel !== "lh:chatgpt") return;
+   if (m.kind === "state" && validStamp(m.stamp) && m.stamp.revision >= state.stamp.revision) {
+    acceptStatus(state, { ...m.stamp, status: m.status });
+    if (m.invalidate === "all") { state.catalogRequest++; clearCatalog(state); }
+    if (m.invalidate === "all" || m.invalidate === "background") {
+     for (const [controller, background] of state.controllers) if (m.invalidate === "all" || background) controller.abort(m.invalidate === "all" ? "all" : "background");
+    }
+    for (const listener of state.listeners) listener();
+   }
+   if (m.id) {
+    const call = state.pending.get(m.id);
+    if (!call || call.op !== m.op) return;
+    clearTimeout(call.timer); state.pending.delete(m.id);
+    if (m.ok === true) call.resolve(m.value);
+    else call.reject(new Error(m.error ?? "ChatGPT connection failed."));
+   }
+  });
+  process.on("disconnect", () => {
+   state.status = unavailable(); clearCatalog(state); state.catalogRequest++;
+   for (const controller of state.controllers.keys()) controller.abort("all");
+   for (const call of state.pending.values()) { clearTimeout(call.timer); call.reject(new Error("The desktop connection closed.")); }
+   state.pending.clear();
+   for (const listener of state.listeners) listener();
+  });
+ }
+ return global[key];
+}
+function clearCatalog(state: BridgeState) { state.models = []; state.catalogAccount = null; }
+function acceptStatus(state: BridgeState, value: StatusValue) {
+ if (!validStamp(value) || value.revision < state.stamp.revision) return;
+ if (value.epoch !== state.stamp.epoch || value.authorityRevision !== state.stamp.authorityRevision || value.status.activeId !== state.status.activeId) clearCatalog(state);
+ state.status = value.status; state.stamp = { epoch: value.epoch, authorityRevision: value.authorityRevision, revision: value.revision };
+}
+/** Runtime-only inherited IPC. No HTTP credential route or renderer method. */
+export function chatGPTRequest<T>(op: "status" | "models" | "lease", fields: Record<string, unknown> = {}): Promise<T> {
+ const state = bridge();
+ if (!process.send || !process.connected) return Promise.reject(new Error("ChatGPT sign-in is available only in the desktop app with a private server connection."));
+ return new Promise<T>((resolve, reject) => {
+  const id = `${process.pid}:${++state.counter}`;
+  const timer = setTimeout(() => { state.pending.delete(id); reject(new Error("The ChatGPT desktop connection timed out.")); }, 45000);
+  state.pending.set(id, { op, resolve: (v) => resolve(v as T), reject, timer });
+  process.send!({ ...fields, channel: "lh:chatgpt", id, op }, (error) => {
+   if (error) { clearTimeout(timer); state.pending.delete(id); reject(new Error("The ChatGPT desktop connection closed.")); }
+  });
+ });
+}
+export function chatGPTStatus(): ChatGPTStatus { return bridge().status; }
+export function chatGPTConnected(): boolean { const s = chatGPTStatus(); return s.available && s.accounts.some((a) => a.id === s.activeId && a.connected && a.planEnabled === true); }
+export function chatGPTBackgroundAllowed(): boolean { const s = chatGPTStatus(); return chatGPTConnected() && s.accounts.some((a) => a.id === s.activeId && a.background); }
+export function onChatGPTChange(listener: () => void): () => void { bridge().listeners.add(listener); return () => { bridge().listeners.delete(listener); }; }
+export function chatGPTModels(): Model<"openai-responses">[] { return bridge().models; }
+export function chatGPTModel(id: string): Model<"openai-responses"> | undefined { return chatGPTModels().find((m) => m.id === id); }
+export function chatGPTScope(selectedModelId: string): AdmissionScope {
+ const state = bridge();
+ if (!chatGPTConnected() || !state.catalogAccount || state.catalogAccount !== state.status.activeId || !chatGPTModel(selectedModelId)) throw new Error("ChatGPT account model catalog is unavailable");
+ return { accountId: state.catalogAccount, epoch: state.stamp.epoch, authorityRevision: state.stamp.authorityRevision, selectedModelId };
+}
+export function chatGPTScopeCurrent(scope: AdmissionScope): boolean {
+ const s = bridge();
+ return !!process.connected && chatGPTConnected() && scope.accountId === s.catalogAccount && scope.accountId === s.status.activeId && scope.epoch === s.stamp.epoch && scope.authorityRevision === s.stamp.authorityRevision && !!chatGPTModel(scope.selectedModelId);
+}
+export async function chatGPTLease(scope: AdmissionScope, background: boolean): Promise<LeaseValue> {
+ if (!chatGPTScopeCurrent(scope)) throw new Error("ChatGPT catalog authority changed");
+ const value = await chatGPTRequest<LeaseValue>("lease", { background, expectedAccountId: scope.accountId, expectedAuthorityRevision: scope.authorityRevision, selectedModelId: scope.selectedModelId });
+ if (!validStamp(value) || !chatGPTScopeCurrent(scope) || value.epoch !== scope.epoch || value.authorityRevision !== scope.authorityRevision || value.accountId !== scope.accountId || typeof value.accessToken !== "string" || !value.accessToken || value.accessToken.startsWith("sk-")) throw new Error("ChatGPT lease is stale");
+ return value;
+}
+/** Safe status confirmation can establish consent-only denial, never parse error text. */
+export async function chatGPTDenial(scope: AdmissionScope | undefined, controller: AbortController, callerAborted: boolean, background: boolean): Promise<NoDispatchKind> {
+ if (callerAborted || controller.signal.reason === "all") return "all";
+ if (!scope || !chatGPTScopeCurrent(scope)) return "other";
+ if (controller.signal.reason === "background") return "background";
+ if (!background) return "other";
+ try { const value = await chatGPTRequest<StatusValue>("status"); acceptStatus(bridge(), value); }
+ catch { return "other"; }
+ if (controller.signal.reason === "all" || !chatGPTScopeCurrent(scope)) return "all";
+ return !chatGPTBackgroundAllowed() ? "background" : "other";
+}
+/** Preserve account order; unknown capabilities/windows are never fabricated. */
+export async function refreshChatGPTModels(supported: Model<"openai-responses">[]): Promise<void> {
+ const state = bridge();
+ if (!process.send || !process.connected) { clearCatalog(state); return; }
+ const sequence = ++state.catalogRequest;
+ try {
+  const value = await chatGPTRequest<ModelsValue>("models");
+  if (sequence !== state.catalogRequest || !validStamp(value)) return;
+  acceptStatus(state, value);
+  if (value.epoch !== state.stamp.epoch || value.authorityRevision !== state.stamp.authorityRevision || value.accountId !== state.status.activeId || !chatGPTConnected()) return;
+  state.catalogAccount = value.accountId;
+  state.models = value.models.flatMap((entry) => {
+   const known = supported.find((m) => m.id === entry.id && m.api === "openai-responses");
+   return known ? [{ ...known, provider: "chatgpt", name: entry.name, headers: undefined, samplingParams: undefined, cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 } }] : [];
+  });
+ } catch { if (sequence === state.catalogRequest) clearCatalog(state); }
+}
+export function chatGPTAbortController(background = true): { controller: AbortController; release: () => void } {
+ const state = bridge(); const controller = new AbortController(); state.controllers.set(controller, background);
+ return { controller, release: () => { state.controllers.delete(controller); } };
+}

src/agent/chatgpt-request-cycle.ts

@@ -0,0 +1,122 @@
+import type { AgentEvent, AgentMessage, PrepareRequestContext } from "@earendil-works/pi-agent-core";
+import type { AssistantMessage } from "@earendil-works/pi-ai";
+import { nativeInputCause, needsContinuation, type LhMessage } from "../core/native-messages.ts";
+
+export interface AdmissionScope { accountId: string; epoch: number; authorityRevision: number; selectedModelId: string }
+export type NoDispatchKind = "background" | "all" | "other";
+export interface ChatGPTDispatchLifecycle {
+ captured(scope: AdmissionScope): void;
+ beforeDispatch(scope: AdmissionScope): void;
+ noDispatch(kind: NoDispatchKind): void;
+}
+interface CycleOptions {
+ initialMessages: AgentMessage[];
+ staged?: AgentMessage[];
+ foregroundOwed: boolean;
+ purpose: { background: boolean };
+ scopeCurrent: (scope: AdmissionScope) => boolean;
+ commitStaged?: () => void;
+ restoreAutomatic: (automaticIds: ReadonlySet<string>, retainedIds: ReadonlySet<string>) => void;
+}
+const idOf = (m: AgentMessage) => (m as { id?: string }).id;
+/** One logical turn, not an SDK loop. No credentials, persisted provenance or wire protocol. */
+export class ChatGPTRequestCycle implements ChatGPTDispatchLifecycle {
+ private release?: (event: AgentEvent) => void;
+ private held: AgentEvent[] = [];
+ private snapshot: AgentMessage[] = [];
+ private seen: Set<string>;
+ private automatic = new Set<string>();
+ private committed = false;
+ private denied?: NoDispatchKind;
+ private scope?: AdmissionScope;
+ private originalScope?: AdmissionScope;
+ private foregroundOwed: boolean;
+ private inheritedBackground: boolean;
+ private resumed = false;
+ private userAborted = false;
+ private stagedCommitted = false;
+ private started = false;
+ private anyDispatch = false;
+ private firstResolved = false;
+ private resolveAdmission!: (admitted: boolean) => void;
+ readonly admission = new Promise<boolean>((resolve) => { this.resolveAdmission = resolve; });
+ constructor(private readonly options: CycleOptions) {
+  this.seen = new Set(options.initialMessages.map(idOf).filter((id): id is string => !!id));
+  for (const message of options.staged ?? []) { const id = idOf(message); if (id) this.seen.delete(id); }
+  this.foregroundOwed = options.foregroundOwed;
+  this.inheritedBackground = options.purpose.background;
+ }
+ get dispatched(): boolean { return this.anyDispatch; }
+ get deniedBeforeDispatch(): boolean { return this.denied !== undefined; }
+ abort(): void { this.userAborted = true; }
+ bindRelease(release: (event: AgentEvent) => void): () => void { this.release = release; return () => { this.release = undefined; }; }
+ prepare(request: PrepareRequestContext): void {
+  const messages = request.context.messages as AgentMessage[];
+  this.snapshot = [...messages]; this.automatic = new Set();
+  let learner = false;
+  for (const m of messages) {
+   const id = idOf(m); if (!id || this.seen.has(id)) continue;
+   const cause = nativeInputCause(m, messages.slice(0, messages.indexOf(m)));
+   if (cause === "automatic") this.automatic.add(id);
+   if (cause === "learner") learner = true;
+  }
+  if (this.automatic.size > 0) this.options.purpose.background = true;
+  const retained = messages.filter((m) => !this.automatic.has(idOf(m) ?? ""));
+  this.foregroundOwed ||= learner || (!this.inheritedBackground && needsContinuation(retained as LhMessage[]));
+ }
+ hold(event: AgentEvent): boolean {
+  if (event.type === "agent_start") { if (this.started) return true; this.started = true; return false; }
+  if (event.type === "agent_end") return true; // settle only after native activeRun is cleared
+  if (event.type === "turn_start") {
+   this.committed = false; this.denied = undefined; this.scope = undefined;
+   // A retained learner batch from the denied cycle already owns its turn_start.
+   if (!this.held.some((e) => e.type === "turn_start")) this.held.push(event);
+   return true;
+  }
+  if (!this.committed) { this.held.push(event); return true; }
+  return false;
+ }
+ captured(scope: AdmissionScope): void {
+  if (this.originalScope && this.resumed && JSON.stringify(scope) !== JSON.stringify(this.originalScope)) throw new Error("ChatGPT recovery scope changed");
+  this.scope = { ...scope };
+  this.originalScope ??= { ...scope };
+ }
+ beforeDispatch(scope: AdmissionScope): void {
+  if (this.userAborted || this.denied || this.committed || !this.scope || JSON.stringify(scope) !== JSON.stringify(this.scope) || !this.options.scopeCurrent(scope)) throw new Error("ChatGPT admission cancelled");
+  if (!this.stagedCommitted && (this.options.staged?.length ?? 0) > 0) {
+   this.options.commitStaged?.(); // synchronous durable prefix, before visibility and fetch
+   this.stagedCommitted = true;
+   const stagedEvents: AgentEvent[] = this.options.staged!.flatMap((message) => [{ type: "message_start" as const, message }, { type: "message_end" as const, message }]);
+   const boundary = this.held.findIndex((event) => event.type === "turn_start");
+   this.held.splice(boundary < 0 ? 0 : boundary + 1, 0, ...stagedEvents);
+  }
+  this.committed = true; this.anyDispatch = true; this.foregroundOwed = false;
+  this.inheritedBackground = this.options.purpose.background;
+  for (const m of this.snapshot) { const id = idOf(m); if (id) this.seen.add(id); }
+  const events = this.held; this.held = [];
+  for (const event of events) this.release?.(event);
+  this.resolveFirst(true);
+ }
+ noDispatch(kind: NoDispatchKind): void { if (!this.committed) this.denied = this.userAborted ? "all" : kind; }
+ acceptsUsage(_message: AssistantMessage): boolean { return this.committed; }
+ private resolveFirst(value: boolean) { if (!this.firstResolved) { this.firstResolved = true; this.resolveAdmission(value); } }
+ finishNativeCycle(): { action: "settle" | "resumeForeground"; messages: AgentMessage[] } {
+  if (!this.denied) { this.resolveFirst(this.anyDispatch); return { action: "settle", messages: [] }; }
+  const messages = this.snapshot.filter((m) => !this.automatic.has(idOf(m) ?? ""));
+  const retainedIds = new Set(messages.map(idOf).filter((id): id is string => !!id));
+  this.options.restoreAutomatic(this.automatic, retainedIds);
+  this.resolveFirst(false); // queue deferral is installed before initial fire cargo returns
+  this.held = this.held.filter((e) => (e.type === "turn_start") || ((e.type === "message_start" || e.type === "message_end") && e.message.role !== "assistant" && !this.automatic.has(idOf(e.message) ?? "")));
+  const retry = this.denied === "background" && !this.userAborted && !this.resumed && this.foregroundOwed && !!this.scope && this.options.scopeCurrent(this.scope);
+  if (retry) {
+   this.resumed = true; this.options.purpose.background = false;
+   this.seen = retainedIds;
+   this.denied = undefined;
+   // Staged automatic input must not be resurrected on foreground recovery.
+   this.options.staged = undefined;
+   return { action: "resumeForeground", messages };
+  }
+  this.held = [];
+  return { action: "settle", messages };
+ }
+}

src/agent/chatgpt-stream.ts

@@ -0,0 +1,118 @@
+import type { Api, AssistantMessage, Model, SimpleStreamOptions, StreamFunction, TranscriptContext } from "@earendil-works/pi-ai";
+import { streamSimple } from "@earendil-works/pi-ai/api/openai-responses";
+import { AssistantMessageEventStream } from "@earendil-works/pi-ai/utils/event-stream";
+import { transformMessages } from "@earendil-works/pi-ai/api/transform-messages";
+import { chatGPTAbortController, chatGPTScope, chatGPTScopeCurrent, chatGPTLease, chatGPTDenial } from "./chatgpt-bridge.ts";
+import type { AdmissionScope, ChatGPTDispatchLifecycle } from "./chatgpt-request-cycle.ts";
+
+const USAGE = "https://chatgpt.com/settings/usage";
+export function chatGPTStreamError(message: string): string {
+	if (/subscription_sharing_usage_limit_exceeded|quota|429/.test(message)) return `ChatGPT plan usage limit reached. Manage usage at ${USAGE}. No API-key fallback was used.`;
+	if (/scope|authorization_context|permission/.test(message)) return "ChatGPT plan permission is missing. Reconnect in Connections and authorize plan use. No API-key fallback was used.";
+	if (/expired|invalid_grant|401|invalid_user/.test(message)) return "Your ChatGPT session expired. Reconnect in Connections. No API-key fallback was used.";
+	if (/background|Background/.test(message)) return "Background use is off for this ChatGPT account. Enable Allow background use in Connections to run this work. No API-key fallback was used.";
+	if (/eligible|403/.test(message)) return "This ChatGPT account or workspace does not permit this request. Check ChatGPT settings; reconnecting may not change eligibility.";
+	if (/cancel|abort|closed|disconnect/i.test(message)) return "ChatGPT request cancelled or connection closed. No API-key fallback was used.";
+	if (/catalog|model/.test(message)) return "The selected model is unavailable for this ChatGPT account. Refresh Connections and explicitly choose an available model. No API-key fallback was used.";
+	if (/incomplete|terminal|interrupted|stop reason/.test(message)) return "ChatGPT returned an incomplete response. Retry when ready; no API-key fallback was used.";
+	return "ChatGPT request failed. Check your connection and ChatGPT usage, then retry. No API-key fallback was used.";
+}
+
+/** Restrict the body at the final payload seam, even if caller samplingParams add fields. */
+export function chatGPTPayload(raw: unknown, selectedModelId: string): Record<string, unknown> {
+	const payload = raw as Record<string, unknown>;
+	const out: Record<string, unknown> = { model: selectedModelId, store: false, stream: true };
+	for (const key of ["instructions", "reasoning", "include", "text", "tool_choice", "parallel_tool_calls", "prompt_cache_key"]) {
+		if (payload[key] !== undefined) out[key] = payload[key];
+	}
+	if (!Array.isArray(payload.input)) throw new Error("ChatGPT requires stateless input history");
+	out.input = payload.input.map((item) => item?.role === "system" ? { ...item, role: "developer" } : item);
+	if (Array.isArray(payload.tools) && payload.tools.length) {
+		if (payload.tools.some((tool) => tool.type !== "function" && tool.type !== "custom")) throw new Error("Unsupported ChatGPT tool capability");
+		out.tools = [{ type: "namespace", name: "learnharness", description: "LearnHarness local learning tools", tools: payload.tools }];
+	}
+	return out;
+}
+
+/** Only ChatGPT history is translated. Mixed API-key history keeps its actual identity. */
+export function chatGPTHistory(context: TranscriptContext, logicalModel: Model<Api>): TranscriptContext {
+	// Apply Pi's normal foreign-model signature treatment BEFORE the alias
+	// becomes openai, or real API-key turns could look like subscription turns.
+	const messages = transformMessages(context.messages, logicalModel);
+	return { ...context, messages: messages.map((m) => m.role === "assistant" && m.provider === "chatgpt" ? { ...m, provider: "openai" } : m) };
+}
+function empty(model: Model<Api>): AssistantMessage {
+	return { role: "assistant", api: model.api, provider: "chatgpt", model: model.id, content: [], stopReason: "error", timestamp: Date.now(), usage: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, totalTokens: 0, cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 } } };
+}
+export function chatGPTStream(background: () => boolean, lifecycle?: ChatGPTDispatchLifecycle): StreamFunction<Api, SimpleStreamOptions> {
+	return (logicalModel, context, options) => {
+		const output = new AssistantMessageEventStream();
+		const purpose = background();
+		const { controller, release } = chatGPTAbortController(purpose);
+		const signal = options?.signal ? AbortSignal.any([options.signal, controller.signal]) : controller.signal;
+		void (async () => {
+			let partial = empty(logicalModel);
+			let completed = false;
+			let failure = "";
+			let scope: AdmissionScope | undefined;
+			let committed = false;
+			let terminal = false;
+			const check = () => { if (signal.aborted || !scope || !chatGPTScopeCurrent(scope)) throw new Error("cancelled or catalog authority changed"); };
+			try {
+				if (logicalModel.provider !== "chatgpt") throw new Error("ChatGPT account catalog has no selected model");
+				if (Object.keys(options?.headers ?? {}).some((key) => key.toLowerCase() === "authorization")) throw new Error("Authorization override is forbidden");
+				scope = chatGPTScope(logicalModel.id);
+				lifecycle?.captured(scope);
+				const credentials = await chatGPTLease(scope, purpose);
+				check();
+				// NEVER dispatch through Models under openai: that would expose ambient API-key fallback.
+				const transport: Model<"openai-responses"> = { ...logicalModel, api: "openai-responses", provider: "openai", baseUrl: "https://api.openai.com/v1", headers: undefined, samplingParams: undefined, compat: { ...logicalModel.compat, supportsDeveloperRole: true } };
+				const delegate = options?.fetch ?? globalThis.fetch;
+				const source = streamSimple(transport, chatGPTHistory(context, logicalModel), {
+					...options,
+					apiKey: credentials.accessToken,
+					signal,
+					maxRetries: 0,
+					headers: { "OpenAI-Beta": null, ...options?.headers },
+					fetch: (input, init) => {
+						check();
+						if (committed) throw new Error("ChatGPT does not retry admitted requests");
+						lifecycle?.beforeDispatch(scope!);
+						committed = true;
+						return delegate(input, init);
+					},
+					onPayload: async (raw, model) => {
+						const replacement = await options?.onPayload?.(raw, model);
+						const payload = chatGPTPayload(replacement === undefined ? raw : replacement, logicalModel.id);
+						check();
+						return payload;
+					},
+					onProviderStreamEvent: async (raw, model) => {
+						const event = raw as { type?: string; code?: string; error?: { code?: string }; response?: { error?: { code?: string }; status?: string } };
+						if (event.type === "response.completed" && event.response?.status === "completed" && !event.response?.error) completed = true;
+						if (["response.incomplete", "response.failed", "error"].includes(event.type ?? "") || event.response?.error) failure = event.code || event.error?.code || event.response?.error?.code || "incomplete terminal response";
+						await options?.onProviderStreamEvent?.(raw, model);
+					},
+				});
+				for await (const event of source) {
+					const message = "partial" in event ? event.partial : event.type === "done" ? event.message : event.error;
+					// Copy the message, not its tool/reasoning blocks: IDs/signatures remain byte-for-byte intact.
+					partial = { ...message, provider: "chatgpt" };
+					if (event.type === "done") {
+						if (!completed || failure) throw new Error(failure || "missing terminal response.completed");
+						terminal = true;
+						output.push({ ...event, message: partial });
+					} else if (event.type === "error") {
+						throw new Error(failure || event.error.errorMessage || "ChatGPT stream error");
+					} else output.push({ ...event, partial });
+				}
+			} catch (err) {
+				if (terminal) return;
+				if (!committed) lifecycle?.noDispatch(await chatGPTDenial(scope, controller, options?.signal?.aborted === true, purpose));
+				partial = { ...partial, stopReason: signal.aborted ? "aborted" : "error", errorMessage: chatGPTStreamError(err instanceof Error ? err.message : "") };
+				output.push({ type: "error", reason: partial.stopReason as "error" | "aborted", error: partial });
+			} finally { release(); output.end(); }
+		})();
+		return output;
+	};
+}

