@selesai/code 0.13.13 → 0.13.14

package/dist/extensions/pi-subagents/src/profiles/profiles.ts

@@ -0,0 +1,662 @@
+import * as fs from "node:fs";
+import * as os from "node:os";
+import * as path from "node:path";
+import type { ExtensionAPI, ExtensionContext } from "@selesai/code";
+import { BUILTIN_AGENT_NAMES } from "../agents/agents.ts";
+import { getPiSpawnCommand } from "../runs/shared/pi-spawn.ts";
+import { findModelInfo, getSupportedThinkingLevels, splitKnownThinkingSuffix, toModelInfo } from "../shared/model-info.ts";
+import { getAgentDir } from "../shared/utils.ts";
+
+export const DEFAULT_PROVIDER_MODELS_MAX_AGE_DAYS = 7;
+
+type BuiltinAgentName = typeof BUILTIN_AGENT_NAMES[number];
+export type ProfileKind = "quota" | "quality";
+export type ProbeStatus = "ok" | "unavailable" | "auth" | "timeout" | "error" | "skipped";
+export type CostTier = "cheap" | "medium" | "expensive";
+export type QualityTier = "weak" | "medium" | "strong";
+export type LatencyTier = "fast" | "medium" | "slow";
+export type RecommendedRoleTier = "cheap" | "medium" | "strong";
+
+interface ProfileAgentOverride {
+	model?: string;
+	thinking?: string | false;
+	fallbackModels?: string[] | false;
+}
+
+export interface SubagentProfileFile {
+	subagents: {
+		agentOverrides: Record<string, ProfileAgentOverride>;
+		disableBuiltins?: boolean;
+		[key: string]: unknown;
+	};
+}
+
+export type ClassificationSource = "official-metadata" | "heuristic-name";
+
+export interface ProviderModelCatalogModel {
+	id: string;
+	fullId: string;
+	observed: {
+		availableInRegistry: boolean;
+		name?: string;
+		reasoning?: boolean;
+		thinkingLevels: string[];
+		contextWindow?: number;
+		maxTokens?: number;
+		cost?: {
+			input?: number;
+			output?: number;
+			cacheRead?: number;
+			cacheWrite?: number;
+		};
+		probe: {
+			status: ProbeStatus;
+			checkedAt: string;
+			message?: string;
+		};
+	};
+	derived: {
+		profileRank: number;
+		costTier: CostTier;
+		qualityTier: QualityTier;
+		latencyTier: LatencyTier;
+		recommendedRoleTier: RecommendedRoleTier;
+		recommendedAgents: BuiltinAgentName[];
+		classificationSources: ClassificationSource[];
+	};
+	warnings: string[];
+	notes: string[];
+}
+
+export interface ProviderModelCatalogFile {
+	provider: string;
+	refreshedAt: string;
+	maxAgeDays: number;
+	sources: string[];
+	models: ProviderModelCatalogModel[];
+}
+
+export interface ProfileCheckResult {
+	profileName: string;
+	filePath: string;
+	results: Array<{
+		agent: string;
+		model: string;
+		inRegistry: boolean;
+		probe: { status: ProbeStatus; message?: string };
+	}>;
+}
+
+function readJsonObjectFile(filePath: string): Record<string, unknown> {
+	const raw = fs.readFileSync(filePath, "utf-8");
+	const parsed = JSON.parse(raw) as unknown;
+	if (!parsed || typeof parsed !== "object" || Array.isArray(parsed)) {
+		throw new Error(`File '${filePath}' must contain a JSON object.`);
+	}
+	return parsed as Record<string, unknown>;
+}
+
+function writeJsonFile(filePath: string, value: unknown): void {
+	fs.mkdirSync(path.dirname(filePath), { recursive: true });
+	fs.writeFileSync(filePath, `${JSON.stringify(value, null, 2)}\n`, "utf-8");
+}
+
+const SAFE_PATH_TOKEN = /^[A-Za-z0-9][A-Za-z0-9._-]*$/;
+
+function normalizePathToken(value: string, label: string): string {
+	const trimmed = value.trim();
+	if (!trimmed) throw new Error(`${label} is required.`);
+	if (!SAFE_PATH_TOKEN.test(trimmed) || trimmed === "." || trimmed === ".." || trimmed.includes("/") || trimmed.includes("\\")) {
