@agi-cli/server 0.1.136 → 0.1.137

package/package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@agi-cli/server",
-	"version": "0.1.136",
+	"version": "0.1.137",
 	"description": "HTTP API server for AGI CLI",
 	"type": "module",
 	"main": "./src/index.ts",
@@ -29,8 +29,8 @@
 		"typecheck": "tsc --noEmit"
 	},
 	"dependencies": {
-		"@agi-cli/sdk": "0.1.136",
-		"@agi-cli/database": "0.1.136",
+		"@agi-cli/sdk": "0.1.137",
+		"@agi-cli/database": "0.1.137",
 		"drizzle-orm": "^0.44.5",
 		"hono": "^4.9.9",
 		"zod": "^4.1.8"

package/src/openapi/schemas.ts

@@ -64,6 +64,8 @@ export const schemas = {
 			lastActiveAt: { type: 'integer', format: 'int64', nullable: true },
 			totalInputTokens: { type: 'integer', nullable: true },
 			totalOutputTokens: { type: 'integer', nullable: true },
+			totalCachedTokens: { type: 'integer', nullable: true },
+			totalCacheCreationTokens: { type: 'integer', nullable: true },
 			totalToolTimeMs: { type: 'integer', nullable: true },
 			toolCounts: {
 				type: 'object',

package/src/routes/research.ts

@@ -34,6 +34,10 @@ export function registerResearchRoutes(app: Hono) {
 				lastActiveAt: sessions.lastActiveAt,
 				provider: sessions.provider,
 				model: sessions.model,
+				totalInputTokens: sessions.totalInputTokens,
+				totalOutputTokens: sessions.totalOutputTokens,
+				totalCachedTokens: sessions.totalCachedTokens,
+				totalCacheCreationTokens: sessions.totalCacheCreationTokens,
 			})
 			.from(sessions)
 			.where(
@@ -112,6 +116,7 @@ export function registerResearchRoutes(app: Hono) {
 			totalInputTokens: null,
 			totalOutputTokens: null,
 			totalCachedTokens: null,
+			totalCacheCreationTokens: null,
 			totalReasoningTokens: null,
 			totalToolTimeMs: null,
 			toolCountsJson: null,
@@ -332,6 +337,7 @@ export function registerResearchRoutes(app: Hono) {
 			totalInputTokens: null,
 			totalOutputTokens: null,
 			totalCachedTokens: null,
+			totalCacheCreationTokens: null,
 			totalReasoningTokens: null,
 			totalToolTimeMs: null,
 			toolCountsJson: null,

package/src/runtime/message/compaction-limits.ts

@@ -20,7 +20,11 @@ export interface ModelLimits {
 export function isOverflow(tokens: TokenUsage, limits: ModelLimits): boolean {
 	if (limits.context === 0) return false;
 
-	const count = tokens.input + (tokens.cacheRead ?? 0) + tokens.output;
+	const count =
+		tokens.input +
+		(tokens.cacheRead ?? 0) +
+		(tokens.cacheWrite ?? 0) +
+		tokens.output;
 	const usableContext = limits.context - limits.output;
 
 	return count > usableContext;

package/src/runtime/session/branch.ts

@@ -119,6 +119,7 @@ export async function createBranch({
 			completionTokens: msg.completionTokens,
 			totalTokens: msg.totalTokens,
 			cachedInputTokens: msg.cachedInputTokens,
+			cacheCreationInputTokens: msg.cacheCreationInputTokens,
 			reasoningTokens: msg.reasoningTokens,
 			error: msg.error,
 			errorType: msg.errorType,
@@ -163,6 +164,7 @@ export async function createBranch({
 		totalInputTokens: null,
 		totalOutputTokens: null,
 		totalCachedTokens: null,
+		totalCacheCreationTokens: null,
 		totalReasoningTokens: null,
 		totalToolTimeMs: null,
 		toolCountsJson: null,

package/src/runtime/session/db-operations.ts

@@ -1,6 +1,7 @@
 import type { getDb } from '@agi-cli/database';
 import { messages, messageParts, sessions } from '@agi-cli/database/schema';
 import { eq } from 'drizzle-orm';
+import { catalog, type ProviderId } from '@agi-cli/sdk';
 import type { RunOpts } from './queue.ts';
 
 export type UsageData = {
@@ -8,6 +9,7 @@ export type UsageData = {
 	outputTokens?: number;
 	totalTokens?: number;
 	cachedInputTokens?: number;
+	cacheCreationInputTokens?: number;
 	reasoningTokens?: number;
 };
 