src/agent/models.ts

@@ -12,6 +12,9 @@ import type { ThinkingLevel } from "@earendil-works/pi-agent-core";
 import { type ModelRole } from "../../config.ts";
 import { DEFAULT_THINKING_ORDER } from "../core/settings.ts";
 import packageMetadata from "../../package.json" with { type: "json" };
+import { chatGPTModel } from "./chatgpt-bridge.ts";
+import { chatGPTStream } from "./chatgpt-stream.ts";
+import type { ChatGPTDispatchLifecycle } from "./chatgpt-request-cycle.ts";
 
 /** App identity is immutable and shared by Models and every native Agent.
  * Agent's default Pi attribution otherwise overrides the Models identity. */
@@ -93,9 +96,9 @@ function defaultThinkingLevel(model: Model<Api>): ThinkingLevel {
  *  key is resolved by pi-ai itself from the environment per request (the
  *  keys seam reflects file-saved keys into env), so the engine never
  *  hardcodes OpenRouter's. */
-export function piEngineForSpec(spec: ResolvedModelSpec): PiEngine {
+export function piEngineForSpec(spec: ResolvedModelSpec, background: boolean | (() => boolean) = true, lifecycle?: ChatGPTDispatchLifecycle): PiEngine {
 	const models: Models = piModels();
-	const model = models.getModel(spec.provider, spec.id);
+	const model = spec.provider === "chatgpt" ? chatGPTModel(spec.id) : models.getModel(spec.provider, spec.id);
 	if (!model) {
 		throw new Error(`${spec.provider} catalog has no model ${spec.id}`);
 	}
@@ -111,11 +114,10 @@ export function piEngineForSpec(spec: ResolvedModelSpec): PiEngine {
 		// bound streamSimple: the agent loop forwards state.thinkingLevel as
 		// `reasoning` (undefined = off on pi-ai's scale); the clamp above keeps
 		// the pin honest even when config or settings drift off the map.
-		streamFn: (requestModel, context, options) =>
-			models.streamSimple(requestModel, context, {
-				...options,
-				reasoning: level === "off" ? undefined : level,
-			}),
+		streamFn: (requestModel, context, options) => {
+			const stream = spec.provider === "chatgpt" ? chatGPTStream(typeof background === "function" ? background : () => background, lifecycle) : models.streamSimple.bind(models);
+			return stream(requestModel, context, { ...options, reasoning: level === "off" ? undefined : level });
+		},
 	};
 }
 

src/agent/subagents/headless.ts

@@ -137,6 +137,9 @@ export function browserServiceSpawnEnv(
 	const out: Record<string, string | undefined> = { ...env };
 	delete out.ELECTRON_RUN_AS_NODE;
 	delete out.LH_DEV_URL;
+	// The artifact renderer is not an authorized app API client.
+	delete out.LH_DESKTOP_API_SECRET;
+	delete out.LH_DESKTOP_SERVER;
 	return out;
 }
 

src/agent/teacher-agent.ts

@@ -18,7 +18,8 @@ import { practiceTools, type PracticeToolContext } from "./tools/practice.ts";
 import { healToolRecord } from "./tools/heal.ts";
 import { quizInputObject, quizInputProblems } from "../core/question-core.ts";
 import { QUESTION_TOOLS, type EnabledTools } from "../core/pending.ts";
-import { toLlmMessages } from "../core/native-messages.ts";
+import { toLlmMessages, nativeInputCause } from "../core/native-messages.ts";
+import type { ChatGPTRequestCycle } from "./chatgpt-request-cycle.ts";
 
 /**
  * The two teaching-side agents (docs/probe-teach-split.md): the PROBER and
@@ -97,6 +98,8 @@ export type StepUsageCollector = (usage: Usage) => void;
 export interface AgentBuildOptions {
 	/** Model + stream function — tests inject a scripted faux engine. */
 	engine?: PiEngine;
+	callPurpose?: { background: boolean };
+	requestCycle?: ChatGPTRequestCycle;
 	/** The initial transcript, as pi messages (the record IS agent state). */
 	messages?: AgentMessage[];
 	/** Money channel: one call per model request. */
@@ -178,6 +181,7 @@ function baseAgentOptions(kind: AgentKind, opts: AgentBuildOptions, tools: Agent
 	// LH_PROVIDER_EVENTS=1 and the caller handed a chatId; undefined is
 	// the normal no-observer path (tests, env-off).
 	const providerEvents = opts.chatId ? providerEventObserver(opts.chatId, kind) : undefined;
+	const seenInputs = new Set((opts.messages ?? []).map((m) => (m as { id?: string }).id).filter(Boolean));
 	const agent = new Agent({
 		identity: APP_IDENTITY,
 		sessionId: opts.chatId,
@@ -211,13 +215,31 @@ function baseAgentOptions(kind: AgentKind, opts: AgentBuildOptions, tools: Agent
 		// an abort's dangling question call gets its INTERRUPTED result. It
 		// must never throw (pi's convertToLlm contract).
 		convertToLlm: toLlmMessages,
+		prepareRequest: (request) => {
+			if (opts.callPurpose) {
+				let learner = false, automatic = false;
+				const messages = request.context.messages;
+				for (const message of messages) {
+					const id = (message as { id?: string }).id;
+					const cause = nativeInputCause(message, messages.slice(0, messages.indexOf(message)));
+					if (cause === "neutral") continue;
+					if (id && seenInputs.has(id)) continue;
+					if (id) seenInputs.add(id);
+					if (cause === "learner" && id) learner = true;
+					else automatic = true; // missing/unrecognized identity never grants foreground
+				}
+				if (automatic) opts.callPurpose.background = true;
+				else if (learner) opts.callPurpose.background = false;
+			}
+			opts.requestCycle?.prepare(request);
+		},
 	});
 	if (opts.onUsage) {
 		agent.subscribe((event: AgentEvent) => {
 			if (event.type === "message_end" && event.message.role === "assistant" && "usage" in event.message) {
 				// A cancelled call is not a billed step (partial/estimated
 				// usage on an aborted message is not money spent).
-				if (event.message.stopReason === "aborted") return;
+				if (event.message.stopReason === "aborted" || (opts.requestCycle && !opts.requestCycle.acceptsUsage(event.message))) return;
 				opts.onUsage!(event.message.usage);
 			}
 		});

src/core/contracts/chatgpt.ts

@@ -0,0 +1,18 @@
+/** Client-safe status only. Credentials and OAuth transaction data never enter this contract. */
+export interface ChatGPTStatus {
+	available: boolean;
+	pending: boolean;
+	activeId: string | null;
+	accounts: { id: string; email: string; registration: string; connected: boolean; background: boolean; planEnabled?: boolean }[];
+	error: string | null;
+	catalogError?: string | null;
+}
+export interface ChatGPTControlBridge {
+	status(): Promise<ChatGPTStatus>;
+	login(id?: string): Promise<ChatGPTStatus>;
+	cancel(): Promise<ChatGPTStatus>;
+	select(id: string): Promise<ChatGPTStatus>;
+	logout(id: string): Promise<ChatGPTStatus>;
+	consent(id: string, enabled: boolean): Promise<ChatGPTStatus>;
+	onChange(callback: () => void): () => void;
+}

src/core/cost-summary.ts

@@ -18,10 +18,12 @@
  */
 
 import type { CallUsage, EmissionStats, ExtractionRunUsage } from "./contracts.ts";
-import { callCost, sumSteps } from "./prices.ts";
+import { callCost, sumSteps, isChatGPTPlan } from "./prices.ts";
 
 /** One pipeline stage's cumulative cost across every run. */
 export interface StageCost {
+	/** ChatGPT plan stage-runs, excluded from API dollars and unpriced counts. */
+	subscriptionRuns?: number;
 	/** Inferred USD summed over PRICED runs (partial when unpricedRuns > 0). */
 	dollars: number;
 	/** Stage-runs whose model is missing from PRICES (render "—", never zero). */
@@ -34,6 +36,7 @@ export interface StageCost {
 
 /** One run in the recent-runs list (extraction or naming). */
 export interface RecentRun {
+	subscriptionRuns?: number;
 	startedAt: number;
 	/** ISO wall-clock. */
 	at: string;
@@ -51,9 +54,10 @@ export interface RecentRun {
 
 /** The cost pill's whole payload: stage totals + recent runs. */
 export interface GraphCostSummary {
+	subscriptionRuns?: number;
 	/** Inferred USD across all priced stages of all runs. */
 	totalDollars: number;
-	/** Runs where EVERY stage was unpriced (totalDollars excludes them). */
+	/** Runs with no priced API stage and at least one unknown stage. */
 	unpricedRuns: number;
 	/** The billable pipeline stages. (Wave 7, owner ruling 2026-09-23:
 	 *  the embedding stage is gone with the provider embedder; the
@@ -89,6 +93,7 @@ const emptyStage = (): StageCost => ({ dollars: 0, unpricedRuns: 0, runs: 0, tok
 /** One stage's dollars over its steps; null when the model is unpriced
  *  (all-or-nothing: a stage runs one model). */
 function stageDollars(model: string, steps: readonly CallUsage[]): number | null {
+	if (isChatGPTPlan(model)) return null;
 	let total = 0;
 	for (const step of steps) {
 		const c = callCost(step, model);
@@ -121,13 +126,15 @@ export function graphCostSummary(input: CostSummaryInput): GraphCostSummary {
 		extraction.runs++;
 		mergeTokens(extraction.tokens, stages.extractor.steps);
 		const extDollars = stageDollars(stages.extractor.model, stages.extractor.steps);
-		if (extDollars === null) extraction.unpricedRuns++;
+		if (isChatGPTPlan(stages.extractor.model)) extraction.subscriptionRuns = (extraction.subscriptionRuns ?? 0) + 1;
+		else if (extDollars === null) extraction.unpricedRuns++;
 		else extraction.dollars += extDollars;
 
 		audit.runs++;
 		mergeTokens(audit.tokens, stages.auditor.steps);
 		const audDollars = stageDollars(stages.auditor.model, stages.auditor.steps);
-		if (audDollars === null) audit.unpricedRuns++;
+		if (isChatGPTPlan(stages.auditor.model)) audit.subscriptionRuns = (audit.subscriptionRuns ?? 0) + 1;
+		else if (audDollars === null) audit.unpricedRuns++;
 		else audit.dollars += audDollars;
 
 		// Run-level: null only when EVERY stage is unpriced.
@@ -135,8 +142,10 @@ export function graphCostSummary(input: CostSummaryInput): GraphCostSummary {
 		const runDollars = allStages.every((d) => d === null)
 			? null
 			: allStages.reduce<number>((a, d) => a + (d ?? 0), 0);
+		const subscriptionRuns = [stages.extractor, stages.auditor].filter((stage) => isChatGPTPlan(stage.model)).length;
 		const allSteps = [...stages.extractor.steps, ...stages.auditor.steps];
 		runs.push({
+			...(subscriptionRuns ? { subscriptionRuns } : {}),
 			startedAt: run.startedAt,
 			at: run.at,
 			durationMs: run.durationMs,
@@ -159,8 +168,10 @@ export function graphCostSummary(input: CostSummaryInput): GraphCostSummary {
 	const capped = runs.slice(0, input.recentCap ?? 10);
 
 	// A run is "unpriced" when it billed nothing countable at all.
-	const unpricedRuns = runs.filter((r) => r.dollars === null).length;
+	const unpricedRuns = runs.filter((r) => r.dollars === null && (r.subscriptionRuns ?? 0) < 2).length;
+	const subscriptionRuns = (extraction.subscriptionRuns ?? 0) + (audit.subscriptionRuns ?? 0);
 	return {
+		...(subscriptionRuns ? { subscriptionRuns } : {}),
 		totalDollars: extraction.dollars + audit.dollars,
 		unpricedRuns,
 		stages: { extraction, audit },

src/core/index.ts

@@ -48,6 +48,7 @@ export * from "./input-events.ts";
 export * from "./view.ts";
 export * from "./tool-parts.ts";
 export * from "./prices.ts";
+export type { ChatGPTStatus, ChatGPTControlBridge } from "./contracts/chatgpt.ts";
 export * from "./cost-summary.ts";
 export * from "./fuzzy.ts";
 // settings.ts + settings-resolution.ts (settings-storage wave, 2026-09-19):

src/core/native-messages.ts

@@ -674,6 +674,57 @@ export function endsInDismissal(messages: readonly LhMessage[]): boolean {
 	return last !== undefined && isDismissedResult(last);
 }
 
+export type RequestCause = "foreground" | "background";
+export type ReconstructedCause = RequestCause | "ambiguous";
+export type NativeInputCause = "learner" | "automatic" | "neutral" | "unrecognizable";
+/** Native causes only: flattened user roles, IDs' prefixes and prose confer no permission. */
+export function nativeInputCause(message: AgentMessage, history: readonly AgentMessage[]): NativeInputCause {
+	if (message.role === "user") return "learner";
+	if (message.role === "system" || message.role === "assistant") return "neutral";
+	if (message.role === "custom") {
+		if (["probe-handoff", "subagent-steer", "subagent-abort"].includes(message.customType)) return "automatic";
+		if (["practice-thinking", "practice-next"].includes(message.customType)) return "learner";
+		return ["graph-context", "plan-block", "problem-dismissed", "compaction"].includes(message.customType) ? "neutral" : "unrecognizable";
+	}
+	if (message.role === "toolResult") {
+		if (!(QUESTION_TOOLS as readonly string[]).includes(message.toolName) || message.isError || message.details == null || (message.details as { dismissed?: boolean }).dismissed) return "neutral";
+		const details = message.details as Partial<QuestionAnswerData>;
+		if (!["quiz", "ask", "explain"].includes(details.kind ?? "") || typeof details.questionText !== "string" || typeof details.dontKnow !== "boolean" || !Array.isArray(details.answers) || typeof details.answeredAt !== "string") return "unrecognizable";
+		const calls = history.flatMap((m) => m.role === "assistant" ? m.content.filter((b) => b.type === "toolCall" && b.id === message.toolCallId) : []);
+		return calls.length === 1 && calls[0]?.type === "toolCall" && calls[0].name === message.toolName ? "learner" : "unrecognizable";
+	}
+	return "unrecognizable";
+}
+/** Full persisted native record, including pre-compaction causes. Never guessed foreground. */
+export function reconstructRequestCause(messages: readonly LhMessage[]): ReconstructedCause {
+	let inherited: ReconstructedCause = "ambiguous";
+	let triggers: NativeInputCause[] = [];
+	const seen = new Map<string, string>();
+	const history: AgentMessage[] = [];
+	const resolve = (): ReconstructedCause => {
+		if (triggers.includes("unrecognizable")) return "ambiguous";
+		const lastLearner = triggers.lastIndexOf("learner"), lastAutomatic = triggers.lastIndexOf("automatic");
+		if (lastLearner < 0 && lastAutomatic < 0) return inherited;
+		if (lastAutomatic < 0) return "foreground";
+		if (lastLearner < lastAutomatic) return "background";
+		return "ambiguous";
+	};
+	for (const message of messages) {
+		const cause = nativeInputCause(message, history);
+		const id = (message as { id?: string }).id;
+		if (id) {
+			const value = JSON.stringify(message);
+			if (seen.has(id)) { if (seen.get(id) !== value) triggers.push("unrecognizable"); continue; }
+			seen.set(id, value);
+		} else if (cause !== "neutral") triggers.push("unrecognizable");
+		if (message.role === "assistant" && !(message.provider === "synthetic" && message.api === "faux" && message.model === "synthetic-conclude" && message.content.some((b) => b.type === "toolCall" && b.name === "conclude" && b.arguments?.outcome === "dismissed"))) {
+			inherited = resolve(); triggers = [];
+		} else if (cause !== "neutral") triggers.push(cause);
+		history.push(message);
+	}
+	return resolve();
+}
+
 export function needsContinuation(messages: readonly LhMessage[]): boolean {
 	const last = messages[messages.length - 1];
 	if (!last || (last as { role?: string }).role !== "toolResult") return false;

src/core/prices.ts

@@ -1661,6 +1661,8 @@ export function sumSteps(steps: readonly CallUsage[]): CallUsage {
 /** One model call's inferred cost in USD — null when the model is unknown
  *  or the usage fields are absent (never guess a price). Cached input bills
  *  at the cache-read rate; cache writes bill separately where priced. */
+export function isChatGPTPlan(model: string | undefined): boolean { return model?.startsWith("chatgpt:") === true; }
+
 export function callCost(
 	usage: { inputTokens?: number; outputTokens?: number; cacheReadTokens?: number; cacheWriteTokens?: number },
 	model: string,
@@ -1681,6 +1683,7 @@ export function callCost(
 
