feat(coding-agent): added /extended-context slash command and pass settings to model registry

- Add the `/extended-context` builtin slash command to toggle premium long-context windows.
- Update `ModelRegistry` to accept settings and apply context-window caps using the finalized configuration.
- Add tests covering extended context toggle behavior and session context window capping.

packages/coding-agent/CHANGELOG.md

@@ -11,6 +11,7 @@
 - Backgroundable Python — `eval` cells can run async and auto-background like `bash`, with configurable thresholds.
 - Local Claude token counting — Anthropic-family tokens now count via a native local tokenizer, and every counter (session maintenance, advisor, stats, context tools) uses the active model's own tokenizer.
 - `extendedContext` setting — pick whether models with premium long-context pricing (272K/1M tiers on Codex-class models) use the extended window or compact early and stay on standard pricing.
+- `/extended-context` — toggle premium long-context windows without leaving the session.
 - Speculative compaction — with `compaction.asyncEnabled`, all compaction modes compact in parallel while the session continues, then splice the result in instantly.
 - `tokenizer` property on custom models and `modelOverrides` to pin the tokenizer family for proxy models.
 - `qwenTemplateReasoningEffort` in `models.yml` `compat` to disable the Qwen 3.8+ reasoning-effort template parameter for strict local servers.

packages/coding-agent/src/config/model-registry.ts

@@ -108,7 +108,7 @@ export {
 
 import { ModelsConfigFile, type ProviderValidationModel, validateProviderConfiguration } from "./models-config";
 import type { ModelOverride, ModelsConfig, ProviderAuthMode } from "./models-config-schema";
-import { settings } from "./settings";
+import { settings, type Settings } from "./settings";
 
 // DeviceCheck attestation (`x-oai-attestation`) for ChatGPT-OAuth Codex
 // requests; the pi-ai provider resolves it just-in-time per request.
@@ -132,21 +132,21 @@ interface CustomModelsResult {
  */
 type ModifyModelsHook = (models: Model<Api>[], credentials: OAuthCredentials) => Model<Api>[];
 
-function getDisabledProviderIdsFromSettings(): Set<string> {
+function getDisabledProviderIdsFromSettings(settingsInstance?: Settings): Set<string> {
 	try {
-		return new Set(settings.get("disabledProviders"));
+		return new Set((settingsInstance ?? settings).get("disabledProviders"));
 	} catch {
 		return new Set();
 	}
 }
 
 /**
- * Whether premium long-context windows are enabled. Defaults to true when the
- * settings singleton is not initialized (SDK embedding, early boot).
+ * Whether premium long-context windows are enabled. Defaults to true when no
+ * settings source is available (SDK embedding, early boot).
  */
-function isExtendedContextEnabledFromSettings(): boolean {
+function isExtendedContextEnabledFromSettings(settingsInstance?: Settings): boolean {
 	try {
-		return settings.get("extendedContext");
+		return (settingsInstance ?? settings).get("extendedContext");
 	} catch {
 		return true;
 	}
@@ -209,6 +209,7 @@ export class ModelRegistry {
 	#runtimeModelManagers: Map<string, { options: ModelManagerOptions<Api>; sourceId: string }> = new Map();
 	#ignoreLocalModelConfig: boolean;
 	#fetch: FetchImpl;
+	#settings: Settings | undefined;
 
 	#resolveCommandBackedApiKey(provider: string, options?: { forceCommandRefresh?: boolean }): CommandApiKeyResolution {
 		const keyConfig = this.#customProviderApiKeys.get(provider);
@@ -252,10 +253,13 @@ export class ModelRegistry {
 			 * must never apply client-side credential or routing overrides.
 			 */
 			ignoreLocalModelConfig?: boolean;
+			/** Settings source for availability and context-window policies. */
+			settings?: Settings;
 			fetch?: FetchImpl;
 		},
 	) {
 		this.#ignoreLocalModelConfig = options?.ignoreLocalModelConfig ?? false;
+		this.#settings = options?.settings;
 		this.#fetch =
 			options?.fetch ??
 			(isBunTestRuntime()
@@ -886,7 +890,7 @@ export class ModelRegistry {
 	}
 
