pi-nvidia-nim 1.1.18 → 1.1.19

package/README.md

@@ -1,6 +1,6 @@
 # pi-nvidia-nim
 
-NVIDIA NIM API provider extension for [pi coding agent](https://github.com/badlogic/pi-mono) - access 100+ models from [build.nvidia.com](https://build.nvidia.com) including DeepSeek V3.2, Kimi K2.5, MiniMax M2.1, GLM-5, GLM-4.7, Qwen3, Llama 4, and many more.
+NVIDIA NIM API provider extension for [pi coding agent](https://github.com/badlogic/pi-mono) - access 100+ models from [build.nvidia.com](https://build.nvidia.com) including DeepSeek V4 Flash/Pro, DeepSeek V3.2, Kimi K2.6, MiniMax M2.1, GLM-5, GLM-4.7, Qwen3, Llama 4, and many more.
 
 https://github.com/user-attachments/assets/f44773e4-9bf8-4bb5-a9c0-d5938030701c
 
@@ -17,10 +17,14 @@ https://github.com/user-attachments/assets/f44773e4-9bf8-4bb5-a9c0-d5938030701c
 ### 2. Set Your API Key
 
 ```bash
+# Preferred by this extension
 export NVIDIA_NIM_API_KEY=nvapi-your-key-here
+
+# Also supported, matching NVIDIA's website examples
+export NVIDIA_API_KEY=nvapi-your-key-here
 ```
 
-Add this to your `~/.bashrc`, `~/.zshrc`, or shell profile to persist it.
+Add one of these to your `~/.bashrc`, `~/.zshrc`, or shell profile to persist it.
 
 ### 3. Install the Extension
 
@@ -53,10 +57,10 @@ Once loaded, NVIDIA NIM models appear in the `/model` selector under the `nvidia
 
 ```bash
 # Use a specific NIM model directly
-pi --provider nvidia-nim --model "deepseek-ai/deepseek-v3.2"
+pi --provider nvidia-nim --model "deepseek-ai/deepseek-v4-flash"
 
 # With thinking enabled
-pi --provider nvidia-nim --model "deepseek-ai/deepseek-v3.2" --thinking low
+pi --provider nvidia-nim --model "deepseek-ai/deepseek-v4-flash" --thinking high
 
 # Limit model cycling to NIM models
 pi --models "nvidia-nim/*"
@@ -70,11 +74,12 @@ NVIDIA NIM models use a non-standard `chat_template_kwargs` parameter to enable
 
 When you change the thinking level in pi (`Shift+Tab` to cycle), the extension:
 
-1. **Maps `"minimal"` → `"low"`** - NIM only accepts `low`, `medium`, `high` (not `minimal`). Selecting "minimal" in pi works fine; it's silently mapped.
+1. **Maps thinking levels** to values each NIM model accepts. For DeepSeek V4, `xhigh` maps to `max`; lower enabled levels use `high`.
 2. **Injects `chat_template_kwargs`** per model to actually enable thinking:
+   - DeepSeek V4: `{ thinking: true, reasoning_effort: "high" | "max" }`
    - DeepSeek V3.x, R1 distills: `{ thinking: true }`
    - GLM-5, GLM-4.7: `{ enable_thinking: true, clear_thinking: false }`
-   - Kimi K2.5, K2-thinking: `{ thinking: true }`
+   - Kimi K2.6, K2-thinking: `{ thinking: true }`
    - Qwen3, QwQ: `{ enable_thinking: true }`
 3. **Explicitly disables thinking** when the level is "off" for models that think by default (e.g., GLM-5, GLM-4.7).
 4. **Uses `system` role** instead of `developer` for all NIM models - the `developer` role combined with `chat_template_kwargs` causes 500 errors on NIM.
@@ -84,22 +89,25 @@ When you change the thinking level in pi (`Shift+Tab` to cycle), the extension:
 | pi Level | NIM Mapping | Effect |
 |----------|-------------|--------|
 | off | No kwargs (or explicit disable) | No reasoning output |
-| minimal | Mapped to "low" | Thinking enabled |
-| low | low | Thinking enabled |
-| medium | medium | Thinking enabled |
+| minimal | low, or high for DeepSeek V4 | Thinking enabled |
+| low | low, or high for DeepSeek V4 | Thinking enabled |
+| medium | medium, or high for DeepSeek V4 | Thinking enabled |
 | high | high | Thinking enabled |
+| xhigh | high, or max for DeepSeek V4 | Maximum supported thinking |
 
 ## Available Models
 
-The extension ships with curated metadata for 39 featured models. At startup, it also queries the NVIDIA NIM API to discover additional models automatically.