 /** One turn's inferred cost — the sum over its steps; null when unpriceable. */
 export function turnCost(turn: TurnUsage): number | null {
+	if (isChatGPTPlan(turn.model)) return null;
 	let total = 0;
 	for (const step of turn.steps) {
 		const c = callCost(step, turn.model);
@@ -1693,15 +1696,17 @@ export function turnCost(turn: TurnUsage): number | null {
 /** Fold a chat record's turn log: total inferred session cost. Unknown-model
  *  or partial-usage turns don't poison the total — they count as unpriced
  *  and are reported alongside, so the UI can show "n turns unpriced". */
-export function sessionCost(record: Pick<ChatRecord, "turns">): { cost: number; unpriced: number } {
+export function sessionCost(record: Pick<ChatRecord, "turns">): { cost: number; unpriced: number; subscription?: number } {
 	let cost = 0;
 	let unpriced = 0;
+	let subscription = 0;
 	for (const turn of record.turns ?? []) {
+		if (isChatGPTPlan(turn.model)) { subscription++; continue; }
 		const c = turnCost(turn);
 		if (c === null) unpriced++;
 		else cost += c;
 	}
-	return { cost, unpriced };
+	return { cost, unpriced, ...(subscription ? { subscription } : {}) };
 }
 
 /** The session's context-window footprint: the LAST model call's

src/server/chat-handler.ts

@@ -7,6 +7,10 @@ import {
 	type TeachingAgent,
 } from "../agent/teacher-agent.ts";
 import { piEngineForSpec, type PiEngine } from "../agent/models.ts";
+import { chatGPTBackgroundAllowed, chatGPTScopeCurrent } from "../agent/chatgpt-bridge.ts";
+import { chatGPTStreamError } from "../agent/chatgpt-stream.ts";
+import { ChatGPTRequestCycle, type ChatGPTDispatchLifecycle } from "../agent/chatgpt-request-cycle.ts";
+import { reconstructRequestCause } from "../core/native-messages.ts";
 import { createJobRunner } from "../agent/subagents/runner.ts";
 import { createExtractionPipeline } from "../agent/subagents/extraction.ts";
 import type { ModelRole } from "../../config.ts";
@@ -15,6 +19,7 @@ import {
 	noteSteerVisible,
 	reconcileUserSteers,
 	registerLiveAgent,
+	restoreUncommittedAutomatic,
 	setJobRespawn,
 	setTurnRunner,
 	settleSteers,
@@ -41,7 +46,7 @@ import { persistTurn, triggerExtraction } from "./turn-service.ts";
 import { hasProbeDone, danglingQuestionOf } from "../core/native-messages.ts";
 import { composeInstructions } from "../agent/teacher-agent.ts";
 import { ensureCompacted, kickCompaction, effectiveThresholdTokens } from "./compaction.ts";
-import { catalogContextLimits } from "./model-catalog.ts";
+import { catalogContextLimits, refreshConnectionCatalog } from "./model-catalog.ts";
 import { emit } from "../core/events.ts";
 import { settledTurnResponse } from "./turn-runs.ts";
 import { asAgentMessages } from "../core/native-messages.ts";
@@ -201,6 +206,12 @@ export async function handleChatInput(
 	// persist (nothing durable changes); wake-pipeline turns ride the same
 	// path. The affected tier is strong — the conversation chain's own tier
 	// (a chat.model override still resolves through it).
+	if (conversation.provider === "chatgpt") {
+		await refreshConnectionCatalog();
+		if (input.type === "steers" && !chatGPTBackgroundAllowed()) {
+			return Response.json({ error: "Background use is off for this ChatGPT account. Automatic follow-up remains deferred in its queue." }, { status: 403 });
+		}
+	}
 	if (!providerKeyPresent(conversation.provider)) {
 		const envelope = providerKeyMissingEnvelope(conversation.provider, ["strong"]);
 		console.info(`[chat ${chatId}] turn refused — ${JSON.stringify(envelope)}`);
@@ -216,12 +227,22 @@ export async function handleChatInput(
 	// placement mirrors the BYOK/key gates — BEFORE appends and
 	// beginTurn, no liveness effects — and the record re-reads after
 	// because compaction appends the marker.
-	const conversationEngine = piEngineForSpec(conversation);
+	const callPurpose = { background: input.type === "steers" || (input.type === "continue" && reconstructRequestCause(record.messages) !== "foreground") };
+	let requestCycle: ChatGPTRequestCycle | undefined;
+	const lifecycle: ChatGPTDispatchLifecycle = {
+		captured: (scope) => requestCycle?.captured(scope),
+		beforeDispatch: (scope) => requestCycle?.beforeDispatch(scope),
+		noDispatch: (kind) => requestCycle?.noDispatch(kind),
+	};
+	let conversationEngine: PiEngine;
+	try { conversationEngine = piEngineForSpec(conversation, () => callPurpose.background, lifecycle); }
+	catch { return Response.json({ error: conversation.provider === "chatgpt" ? chatGPTStreamError("model catalog") : "The selected model is unavailable." }, { status: 503 }); }
 	const modelIdForCompaction = `${conversationEngine.model.provider}:${conversationEngine.model.id}`;
 	await ensureCompacted(
 		{
 			chatStore: stores.chatStore,
-			engine: conversationEngine,
+			engine: piEngineForSpec(conversation, true),
+			canStart: () => conversation.provider !== "chatgpt" || chatGPTBackgroundAllowed(),
 			modelId: modelIdForCompaction,
 			kind: record.kind === "practice" ? "practice" : hasProbeDone(record.messages) ? "teach" : "probe",
 			thresholdTokens: effectiveThresholdTokens(
@@ -287,12 +308,14 @@ export async function handleChatInput(
 	// ride this same write-ahead - the pump de-queued them at fire, so this
 	// is the ONLY copy ever appended.)
 	const merged = [...record.messages, ...appends];
+	const stagedAutomatic = conversation.provider === "chatgpt" && input.type === "steers";
+	let persistBaseCount = stagedAutomatic ? record.messages.length : merged.length;
 	try {
-		stores.chatStore.saveMessages(chatId, merged);
+		if (!stagedAutomatic) stores.chatStore.saveMessages(chatId, merged);
 		// Dock fix (owner ruling 2026-09-26): persisted markers are visible
 		// in the record — flip their jobs now (the wake's steers ride this
 		// write-ahead; the delivering turn may run for minutes).
-		for (const appended of appends) {
+		for (const appended of stagedAutomatic ? [] : appends) {
 			if ((appended as { role?: string }).role === "custom") noteSteerVisible(appended as SteerMessage);
 		}
 	} catch (err) {
@@ -339,8 +362,19 @@ export async function handleChatInput(
 	// enabled). The definitive toggleable set is TOGGLEABLE_TOOLS — the
 	// sub-agent JOB tools (research/demo/visual/pdf; adjudicated with the
 	// user: quiz/ask/fs/plan/probe_done are always-on).
+	if (conversation.provider === "chatgpt") requestCycle = new ChatGPTRequestCycle({
+		initialMessages: asAgentMessages(merged),
+		staged: stagedAutomatic ? asAgentMessages(appends) : undefined,
+		foregroundOwed: !callPurpose.background,
+		purpose: callPurpose,
+		scopeCurrent: chatGPTScopeCurrent,
+		commitStaged: () => { stores.chatStore.saveMessages(chatId, merged); persistBaseCount = merged.length; },
+		restoreAutomatic: (automaticIds, retainedIds) => restoreUncommittedAutomatic(chatId, constructed, automaticIds, retainedIds),
+	});
 	const agentOpts: AgentBuildOptions = {
 		engine,
+		callPurpose,
+		requestCycle,
 		messages: asAgentMessages(merged),
 		onUsage: collectUsage,
 		toolToggles: effective.agent.tools,
@@ -439,6 +473,7 @@ export async function handleChatInput(
 		providerEventNote(chatId, "turn", `run start (base: ${merged.length} messages)`);
 		const handle = runPiTurn({
 			agent: constructed,
+			requestCycle,
 			// The coalescing window seam (COMPACTION LAW, 2026-10-09): tests
 			// tighten it via LH_UPDATE_FLUSH_MS; production rides the 100ms
 			// default in ui-stream.
@@ -454,13 +489,14 @@ export async function handleChatInput(
 			// route through here: they ride the error-result path and reach the
 			// transcript directly (the retry context the model saw).
 			onError: (error) => {
+				if (conversation.provider === "chatgpt") return chatGPTStreamError(String(error));
 				console.error(`[chat ${chatId}] turn error:`, error);
 				return "An error occurred.";
 			},
 			onFinish: (turnMessages, isAborted) => {
 				// Wave B: the record write is the agent's own final state —
 				// the new messages are the slice past the merged base.
-				const stamped = persistTurn(stores.chatStore, chatId, merged.length, turnMessages, isAborted);
+				const stamped = persistTurn(stores.chatStore, chatId, persistBaseCount, turnMessages, isAborted);
 				// usage log append (idempotent per turn id) — the stats bar
 				// refolds this on the client after the turn settles. The
 				// provenance message is the turn's FIRST new assistant message.
@@ -479,7 +515,7 @@ export async function handleChatInput(
 				// PRACTICE (owner, 2026-10-04): practice sessions write nothing —
 				// no extraction, no graph, no coverage (v1: the session itself
 				// is the deliverable).
-				if (record.kind !== "practice" && extractionEnabled(effective.graph)) {
+				if ((!requestCycle || requestCycle.dispatched) && record.kind !== "practice" && extractionEnabled(effective.graph)) {
 					// Fresh per-turn read of the graph section's chunk budget
 					// (extraction-UX redesign, 2026-09-24): the trigger compares
 					// against it and the pipeline's batches share it.
@@ -520,18 +556,20 @@ export async function handleChatInput(
 				// beneath-threshold record is a silent no-op. Skipped under
 				// tests (the extraction precedent — scripted chats are
 				// fixtures; the service itself is pinned directly).
-				if (process.env.NODE_ENV !== "test")
+				if (process.env.NODE_ENV !== "test" && (!requestCycle || requestCycle.dispatched))
 					kickCompaction(() => {
 						const fresh = effectiveSettings();
 						const spec = resolveConversationSpec(stores.chatStore.getChat(chatId)?.settings, fresh.models.strong);
-						if (!spec) return null;
-						const kickEngine = piEngineForSpec(spec);
+						if (!spec || (spec.provider === "chatgpt" && !chatGPTBackgroundAllowed())) return null;
+						let kickEngine: PiEngine;
+						try { kickEngine = piEngineForSpec(spec, true); } catch { return null; }
 						const kickModelId = `${kickEngine.model.provider}:${kickEngine.model.id}`;
 						const freshRecord = stores.chatStore.getChat(chatId);
 						if (!freshRecord) return null;
 						return {
 							chatStore: stores.chatStore,
 							engine: kickEngine,
+							canStart: () => spec.provider !== "chatgpt" || chatGPTBackgroundAllowed(),
 							modelId: kickModelId,
 							kind: freshRecord.kind === "practice" ? "practice" : hasProbeDone(freshRecord.messages) ? "teach" : "probe",
 							thresholdTokens: effectiveThresholdTokens(
@@ -558,6 +596,7 @@ export async function handleChatInput(
 		// must cost zero permanently-held slots beyond the one live view. The
 		// turn itself is a server-side process (ruling 2026-09-25) — it never
 		// needed this connection, only a view did.
+		if (stagedAutomatic && requestCycle && !(await requestCycle.admission)) return Response.json({ error: "Automatic follow-up remains deferred in its queue." }, { status: 403 });
 		return Response.json({ ok: true });
 	} catch (err) {
 		// The stream never started, so no onFinish will fire — clear the registry

src/server/compaction.ts

@@ -89,6 +89,8 @@ async function subagentSummarizer(prompt: string, engine: PiEngine): Promise<{ s
 /** Everything the service needs from its host (chat-handler wires the
  *  production versions; tests inject). */
 export interface CompactionDeps {
+	/** Eligibility only for NEW runs; structural waits still apply. */
+	canStart?: () => boolean;
 	chatStore: ChatStore;
 	/** The chat's conversation engine — also the summarizer's. */
 	engine: PiEngine;
@@ -267,6 +269,7 @@ export async function ensureCompacted(deps: CompactionDeps, chatId: string): Pro
 	// injected summarizer.
 	if (process.env.NODE_ENV === "test" && !deps.summarize) return;
 	await awaitCompaction(chatId);
+	if (deps.canStart?.() === false) return;
 	const record = deps.chatStore.getChat(chatId);
 	if (!record) return;
 	if (projectedContextTokens(record.messages, deps.systemPromptTokens) <= deps.thresholdTokens) return;
@@ -291,7 +294,7 @@ export async function ensureCompacted(deps: CompactionDeps, chatId: string): Pro
 export function kickCompaction(buildDeps: () => CompactionDeps | null, chatId: string): void {
 	if (state().inFlight.has(chatId)) return;
 	const deps = buildDeps();
-	if (!deps) return;
+	if (!deps || deps.canStart?.() === false) return;
 	const record = deps.chatStore.getChat(chatId);
 	if (!record) return;
 	if (projectedContextTokens(record.messages, deps.systemPromptTokens) <= deps.thresholdTokens) return;

src/server/desktop-access.ts

@@ -0,0 +1,17 @@
+import { timingSafeEqual } from "node:crypto";
+
+/** Per-launch desktop HTTP capability. This is not an OS-user malware
+ * boundary and is not an OAuth credential. Browser-only mode is unchanged. */
+export function desktopAPIAuthorized(
+	headers: Headers,
+	env: Record<string, string | undefined> = process.env,
+): boolean {
+	const expected = env.LH_DESKTOP_API_SECRET;
+	if (env.LH_DESKTOP_SERVER !== "1" && !expected) return true;
+	if (!expected || expected.length < 64) return false;
+	const supplied = headers.get("x-learnharness-desktop");
+	if (!supplied) return false;
+	const actual = Buffer.from(supplied);
+	const wanted = Buffer.from(expected);
+	return actual.length === wanted.length && timingSafeEqual(actual, wanted);
+}

src/server/jobs.ts

@@ -66,7 +66,9 @@ import {
 import type { TurnInput } from "../core/input-events.ts";
 import { danglingQuestionOf, endsInDismissal, type LhMessage } from "../core/native-messages.ts";
 import { fireJobAbort, isJobLive } from "../agent/subagents/aborts.ts";
-import { stores } from "./stores.ts";
+import { stores, effectiveSettings } from "./stores.ts";
+import { resolveConversationSpec } from "../core/settings-resolution.ts";
+import { chatGPTBackgroundAllowed, onChatGPTChange } from "../agent/chatgpt-bridge.ts";
 import { isTurnRunning } from "./turn-liveness.ts";
 import type { Agent } from "@earendil-works/pi-agent-core";
 
@@ -111,6 +113,8 @@ interface WakeState {
 	pumping: Set<string>;
 	/** Throttle bookkeeping for the stale-job sweep. */
 	lastSweepAt: number;
+	stopConnectionWatch?: () => void;
+	automaticDeferred: WeakSet<Agent>;
 }
 
 interface LhWakeGlobal {
@@ -119,7 +123,7 @@ interface LhWakeGlobal {
 
 function state(): WakeState {
 	const g = globalThis as typeof globalThis & LhWakeGlobal;
-	return (g.__lhWakePipeline ??= { queues: new Map(), liveAgents: new Map(), userSteers: new Map(), pumping: new Set(), lastSweepAt: 0 });
+	return (g.__lhWakePipeline ??= { queues: new Map(), liveAgents: new Map(), userSteers: new Map(), pumping: new Set(), lastSweepAt: 0, automaticDeferred: new WeakSet() });
 }
 
 /** chat-handler wires this at module init: its adapter around handleChatRequest. */
@@ -210,9 +214,18 @@ export function deliverJobResult(steer: SteerMessage): Promise<void> {
  *  beats the loop's synchronous start-poll drain and the marker rides the
  *  run's FIRST LLM call (the native equivalent of the old continuation
  *  coalescing). */
+export function automaticFollowUpAllowed(chatId: string, agent?: Agent): boolean {
+	if (agent && state().automaticDeferred.has(agent)) return false;
+	const provider = agent?.state?.model?.provider ?? resolveConversationSpec(stores.chatStore.getChat(chatId)?.settings, effectiveSettings().models.strong)?.provider;
+	return provider !== "chatgpt" || chatGPTBackgroundAllowed();
+}
+
 function handOffEligible(chatId: string, agent: Agent): void {
 	const queue = state().queues.get(chatId);
-	if (queue && queue.length > 0) {
+	// Keep automatic results in their owning registry. Never mix them into a
+	// foreground batch without express account consent. Learner steers below
+	// remain deliverable, including the first poll of an answer run.
+	if (automaticFollowUpAllowed(chatId, agent) && queue && queue.length > 0) {
 		for (const entry of queue) {
 			if (entry.handedTo !== agent && entry.steer.customType !== "probe-handoff") {
 				agent.steer(entry.steer);
@@ -290,9 +303,25 @@ export function reconcileUserSteers(chatId: string, persisted: readonly LhMessag
 /** Clear the live agent at settlement — MUST run before settleSteers so the
  *  reconciliation's re-pump can never hand a marker to the dying agent. */
 export function unregisterLiveAgent(chatId: string): void {
+	const agent = state().liveAgents.get(chatId);
+	if (agent) state().automaticDeferred.delete(agent);
 	state().liveAgents.delete(chatId);
 }
 
+/** Restore exact registry objects after native drain but before dispatch.
+ * The vehicle is disposable; result ownership/promises are not. */
+export function restoreUncommittedAutomatic(chatId: string, agent: Agent, automaticIds: ReadonlySet<string>, retainedIds: ReadonlySet<string>): void {
+	const s = state();
+	s.automaticDeferred.add(agent);
+	for (const entry of s.queues.get(chatId) ?? []) {
+		if (entry.handedTo === agent && automaticIds.has(entry.steer.id)) entry.handedTo = undefined;
+	}
+	agent.clearSteeringQueue();
+	for (const entry of s.userSteers.get(chatId) ?? []) {
+		if (!retainedIds.has(entry.message.id)) { agent.steer(entry.message); entry.handedTo = agent; }
+	}
+}
+
 /** TURN SETTLEMENT RECONCILIATION (Wave C): the handler calls this AFTER
  * unregisterLiveAgent, with the run's persisted messages. Two jobs:
  *
@@ -419,6 +448,7 @@ async function pump(chatId: string): Promise<void> {
 	const s = state();
 	if (s.pumping.has(chatId)) return;
 	s.pumping.add(chatId);
+	let parked = false;
 	try {
 		for (let attempt = 0; ; ) {
 			const queue = s.queues.get(chatId);
@@ -452,6 +482,7 @@ async function pump(chatId: string): Promise<void> {
 				return;
 			}
 			s.queues.set(chatId, stillPending);
+			if (!automaticFollowUpAllowed(chatId, s.liveAgents.get(chatId))) { parked = true; return; }
 
