pi-free 2.0.2 → 2.0.5

package/providers/nvidia/nvidia.ts

@@ -1,415 +1,474 @@
-/**
- * NVIDIA NIM Provider Extension
- *
- * Provides access to NVIDIA-hosted large models via integrate.api.nvidia.com.
- * All models use NVIDIA's free credit system — requires NVIDIA_API_KEY.
- * Get a free key at: https://build.nvidia.com
- *
- * Small models (< 70B) are filtered out to keep the list focused on useful
- * chat/coding models. Non-chat models (embedding, speech-to-text, OCR,
- * image-gen) are filtered by their modalities (output must be ["text"],
- * input must include "text").
- *
- * Responds to global free-only filter for free/paid model filtering.
- */
-
-import type {
-	ExtensionAPI,
-	ProviderModelConfig,
-} from "@mariozechner/pi-coding-agent";
-import {
-	applyHidden,
-	getNvidiaApiKey,
-	loadConfigFile,
-	PROVIDER_NVIDIA,
-	saveConfig,
-} from "../../config.ts";
-import {
-	BASE_URL_NVIDIA,
-	DEFAULT_FETCH_TIMEOUT_MS,
-	NVIDIA_MIN_SIZE_B,
-	URL_MODELS_DEV,
-} from "../../constants.ts";
-import { registerWithGlobalToggle } from "../../lib/registry.ts";
-import type { ModelsDevModel, ModelsDevProvider } from "../../lib/types.ts";
-import {
-	fetchWithRetry,
-	fetchWithTimeout,
-	isUsableModel,
-} from "../../lib/util.ts";
-import { createReRegister, enhanceWithCI } from "../../provider-helper.ts";
-
-// =============================================================================
-// Non-chat model heuristics for models not in models.dev
-// =============================================================================
-
-const NVIDIA_NON_CHAT_PATTERNS: RegExp[] = [
-	/embed(?!.*instruct)/i,
-	/whisper/i,
-	/reward/i,
-	/ocr(?!.*instruct)/i,
-	/safety-guard|content-safety|nemoguard/i,
-	/retriever-parse|nemotron-parse(?!.*instruct)/i,
-	/detector/i,
-	/deplot/i,
-	/nvclip/i,
-	/vila$/i,
-	/neva(?!.*instruct)/i,
-	/translate/i,
-	/cosmos-reason/i,
-	/kosmos/i,
-	/bge-/i,
-	/arctic-embed/i,
-	/gliner/i,
-	/nv-embed/i,
-	/embedqa/i,
-	/embedcode/i,
-];
-
-/**
- * Models that appear in NVIDIA's /v1/models but return 404 "Function not found"
- * on /v1/chat/completions. These are listed but not actually provisioned for
- * hosted chat inference. Community-reported; add new IDs as they surface.
- *
- * Users can also hide individual models via hidden_models in ~/.pi/free.json.
- */
-const NVIDIA_KNOWN_404_MODELS: ReadonlySet<string> = new Set([
-	"01-ai/yi-large",
-	"adept/fuyu-8b",
-	"ai21labs/jamba-1.5-large-instruct",
-	"aisingapore/sea-lion-7b-instruct",
-	"baai/bge-m3",
-	"bigcode/starcoder2-15b",
-	"databricks/dbrx-instruct",
-	"deepseek-ai/deepseek-coder-6.7b-instruct",
-	"google/codegemma-1.1-7b",
-	"google/codegemma-7b",
-	"google/deplot",
-	"google/gemma-2b",
-	"google/recurrentgemma-2b",
-	"ibm/granite-3.0-3b-a800m-instruct",
-	"ibm/granite-3.0-8b-instruct",
-	"ibm/granite-34b-code-instruct",
-	"ibm/granite-8b-code-instruct",
-	"meta/codellama-70b",
-	"meta/llama2-70b",
-	"microsoft/kosmos-2",
-	"microsoft/phi-3-vision-128k-instruct",
-	"microsoft/phi-3.5-moe-instruct",
-	"mistralai/codestral-22b-instruct-v0.1",
-	"mistralai/mistral-7b-instruct-v0.3",
-	"mistralai/mistral-large",
-	"mistralai/mistral-large-2-instruct",
-	"mistralai/mixtral-8x22b-v0.1",