 	#addImplicitDiscoverableProviders(configuredProviders: Set<string>): void {
-		const disabledProviders = getDisabledProviderIdsFromSettings();
+		const disabledProviders = getDisabledProviderIdsFromSettings(this.#settings);
 		if (!configuredProviders.has("ollama") && !disabledProviders.has("ollama")) {
 			this.#discoverableProviders.push({
 				provider: "ollama",
@@ -1058,7 +1062,7 @@ export class ModelRegistry {
 		strategy: ModelRefreshStrategy,
 		providerFilter?: ReadonlySet<string>,
 	): Promise<void> {
-		const disabledProviders = getDisabledProviderIdsFromSettings();
+		const disabledProviders = getDisabledProviderIdsFromSettings(this.#settings);
 		const selectedDiscoverableProviders = (
 			providerFilter
 				? this.#discoverableProviders.filter(provider => providerFilter.has(provider.provider))
@@ -1367,7 +1371,7 @@ export class ModelRegistry {
 					}),
 			},
 		];
-		const disabledProviders = getDisabledProviderIdsFromSettings();
+		const disabledProviders = getDisabledProviderIdsFromSettings(this.#settings);
 		const standardProviderDescriptors = PROVIDER_DESCRIPTORS.filter(descriptor => {
 			if (disabledProviders.has(descriptor.providerId)) return false;
 			if (configuredDiscoveryProviders.has(descriptor.providerId)) return false;
@@ -1592,7 +1596,7 @@ export class ModelRegistry {
 		});
 	}
 	#applyHardcodedModelPolicies(models: Model<Api>[]): Model<Api>[] {
-		const extendedContext = isExtendedContextEnabledFromSettings();
+		const extendedContext = isExtendedContextEnabledFromSettings(this.#settings);
 		return models.map(model => {
 			// Extended context off: cap models with a premium long-context price
 			// tier (e.g. GPT-5.6 bills 2x input above 272K) at the standard-pricing
@@ -1683,7 +1687,7 @@ export class ModelRegistry {
 	 * full bundled catalog (thousands of models, ~50 providers).
 	 */
 	#createProviderAvailabilityCheck(): (provider: string) => boolean {
-		const disabledProviders = getDisabledProviderIdsFromSettings();
+		const disabledProviders = getDisabledProviderIdsFromSettings(this.#settings);
 		const byProvider = new Map<string, boolean>();
 		return provider => {
 			let available = byProvider.get(provider);
@@ -1752,7 +1756,7 @@ export class ModelRegistry {
 	}
 
 	getDiscoverableProviders(): string[] {
-		const disabledProviders = getDisabledProviderIdsFromSettings();
+		const disabledProviders = getDisabledProviderIdsFromSettings(this.#settings);
 		return this.#discoverableProviders
 			.filter(provider => !disabledProviders.has(provider.provider))
 			.map(provider => provider.provider);
@@ -1770,7 +1774,7 @@ export class ModelRegistry {
 	hasProvider(providerId: string): boolean {
 		const providerModels = this.#hasFullSnapshot ? this.#models : this.#composeStaticModels(new Set([providerId]));
 		if (providerModels.some(model => model.provider === providerId)) return true;
-		if (getDisabledProviderIdsFromSettings().has(providerId)) return false;
+		if (getDisabledProviderIdsFromSettings(this.#settings).has(providerId)) return false;
 		return (
 			this.#discoverableProviders.some(provider => provider.provider === providerId) ||
 			this.#runtimeModelManagers.has(providerId)

packages/coding-agent/src/main.ts

@@ -1289,9 +1289,9 @@ export async function runRootCommand(
 	// tree; declare it so headless subagent optimizations (e.g. skipping replan
 	// title refresh) can tell a focusable process from a print/RPC/eval one.
 	setInteractiveHost(isInteractive);
-	// Create AuthStorage and ModelRegistry upfront. A configured-but-unreachable
-	// auth broker throws here; convert it to an actionable stderr message + clean
-	// exit instead of a raw uncaught stack trace (issue #8096).
+	// Create AuthStorage upfront. A configured-but-unreachable auth broker throws
+	// here; convert it to an actionable stderr message + clean exit instead of a
+	// raw uncaught stack trace (issue #8096).
 	let authStorage: AuthStorage;
 	try {
 		authStorage = await logger.time("discoverAuthStorage", deps.discoverAuthStorage ?? discoverAuthStorage);
@@ -1301,7 +1301,6 @@ export async function runRootCommand(
 		process.stderr.write(`${chalk.red(`Error: ${message}`)}\n`);
 		process.exit(1);
 	}
-	const modelRegistry = logger.time("modelRegistry:init", () => new ModelRegistry(authStorage));
 