 			// VEHICLE 1 — THE LIVE RUN: pi's steering queue. The agent of the
 			// streaming turn takes every eligible marker; pi injects at the
@@ -525,7 +556,8 @@ async function pump(chatId: string): Promise<void> {
 			// hands them to that turn once the text-claim lands — until then,
 			// settlement re-pumps and lands them right after it). The pause is
 			// derived from the tail, so any user message un-holds by construction.
-			if (endsInDismissal(fresh.messages)) {
+			if (endsInDismissal(fresh.messages) || !automaticFollowUpAllowed(chatId)) {
+				parked = true;
 				return;
 			}
 
@@ -544,6 +576,15 @@ async function pump(chatId: string): Promise<void> {
 				console.error(`[steer ${steers.map((x) => x.id).join(",")}] wake turn crashed:`, err);
 				return -1;
 			});
+			if (status === 403) {
+				// Consent may have changed after the last gate. Preserve completed
+				// job/result ownership and wait for an explicit connection change.
+				const back = s.queues.get(chatId) ?? [];
+				back.push(...fire);
+				s.queues.set(chatId, back);
+				parked = true;
+				return;
+			}
 			if (status === 409) {
 				const back = s.queues.get(chatId) ?? [];
 				back.push(...fire);
@@ -573,10 +614,30 @@ async function pump(chatId: string): Promise<void> {
 		// a live run are NOT stragglers (settleSteers owns their resurrection)
 		// — restarting for them would just park in waitForIdle forever.
 		const q = state().queues.get(chatId) ?? [];
-		if (q.some((e) => e.handedTo === undefined)) void pump(chatId);
+		if (!parked && automaticFollowUpAllowed(chatId, state().liveAgents.get(chatId)) && q.some((e) => e.handedTo === undefined)) void pump(chatId);
 	}
 }
 
+// Connection changes are the only wake for consent-deferred automatic work.
+// No retry timer, inference, or API-key fallback is started while consent is off.
+state().stopConnectionWatch?.();
+state().stopConnectionWatch = onChatGPTChange(() => {
+	const s = state();
+	for (const [chatId, agent] of s.liveAgents) {
+		if (agent.state?.model?.provider !== "chatgpt" || automaticFollowUpAllowed(chatId, agent)) continue;
+		const consumed = new Set(agent.state.messages.map((m) => (m as { id?: string }).id));
+		const pending = (s.queues.get(chatId) ?? []).filter((entry) => entry.handedTo === agent && !consumed.has(entry.steer.id));
+		if (pending.length === 0) continue;
+		agent.clearSteeringQueue();
+		for (const entry of pending) entry.handedTo = undefined;
+		for (const entry of s.userSteers.get(chatId) ?? []) if (!consumed.has(entry.message.id)) entry.handedTo = undefined;
+		handOffEligible(chatId, agent); // only unconsumed learner input; never replay consumed steers
+	}
+	for (const chatId of s.queues.keys()) {
+		if (automaticFollowUpAllowed(chatId, s.liveAgents.get(chatId))) void pump(chatId);
+	}
+});
+
 /** Flip the job to its final status now that the lesson has resumed. */
 function finalizeJob(jobId: string): void {
 	const job = stores.jobStore.getJob(jobId);

src/server/keys.ts

@@ -52,6 +52,7 @@ import {
 	validateKeyShape,
 } from "../core/auth-keys.ts";
 import { authStore } from "../store/auth.ts";
+import { chatGPTConnected } from "../agent/chatgpt-bridge.ts";
 
 // The once-per-process boot record — globalThis anchor (HMR re-imports the
 // module; the snapshot must not re-run and misread fill-gapped vars as real
@@ -200,6 +201,7 @@ export function keyValue(id: string): string | undefined {
  * (gate by what would run). Providers absent from AUTH_PROVIDER_KEYS take
  * no key at all: nothing to gate, present by definition. */
 export function providerKeyPresent(provider: string): boolean {
+	if (provider === "chatgpt") return chatGPTConnected();
 	const info = AUTH_PROVIDER_KEYS.find((p) => p.id === provider);
 	return info ? keyValue(info.envVar) !== undefined : true;
 }

src/server/model-catalog.ts

@@ -20,6 +20,15 @@
 import { clampThinkingLevel, getSupportedThinkingLevels, type Api, type Model } from "@earendil-works/pi-ai";
 import type { ThinkingLevel } from "@earendil-works/pi-agent-core";
 import { piModels } from "../agent/models.ts";
+import { chatGPTConnected, chatGPTModel, chatGPTModels, refreshChatGPTModels } from "../agent/chatgpt-bridge.ts";
+
+export async function refreshConnectionCatalog(): Promise<void> {
+	await refreshChatGPTModels(piModels().getModels("openai").filter((m): m is Model<"openai-responses"> => m.api === "openai-responses"));
+}
+function connectionModel(provider: string, id: string): Model<Api> | undefined {
+	return provider === "chatgpt" ? chatGPTModel(id) : piModels().getModel(provider, id);
+}
+function connectionProviderIds(): string[] { return [...piModels().getProviders().map((p) => p.id), "chatgpt"]; }
 import { settingsStore } from "../store/settings.ts";
 import { effectiveSettings } from "./stores.ts";
 import { keyValue } from "./keys.ts";
@@ -125,7 +134,7 @@ export function catalogProviders(): CatalogProvider[] {
 		id: p.id,
 		envVar: p.envVar,
 		keyPresent: keyValue(p.envVar) !== undefined,
-	})).sort((a, b) => a.id.localeCompare(b.id));
+	})).concat([{ id: "chatgpt", envVar: "", keyPresent: chatGPTConnected() }]).sort((a, b) => a.id.localeCompare(b.id));
 }
 
 /** The providers whose key is present — the picker federates exactly
@@ -171,6 +180,7 @@ export function catalogContextLimits(): Map<string, number> {
 			index.set(`${provider.id}:${model.id}`, model.contextWindow);
 		}
 	}
+	for (const model of chatGPTModels()) index.set(`chatgpt:${model.id}`, model.contextWindow);
 	return index;
 }
 
@@ -185,7 +195,8 @@ export function modelCatalog(): ModelCatalogResponse {
 		.filter((p) => keyed.has(p.id))
 		.flatMap((p) => piModels().getModels(p.id))
 		.map(toModelDto)
-		.sort((a, b) => (a.provider === b.provider ? a.id.localeCompare(b.id) : a.provider.localeCompare(b.provider)));
+		.sort((a, b) => (a.provider === b.provider ? a.id.localeCompare(b.id) : a.provider.localeCompare(b.provider)))
+		.concat(chatGPTModels().map(toModelDto));
 	return {
 		models,
 		defaults: catalogDefaults(),
@@ -207,13 +218,13 @@ export type SettingsValidation =
  *    the model the picker writes against (settings.model, else config's
  *    prober/teacher pin; one controller, so the two specs share it). */
 export function validateChatSettings(settings: ChatSettings): SettingsValidation {
-	const providers = piModels().getProviders().map((p) => p.id);
+	const providers = connectionProviderIds();
 	if (settings.model) {
 		const { provider, id } = settings.model;
 		if (!providers.includes(provider)) {
 			return { ok: false, error: `Unknown provider "${provider}" — configured: ${providers.join(", ")}` };
 		}
-		if (!piModels().getModel(provider, id)) {
+		if (!connectionModel(provider, id)) {
 			return { ok: false, error: `No model "${id}" in the ${provider} catalog` };
 		}
 	}
@@ -237,13 +248,13 @@ export function validateModelsSettings(settings: ModelsSettings): SettingsValida
 }
 
 function validateSlotSpec(name: ModelSlotName, slot: { model?: { provider: string; id: string }; thinkingLevel?: ThinkingLevel }): SettingsValidation {
-	const providers = piModels().getProviders().map((p) => p.id);
+	const providers = connectionProviderIds();
 	if (slot.model) {
 		const { provider, id } = slot.model;
 		if (!providers.includes(provider)) {
 			return { ok: false, error: `Unknown provider "${provider}" — configured: ${providers.join(", ")}` };
 		}
-		if (!piModels().getModel(provider, id)) {
+		if (!connectionModel(provider, id)) {
 			return { ok: false, error: `No model "${id}" in the ${provider} catalog` };
 		}
 	}
@@ -265,7 +276,7 @@ function validateSlotSpec(name: ModelSlotName, slot: { model?: { provider: strin
  *  or the level already supported — the existence gate rejects unknown
  *  models separately, so this only ever sees real ones in practice). */
 function clampLevelFor(provider: string, id: string, level: ThinkingLevel): ThinkingLevel {
-	const model = piModels().getModel(provider, id);
+	const model = connectionModel(provider, id);
 	return model ? clampThinkingLevel(model, level) : level;
 }
 

src/server/pi/ui-stream.ts

@@ -2,6 +2,7 @@ import type { Agent, AgentEvent, AgentMessage } from "@earendil-works/pi-agent-c
 import type { SteerMessage } from "../../core/contracts.ts";
 import { noteSteerVisible } from "../jobs.ts";
 import { ToolCallIdUniquifier } from "./call-ids.ts";
+import type { ChatGPTRequestCycle } from "../../agent/chatgpt-request-cycle.ts";
 
 /**
  * THE NATIVE WIRE (Wave D, owner-approved 2026-09-26): one agent turn, streamed
@@ -71,6 +72,7 @@ export type TurnWireEvent =
 	| { type: "error"; errorText: string };
 
 export interface PiTurnSpec {
+	requestCycle?: ChatGPTRequestCycle;
 	/** The per-turn agent (stateless engine — built from the record). */
 	agent: Agent;
 	/** The handler's error-masking policy for LOOP-level failures (the
@@ -117,7 +119,7 @@ export interface PiTurnHandle {
 }
 
 export function runPiTurn(spec: PiTurnSpec): PiTurnHandle {
-	const { agent, onError, onFinish, usedToolCallIds, updateFlushMs = 100 } = spec;
+	const { agent, onError, onFinish, usedToolCallIds, updateFlushMs = 100, requestCycle } = spec;
 	const uniquifier = usedToolCallIds ? new ToolCallIdUniquifier(usedToolCallIds) : undefined;
 
 	let settled = false;
@@ -248,11 +250,24 @@ export function runPiTurn(spec: PiTurnSpec): PiTurnHandle {
 			settleWith();
 		}
 	};
-	agent.subscribe(handle);
+	requestCycle?.bindRelease(handle);
+	agent.subscribe((event) => {
+		if (!requestCycle?.hold(event)) handle(event);
+	});
 
 	void (async () => {
 		try {
-			await agent.continue();
+			do {
+				await agent.continue();
+				if (!requestCycle) break;
+				const denied = requestCycle.deniedBeforeDispatch;
+				const outcome = requestCycle.finishNativeCycle();
+				if (denied) agent.state.messages = outcome.messages;
+				if (outcome.action !== "resumeForeground") {
+					handle({ type: "agent_end", messages: [...agent.state.messages] });
+					break;
+				}
+			} while (!sawAbort);
 		} catch (err) {
 			// A loop-level failure (not a provider error — those ride the
 			// stream as stopReason "error"): surface, close, persist. If the
@@ -318,6 +333,7 @@ export function runPiTurn(spec: PiTurnSpec): PiTurnHandle {
 		abort() {
 			if (settled) return;
 			sawAbort = true;
+			requestCycle?.abort();
 			agent.abort();
 		},
 	};