-	"nv-mistralai/mistral-nemo-12b-instruct",
-	"nvidia/cosmos-reason2-8b",
-	"nvidia/embed-qa-4",
-	"nvidia/llama-3.1-nemotron-51b-instruct",
-	"nvidia/llama-3.1-nemotron-70b-instruct",
-	"nvidia/llama-3.1-nemotron-ultra-253b-v1",
-	"nvidia/llama-3.2-nemoretriever-1b-vlm-embed-v1",
-	"nvidia/llama-3.2-nemoretriever-300m-embed-v1",
-	"nvidia/llama-3.2-nv-embedqa-1b-v1",
-	"nvidia/llama-3.2-nv-embedqa-1b-v2",
-	"nvidia/llama-nemotron-embed-1b-v2",
-	"nvidia/llama-nemotron-embed-vl-1b-v2",
-	"nvidia/llama3-chatqa-1.5-70b",
-	"nvidia/mistral-nemo-minitron-8b-8k-instruct",
-	"nvidia/nemotron-4-340b-instruct",
-	"nvidia/nemotron-4-340b-reward",
-	"nvidia/nemotron-nano-3-30b-a3b",
-	"nvidia/neva-22b",
-	"nvidia/nv-embed-v1",
-	"nvidia/nv-embedcode-7b-v1",
-	"nvidia/nv-embedqa-e5-v5",
-	"nvidia/nv-embedqa-mistral-7b-v2",
-	"nvidia/nvclip",
-	"nvidia/riva-translate-4b-instruct",
-	"snowflake/arctic-embed-l",
-	"writer/palmyra-creative-122b",
-	"writer/palmyra-fin-70b-32k",
-	"writer/palmyra-med-70b",
-	"writer/palmyra-med-70b-32k",
-	"zyphra/zamba2-7b-instruct",
-]);
-
-/**
- * Infer model metadata from a NVIDIA model ID for models not present in
- * models.dev. Returns null if the ID matches known non-chat patterns.
- */
-function inferModelFromId(id: string): ModelsDevModel | null {
-	for (const pattern of NVIDIA_NON_CHAT_PATTERNS) {
-		if (pattern.test(id)) return null;
-	}
-
-	const name = id
-		.split("/")
-		.pop()!
-		.replace(/-/g, " ")
-		.replace(/\b\w/g, (c) => c.toUpperCase())
-		.replace(/\b(\d+(?:\.\d+)?)b\b/gi, "$1B");
-
-	const hasVision = /vision|multimodal|vl/i.test(id);
-	const hasReasoning = /reason|r1|thinking/i.test(id);
-
-	return {
-		id,
-		name,
-		reasoning: hasReasoning,
-		limit: { context: 128_000, output: 4096 },
-		modalities: {
-			input: hasVision ? ["text", "image"] : ["text"],
-			output: ["text"],
-		},
-		cost: { input: 0, output: 0 },
-	};
-}
-
-// =============================================================================
-// Fetch + map
-// =============================================================================
-
-async function fetchNvidiaModels(
-	apiKey?: string,
-): Promise<ProviderModelConfig[]> {
-	// ── 1. Query NVIDIA's actual API (source of truth) ─────────────────
-	let apiModelIds = new Set<string>();
-	if (apiKey) {
-		try {
-			const response = await fetchWithRetry(
-				`${BASE_URL_NVIDIA}/models`,
-				{
-					headers: {
-						Authorization: `Bearer ${apiKey}`,
-						"User-Agent": "pi-free-providers",
-					},
-				},
-				3,
-				1000,
-				DEFAULT_FETCH_TIMEOUT_MS,
-			);
-			if (response.ok) {
-				const json = (await response.json()) as {
-					data?: Array<{ id: string }>;
-				};
-				if (json.data) {
-					apiModelIds = new Set(json.data.map((m) => m.id));
-				}
-			}
-		} catch (error) {
-			console.error("[nvidia] Failed to fetch models from NVIDIA API", error);
-		}
-	}
-
-	// ── 2. Fetch models.dev for rich metadata (cost, limits, etc.) ─────
-	const devModels = new Map<string, ModelsDevModel>();
-	try {
-		const response = await fetchWithRetry(
-			URL_MODELS_DEV,
-			{
-				headers: { "User-Agent": "pi-free-providers" },
-			},
-			3,
-			1000,
-			DEFAULT_FETCH_TIMEOUT_MS,
-		);
-		if (response.ok) {