 	const settingsInstance =
 		deps.settings ?? (await logger.time("settings:init", Settings.init, { cwd, configFiles: parsedArgs.config }));
@@ -1319,6 +1318,13 @@ export async function runRootCommand(
 	} else if (parsedArgs.mode === "acp") {
 		applyAcpDefaultSettingOverrides(settingsInstance);
 	}
+
+	// The registry composes policy-dependent metadata synchronously, including
+	// extended-context window caps, so it must receive the finalized settings.
+	const modelRegistry = logger.time(
+		"modelRegistry:init",
+		() => new ModelRegistry(authStorage, undefined, { settings: settingsInstance }),
+	);
 	if (parsedArgs.noPty || parsedArgs.mode === "rpc-ui") {
 		Bun.env.PI_NO_PTY = "1";
 	}

packages/coding-agent/src/sdk.ts

@@ -1241,12 +1241,19 @@ async function createAgentSessionScoped(options: CreateAgentSessionOptions): Pro
 	registerSshCleanup();
 	registerEvalCleanup();
 
+	const settings = await (options.settings ??
+		options.settingsManager ??
+		logger.time("settings", Settings.init, { cwd, agentDir }));
+	logger.time("initializeWithSettings", initializeWithSettings, settings);
+
 	// Pin authStorage to modelRegistry.authStorage: ModelRegistry.getApiKey() routes refresh
 	// failures through that instance, so any divergent storage handed to the bridge / mcpManager
 	// / session would silently miss credential_disabled events.
 	const modelRegistry =
 		options.modelRegistry ??
-		new ModelRegistry(options.authStorage ?? (await logger.time("discoverModels", discoverAuthStorage, agentDir)));
+		new ModelRegistry(options.authStorage ?? (await logger.time("discoverModels", discoverAuthStorage, agentDir)), undefined, {
+			settings,
+		});
 	// Track whether we internally created the authStorage so we can close it
 	// if construction fails before the session takes ownership.
 	const ownsAuthStorage = !options.authStorage && !options.modelRegistry;
@@ -1270,10 +1277,6 @@ async function createAgentSessionScoped(options: CreateAgentSessionOptions): Pro
 			startupCredentialDisabledEvents.push(event);
 		}
 	});
-	const settings = await (options.settings ??
-		options.settingsManager ??
-		logger.time("settings", Settings.init, { cwd, agentDir }));
-	logger.time("initializeWithSettings", initializeWithSettings, settings);
 	if (!options.modelRegistry) {
 		modelRegistry.refreshInBackground();
 	}

packages/coding-agent/src/slash-commands/builtin-modes.ts

@@ -5,7 +5,7 @@ import {
 	getModelMatchPreferences,
 	resolveCliModel,
 } from "../config/model-resolver";
-import type { SettingPath } from "../config/settings";
+import type { SettingPath, Settings } from "../config/settings";
 import { describeLoopLimitRuntime } from "../modes/loop-limit";
 import type { InteractiveModeContext } from "../modes/types";
 import type { AgentSession } from "../session/agent-session";
@@ -54,6 +54,32 @@ function formatFastModeStatus(session: AgentSession): string {
 	return session.isFastModeEnabled() ? "on" : "off";
 }
 
+/** `/extended-context status` label for the premium long-context window setting. */
+function formatExtendedContextStatus(settings: Settings): string {
+	return settings.get("extendedContext") ? "on" : "off";
+}
+
+/** Applies an `/extended-context` argument and returns its operator feedback. */
+function applyExtendedContextCommand(settings: Settings, args: string): string | undefined {
+	const arg = args.trim().toLowerCase();
+	const current = settings.get("extendedContext");
+	if (!arg || arg === "toggle") {
+		const enabled = !current;
+		settings.set("extendedContext", enabled);
+		return `Extended context ${enabled ? "enabled" : "disabled"}.`;
+	}
+	if (arg === "on") {
+		settings.set("extendedContext", true);
+		return "Extended context enabled.";
+	}
+	if (arg === "off") {
+		settings.set("extendedContext", false);
+		return "Extended context disabled.";
+	}
+	if (arg === "status") return `Extended context is ${formatExtendedContextStatus(settings)}.`;
+	return undefined;
+}
+
 /** Detailed, session-effective `/computer status` diagnostics. */
 async function formatComputerUseStatus(session: AgentSession): Promise<string> {
 	const enabled = session.settings.get("computer.enabled");
@@ -433,6 +459,32 @@ export const BUILTIN_MODE_SLASH_COMMANDS: ReadonlyArray<SlashCommandSpec> = [
 			runtime.ctx.editor.setText("");
 		},
 	},
+	{
+		name: "extended-context",
+		description: "Toggle premium long-context windows",
+		acpDescription: "Toggle extended context",
+		acpInputHint: "[on|off|status]",
+		subcommands: [
+			{ name: "on", description: "Enable premium long-context windows" },
+			{ name: "off", description: "Use standard-pricing context windows" },
+			{ name: "status", description: "Show extended context status" },
+		],
+		allowArgs: true,
+		getTuiAutocompleteDescription: runtime =>
+			`Extended context: ${formatExtendedContextStatus(runtime.ctx.settings)}`,
+		handle: async (command, runtime) => {
+			const output = applyExtendedContextCommand(runtime.settings, command.args);
+			if (!output) return usage("Usage: /extended-context [on|off|status]", runtime);
+			await runtime.output(output);
+			return commandConsumed();
+		},
+		handleTui: (command, runtime) => {
+			const output = applyExtendedContextCommand(runtime.ctx.settings, command.args);
+			refreshStatusLine(runtime.ctx);
+			runtime.ctx.showStatus(output ?? "Usage: /extended-context [on|off|status]");
+			runtime.ctx.editor.setText("");
+		},
+	},
 	{
 		name: "computer",
 		description: "Toggle the native computer-use tool for this session",