+The extension ships with curated metadata for 42 featured models. At startup, it also queries the NVIDIA NIM API to discover additional models automatically.
 
 ### Featured Models
 
 | Model | Reasoning | Vision | Context |
 |-------|-----------|--------|---------|
+| `deepseek-ai/deepseek-v4-flash` | ✅ | | 1M |
+| `deepseek-ai/deepseek-v4-pro` | ✅ | | 1M |
 | `deepseek-ai/deepseek-v3.2` | ✅ | | 128K |
 | `deepseek-ai/deepseek-v3.1` | ✅ | | 128K |
-| `moonshotai/kimi-k2.5` | ✅ | | 256K |
+| `moonshotai/kimi-k2.6` | ✅ | | 256K |
 | `moonshotai/kimi-k2-thinking` | ✅ | | 128K |
 | `minimaxai/minimax-m2.1` | | | 1M |
 | `z-ai/glm5` | ✅ | | 128K |
@@ -121,7 +129,7 @@ The extension ships with curated metadata for 39 featured models. At startup, it
 
 ### Tool Calling
 
-All major models support OpenAI-compatible tool calling. Tested and confirmed working with DeepSeek V3.2, GLM-5, GLM-4.7, Qwen3, Kimi K2.5, and others.
+All major models support OpenAI-compatible tool calling. Tested and confirmed working with DeepSeek V4/V3.2, GLM-5, GLM-4.7, Qwen3, Kimi K2.6, and others.
 
 ## How It Works
 
@@ -130,13 +138,13 @@ This extension uses `pi.registerProvider()` to register NVIDIA NIM as a custom p
 The custom streamer:
 1. Intercepts the request payload via `onPayload` callback
 2. Injects `chat_template_kwargs` for models that need it to enable thinking
-3. Maps unsupported thinking levels (`minimal` → `low`)
+3. Maps unsupported thinking levels to NIM-compatible values (`minimal` → `low`; `xhigh` → `high` or DeepSeek V4 `max`)
 4. Suppresses `reasoning_effort` for models that don't respond to it (e.g., DeepSeek without kwargs)
 5. Uses the standard OpenAI SSE streaming format - pi already parses `reasoning_content` and `reasoning` fields from streaming deltas
 
 ## Configuration
 
-The only configuration needed is the `NVIDIA_NIM_API_KEY` environment variable. All models on NVIDIA NIM are free during the preview period (with rate limits).
+The only configuration needed is either the `NVIDIA_NIM_API_KEY` or `NVIDIA_API_KEY` environment variable. All models on NVIDIA NIM are free during the preview period (with rate limits).
 
 ## Notes
 

package/index.ts

@@ -6,7 +6,7 @@
  *
  * Setup:
  *   1. Get an API key from https://build.nvidia.com
- *   2. Export it: export NVIDIA_NIM_API_KEY=nvapi-...
+ *   2. Export it: export NVIDIA_NIM_API_KEY=nvapi-... (or NVIDIA_API_KEY=nvapi-...)
  *   3. Load the extension:
  *      pi -e ./path/to/pi-nvidia-nim
  *      # or install as a package:
@@ -22,17 +22,20 @@
  * parameters:
  *
  * - DeepSeek V3.x: `chat_template_kwargs: { thinking: true }`
+ * - DeepSeek V4:   `chat_template_kwargs: { thinking: true, reasoning_effort: "high" | "max" }`
  * - GLM-5/4.7:     `chat_template_kwargs: { enable_thinking: true, clear_thinking: false }`
  * - Kimi K2.5:     `chat_template_kwargs: { thinking: true }` (also accepts reasoning_effort)
  * - Qwen3:         `chat_template_kwargs: { enable_thinking: true }`
  *
- * NIM only accepts `reasoning_effort` values of "low", "medium", "high" - NOT
- * "minimal". The extension maps pi's "minimal" level to "low" automatically.
+ * NIM only accepts selected `reasoning_effort` values. The extension maps pi's
+ * provider-agnostic levels to the values each NIM model accepts.
  *
  * Some models (e.g., GLM-5, GLM-4.7) always produce reasoning output regardless of
  * thinking settings.