@@ -16,8 +18,81 @@ export type ProviderMetadata = Record<string, unknown> & {
 		cachedPromptTokens?: number;
 		[key: string]: unknown;
 	};
+	anthropic?: {
+		cacheCreationInputTokens?: number;
+		cacheReadInputTokens?: number;
+		[key: string]: unknown;
+	};
 };
 
+export function normalizeUsage(
+	usage: UsageData,
+	providerMetadata: ProviderMetadata | undefined,
+	provider: ProviderId,
+): UsageData {
+	const rawInputTokens = Number(usage.inputTokens ?? 0);
+	const outputTokens = Number(usage.outputTokens ?? 0);
+	const reasoningTokens = Number(usage.reasoningTokens ?? 0);
+
+	const cachedInputTokens =
+		usage.cachedInputTokens != null
+			? Number(usage.cachedInputTokens)
+			: providerMetadata?.openai?.cachedPromptTokens != null
+				? Number(providerMetadata.openai.cachedPromptTokens)
+				: providerMetadata?.anthropic?.cacheReadInputTokens != null
+					? Number(providerMetadata.anthropic.cacheReadInputTokens)
+					: undefined;
+
+	const cacheCreationInputTokens =
+		usage.cacheCreationInputTokens != null
+			? Number(usage.cacheCreationInputTokens)
+			: providerMetadata?.anthropic?.cacheCreationInputTokens != null
+				? Number(providerMetadata.anthropic.cacheCreationInputTokens)
+				: undefined;
+
+	const cachedValue = cachedInputTokens ?? 0;
+	const cacheCreationValue = cacheCreationInputTokens ?? 0;
+
+	let inputTokens = rawInputTokens;
+	if (provider === 'openai') {
+		inputTokens = Math.max(0, rawInputTokens - cachedValue);
+	} else if (provider === 'anthropic') {
+		inputTokens = Math.max(0, rawInputTokens - cacheCreationValue);
+	}
+
+	return {
+		inputTokens,
+		outputTokens,
+		cachedInputTokens,
+		cacheCreationInputTokens,
+		reasoningTokens,
+	};
+}
+
+export function resolveUsageProvider(
+	provider: ProviderId,
+	model: string,
+): ProviderId {
+	if (
+		provider !== 'solforge' &&
+		provider !== 'openrouter' &&
+		provider !== 'opencode'
+	) {
+		return provider;
+	}
+	const entry = catalog[provider];
+	const normalizedModel = model.includes('/') ? model.split('/').at(-1) : model;
+	const modelEntry = entry?.models.find(
+		(m) => m.id?.toLowerCase() === normalizedModel?.toLowerCase(),
+	);
+	const npm = modelEntry?.provider?.npm ?? '';
+	if (npm.includes('openai')) return 'openai';
+	if (npm.includes('anthropic')) return 'anthropic';
+	if (npm.includes('google')) return 'google';
+	if (npm.includes('zai')) return 'zai';
+	return provider;
+}
+
 /**
  * Updates session token counts incrementally after each step.
  * Note: onStepFinish.usage is CUMULATIVE per message, so we compute DELTA and add to session.
@@ -30,6 +105,13 @@ export async function updateSessionTokensIncremental(
 ) {
 	if (!usage || !db) return;
 
