feat: add lite mode for smaller context in local models

- Add liteMode option to SystemPromptSettings type
- Create condensed tool descriptions for lite mode
- Add lite mode handling to system prompt sections (rules, capabilities)
- Reduce system prompt from ~14k to ~4k tokens when lite mode enabled
- Add UI checkbox in Context Management settings for lite mode
- Optimized for local models with limited context windows (32k, 64k)

Addresses #9351

src/core/prompts/sections/capabilities.ts

@@ -16,6 +16,18 @@ export function getCapabilitiesSection(
 	codeIndexManager?: CodeIndexManager,
 	settings?: SystemPromptSettings,
 ): string {
+	// Return simplified capabilities for lite mode
+	if (settings?.liteMode) {
+		return `====
+
+CAPABILITIES
+
+- CLI commands, file operations (list/read/write), search, definitions
+- Project structure in environment_details
+- Each command runs in new terminal
+${mcpHub ? "- MCP servers provide additional tools\n" : ""}`
+	}
+
 	// Get available tools from relevant groups
 	const availableEditTools = getAvailableToolsInGroup(
 		"edit",

src/core/prompts/sections/rules.ts

@@ -86,6 +86,20 @@ export function getRulesSection(
 	codeIndexManager?: CodeIndexManager,
 	settings?: SystemPromptSettings,
 ): string {
+	// Return simplified rules for lite mode
+	if (settings?.liteMode) {
+		return `====
+
+RULES
+
+- Base directory: ${cwd.toPosix()}
+- Use relative paths from base directory
+- Read files before editing
+- Wait for user confirmation after each tool use
+- Use attempt_completion to present final results
+- Be direct and technical, not conversational`
+	}
+
 	const isCodebaseSearchAvailable =
 		codeIndexManager &&
 		codeIndexManager.isFeatureEnabled &&
@@ -132,6 +146,7 @@ export function getRulesSection(
 	// Determine whether to use XML tool references based on protocol
 	const effectiveProtocol = getEffectiveProtocol(settings?.toolProtocol)
 
+	// Full rules for regular mode
 	return `====
 
 RULES

src/core/prompts/tools/index.ts

@@ -28,10 +28,38 @@ import { getRunSlashCommandDescription } from "./run-slash-command"
 import { getGenerateImageDescription } from "./generate-image"
 import { CodeIndexManager } from "../../../services/code-index/manager"
 