+		throw new Error(`${label} must be a safe file name using only letters, numbers, dots, underscores, and hyphens.`);
+	}
+	return trimmed;
+}
+
+function normalizeProfileName(name: string): string {
+	const trimmed = name.trim();
+	const stem = trimmed.endsWith(".json") ? trimmed.slice(0, -5) : trimmed;
+	return normalizePathToken(stem, "Profile name");
+}
+
+function normalizeProviderName(provider: string): string {
+	return normalizePathToken(provider, "Provider");
+}
+
+function validateSubagentProfile(filePath: string, parsed: Record<string, unknown>): SubagentProfileFile {
+	const subagents = parsed.subagents;
+	if (!subagents || typeof subagents !== "object" || Array.isArray(subagents)) {
+		throw new Error(`Profile '${filePath}' must contain a 'subagents' object.`);
+	}
+	const agentOverrides = (subagents as Record<string, unknown>).agentOverrides;
+	if (!agentOverrides || typeof agentOverrides !== "object" || Array.isArray(agentOverrides)) {
+		throw new Error(`Profile '${filePath}' must contain 'subagents.agentOverrides' as an object.`);
+	}
+	for (const [name, value] of Object.entries(agentOverrides)) {
+		if (!value || typeof value !== "object" || Array.isArray(value)) {
+			throw new Error(`Profile '${filePath}' has invalid override '${name}'; expected an object.`);
+		}
+		const override = value as Record<string, unknown>;
+		const model = override.model;
+		if (model !== undefined && typeof model !== "string") {
+			throw new Error(`Profile '${filePath}' has invalid model for '${name}'; expected a string.`);
+		}
+		const thinking = override.thinking;
+		if (thinking !== undefined && thinking !== false && typeof thinking !== "string") {
+			throw new Error(`Profile '${filePath}' has invalid thinking for '${name}'; expected a string or false.`);
+		}
+		const fallbackModels = override.fallbackModels;
+		if (fallbackModels !== undefined && fallbackModels !== false && (!Array.isArray(fallbackModels) || fallbackModels.some((item) => typeof item !== "string"))) {
+			throw new Error(`Profile '${filePath}' has invalid fallbackModels for '${name}'; expected an array of strings or false.`);
+		}
+	}
+	const disableBuiltins = (subagents as Record<string, unknown>).disableBuiltins;
+	if (disableBuiltins !== undefined && typeof disableBuiltins !== "boolean") {
+		throw new Error(`Profile '${filePath}' has invalid subagents.disableBuiltins; expected a boolean.`);
+	}
+	return parsed as unknown as SubagentProfileFile;
+}
+
+function getUserSettingsPath(): string {
+	return path.join(getAgentDir(), "settings.json");
+}
+
+function readSettingsFile(filePath: string): Record<string, unknown> {
+	if (!fs.existsSync(filePath)) return {};
+	return readJsonObjectFile(filePath);
+}
+
+function extractVersionScore(id: string): number {
+	const match = id.match(/(\d+(?:\.\d+)?)/g);
+	if (!match || match.length === 0) return 0;
+	return Math.max(...match.map((value) => Number.parseFloat(value)).filter((value) => Number.isFinite(value)));
+}
+
+function modelNameTokens(modelName: string): string[] {
+	return modelName
+		.toLowerCase()
+		.replace(/([a-z])([0-9])/g, "$1 $2")
+		.replace(/([0-9])([a-z])/g, "$1 $2")
+		.split(/[^a-z0-9.]+/)
+		.filter(Boolean);
+}
+
+function inferProfileBand(modelName: string): 0 | 1 | 2 | 3 | 4 {
+	const tokens = new Set(modelNameTokens(modelName));
+	if (["spark", "flash", "nano", "tiny", "instant"].some((token) => tokens.has(token))) return 0;
+	if (["mini", "haiku", "small"].some((token) => tokens.has(token))) return 1;
+	if (["opus", "max", "ultra", "pro"].some((token) => tokens.has(token))) return 4;
+	if (["sonnet", "turbo", "plus"].some((token) => tokens.has(token))) return 3;
