fix(llms): report prompt-only input token usage totals

Subtract cache read and write tokens from reported input token counts in
runtime usage events and run results to avoid double counting cached tokens.

Also add reasoningTokens to AgentUsageEvent so usage metadata can represent
reasoning token counts when providers return them.

apps/cli/src/runtime/run-agent.ts

@@ -362,7 +362,15 @@ export async function runAgent(
 				type: "run_result",
 				finishReason: result.finishReason,
 				iterations: result.iterations,
-				usage,
+				usage: usage && {
+					...usage,
+					inputTokens: Math.max(
+						0,
+						usage.inputTokens -
+							(usage.cacheReadTokens ?? 0) -
+							(usage.cacheWriteTokens ?? 0),
+					),
+				},
 				...(aggregateUsage ? { aggregateUsage } : {}),
 				durationMs: result.durationMs,
 				text: result.text,

sdk/packages/core/src/runtime/orchestration/runtime-event-adapter.ts

@@ -336,17 +336,27 @@ export class RuntimeEventAdapter {
 			cacheWriteTokens: next.cacheWriteTokens,
 			totalCost: next.totalCost,
 		};
+		// Emit prompt-only inputTokens (excluding cache reads/writes) so consumers
+		// can explicitly add cache_read themselves, consistent with other agents.
+		const promptOnlyDeltaInput = Math.max(
+			0,
+			deltaInput - Math.max(0, deltaCacheRead) - Math.max(0, deltaCacheWrite),
+		);
+		const totalPromptOnlyInput = Math.max(
+			0,
+			next.inputTokens - next.cacheReadTokens - next.cacheWriteTokens,
+		);
 		return [
 			{
 				type: "usage",
-				inputTokens: Math.max(0, deltaInput),
+				inputTokens: promptOnlyDeltaInput,
 				outputTokens: Math.max(0, deltaOutput),
 				cacheReadTokens:
 					deltaCacheRead === 0 ? undefined : Math.max(0, deltaCacheRead),
 				cacheWriteTokens:
 					deltaCacheWrite === 0 ? undefined : Math.max(0, deltaCacheWrite),
 				cost: deltaCost === 0 ? undefined : deltaCost,
-				totalInputTokens: next.inputTokens,
+				totalInputTokens: totalPromptOnlyInput,
 				totalOutputTokens: next.outputTokens,
 				totalCacheReadTokens:
 					next.cacheReadTokens === 0 ? undefined : next.cacheReadTokens,

sdk/packages/shared/src/agents/types.ts

@@ -140,6 +140,8 @@ export interface AgentUsageEvent extends AgentEventMetadata {
 	cacheReadTokens?: number;
 	/** Tokens written to cache */
 	cacheWriteTokens?: number;
+	/** Number of tokens used for reasoning */
+	reasoningTokens?: number;
 	/** Cost for this turn */
 	cost?: number;