@f5-sales-demo/xcsh 21.49.4 → 22.0.0

package/README.md

@@ -10,9 +10,6 @@ For installation, setup, provider configuration, model roles, slash commands, an
 Package-specific references:
 
 - [CHANGELOG](./CHANGELOG.md)
-- [MCP configuration guide](../../docs/mcp-config.md)
-- [MCP runtime lifecycle](../../docs/mcp-runtime-lifecycle.md)
-- [MCP server/tool authoring](../../docs/mcp-server-tool-authoring.md)
 - [Remote-control implementation](./src/remote-control/IMPLEMENTATION.md)
 - [Remote-control acceptance](./src/remote-control/ACCEPTANCE.md)
 - [CONTRIBUTING](../../CONTRIBUTING.md)

package/package.json

@@ -1,7 +1,7 @@
 {
 	"type": "module",
 	"name": "@f5-sales-demo/xcsh",
-	"version": "21.49.4",
+	"version": "22.0.0",
 	"description": "Coding agent CLI with read, bash, edit, write tools and session management",
 	"homepage": "https://github.com/f5-sales-demo/xcsh",
 	"author": "Can Boluk",
@@ -67,13 +67,13 @@
 	},
 	"dependencies": {
 		"@agentclientprotocol/sdk": "1.4.0",
-		"@f5-sales-demo/pi-agent-core": "21.49.4",
-		"@f5-sales-demo/pi-ai": "21.49.4",
-		"@f5-sales-demo/pi-natives": "21.49.4",
-		"@f5-sales-demo/pi-resource-management": "21.49.4",
-		"@f5-sales-demo/pi-tui": "21.49.4",
-		"@f5-sales-demo/pi-utils": "21.49.4",
-		"@f5-sales-demo/xcsh-stats": "21.49.4",
+		"@f5-sales-demo/pi-agent-core": "22.0.0",
+		"@f5-sales-demo/pi-ai": "22.0.0",
+		"@f5-sales-demo/pi-natives": "22.0.0",
+		"@f5-sales-demo/pi-resource-management": "22.0.0",
+		"@f5-sales-demo/pi-tui": "22.0.0",
+		"@f5-sales-demo/pi-utils": "22.0.0",
+		"@f5-sales-demo/xcsh-stats": "22.0.0",
 		"@mozilla/readability": "^0.6",
 		"@resvg/resvg-wasm": "2.6.2",
 		"@sinclair/typebox": "0.34.52",

package/src/browser/headless-bridge.ts

@@ -180,7 +180,6 @@ export async function startHeadlessChatBridge(
 					hasUI: false,
 					toolNames: [...new Set([...BROWSER_TOOL_NAMES, ...EXTENSION_AGENT_TOOL_NAMES])],
 					customTools: createExtensionBridgeTools(bridge),
-					enableMCP: false,
 					enableLsp: false,
 					disableExtensionDiscovery: true,
 					...(process.env.XCSH_BENCH_EXTENSION
@@ -201,7 +200,6 @@ export async function startHeadlessChatBridge(
 				sessionManager: SessionManager.inMemory(cwd),
 				toolNames: [...OFFICE_TOOL_NAMES],
 				customTools: [],
-				enableMCP: false,
 				enableLsp: false,
 				disableExtensionDiscovery: true,
 				bundledExtensions: ["sandbox-guard"],

package/src/browser/host-profiles.ts

@@ -102,7 +102,7 @@ SAFETY — NEVER DO THESE:
 const OFFICE_NATIVE_TOOLS_NOTE = `
 NATIVE TOOLS: Beyond the document host tools, you have xcsh's full local toolset — \`bash\` (run shell commands, including CLIs like \`az\`, \`gh\`, \`terraform\`, \`git\` when installed and authenticated), file tools (\`read\`/\`write\`/\`edit\`), and \`grep\` — plus any skills available in this workspace. Reach for them when the task genuinely needs them (pull live data with a CLI, read a local file the user points you at). Prefer the document host tools for document work. Your file tools and shell are confined to the folder xcsh was launched from.
 