-			const json = (await response.json()) as Record<string, ModelsDevProvider>;
-			const provider = Object.values(json).find((p) => p?.id === "nvidia");
-			if (provider?.models) {
-				for (const m of Object.values(provider.models)) {
-					devModels.set(m.id, m);
-				}
-			}
-		}
-	} catch (error) {
-		console.error("[nvidia] Failed to fetch models.dev", error);
-	}
-
-	// ── 3. Build unified list (NVIDIA API wins; fallback to models.dev) ─
-	const modelIds =
-		apiModelIds.size > 0 ? [...apiModelIds] : [...devModels.keys()];
-
-	const result = applyHidden(
-		modelIds
-			.map((id) => {
-				const dev = devModels.get(id);
-				if (dev) return dev;
-				return inferModelFromId(id);
-			})
-			.filter((m): m is ModelsDevModel => m !== null)
-			.filter((m) => isUsableModel(m.id, NVIDIA_MIN_SIZE_B))
-			.filter((m) => {
-				const modalities = m.modalities;
-				if (modalities) {
-					const output = modalities.output ?? [];
-					const input = modalities.input ?? [];
-					if (!output.includes("text")) return false;
-					if (!input.includes("text")) return false;
-				}
-				return true;
-			})
-			// Filter out known 404 models (listed but not provisioned for chat)
-			.filter((m) => {
-				if (NVIDIA_KNOWN_404_MODELS.has(m.id)) {
-					return false;
-				}
-				return true;
-			})
-			// NVIDIA is freemium — all models are usable with free credits.
-			// No cost filtering applied.
-			.map(
-				(m): ProviderModelConfig => ({
-					id: m.id,
-					name: m.name,
-					reasoning: m.reasoning,
-					input: m.modalities?.input?.includes("image")
-						? ["text", "image"]
-						: ["text"],
-					cost: {
-						input: m.cost?.input ?? 0,
-						output: m.cost?.output ?? 0,
-						cacheRead: m.cost?.cache_read ?? 0,
-						cacheWrite: m.cost?.cache_write ?? 0,
-					},
-					contextWindow: m.limit.context,
-					maxTokens: m.limit.output,
-				}),
-			),
-		PROVIDER_NVIDIA,
-	);
-
-	return result;
-}
-
-// =============================================================================
-// Extension Entry Point
-// =============================================================================
-
-/**
- * Probe a single NVIDIA model with a minimal chat request.
- * Returns true if the model is routable (not 404), false if it 404s.
- */
-async function probeNvidiaModel(
-	apiKey: string,
-	modelId: string,
-): Promise<boolean> {
-	try {
-		const response = await fetchWithTimeout(
-			`${BASE_URL_NVIDIA}/chat/completions`,
-			{
-				method: "POST",
-				headers: {
-					Authorization: `Bearer ${apiKey}`,
-					"Content-Type": "application/json",
-					"User-Agent": "pi-free-providers",
-				},
-				body: JSON.stringify({
-					model: modelId,
-					messages: [{ role: "user", content: "hi" }],
-					max_tokens: 1,
-				}),
-			},
-			10000, // 10 second timeout
-		);
-		// 404 = function not found (model not provisioned)
-		// 200/400/401/etc = at least routable
-		return response.status !== 404;
-	} catch {
-		return true; // Network errors / timeouts are not "model not found"
-	}
-}
-
-export default async function (pi: ExtensionAPI) {
-	const apiKey = getNvidiaApiKey();
-	const hasKey = !!apiKey;
-
-	let allModels: ProviderModelConfig[] = [];
-
-	try {
-		allModels = await fetchNvidiaModels(apiKey);
