feat: implement Small Model Mode (Phases 1 & 2) #391

.nanocoder/tasks.json

@@ -0,0 +1 @@
+[]

biome.json

@@ -1,5 +1,5 @@
 {
-	"$schema": "https://biomejs.dev/schemas/2.3.14/schema.json",
+	"$schema": "https://biomejs.dev/schemas/2.4.6/schema.json",
 	"vcs": {
 		"enabled": true,
 		"clientKind": "git",

package.json

@@ -59,13 +59,14 @@
 		"assets/nanocoder-vscode.vsix"
 	],
 	"dependencies": {
-		"@ai-sdk/anthropic": "^3.0.37",
-		"@ai-sdk/google": "^3.0.13",
+		"@ai-sdk/anthropic": "^3.0.58",
+		"@ai-sdk/google": "^3.0.43",
 		"@ai-sdk/openai": "^3.0.41",
 		"@ai-sdk/openai-compatible": "2.0.31",
 		"@anthropic-ai/tokenizer": "^0.0.4",
 		"@modelcontextprotocol/sdk": "^1.26.0",
 		"@nanocollective/get-md": "^1.0.2",
+
 		"ai": "6.0.116",
 		"chalk": "^5.2.0",
 		"cheerio": "^1.1.2",
@@ -95,6 +96,7 @@
 	},
 	"devDependencies": {
 		"@ava/typescript": "^6.0.0",
+		"@types/wrap-ansi": "^8.1.0",
 		"@biomejs/biome": "^2.3.10",
 		"@types/node": "^25.0.3",
 		"@types/react": "^19.0.0",

source/ai-sdk-client/ai-sdk-client.ts

@@ -168,4 +168,10 @@ export class AISDKClient implements LLMClient {
 		// No internal state to clear
 		return Promise.resolve();
 	}
+
+	getTimeout(): number | undefined {
+		return (
+			this.providerConfig.socketTimeout ?? this.providerConfig.requestTimeout
+		);
+	}
 }

source/ai-sdk-client/chat/chat-handler.ts

@@ -37,6 +37,32 @@ import {
 	createPrepareStepHandler,
 } from './streaming-handler.js';
 
+/**
+ * Recursively removes 'description' and 'example' fields from a JSON schema
+ * to reduce token usage and cognitive load for small models.
+ */
+// biome-ignore lint/suspicious/noExplicitAny: Necessary for generic JSON schema transformation
+export function simplifyToolSchema(schema: any): any {
+	if (!schema || typeof schema !== 'object') {
+		return schema;
+	}
+
+	if (Array.isArray(schema)) {
+		return schema.map(item => simplifyToolSchema(item));
+	}
+
+	// biome-ignore lint/suspicious/noExplicitAny: Final schema object build
+	const result: any = {};
+	for (const key of Object.keys(schema)) {
+		// Skip description and example fields
+		if (key === 'description' || key === 'example' || key === 'examples') {
+			continue;
+		}
+		result[key] = simplifyToolSchema(schema[key]);
+	}
+	return result;
+}
+
 export interface ChatHandlerParams {
 	model: LanguageModel;
 	currentModel: string;
@@ -131,6 +157,27 @@ export async function handleChat(
 				);
 			}
 
+			// Apply schema simplification if enabled
+			const smmConfig = providerConfig.smallModelMode;
+			if (smmConfig?.enabled && smmConfig.simplifiedSchemas) {
+				effectiveTools = Object.fromEntries(
+					Object.entries(effectiveTools).map(([name, toolDef]) => {
+						// biome-ignore lint/suspicious/noExplicitAny: Cast to access inputSchema which is an internal AI SDK type often nested deeply
+						const toolAny = toolDef as any;
+						return [
+							name,
+							{
+								...toolDef,
+								inputSchema: simplifyToolSchema(toolAny.inputSchema),
+							} as AISDKCoreTool,
+						];
+					}),
+				);
+				logger.debug('Simplified tool schemas for small model mode', {
+					correlationId,
+				});
+			}
+
 			// Tools are already in AI SDK format - use directly
 			const aiTools = shouldDisableTools
 				? undefined
@@ -255,17 +302,63 @@ export async function handleChat(
 