tests/chatgpt-auth.test.ts

@@ -0,0 +1,343 @@
+import { test } from "node:test";
+import assert from "node:assert/strict";
+import { createRequire } from "node:module";
+import { createHash } from "node:crypto";
+import fs from "node:fs/promises";
+import os from "node:os";
+import path from "node:path";
+import { exportJWK, generateKeyPair, SignJWT } from "jose";
+const require = createRequire(import.meta.url);
+const { createChatGPT, encryptedStore, allowedAuthEndpoint, SCOPES } = require("../electron/chatgpt.cjs");
+const { privateKey, publicKey } = await generateKeyPair("RS256");
+const jwk = { ...await exportJWK(publicKey), kid: "scratch-key", alg: "RS256", use: "sig" };
+const tokenUrl = "https://auth.openai.com/api/accounts/oauth/token";
+const initial = () => ({ version: 1, hostId: "urn:uuid:00000000-0000-4000-8000-000000000001", activeId: null as string | null, registrations: [] as any[] });
+async function signed(client: string, nonce: string, subject = "scratch-user", overrides: Record<string, unknown> = {}) {
+	return new SignJWT({ nonce, email: "dummy@example.invalid", ...overrides }).setProtectedHeader({ alg: "RS256", kid: "scratch-key" }).setIssuer("https://auth.openai.com").setSubject(subject).setAudience(client).setIssuedAt().setExpirationTime("1h").sign(privateKey);
+}
+function deferred<T = void>() { let resolve!: (value: T) => void; const promise = new Promise<T>((done) => { resolve = done; }); return { resolve, promise }; }
+function fixture(options: { state?: any; expires?: number; timeoutMs?: number; token?: (form: URLSearchParams, url: URL) => Promise<Response>; write?: (next: any) => Promise<void>; secure?: () => void; change?: (push: any) => void; provider?: (target: string, init: RequestInit) => Promise<Response | undefined> } = {}) {
+	let saved = structuredClone(options.state ?? initial());
+	let url!: URL;
+	let writeCount = 0;
+	const changed = deferred();
+	const requests: { url: string; form?: URLSearchParams }[] = [];
+	const service = createChatGPT({
+		store: { secure() { options.secure?.(); }, async read() { return structuredClone(saved); }, async write(next: any) { writeCount++; await options.write?.(next); saved = structuredClone(next); } },
+		openExternal: async (value: string) => { url = new URL(value); },
+		timeoutMs: options.timeoutMs ?? 5000,
+		onChange: (push: any) => { options.change?.(push); if (!push.status.pending) changed.resolve(); },
+		fetch: async (target: string, init: RequestInit = {}) => {
+			const form = init.body instanceof URLSearchParams ? init.body : undefined;
+			requests.push({ url: target, form });
+			const response = await options.provider?.(target, init); if (response) return response;
+			if (target.endsWith("jwks.json")) return Response.json({ keys: [jwk] });
+			if (target === tokenUrl) {
+				if (options.token) return options.token(form!, url);
+				return Response.json({ access_token: "scratch-oauth-token", refresh_token: "scratch-refresh", id_token: await signed(form!.get("client_id")!, url.searchParams.get("nonce")!), token_type: "Bearer", expires_in: options.expires ?? 3600, scope: SCOPES.join(" ") });
+			}
+			if (target.endsWith("openid-configuration")) return Response.json({ issuer: "https://auth.openai.com", revocation_endpoint: "https://auth.openai.com/api/accounts/oauth/revoke" });
+			if (target.endsWith("/revoke")) return new Response(null, { status: 200 });
+			if (target.endsWith("/models")) return Response.json({ models: [{ slug: "gpt-5", display_name: "Scratch model", visibility: "list" }, { slug: "hidden", display_name: "Hidden", visibility: "hide" }] });
+			throw new Error("Unexpected network request in scratch fixture");
+		},
+	});
+	const rawLease = service.lease;
+	let discovered: Promise<any> | undefined;
+	service.lease = async (background: boolean) => {
+		discovered ??= service.models();
+		const catalog = await discovered;
+		return rawLease(background, { expectedAccountId: catalog.accountId ?? saved.activeId ?? "scratch-missing", expectedAuthorityRevision: catalog.authorityRevision, selectedModelId: "gpt-5" });
+	};
+	return { service, rawLease, requests, changed, get url() { return url; }, get saved() { return saved; }, get writeCount() { return writeCount; }, async callback(params: Record<string, string> = {}) {
+		const callback = new URL(url.searchParams.get("redirect_uri")!);
+		callback.search = new URLSearchParams({ state: url.searchParams.get("state")!, code: "scratch-code", client_id: "oaiapp_scratch", ...params }).toString();
+		return fetch(callback);
+	} };
+}
+async function loggedIn(options: Parameters<typeof fixture>[0] = {}) {
+	const f = fixture(options);
+	await f.service.login();
+	await f.callback();
+	await f.changed.promise;
+	return f;
+}
+
+test("PKCE/state/nonce/host identity and exact ephemeral 127.0.0.1 callback; no secret renderer snapshot", async () => {
+	const f = fixture();
+	try {
+		await f.service.login();
+		const params = f.url.searchParams;
+		assert.equal(params.get("client_id"), "dynamic_agent_client");
+		assert.equal(params.get("agent_name_hint"), "LearnHarness");
+		assert.equal(params.get("ext_agent_host_id"), f.saved.hostId);
+		assert.equal(new URL(params.get("redirect_uri")!).hostname, "127.0.0.1");
+		assert.equal(new URL(params.get("redirect_uri")!).pathname, "/auth/callback");
+		assert.equal((await f.callback({ state: "wrong" })).status, 400);
+		assert.equal(f.requests.length, 0, "unverified callback must never exchange a code");
+		await f.callback(); await f.changed.promise;
+		const form = f.requests.find((r) => r.url === tokenUrl)!.form!;
+		assert.equal(form.get("redirect_uri"), params.get("redirect_uri"));
+		assert.equal(form.get("client_id"), "oaiapp_scratch");
+		assert.equal(createHash("sha256").update(form.get("code_verifier")!).digest("base64url"), params.get("code_challenge"));
+		assert.equal(form.get("resource"), "https://api.openai.com/v1");
+		const status = await f.service.status();
+		assert.equal(status.accounts[0].background, false);
+		assert.equal(status.activeId, f.saved.activeId);
+		assert(!JSON.stringify(status).includes("scratch-oauth-token"));
+		assert(!JSON.stringify(status).includes("scratch-refresh"));
+		assert(!JSON.stringify(status).includes("idToken"));
+		assert.equal((await f.service.lease(false)).accessToken, "scratch-oauth-token");
+		await assert.rejects(f.service.lease(true), /background/);
+	} finally { f.service.close(); }
+});
+
+test("returning sign-in reuses issued client/host and fresh PKCE; rejects account/client mismatch without replacing active credentials", async () => {
+	const f = await loggedIn();
+	try {
+		const first = f.url;
+		const status = await f.service.status();
+		await f.service.login(status.activeId);
+		assert.equal(f.url.searchParams.get("client_id"), "oaiapp_scratch");
+		assert.equal(f.url.searchParams.get("agent_name_hint"), null);
+		assert(f.url.searchParams.get("id_token_hint"));
+		assert.notEqual(f.url.searchParams.get("state"), first.searchParams.get("state"));
+		assert.notEqual(f.url.searchParams.get("nonce"), first.searchParams.get("nonce"));
+		assert.equal(f.url.searchParams.get("ext_agent_host_id"), first.searchParams.get("ext_agent_host_id"));
+		await f.callback({ client_id: "oaiapp_other" });
+		await new Promise((resolve) => setImmediate(resolve));
+		assert.equal(f.saved.registrations.length, 1);
+		assert.equal((await f.service.lease(false)).accessToken, "scratch-oauth-token");
+	} finally { f.service.close(); }
+});
+
+for (const [label, claims] of [["nonce", { nonce: "wrong" }], ["issuer", { iss: "https://evil.invalid" }], ["audience", { aud: "oaiapp_wrong" }], ["expiry", { exp: 1 }]] as const) {
+	test(`rejects invalid ID-token ${label}`, async () => {
+		const f = fixture({ token: async (form, url) => {
+			const jwt = new SignJWT(Object.assign({ nonce: url.searchParams.get("nonce")!, email: "dummy@example.invalid", iss: "https://auth.openai.com", aud: form.get("client_id")!, sub: "scratch-user", iat: Math.floor(Date.now() / 1000), exp: Math.floor(Date.now() / 1000) + 3600 }, claims));
+			return Response.json({ access_token: "scratch-access", refresh_token: "scratch-refresh", id_token: await jwt.setProtectedHeader({ alg: "RS256", kid: "scratch-key" }).sign(privateKey), token_type: "Bearer", expires_in: 3600, scope: SCOPES.join(" ") });
+		} });
+		try { await f.service.login(); await f.callback(); await f.changed.promise; assert.equal(f.saved.activeId, null); assert.equal(f.saved.registrations.length, 0); }
+		finally { f.service.close(); }
+	});
+}
+
+test("cancel and timeout close the listener without exchanging or replacing an active account", async () => {
+	const f = fixture({ timeoutMs: 30 });
+	try {
+		await f.service.login(); const callback = f.url.searchParams.get("redirect_uri")!;
+		f.service.cancelLogin();
+		await assert.rejects(fetch(callback));
+		assert.equal(f.saved.activeId, null);
+		const timeout = fixture({ timeoutMs: 15 });
+		try {
+			await timeout.service.login();
+			await timeout.changed.promise;
+			assert.equal((await timeout.service.status()).pending, false);
+			assert.equal(timeout.requests.length, 0);
+		} finally { timeout.service.close(); }
+		assert.equal(f.requests.length, 0);
+	} finally { f.service.close(); }
+});
+
+test("missing plan scope retains verified registration but refuses inference and catalog", async () => {
+	const f = await loggedIn({ token: async (form, url) => Response.json({ id_token: await signed(form.get("client_id")!, url.searchParams.get("nonce")!), scope: "openid profile email", token_type: "Bearer" }) });
+	try {
+		assert.equal(f.saved.registrations.length, 1);
+		assert.equal((await f.service.status()).accounts[0].planEnabled, false);
+		await assert.rejects(f.service.lease(false), /scope/);
+		assert.deepEqual((await f.service.models()).models, []);
+	} finally { f.service.close(); }
+});
+
+test("durable encrypted save failure never activates a pending login", async () => {
+	const f = fixture({ write: async () => { throw new Error("scratch disk failure containing no real data"); } });
+	try {
+		await f.service.login(); await f.callback(); await f.changed.promise;
+		assert.equal(f.saved.activeId, null);
+		await assert.rejects(f.service.lease(false), /disconnected/);
+	} finally { f.service.close(); }
+});
+
+test("serialized concurrent refresh rotates once and persists before leases", async () => {
+	let refreshes = 0;
+	const f = await loggedIn({ expires: 1, token: async (form, url) => {
+		if (form.get("grant_type") === "refresh_token") { refreshes++; assert.equal(form.get("scope"), null); return Response.json({ access_token: "scratch-new-access", refresh_token: "scratch-rotated", token_type: "Bearer", expires_in: 3600 }); }
+		return Response.json({ access_token: "scratch-old-access", refresh_token: "scratch-old-refresh", id_token: await signed(form.get("client_id")!, url.searchParams.get("nonce")!), token_type: "Bearer", expires_in: 1, scope: SCOPES.join(" ") });
+	} });
+	try {
+		const leases = await Promise.all([f.service.lease(false), f.service.lease(false), f.service.lease(false)]);
+		assert.equal(refreshes, 1);
+		assert(leases.every((lease) => lease.accessToken === "scratch-new-access"));
+		assert.equal(f.saved.registrations[0].tokens.refreshToken, "scratch-rotated");
+	} finally { f.service.close(); }
+});
+
+test("logout wins an in-flight refresh and retains only account/client/host mapping", async () => {
+	const started = deferred(); const release = deferred();
+	const f = await loggedIn({ token: async (form, url) => {
+		if (form.get("grant_type") === "refresh_token") { started.resolve(); await release.promise; return Response.json({ access_token: "scratch-late", refresh_token: "scratch-rotated", token_type: "Bearer", expires_in: 3600 }); }
+		return Response.json({ access_token: "scratch-old", refresh_token: "scratch-refresh", id_token: await signed(form.get("client_id")!, url.searchParams.get("nonce")!), token_type: "Bearer", expires_in: 1, scope: SCOPES.join(" ") });
+	} });
+	try {
+		const id = f.saved.activeId;
+		const refresh = f.service.lease(false); const rejected = assert.rejects(refresh);
+		await started.promise;
+		const logout = f.service.logout(id);
+		await new Promise((resolve) => setImmediate(resolve)); release.resolve();
+		await rejected; await logout;
+		assert.equal(f.saved.activeId, null);
+		assert.equal(f.saved.registrations[0].clientId, "oaiapp_scratch");
+		assert.equal(f.saved.registrations[0].tokens, undefined);
+		assert(f.requests.some((r) => r.form?.get("token_type_hint") === "refresh_token"));
+		await f.service.login(id);
+		assert.equal(f.url.searchParams.get("id_token_hint"), null);
+		assert.equal(f.url.searchParams.get("client_id"), "oaiapp_scratch");
+	} finally { f.service.close(); }
+});
+
+test("account catalog is scoped, visibility filtered, and cleared on disconnect", async () => {
+	const f = await loggedIn();
+	try {
+		assert.deepEqual((await f.service.models()).models, [{ id: "gpt-5", name: "Scratch model" }]);
+		await f.service.logout(f.saved.activeId);
+		assert.deepEqual((await f.service.models()).models, []);
+	} finally { f.service.close(); }
+});
+
+test("safeStorage refuses basic_text/unavailable; encrypted file owner-only, atomic, no plaintext fallback", async () => {
+	const dir = await fs.mkdtemp(path.join(os.tmpdir(), "lh-chatgpt-storage-"));
+	const file = path.join(dir, "credentials.enc.json");
+	const secure = { isEncryptionAvailable: () => true, getSelectedStorageBackend: () => "keychain", encryptString: (s: string) => Buffer.from(s.split("").reverse().join("")), decryptString: (b: Buffer) => b.toString().split("").reverse().join("") };
+	try {
+		const store = encryptedStore(file, secure);
+		const data = { ...initial(), marker: "scratch-secret-never-plaintext" };
+		await store.write(data);
+		assert(!(await fs.readFile(file, "utf8")).includes(data.marker));
+		assert.deepEqual(await store.read(), data);
+		assert.equal((await fs.stat(file)).mode & 0o777, 0o600);
+		assert.deepEqual(await fs.readdir(dir), ["credentials.enc.json"]);
+		for (const backend of [{ ...secure, getSelectedStorageBackend: () => "basic_text" }, { ...secure, isEncryptionAvailable: () => false }]) {
+			await assert.rejects(encryptedStore(file, backend).read(), /secure_storage/);
+			await assert.rejects(encryptedStore(file, backend).write(data), /secure_storage/);
+		}
+		assert(!allowedAuthEndpoint("https://auth.openai.com.evil.invalid/revoke"));
+		assert(!allowedAuthEndpoint("http://127.0.0.1/revoke"));
+		assert(!allowedAuthEndpoint("https://auth.openai.com@evil.invalid/revoke"));
+	} finally { await fs.rm(dir, { recursive: true, force: true }); }
+});
+
+test("invalid rotating replacement is durably revocation-only; verified mapping and prior validated ID survive restart", async () => {
+	const f = await loggedIn({ expires: 1, token: async (form, url) => form.get("grant_type") === "refresh_token"
+		? Response.json({ access_token: "never-save-this-access", refresh_token: "likely-latest-revoke-handle", id_token: "invalid.signature.jwt", token_type: "Bearer", expires_in: 3600 })
+		: Response.json({ access_token: "old-verified-access", refresh_token: "old-refresh", id_token: await signed(form.get("client_id")!, url.searchParams.get("nonce")!), token_type: "Bearer", expires_in: 1, scope: SCOPES.join(" ") }) });
+	try {
+		const before = structuredClone(f.saved.registrations[0]);
+		await assert.rejects(f.service.lease(false));
+		const restricted = f.saved.registrations[0];
+		assert.equal(restricted.reconnectRequired, true);
+		for (const field of ["id", "issuer", "subject", "email", "clientId"]) assert.equal(restricted[field], before[field]);
+		assert.deepEqual(restricted.tokens, { ...before.tokens, refreshToken: "likely-latest-revoke-handle" });
+		assert(!JSON.stringify(f.saved).includes("never-save-this-access"));
+		const restarted = fixture({ state: f.saved });
+		try {
+			await restarted.service.status();
+			await restarted.service.select(restricted.id);
+			restarted.service.resume(); await restarted.service.consent(restricted.id, true);
+			await assert.rejects(restarted.service.lease(false));
+			assert.equal((await restarted.service.status()).accounts[0].connected, false);
+			assert.deepEqual((await restarted.service.models()).models, []);
+			assert.equal(restarted.requests.filter((r) => r.url === tokenUrl).length, 0, "restricted material cannot refresh for inference/catalog");
+			await restarted.service.logout(restricted.id);
+			assert.equal(restarted.requests.find((r) => r.form?.has("token_type_hint"))!.form!.get("token"), "likely-latest-revoke-handle");
+			assert.equal(restarted.saved.registrations[0].reconnectRequired, true);
+			await restarted.service.login(restricted.id); await restarted.callback(); await restarted.changed.promise;
+			// Wait for callback's asynchronous encrypted activation, not the prior status push.
+			for (let i = 0; i < 30 && restarted.saved.registrations[0].reconnectRequired; i++) await new Promise((resolve) => setImmediate(resolve));
+			assert.equal(restarted.saved.registrations[0].reconnectRequired, undefined);
+		} finally { restarted.service.close(); }
+	} finally { f.service.close(); }
+});
+
+test("detected secure backend loss emits full invalidation, denies all leases and cannot be repaired by consent/resume", async () => {
+	let lost = false; const pushes: any[] = [];
+	const f = await loggedIn({ secure: () => { if (lost) throw Object.assign(new Error("secure_storage"), { code: "secure_storage" }); }, change: (push) => pushes.push(push) });
+	try {
+		await f.service.lease(false); const epoch = f.service.stamp().epoch;
+		lost = true;
+		assert.equal((await f.service.status()).available, false);
+		assert(pushes.some((p) => p.invalidate === "all" && p.stamp.epoch > epoch));
+		await assert.rejects(f.service.lease(false)); await assert.rejects(f.service.lease(true));
+		lost = false; f.service.resume(); await f.service.consent(f.saved.activeId, true);
+		await assert.rejects(f.service.lease(false));
+		assert.equal((await f.service.status()).accounts[0].connected, false);
+	} finally { f.service.close(); }
+});
+
+test("strict encrypted schema rejects malformed marker/duplicates/token/active identity without changing original bytes", async () => {
+	const f = await loggedIn(); f.service.close();
+	const dir = await fs.mkdtemp(path.join(os.tmpdir(), "lh-chatgpt-corrupt-"));
+	const file = path.join(dir, "credentials.enc.json");
+	const secure = { isEncryptionAvailable: () => true, encryptString: (s: string) => Buffer.from(s), decryptString: (b: Buffer) => b.toString() };
+	try {
+		for (const mutate of [(s: any) => s.registrations[0].reconnectRequired = false, (s: any) => s.registrations[0].reconnectRequired = "true", (s: any) => s.registrations.push(s.registrations[0]), (s: any) => s.activeId = "missing", (s: any) => s.registrations[0].tokens.scopes = "not-array", (s: any) => s.registrations[0].tokens.expiresAt = -1]) {
+			const state = structuredClone(f.saved); mutate(state);
+			const bytes = JSON.stringify({ version: 1, encrypted: Buffer.from(JSON.stringify(state)).toString("base64") });
+			await fs.writeFile(file, bytes);
+			const service = createChatGPT({ store: encryptedStore(file, secure), openExternal: () => assert.fail("no OAuth"), fetch: () => assert.fail("no provider") });
+			try { assert.equal((await service.status()).available, false); await assert.rejects(service.login()); assert.equal(await fs.readFile(file, "utf8"), bytes); }
+			finally { service.close(); }
+		}
+	} finally { await fs.rm(dir, { recursive: true, force: true }); }
+});
+
+test("consent OFF wins an older ON save; foreground catalog authority remains unchanged", async () => {
+	let hold = false; const entered = deferred(); const release = deferred();
+	const f = await loggedIn({ write: async (next) => { if (hold && next.registrations[0]?.background) { entered.resolve(); await release.promise; } } });
+	try {
+		const catalog = await f.service.models(); const id = f.saved.activeId;
+		hold = true;
+		const on = f.service.consent(id, true); await entered.promise;
+		const off = f.service.consent(id, false);
+		assert.equal((await f.service.status()).accounts[0].background, false);
+		release.resolve(); await on; await off;
+		assert.equal(f.saved.registrations[0].background, false);
+		assert.equal(f.service.stamp().authorityRevision, catalog.authorityRevision);
+		assert.equal((await f.rawLease(false, { expectedAccountId: id, expectedAuthorityRevision: catalog.authorityRevision, selectedModelId: "gpt-5" })).accountId, id);
+		await assert.rejects(f.rawLease(true, { expectedAccountId: id, expectedAuthorityRevision: catalog.authorityRevision, selectedModelId: "gpt-5" }), /background/);
+	} finally { f.service.close(); }
+});
+
+test("cancellation during verified durable save cannot activate late credentials", async () => {
+	const entered = deferred(); const release = deferred(); let blocked = true;
+	const f = fixture({ write: async (next) => { if (blocked && next.registrations.length) { entered.resolve(); await release.promise; blocked = false; } } });
+	try {
+		await f.service.login(); await f.callback(); await entered.promise;
+		f.service.cancelLogin(); release.resolve();
+		for (let i = 0; i < 30 && f.writeCount < 2; i++) await new Promise((resolve) => setImmediate(resolve));
+		assert.equal(f.saved.activeId, null);
+		assert.equal((await f.service.status()).accounts.length, 0);
+	} finally { f.service.close(); }
+});
+
+test("scope-disabled credentials remain revocable and unconfirmed revoke never claims success", async () => {
+	const f = await loggedIn({ token: async (form, url) => Response.json({ id_token: await signed(form.get("client_id")!, url.searchParams.get("nonce")!), refresh_token: "scope-disabled-refresh", scope: "openid profile email", token_type: "Bearer" }), provider: async (url) => url.endsWith("/revoke") ? new Response(null, { status: 500 }) : undefined });
+	try {
+		await assert.rejects(f.service.lease(false));
+		const status = await f.service.logout(f.saved.activeId);
+		assert.match(status.error, /Remote revocation was not confirmed/);
+		assert.equal(f.requests.filter((r) => r.form?.get("token") === "scope-disabled-refresh").length, 2);
+		assert.equal(f.saved.registrations[0].tokens, undefined);
+	} finally { f.service.close(); }
+});
+
+test("private leases bind active account, authority revision and raw visible model membership", async () => {
+	const f = await loggedIn();
+	try {
+		const catalog = await f.service.models();
+		const expected = { expectedAccountId: catalog.accountId, expectedAuthorityRevision: catalog.authorityRevision, selectedModelId: "gpt-5" };
+		await assert.rejects(f.rawLease(false));
+		for (const mutation of [{ expectedAccountId: "another" }, { expectedAuthorityRevision: catalog.authorityRevision - 1 }, { selectedModelId: "hidden" }]) await assert.rejects(f.rawLease(false, { ...expected, ...mutation }));
+		assert.equal((await f.rawLease(false, expected)).accountId, catalog.accountId);
+	} finally { f.service.close(); }
+});

tests/chatgpt-background.test.ts

@@ -0,0 +1,80 @@
+import { test, after } from "node:test";
+import assert from "node:assert/strict";
+import fs from "node:fs";
+import os from "node:os";
+import path from "node:path";
+import { Agent, type AgentMessage } from "@earendil-works/pi-agent-core";
+import { AssistantMessageEventStream } from "@earendil-works/pi-ai/utils/event-stream";
+import { piModels } from "../src/agent/models.ts";
+import { toLlmMessages } from "../src/core/native-messages.ts";
+import { buildSubagentSteer } from "../src/core/contracts.ts";
+const dir = fs.mkdtempSync(path.join(os.tmpdir(), "lh-chatgpt-background-"));
+process.env.LH_DATA_DIR = path.join(dir, "data");
+process.env.LH_AUTH_FILE = path.join(dir, "scratch-auth.json");
+fs.writeFileSync(process.env.LH_AUTH_FILE, "{}");
+const { stores } = await import("../src/server/stores.ts");
+const { beginTurn, endTurn } = await import("../src/server/turn-liveness.ts");
+const { deliverJobResult, registerLiveAgent, unregisterLiveAgent, setTurnRunner, settleSteers, steerLiveChat } = await import("../src/server/jobs.ts");
+const { persistTurn } = await import("../src/server/turn-service.ts");
+const ipc = (value: unknown) => (process.emit as (name: string, value: unknown) => boolean)("message", value);
+let revision = 1;
+function connection(background: boolean) {
+	ipc({ channel: "lh:chatgpt", kind: "state", stamp: { epoch: 1, authorityRevision: 1, revision: revision++ }, invalidate: background ? "none" : "background", status: { available: true, pending: false, activeId: "scratch-account", accounts: [{ id: "scratch-account", connected: true, planEnabled: true, background }], error: null } });
+}
+after(() => { fs.rmSync(dir, { recursive: true, force: true }); });
+
+for (const revokedAfterQueue of [false, true]) {
+	test(`mixed learner+automatic batch: foreground runs, automatic result stays owned/deferred without consent (${revokedAfterQueue ? "revoked after handoff" : "initially off"})`, { timeout: 5000 }, async () => {
+		connection(revokedAfterQueue);
+		const chat = stores.chatStore.createChat("Scratch mixed queue");
+		stores.chatStore.saveSettings(chat.id, { model: { provider: "chatgpt", id: "gpt-5" } });
+		const job = stores.jobStore.createJob({ kind: "research", chatId: chat.id, toolCallId: "scratch-job-call" });
+		const result = { brief: "Scratch completed job result", sources: [] };
+		stores.jobStore.updateJob(job.id, { status: "running", result });
+		const marker = buildSubagentSteer(job, result);
+		const contexts: string[] = [];
+		const native = piModels().getModel("openai", "gpt-5")!;
+		const model = { ...native, provider: "chatgpt" };
+		const agent = new Agent({ initialState: { model, messages: [{ role: "user", content: "Initial scratch learner input", timestamp: 1, id: "scratch-initial" } as AgentMessage] }, convertToLlm: toLlmMessages,
+			streamFn: () => {
+				contexts.push(JSON.stringify(agent.state.messages));
+				const stream = new AssistantMessageEventStream();
+				const output = { role: "assistant" as const, provider: "chatgpt", api: model.api, model: model.id, content: [{ type: "text" as const, text: "Scratch foreground response" }], stopReason: "stop" as const, timestamp: 2, usage: { input: 2, output: 1, cacheRead: 0, cacheWrite: 0, totalTokens: 3, cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 } } };
+				stream.push({ type: "start", partial: output }); stream.push({ type: "done", reason: "stop", message: output }); stream.end();
+				return stream;
+			},
+		});
+		let wakes = 0;
+		setTurnRunner(async (chatId, input) => {
+			wakes++;
+			assert.equal(input.type, "steers");
+			if (input.type === "steers") stores.chatStore.saveMessages(chatId, [...stores.chatStore.getChat(chatId)!.messages, ...input.steers]);
+			return 200;
+		});
+		assert(beginTurn(chat.id));
+		registerLiveAgent(chat.id, agent);
+		const delivered = deliverJobResult(marker);
+		const learner = { id: "scratch-live-learner", role: "user", content: "Please continue my learning", timestamp: 3 } as never;
+		assert(steerLiveChat(chat.id, learner));
+		if (revokedAfterQueue) connection(false);
+		await agent.continue();
+		assert.equal(contexts.length, 1, "ordinary foreground tool-free call remains usable");
+		assert(contexts[0]!.includes("Please continue my learning"));
+		assert(!contexts[0]!.includes(marker.id), "automatic marker must not enter the foreground batch");
+		assert(!agent.hasQueuedMessages());
+		// Ordinary lease/catalog replies are not consent changes and must never
+		// reset/replay a learner steer that the agent has already consumed.
+		ipc({ channel: "lh:chatgpt", id: "ordinary-reply-with-no-pending-call", value: { accessToken: "scratch-only" } });
+		assert(!agent.hasQueuedMessages());
+		const persisted = persistTurn(stores.chatStore, chat.id, 0, agent.state.messages, false);
+		unregisterLiveAgent(chat.id); endTurn(chat.id); settleSteers(chat.id, persisted);
+		await new Promise((resolve) => setImmediate(resolve));
+		assert.equal(wakes, 0, "no background spin or spend while consent is absent");
+		assert.equal(stores.jobStore.getJob(job.id)!.status, "running", "delivery ownership must not be finalized or failed by deferral");
+		assert.deepEqual(stores.jobStore.getJob(job.id)!.result, result);
+		connection(true);
+		await delivered;
+		assert.equal(wakes, 1, "explicit consent resumes the preserved result exactly once");
+		assert.equal(stores.jobStore.getJob(job.id)!.status, "done");
+	});
+}