-	} catch (error) {
-		console.error("[nvidia] Failed to fetch models at startup", error);
-		return;
-	}
-
-	// Store both sets for global toggle (same list — NVIDIA is freemium)
-	const stored = { free: allModels, all: allModels };
-
-	// Create re-register function
-	const reRegister = createReRegister(pi, {
-		providerId: PROVIDER_NVIDIA,
-		baseUrl: BASE_URL_NVIDIA,
-		apiKey: apiKey || "NVIDIA_API_KEY",
-	});
-
-	// Register with global toggle system
-	registerWithGlobalToggle(PROVIDER_NVIDIA, stored, reRegister, hasKey);
-
-	// Register initial models (global toggle will apply filter if needed)
-	const initialModels = allModels;
-	pi.registerProvider(PROVIDER_NVIDIA, {
-		baseUrl: BASE_URL_NVIDIA,
-		apiKey: apiKey || "NVIDIA_API_KEY",
-		api: "openai-completions" as const,
-		authHeader: true,
-		headers: {
-			"User-Agent": "pi-free-providers",
-		},
-		models: enhanceWithCI(initialModels),
-	});
-
-	// ── Probe command: test all registered models for 404s ─────────────
-	pi.registerCommand("probe-nvidia", {
-		description: "Test all NVIDIA models for 404 'Function not found' errors",
-		handler: async (_args, ctx) => {
-			if (!apiKey) {
-				ctx.ui.notify("NVIDIA_API_KEY not set", "error");
-				return;
-			}
-
-			const modelsToTest = allModels;
-			ctx.ui.notify(`Probing ${modelsToTest.length} NVIDIA models…`, "info");
-
-			const notFound: string[] = [];
-			const batchSize = 5;
-
-			for (let i = 0; i < modelsToTest.length; i += batchSize) {
-				const batch = modelsToTest.slice(i, i + batchSize);
-				const results = await Promise.all(
-					batch.map(async (m) => {
-						const ok = await probeNvidiaModel(apiKey, m.id);
-						return { id: m.id, ok };
-					}),
-				);
-				for (const r of results) {
-					if (!r.ok) notFound.push(r.id);
-				}
-			}
-
-			if (notFound.length === 0) {
-				ctx.ui.notify("All NVIDIA models are routable ✅", "info");
-				return;
-			}
-
-			// Auto-hide 404 models in config (provider-scoped)
-			const config = loadConfigFile();
-			const existingHidden = new Set(config.hidden_models ?? []);
-			for (const id of notFound) existingHidden.add(`${PROVIDER_NVIDIA}/${id}`);
-			saveConfig({ hidden_models: Array.from(existingHidden) });
-
-			// Re-register so hidden models disappear immediately
-			const filtered = await fetchNvidiaModels(apiKey);
-			stored.free = filtered;
-			stored.all = filtered;
-			reRegister(filtered);
-
-			ctx.ui.notify(
-				`Found ${notFound.length} broken models (auto-hidden):\n${notFound.join("\n")}`,
-				"warning",
-			);
-		},
-	});
-
-	// Registration complete - models registered silently (use LOG_LEVEL=info to see details)
-}
+/**
+ * NVIDIA NIM Provider Extension
+ *
+ * Provides access to NVIDIA-hosted large models via integrate.api.nvidia.com.
+ * All models use NVIDIA's free credit system — requires NVIDIA_API_KEY.
+ * Get a free key at: https://build.nvidia.com
+ *
+ * Small models (< 70B) are filtered out to keep the list focused on useful
+ * chat/coding models. Non-chat models (embedding, speech-to-text, OCR,
+ * image-gen) are filtered by their modalities (output must be ["text"],
+ * input must include "text").
+ *
+ * Responds to global free-only filter for free/paid model filtering.
