@agi-cli/server 0.1.108 → 0.1.109

package/package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@agi-cli/server",
-	"version": "0.1.108",
+	"version": "0.1.109",
 	"description": "HTTP API server for AGI CLI",
 	"type": "module",
 	"main": "./src/index.ts",
@@ -29,8 +29,8 @@
 		"typecheck": "tsc --noEmit"
 	},
 	"dependencies": {
-		"@agi-cli/sdk": "0.1.108",
-		"@agi-cli/database": "0.1.108",
+		"@agi-cli/sdk": "0.1.109",
+		"@agi-cli/database": "0.1.109",
 		"drizzle-orm": "^0.44.5",
 		"hono": "^4.9.9",
 		"zod": "^4.1.8"

package/src/routes/session-messages.ts

@@ -120,6 +120,8 @@ export function registerSessionMessagesRoutes(app: Hono) {
 				typeOf: typeof userContext,
 			});
 
+			const reasoning = body?.reasoning === true;
+
 			// Validate model capabilities if tools are allowed for this agent
 			const wantsToolCalls = true; // agent toolset may be non-empty
 			try {
@@ -152,6 +154,7 @@ export function registerSessionMessagesRoutes(app: Hono) {
 				content,
 				oneShot: Boolean(body?.oneShot),
 				userContext,
+				reasoning,
 			});
 			return c.json({ messageId: assistantMessageId }, 202);
 		} catch (error) {

package/src/runtime/message-service.ts

@@ -22,6 +22,7 @@ type DispatchOptions = {
 	content: string;
 	oneShot?: boolean;
 	userContext?: string;
+	reasoning?: boolean;
 };
 
 export async function dispatchAssistantMessage(
@@ -37,6 +38,7 @@ export async function dispatchAssistantMessage(
 		content,
 		oneShot,
 		userContext,
+		reasoning,
 	} = options;
 
 	// DEBUG: Log userContext in dispatch
@@ -108,6 +110,7 @@ export async function dispatchAssistantMessage(
 			projectRoot: cfg.projectRoot,
 			oneShot: Boolean(oneShot),
 			userContext,
+			reasoning,
 		},
 		runSessionLoop,
 	);

package/src/runtime/runner.ts

@@ -298,13 +298,44 @@ async function runAssistant(opts: RunOpts) {
 	let accumulated = '';
 	let stepIndex = 0;
 
+	// Build provider options for reasoning/extended thinking
+	const providerOptions: Record<string, unknown> = {};
+	const THINKING_BUDGET = 16000;
+	// When reasoning is enabled for Anthropic, the API requires max_tokens to fit
+	// both thinking tokens AND response tokens. AI SDK adds budgetTokens to maxOutputTokens,
+	// so we need to reduce maxOutputTokens to leave room for thinking.
+	let effectiveMaxOutputTokens = maxOutputTokens;
+
+	if (opts.reasoning) {
+		if (opts.provider === 'anthropic') {
+			providerOptions.anthropic = {
+				thinking: { type: 'enabled', budgetTokens: THINKING_BUDGET },
+			};
+			// Reduce max output to leave room for thinking budget
+			if (maxOutputTokens && maxOutputTokens > THINKING_BUDGET) {
+				effectiveMaxOutputTokens = maxOutputTokens - THINKING_BUDGET;
+			}
+		} else if (opts.provider === 'openai') {
+			providerOptions.openai = {
+				reasoningSummary: 'auto',
+			};
+		} else if (opts.provider === 'google') {
+			providerOptions.google = {
+				thinkingConfig: { thinkingBudget: THINKING_BUDGET },
+			};
+		}
+	}
+
 	try {
 		const result = streamText({
 			model,
 			tools: toolset,
 			...(cachedSystem ? { system: cachedSystem } : {}),
 			messages: optimizedMessages,
-			...(maxOutputTokens ? { maxOutputTokens } : {}),
+			...(effectiveMaxOutputTokens
+				? { maxOutputTokens: effectiveMaxOutputTokens }
+				: {}),
+			...(Object.keys(providerOptions).length > 0 ? { providerOptions } : {}),
 			abortSignal: opts.abortSignal,
 			stopWhen: hasToolCall('finish'),
 			onStepFinish,

package/src/runtime/session-queue.ts

@@ -9,6 +9,7 @@ export type RunOpts = {
 	projectRoot: string;
 	oneShot?: boolean;
 	userContext?: string;
+	reasoning?: boolean;
 	abortSignal?: AbortSignal;
 };