 			// Check if error indicates tool support issue and we haven't retried
 			if (!skipTools && isToolSupportError(error)) {
-				logger.warn('Tool support error detected, retrying without tools', {
-					model: currentModel,
-					error: error instanceof Error ? error.message : error,
-					correlationId,
-					provider: providerConfig.name,
-				});
+				const smmConfig = providerConfig.smallModelMode;
+
+				// Case 1: Already in SMM, but hit a tool error.
+				// Enable aggressive simplification and slim prompt for the retry.
+				if (smmConfig?.enabled) {
+					logger.warn(
+						'Tool support error in Small Model Mode, retrying with maximized simplification',
+						{
+							model: currentModel,
+							correlationId,
+						},
+					);
+
+					const maxSimpConfig = {
+						...providerConfig,
+						smallModelMode: {
+							...smmConfig,
+							simplifiedSchemas: true,
+							slimPrompt: true,
+							// If minimal profile isn't already active, maybe try it?
+							// For now just stick to maximizing simplification.
+						},
+					};
+
+					return await handleChat({
+						...params,
+						providerConfig: maxSimpConfig,
+						skipTools: true, // This currently disables native tools entirely.
+						// Wait, skipTools=true in handleChat makes it use XML for ALL tools.
+						// That might actually be better for small models failing native tools.
+					});
+				}
+
+				// Case 2: Not in SMM. Enable it for the retry.
+				logger.warn(
+					'Tool support error detected, retrying with Small Model Mode enabled',
+					{
+						model: currentModel,
+						error: error instanceof Error ? error.message : error,
+						correlationId,
+						provider: providerConfig.name,
+					},
+				);
+
+				const smmRetryConfig = {
+					...providerConfig,
+					smallModelMode: {
+						enabled: true,
+						simplifiedSchemas: true,
+						slimPrompt: true,
+					},
+				};
 
-				// Retry without tools
 				return await handleChat({
 					...params,
-					skipTools: true, // Mark that we're retrying
+					providerConfig: smmRetryConfig,
+					skipTools: true, // Use XML fallback for retry
 				});
 			}
 

source/app/App.tsx

@@ -1,11 +1,11 @@
 import {Box, Text, useApp} from 'ink';
 import Spinner from 'ink-spinner';
 import React, {useEffect, useMemo} from 'react';
+import {shouldRenderWelcome} from '@/app';
 import {createStaticComponents} from '@/app/components/app-container';
 import {ChatHistory} from '@/app/components/chat-history';
 import {ChatInput} from '@/app/components/chat-input';
 import {ModalSelectors} from '@/app/components/modal-selectors';
-import {shouldRenderWelcome} from '@/app/helpers';
 import type {AppProps} from '@/app/types';
 import {FileExplorer} from '@/components/file-explorer';
 import {IdeSelector} from '@/components/ide-selector';

source/app/prompts/main-prompt-slim.md

@@ -0,0 +1,13 @@
+You are Nanocoder, a terminal-based AI coding assistant. Assist the user with software development tasks using tools.
+
+## RULES
+- Be concise. Focus on technical accuracy.
+- Use `create_task` to break down multi-step work before starting.
+- Never use bash tools (`execute_bash`) for exploring files.
+- ALWAYS use native tools: `find_files`, `search_file_contents`, `read_file`, `list_directory`.
+- Make targeted edits with `string_replace` or `write_file`.
+- Chain your tool calls sequentially. Never stop working after one tool call unless the user's task is fully complete.
+- Verify your edits.
+
+<!-- DYNAMIC_SYSTEM_INFO_START -->
+<!-- DYNAMIC_SYSTEM_INFO_END -->