+ */
+
+import type {
+	ExtensionAPI,
+	ProviderModelConfig,
+} from "@mariozechner/pi-coding-agent";
+import {
+	applyHidden,
+	getNvidiaApiKey,
+	loadConfigFile,
+	PROVIDER_NVIDIA,
+	saveConfig,
+} from "../../config.ts";
+import {
+	BASE_URL_NVIDIA,
+	DEFAULT_FETCH_TIMEOUT_MS,
+	NVIDIA_MIN_SIZE_B,
+	URL_MODELS_DEV,
+} from "../../constants.ts";
+import { createLogger } from "../../lib/logger.ts";
+import { isFreeModel, registerWithGlobalToggle } from "../../lib/registry.ts";
+import type { ModelsDevModel, ModelsDevProvider } from "../../lib/types.ts";
+import {
+	fetchWithRetry,
+	fetchWithTimeout,
+	isUsableModel,
+} from "../../lib/util.ts";
+import { createReRegister, enhanceWithCI } from "../../provider-helper.ts";
+
+// =============================================================================
+// Non-chat model heuristics for models not in models.dev
+// =============================================================================
+
+const NVIDIA_NON_CHAT_PATTERNS: RegExp[] = [
+	/embed(?!.*instruct)/i,
+	/whisper/i,
+	/reward/i,
+	/ocr(?!.*instruct)/i,
+	/safety-guard|content-safety|nemoguard/i,
+	/retriever-parse|nemotron-parse(?!.*instruct)/i,
+	/detector/i,
+	/deplot/i,
+	/nvclip/i,
+	/vila$/i,
+	/neva(?!.*instruct)/i,
+	/translate/i,
+	/cosmos-reason/i,
+	/kosmos/i,
+	/bge-/i,
+	/arctic-embed/i,
+	/gliner/i,
+	/nv-embed/i,
+	/embedqa/i,
+	/embedcode/i,
+];
+
+/**
+ * Models that appear in NVIDIA's /v1/models but return 404 "Function not found"
+ * on /v1/chat/completions. These are listed but not actually provisioned for
+ * hosted chat inference. Community-reported; add new IDs as they surface.
+ *
+ * Users can also hide individual models via hidden_models in ~/.pi/free.json.
+ */
+const NVIDIA_KNOWN_404_MODELS: ReadonlySet<string> = new Set([
+	"01-ai/yi-large",
+	"adept/fuyu-8b",
+	"ai21labs/jamba-1.5-large-instruct",
+	"aisingapore/sea-lion-7b-instruct",
+	"baai/bge-m3",
+	"bigcode/starcoder2-15b",
+	"databricks/dbrx-instruct",
+	"deepseek-ai/deepseek-coder-6.7b-instruct",
+	"google/codegemma-1.1-7b",
+	"google/codegemma-7b",
+	"google/deplot",
+	"google/gemma-2b",
+	"google/recurrentgemma-2b",
+	"ibm/granite-3.0-3b-a800m-instruct",
+	"ibm/granite-3.0-8b-instruct",
+	"ibm/granite-34b-code-instruct",
+	"ibm/granite-8b-code-instruct",
+	"meta/codellama-70b",
+	"meta/llama2-70b",
+	"microsoft/kosmos-2",
+	"microsoft/phi-3-vision-128k-instruct",
+	"microsoft/phi-3.5-moe-instruct",
+	"mistralai/codestral-22b-instruct-v0.1",
+	"mistralai/mistral-7b-instruct-v0.3",
+	"mistralai/mistral-large",
+	"mistralai/mistral-large-2-instruct",
+	"mistralai/mixtral-8x22b-v0.1",
+	"nv-mistralai/mistral-nemo-12b-instruct",
+	"nvidia/cosmos-reason2-8b",
+	"nvidia/embed-qa-4",
+	"nvidia/llama-3.1-nemotron-51b-instruct",
+	"nvidia/llama-3.1-nemotron-70b-instruct",
+	"nvidia/llama-3.1-nemotron-ultra-253b-v1",
+	"nvidia/llama-3.2-nemoretriever-1b-vlm-embed-v1",
+	"nvidia/llama-3.2-nemoretriever-300m-embed-v1",
+	"nvidia/llama-3.2-nv-embedqa-1b-v1",
+	"nvidia/llama-3.2-nv-embedqa-1b-v2",
+	"nvidia/llama-nemotron-embed-1b-v2",
+	"nvidia/llama-nemotron-embed-vl-1b-v2",
+	"nvidia/llama3-chatqa-1.5-70b",
+	"nvidia/mistral-nemo-minitron-8b-8k-instruct",
+	"nvidia/nemotron-4-340b-instruct",
+	"nvidia/nemotron-4-340b-reward",
+	"nvidia/nemotron-nano-3-30b-a3b",
+	"nvidia/neva-22b",
+	"nvidia/nv-embed-v1",
+	"nvidia/nv-embedcode-7b-v1",
+	"nvidia/nv-embedqa-e5-v5",
+	"nvidia/nv-embedqa-mistral-7b-v2",
+	"nvidia/nvclip",
+	"nvidia/riva-translate-4b-instruct",
+	"snowflake/arctic-embed-l",
+	"writer/palmyra-creative-122b",
+	"writer/palmyra-fin-70b-32k",
+	"writer/palmyra-med-70b",
+	"writer/palmyra-med-70b-32k",
+	"zyphra/zamba2-7b-instruct",
+]);
+
+/**
+ * Infer model metadata from a NVIDIA model ID for models not present in
+ * models.dev. Returns null if the ID matches known non-chat patterns.