  */
 
+import { existsSync, readFileSync } from "node:fs";
+import { join } from "node:path";
 import type {
 	Api,
 	AssistantMessageEventStream,
@@ -41,7 +44,7 @@ import type {
 	SimpleStreamOptions,
 } from "@mariozechner/pi-ai";
 import { streamSimpleOpenAICompletions } from "@mariozechner/pi-ai";
-import type { ExtensionAPI } from "@mariozechner/pi-coding-agent";
+import { getAgentDir, type ExtensionAPI } from "@mariozechner/pi-coding-agent";
 
 // =============================================================================
 // Constants
@@ -49,6 +52,8 @@ import type { ExtensionAPI } from "@mariozechner/pi-coding-agent";
 
 const NVIDIA_NIM_BASE_URL = "https://integrate.api.nvidia.com/v1";
 const NVIDIA_NIM_API_KEY_ENV = "NVIDIA_NIM_API_KEY";
+const NVIDIA_API_KEY_ENV = "NVIDIA_API_KEY";
+const NVIDIA_API_KEY_ENV_NAMES = [NVIDIA_NIM_API_KEY_ENV, NVIDIA_API_KEY_ENV] as const;
 const PROVIDER_NAME = "nvidia-nim";
 
 // =============================================================================
@@ -70,10 +75,22 @@ interface ThinkingConfig {
 	disableKwargs?: Record<string, unknown>;
 	/** If true, also send reasoning_effort alongside chat_template_kwargs */
 	sendReasoningEffort?: boolean;
+	/** If true, include a model-specific reasoning_effort inside chat_template_kwargs */
+	includeReasoningEffortInKwargs?: boolean;
 }
 
 const THINKING_CONFIGS: Record<string, ThinkingConfig> = {
 	// DeepSeek models need chat_template_kwargs - reasoning_effort alone doesn't trigger thinking
+	"deepseek-ai/deepseek-v4-flash": {
+		enableKwargs: { thinking: true },
+		disableKwargs: { thinking: false },
+		includeReasoningEffortInKwargs: true,
+	},
+	"deepseek-ai/deepseek-v4-pro": {
+		enableKwargs: { thinking: true },
+		disableKwargs: { thinking: false },
+		includeReasoningEffortInKwargs: true,
+	},
 	"deepseek-ai/deepseek-v3.2": {
 		enableKwargs: { thinking: true },
 		disableKwargs: { thinking: false },
@@ -112,7 +129,7 @@ const THINKING_CONFIGS: Record<string, ThinkingConfig> = {
 		disableKwargs: { enable_thinking: false },
 	},
 	// Kimi models: chat_template_kwargs works, reasoning_effort also works
-	"moonshotai/kimi-k2.5": {
+	"moonshotai/kimi-k2.6": {
 		enableKwargs: { thinking: true },
 		disableKwargs: { thinking: false },
 		sendReasoningEffort: true,
@@ -236,6 +253,8 @@ const CONTEXT_WINDOWS: Record<string, number> = {
 	"deepseek-ai/deepseek-v3.1": 131072,
 	"deepseek-ai/deepseek-v3.1-terminus": 131072,
 	"deepseek-ai/deepseek-v3.2": 131072,
+	"deepseek-ai/deepseek-v4-flash": 1048576,
+	"deepseek-ai/deepseek-v4-pro": 1048576,
 	"deepseek-ai/deepseek-r1-distill-llama-8b": 131072,
 	"deepseek-ai/deepseek-r1-distill-qwen-14b": 131072,
 	"deepseek-ai/deepseek-r1-distill-qwen-32b": 131072,
@@ -245,7 +264,7 @@ const CONTEXT_WINDOWS: Record<string, number> = {
 	"moonshotai/kimi-k2-instruct": 131072,
 	"moonshotai/kimi-k2-instruct-0905": 131072,
 	"moonshotai/kimi-k2-thinking": 131072,
-	"moonshotai/kimi-k2.5": 262144,
+	"moonshotai/kimi-k2.6": 262144,
 	// MiniMax
 	"minimaxai/minimax-m2": 1048576,
 	"minimaxai/minimax-m2.1": 1048576,
@@ -369,7 +388,9 @@ const MAX_TOKENS: Record<string, number> = {
 	"deepseek-ai/deepseek-v3.1": 16384,
 	"deepseek-ai/deepseek-v3.1-terminus": 16384,
 	"deepseek-ai/deepseek-v3.2": 16384,
