feat(coding-agent): added provider extension support to CLI tools

- Enable shared extension provider loading in bench and dry-balance CLI commands to ensure custom providers are registered.
- Surface benchmark failures for empty streams that return no content and zero usage tokens instead of treating them as successful.

packages/coding-agent/src/cli/bench-cli.ts

@@ -25,7 +25,7 @@ import {
 } from "../config/model-resolver";
 import { Settings } from "../config/settings";
 import benchPrompt from "../prompts/bench.md" with { type: "text" };
-import { discoverAuthStorage } from "../sdk";
+import { discoverAuthStorage, loadCliExtensionProviders } from "../sdk";
 import { resolveThinkingLevelForModel, shouldDisableReasoning, toReasoningEffort } from "../thinking";
 
 const DEFAULT_RUNS = 1;
@@ -145,6 +145,23 @@ function isFirstTokenEvent(event: AssistantMessageEvent): boolean {
 	}
 }
 
+/** Final message carries visible output — non-empty text/thinking or a tool call. */
+function hasVisibleFinalContent(message: AssistantMessage): boolean {
+	return message.content.some(block => {
+		switch (block.type) {
+			case "text":
+				return block.text.length > 0;
+			case "thinking":
+				return block.thinking.length > 0;
+			case "redactedThinking":
+			case "toolCall":
+				return true;
+			default:
+				return false;
+		}
+	});
+}
+
 /**
  * Tokens/s over the generation window (duration minus TTFT) so queue/prefill
  * latency does not dilute throughput. Falls back to total duration when the
@@ -232,6 +249,18 @@ async function runBenchRequest(
 		const rawTtft = message.ttft ?? (firstTokenAt === undefined ? durationMs : firstTokenAt - startedAt);
 		const ttftMs = Number.isFinite(rawTtft) && rawTtft > 0 ? rawTtft : 0;
 		const outputTokens = Number.isFinite(message.usage.output) && message.usage.output > 0 ? message.usage.output : 0;
+		// A run that streamed no content (no delta/end event set firstTokenAt),
+		// carries no visible final content, and measured no output tokens
+		// benchmarked nothing — a genuinely empty stream (e.g. a gateway that 200s
+		// with an empty body). Surface it as a failure instead of a misleading
+		// 0-token "✓". Streaming and buffered providers that produce content keep
+		// passing even when usage is omitted.
+		if (firstTokenAt === undefined && outputTokens === 0 && !hasVisibleFinalContent(message)) {
+			return {
+				ok: false,
+				error: `provider returned no output (0 tokens, empty stream; stop reason: ${message.stopReason ?? "unknown"})`,
+			};
+		}
 		return {
 			ok: true,
 			ttftMs,
@@ -328,8 +357,10 @@ export function formatBenchTable(summary: BenchSummary): string {
 async function createDefaultRuntime(): Promise<BenchRuntime> {
 	const authStorage = await discoverAuthStorage();
 	try {
-		const settings = await Settings.init({ cwd: getProjectDir() });
+		const cwd = getProjectDir();
+		const settings = await Settings.init({ cwd });
 		const modelRegistry = new ModelRegistry(authStorage);
+		await loadCliExtensionProviders(modelRegistry, settings, cwd);
 		return {
 			modelRegistry,
 			settings,

packages/coding-agent/src/cli/dry-balance-cli.ts

@@ -26,7 +26,7 @@ import {
 } from "../config/model-resolver";
 import { Settings } from "../config/settings";
 import dryBalanceBenchPrompt from "../prompts/dry-balance-bench.md" with { type: "text" };
-import { discoverAuthStorage } from "../sdk";
+import { discoverAuthStorage, loadCliExtensionProviders } from "../sdk";
 
 const DEFAULT_SAMPLE_COUNT = 100;
 const DEFAULT_CONCURRENCY = 32;
@@ -523,8 +523,10 @@ async function runBenchTargets(
 async function createDefaultRuntime(): Promise<DryBalanceRuntime> {
 	const authStorage = await discoverAuthStorage();
 	try {
-		const settings = await Settings.init({ cwd: getProjectDir() });
+		const cwd = getProjectDir();
+		const settings = await Settings.init({ cwd });
 		const modelRegistry = new ModelRegistry(authStorage);
+		await loadCliExtensionProviders(modelRegistry, settings, cwd);
 		return {
 			modelRegistry,
 			settings,

packages/coding-agent/src/sdk.ts

@@ -687,6 +687,37 @@ export async function loadSessionExtensions(
 	return result;
 }
 
+/**
+ * Load discovered/configured extensions and register their providers into
+ * `modelRegistry`, then discover the dynamic provider catalogs. One-shot CLIs
+ * (`omp bench`, dry-balance) build a bare {@link ModelRegistry} that only knows
+ * built-in catalog providers; without this, providers contributed by an
+ * extension (e.g. a custom OpenAI-compatible provider under
+ * `~/.omp/agent/extensions/`) never reach model resolution. Mirrors the
+ * session / `omp models` path: drain the queued provider registrations, then
+ * `refreshRuntimeProviders` so dynamically-discovered models exist before
+ * selectors are resolved.
+ */
+export async function loadCliExtensionProviders(
+	modelRegistry: ModelRegistry,
+	settings: Settings,
+	cwd: string,
+	options: Pick<CreateAgentSessionOptions, "disableExtensionDiscovery" | "additionalExtensionPaths"> = {},
+): Promise<void> {
+	const eventBus = new EventBus();
+	const extensionsResult = await loadSessionExtensions(options, cwd, settings, eventBus);
+	const activeSources = extensionsResult.extensions.map(extension => extension.path);
+	modelRegistry.syncExtensionSources(activeSources);
+	for (const sourceId of new Set(activeSources)) {
+		modelRegistry.clearSourceRegistrations(sourceId);
+	}
+	for (const { name, config, sourceId } of extensionsResult.runtime.pendingProviderRegistrations) {
+		modelRegistry.registerProvider(name, config, sourceId);
+	}
+	extensionsResult.runtime.pendingProviderRegistrations = [];
+	await modelRegistry.refreshRuntimeProviders();
+}
+
 /**
  * Discover skills from cwd and agentDir.
  */