+ */
+function inferModelFromId(id: string): ModelsDevModel | null {
+	for (const pattern of NVIDIA_NON_CHAT_PATTERNS) {
+		if (pattern.test(id)) return null;
+	}
+
+	const name = id
+		.split("/")
+		.pop()!
+		.replace(/-/g, " ")
+		.replace(/\b\w/g, (c) => c.toUpperCase())
+		.replace(/\b(\d+(?:\.\d+)?)b\b/gi, "$1B");
+
+	const hasVision = /vision|multimodal|vl/i.test(id);
+	const hasReasoning = /reason|r1|thinking/i.test(id);
+
+	return {
+		id,
+		name,
+		reasoning: hasReasoning,
+		limit: { context: 128_000, output: 4096 },
+		modalities: {
+			input: hasVision ? ["text", "image"] : ["text"],
+			output: ["text"],
+		},
+		cost: { input: 0, output: 0 },
+	};
+}
+
+// =============================================================================
+// Fetch + map
+// =============================================================================
+
+async function fetchNvidiaApiModelIds(apiKey: string): Promise<Set<string>> {
+	try {
+		const response = await fetchWithRetry(
+			`${BASE_URL_NVIDIA}/models`,
+			{
+				headers: {
+					Authorization: `Bearer ${apiKey}`,
+					"User-Agent": "pi-free-providers",
+				},
+			},
+			3,
+			1000,
+			DEFAULT_FETCH_TIMEOUT_MS,
+		);
+		if (response.ok) {
+			const json = (await response.json()) as {
+				data?: Array<{ id: string }>;
+			};
+			if (json.data) {
+				return new Set(json.data.map((m) => m.id));
+			}
+		}
+	} catch (error) {
+		console.error("[nvidia] Failed to fetch models from NVIDIA API", error);
+	}
+	return new Set();
+}
+
+async function fetchModelsDevMetadata(): Promise<Map<string, ModelsDevModel>> {
+	const devModels = new Map<string, ModelsDevModel>();
+	try {
+		const response = await fetchWithRetry(
+			URL_MODELS_DEV,
+			{
+				headers: { "User-Agent": "pi-free-providers" },
+			},
+			3,
+			1000,
+			DEFAULT_FETCH_TIMEOUT_MS,
+		);
+		if (response.ok) {
+			const json = (await response.json()) as Record<string, ModelsDevProvider>;
+			const provider = Object.values(json).find((p) => p?.id === "nvidia");
+			if (provider?.models) {
+				for (const m of Object.values(provider.models)) {
+					devModels.set(m.id, m);
+				}
+			}
+		}
+	} catch (error) {
+		console.error("[nvidia] Failed to fetch models.dev", error);
+	}
+	return devModels;
+}
+
+function isChatModel(m: ModelsDevModel): boolean {
+	const modalities = m.modalities;
+	if (!modalities) return true;
+	const output = modalities.output ?? [];
+	const input = modalities.input ?? [];
+	return output.includes("text") && input.includes("text");
+}
+
+async function fetchNvidiaModels(
+	apiKey?: string,
+): Promise<ProviderModelConfig[]> {
+	// ── 1. Query NVIDIA's actual API (source of truth) ─────────────────
+	const apiModelIds = apiKey
+		? await fetchNvidiaApiModelIds(apiKey)
+		: new Set<string>();
+
+	// ── 2. Fetch models.dev for rich metadata (cost, limits, etc.) ─────
+	const devModels = await fetchModelsDevMetadata();
+
+	// ── 3. Build unified list (NVIDIA API wins; fallback to models.dev) ─
+	const modelIds =
+		apiModelIds.size > 0 ? [...apiModelIds] : [...devModels.keys()];
+
+	const result = applyHidden(
+		modelIds
+			.map((id) => devModels.get(id) ?? inferModelFromId(id))
+			.filter((m): m is ModelsDevModel => m !== null)
+			.filter((m) => isUsableModel(m.id, NVIDIA_MIN_SIZE_B))
+			.filter(isChatModel)
+			.filter((m) => !NVIDIA_KNOWN_404_MODELS.has(m.id))
+			// NVIDIA is freemium — all models are usable with free credits.