+	return 2;
+}
+
+interface ModelClassificationInput {
+	id: string;
+	name?: string;
+	reasoning?: boolean;
+	contextWindow?: number;
+	maxTokens?: number;
+	cost?: {
+		input?: number;
+		output?: number;
+		cacheRead?: number;
+		cacheWrite?: number;
+	};
+}
+
+interface NumericStats {
+	min: number;
+	max: number;
+}
+
+interface ClassificationContext {
+	cost?: NumericStats;
+	contextWindow?: NumericStats;
+	maxTokens?: NumericStats;
+}
+
+function combinedCost(cost: ModelClassificationInput["cost"]): number | undefined {
+	if (!cost) return undefined;
+	const values = [cost.input, cost.output, cost.cacheRead, cost.cacheWrite].filter((value): value is number => typeof value === "number" && Number.isFinite(value));
+	if (values.length === 0) return undefined;
+	return values.reduce((sum, value) => sum + value, 0);
+}
+
+function collectStats(values: Array<number | undefined>): NumericStats | undefined {
+	const filtered = values.filter((value): value is number => typeof value === "number" && Number.isFinite(value));
+	if (filtered.length === 0) return undefined;
+	return { min: Math.min(...filtered), max: Math.max(...filtered) };
+}
+
+function normalize(value: number | undefined, stats: NumericStats | undefined): number | undefined {
+	if (value === undefined || !stats) return undefined;
+	if (stats.max <= stats.min) return 0.5;
+	return (value - stats.min) / (stats.max - stats.min);
+}
+
+function buildClassificationContext(models: ModelClassificationInput[]): ClassificationContext {
+	return {
+		cost: collectStats(models.map((model) => combinedCost(model.cost))),
+		contextWindow: collectStats(models.map((model) => model.contextWindow)),
+		maxTokens: collectStats(models.map((model) => model.maxTokens)),
+	};
+}
+
+function rankToCostTier(rank: number): CostTier {
+	if (rank <= 0.33) return "cheap";
+	if (rank <= 0.66) return "medium";
+	return "expensive";
+}
+
+function scoreToQualityTier(score: number): QualityTier {
+	if (score <= 0.33) return "weak";
+	if (score <= 0.66) return "medium";
+	return "strong";
+}
+
+function qualityTierToRoleTier(quality: QualityTier, cost: CostTier): RecommendedRoleTier {
+	if (quality === "strong") return "strong";
+	if (quality === "medium") return cost === "cheap" ? "cheap" : "medium";
+	return "cheap";
+}
+
+function agentsForRoleTier(roleTier: RecommendedRoleTier): BuiltinAgentName[] {
+	if (roleTier === "cheap") return ["scout", "delegate"];
+	if (roleTier === "medium") return ["researcher", "reviewer"];
+	return ["worker", "reviewer", "oracle"];
+}
+
+function classifyModel(input: ModelClassificationInput, context: ClassificationContext): {
+	profileRank: number;
+	costTier: CostTier;
+	qualityTier: QualityTier;
+	latencyTier: LatencyTier;
+	recommendedRoleTier: RecommendedRoleTier;
+	recommendedAgents: BuiltinAgentName[];
+	classificationSources: ClassificationSource[];
+} {
+	const modelName = input.name?.trim() || input.id;
+	const tokens = new Set(modelNameTokens(modelName));
+	const band = inferProfileBand(modelName);
+	const versionScore = extractVersionScore(input.id);
+	const costNorm = normalize(combinedCost(input.cost), context.cost);
+	const contextNorm = normalize(input.contextWindow, context.contextWindow);
+	const maxTokensNorm = normalize(input.maxTokens, context.maxTokens);
+	const hasOfficialMetadata = costNorm !== undefined || contextNorm !== undefined || maxTokensNorm !== undefined;
+	const classificationSources: ClassificationSource[] = hasOfficialMetadata
+		? ["official-metadata", "heuristic-name"]
+		: ["heuristic-name"];