packages/coding-agent/test/acp-builtins.test.ts

@@ -243,6 +243,24 @@ describe("ACP builtin slash commands", () => {
 		expect(output).toEqual(["Fast mode is off."]);
 	});
 
+	it("toggles extended context with explicit controls and reports state", async () => {
+		const { output, runtime } = createRuntime();
+
+		expect(await executeAcpBuiltinSlashCommand("/extended-context off", runtime)).toEqual({ consumed: true });
+		expect(runtime.settings.get("extendedContext")).toBe(false);
+		expect(await executeAcpBuiltinSlashCommand("/extended-context on", runtime)).toEqual({ consumed: true });
+		expect(runtime.settings.get("extendedContext")).toBe(true);
+		expect(await executeAcpBuiltinSlashCommand("/extended-context", runtime)).toEqual({ consumed: true });
+		expect(runtime.settings.get("extendedContext")).toBe(false);
+		expect(await executeAcpBuiltinSlashCommand("/extended-context status", runtime)).toEqual({ consumed: true });
+		expect(output).toEqual([
+			"Extended context disabled.",
+			"Extended context enabled.",
+			"Extended context disabled.",
+			"Extended context is off.",
+		]);
+	});
+
 	it("forces a tool and returns remaining prompt text", async () => {
 		const { output, runtime } = createRuntime();
 

packages/coding-agent/test/available-commands.test.ts

@@ -47,6 +47,7 @@ describe("buildAvailableSlashCommands", () => {
 		expect(byName["reset-usage"]).toBeUndefined();
 
 		expect(byName.fast.description).toBe("Toggle fast mode");
+		expect(byName["extended-context"].description).toBe("Toggle extended context");
 		expect(byName["ext:hello"].description).toBe("Extension hello");
 		expect(byName["custom:hello"].description).toBe("Custom hello");
 		expect(byName["server:prompt"].description).toBe("MCP prompt");

packages/coding-agent/test/sdk-model-selection.test.ts

@@ -1029,6 +1029,40 @@ describe("createAgentSession deferred model pattern resolution", () => {
 		}
 	});
 
+	test("caps premium Codex context before a new session starts", async () => {
+		const authStorage = createInMemoryAuthStorage();
+		authStoragesToClose.push(authStorage);
+		authStorage.setRuntimeApiKey("openai-codex", "codex-oauth-token");
+
+		const { session } = await createAgentSession({
+			cwd: tempDir,
+			agentDir: tempDir,
+			authStorage,
+			settings: Settings.isolated({ extendedContext: false }),
+			sessionManager: SessionManager.inMemory(),
+			disableExtensionDiscovery: true,
+			skills: [],
+			contextFiles: [],
+			promptTemplates: [],
+			slashCommands: [],
+			enableMCP: false,
+			enableLsp: false,
+			skipPythonPreflight: true,
+			rules: [],
+			preloadedCustomToolPaths: [],
+			toolNames: ["read"],
+			modelPattern: "openai-codex/gpt-5.6-sol",
+		});
+
+		try {
+			expect(session.model?.provider).toBe("openai-codex");
+			expect(session.model?.id).toBe("gpt-5.6-sol");
+			expect(session.model?.contextWindow).toBe(272_000);
+		} finally {
+			await session.dispose();
+		}
+	});
+
 	test("restores role model max selector from extension provider after startup resume", async () => {
 		const defaultModel = getBundledModel("anthropic", "claude-sonnet-4-5");
 		if (!defaultModel) {