-This pane runs NO MCP servers and exposes NO plugin-provided TOOLS — the tools listed above (plus the document host tools) are everything you can call. To use a cloud or SCM CLI, invoke it directly with \`bash\` (e.g. \`az account show\`, \`gh repo view\`) rather than hunting for an MCP server that would provide it. The user sees your narration, so don't describe a missing MCP tool as a failure — just use the CLI.
+This pane exposes no plugin-provided tools — the tools listed above (plus the document host tools) are everything you can call. To use a cloud or SCM CLI, invoke it directly with \`bash\` (e.g. \`az account show\`, \`gh repo view\`).
 
 Plugin RESOURCES are a different thing and they ARE available to you: an installed plugin's skills, slash commands, schemas, templates and engines. Read them through \`xcsh://plugin/<name>\` (summary), \`xcsh://plugin/<name>/<key>\` (a declared resource such as \`schema\`), or \`xcsh://plugin/<name>/file/<path>\` (any file in the plugin), and run a plugin's engine with \`bash\` (e.g. \`bun xcsh://plugin/<name>/file/engine/cli.ts <command>\`). When a plugin declares a deterministic engine, use it — never recompute by hand what the engine computes.
 

package/src/capability/extension.ts

@@ -1,10 +1,9 @@
 /**
  * Extensions Capability
  *
- * Gemini-style extensions that provide MCP servers, tools, and context.
+ * Gemini-style extensions that provide tools and context.
  */
 import { defineCapability } from ".";
-import type { MCPServer } from "./mcp";
 import type { SourceMeta } from "./types";
 
 /**
@@ -13,7 +12,6 @@ import type { SourceMeta } from "./types";
 export interface ExtensionManifest {
 	name?: string;
 	description?: string;
-	mcpServers?: Record<string, Omit<MCPServer, "name" | "_source">>;
 	tools?: unknown[];
 	context?: unknown;
 }
@@ -37,7 +35,7 @@ export interface Extension {
 export const extensionCapability = defineCapability<Extension>({
 	id: "extensions",
 	displayName: "Extensions",
-	description: "Gemini-style extensions providing MCP servers, tools, and context",
+	description: "Gemini-style extensions providing tools and context",
 	key: ext => ext.name,
 	validate: ext => {
 		if (!ext.name) return "Missing extension name";

package/src/capability/types.ts

@@ -2,8 +2,8 @@
  * Core types for the capability-based config discovery system.
  *
  * This architecture inverts control: instead of callers knowing about paths like
- * `.xcsh`, `.codex`, `.gemini`, they simply ask for `load("mcps")` and get back
- * a unified array of MCP servers.
+ * `.xcsh`, `.codex`, `.gemini`, they simply ask for a capability and get back
+ * a unified array of items.
  */
 
 /**
@@ -33,7 +33,7 @@ export interface LoadResult<T> {
  * A provider that can load items for a capability.
  */
 export interface Provider<T> {
-	/** Unique provider ID (e.g., "xcsh", "mcp-json", "agents-md") */
+	/** Unique provider ID (e.g., "xcsh", "codex", "agents-md") */
 	id: string;
 
 	/** Human-readable name for UI display (e.g., "xcsh", "OpenAI Codex") */
@@ -47,7 +47,7 @@ export interface Provider<T> {
 	 * Suggested ranges:
 	 *   100+ : Primary providers (xcsh, pi)
 	 *   50-99: Tool-specific providers (claude, codex, gemini)
-	 *   1-49 : Shared standards (mcp-json, agents-md)
+	 *   1-49 : Shared standards (agents-md)
 	 */
 	priority: number;
 