+	const heuristicBase = band / 4;
+	const qualitySignals = [
+		heuristicBase,
+		...(contextNorm !== undefined ? [contextNorm] : []),
+		...(maxTokensNorm !== undefined ? [maxTokensNorm] : []),
+		...(input.reasoning === true ? [1] : []),
+		...(input.reasoning === false ? [0] : []),
+	];
+	const latencyHintsFast = tokens.has("highspeed") || tokens.has("flash") || tokens.has("instant") || tokens.has("turbo");
+	const latencyHintsSlow = tokens.has("pro") || tokens.has("ultra") || tokens.has("opus") || tokens.has("max");
+	let qualityScore = qualitySignals.reduce((sum, value) => sum + value, 0) / qualitySignals.length;
+	if (latencyHintsFast) {
+		qualityScore -= 0.2;
+	}
+	qualityScore = Math.max(0, Math.min(1, qualityScore));
+	const costTier = costNorm !== undefined
+		? rankToCostTier(costNorm)
+		: band === 0 ? "cheap" : band >= 3 ? "expensive" : "medium";
+	const qualityTier = scoreToQualityTier(qualityScore);
+	const latencyTier: LatencyTier = latencyHintsFast
+		? "fast"
+		: latencyHintsSlow
+			? "slow"
+			: costNorm !== undefined
+				? (costNorm <= 0.33 ? "fast" : costNorm <= 0.66 ? "medium" : "slow")
+				: (band <= 1 ? "fast" : band >= 3 ? "slow" : "medium");
+	const recommendedRoleTier = qualityTierToRoleTier(qualityTier, costTier);
+	const latencyPenalty = latencyHintsFast ? 125 : 0;
+	const profileRank = Math.round((qualityScore * 100) * 10) + Math.round(versionScore * 25) - latencyPenalty;
+	return {
+		profileRank,
+		costTier,
+		qualityTier,
+		latencyTier,
+		recommendedRoleTier,
+		recommendedAgents: agentsForRoleTier(recommendedRoleTier),
+		classificationSources,
+	};
+}
+
+function resolveProbeStatus(text: string, timedOut: boolean): ProbeStatus {
+	if (timedOut) return "timeout";
+	if (!text) return "error";
+	if (/(unauthori[sz]ed|forbidden|api key|auth|billing|credit|quota)/i.test(text)) return "auth";
+	if (/(not found|unknown model|model unavailable|model disabled|unsupported model|unavailable)/i.test(text)) return "unavailable";
+	return "error";
+}
+
+async function probeModel(
+	pi: Pick<ExtensionAPI, "exec"> | { exec?: ExtensionAPI["exec"] },
+	fullId: string,
+): Promise<{ status: ProbeStatus; message?: string }> {
+	if (typeof pi.exec !== "function") {
+		return { status: "skipped", message: "pi.exec is unavailable in this runtime." };
+	}
+	const spawnSpec = getPiSpawnCommand(["-p", "--model", fullId, "--no-tools", 'Reply with exactly "OK".']);
+	const result = await pi.exec(spawnSpec.command, spawnSpec.args, {
+		cwd: os.tmpdir(),
+		timeout: 45_000,
+	} as Record<string, unknown>);
+	const stdout = typeof result.stdout === "string" ? result.stdout.trim() : "";
+	const stderr = typeof result.stderr === "string" ? result.stderr.trim() : "";
+	const combined = [stderr, stdout].filter(Boolean).join("\n").trim();
+	if (result.code === 0) return { status: "ok", message: stdout || "Probe succeeded." };
+	return { status: resolveProbeStatus(combined, result.killed === true), message: combined || `Probe exited with code ${result.code ?? "unknown"}.` };
+}
+
+function roundIndex(count: number, position: number): number {
+	if (count <= 1) return 0;
+	return Math.max(0, Math.min(count - 1, Math.round((count - 1) * position)));
+}
+
+function profilePositions(kind: ProfileKind): { cheap: number; medium: number; strong: number } {
+	return kind === "quota"
+		? { cheap: 0, medium: 1 / 3, strong: 2 / 3 }
+		: { cheap: 1 / 3, medium: 2 / 3, strong: 1 };
+}
+
+function pickTierModels(models: ProviderModelCatalogModel[], kind: ProfileKind): { cheap: string; medium: string; strong: string } {