tests/chatgpt-catalog.test.ts

@@ -0,0 +1,48 @@
+import { test, before, after } from "node:test";
+import assert from "node:assert/strict";
+import { piModels } from "../src/agent/models.ts";
+import { refreshChatGPTModels, chatGPTModels, chatGPTStatus, chatGPTScope, chatGPTScopeCurrent, chatGPTLease, chatGPTAbortController } from "../src/agent/chatgpt-bridge.ts";
+const send = Object.getOwnPropertyDescriptor(process, "send"), connected = Object.getOwnPropertyDescriptor(process, "connected");
+const pending: any[] = [];
+const model = piModels().getModels("openai").find((m) => m.api === "openai-responses")!;
+const ipc = (m: unknown) => (process.emit as (name: string, value: unknown) => boolean)("message", m);
+const status = (account: string, background = true) => ({ available: true, pending: false, activeId: account, accounts: [{ id: account, connected: true, background, planEnabled: true }], error: null });
+const value = (name: string, revision: number, account = "A", authorityRevision = 1) => ({ epoch: account === "A" ? 1 : 2, authorityRevision, revision, status: status(account), accountId: account, models: [{ id: model.id, name }] });
+function reply(request: any, v: unknown) { ipc({ channel: "lh:chatgpt", id: request.id, op: request.op, ok: true, value: v }); }
+before(() => {
+ Object.defineProperty(process, "connected", { configurable: true, value: true });
+ Object.defineProperty(process, "send", { configurable: true, value: (m: unknown, cb?: () => void) => { pending.push(m); cb?.(); return true; } });
+});
+after(() => {
+ if (send) Object.defineProperty(process, "send", send); else delete process.send;
+ if (connected) Object.defineProperty(process, "connected", connected); else delete (process as { connected?: boolean }).connected;
+});
+test("overlapping catalog requests only install the newest generation", async () => {
+ const old = refreshChatGPTModels([model as never]), latest = refreshChatGPTModels([model as never]);
+ reply(pending[1], value("newest", 2)); await latest;
+ reply(pending[0], value("stale", 1)); await old;
+ assert.equal(chatGPTModels()[0]?.name, "newest");
+});
+test("consent-only status does not discard current-authority catalog or overwrite newer consent from an old status embedded in models", async () => {
+ const refresh = refreshChatGPTModels([model as never]); const request = pending.at(-1);
+ ipc({ channel: "lh:chatgpt", kind: "state", stamp: { epoch: 1, authorityRevision: 1, revision: 3 }, status: status("A", false), invalidate: "background" });
+ reply(request, value("same authority", 2)); await refresh;
+ assert.equal(chatGPTModels()[0]?.name, "same authority"); assert.equal(chatGPTStatus().accounts[0]?.background, false);
+ const scope = chatGPTScope(model.id); assert(chatGPTScopeCurrent(scope));
+ const foreground = chatGPTAbortController(false), background = chatGPTAbortController(true);
+ ipc({ channel: "lh:chatgpt", kind: "state", stamp: { epoch: 1, authorityRevision: 1, revision: 4 }, status: status("A", false), invalidate: "background" });
+ ipc({ channel: "lh:chatgpt", kind: "state", stamp: { epoch: 1, authorityRevision: 1, revision: 5 }, status: status("A", true), invalidate: "none" });
+ assert(!foreground.controller.signal.aborted); assert(background.controller.signal.aborted); assert(chatGPTScopeCurrent(scope));
+ foreground.release(); background.release();
+});
+test("account/full and catalog authority changes reject outstanding old leases without silently switching identity", async () => {
+ const scope = chatGPTScope(model.id); const lease = chatGPTLease(scope, false); const request = pending.at(-1);
+ ipc({ channel: "lh:chatgpt", kind: "state", stamp: { epoch: 2, authorityRevision: 2, revision: 6 }, status: status("B"), invalidate: "all" });
+ assert.equal(chatGPTModels().length, 0); assert(!chatGPTScopeCurrent(scope));
+ reply(request, { epoch: 1, authorityRevision: 1, revision: 5, accountId: "A", accessToken: "dummy-old" });
+ await assert.rejects(lease, /stale/);
+ const refresh = refreshChatGPTModels([model as never]); reply(pending.at(-1), value("B", 7, "B", 2)); await refresh;
+ const newScope = chatGPTScope(model.id); assert.equal(newScope.accountId, "B");
+ ipc({ channel: "lh:chatgpt", kind: "state", stamp: { epoch: 2, authorityRevision: 3, revision: 8 }, status: status("B"), invalidate: "none" });
+ assert.equal(chatGPTModels().length, 0); await assert.rejects(chatGPTLease(newScope, false));
+});

tests/chatgpt-request-cycle.test.ts

@@ -0,0 +1,154 @@
+import { test, before, after } from "node:test";
+import assert from "node:assert/strict";
+import { type Model } from "@earendil-works/pi-ai";
+import { Type } from "typebox";
+import type { AgentEvent, AgentMessage } from "@earendil-works/pi-agent-core";
+import { piModels, piEngineForSpec, APP_IDENTITY } from "../src/agent/models.ts";
+import { refreshChatGPTModels, chatGPTScopeCurrent } from "../src/agent/chatgpt-bridge.ts";
+import { ChatGPTRequestCycle } from "../src/agent/chatgpt-request-cycle.ts";
+import { createTeacherAgent } from "../src/agent/teacher-agent.ts";
+import fs from "node:fs";
+import os from "node:os";
+import path from "node:path";
+const scratch = fs.mkdtempSync(path.join(os.tmpdir(), "lh-chatgpt-native-"));
+process.env.LH_DATA_DIR = path.join(scratch, "data");
+process.env.LH_AUTH_FILE = path.join(scratch, "auth.json");
+fs.writeFileSync(process.env.LH_AUTH_FILE, "{}");
+const { runPiTurn } = await import("../src/server/pi/ui-stream.ts");
+import { newUserMessage, newNextMessage, newThinkingMessage, newAnswerResult, reconstructRequestCause, concludedByDismissalPair, type LhMessage } from "../src/core/native-messages.ts";
+
+const originalSend = Object.getOwnPropertyDescriptor(process, "send");
+const originalConnected = Object.getOwnPropertyDescriptor(process, "connected");
+const realFetch = globalThis.fetch;
+const model = piModels().getModel("openai", "gpt-5") as Model<"openai-responses">;
+let epoch = 1, revision = 1, background = true;
+let cut: "lease" | "payload" | "full" | "afterFetch" | "none" = "none";
+let leases: boolean[] = [], fetches = 0, toolFirst = false;
+const status = () => ({ available: true, pending: false, activeId: "scratch", accounts: [{ id: "scratch", registration: "oaiapp_scratch", email: "dummy@example.invalid", connected: true, planEnabled: true, background }], error: null });
+const ipc = (m: unknown) => (process.emit as (name: string, m: unknown) => boolean)("message", m);
+function revoke(full = false) {
+ background = false; if (full) epoch++;
+ ipc({ channel: "lh:chatgpt", kind: "state", stamp: { epoch, authorityRevision: epoch, revision: ++revision }, status: status(), invalidate: full ? "all" : "background" });
+}
+before(() => {
+ Object.defineProperty(process, "connected", { configurable: true, value: true });
+ Object.defineProperty(process, "send", { configurable: true, value: (m: any, callback?: () => void) => {
+  if (m.op === "lease") {
+   leases.push(m.background);
+   assert.equal(m.expectedAccountId, "scratch"); assert.equal(m.expectedAuthorityRevision, epoch); assert.equal(m.selectedModelId, model.id);
+   if (m.background && (cut === "lease" || cut === "full")) { const full = cut === "full"; cut = "none"; revoke(full); }
+  }
+  const value = { epoch, authorityRevision: epoch, revision, ...(m.op === "lease" ? { accountId: "scratch", accessToken: "dummy-oauth-native" } : { status: status(), accountId: "scratch", models: [{ id: model.id, name: "Scratch" }] }) };
+  queueMicrotask(() => ipc({ channel: "lh:chatgpt", id: m.id, op: m.op, ok: true, value })); callback?.(); return true;
+ } });
+ globalThis.fetch = async (input, init) => {
+  const request = new Request(input, init);
+  assert.equal(request.url, "https://api.openai.com/v1/responses");
+  assert.equal(request.headers.get("authorization"), "Bearer dummy-oauth-native");
+  assert.equal(request.headers.get("user-agent"), APP_IDENTITY.userAgent.custom);
+  const body = JSON.parse(await request.text()); assert.equal(body.model, model.id); assert.equal(body.store, false);
+  fetches++;
+  if (cut === "afterFetch") { cut = "none"; revoke(); throw new Error("admitted local abort"); }
+  if (toolFirst && fetches === 1) {
+   const call = { type: "function_call", id: "fc_again", call_id: "call_again", name: "again", namespace: "learnharness", arguments: "{}", status: "completed" };
+   return new Response([{ type: "response.output_item.added", output_index: 0, item: call }, { type: "response.output_item.done", output_index: 0, item: call }, { type: "response.completed", response: { status: "completed", output: [call], usage: { input_tokens: 4, output_tokens: 2, total_tokens: 6 } } }].map((e) => `data: ${JSON.stringify(e)}\n\n`).join(""), { headers: { "content-type": "text/event-stream" } });
+  }
+  const item = { id: "msg_native", type: "message", role: "assistant", content: [{ type: "output_text", text: "Accepted learner", annotations: [] }] };
+  const events = [{ type: "response.created", response: { id: "resp_native" } }, { type: "response.output_item.added", output_index: 0, item: { ...item, content: [] } }, { type: "response.content_part.added", output_index: 0, content_index: 0, part: { type: "output_text", text: "", annotations: [] } }, { type: "response.output_text.delta", output_index: 0, content_index: 0, delta: "Accepted learner" }, { type: "response.output_item.done", output_index: 0, item }, { type: "response.completed", response: { status: "completed", output: [item], usage: { input_tokens: 4, output_tokens: 2, total_tokens: 6 } } }];
+  return new Response(events.map((e) => `data: ${JSON.stringify(e)}\n\n`).join(""), { headers: { "content-type": "text/event-stream" } });
+ };
+});
+after(() => {
+ globalThis.fetch = realFetch;
+ fs.rmSync(scratch, { recursive: true, force: true });
+ if (originalSend) Object.defineProperty(process, "send", originalSend); else delete process.send;
+ if (originalConnected) Object.defineProperty(process, "connected", originalConnected); else delete (process as { connected?: boolean }).connected;
+});
+const automatic = (id = "auto") => ({ id, role: "custom" as const, customType: "subagent-steer" as const, content: "Automatic result", details: { jobId: "dummy-job" }, display: true, timestamp: 2 });
+async function nativeRun(where: typeof cut, staged: boolean, withLearner = true, commitThrows = false, lateAutomatic = false, withTool = false) {
+ epoch++; revision++; background = true; cut = where; leases = []; fetches = 0; toolFirst = withTool;
+ await refreshChatGPTModels([model]);
+ const seed = newUserMessage("Initial learner"); const marker = automatic(); const learner = { ...newUserMessage("New learner"), id: "learner-late" };
+ const messages = staged ? [seed, marker] : [seed];
+ const purpose = { background: staged };
+ let restored = 0, committed = 0, usage = 0, settled = 0, executed = 0, handedLate = false;
+ let final: AgentMessage[] = [];
+ let resolve!: () => void; const finished = new Promise<void>((done) => { resolve = done; });
+ const cycle = new ChatGPTRequestCycle({ initialMessages: messages, staged: staged ? [marker] : undefined, foregroundOwed: !staged, purpose, scopeCurrent: chatGPTScopeCurrent,
+  commitStaged: () => { if (commitThrows) throw new Error("scratch durable write failure"); committed++; },
+  restoreAutomatic: (ids, retained) => { restored++; assert(ids.has(marker.id)); assert(!retained.has(marker.id)); agent.clearSteeringQueue(); },
+ });
+ const raw = piEngineForSpec({ provider: "chatgpt", id: model.id }, () => purpose.background, cycle);
+ const engine = { ...raw, streamFn: (m: any, context: any, opts: any) => raw.streamFn(m, context, { ...opts, onPayload: async (payload: unknown) => { if (lateAutomatic && !handedLate) { handedLate = true; agent.steer(marker); } if (cut === "payload") { cut = "none"; await Promise.resolve(); revoke(); } return payload; } }) };
+ const agent = createTeacherAgent(undefined, { engine, messages, chatId: "scratch-native-session", tools: withTool ? { again: { name: "again", label: "again", description: "scratch", parameters: Type.Object({}), execute: async () => { executed++; return { content: [{ type: "text" as const, text: "Done" }], details: {} }; } } } : {}, callPurpose: purpose, requestCycle: cycle, onUsage: () => usage++ });
+ if (!staged && !lateAutomatic) agent.steer(marker);
+ if (withLearner) agent.steer(learner);
+ const handle = runPiTurn({ agent, requestCycle: cycle, onFinish: (m) => { settled++; final = m; resolve(); } });
+ const firstView = new Response(handle.subscribe()).text();
+ await finished;
+ const wire = await firstView; const replay = await new Response(handle.subscribe()).text();
+ assert.equal(settled, 1);
+ assert.equal(replay.includes('"id":"auto"'), wire.includes('"id":"auto"'));
+ assert.equal(replay.split('"type":"agent_end"').length, 2);
+ const events = wire.split("\n\n").filter((line) => line.startsWith("data: {")).map((line) => JSON.parse(line.slice(6)) as AgentEvent);
+ assert.equal(events.filter((e) => e.type === "agent_end").length, 1);
+ assert.equal(events.filter((e) => e.type === "agent_start").length, 1);
+ return { cycle, agent, final, wire, events, restored, committed, usage, executed, leases: [...leases], fetches };
+}
+for (const where of ["lease", "payload"] as const) for (const staged of [false, true]) {
+ test(`real native drain -> ${where} denial -> same Agent foreground recovery, staged=${staged}`, { timeout: 5000 }, async () => {
+  const r = await nativeRun(where, staged);
+  assert.equal(r.fetches, 1); assert.deepEqual(r.leases, [true, false]); assert.equal(r.usage, 1); assert.equal(r.restored, 1);
+  assert(!r.wire.includes('"id":"auto"')); assert(!r.wire.includes('"stopReason":"error"'));
+  assert.equal(r.final.filter((m) => (m as { id?: string }).id === "learner-late").length, 1);
+  assert.equal(r.events.filter((e) => e.type === "message_end" && (e.message as { id?: string }).id === "learner-late").length, 1);
+  assert.equal(r.events.filter((e) => e.type === "turn_start").length, 1);
+  assert.equal(await r.cycle.admission, false);
+ });
+}
+test("initial automatic denial without learner creates no usage/visible marker/durable prefix", async () => {
+ const r = await nativeRun("lease", true, false); assert.equal(r.fetches, 0); assert.equal(r.usage, 0); assert.equal(r.committed, 0); assert.equal(await r.cycle.admission, false); assert(!r.wire.includes('"id":"auto"'));
+});
+test("full invalidation never internally retries under a new authority", async () => {
+ const r = await nativeRun("full", true); assert.equal(r.fetches, 0); assert.equal(r.usage, 0); assert.deepEqual(r.leases, [true]);
+});
+test("successful staged prefix commits once before native dispatch and is visible exactly once", async () => {
+ const r = await nativeRun("none", true, false); assert.equal(r.fetches, 1); assert.equal(r.committed, 1); assert.equal(await r.cycle.admission, true);
+ assert.equal(r.events.filter((e) => e.type === "message_end" && (e.message as { id?: string }).id === "auto").length, 1);
+});
+test("durable staged input failure prevents dispatch and does not infer foreground replay", async () => {
+ const r = await nativeRun("none", true, false, true); assert.equal(r.fetches, 0); assert.equal(r.usage, 0); assert.equal(await r.cycle.admission, false);
+});
+test("invalidation after fetch commitment cannot replay an admitted request", async () => {
+ const r = await nativeRun("afterFetch", false); assert.equal(r.fetches, 1); assert.equal(r.restored, 0); assert.deepEqual(r.leases, [true]); assert.equal(await r.cycle.admission, true);
+});
+
+test("automatic arrival after completed foreground prose cannot manufacture another foreground request", async () => {
+ const r = await nativeRun("lease", false, false, false, true);
+ assert.equal(r.fetches, 1); assert.equal(r.usage, 1); assert.equal(r.restored, 1); assert.deepEqual(r.leases, [false, true]);
+ assert(!r.wire.includes('"id":"auto"'));
+});
+test("unfinished foreground ordinary-tool chain recovers without rerunning the completed tool", async () => {
+ const r = await nativeRun("lease", false, false, false, true, true);
+ assert.equal(r.fetches, 2); assert.equal(r.executed, 1); assert.equal(r.usage, 2); assert.deepEqual(r.leases, [false, true, false]);
+ assert.equal(r.final.filter((m) => m.role === "toolResult").length, 1);
+ assert(!r.wire.includes('"id":"auto"'));
+});
+
+test("full native record reconstruction preserves original causes and reports genuine ambiguity", () => {
+ const user = newUserMessage("Learner"); const auto = automatic();
+ const assistant = { role: "assistant", id: "assistant", api: "openai-responses", provider: "chatgpt", model: model.id, content: [{ type: "toolCall", id: "call", name: "read", arguments: {} }], stopReason: "toolUse", timestamp: 3, usage: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, totalTokens: 0, cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 } } } as LhMessage;
+ const result = { role: "toolResult", id: "result", toolName: "read", toolCallId: "call", content: [], details: {}, isError: false, timestamp: 4 } as LhMessage;
+ for (const [origin, expected] of [[user, "foreground"], [auto, "background"]] as const) assert.equal(reconstructRequestCause(JSON.parse(JSON.stringify([origin, assistant, result]))), expected);
+ assert.equal(reconstructRequestCause([auto, user, assistant, result]), "ambiguous");
+ assert.equal(reconstructRequestCause([user, auto, assistant, result]), "background");
+ assert.equal(reconstructRequestCause([assistant, result]), "ambiguous");
+ assert.equal(reconstructRequestCause([auto, newThinkingMessage("problem", "working")]), "ambiguous");
+ assert.equal(reconstructRequestCause([auto, assistant, newNextMessage()]), "foreground");
+ assert.equal(reconstructRequestCause([user, ...concludedByDismissalPair("problem")]), "foreground");
+ const question = { ...assistant, content: [{ type: "toolCall", id: "question", name: "quiz", arguments: {} }] } as LhMessage;
+ assert.equal(reconstructRequestCause([auto, question, newAnswerResult("question", "quiz", { kind: "quiz", questionText: "Scratch", dontKnow: false, answers: [], answeredAt: "2026-10-24T00:00:00Z" })]), "foreground");
+ assert.equal(reconstructRequestCause([auto, question, { ...newAnswerResult("question", "quiz", {}), details: null }]), "background");
+ assert.equal(reconstructRequestCause([auto, question, newAnswerResult("orphan", "quiz", {})]), "ambiguous");
+ assert.equal(reconstructRequestCause([user, { ...user, content: "conflict" }]), "ambiguous");
+});