+			// No cost filtering applied.
+			.map(
+				(m): ProviderModelConfig => ({
+					id: m.id,
+					name: m.name,
+					reasoning: m.reasoning,
+					input: m.modalities?.input?.includes("image")
+						? ["text", "image"]
+						: ["text"],
+					cost: {
+						input: m.cost?.input ?? 0,
+						output: m.cost?.output ?? 0,
+						cacheRead: m.cost?.cache_read ?? 0,
+						cacheWrite: m.cost?.cache_write ?? 0,
+					},
+					contextWindow: m.limit.context,
+					maxTokens: m.limit.output,
+				}),
+			),
+		PROVIDER_NVIDIA,
+	);
+
+	return result;
+}
+
+// =============================================================================
+// Extension Entry Point
+// =============================================================================
+
+/**
+ * Probe a single NVIDIA model with a minimal chat request.
+ * Returns true if the model is routable (not 404), false if it 404s.
+ */
+async function probeNvidiaModel(
+	apiKey: string,
+	modelId: string,
+): Promise<boolean> {
+	try {
+		const response = await fetchWithTimeout(
+			`${BASE_URL_NVIDIA}/chat/completions`,
+			{
+				method: "POST",
+				headers: {
+					Authorization: `Bearer ${apiKey}`,
+					"Content-Type": "application/json",
+					"User-Agent": "pi-free-providers",
+				},
+				body: JSON.stringify({
+					model: modelId,
+					messages: [{ role: "user", content: "hi" }],
+					max_tokens: 1,
+				}),
+			},
+			10000, // 10 second timeout
+		);
+		// 404 = function not found (model not provisioned)
+		// 200/400/401/etc = at least routable
+		return response.status !== 404;
+	} catch {
+		return true; // Network errors / timeouts are not "model not found"
+	}
+}
+
+const _nvidiaLogger = createLogger("nvidia");
+
+/**
+ * Run probe on a list of models and auto-hide 404s.
+ * Shared between the /probe-nvidia command and auto-probe on session_start.
+ */
+async function runNvidiaProbe(
+	apiKey: string,
+	modelsToTest: ProviderModelConfig[],
+	stored: { free: ProviderModelConfig[]; all: ProviderModelConfig[] },
+	reRegister: (models: ProviderModelConfig[]) => void,
+): Promise<void> {
+	const notFound: string[] = [];
+	const batchSize = 5;
+
+	for (let i = 0; i < modelsToTest.length; i += batchSize) {
+		const batch = modelsToTest.slice(i, i + batchSize);
+		const results = await Promise.all(
+			batch.map(async (m) => {
+				const ok = await probeNvidiaModel(apiKey, m.id);
+				return { id: m.id, ok };
+			}),
+		);
+		for (const r of results) {
+			if (!r.ok) notFound.push(r.id);
+		}
+	}
+
+	if (notFound.length === 0) {
+		_nvidiaLogger.info("Auto-probe: all NVIDIA models are routable");
+		return;
+	}
+
+	// Auto-hide 404 models in config (provider-scoped)
+	const cfg = loadConfigFile();
+	const existingHidden = new Set(cfg.hidden_models ?? []);
+	for (const id of notFound) existingHidden.add(`${PROVIDER_NVIDIA}/${id}`);
+	saveConfig({ hidden_models: Array.from(existingHidden) });
+
+	// Re-register so hidden models disappear immediately
+	const filtered = await fetchNvidiaModels(apiKey);
+	stored.free = filtered;
+	stored.all = filtered;
+	reRegister(filtered);
+
+	_nvidiaLogger.info(
+		`Auto-probe: found ${notFound.length} broken models (auto-hidden)`,
+	);
+}
+
+export default async function (pi: ExtensionAPI) {