+	if (models.length === 0) throw new Error("No provider models are available for profile generation.");
+	const selectionPool = kind === "quota" && models.length > 1
+		? models.slice(0, -1)
+		: models;
+	const positions = profilePositions(kind);
+	return {
+		cheap: selectionPool[roundIndex(selectionPool.length, positions.cheap)]!.fullId,
+		medium: selectionPool[roundIndex(selectionPool.length, positions.medium)]!.fullId,
+		strong: selectionPool[roundIndex(selectionPool.length, positions.strong)]!.fullId,
+	};
+}
+
+function observedCombinedCost(model: ProviderModelCatalogModel): number | undefined {
+	return combinedCost(model.observed.cost);
+}
+
+function dominatesModel(a: ProviderModelCatalogModel, b: ProviderModelCatalogModel): boolean {
+	const costA = observedCombinedCost(a);
+	const costB = observedCombinedCost(b);
+	if (costA === undefined || costB === undefined) return false;
+	if (costA > costB) return false;
+	if (a.derived.profileRank < b.derived.profileRank) return false;
+	if ((a.observed.reasoning === true ? 1 : 0) < (b.observed.reasoning === true ? 1 : 0)) return false;
+	if ((a.observed.contextWindow ?? 0) < (b.observed.contextWindow ?? 0)) return false;
+	if ((a.observed.maxTokens ?? 0) < (b.observed.maxTokens ?? 0)) return false;
+	return costA < costB
+		|| a.derived.profileRank > b.derived.profileRank
+		|| (a.observed.reasoning === true && b.observed.reasoning !== true)
+		|| (a.observed.contextWindow ?? 0) > (b.observed.contextWindow ?? 0)
+		|| (a.observed.maxTokens ?? 0) > (b.observed.maxTokens ?? 0);
+}
+
+function filterDominatedModels(models: ProviderModelCatalogModel[]): ProviderModelCatalogModel[] {
+	return models.filter((candidate, index) => !models.some((other, otherIndex) => otherIndex !== index && dominatesModel(other, candidate)));
+}
+
+function buildProfileFile(models: { cheap: string; medium: string; strong: string }): SubagentProfileFile {
+	return {
+		subagents: {
+			agentOverrides: {
+				scout: { model: models.cheap },
+				delegate: { model: models.cheap },
+				researcher: { model: models.medium },
+				worker: { model: models.strong },
+				reviewer: { model: models.strong },
+				oracle: { model: models.strong },
+			},
+		},
+	};
+}
+
+function catalogModelIsUsable(model: ProviderModelCatalogModel): boolean {
+	return model.observed.availableInRegistry && model.observed.probe.status !== "unavailable" && model.observed.probe.status !== "auth" && model.observed.probe.status !== "timeout" && model.observed.probe.status !== "error";
+}
+
+function modelUsesHeuristicClassification(model: ProviderModelCatalogModel): boolean {
+	return model.derived.classificationSources.includes("heuristic-name")
+		&& !model.derived.classificationSources.includes("official-metadata");
+}
+
+function warningLineForHeuristicFallback(): string {
+	return "Classification fell back to name heuristics.";
+}
+
+export function countHeuristicFallbackModels(catalog: ProviderModelCatalogFile): number {
+	return catalog.models.filter(modelUsesHeuristicClassification).length;
+}
+
+function resolveProfilePath(name: string): string {
+	const dir = ensureSubagentProfilesDir();
+	return path.join(dir, `${normalizeProfileName(name)}.json`);
+}
+
+export function getSubagentProfilesRootDir(): string {
+	return path.join(getAgentDir(), "profiles", "pi-subagents");
+}
+
+export function getSubagentProfilesDir(): string {
+	return getSubagentProfilesRootDir();
+}
+
+export function ensureSubagentProfilesDir(): string {
+	const dir = getSubagentProfilesDir();
+	fs.mkdirSync(dir, { recursive: true });