+// Import lite descriptions
+import {
+	getLiteReadFileDescription,
+	getLiteWriteToFileDescription,
+	getLiteSearchFilesDescription,
+	getLiteListFilesDescription,
+	getLiteExecuteCommandDescription,
+	getLiteInsertContentDescription,
+	getLiteListCodeDefinitionNamesDescription,
+	getLiteAskFollowupQuestionDescription,
+	getLiteAttemptCompletionDescription,
+	getLiteBrowserActionDescription,
+	getLiteSwitchModeDescription,
+	getLiteNewTaskDescription,
+	getLiteUpdateTodoListDescription,
+	getLiteFetchInstructionsDescription,
+	getLiteApplyDiffDescription,
+	getLiteCodebaseSearchDescription,
+	getLiteUseMcpToolDescription,
+	getLiteAccessMcpResourceDescription,
+	getLiteGenerateImageDescription,
+	getLiteRunSlashCommandDescription,
+} from "./lite-descriptions"
+
 // Map of tool names to their description functions
 const toolDescriptionMap: Record<string, (args: ToolArgs) => string | undefined> = {
-	execute_command: (args) => getExecuteCommandDescription(args),
+	execute_command: (args) => {
+		if (args.settings?.liteMode) return getLiteExecuteCommandDescription()
+		return getExecuteCommandDescription(args)
+	},
 	read_file: (args) => {
+		if (args.settings?.liteMode) return getLiteReadFileDescription(args)
 		// Check if the current model should use the simplified read_file tool
 		const modelId = args.settings?.modelId
 		if (modelId && shouldUseSingleFileRead(modelId)) {
@@ -39,25 +67,80 @@ const toolDescriptionMap: Record<string, (args: ToolArgs) => string | undefined>
 		}
 		return getReadFileDescription(args)
 	},
-	fetch_instructions: (args) => getFetchInstructionsDescription(args.settings?.enableMcpServerCreation),
-	write_to_file: (args) => getWriteToFileDescription(args),
-	search_files: (args) => getSearchFilesDescription(args),
-	list_files: (args) => getListFilesDescription(args),
-	list_code_definition_names: (args) => getListCodeDefinitionNamesDescription(args),
-	browser_action: (args) => getBrowserActionDescription(args),
-	ask_followup_question: () => getAskFollowupQuestionDescription(),
-	attempt_completion: (args) => getAttemptCompletionDescription(args),
-	use_mcp_tool: (args) => getUseMcpToolDescription(args),
-	access_mcp_resource: (args) => getAccessMcpResourceDescription(args),
-	codebase_search: (args) => getCodebaseSearchDescription(args),
-	switch_mode: () => getSwitchModeDescription(),
-	new_task: (args) => getNewTaskDescription(args),
-	insert_content: (args) => getInsertContentDescription(args),
-	apply_diff: (args) =>
-		args.diffStrategy ? args.diffStrategy.getToolDescription({ cwd: args.cwd, toolOptions: args.toolOptions }) : "",
-	update_todo_list: (args) => getUpdateTodoListDescription(args),
-	run_slash_command: () => getRunSlashCommandDescription(),
-	generate_image: (args) => getGenerateImageDescription(args),
+	fetch_instructions: (args) => {
+		if (args.settings?.liteMode) return getLiteFetchInstructionsDescription()
+		return getFetchInstructionsDescription(args.settings?.enableMcpServerCreation)
+	},
+	write_to_file: (args) => {
+		if (args.settings?.liteMode) return getLiteWriteToFileDescription()
+		return getWriteToFileDescription(args)
+	},
+	search_files: (args) => {
+		if (args.settings?.liteMode) return getLiteSearchFilesDescription()
+		return getSearchFilesDescription(args)
+	},
+	list_files: (args) => {
+		if (args.settings?.liteMode) return getLiteListFilesDescription()
+		return getListFilesDescription(args)
+	},
+	list_code_definition_names: (args) => {
+		if (args.settings?.liteMode) return getLiteListCodeDefinitionNamesDescription()
+		return getListCodeDefinitionNamesDescription(args)
+	},
+	browser_action: (args) => {
+		if (args.settings?.liteMode) return getLiteBrowserActionDescription()
+		return getBrowserActionDescription(args)
+	},
+	ask_followup_question: (args) => {
+		if (args.settings?.liteMode) return getLiteAskFollowupQuestionDescription()
+		return getAskFollowupQuestionDescription()
+	},
+	attempt_completion: (args) => {
+		if (args.settings?.liteMode) return getLiteAttemptCompletionDescription()
+		return getAttemptCompletionDescription(args)
+	},
+	use_mcp_tool: (args) => {
+		if (args.settings?.liteMode) return getLiteUseMcpToolDescription()
+		return getUseMcpToolDescription(args)
+	},
+	access_mcp_resource: (args) => {
+		if (args.settings?.liteMode) return getLiteAccessMcpResourceDescription()
+		return getAccessMcpResourceDescription(args)
+	},
+	codebase_search: (args) => {
+		if (args.settings?.liteMode) return getLiteCodebaseSearchDescription()
+		return getCodebaseSearchDescription(args)
+	},
+	switch_mode: (args) => {
+		if (args.settings?.liteMode) return getLiteSwitchModeDescription()
+		return getSwitchModeDescription()
+	},
+	new_task: (args) => {
+		if (args.settings?.liteMode) return getLiteNewTaskDescription()
+		return getNewTaskDescription(args)
+	},
+	insert_content: (args) => {
+		if (args.settings?.liteMode) return getLiteInsertContentDescription()
+		return getInsertContentDescription(args)
+	},
+	apply_diff: (args) => {
+		if (args.settings?.liteMode) return getLiteApplyDiffDescription(args.diffStrategy)
+		return args.diffStrategy
+			? args.diffStrategy.getToolDescription({ cwd: args.cwd, toolOptions: args.toolOptions })
+			: ""
+	},
+	update_todo_list: (args) => {
+		if (args.settings?.liteMode) return getLiteUpdateTodoListDescription()
+		return getUpdateTodoListDescription(args)
+	},
+	run_slash_command: (args) => {
+		if (args.settings?.liteMode) return getLiteRunSlashCommandDescription()
+		return getRunSlashCommandDescription()
+	},
+	generate_image: (args) => {
+		if (args.settings?.liteMode) return getLiteGenerateImageDescription()
+		return getGenerateImageDescription(args)
+	},
 }
 