+	const usageProvider = resolveUsageProvider(opts.provider, opts.model);
+	const normalizedUsage = normalizeUsage(
+		usage,
+		providerMetadata,
+		usageProvider,
+	);
+
 	// Read session totals
 	const sessRows = await db
 		.select()
@@ -42,6 +124,7 @@ export async function updateSessionTokensIncremental(
 	const priorInputSess = Number(sess.totalInputTokens ?? 0);
 	const priorOutputSess = Number(sess.totalOutputTokens ?? 0);
 	const priorCachedSess = Number(sess.totalCachedTokens ?? 0);
+	const priorCacheCreationSess = Number(sess.totalCacheCreationTokens ?? 0);
 	const priorReasoningSess = Number(sess.totalReasoningTokens ?? 0);
 
 	// Read current message totals to compute delta
@@ -54,38 +137,47 @@ export async function updateSessionTokensIncremental(
 	const priorPromptMsg = Number(msg?.promptTokens ?? 0);
 	const priorCompletionMsg = Number(msg?.completionTokens ?? 0);
 	const priorCachedMsg = Number(msg?.cachedInputTokens ?? 0);
+	const priorCacheCreationMsg = Number(msg?.cacheCreationInputTokens ?? 0);
 	const priorReasoningMsg = Number(msg?.reasoningTokens ?? 0);
 
 	// Treat usage as cumulative per-message for this step
 	const cumPrompt =
-		usage.inputTokens != null ? Number(usage.inputTokens) : priorPromptMsg;
+		normalizedUsage.inputTokens != null
+			? Number(normalizedUsage.inputTokens)
+			: priorPromptMsg;
 	const cumCompletion =
-		usage.outputTokens != null
-			? Number(usage.outputTokens)
+		normalizedUsage.outputTokens != null
+			? Number(normalizedUsage.outputTokens)
 			: priorCompletionMsg;
 	const cumReasoning =
-		usage.reasoningTokens != null
-			? Number(usage.reasoningTokens)
+		normalizedUsage.reasoningTokens != null
+			? Number(normalizedUsage.reasoningTokens)
 			: priorReasoningMsg;
 
 	const cumCached =
-		usage.cachedInputTokens != null
-			? Number(usage.cachedInputTokens)
-			: providerMetadata?.openai?.cachedPromptTokens != null
-				? Number(providerMetadata.openai.cachedPromptTokens)
-				: priorCachedMsg;
+		normalizedUsage.cachedInputTokens != null
+			? Number(normalizedUsage.cachedInputTokens)
+			: priorCachedMsg;
+
+	const cumCacheCreation =
+		normalizedUsage.cacheCreationInputTokens != null
+			? Number(normalizedUsage.cacheCreationInputTokens)
+			: priorCacheCreationMsg;
 
 	// Compute deltas for this step; clamp to 0 in case provider reports smaller values
 	const deltaInput = Math.max(0, cumPrompt - priorPromptMsg);
 	const deltaOutput = Math.max(0, cumCompletion - priorCompletionMsg);
 	const deltaCached = Math.max(0, cumCached - priorCachedMsg);
+	const deltaCacheCreation = Math.max(
+		0,
+		cumCacheCreation - priorCacheCreationMsg,
+	);
 	const deltaReasoning = Math.max(0, cumReasoning - priorReasoningMsg);
 
-	// Note: AI SDK's inputTokens already excludes cached tokens for Anthropic,
-	// so we don't need to subtract deltaCached here. Just accumulate directly.
 	const nextInputSess = priorInputSess + deltaInput;
 	const nextOutputSess = priorOutputSess + deltaOutput;
 	const nextCachedSess = priorCachedSess + deltaCached;
+	const nextCacheCreationSess = priorCacheCreationSess + deltaCacheCreation;
 	const nextReasoningSess = priorReasoningSess + deltaReasoning;
 
 	await db
@@ -94,6 +186,7 @@ export async function updateSessionTokensIncremental(
 			totalInputTokens: nextInputSess,
 			totalOutputTokens: nextOutputSess,
 			totalCachedTokens: nextCachedSess,
+			totalCacheCreationTokens: nextCacheCreationSess,
 			totalReasoningTokens: nextReasoningSess,
 		})
 		.where(eq(sessions.id, opts.sessionId));
@@ -144,6 +237,13 @@ export async function updateMessageTokensIncremental(
 ) {
 	if (!usage || !db) return;
 