source/config/index.ts

@@ -145,12 +145,130 @@ function tryLoadAutoCompactFromPath(
 	return null;
 }
 
+// Try to load small model mode config from a specific path
+// Returns the config if found and valid, null otherwise
+function tryLoadSmallModelModeFromPath(
+	configPath: string,
+	defaults: NonNullable<AppConfig['smallModelMode']>,
+): AppConfig['smallModelMode'] | null {
+	if (!existsSync(configPath)) {
+		return null;
+	}
+
+	try {
+		const rawData = readFileSync(configPath, 'utf-8');
+		const config = JSON.parse(rawData);
+		const smallModelMode =
+			config.nanocoder?.smallModelMode || config.smallModelMode;
+
+		if (smallModelMode && typeof smallModelMode === 'object') {
+			return {
+				enabled:
+					smallModelMode.enabled !== undefined
+						? Boolean(smallModelMode.enabled)
+						: defaults.enabled,
+				slimPrompt:
+					smallModelMode.slimPrompt !== undefined
+						? Boolean(smallModelMode.slimPrompt)
+						: defaults.slimPrompt,
+				toolProfile:
+					smallModelMode.toolProfile !== undefined
+						? smallModelMode.toolProfile
+						: defaults.toolProfile,
+				maxToolsPerTurn:
+					typeof smallModelMode.maxToolsPerTurn === 'number'
+						? smallModelMode.maxToolsPerTurn
+						: defaults.maxToolsPerTurn,
+				aggressiveCompact:
+					smallModelMode.aggressiveCompact !== undefined
+						? Boolean(smallModelMode.aggressiveCompact)
+						: defaults.aggressiveCompact,
+				simplifiedSchemas:
+					smallModelMode.simplifiedSchemas !== undefined
+						? Boolean(smallModelMode.simplifiedSchemas)
+						: defaults.simplifiedSchemas,
+				plannerModel: smallModelMode.plannerModel || defaults.plannerModel,
+			};
+		}
+	} catch (error) {
+		logError(
+			`Failed to load small model mode config from ${configPath}: ${String(error)}`,
+		);
+	}
+
+	return null;
+}
+
+// Load small model mode configuration
+function loadSmallModelModeConfig(): AppConfig['smallModelMode'] {
+	const defaults: NonNullable<AppConfig['smallModelMode']> = {
+		enabled: false,
+		slimPrompt: true,
+		toolProfile: 'minimal',
+		maxToolsPerTurn: 1,
+		aggressiveCompact: true,
+		simplifiedSchemas: true,
+	};
+
+	// Try to load from project-level config first
+	const projectConfigPath = join(process.cwd(), 'agents.config.json');
+	const projectConfig = tryLoadSmallModelModeFromPath(
+		projectConfigPath,
+		defaults,
+	);
+	if (projectConfig) {
+		return projectConfig;
+	}
+
+	// Try global config
+	const configDir = getConfigPath();
+	const globalConfigPath = join(configDir, 'agents.config.json');
+	const globalConfig = tryLoadSmallModelModeFromPath(
+		globalConfigPath,
+		defaults,
+	);
+	if (globalConfig) {
+		return globalConfig;
+	}
+
+	// Fallback to home directory
+	const homePath = join(homedir(), '.agents.config.json');
+	const homeConfig = tryLoadSmallModelModeFromPath(homePath, defaults);
+	if (homeConfig) {
+		return homeConfig;
+	}
+
+	return defaults;
+}
+
+/**
+ * Checks if a model name matches known "small" model patterns (< 10B parameters)
+ */
+export function isSmallModel(modelName: string): boolean {
+	const smallModelPatterns = [
+		/llama-?3\.?[12]?.*?[138]b/i,
+		/gemma-?2?.*?[279]b/i,
+		/phi-?[34]/i,
+		/qwen.*?([0-9.]+)b/i, // Broadly catch qwen small models
+		/mistral.*?7b/i,
+		/deepseek.*?coder.*?1\.3b/i,
+		/stable-?lm-?3b/i,
+	];
+
+	return smallModelPatterns.some(pattern => pattern.test(modelName));
+}
+
 // Load auto-compact configuration and Returns default config if not specified
-function loadAutoCompactConfig(): AutoCompactConfig {
+function loadAutoCompactConfig(
+	smallModelMode?: AppConfig['smallModelMode'],
+): AutoCompactConfig {
+	const isAggressiveDefault =
+		smallModelMode?.enabled && smallModelMode?.aggressiveCompact;
+
 	const defaults: AutoCompactConfig = {
 		enabled: true,
-		threshold: 60,
-		mode: 'conservative',
+		threshold: isAggressiveDefault ? 40 : 60,
+		mode: isAggressiveDefault ? 'aggressive' : 'conservative',
 		notifyUser: true,
 	};
 
@@ -295,16 +413,38 @@ function loadAppConfig(): AppConfig {
 	const mcpServersWithSource = loadAllMCPConfigs();
 	const mcpServers = mcpServersWithSource.map(item => item.server);
 
-	// Load auto-compact configuration
-	const autoCompact = loadAutoCompactConfig();
+	// Load global small model mode configuration
+	const smallModelMode = loadSmallModelModeConfig();
+
+	// Load auto-compact configuration, passing small model mode to inform defaults
+	const autoCompact = loadAutoCompactConfig(smallModelMode);
 
 	// Load session configuration
 	const sessions = loadSessionConfig();
 
+	// Auto-detect small model mode for providers if not explicitly configured
+	const augmentedProviders = providers.map(provider => {
+		if (provider.smallModelMode?.enabled === undefined) {
+			// Check if the primary model (first one) is a small model
+			const primaryModel = provider.models[0];
+			if (primaryModel && isSmallModel(primaryModel)) {
+				return {
+					...provider,
+					smallModelMode: {
+						...smallModelMode, // Use global defaults
+						enabled: true,
+					},
+				};
+			}
+		}
+		return provider;
+	});
+
 	return {
-		providers,
+		providers: augmentedProviders,
 		mcpServers,
 		autoCompact,
+		smallModelMode,
 		sessions,
 	};
 }