@@ -110,10 +110,10 @@ export interface CapabilityResult<T> {
  * Definition of a capability.
  */
 export interface Capability<T> {
-	/** Capability ID (e.g., "mcps", "skills", "context-files") */
+	/** Capability ID (e.g., "skills", "context-files") */
 	id: string;
 
-	/** Human-readable name for UI display (e.g., "MCP Servers", "Skills") */
+	/** Human-readable name for UI display (e.g., "Skills", "Context Files") */
 	displayName: string;
 
 	/** Short description for settings/status UI */

package/src/cli/args.ts

@@ -3,7 +3,6 @@
  */
 import { type Effort, THINKING_EFFORTS } from "@f5-sales-demo/pi-ai";
 import { APP_NAME, CONFIG_DIR_NAME, logger } from "@f5-sales-demo/pi-utils";
-import { CliUsageError } from "@f5-sales-demo/pi-utils/cli";
 import chalk from "chalk";
 import { parseEffort } from "../thinking";
 import {
@@ -50,8 +49,6 @@ export interface Args {
 	models?: string[];
 	tools?: string[];
 	noTools?: boolean;
-	mcp?: boolean;
-	noMcp?: boolean;
 	noLsp?: boolean;
 	noPty?: boolean;
 	hooks?: string[];
@@ -161,12 +158,6 @@ const APPLY: Record<LaunchFlagName, (result: Args, value: string | true) => void
 	"no-tools": r => {
 		r.noTools = true;
 	},
-	mcp: r => {
-		r.mcp = true;
-	},
-	"no-mcp": r => {
-		r.noMcp = true;
-	},
 	"no-lsp": r => {
 		r.noLsp = true;
 	},
@@ -404,8 +395,6 @@ export function parseArgs(args: string[], extensionFlags?: ExtensionFlagRegistry
 		result.unrecognizedFlags.push({ token, name: name ?? token.replace(/^-+/, "") });
 	}
 
-	if (result.mcp && result.noMcp) throw new CliUsageError("--mcp cannot be combined with --no-mcp");
-	if (result.mcp && result.noTools) throw new CliUsageError("--mcp cannot be combined with --no-tools");
 	return result;
 }
 

package/src/cli/flag-spec.ts

@@ -79,8 +79,6 @@ export const LAUNCH_FLAGS = defineFlags({
 	"provider-session-id": { arity: "value", description: "Resume a provider-side session", hidden: true },
 	models: { arity: "value", description: "Comma-separated model patterns for Ctrl+P cycling" },
 	"no-tools": { arity: "boolean", description: "Disable all built-in tools" },
-	mcp: { arity: "boolean", description: "Enable MCP servers for this run" },
-	"no-mcp": { arity: "boolean", description: "Disable MCP server discovery and tools" },
 	"no-lsp": { arity: "boolean", description: "Disable LSP tools, formatting, and diagnostics" },
 	"no-pty": { arity: "boolean", description: "Disable PTY-based interactive bash execution" },
 	tools: { arity: "value", description: "Comma-separated list of tools to enable (default: all)" },

package/src/commit/agentic/agent.ts

@@ -63,7 +63,6 @@ export async function runCommitAgentSession(input: CommitAgentInput): Promise<Co
 		systemPrompt,
 		customTools: tools,
 		enableLsp: false,
-		enableMCP: false,
 		hasUI: false,
 		spawns,
 		toolNames: ["__none__"],

package/src/commit/agentic/index.ts

@@ -31,8 +31,8 @@ export async function runAgenticCommit(args: CommitCommandArgs): Promise<void> {
 	process.stdout.write("● Resolving model...\n");
 	const modelRegistry = new ModelRegistry(authStorage, undefined, {
 		getProviderOrder: () => settings.get("modelProviderOrder"),
-		getOpenAICodexMaxContext: () => settings.get("providers.openaiCodexMaxContext"),
-		getLiteLLMMaxContext: () => settings.get("providers.litellmMaxContext"),
+		getOpenAIContextTier: () => settings.get("providers.openaiContextTier"),
+		getLiteLLMContextTier: () => settings.get("providers.litellmContextTier"),
 	});
 	await modelRegistry.refresh();
 	const stagedFilesPromise = (async () => {