tests/chatgpt-stream.test.ts

@@ -0,0 +1,174 @@
+import { test, before, after } from "node:test";
+import assert from "node:assert/strict";
+import { normalizeContext, type AssistantMessage, type Model } from "@earendil-works/pi-ai";
+import { Type } from "typebox";
+import { piModels, piEngineForSpec } from "../src/agent/models.ts";
+import { chatGPTAbortController, chatGPTModels, refreshChatGPTModels } from "../src/agent/chatgpt-bridge.ts";
+import { chatGPTHistory, chatGPTPayload } from "../src/agent/chatgpt-stream.ts";
+
+const originalSend = Object.getOwnPropertyDescriptor(process, "send");
+const originalConnected = Object.getOwnPropertyDescriptor(process, "connected");
+const realFetch = globalThis.fetch;
+const originalKey = process.env.OPENAI_API_KEY;
+let leaseError: string | null = null;
+let requests: { authorization: string | null; body: Record<string, any>; url: string }[] = [];
+let responseEvents: unknown[] = [];
+const ipc = (value: unknown) => (process.emit as (name: string, value: unknown) => boolean)("message", value);
+const model = piModels().getModel("openai", "gpt-5") as Model<"openai-responses">;
+const status = { available: true, pending: false, activeId: "scratch-account", accounts: [{ id: "scratch-account", email: "dummy@example.invalid", registration: "oaiapp_scratch", connected: true, planEnabled: true, background: false }], error: null };
+before(async () => {
+	assert(model);
+	process.env.OPENAI_API_KEY = "sk-scratch-must-never-be-used";
+	Object.defineProperty(process, "connected", { configurable: true, value: true });
+	Object.defineProperty(process, "send", { configurable: true, value: (message: any, callback?: () => void) => {
+		queueMicrotask(() => {
+			ipc({ channel: "lh:chatgpt", id: message.id, op: message.op,
+				...(message.op === "lease" && leaseError ? { ok: false, error: leaseError, stamp: { epoch: 1, authorityRevision: 1, revision: 1 } } : { ok: true, value: message.op === "lease" ? { accessToken: "scratch-oauth-only", accountId: "scratch-account", epoch: 1, authorityRevision: 1, revision: 1 } : { epoch: 1, authorityRevision: 1, revision: 1, status, accountId: "scratch-account", models: [{ id: model.id, name: "Scratch account model" }, { id: "unknown-capabilities", name: "Do not invent capabilities" }] } }),
+			});
+		});
+		callback?.(); return true;
+	} });
+	await refreshChatGPTModels([model]);
+	globalThis.fetch = async (input, init) => {
+		const req = new Request(input, init);
+		requests.push({ authorization: req.headers.get("authorization"), body: JSON.parse(await req.text()), url: req.url });
+		return new Response(responseEvents.map((event) => `data: ${JSON.stringify(event)}\n\n`).join(""), { headers: { "content-type": "text/event-stream" } });
+	};
+});
+after(() => {
+	globalThis.fetch = realFetch;
+	if (originalSend) Object.defineProperty(process, "send", originalSend); else delete process.send;
+	if (originalConnected) Object.defineProperty(process, "connected", originalConnected); else delete (process as { connected?: boolean }).connected;
+	if (originalKey === undefined) delete process.env.OPENAI_API_KEY; else process.env.OPENAI_API_KEY = originalKey;
+});
+const usage = { input_tokens: 10, output_tokens: 4, total_tokens: 14, input_tokens_details: { cached_tokens: 0 }, output_tokens_details: { reasoning_tokens: 0 } };
+const complete = (output: unknown[] = []) => ({ type: "response.completed", response: { id: "resp_scratch", status: "completed", output, usage } });
+const context = () => normalizeContext({ systemPrompt: "Scratch teacher instructions", messages: [{ role: "user", content: [{ type: "text", text: "Scratch learning material" }, { type: "image", data: "c2NyYXRjaA==", mimeType: "image/png" }], timestamp: 1 }], tools: [{ name: "scratch_tool", description: "Scratch local tool", parameters: Type.Object({ value: Type.String() }) }] });
+
+test("catalog intersects supported capabilities, without inventing model limits", () => {
+	assert.equal(chatGPTModels().length, 1);
+	assert.equal(chatGPTModels()[0]!.provider, "chatgpt");
+	assert.equal(chatGPTModels()[0]!.contextWindow, model.contextWindow);
+	assert.deepEqual(chatGPTModels()[0]!.input, model.input);
+});
+
+test("alias resolves only OAuth; native Responses body is stateless/namespaced, image/history preserved and logical output retained", async () => {
+	requests = [];
+	const tool = { type: "function_call", id: "fc_scratch", call_id: "call_scratch", name: "scratch_tool", namespace: "learnharness", arguments: '{"value":"yes"}', status: "completed" };
+	const reasoning = { type: "reasoning", id: "rs_scratch", summary: [], encrypted_content: "scratch-encrypted-reasoning" };
+	responseEvents = [{ type: "response.created", response: { id: "resp_scratch" } }, { type: "response.output_item.added", output_index: 0, item: reasoning }, { type: "response.output_item.done", output_index: 0, item: reasoning }, { type: "response.output_item.added", output_index: 1, item: tool }, { type: "response.output_item.done", output_index: 1, item: tool }, complete([reasoning, tool])];
+	const engine = piEngineForSpec({ provider: "chatgpt", id: model.id }, false);
+	const stream = engine.streamFn(engine.model, context(), { temperature: 0.7, maxTokens: 1000, samplingParams: { top_p: 0.5 }, onPayload: (payload) => ({ ...payload as object, model: "caller-cannot-change", dangerous: true }) });
+	const events = [];
+	for await (const event of stream) { events.push(event); if ("partial" in event) assert.equal(event.partial.provider, "chatgpt"); }
+	const result = await stream.result();
+	assert.equal(result.provider, "chatgpt"); assert.equal(result.stopReason, "toolUse");
+	assert.equal(requests.length, 1); assert.equal(requests[0]!.url, "https://api.openai.com/v1/responses");
+	assert.equal(requests[0]!.authorization, "Bearer scratch-oauth-only");
+	const body = requests[0]!.body;
+	assert.equal(body.store, false); assert.equal(body.stream, true);
+	for (const key of ["temperature", "top_p", "max_output_tokens", "previous_response_id", "background", "metadata", "prompt_cache_retention"]) assert.equal(body[key], undefined, key);
+	assert.equal(body.input[0].role, "developer");
+	assert.equal(body.input[1].content[1].type, "input_image");
+	assert.equal(body.tools[0].type, "namespace"); assert.equal(body.tools[0].tools[0].name, "scratch_tool");
+	const call = result.content.find((block) => block.type === "toolCall")!;
+	assert.equal(call.id, "call_scratch|fc_scratch"); assert.equal(call.namespace, "learnharness");
+	const thinking = result.content.find((block) => block.type === "thinking")!;
+	assert.equal(thinking.thinkingSignature, JSON.stringify(reasoning));
+	const history = normalizeContext({ messages: [...context().messages, result, { role: "toolResult", toolCallId: call.id, toolName: call.name, isError: false, content: [{ type: "text", text: "Scratch tool result" }], timestamp: 2 }] });
+	responseEvents = [complete()];
+	await engine.streamFn(engine.model, history).result();
+	const replay = requests[1]!.body.input;
+	assert(replay.some((item: any) => item.encrypted_content === "scratch-encrypted-reasoning"));
+	assert(replay.some((item: any) => item.type === "function_call" && item.id === "fc_scratch" && item.namespace === "learnharness"));
+	assert(replay.some((item: any) => item.type === "function_call_output" && item.call_id === "call_scratch"));
+});
+
+test("mixed history never blanket-relabels API providers or reuses foreign reasoning signatures", () => {
+	const native: AssistantMessage = { role: "assistant", provider: "openai", api: "openai-responses", model: model.id, content: [{ type: "thinking", thinking: "", thinkingSignature: '{"encrypted_content":"API-only"}', redacted: true }], stopReason: "stop", usage: { input: 1, output: 0, cacheRead: 0, cacheWrite: 0, totalTokens: 1, cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 } }, timestamp: 1 };
+	const own = { ...native, provider: "chatgpt", content: [{ type: "thinking" as const, thinking: "", thinkingSignature: '{"encrypted_content":"ChatGPT-only"}', redacted: true }] };
+	const transformed = chatGPTHistory(normalizeContext({ messages: [native, own] }), { ...model, provider: "chatgpt" });
+	assert.equal((transformed.messages[0] as AssistantMessage).provider, "openai");
+	assert.deepEqual((transformed.messages[0] as AssistantMessage).content, []);
+	assert.equal((transformed.messages[1] as AssistantMessage).content[0]!.type, "thinking");
+	assert.equal(native.content.length, 1, "source history must not be mutated");
+});
+
+test("no terminal completed is never success; late quota/incomplete streams are actionable/redacted", async () => {
+	const engine = piEngineForSpec({ provider: "chatgpt", id: model.id }, false);
+	for (const events of [[{ type: "response.created", response: { id: "resp_partial" } }], [{ type: "response.incomplete", response: { status: "incomplete", output: [], incomplete_details: { reason: "max_output_tokens" }, usage } }], [{ type: "error", code: "subscription_sharing_usage_limit_exceeded", message: "secret scratch provider diagnostic" }]]) {
+		responseEvents = events;
+		const result = await engine.streamFn(engine.model, context()).result();
+		assert.equal(result.stopReason, "error"); assert.equal(result.provider, "chatgpt");
+		assert(!result.errorMessage?.includes("secret scratch provider diagnostic"));
+		if (JSON.stringify(events).includes("usage_limit")) assert(result.errorMessage?.includes("https://chatgpt.com/settings/usage"));
+	}
+});
+
+test("OAuth failure and background denial NEVER use ambient OPENAI_API_KEY", async () => {
+	requests = [];
+	for (const error of ["expired ChatGPT session", "Background use is off", "catalog unavailable", "quota exceeded"]) {
+		leaseError = error;
+		const engine = piEngineForSpec({ provider: "chatgpt", id: model.id }, true);
+		assert.equal((await engine.streamFn(engine.model, context()).result()).stopReason, "error");
+	}
+	leaseError = null;
+	assert.equal(requests.length, 0);
+});
+
+test("consent revocation cancels only applicable background streams and preserves foreground", () => {
+	const foreground = chatGPTAbortController(false);
+	const background = chatGPTAbortController(true);
+	ipc({ channel: "lh:chatgpt", kind: "state", stamp: { epoch: 1, authorityRevision: 1, revision: 2 }, status, invalidate: "background" });
+	assert.equal(foreground.controller.signal.aborted, false);
+	assert.equal(background.controller.signal.aborted, true);
+	foreground.release(); background.release();
+});
+
+test("payload sanitizer omits unsupported fields and rejects unsupported hosted tools", () => {
+	assert.deepEqual(chatGPTPayload({ model: "scratch", input: [], stream: false, store: true, metadata: { test: true }, previous_response_id: "old", background: true }, "scratch"), { model: "scratch", input: [], stream: true, store: false });
+	assert.throws(() => chatGPTPayload({ input: [], tools: [{ type: "tool_search" }] }, "scratch"), /Unsupported/);
+});
+
+test("native identity/session/cache/fetch and awaited caller observers survive enforcement; terminal outcome exactly once", async () => {
+	const { APP_IDENTITY } = await import("../src/agent/models.ts");
+	const engine = piEngineForSpec({ provider: "chatgpt", id: model.id }, false);
+	let headers = new Headers(), payload: any, observed = 0, responses = 0, delegated = 0;
+	responseEvents = [complete()];
+	const stream = engine.streamFn(engine.model, context(), {
+		identity: APP_IDENTITY, sessionId: "exact-native-chat-session", timeoutMs: 12345,
+		headers: { "x-native-observer": "retained", "x-suppressed": null },
+		fetch: async (input, init) => { delegated++; const req = new Request(input, init); headers = req.headers; payload = JSON.parse(await req.text()); return new Response(`data: ${JSON.stringify(complete())}\n\n`, { headers: { "content-type": "text/event-stream" } }); },
+		onPayload: async (raw, native) => { await Promise.resolve(); assert.equal(native.provider, "openai"); return { ...raw as object, model: "not-the-selected-model", store: true, metadata: { forbidden: true } }; },
+		onResponse: async (_response, native) => { await Promise.resolve(); assert.equal(native.provider, "openai"); responses++; },
+		onProviderStreamEvent: async (_raw, native) => { await Promise.resolve(); assert.equal(native.provider, "openai"); observed++; },
+	});
+	const events = []; for await (const event of stream) events.push(event);
+	assert.equal(delegated, 1); assert.equal(observed, 1); assert.equal(responses, 1);
+	assert.equal(headers.get("user-agent"), APP_IDENTITY.userAgent.custom);
+	assert.equal(headers.get("session_id"), "exact-native-chat-session");
+	assert.equal(headers.get("x-client-request-id"), "exact-native-chat-session");
+	assert.equal(headers.get("x-native-observer"), "retained"); assert.equal(headers.get("x-suppressed"), null);
+	assert.equal(payload.prompt_cache_key, "exact-native-chat-session"); assert.equal(payload.model, model.id); assert.equal(payload.metadata, undefined);
+	assert.equal(events.filter((e) => e.type === "done" || e.type === "error").length, 1);
+	const failed = engine.streamFn(engine.model, context(), { onProviderStreamEvent: async () => { throw new Error("private observer detail"); } });
+	const failedEvents = []; for await (const event of failed) failedEvents.push(event);
+	assert.equal(failedEvents.filter((e) => e.type === "error").length, 1); assert.equal(failedEvents.filter((e) => e.type === "done").length, 0);
+	assert(!(await failed.result()).errorMessage?.includes("private observer detail"));
+});
+
+test("case-insensitive Authorization overrides including null are refused before native fetch", async () => {
+	const engine = piEngineForSpec({ provider: "chatgpt", id: model.id }, false);
+	const before = requests.length;
+	for (const headers of [{ Authorization: "Bearer wrong" }, { authorization: null }, { AUTHORIZATION: "" }] as Record<string, string | null>[]) {
+		assert.equal((await engine.streamFn(engine.model, context(), { headers }).result()).stopReason, "error");
+	}
+	assert.equal(requests.length, before);
+});
+
+test("full secure-storage invalidation aborts foreground and background controllers and clears capabilities", () => {
+	const foreground = chatGPTAbortController(false), background = chatGPTAbortController(true);
+	ipc({ channel: "lh:chatgpt", kind: "state", stamp: { epoch: 2, authorityRevision: 2, revision: 3 }, status: { ...status, available: false }, invalidate: "all" });
+	assert(foreground.controller.signal.aborted); assert(background.controller.signal.aborted); assert.equal(chatGPTModels().length, 0);
+	foreground.release(); background.release();
+});