-	"moonshotai/kimi-k2.5": 16384,
+	"deepseek-ai/deepseek-v4-flash": 16384,
+	"deepseek-ai/deepseek-v4-pro": 16384,
+	"moonshotai/kimi-k2.6": 16384,
 	"moonshotai/kimi-k2-instruct": 8192,
 	"moonshotai/kimi-k2-thinking": 16384,
 	"minimaxai/minimax-m2": 8192,
@@ -393,10 +414,12 @@ const MAX_TOKENS: Record<string, number> = {
 
 const FEATURED_MODELS = [
 	// Flagship / frontier
+	"deepseek-ai/deepseek-v4-flash",
+	"deepseek-ai/deepseek-v4-pro",
 	"deepseek-ai/deepseek-v3.2",
 	"deepseek-ai/deepseek-v3.1",
 	"deepseek-ai/deepseek-v3.1-terminus",
-	"moonshotai/kimi-k2.5",
+	"moonshotai/kimi-k2.6",
 	"moonshotai/kimi-k2-thinking",
 	"moonshotai/kimi-k2-instruct",
 	"moonshotai/kimi-k2-instruct-0905",
@@ -450,11 +473,148 @@ const FEATURED_MODELS = [
  * Custom streamSimple that wraps the standard OpenAI completions streamer.
  *
  * Fixes for NVIDIA NIM:
- * 1. Maps pi's "minimal" thinking level → "low" (NIM only accepts low/medium/high)
+ * 1. Maps pi's thinking levels to values accepted by NVIDIA NIM
  * 2. Strips reasoning_effort for models where it doesn't trigger thinking
  * 3. Injects chat_template_kwargs per model to actually enable thinking
  * 4. Uses onPayload callback to mutate request params before they're sent
  */
+type NimApiKeyEnvName = (typeof NVIDIA_API_KEY_ENV_NAMES)[number];
+type AuthStorageLike = {
+	get?: (provider: string) => unknown;
+};
+
+interface NimApiKeyCredential {
+	type: "api_key";
+	key: string;
+}
+
+function getNimApiKeyEnv(): NimApiKeyEnvName | undefined {
+	return NVIDIA_API_KEY_ENV_NAMES.find((envName) => !!process.env[envName]);
+}
+
+function getNimApiKey(): string | undefined {
+	const envName = getNimApiKeyEnv();
+	const apiKey = envName ? process.env[envName] : undefined;
+	return apiKey?.trim() || undefined;
+}
+
+function isNimApiKeyEnvName(value: string): value is NimApiKeyEnvName {
+	return NVIDIA_API_KEY_ENV_NAMES.includes(value as NimApiKeyEnvName);
+}
+
+function isNimApiKeyEnvValue(value: string): boolean {
+	return NVIDIA_API_KEY_ENV_NAMES.some((envName) => process.env[envName]?.trim() === value);
+}
+
+function isNimApiKeyCredential(credential: unknown): credential is NimApiKeyCredential {
+	return (
+		typeof credential === "object" &&
+		credential !== null &&
+		(credential as { type?: unknown }).type === "api_key" &&
+		typeof (credential as { key?: unknown }).key === "string"
+	);
+}
+
+function readStoredNimApiKeyConfig(): string | undefined {
+	try {
+		const authPath = join(getAgentDir(), "auth.json");
+		if (!existsSync(authPath)) return undefined;
+
+		const data = JSON.parse(readFileSync(authPath, "utf-8")) as Record<string, unknown>;
+		const credential = data[PROVIDER_NAME];
+		return isNimApiKeyCredential(credential) ? credential.key : undefined;
+	} catch {
+		return undefined;
+	}
+}
+
+function getStoredNimApiKeyConfig(authStorage?: AuthStorageLike): string | undefined {
+	if (authStorage) {
+		try {
+			const credential = authStorage.get?.(PROVIDER_NAME);
+			return isNimApiKeyCredential(credential) ? credential.key : undefined;
+		} catch {
+			return undefined;
+		}
+	}
+
+	return readStoredNimApiKeyConfig();
+}
+
+function hasStoredNimCommandCredential(authStorage?: AuthStorageLike): boolean {
+	return getStoredNimApiKeyConfig(authStorage)?.startsWith("!") ?? false;