package/src/commit/pipeline.ts

@@ -44,8 +44,8 @@ async function runLegacyCommitCommand(args: CommitCommandArgs): Promise<void> {
 	const authStorage = await discoverAuthStorage();
 	const modelRegistry = new ModelRegistry(authStorage, undefined, {
 		getProviderOrder: () => settings.get("modelProviderOrder"),
-		getOpenAICodexMaxContext: () => settings.get("providers.openaiCodexMaxContext"),
-		getLiteLLMMaxContext: () => settings.get("providers.litellmMaxContext"),
+		getOpenAIContextTier: () => settings.get("providers.openaiContextTier"),
+		getLiteLLMContextTier: () => settings.get("providers.litellmContextTier"),
 	});
 	await modelRegistry.refresh();
 

package/src/config/auto-config.ts

@@ -20,7 +20,7 @@ import { hardenAgentConfigFileSync, writeAgentConfigFileSync } from "./agent-con
 import { DEFAULT_MODEL_ROLE } from "./settings-schema";
 
 /** Current config schema version. Bump when the generated format changes. */
-export const CURRENT_CONFIG_VERSION = 9;
+export const CURRENT_CONFIG_VERSION = 11;
 const LITELLM_CONFIG_DIR_MODE = 0o700;
 const LITELLM_MODELS_FILE_MODE = 0o600;
 const GENERATED_LITELLM_MARKER = "# Auto-generated by xcsh for LiteLLM proxy";
@@ -123,6 +123,7 @@ export function generateModelsYml(baseUrl: string, options?: GenerateModelsYmlOp
 		"      - gpt-6-astra",
 		"      - gpt-5.6-luna",
 		"      - gpt-5.6-sol",
+		"      - gpt-5.5",
 		"    models:",
 		"      - id: gpt-6-luna",
 		"        name: GPT-6 Luna",
@@ -130,7 +131,8 @@ export function generateModelsYml(baseUrl: string, options?: GenerateModelsYmlOp
 		`        baseUrl: "${responsesBaseUrl}"`,
 		"      - id: gpt-5.6-terra",
 		"        name: GPT-5.6 Terra",
-		"        api: openai-completions",
+		"        api: openai-responses",
+		`        baseUrl: "${responsesBaseUrl}"`,
 		"      - id: gpt-6-sol",
 		"        name: GPT-6 Sol",
 		"        api: openai-responses",
@@ -141,10 +143,16 @@ export function generateModelsYml(baseUrl: string, options?: GenerateModelsYmlOp
 		`        baseUrl: "${responsesBaseUrl}"`,
 		"      - id: gpt-5.6-luna",
 		"        name: GPT-5.6 Luna",
-		"        api: openai-completions",
+		"        api: openai-responses",
+		`        baseUrl: "${responsesBaseUrl}"`,
 		"      - id: gpt-5.6-sol",
 		"        name: GPT-5.6 Sol",
-		"        api: openai-completions",
+		"        api: openai-responses",
+		`        baseUrl: "${responsesBaseUrl}"`,
+		"      - id: gpt-5.5",
+		"        name: GPT-5.5",
+		"        api: openai-responses",
+		`        baseUrl: "${responsesBaseUrl}"`,
 		"    picker:",
 		"      groupId: litellm",
 		"      groupLabel: LiteLLM",
@@ -251,6 +259,23 @@ export function generateModelsYml(baseUrl: string, options?: GenerateModelsYmlOp
 		"          supportsTemperature: false",
 		"      gpt-5.6-luna:",
 		"        reasoning: true",
+		"        input:",
+		"          - text",
+		"          - image",
+		"        contextWindow: 1050000",
+		"        maxTokens: 128000",
+		"        compat:",
+		"          supportsTemperature: false",
+		"        cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 }",
+		"      gpt-5.5:",
+		"        reasoning: true",
+		"        input:",
+		"          - text",
+		"          - image",
+		"        contextWindow: 1050000",
+		"        maxTokens: 128000",
+		"        compat:",
+		"          supportsTemperature: false",
 		"        cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 }",
 	];
 

package/src/config/model-catalog.ts

@@ -10,9 +10,16 @@ function compareVersions(left: string, right: string): number {
 	return 0;
 }
 