+	return dir;
+}
+
+export function getProviderModelsDir(): string {
+	return path.join(getSubagentProfilesRootDir(), "providers");
+}
+
+export function ensureProviderModelsDir(): string {
+	const dir = getProviderModelsDir();
+	fs.mkdirSync(dir, { recursive: true });
+	return dir;
+}
+
+export function getProviderModelsPath(provider: string): string {
+	return path.join(ensureProviderModelsDir(), `${normalizeProviderName(provider)}.models.json`);
+}
+
+export function listSubagentProfiles(): string[] {
+	const dir = ensureSubagentProfilesDir();
+	return fs.readdirSync(dir, { withFileTypes: true })
+		.filter((entry) => entry.isFile() && entry.name.endsWith(".json"))
+		.map((entry) => entry.name.slice(0, -5))
+		.sort((a, b) => a.localeCompare(b));
+}
+
+export function readSubagentProfile(name: string): { filePath: string; profile: SubagentProfileFile } {
+	const filePath = resolveProfilePath(name);
+	if (!fs.existsSync(filePath)) throw new Error(`Profile not found: ${name}`);
+	const parsed = readJsonObjectFile(filePath);
+	return { filePath, profile: validateSubagentProfile(filePath, parsed) };
+}
+
+export function applySubagentProfile(name: string): { filePath: string; settingsPath: string } {
+	const { filePath, profile } = readSubagentProfile(name);
+	const settingsPath = getUserSettingsPath();
+	const settings = readSettingsFile(settingsPath);
+	const existing = settings.subagents && typeof settings.subagents === "object" && !Array.isArray(settings.subagents)
+		? settings.subagents as Record<string, unknown>
+		: {};
+	// A profile owns the complete agent mapping, but unrelated subagent settings
+	// (notably disableBuiltins, modelScope, watchdog, etc.) survive profile switches.
+	settings.subagents = {
+		...existing,
+		...profile.subagents,
+		agentOverrides: profile.subagents.agentOverrides,
+	};
+	writeJsonFile(settingsPath, settings);
+	return { filePath, settingsPath };
+}
+
+export function readProviderModelCatalog(provider: string): ProviderModelCatalogFile | null {
+	const filePath = getProviderModelsPath(provider);
+	if (!fs.existsSync(filePath)) return null;
+	return readJsonObjectFile(filePath) as unknown as ProviderModelCatalogFile;
+}
+
+export function isProviderModelCatalogStale(catalog: ProviderModelCatalogFile, maxAgeDays = DEFAULT_PROVIDER_MODELS_MAX_AGE_DAYS): boolean {
+	const refreshedAt = Date.parse(catalog.refreshedAt);
+	if (!Number.isFinite(refreshedAt)) return true;
+	const maxAgeMs = maxAgeDays * 24 * 60 * 60 * 1000;
+	return Date.now() - refreshedAt > maxAgeMs;
+}
+
+export async function refreshProviderModelCatalog(
+	pi: Pick<ExtensionAPI, "exec"> | { exec?: ExtensionAPI["exec"] },
+	ctx: Pick<ExtensionContext, "cwd" | "modelRegistry">,
+	provider: string,
+	options: { force?: boolean; maxAgeDays?: number; probe?: boolean } = {},
+): Promise<{ filePath: string; catalog: ProviderModelCatalogFile; reused: boolean; heuristicFallbackCount: number }> {
+	const normalizedProvider = normalizeProviderName(provider);
+	const maxAgeDays = options.maxAgeDays ?? DEFAULT_PROVIDER_MODELS_MAX_AGE_DAYS;
+	const filePath = getProviderModelsPath(normalizedProvider);
+	if (!options.force) {
+		const existing = readProviderModelCatalog(normalizedProvider);
+		if (existing && !isProviderModelCatalogStale(existing, maxAgeDays)) {
+			return { filePath, catalog: existing, reused: true, heuristicFallbackCount: countHeuristicFallbackModels(existing) };
+		}
+	}
+
+	const availableModels = ctx.modelRegistry.getAvailable().filter((model) => model.provider === normalizedProvider);
+	if (availableModels.length === 0) {