+	const usageProvider = resolveUsageProvider(opts.provider, opts.model);
+	const normalizedUsage = normalizeUsage(
+		usage,
+		providerMetadata,
+		usageProvider,
+	);
+
 	const msgRows = await db
 		.select()
 		.from(messages)
@@ -154,28 +254,35 @@ export async function updateMessageTokensIncremental(
 		const priorPrompt = Number(msg.promptTokens ?? 0);
 		const priorCompletion = Number(msg.completionTokens ?? 0);
 		const priorCached = Number(msg.cachedInputTokens ?? 0);
+		const priorCacheCreation = Number(msg.cacheCreationInputTokens ?? 0);
 		const priorReasoning = Number(msg.reasoningTokens ?? 0);
 
 		// Treat usage as cumulative per-message - REPLACE not ADD
 		const cumPrompt =
-			usage.inputTokens != null ? Number(usage.inputTokens) : priorPrompt;
+			normalizedUsage.inputTokens != null
+				? Number(normalizedUsage.inputTokens)
+				: priorPrompt;
 		const cumCompletion =
-			usage.outputTokens != null ? Number(usage.outputTokens) : priorCompletion;
+			normalizedUsage.outputTokens != null
+				? Number(normalizedUsage.outputTokens)
+				: priorCompletion;
 		const cumReasoning =
-			usage.reasoningTokens != null
-				? Number(usage.reasoningTokens)
+			normalizedUsage.reasoningTokens != null
+				? Number(normalizedUsage.reasoningTokens)
 				: priorReasoning;
 
 		const cumCached =
-			usage.cachedInputTokens != null
-				? Number(usage.cachedInputTokens)
-				: providerMetadata?.openai?.cachedPromptTokens != null
-					? Number(providerMetadata.openai.cachedPromptTokens)
-					: priorCached;
+			normalizedUsage.cachedInputTokens != null
+				? Number(normalizedUsage.cachedInputTokens)
+				: priorCached;
+
+		const cumCacheCreation =
+			normalizedUsage.cacheCreationInputTokens != null
+				? Number(normalizedUsage.cacheCreationInputTokens)
+				: priorCacheCreation;
 
-		// Note: AI SDK's totalTokens excludes cachedInputTokens for Anthropic,
-		// so we always compute total ourselves to include all token types.
-		const cumTotal = cumPrompt + cumCompletion + cumCached + cumReasoning;
+		const cumTotal =
+			cumPrompt + cumCompletion + cumCached + cumCacheCreation + cumReasoning;
 
 		await db
 			.update(messages)
@@ -184,6 +291,7 @@ export async function updateMessageTokensIncremental(
 				completionTokens: cumCompletion,
 				totalTokens: cumTotal,
 				cachedInputTokens: cumCached,
+				cacheCreationInputTokens: cumCacheCreation,
 				reasoningTokens: cumReasoning,
 			})
 			.where(eq(messages.id, opts.assistantMessageId));

package/src/runtime/session/manager.ts

@@ -51,6 +51,7 @@ export async function createSession({
 		totalInputTokens: null,
 		totalOutputTokens: null,
 		totalCachedTokens: null,
+		totalCacheCreationTokens: null,
 		totalReasoningTokens: null,
 		totalToolTimeMs: null,
 		toolCountsJson: null,

package/src/runtime/stream/finish-handler.ts

@@ -13,6 +13,10 @@ import {
 } from '../message/compaction.ts';
 import { debugLog } from '../debug/index.ts';
 import type { FinishEvent } from './types.ts';
+import {
+	normalizeUsage,
+	resolveUsageProvider,
+} from '../session/db-operations.ts';
 
 export function createFinishHandler(
 	opts: RunOpts,
@@ -73,8 +77,17 @@ export function createFinishHandler(
 					outputTokens: Number(sessRows[0].completionTokens ?? 0),
 					totalTokens: Number(sessRows[0].totalTokens ?? 0),
 					cachedInputTokens: Number(sessRows[0].cachedInputTokens ?? 0),
+					cacheCreationInputTokens: Number(
+						sessRows[0].cacheCreationInputTokens ?? 0,
+					),
 				}
-			: fin.usage;
+			: fin.usage
+				? normalizeUsage(
+						fin.usage,
+						undefined,
+						resolveUsageProvider(opts.provider, opts.model),
+					)
+				: undefined;
 