 export function getToolDescriptionsForMode(

src/core/prompts/tools/lite-descriptions.ts

@@ -0,0 +1,127 @@
+import { ToolArgs } from "./types"
+import { DiffStrategy } from "../../../shared/tools"
+
+/**
+ * Returns a compact tool description without examples for lite mode
+ */
+export function getLiteReadFileDescription(args: ToolArgs): string {
+	const maxReads = args.settings?.maxConcurrentFileReads ?? 5
+	return `## read_file
+Read file contents (max ${maxReads} files/request). Line-numbered output.
+Params: path (required), line_range (optional)`
+}
+
+export function getLiteWriteToFileDescription(): string {
+	return `## write_to_file
+Create/overwrite file with content.
+Params: path, content, line_count (required)`
+}
+
+export function getLiteSearchFilesDescription(): string {
+	return `## search_files
+Regex search in directory.
+Params: path, regex (required), file_pattern (optional)`
+}
+
+export function getLiteListFilesDescription(): string {
+	return `## list_files
+List directory contents.
+Params: path (required), recursive (optional)`
+}
+
+export function getLiteExecuteCommandDescription(): string {
+	return `## execute_command
+Execute CLI command.
+Params: command (required), cwd (optional)`
+}
+
+export function getLiteInsertContentDescription(): string {
+	return `## insert_content
+Insert lines at specific position.
+Params: path, line, content (required)`
+}
+
+export function getLiteListCodeDefinitionNamesDescription(): string {
+	return `## list_code_definition_names
+List code definitions (classes, functions, etc).
+Params: path (required)`
+}
+
+export function getLiteAskFollowupQuestionDescription(): string {
+	return `## ask_followup_question
+Ask user for clarification.
+Params: question, follow_up with 2-4 suggest tags`
+}
+
+export function getLiteAttemptCompletionDescription(): string {
+	return `## attempt_completion
+Present final result after task completion.
+Params: result (required)`
+}
+
+export function getLiteBrowserActionDescription(): string {
+	return `## browser_action
+Browser interaction: screenshot, click, type, scroll.
+Params: action (required), coordinate/text/direction/amount based on action`
+}
+
+export function getLiteSwitchModeDescription(): string {
+	return `## switch_mode
+Switch to different mode.
+Params: mode_slug (required), reason (optional)`
+}
+
+export function getLiteNewTaskDescription(): string {
+	return `## new_task
+Create new task in specified mode.
+Params: mode, message (required)`
+}
+
+export function getLiteUpdateTodoListDescription(): string {
+	return `## update_todo_list
+Update TODO checklist.
+Format: [ ] pending, [x] completed, [-] in progress`
+}
+
+export function getLiteFetchInstructionsDescription(): string {
+	return `## fetch_instructions
+Get task instructions.
+Params: task (required) - create_mcp_server or create_mode`
+}
+
+export function getLiteApplyDiffDescription(diffStrategy?: DiffStrategy): string {
+	if (!diffStrategy) return ""
+	return `## apply_diff
+Apply targeted edits to existing file.
+Params: path, diff (SEARCH/REPLACE blocks with :start_line:)`
+}
+
+export function getLiteCodebaseSearchDescription(): string {
+	return `## codebase_search
+Semantic search for relevant code.
+Params: query (required)`
+}
+
+export function getLiteUseMcpToolDescription(): string {
+	return `## use_mcp_tool
+Use MCP server tool.
+Params: server_name, tool_name, arguments (required)`
+}
+
+export function getLiteAccessMcpResourceDescription(): string {
+	return `## access_mcp_resource
+Access MCP server resource.
+Params: server_name, uri (required), arguments (optional)`
+}
+
+export function getLiteGenerateImageDescription(): string {
+	return `## generate_image
+Generate image using AI.
+Params: prompt (required), size (optional)`
+}
+
+export function getLiteRunSlashCommandDescription(): string {
+	return `## run_slash_command
+Run a VS Code slash command.
+Params: command, args (required)`
+}