+		throw new Error(`No models found in the current registry for provider '${normalizedProvider}'.`);
+	}
+
+	const observedModels = [] as Array<{
+		rawModel: typeof availableModels[number];
+		modelRecord: Record<string, unknown> & { provider: string; id: string; name?: string };
+		fullId: string;
+		probe: { status: ProbeStatus; message?: string };
+	}>;
+	for (const rawModel of availableModels) {
+		const modelRecord = rawModel as unknown as Record<string, unknown> & { provider: string; id: string; name?: string };
+		const fullId = `${modelRecord.provider}/${modelRecord.id}`;
+		const probe = options.probe === false
+			? { status: "skipped" as const, message: "Live probing disabled." }
+			: await probeModel(pi, fullId);
+		observedModels.push({ rawModel, modelRecord, fullId, probe });
+	}
+	const classificationContext = buildClassificationContext(observedModels.map(({ modelRecord }) => ({
+		id: modelRecord.id,
+		...(typeof modelRecord.name === "string" ? { name: modelRecord.name } : {}),
+		...(typeof modelRecord.reasoning === "boolean" ? { reasoning: modelRecord.reasoning } : {}),
+		...(typeof modelRecord.contextWindow === "number" ? { contextWindow: modelRecord.contextWindow } : {}),
+		...(typeof modelRecord.maxTokens === "number" ? { maxTokens: modelRecord.maxTokens } : {}),
+		...(modelRecord.cost && typeof modelRecord.cost === "object" ? { cost: modelRecord.cost as ProviderModelCatalogModel["observed"]["cost"] } : {}),
+	})));
+	const models: ProviderModelCatalogModel[] = [];
+	for (const { rawModel, modelRecord, fullId, probe } of observedModels) {
+		const classification = classifyModel({
+			id: modelRecord.id,
+			...(typeof modelRecord.name === "string" ? { name: modelRecord.name } : {}),
+			...(typeof modelRecord.reasoning === "boolean" ? { reasoning: modelRecord.reasoning } : {}),
+			...(typeof modelRecord.contextWindow === "number" ? { contextWindow: modelRecord.contextWindow } : {}),
+			...(typeof modelRecord.maxTokens === "number" ? { maxTokens: modelRecord.maxTokens } : {}),
+			...(modelRecord.cost && typeof modelRecord.cost === "object" ? { cost: modelRecord.cost as ProviderModelCatalogModel["observed"]["cost"] } : {}),
+		}, classificationContext);
+		const warnings = classification.classificationSources.includes("heuristic-name") && !classification.classificationSources.includes("official-metadata")
+			? [warningLineForHeuristicFallback()]
+			: [];
+		models.push({
+			id: modelRecord.id,
+			fullId,
+			observed: {
+				availableInRegistry: true,
+				...(typeof modelRecord.name === "string" ? { name: modelRecord.name } : {}),
+				...(typeof modelRecord.reasoning === "boolean" ? { reasoning: modelRecord.reasoning } : {}),
+				thinkingLevels: getSupportedThinkingLevels(toModelInfo(rawModel)).map((level) => level),
+				...(typeof modelRecord.contextWindow === "number" ? { contextWindow: modelRecord.contextWindow } : {}),
+				...(typeof modelRecord.maxTokens === "number" ? { maxTokens: modelRecord.maxTokens } : {}),
+				...(modelRecord.cost && typeof modelRecord.cost === "object" ? { cost: modelRecord.cost as ProviderModelCatalogModel["observed"]["cost"] } : {}),
+				probe: {
+					status: probe.status,
+					checkedAt: new Date().toISOString(),
+					...(probe.message ? { message: probe.message } : {}),
+				},
+			},
+			derived: classification,
+			warnings,
+			notes: [],
+		});
+	}
+	models.sort((a, b) => a.derived.profileRank - b.derived.profileRank || a.fullId.localeCompare(b.fullId));
+	const catalog: ProviderModelCatalogFile = {
+		provider: normalizedProvider,