+}
+
+function getStoredResolvedNimApiKey(authStorage?: AuthStorageLike): string | undefined {
+	const configuredApiKey = getStoredNimApiKeyConfig(authStorage)?.trim();
+	if (!configuredApiKey || configuredApiKey.startsWith("!")) return undefined;
+
+	const envValue = process.env[configuredApiKey]?.trim();
+	return envValue || configuredApiKey;
+}
+
+function normalizeResolvedNimApiKey(apiKey: string | undefined): string | undefined {
+	if (apiKey === undefined) return undefined;
+
+	const trimmed = apiKey.trim();
+	if (!trimmed) {
+		throw new Error("NVIDIA NIM API key resolved to an empty value.");
+	}
+
+	return trimmed;
+}
+
+function resolveNimApiKey(apiKey: string | undefined, authStorage?: AuthStorageLike): string | undefined {
+	const resolvedApiKey = normalizeResolvedNimApiKey(apiKey);
+	const hasStoredCommandCredential = hasStoredNimCommandCredential(authStorage);
+
+	if (
+		hasStoredCommandCredential &&
+		(!resolvedApiKey || isNimApiKeyEnvName(resolvedApiKey) || isNimApiKeyEnvValue(resolvedApiKey))
+	) {
+		throw new Error("NVIDIA NIM API key command resolved to an empty value.");
+	}
+
+	const storedApiKey = getStoredResolvedNimApiKey(authStorage);
+	if (storedApiKey && (!resolvedApiKey || isNimApiKeyEnvName(resolvedApiKey))) return storedApiKey;
+
+	if (resolvedApiKey && !isNimApiKeyEnvName(resolvedApiKey)) return resolvedApiKey;
+
+	return getNimApiKey();
+}
+
+function resolveRequiredNimApiKey(apiKey: string | undefined): string {
+	const resolvedApiKey = resolveNimApiKey(apiKey);
+	if (resolvedApiKey) return resolvedApiKey;
+
+	throw new Error(
+		`NVIDIA NIM: no API key configured. Set ${NVIDIA_NIM_API_KEY_ENV} or ${NVIDIA_API_KEY_ENV}. ` +
+		`Get a free API key at https://build.nvidia.com and export it: ` +
+		`export ${NVIDIA_NIM_API_KEY_ENV}=nvapi-...`,
+	);
+}
+
+function mapNimTopLevelReasoning(reasoning: SimpleStreamOptions["reasoning"]): SimpleStreamOptions["reasoning"] {
+	if (reasoning === "minimal") return "low";
+	if (reasoning === "xhigh") return "high";
+	return reasoning;
+}
+
+function mapDeepSeekV4Reasoning(reasoning: SimpleStreamOptions["reasoning"]): "high" | "max" {
+	return reasoning === "xhigh" ? "max" : "high";
+}
+
+function buildThinkingKwargs(
+	thinkingConfig: ThinkingConfig,
+	reasoning: SimpleStreamOptions["reasoning"],
+): Record<string, unknown> {
+	const kwargs = { ...thinkingConfig.enableKwargs };
+	if (thinkingConfig.includeReasoningEffortInKwargs) {
+		kwargs.reasoning_effort = mapDeepSeekV4Reasoning(reasoning);
+	}
+	return kwargs;
+}
+
 function nimStreamSimple(
 	model: Model<Api>,
 	context: Context,
@@ -464,12 +624,9 @@ function nimStreamSimple(
 	const reasoning = options?.reasoning;
 	const isThinkingEnabled = !!reasoning;
 
-	// Map "minimal" → "low" since NIM rejects "minimal" with a 400 error.
-	// NIM only accepts: "low", "medium", "high"
-	let mappedReasoning = reasoning;
-	if (reasoning === "minimal") {
-		mappedReasoning = "low";
-	}
+	// Map provider-agnostic pi levels to NIM's accepted top-level values.
+	// Model-specific chat_template_kwargs may apply a different mapping below.
+	const mappedReasoning = mapNimTopLevelReasoning(reasoning);
 
 	// For models that have a thinking config: we handle thinking via chat_template_kwargs.