-/** Keep historical catalogs available while presenting only current user-facing families. */
-const DEFAULT_OPENAI_CODEX_MODEL_IDS = new Set(["gpt-5.6-terra", "gpt-6-astra", "gpt-6-luna", "gpt-6-sol"]);
-const CURRENT_LITELLM_OPENAI_MODEL_IDS = new Set(["gpt-6-luna", "gpt-5.6-terra", "gpt-6-sol", "gpt-6-astra"]);
+/** Present the exact model-interaction contract when each provider advertises the model. */
+const CODEX_INTERACTION_MODEL_IDS = new Set([
+	"gpt-6-astra",
+	"gpt-6-sol",
+	"gpt-6-luna",
+	"gpt-5.6-sol",
+	"gpt-5.6-terra",
+	"gpt-5.6-luna",
+	"gpt-5.5",
+]);
 
 /** Shared catalog boundary for the TUI and attached human interfaces. */
 export function filterCurrentBrowserModels(models: readonly Model[]): Model[] {
@@ -39,9 +46,10 @@ export function filterCurrentBrowserModels(models: readonly Model[]): Model[] {
 	return models.filter(model => {
 		const gpt = model.id.match(/^gpt-(\d+(?:\.\d+)?)(?:-|$)/i);
 		if (gpt?.[1]) {
+			if (model.provider === "openai-codex" || model.provider === "litellm") {
+				return CODEX_INTERACTION_MODEL_IDS.has(model.id.toLowerCase());
+			}
 			if (compareVersions(gpt[1], "5.6") < 0) return false;
-			if (model.provider === "openai-codex") return DEFAULT_OPENAI_CODEX_MODEL_IDS.has(model.id.toLowerCase());
-			if (model.provider === "litellm") return CURRENT_LITELLM_OPENAI_MODEL_IDS.has(model.id.toLowerCase());
 		}
 		const claude = model.id.match(/^claude-(opus|sonnet|haiku)-(\d+)(?:[.-](\d+))?(?:-|$)/i);
 		if (claude?.[1] && claude[2]) {

package/src/config/model-registry.ts

@@ -2,7 +2,9 @@ import * as path from "node:path";
 import {
 	type Api,
 	type AssistantMessageEventStream,
+	applyCodexInteractionMetadata,
 	applyGeneratedModelPolicies,
+	type CodexContextTier,
 	type Context,
 	createModelManager,
 	DEFAULT_LOCAL_TOKEN,
@@ -23,6 +25,7 @@ import {
 	readModelCache,
 	registerCustomApi,
 	registerOAuthProvider,
+	resolveCodexContextBudget,
 	type SimpleStreamOptions,
 	type ThinkingConfig,
 	unregisterCustomApis,
@@ -50,17 +53,14 @@ export type { CanonicalModelIndex, CanonicalModelRecord, CanonicalModelVariant,
 
 export const kNoAuth = NO_AUTH_API_KEY;
 
-const OPENAI_CODEX_STANDARD_CONTEXT_WINDOW = 272_000;
-const OPENAI_CODEX_MAX_CONTEXT_WINDOW = 1_050_000;
-const OPENAI_CODEX_MAX_CONTEXT_MODEL_IDS = new Set(["gpt-6-luna", "gpt-6-sol"]);
-const LITELLM_STANDARD_CONTEXT_WINDOW = 272_000;
-const LITELLM_MAX_CONTEXT_WINDOW = 1_050_000;
-const LITELLM_MAX_CONTEXT_MODEL_IDS = new Set(["gpt-6-luna", "gpt-5.6-terra", "gpt-6-sol", "gpt-6-astra"]);
-
 export function isAuthenticated(apiKey: string | undefined | null): apiKey is string {
 	return Boolean(apiKey) && apiKey !== kNoAuth;
 }
 