src/core/prompts/types.ts

@@ -10,4 +10,5 @@ export interface SystemPromptSettings {
 	useAgentRules: boolean
 	newTaskRequireTodos: boolean
 	toolProtocol?: ToolProtocol
+	liteMode?: boolean
 }

src/shared/ExtensionMessage.ts

@@ -349,6 +349,7 @@ export type ExtensionState = Pick<
 	remoteControlEnabled: boolean
 	taskSyncEnabled: boolean
 	featureRoomoteControlEnabled: boolean
+	liteMode?: boolean
 }
 
 export interface ClineSayTool {

webview-ui/src/components/settings/ContextManagementSettings.tsx

@@ -30,6 +30,7 @@ type ContextManagementSettingsProps = HTMLAttributes<HTMLDivElement> & {
 	includeCurrentTime?: boolean
 	includeCurrentCost?: boolean
 	maxGitStatusFiles?: number
+	liteMode?: boolean
 	setCachedStateField: SetCachedStateField<
 		| "autoCondenseContext"
 		| "autoCondenseContextPercent"
@@ -47,6 +48,7 @@ type ContextManagementSettingsProps = HTMLAttributes<HTMLDivElement> & {
 		| "includeCurrentTime"
 		| "includeCurrentCost"
 		| "maxGitStatusFiles"
+		| "liteMode"
 	>
 }
 
@@ -69,6 +71,7 @@ export const ContextManagementSettings = ({
 	includeCurrentTime,
 	includeCurrentCost,
 	maxGitStatusFiles,
+	liteMode,
 	className,
 	...props
 }: ContextManagementSettingsProps) => {
@@ -411,6 +414,20 @@ export const ContextManagementSettings = ({
 						{t("settings:contextManagement.includeCurrentCost.description")}
 					</div>
 				</div>
+
+				<div>
+					<VSCodeCheckbox
+						checked={liteMode}
+						onChange={(e: any) => setCachedStateField("liteMode", e.target.checked)}
+						data-testid="lite-mode-checkbox">
+						<label className="block font-medium mb-1">
+							{t("settings:contextManagement.liteMode.label")}
+						</label>
+					</VSCodeCheckbox>
+					<div className="text-vscode-descriptionForeground text-sm mt-1 mb-3">
+						{t("settings:contextManagement.liteMode.description")}
+					</div>
+				</div>
 			</Section>
 			<Section className="pt-2">
 				<VSCodeCheckbox

webview-ui/src/components/settings/SettingsView.tsx

@@ -204,6 +204,7 @@ const SettingsView = forwardRef<SettingsViewRef, SettingsViewProps>(({ onDone, t
 		includeCurrentTime,
 		includeCurrentCost,
 		maxGitStatusFiles,
+		liteMode,
 	} = cachedState
 
 	const apiConfiguration = useMemo(() => cachedState.apiConfiguration ?? {}, [cachedState.apiConfiguration])
@@ -767,6 +768,7 @@ const SettingsView = forwardRef<SettingsViewRef, SettingsViewProps>(({ onDone, t
 							includeCurrentTime={includeCurrentTime}
 							includeCurrentCost={includeCurrentCost}
 							maxGitStatusFiles={maxGitStatusFiles}
+							liteMode={liteMode}
 							setCachedStateField={setCachedStateField}
 						/>
 					)}

webview-ui/src/context/ExtensionStateContext.tsx

@@ -165,6 +165,8 @@ export interface ExtensionStateContextType extends ExtensionState {
 	setIncludeCurrentTime: (value: boolean) => void
 	includeCurrentCost?: boolean
 	setIncludeCurrentCost: (value: boolean) => void
+	liteMode?: boolean
+	setLiteMode: (value: boolean) => void
 }
 
 export const ExtensionStateContext = createContext<ExtensionStateContextType | undefined>(undefined)
@@ -596,6 +598,8 @@ export const ExtensionStateContextProvider: React.FC<{ children: React.ReactNode
 		setIncludeCurrentTime,
 		includeCurrentCost,
 		setIncludeCurrentCost,
+		liteMode: state.liteMode,
+		setLiteMode: (value) => setState((prevState) => ({ ...prevState, liteMode: value })),
 	}
 
 	return <ExtensionStateContext.Provider value={contextValue}>{children}</ExtensionStateContext.Provider>