 		const costUsd = usage
 			? estimateModelCostUsd(opts.provider, opts.model, usage)
@@ -89,6 +102,9 @@ export function createFinishHandler(
 						output: usage.outputTokens ?? 0,
 						cacheRead:
 							(usage as { cachedInputTokens?: number }).cachedInputTokens ?? 0,
+						cacheWrite:
+							(usage as { cacheCreationInputTokens?: number })
+								.cacheCreationInputTokens ?? 0,
 					};
 
 					if (isOverflow(tokenUsage, limits)) {

package/src/runtime/stream/types.ts

@@ -8,7 +8,15 @@ export type StepFinishEvent = {
 };
 
 export type FinishEvent = {
-	usage?: Pick<UsageData, 'inputTokens' | 'outputTokens' | 'totalTokens'>;
+	usage?: Pick<
+		UsageData,
+		| 'inputTokens'
+		| 'outputTokens'
+		| 'totalTokens'
+		| 'cachedInputTokens'
+		| 'cacheCreationInputTokens'
+		| 'reasoningTokens'
+	>;
 	finishReason?: string;
 };
 

package/src/tools/database/get-parent-session.ts

@@ -78,7 +78,10 @@ export function buildGetParentSessionTool(
 				}
 
 				const totalTokens =
-					(session.totalInputTokens ?? 0) + (session.totalOutputTokens ?? 0);
+					(session.totalInputTokens ?? 0) +
+					(session.totalOutputTokens ?? 0) +
+					(session.totalCachedTokens ?? 0) +
+					(session.totalCacheCreationTokens ?? 0);
 
 				const stats = {
 					totalMessages: msgCountResult[0]?.count ?? 0,
@@ -87,6 +90,8 @@ export function buildGetParentSessionTool(
 					totalTokens,
 					totalInputTokens: session.totalInputTokens ?? 0,
 					totalOutputTokens: session.totalOutputTokens ?? 0,
+					totalCachedTokens: session.totalCachedTokens ?? 0,
+					totalCacheCreationTokens: session.totalCacheCreationTokens ?? 0,
 				};
 
 				let messagesData:

package/src/tools/database/get-session-context.ts

@@ -75,7 +75,10 @@ export function buildGetSessionContextTool(projectRoot: string) {
 				}
 
 				const totalTokens =
-					(session.totalInputTokens ?? 0) + (session.totalOutputTokens ?? 0);
+					(session.totalInputTokens ?? 0) +
+					(session.totalOutputTokens ?? 0) +
+					(session.totalCachedTokens ?? 0) +
+					(session.totalCacheCreationTokens ?? 0);
 
 				const stats = {
 					totalMessages: msgCountResult[0]?.count ?? 0,
@@ -84,6 +87,8 @@ export function buildGetSessionContextTool(projectRoot: string) {
 					totalTokens,
 					totalInputTokens: session.totalInputTokens ?? 0,
 					totalOutputTokens: session.totalOutputTokens ?? 0,
+					totalCachedTokens: session.totalCachedTokens ?? 0,
+					totalCacheCreationTokens: session.totalCacheCreationTokens ?? 0,
 				};
 
 				let messagesData:

package/src/tools/database/query-sessions.ts

@@ -82,6 +82,8 @@ export function buildQuerySessionsTool(projectRoot: string) {
 						lastActiveAt: sessions.lastActiveAt,
 						totalInputTokens: sessions.totalInputTokens,
 						totalOutputTokens: sessions.totalOutputTokens,
+						totalCachedTokens: sessions.totalCachedTokens,
+						totalCacheCreationTokens: sessions.totalCacheCreationTokens,
 						sessionType: sessions.sessionType,
 						parentSessionId: sessions.parentSessionId,
 					})