+function positiveInteger(value: unknown): number | undefined {
+	return typeof value === "number" && Number.isFinite(value) && value > 0 ? Math.floor(value) : undefined;
+}
+
 export type ModelRole = "default" | "smol" | "slow" | "vision" | "plan" | "designer" | "commit" | "task";
 
 export interface ModelRoleInfo {
@@ -878,8 +878,8 @@ export class ModelRegistry {
 	#lastDiscoveryWarnings: Map<string, string> = new Map();
 	#hasProbed = false;
 	#refreshQueue: Promise<void> = Promise.resolve();
-	#openAICodexMaxContext = false;
-	#litellmMaxContext = false;
+	#openAIContextTier: CodexContextTier = "standard";
+	#litellmContextTier: CodexContextTier = "standard";
 
 	#enqueueRefresh(operation: () => Promise<void>): Promise<void> {
 		const pending = this.#refreshQueue.then(operation);
@@ -901,8 +901,8 @@ export class ModelRegistry {
 		modelsPath?: string,
 		private readonly options: {
 			getProviderOrder?: () => readonly string[];
-			getOpenAICodexMaxContext?: () => boolean;
-			getLiteLLMMaxContext?: () => boolean;
+			getOpenAIContextTier?: () => CodexContextTier;
+			getLiteLLMContextTier?: () => CodexContextTier;
 		} = {},
 	) {
 		this.#modelsConfigFile = ModelsConfigFile.relocate(modelsPath);
@@ -1055,8 +1055,8 @@ export class ModelRegistry {
 	}
 
 	#loadModels() {
-		this.#openAICodexMaxContext = this.options.getOpenAICodexMaxContext?.() ?? this.#openAICodexMaxContext;
-		this.#litellmMaxContext = this.options.getLiteLLMMaxContext?.() ?? this.#litellmMaxContext;
+		this.#openAIContextTier = this.options.getOpenAIContextTier?.() ?? this.#openAIContextTier;
+		this.#litellmContextTier = this.options.getLiteLLMContextTier?.() ?? this.#litellmContextTier;
 		// Load custom models from models.json first (to know which providers to override)
 		const {
 			models: customModels = [],
@@ -1090,42 +1090,39 @@ export class ModelRegistry {
 		const combined = this.#mergeCustomModels(withConfigModels, this.#runtimeModelOverlays);
 
 		this.#models = this.#applyProviderModelAllowlists(this.#applyModelOverrides(combined, this.#modelOverrides));
-		this.#applyOpenAICodexContextWindow();
-		this.#applyLiteLLMContextWindow();
+		this.#applyOpenAIContextTier();
+		this.#applyLiteLLMContextTier();
 		this.#rebuildCanonicalIndex();
 	}
 
-	#applyOpenAICodexContextWindow(): void {
-		const contextWindow = this.#openAICodexMaxContext
-			? OPENAI_CODEX_MAX_CONTEXT_WINDOW
-			: OPENAI_CODEX_STANDARD_CONTEXT_WINDOW;
+	#applyContextTier(provider: "openai-codex" | "litellm", tier: CodexContextTier): void {
 		for (const model of this.#models) {
-			if (model.provider === "openai-codex" && OPENAI_CODEX_MAX_CONTEXT_MODEL_IDS.has(model.id)) {
-				model.contextWindow = contextWindow;
-			}
+			if (model.provider !== provider) continue;
+			Object.assign(model, applyCodexInteractionMetadata(model));
+			const budget = resolveCodexContextBudget(model.id, tier, model.providerContextWindow);
+			if (budget) Object.assign(model, budget);
 		}
 	}
 
