enhance context retention between tool calls, fix lints

tools/server/webui/src/lib/components/app/chat/ChatMessages/ChatMessageThinkingBlock.svelte

@@ -5,19 +5,22 @@
 	import { buttonVariants } from '$lib/components/ui/button/index.js';
 	import { Card } from '$lib/components/ui/card';
 	import { config } from '$lib/stores/settings.svelte';
+	import type { Snippet } from 'svelte';
 
 	interface Props {
 		class?: string;
 		hasRegularContent?: boolean;
 		isStreaming?: boolean;
 		reasoningContent: string | null;
+		children?: Snippet;
 	}
 
 	let {
 		class: className = '',
 		hasRegularContent = false,
 		isStreaming = false,
-		reasoningContent
+		reasoningContent,
+		children
 	}: Props = $props();
 
 	const currentConfig = config();
@@ -72,9 +75,11 @@
 			<div class="border-t border-muted px-3 pb-3">
 				<div class="pt-3">
 					<div class="text-xs leading-relaxed break-words whitespace-pre-wrap">
-						<slot>
+						{#if children}
+							{@render children()}
+						{:else}
 							{reasoningContent ?? ''}
-						</slot>
+						{/if}
 					</div>
 				</div>
 			</div>

tools/server/webui/src/lib/services/chat.ts

@@ -119,6 +119,7 @@ export class ChatService {
 			messages: normalizedMessages.map((msg: ApiChatMessageData) => ({
 				role: msg.role,
 				content: msg.content,
+				...(msg.reasoning_content ? { reasoning_content: msg.reasoning_content } : {}),
 				...((msg as ApiChatCompletionRequestMessage).tool_call_id
 					? { tool_call_id: (msg as ApiChatCompletionRequestMessage).tool_call_id }
 					: {}),
@@ -602,6 +603,9 @@ export class ChatService {
 			return {
 				role: message.role as ChatRole,
 				content: message.content,
+				...(message.role === 'assistant' && message.thinking
+					? { reasoning_content: message.thinking }
+					: {}),
 				// tool_call_id is only relevant for tool role messages
 				...(message.toolCallId ? { tool_call_id: message.toolCallId } : {}),
 				...(toolCalls ? { tool_calls: toolCalls } : {})
@@ -693,6 +697,9 @@ export class ChatService {
 		return {
 			role: message.role as ChatRole,
 			content: contentParts,
+			...(message.role === 'assistant' && message.thinking
+				? { reasoning_content: message.thinking }
+				: {}),
 			...(message.toolCallId ? { tool_call_id: message.toolCallId } : {}),
 			...(toolCalls ? { tool_calls: toolCalls } : {})
 		};

tools/server/webui/src/lib/types/api.d.ts

@@ -35,6 +35,13 @@ export interface ApiChatMessageData {
 	role: ChatRole;
 	content: string | ApiChatMessageContentPart[];
 	timestamp?: number;
+	/**
+	 * Optional reasoning/thinking content to be sent back to the server.
+	 *
+	 * llama-server accepts this non-OpenAI field and uses it to preserve the model's
+	 * internal "thinking" blocks across tool-call resumptions (notably for gpt-oss).
+	 */
+	reasoning_content?: string;
 	tool_call_id?: string;
 	tool_calls?: ApiChatCompletionToolCallDelta[];
 }
@@ -183,6 +190,7 @@ export interface ApiLlamaCppServerProps {
 export interface ApiChatCompletionRequestMessage {
 	role: ChatRole;
 	content: string | ApiChatMessageContentPart[];
+	reasoning_content?: string;
 	tool_call_id?: string;
 	tool_calls?: ApiChatCompletionToolCallDelta[];
 }

tools/server/webui/tests/client/chatMessage.delete-merged-assistant.test.ts

@@ -75,9 +75,8 @@ describe('ChatMessage delete for merged assistant messages', () => {
 		conversationsStore.activeMessages = allMessages;
 
 		// Avoid touching IndexedDB by stubbing the store call used by getDeletionInfo.
-		const originalGetConversationMessages = conversationsStore.getConversationMessages.bind(
-			conversationsStore
-		);
+		const originalGetConversationMessages =
+			conversationsStore.getConversationMessages.bind(conversationsStore);
 		conversationsStore.getConversationMessages = async () => allMessages;
 
 		const onDelete = vi.fn();
@@ -111,4 +110,3 @@ describe('ChatMessage delete for merged assistant messages', () => {
 		}
 	});
 });
-

tools/server/webui/tests/client/components/TestChatMessageWrapper.svelte

@@ -14,4 +14,3 @@
 		<ChatMessage message={props.message} onDelete={props.onDelete} />
 	</Sidebar.Provider>
 </Tooltip.Provider>
-

tools/server/webui/tests/e2e/tool-output-no-echo.spec.ts

@@ -137,4 +137,6 @@ test('tool output does not echo tool arguments back to the model', async ({ page
 	);
 	expect(assistantWithToolCall).toBeTruthy();
 	expect(JSON.stringify(assistantWithToolCall?.tool_calls ?? null)).toContain('LARGE_CODE_BEGIN');
+	// Preserve the model's reasoning across tool-call resumptions (required for gpt-oss).
+	expect(String(assistantWithToolCall?.reasoning_content ?? '')).toContain('reasoning-step-1');
 });