+	const apiKey = getNvidiaApiKey();
+	const hasKey = !!apiKey;
+
+	let allModels: ProviderModelConfig[] = [];
+
+	try {
+		allModels = await fetchNvidiaModels(apiKey);
+	} catch (error) {
+		console.error("[nvidia] Failed to fetch models at startup", error);
+		return;
+	}
+
+	// Store both sets for global toggle using consistent isFreeModel helper
+	// NVIDIA uses Route B (name-based): only models with "free" in name are marked free
+	const freeModels = allModels.filter((m) =>
+		isFreeModel({ ...m, provider: PROVIDER_NVIDIA }),
+	);
+	const stored = { free: freeModels, all: allModels };
+
+	// Create re-register function
+	const reRegister = createReRegister(pi, {
+		providerId: PROVIDER_NVIDIA,
+		baseUrl: BASE_URL_NVIDIA,
+		apiKey: apiKey || "NVIDIA_API_KEY",
+	});
+
+	// Register with global toggle system
+	registerWithGlobalToggle(PROVIDER_NVIDIA, stored, reRegister, hasKey);
+
+	// Register initial models (global toggle will apply filter if needed)
+	const initialModels = allModels;
+	pi.registerProvider(PROVIDER_NVIDIA, {
+		baseUrl: BASE_URL_NVIDIA,
+		apiKey: apiKey || "NVIDIA_API_KEY",
+		api: "openai-completions" as const,
+		authHeader: true,
+		headers: {
+			"User-Agent": "pi-free-providers",
+		},
+		models: enhanceWithCI(initialModels),
+	});
+
+	// ── Lazy auto-probe on first session_start ──────────────────────
+	let _autoProbeDone = false;
+	pi.on("session_start", async () => {
+		if (_autoProbeDone || !apiKey) return;
+		_autoProbeDone = true;
+		_nvidiaLogger.info("Starting lazy auto-probe of NVIDIA models...");
+		runNvidiaProbe(apiKey, allModels, stored, reRegister).catch((err) => {
+			_nvidiaLogger.warn("Auto-probe failed", {
+				error: err instanceof Error ? err.message : String(err),
+			});
+		});
+	});
+
+	// ── Probe command: test all registered models for 404s ─────────────
+	pi.registerCommand("probe-nvidia", {
+		description: "Test all NVIDIA models for 404 'Function not found' errors",
+		handler: async (_args, ctx) => {
+			if (!apiKey) {
+				ctx.ui.notify("NVIDIA_API_KEY not set", "error");
+				return;
+			}
+
+			const modelsToTest = allModels;
+			ctx.ui.notify(`Probing ${modelsToTest.length} NVIDIA models…`, "info");
+
+			await runNvidiaProbe(apiKey, modelsToTest, stored, reRegister);
+
+			// Check if any were hidden (re-read config)
+			const cfgAfter = loadConfigFile();
+			const newHidden = (cfgAfter.hidden_models ?? []).filter((h) =>
+				h.startsWith(`${PROVIDER_NVIDIA}/`),
+			);
+			if (newHidden.length > 0) {
+				ctx.ui.notify(
+					`Found ${newHidden.length} broken models (auto-hidden):\n${newHidden.join("\n")}`,
+					"warning",
+				);
+			} else {
+				ctx.ui.notify("All NVIDIA models are routable ✅", "info");
+			}
+		},
+	});
+
+	// ── Status bar for provider selection ─────────────────────────
+
+	pi.on("model_select", (_event, ctx) => {
+		if (_event.model?.provider !== PROVIDER_NVIDIA) {
+			ctx.ui.setStatus(`${PROVIDER_NVIDIA}-status`, undefined);
+			return;
+		}
+
+		const count = allModels.length;
+		ctx.ui.setStatus(
+			`${PROVIDER_NVIDIA}-status`,
+			`nvidia: ${count} models (freemium)`,
+		);
+	});
+
+	// Registration complete - models registered silently (use LOG_LEVEL=info to see details)
+}