 	// Suppress reasoning_effort (set reasoning to undefined) unless the model explicitly
@@ -481,16 +638,9 @@ function nimStreamSimple(
 		effectiveReasoning = undefined;
 	}
 
-	// Resolve API key at request time - must pass via options.apiKey because
-	// pi-ai's getEnvApiKey() doesn't know about custom providers like "nvidia-nim"
-	const nimApiKey = process.env["NVIDIA_NIM_API_KEY"];
-	if (!nimApiKey) {
-		throw new Error(
-			`NVIDIA NIM: NVIDIA_NIM_API_KEY environment variable is not set. ` +
-			`Get a free API key at https://build.nvidia.com and export it: ` +
-			`export NVIDIA_NIM_API_KEY=nvapi-...`
-		);
-	}
+	// Use pi's already-resolved provider key when available (auth.json, shell command,
+	// CLI override), and fall back to the two NVIDIA environment variable names.
+	const nimApiKey = resolveRequiredNimApiKey(options?.apiKey);
 
 	const modifiedOptions: SimpleStreamOptions = {
 		...options,
@@ -502,7 +652,7 @@ function nimStreamSimple(
 			if (thinkingConfig) {
 				if (isThinkingEnabled) {
 					// Inject chat_template_kwargs to enable thinking
-					p.chat_template_kwargs = thinkingConfig.enableKwargs;
+					p.chat_template_kwargs = buildThinkingKwargs(thinkingConfig, reasoning);
 				} else if (thinkingConfig.disableKwargs) {
 					// Explicitly disable thinking (some models think by default, e.g. GLM-5/4.7)
 					p.chat_template_kwargs = thinkingConfig.disableKwargs;
@@ -532,7 +682,7 @@ function nimStreamSimple(
 			}
 
 			// Chain to original onPayload if present
-			options?.onPayload?.(params);
+			return options?.onPayload?.(params, model);
 		},
 	};
 
@@ -612,7 +762,34 @@ interface NimApiModel {
 	owned_by: string;
 }
 
-async function fetchNimModels(apiKey: string): Promise<string[]> {
+type NimModelFetchResult =
+	| { ok: true; modelIds: string[] }
+	| { ok: false; reason: "auth" | "transient" | "invalid" | "network" | "other" };
+
+const NIM_DISCOVERY_CREDENTIAL_WARNING =
+	"NVIDIA NIM model discovery skipped: check your nvidia-nim credentials.";
+
+function sanitizeNimLogMessage(message: string): string {
+	return message.replace(/nvapi-[A-Za-z0-9._-]+/g, "nvapi-[REDACTED]");
+}
+
+function notifyNimDiscoveryCredentialWarning(ctx: any): void {
+	ctx?.ui?.notify?.(NIM_DISCOVERY_CREDENTIAL_WARNING, "warning");
+}
+
+async function resolveNimDiscoveryApiKey(ctx: any): Promise<string | undefined> {
+	try {
+		const apiKey = await ctx?.modelRegistry?.getApiKeyForProvider?.(PROVIDER_NAME);
+		return resolveNimApiKey(apiKey, ctx?.modelRegistry?.authStorage);
+	} catch (error) {
+		const message = error instanceof Error ? error.message : String(error);
+		console.warn(`pi-nvidia-nim: ${sanitizeNimLogMessage(message)}`);
+		notifyNimDiscoveryCredentialWarning(ctx);
+		return undefined;
+	}
+}
+
+async function fetchNimModels(apiKey: string): Promise<NimModelFetchResult> {
 	try {
 		const response = await fetch(`${NVIDIA_NIM_BASE_URL}/models`, {
 			headers: {
@@ -622,12 +799,25 @@ async function fetchNimModels(apiKey: string): Promise<string[]> {
 			signal: AbortSignal.timeout(10000),
 		});
 
-		if (!response.ok) return [];
+		if (response.status === 401 || response.status === 403) {
+			return { ok: false, reason: "auth" };
+		}
+
+		if (response.status === 429 || response.status >= 500) {
+			return { ok: false, reason: "transient" };
+		}
+
+		if (!response.ok) return { ok: false, reason: "other" };
+
+		const data = (await response.json()) as { data?: NimApiModel[] };
+		if (!Array.isArray(data.data)) return { ok: false, reason: "invalid" };
 
-		const data = (await response.json()) as { data: NimApiModel[] };
-		return data.data?.map((m) => m.id) ?? [];
+		return {
+			ok: true,
+			modelIds: data.data.map((m) => m.id).filter((id): id is string => typeof id === "string" && id.length > 0),
+		};
 	} catch {
-		return [];
+		return { ok: false, reason: "network" };
 	}
 }
 
@@ -636,27 +826,12 @@ async function fetchNimModels(apiKey: string): Promise<string[]> {
 // =============================================================================
 
 export default function (pi: ExtensionAPI) {
-	// Bail out cleanly if the user hasn't set up an API key yet. Without this
-	// guard, pi's flow can still route requests through this provider's
-	// streamSimple (e.g. when probing available providers, or when the
-	// configured default provider can't be resolved for any reason), which
-	// causes a fatal "NVIDIA_NIM_API_KEY environment variable is not set"
-	// error to be emitted on every `pi` invocation — even when the user
-	// never intended to use NVIDIA NIM. Skipping registration leaves the
-	// extension dormant: no NIM models appear in the picker, no provider is
-	// available to route to, and other providers (OpenAI, Anthropic, custom
-	// OpenAI-compatible endpoints in models.json, etc.) work normally.