-	/** Update GPT-6 subscription context limits without rebuilding the registry. */
-	setOpenAICodexMaxContext(enabled: boolean): void {
-		this.#openAICodexMaxContext = enabled;
-		this.#applyOpenAICodexContextWindow();
+	#applyOpenAIContextTier(): void {
+		this.#applyContextTier("openai-codex", this.#openAIContextTier);
+	}
+
+	/** Update subscription context limits without rebuilding the registry. */
+	setOpenAIContextTier(tier: CodexContextTier): void {
+		this.#openAIContextTier = tier;
+		this.#applyOpenAIContextTier();
 		this.#rebuildCanonicalIndex();
 	}
 
-	#applyLiteLLMContextWindow(): void {
-		const contextWindow = this.#litellmMaxContext ? LITELLM_MAX_CONTEXT_WINDOW : LITELLM_STANDARD_CONTEXT_WINDOW;
-		for (const model of this.#models) {
-			if (model.provider === "litellm" && LITELLM_MAX_CONTEXT_MODEL_IDS.has(model.id)) {
-				model.contextWindow = contextWindow;
-			}
-		}
+	#applyLiteLLMContextTier(): void {
+		this.#applyContextTier("litellm", this.#litellmContextTier);
 	}
 
 	/** Update current internal OpenAI context limits without changing subscription models. */
-	setLiteLLMMaxContext(enabled: boolean): void {
-		this.#litellmMaxContext = enabled;
-		this.#applyLiteLLMContextWindow();
+	setLiteLLMContextTier(tier: CodexContextTier): void {
+		this.#litellmContextTier = tier;
+		this.#applyLiteLLMContextTier();
 		this.#rebuildCanonicalIndex();
 	}
 
@@ -1198,6 +1195,13 @@ export class ModelRegistry {
 		return merged;
 	}
 
+	#filterOverlaysBySuccessfulDiscovery(models: CustomModelOverlay[]): CustomModelOverlay[] {
+		return models.filter(model => {
+			const discovery = this.#providerDiscoveryStates.get(model.provider);
+			return discovery?.status !== "ok" || discovery.models.includes(model.id);
+		});
+	}
+
 	#applyProviderModelAllowlists(models: Model<Api>[]): Model<Api>[] {
 		return models.filter(model => {
 			const allowlist = this.#modelAllowlists.get(model.provider);
@@ -1420,9 +1424,15 @@ export class ModelRegistry {
 		]);
 		const discovered = [...configuredDiscovered, ...builtInDiscovered];
 		// Replace only catalogs which completed successfully, including an empty response.
+		const successfulDiscoveryProviders = providerFilter
+			? [...providerFilter].filter(provider => this.#providerDiscoveryStates.get(provider)?.status === "ok")
+			: [...this.#providerDiscoveryStates.values()]
+					.filter(state => state.status === "ok")
+					.map(state => state.provider);
 		const refreshed = new Set([
 			...selectedDiscoverableProviders.map(provider => provider.provider),
 			...discovered.map(model => model.provider),
+			...successfulDiscoveryProviders,
 		]);
 		const discoveredModels = this.#applyHardcodedModelPolicies(
 			discovered.map(model => {
@@ -1452,12 +1462,18 @@ export class ModelRegistry {
 			model => !refreshed.has(model.provider) || this.#providerDiscoveryStates.get(model.provider)?.status !== "ok",
 		);
 		const resolved = this.#mergeResolvedModels(this.#models, discoveredModels);
-		const withConfigModels = this.#mergeCustomModels(resolved, this.#customModelOverlays);
+		const withConfigModels = this.#mergeCustomModels(
+			resolved,
+			this.#filterOverlaysBySuccessfulDiscovery(this.#customModelOverlays),
+		);
 		// Merge runtime extension models so they survive online discovery completion
-		const combined = this.#mergeCustomModels(withConfigModels, this.#runtimeModelOverlays);
+		const combined = this.#mergeCustomModels(
+			withConfigModels,
+			this.#filterOverlaysBySuccessfulDiscovery(this.#runtimeModelOverlays),
+		);
 		this.#models = this.#applyProviderModelAllowlists(this.#applyModelOverrides(combined, this.#modelOverrides));
-		this.#applyOpenAICodexContextWindow();
-		this.#applyLiteLLMContextWindow();
+		this.#applyOpenAIContextTier();
+		this.#applyLiteLLMContextTier();
 		this.#rebuildCanonicalIndex();
 	}
 