tests/chatgpt-ui.test.tsx

@@ -0,0 +1,83 @@
+import React, { act } from "react";
+(globalThis as Record<string, unknown>).React = React;
+import assert from "node:assert/strict";
+import { test, afterEach } from "node:test";
+import { createRoot, type Root } from "react-dom/client";
+import { parseHTML } from "linkedom";
+import { useModelCatalog, catalogLimits } from "../components/model-picker.tsx";
+import { useZeroKeys, useStrongKeyless } from "../components/setup-cta.tsx";
+import { notifySettingsSaved } from "../components/settings-signal.ts";
+import { SidebarProvider } from "../components/sidebar.tsx";
+import { ChatGPTConnection } from "../components/settings/chatgpt-connection.tsx";
+import { GraphCostControl } from "../components/graph-view.tsx";
+import { graphCostSummary } from "../src/core/cost-summary.ts";
+import type { ModelCatalogResponse, ChatGPTStatus } from "../src/core/index.ts";
+const dom = parseHTML("<html><body></body></html>");
+Object.assign(globalThis, { window: dom.window, document: dom.document, Element: dom.window.Element, HTMLElement: dom.window.HTMLElement, CustomEvent: dom.window.CustomEvent, IS_REACT_ACT_ENVIRONMENT: true, localStorage: { getItem: () => null, setItem() {} } });
+const originalFetch = globalThis.fetch;
+let root: Root | undefined, host: HTMLDivElement | undefined;
+function deferred<T>() { let resolve!: (v: T) => void; const promise = new Promise<T>((done) => { resolve = done; }); return { promise, resolve }; }
+function body(id: string, ready = true): ModelCatalogResponse {
+ return { models: ready ? [{ id, name: id, provider: "chatgpt", input: ["text"], contextWindow: 8192, maxTokens: 1024, reasoning: false, supportedLevels: ["off"], cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 } }] : [], providers: [{ id: "chatgpt", envVar: "", keyPresent: ready }], slots: { strong: { model: { provider: "chatgpt", id: "explicit-pin" } }, cheap: null, vision: null }, defaults: {} } as ModelCatalogResponse;
+}
+let snapshots: Record<string, { catalog?: ModelCatalogResponse; zero: boolean; strong: boolean }> = {};
+function Reader({ id }: { id: string }) {
+ const { catalog } = useModelCatalog();
+ snapshots[id] = { catalog, zero: useZeroKeys(), strong: useStrongKeyless() };
+ return <span data-testid={id}>{catalog ? [...catalogLimits(catalog).keys()].join(",") : "unavailable"}</span>;
+}
+async function mount(children: React.ReactNode) { host = document.createElement("div"); document.body.append(host); root = createRoot(host); await act(async () => root!.render(children)); }
+afterEach(async () => { await act(async () => root?.unmount()); host?.remove(); root = undefined; host = undefined; globalThis.fetch = originalFetch; delete window.lh; snapshots = {}; });
+
+test("shared ChatGPT-only readiness admits without API keys; connection invalidation clears every reader and fences old success/error", async () => {
+ const requests: ReturnType<typeof deferred<Response>>[] = [];
+ globalThis.fetch = async (input) => { assert.equal(input, "/api/models"); const request = deferred<Response>(); requests.push(request); return request.promise; };
+ await mount(<><Reader id="first" /><Reader id="second" /></>);
+ assert.equal(requests.length, 1);
+ await act(async () => requests[0]!.resolve(Response.json(body("A"))));
+ for (const value of Object.values(snapshots)) { assert.equal(value.zero, false); assert.equal(value.strong, false); assert.equal(value.catalog?.models[0]?.id, "A"); }
+ await act(async () => notifySettingsSaved("models"));
+ assert.equal(requests.length, 2);
+ await act(async () => notifySettingsSaved("keys"));
+ assert.equal(requests.length, 3);
+ for (const value of Object.values(snapshots)) assert.equal(value.catalog, undefined);
+ await act(async () => requests[2]!.resolve(Response.json(body("B", false))));
+ await act(async () => requests[1]!.resolve(Response.json(body("stale-A"))));
+ for (const value of Object.values(snapshots)) { assert.equal(value.catalog?.models.length, 0); assert.equal(value.zero, true); assert.equal(value.strong, true); }
+ assert(!document.body.textContent?.includes("chatgpt:A"));
+});
+
+test("plan plus unknown graph rows keep the existing unknown suffix without fabricating API dollars", async () => {
+ const step = { inputTokens: 100, outputTokens: 20, cacheReadTokens: 0, cacheWriteTokens: 0, reasoningTokens: 0 };
+ const plan = { model: "chatgpt:gpt-5", steps: [step] };
+ const unknown = { model: "unknown:missing", steps: [step] };
+ const costs = graphCostSummary({ chats: [{ title: "Scratch", extractions: [{ startedAt: 1, at: "2026-10-24T00:00:00Z", durationMs: 1, stages: { extractor: plan, auditor: unknown } }, { startedAt: 2, at: "2026-10-24T00:00:00Z", durationMs: 1, stages: { extractor: unknown, auditor: plan } }] }], sources: [] });
+ await mount(<GraphCostControl costs={costs} />);
+ await act(async () => { document.querySelector("button")!.click(); });
+ const recent = [...document.querySelectorAll(".border-t .shrink-0.tabular-nums")];
+ assert.equal(recent.length, 2);
+ for (const row of recent) assert.equal(row.textContent, "ChatGPT plan +1 unpriced", "both recent two-stage runs disclose their unknown stage");
+ assert.equal((document.body.textContent?.match(/\+1 unpriced/g) ?? []).length, 4, "both mixed stage rows also retain their unknown suffix");
+ assert(!document.body.textContent?.includes("$0.00"));
+});
+
+test("root-lifetime safe bridge publishes with Connections unmounted; commands/panel pushes do not duplicate global signals", async () => {
+ let pushes: (() => void)[] = [], subscriptions = 0, unsubscriptions = 0, fetches = 0;
+ const safe: ChatGPTStatus = { available: true, pending: false, activeId: "a", accounts: [{ id: "a", email: "dummy@example.invalid", registration: "oaiapp_dummy", connected: true, background: false, planEnabled: true }], error: null };
+ window.lh = { expandWindow() {}, chatgpt: { status: async () => safe, login: async () => safe, cancel: async () => safe, select: async () => safe, logout: async () => safe, consent: async () => safe, onChange: (cb) => { subscriptions++; pushes.push(cb); return () => { unsubscriptions++; pushes = pushes.filter((c) => c !== cb); }; } } };
+ globalThis.fetch = async (input) => { assert.equal(input, "/api/models"); fetches++; return Response.json(body(`model-${fetches}`)); };
+ await mount(<SidebarProvider><Reader id="root" /></SidebarProvider>);
+ assert.equal(subscriptions, 1);
+ const before = fetches;
+ await act(async () => { for (const callback of pushes) callback(); });
+ assert.equal(fetches, before + 1);
+ await act(async () => root!.render(<SidebarProvider><Reader id="root" /><ChatGPTConnection /></SidebarProvider>));
+ assert.equal(subscriptions, 2);
+ const mounted = fetches;
+ await act(async () => { for (const callback of pushes) callback(); });
+ assert.equal(fetches, mounted + 1, "one global publisher even when panel also reads pushed status");
+ await act(async () => root!.render(<SidebarProvider><Reader id="root" /></SidebarProvider>));
+ assert.equal(unsubscriptions, 1);
+ await act(async () => { for (const callback of pushes) callback(); });
+ assert.equal(fetches, mounted + 2);
+});

tests/chatgpt-usage.test.ts

@@ -0,0 +1,45 @@
+import { test } from "node:test";
+import assert from "node:assert/strict";
+import { callCost, sessionCost, turnCost, jobCost } from "../src/core/prices.ts";
+import { graphCostSummary } from "../src/core/cost-summary.ts";
+import type { TurnUsage } from "../src/core/contracts.ts";
+import { planCostLabel } from "../components/usage-format.ts";
+const step = { inputTokens: 1000, outputTokens: 100, cacheReadTokens: undefined, cacheWriteTokens: undefined, reasoningTokens: undefined };
+const turn = (model: string, steps = [step]): TurnUsage => ({ id: "scratch-turn", messageId: "scratch-message", at: "2026-10-19T00:00:00Z", model, steps });
+
+test("subscription usage is neither API dollars nor an unpriced-model mistake, including empty/partial usage", () => {
+	const plan = turn("chatgpt:gpt-5");
+	assert.equal(callCost(step, plan.model), null);
+	assert.equal(turnCost(turn(plan.model, [])), null);
+	assert.equal(jobCost({ usage: { model: plan.model, ...step, steps: 1 } }), null);
+	assert.deepEqual(sessionCost({ turns: [plan, turn(plan.model, [])] }), { cost: 0, unpriced: 0, subscription: 2 });
+	const api = turn("openai:gpt-5");
+	const mixed = sessionCost({ turns: [plan, api, turn("unknown:missing")] });
+	assert.equal(mixed.cost, turnCost(api)); assert.equal(mixed.unpriced, 1); assert.equal(mixed.subscription, 1);
+	assert.equal(planCostLabel(null, true), "ChatGPT plan");
+	assert.match(planCostLabel(mixed.cost, true), /API \+ ChatGPT plan$/);
+});
+
+test("graph extraction and audit preserve subscription identity separately from priced and unknown stages", () => {
+	const base = { startedAt: 1, at: "2026-10-19T00:00:00Z", durationMs: 1, stages: { extractor: { model: "chatgpt:gpt-5", steps: [step] }, auditor: { model: "chatgpt:gpt-5", steps: [] } } };
+	const plan = graphCostSummary({ chats: [{ title: "Scratch", extractions: [base] }], sources: [] });
+	assert.equal(plan.unpricedRuns, 0); assert.equal(plan.subscriptionRuns, 2);
+	assert.equal(plan.stages.extraction.unpricedRuns, 0); assert.equal(plan.stages.audit.unpricedRuns, 0);
+	assert.equal(plan.runs[0]!.dollars, null);
+	const mixed = graphCostSummary({ chats: [{ title: "Scratch", extractions: [{ ...base, stages: { ...base.stages, auditor: { model: "openai:gpt-5", steps: [step] } } }] }], sources: [] });
+	assert.equal(mixed.subscriptionRuns, 1);
+	assert.equal(mixed.totalDollars, callCost(step, "openai:gpt-5"));
+	assert.equal(mixed.runs[0]!.subscriptionRuns, 1);
+	assert.equal(mixed.unpricedRuns, 0);
+	for (const reverse of [false, true]) {
+		const stages = [{ model: "chatgpt:gpt-5", steps: [step] }, { model: "unknown:not-priced", steps: [step] }];
+		if (reverse) stages.reverse();
+		const unknown = graphCostSummary({ chats: [{ title: "Scratch", extractions: [{ ...base, stages: { extractor: stages[0]!, auditor: stages[1]! } }] }], sources: [] });
+		assert.equal(unknown.unpricedRuns, 1); assert.equal(unknown.runs[0]!.dollars, null); assert.equal(unknown.subscriptionRuns, 1);
+		assert.equal(unknown.stages.extraction.unpricedRuns + unknown.stages.audit.unpricedRuns, 1);
+	}
+	const zero = graphCostSummary({ chats: [{ title: "Scratch", extractions: [{ ...base, stages: { ...base.stages, auditor: { model: "openai:gpt-5", steps: [{ ...step, inputTokens: 0, outputTokens: 0 }] } } }] }], sources: [] });
+	assert.equal(zero.runs[0]!.dollars, 0); assert.equal(zero.unpricedRuns, 0); assert.equal(zero.subscriptionRuns, 1);
+	const capped = graphCostSummary({ recentCap: 0, chats: [{ title: "Scratch", extractions: [{ ...base, stages: { ...base.stages, auditor: { model: "unknown:model", steps: [step] } } }] }], sources: [] });
+	assert.equal(capped.runs.length, 0); assert.equal(capped.unpricedRuns, 1);
+});

tests/compaction-service.test.ts

@@ -74,6 +74,23 @@ const depsOf = (chatId: string, thresholdTokens: number): CompactionDeps => ({
 // ── the laws ────────────────────────────────────────────────────────────────
 
 describe("the compaction service", () => {
+	it("background eligibility skips both entrypoints without writes, but still awaits an existing structural run", async () => {
+		const chatId = overThresholdChat("compact: consent off");
+		const before = summarizeCalls;
+		const denied = { ...depsOf(chatId, 40_000), canStart: () => false };
+		await ensureCompacted(denied, chatId); kickCompaction(() => denied, chatId);
+		assert.equal(summarizeCalls, before);
+		assert.equal(activeCompactionOf(chatStore.getChat(chatId)!.messages), null);
+		let release!: () => void, entered!: () => void;
+		const gate = new Promise<void>((done) => { release = done; });
+		const started = new Promise<void>((done) => { entered = done; });
+		const allowed = { ...depsOf(chatId, 40_000), summarize: async (prompt: string) => { entered(); await gate; return depsOf(chatId, 40_000).summarize!(prompt, denied.engine); } };
+		kickCompaction(() => allowed, chatId); await started;
+		let finished = false;
+		const waiter = ensureCompacted(denied, chatId).then(() => { finished = true; });
+		await Promise.resolve(); assert.equal(finished, false);
+		release(); await waiter; assert.equal(finished, true);
+	});
 	it("over-threshold record → marker appended, usage logged, both events emitted", async () => {
 		const chatId = overThresholdChat("compact: basic");
 		const events: string[] = [];

tests/desktop-api.test.ts

@@ -0,0 +1,84 @@
+import { test } from "node:test";
+import assert from "node:assert/strict";
+import { createRequire } from "node:module";
+import { NextRequest } from "next/server";
+// Installed Next 16.3.2 retains the old test-helper export despite proxy.md's new name.
+import { unstable_doesMiddlewareMatch as unstable_doesProxyMatch } from "next/experimental/testing/server";
+import { desktopAPIAuthorized } from "../src/server/desktop-access.ts";
+import { proxy, config } from "../proxy.ts";
+import { browserServiceSpawnEnv } from "../src/agent/subagents/headless.ts";
+const require = createRequire(import.meta.url);
+const { desktopAPIHeaders, HEADER } = require("../electron/desktop-api.cjs");
+const { trustedSender, chatGPTControl } = require("../electron/chatgpt-ipc.cjs");
+const secret = "a".repeat(64);
+const origin = "http://127.0.0.1:12345";
+
+function fixture() {
+	const mainFrame = { url: `${origin}/chat/new` };
+	const contents = { id: 5, mainFrame, isDestroyed: () => false };
+	const win = { webContents: contents, isDestroyed: () => false };
+	const details = { url: `${origin}/api/settings`, webContentsId: contents.id, webContents: contents, frame: mainFrame, initiatorOrigin: origin, requestHeaders: { Accept: "application/json" } };
+	return { mainFrame, contents, win, details };
+}
+
+test("desktop capability injects only trusted app mainframe exact-origin API requests and strips every other case", () => {
+	const { details, win, mainFrame, contents } = fixture();
+	assert.equal(desktopAPIHeaders(details, origin, secret, [win])[HEADER], secret);
+	for (const mutation of [
+		{ url: "https://thirdparty.invalid/api/settings" },
+		{ url: "http://127.0.0.1:12346/api/settings" },
+		{ url: `${origin}/chat/new` },
+		{ frame: { url: `${origin}/chat/new` } }, // forged frame identity
+		{ frame: { url: "about:srcdoc", parent: mainFrame } },
+		{ frame: null },
+		{ webContentsId: 6 },
+		{ webContents: { ...contents } },
+		{ initiatorOrigin: "https://thirdparty.invalid" },
+		{ initiatorOrigin: "null" },
+	]) {
+		const result = desktopAPIHeaders({ ...details, ...mutation, requestHeaders: { "X-LearnHarness-Desktop": secret } }, origin, secret, [win]);
+		assert(!Object.keys(result).some((key) => key.toLowerCase() === HEADER), JSON.stringify(mutation));
+	}
+	assert.equal(desktopAPIHeaders(details, origin, secret, [])[HEADER], undefined, "other WebContents are not app windows");
+	mainFrame.url = "https://thirdparty.invalid/after-navigation";
+	assert.equal(desktopAPIHeaders(details, origin, secret, [win])[HEADER], undefined, "navigation revokes trust even for the same native frame");
+	assert.equal(desktopAPIHeaders(details, origin, null, [win])[HEADER], undefined, "externally launched desktop dev receives no packaged capability");
+});
+
+test("all methods fail closed for desktop missing/wrong secret; API-only matcher covers SSE/settings/spending", () => {
+	assert(desktopAPIAuthorized(new Headers(), {}), "browser-only API-key mode unchanged");
+	assert(!desktopAPIAuthorized(new Headers({ [HEADER]: secret }), { LH_DESKTOP_SERVER: "1" }), "missing server secret must not disable authorization");
+	for (const value of [undefined, "", "wrong", secret + ", " + secret]) {
+		assert(!desktopAPIAuthorized(new Headers(value === undefined ? {} : { [HEADER]: value }), { LH_DESKTOP_SERVER: "1", LH_DESKTOP_API_SECRET: secret }));
+	}
+	assert(desktopAPIAuthorized(new Headers({ [HEADER]: secret }), { LH_DESKTOP_SERVER: "1", LH_DESKTOP_API_SECRET: secret }));
+	for (const pathname of ["/api/events", "/api/chat", "/api/settings/models", "/api/chats/scratch/markdown"]) assert(unstable_doesProxyMatch({ config, nextConfig: {}, url: origin + pathname }));
+	assert(!unstable_doesProxyMatch({ config, nextConfig: {}, url: origin + "/chat/new" }));
+	const originalMarker = process.env.LH_DESKTOP_SERVER;
+	const originalSecret = process.env.LH_DESKTOP_API_SECRET;
+	try {
+		process.env.LH_DESKTOP_SERVER = "1";
+		process.env.LH_DESKTOP_API_SECRET = secret;
+		for (const method of ["GET", "POST", "PUT", "PATCH", "DELETE", "HEAD", "OPTIONS"]) {
+			assert.equal(proxy(new NextRequest(origin + "/api/settings", { method })).status, 403);
+			assert.equal(proxy(new NextRequest(origin + "/api/settings", { method, headers: { [HEADER]: secret } })).headers.get("x-middleware-next"), "1");
+		}
+	} finally {
+		if (originalMarker === undefined) delete process.env.LH_DESKTOP_SERVER; else process.env.LH_DESKTOP_SERVER = originalMarker;
+		if (originalSecret === undefined) delete process.env.LH_DESKTOP_API_SECRET; else process.env.LH_DESKTOP_API_SECRET = originalSecret;
+	}
+});
+
+test("renderer controls have no lease or general token bridge, and require the trusted main frame", async () => {
+	let leased = false;
+	const connection = { lease: async () => { leased = true; return { accessToken: "scratch-private-token" }; } };
+	assert.deepEqual(await chatGPTControl({ op: "lease", background: false }, connection), { error: "Unsupported ChatGPT control." });
+	assert.equal(leased, false);
+	const { contents, mainFrame, win } = fixture();
+	const BrowserWindow = { fromWebContents: (value: unknown) => value === contents ? win : null };
+	assert(trustedSender({ sender: contents, senderFrame: mainFrame }, origin, BrowserWindow));
+	assert(!trustedSender({ sender: contents, senderFrame: { url: mainFrame.url } }, origin, BrowserWindow));
+	mainFrame.url = "https://thirdparty.invalid/";
+	assert(!trustedSender({ sender: contents, senderFrame: mainFrame }, origin, BrowserWindow));
+	assert.equal(browserServiceSpawnEnv({ LH_DESKTOP_API_SECRET: secret, LH_DESKTOP_SERVER: "1" }).LH_DESKTOP_API_SECRET, undefined);
+});

tests/desktop-platform.test.ts

@@ -10,6 +10,7 @@ import { describe, it } from "node:test";
 const require = createRequire(import.meta.url);
 const mainFile = fileURLToPath(new URL("../electron/main.cjs", import.meta.url));
 const mainSource = readFileSync(mainFile, "utf8");
+const mainRequire = createRequire(mainFile);
 
 /** Decline the single-instance lock so normal boot never spawns a server.
  * Then call the REAL createWindow with a stub Electron API and inspect its
@@ -27,7 +28,7 @@ function windowOptions(platform: string) {
 		session: { defaultSession: { setPermissionRequestHandler() {}, setPermissionCheckHandler() {} } },
 	};
 	const context = {
-		require: (id: string) => id === "electron" ? electron : require(id),
+		require: (id: string) => id === "electron" ? electron : mainRequire(id),
 		process: { platform, argv: [], env: {}, on() {} },
 		__dirname: dirname(mainFile),
 	};
@@ -77,7 +78,7 @@ describe("desktop window chrome", () => {
 		runInNewContext(readFileSync(new URL("../electron/preload.cjs", import.meta.url), "utf8"), context);
 		assert.equal(bridge.platform, "win32");
 		assert.deepEqual(Object.keys(bridge).sort(), [
-			"expandWindow", "getZoom", "installUpdate", "onUpdateStatus", "onZoom", "platform",
+			"chatgpt", "expandWindow", "getZoom", "installUpdate", "onUpdateStatus", "onZoom", "platform",
 		]);
 	});
 });

tests/zero-keys.test.tsx

@@ -76,7 +76,7 @@ describe("SetupKeysBody — stage 1's self-contained first-key screen", () => {
 		const html = renderToStaticMarkup(
 			createElement(SetupKeysBody, { ...base, providers: PROVIDERS, provider: null }),
 		);
-		assert.match(html, /Add a provider key\./);
+		assert.match(html, /Or add an API key\./);
 		assert.ok(html.includes("Choose your AI provider and paste its API key. You’ll pick your models next."));
 		assert.doesNotMatch(html, /First, add|keys unlock|—/);
 		assert.ok(html.includes('data-testid="setup-key-provider"'));