-	//
-	// Re-enable by exporting the key:
-	//   export NVIDIA_NIM_API_KEY=nvapi-...
-	if (!process.env[NVIDIA_NIM_API_KEY_ENV]) {
-		// One-line, stderr, only on first load. Not fatal.
-		console.error(
-			`pi-nvidia-nim: ${NVIDIA_NIM_API_KEY_ENV} not set — extension dormant. ` +
-			`Set the key (https://build.nvidia.com) and reload pi to enable NVIDIA NIM models.`
-		);
-		return;
-	}
+	const providerApiKeyConfig = getNimApiKeyEnv() ?? NVIDIA_NIM_API_KEY_ENV;
+
+	// Always register the curated model list. The request path resolves credentials
+	// through pi first (CLI override, auth.json, shell command), then falls back to
+	// NVIDIA_NIM_API_KEY/NVIDIA_API_KEY. This keeps models available even when pi
+	// was launched by a shell that did not source ~/.bashrc or ~/.zshrc.
 
 	// Build the curated model list
 	const modelMap = new Map<string, NimModelEntry>();
@@ -672,7 +847,7 @@ export default function (pi: ExtensionAPI) {
 
 	pi.registerProvider(PROVIDER_NAME, {
 		baseUrl: NVIDIA_NIM_BASE_URL,
-		apiKey: NVIDIA_NIM_API_KEY_ENV,
+		apiKey: providerApiKeyConfig,
 		api: "openai-completions",
 		authHeader: true,
 		models: curatedModels,
@@ -681,11 +856,19 @@ export default function (pi: ExtensionAPI) {
 
 	// On session start, discover additional models from the API
 	pi.on("session_start", async (_event: any, ctx: any) => {
-		const apiKey = process.env[NVIDIA_NIM_API_KEY_ENV];
-		if (!apiKey) return; // belt-and-suspenders: skip if env was unset after load
+		const apiKey = await resolveNimDiscoveryApiKey(ctx);
+		if (!apiKey) return;
 
 		// Fetch live model list
-		const liveModelIds = await fetchNimModels(apiKey);
+		const fetchResult = await fetchNimModels(apiKey);
+		if (!fetchResult.ok) {
+			if (fetchResult.reason === "auth") {
+				notifyNimDiscoveryCredentialWarning(ctx);
+			}
+			return;
+		}
+
+		const liveModelIds = fetchResult.modelIds;
 		if (liveModelIds.length === 0) return;
 
 		let newModelsAdded = 0;
@@ -706,7 +889,7 @@ export default function (pi: ExtensionAPI) {
 			const allModels = Array.from(modelMap.values());
 			ctx.modelRegistry.registerProvider(PROVIDER_NAME, {
 				baseUrl: NVIDIA_NIM_BASE_URL,
-				apiKey: NVIDIA_NIM_API_KEY_ENV,
+				apiKey: getNimApiKeyEnv() ?? NVIDIA_NIM_API_KEY_ENV,
 				api: "openai-completions",
 				authHeader: true,
 				models: allModels,

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "pi-nvidia-nim",
-  "version": "1.1.18",
+  "version": "1.1.19",
   "description": "NVIDIA NIM API provider extension for pi coding agent — access 100+ models from build.nvidia.com",
   "type": "module",
   "files": [
@@ -19,7 +19,8 @@
   "scripts": {
     "clean": "echo 'nothing to clean'",
     "build": "echo 'nothing to build'",
-    "check": "echo 'nothing to check'"
+    "check": "tsc --noEmit",
+    "test": "node --test test/*.test.mjs"
   },
   "devDependencies": {
     "@mariozechner/pi-coding-agent": "^0.52.9",