@@ -2047,7 +2063,13 @@ export class ModelRegistry {
 			throw new Error(`HTTP ${response.status} from ${modelsUrl}`);
 		}
 		const payload = await response.json();
-		const items: Array<{ id: string; contextWindow?: number }> =
+		const items: Array<{
+			id: string;
+			contextWindow?: number;
+			context_window?: unknown;
+			max_input_tokens?: unknown;
+			max_output_tokens?: unknown;
+		}> =
 			providerConfig.provider === "vllm"
 				? parseVllmModelsPayload(payload)
 				: ((payload as { data?: Array<{ id: string }> }).data ?? []);
@@ -2055,6 +2077,10 @@ export class ModelRegistry {
 		for (const item of items) {
 			const id = item.id;
 			if (!id) continue;
+			const advertisedContextWindow =
+				positiveInteger(item.context_window) ?? positiveInteger(item.max_input_tokens) ?? item.contextWindow;
+			const advertisedOutputLimit = positiveInteger(item.max_output_tokens);
+			const contextWindow = advertisedContextWindow ?? 128000;
 			discovered.push(
 				enrichModelThinking({
 					id,
@@ -2065,11 +2091,13 @@ export class ModelRegistry {
 					reasoning: false,
 					input: ["text"],
 					cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 },
-					contextWindow: item.contextWindow ?? 128000,
+					contextWindow,
 					maxTokens:
-						item.contextWindow === undefined
+						advertisedOutputLimit ??
+						(advertisedContextWindow === undefined
 							? 8192
-							: Math.max(1, Math.min(8192, Math.floor(item.contextWindow / 4))),
+							: Math.max(1, Math.min(8192, Math.floor(advertisedContextWindow / 4)))),
+					providerContextWindow: advertisedContextWindow,
 					headers,
 				}),
 			);
@@ -2150,7 +2178,8 @@ export class ModelRegistry {
 		// bundled corrections. Reapply generated-model policy before a cached entry
 		// replaces its bundled peer; explicit user overrides still run afterward.
 		applyGeneratedModelPolicies(models, { preserveDiscoveredThinking });
-		return models.map(model => {
+		return models.map(rawModel => {
+			const model = applyCodexInteractionMetadata(rawModel);
 			if (model.id !== "gpt-5.4" || model.provider === "github-copilot") {
 				return model;
 			}

package/src/config/model-resolver.ts

@@ -1116,6 +1116,20 @@ export function resolveCliModel(options: {
 				error: undefined,
 			};
 		}
+		const thinkingSeparator = pattern.lastIndexOf(":");
+		const exactModelId =
+			thinkingSeparator !== -1 && parseThinkingLevel(pattern.slice(thinkingSeparator + 1))
+				? pattern.slice(0, thinkingSeparator)
+				: pattern;
+		if (resolveProviderModelReference(provider, exactModelId, allModels)) {
+			return {
+				model: undefined,
+				selector: undefined,
+				thinkingLevel: undefined,
+				warning: undefined,
+				error: `Model "${provider}/${exactModelId}" is unavailable. Configure or sign in to provider "${provider}".`,
+			};
+		}
 	}
 
 	const candidates = provider ? availableModels.filter(model => model.provider === provider) : availableModels;