+		refreshedAt: new Date().toISOString(),
+		maxAgeDays,
+		sources: ["runtime-registry", ...(options.probe === false ? [] : ["live-probe"]), "heuristic-classifier"],
+		models,
+	};
+	writeJsonFile(filePath, catalog);
+	return { filePath, catalog, reused: false, heuristicFallbackCount: countHeuristicFallbackModels(catalog) };
+}
+
+export async function generateProfilesForProvider(
+	pi: Pick<ExtensionAPI, "exec"> | { exec?: ExtensionAPI["exec"] },
+	ctx: Pick<ExtensionContext, "cwd" | "modelRegistry">,
+	provider: string,
+	options: { maxAgeDays?: number; forceRefresh?: boolean; probe?: boolean } = {},
+): Promise<{ quotaPath: string; qualityPath: string; catalogPath: string; quotaModels: { cheap: string; medium: string; strong: string }; qualityModels: { cheap: string; medium: string; strong: string }; heuristicFallbackCount: number; selectedHeuristicFallbackCount: number }> {
+	const normalizedProvider = normalizeProviderName(provider);
+	const { filePath: catalogPath, catalog, heuristicFallbackCount } = await refreshProviderModelCatalog(pi, ctx, normalizedProvider, {
+		maxAgeDays: options.maxAgeDays,
+		force: options.forceRefresh,
+		probe: options.probe,
+	});
+	const usableModels = catalog.models.filter(catalogModelIsUsable);
+	const profileModels = filterDominatedModels(usableModels);
+	if (profileModels.length === 0) {
+		throw new Error(`Provider '${normalizedProvider}' has no usable models after filtering.`);
+	}
+	const quotaModels = pickTierModels(profileModels, "quota");
+	const qualityModels = pickTierModels(profileModels, "quality");
+	const dir = ensureSubagentProfilesDir();
+	const quotaPath = path.join(dir, `${normalizedProvider}.quota.json`);
+	const qualityPath = path.join(dir, `${normalizedProvider}.quality.json`);
+	writeJsonFile(quotaPath, buildProfileFile(quotaModels));
+	writeJsonFile(qualityPath, buildProfileFile(qualityModels));
+	const selectedModels = new Set([...Object.values(quotaModels), ...Object.values(qualityModels)]);
+	const selectedHeuristicFallbackCount = profileModels.filter((model) => selectedModels.has(model.fullId) && modelUsesHeuristicClassification(model)).length;
+	return { quotaPath, qualityPath, catalogPath, quotaModels, qualityModels, heuristicFallbackCount, selectedHeuristicFallbackCount };
+}
+
+export async function checkSubagentProfile(
+	pi: Pick<ExtensionAPI, "exec"> | { exec?: ExtensionAPI["exec"] },
+	ctx: Pick<ExtensionContext, "cwd" | "modelRegistry">,
+	name: string,
+): Promise<ProfileCheckResult> {
+	const { filePath, profile } = readSubagentProfile(name);
+	const availableModels = ctx.modelRegistry.getAvailable().map(toModelInfo);
+	const entries = Object.entries(profile.subagents.agentOverrides)
+		.filter(([, value]) => typeof value?.model === "string" && value.model.trim())
+		.map(([agent, value]) => ({ agent, model: value.model!.trim() }));
+	const probeCache = new Map<string, { status: ProbeStatus; message?: string }>();
+	const results: ProfileCheckResult["results"] = [];
+	for (const entry of entries) {
+		const modelInfo = findModelInfo(entry.model, availableModels);
+		const { thinkingSuffix } = splitKnownThinkingSuffix(entry.model);
+		const probeModelId = modelInfo ? `${modelInfo.fullId}${thinkingSuffix}` : entry.model;
+		let probe = probeCache.get(probeModelId);
+		if (!probe) {
+			probe = await probeModel(pi, probeModelId);
+			probeCache.set(probeModelId, probe);
+		}
+		results.push({
+			agent: entry.agent,
+			model: entry.model,
+			inRegistry: modelInfo !== undefined,
+			probe,
+		});
+	}
+	return { profileName: name, filePath, results };
+}