codeep 1.0.8 → 1.0.9

package/dist/utils/agent.js

@@ -7,7 +7,7 @@ import { getProviderBaseUrl, getProviderAuthHeader, supportsNativeTools } from '
 import { startSession, endSession, undoLastAction, undoAllActions, getCurrentSession, getRecentSessions, formatSession } from './history.js';
 import { runAllVerifications, formatErrorsForAgent, hasVerificationErrors, getVerificationSummary } from './verify.js';
 import { gatherSmartContext, formatSmartContext, extractTargetFile } from './smartContext.js';
-import { planTasks, formatTaskPlan } from './taskPlanner.js';
+import { planTasks, getNextTask, formatTaskPlan } from './taskPlanner.js';
 const DEFAULT_OPTIONS = {
     maxIterations: 100, // Increased for large tasks
     maxDuration: 20 * 60 * 1000, // 20 minutes
@@ -27,15 +27,12 @@ function getAgentSystemPrompt(projectContext) {
 - List directory contents
 
 ## IMPORTANT: Follow User Instructions Exactly
-- Do EXACTLY what the user asks - complete the ENTIRE task
-- If user says "create a website" -> create ALL necessary files (HTML, CSS, JS, etc.)
+- Do EXACTLY what the user asks
 - If user says "create folder X" -> use create_directory tool to create folder X
 - If user says "delete file X" -> use delete_file tool to delete file X
-- Do NOT stop after just 1-2 tool calls unless the task is trivially simple
-- Complex tasks (like creating websites) require MANY tool calls to complete
 - The user may write in any language - understand their request and execute it
 - Tool names and parameters must ALWAYS be in English (e.g., "create_directory", not "kreiraj_direktorij")
-- KEEP WORKING until the entire task is finished - do not stop prematurely
+- When you finish a subtask, provide a clear summary and I will give you the next subtask
 
 ## Rules
 1. Always read files before editing them to understand the current content
@@ -72,15 +69,12 @@ function getFallbackSystemPrompt(projectContext) {
     return `You are an AI coding agent with FULL autonomous access to this project.
 
 ## IMPORTANT: Follow User Instructions Exactly
-- Do EXACTLY what the user asks - complete the ENTIRE task
-- If user says "create a website" -> create ALL necessary files (HTML, CSS, JS, etc.)
+- Do EXACTLY what the user asks
 - If user says "create folder X" -> use create_directory tool
 - If user says "delete file X" -> use delete_file tool
-- Do NOT stop after just 1-2 tool calls unless the task is trivially simple
-- Complex tasks (like creating websites) require MANY tool calls to complete
 - The user may write in any language - understand and execute
 - Tool names and parameters must ALWAYS be in English
-- KEEP WORKING until the entire task is finished - do not stop prematurely
+- When you finish a subtask, provide a clear summary and I will give you the next subtask
 
 ## Available Tools
 ${formatToolDefinitions()}
@@ -379,9 +373,13 @@ export async function runAgent(prompt, projectContext, options = {}) {
     const messages = [];
     // Start history session for undo support
     const sessionId = startSession(prompt, projectContext.root || process.cwd());
-    // Task planning phase (if enabled and prompt is complex enough)
+    // Task planning phase (if enabled)
+    // Use planning for complex keywords or multi-word prompts
     let taskPlan = null;
-    if (opts.usePlanning && prompt.split(' ').length > 5) {
+    const complexKeywords = ['create', 'build', 'implement', 'add', 'setup', 'generate', 'make', 'develop'];
+    const hasComplexKeyword = complexKeywords.some(kw => prompt.toLowerCase().includes(kw));
+    const shouldPlan = opts.usePlanning && (prompt.split(' ').length > 3 || hasComplexKeyword);
+    if (shouldPlan) {
         try {
             opts.onIteration?.(0, 'Planning tasks...');
             taskPlan = await planTasks(prompt, {
@@ -391,6 +389,8 @@ export async function runAgent(prompt, projectContext, options = {}) {
             });
             if (taskPlan.tasks.length > 1) {
                 opts.onTaskPlan?.(taskPlan);
+                // Mark first task as in_progress
+                taskPlan.tasks[0].status = 'in_progress';
             }
             else {
                 taskPlan = null; // Single task, no need for planning
@@ -479,41 +479,32 @@ export async function runAgent(prompt, projectContext, options = {}) {
                     toolCalls = textToolCalls;
                 }
             }
-            // If no tool calls, check if this is really the final response
-            // Don't exit on first iteration without tool calls - agent might be thinking
+            // If no tool calls, this is the final response
             if (toolCalls.length === 0) {
-                // Only accept as final response if:
-                // 1. We've done at least some work (iteration > 2)
-                // 2. Agent explicitly indicates completion
-                const completionIndicators = [
-                    'task is complete',
-                    'all files have been created',
-                    'website has been created',
-                    'successfully completed',
-                    'everything is ready',
-                    'all done'
-                ];
-                const lowerContent = content.toLowerCase();
-                const indicatesCompletion = completionIndicators.some(indicator => lowerContent.includes(indicator));
-                if (iteration > 2 && indicatesCompletion) {
-                    // Remove <think>...</think> tags from response
-                    finalResponse = content.replace(/<think>[\s\S]*?<\/think>/gi, '').trim();
-                    break;
-                }
-                else if (iteration <= 2) {
-                    // Too early to quit - remind agent to continue
-                    messages.push({ role: 'assistant', content });
-                    messages.push({
-                        role: 'user',
-                        content: 'Continue with the task. Use the tools to complete what was requested. Do not stop until all files are created and the task is fully complete.'
-                    });
-                    continue;
-                }
-                else {
-                    // Later iteration without completion indicator - accept as final
-                    finalResponse = content.replace(/<think>[\s\S]*?<\/think>/gi, '').trim();
-                    break;
+                // Remove <think>...</think> tags from response (some models include thinking)
+                finalResponse = content.replace(/<think>[\s\S]*?<\/think>/gi, '').trim();
+                // Check if we're using task planning and there are more tasks
+                if (taskPlan && taskPlan.tasks.some(t => t.status === 'pending')) {
+                    const nextTask = getNextTask(taskPlan.tasks);
+                    if (nextTask) {
+                        // Move to next task
+                        nextTask.status = 'in_progress';
+                        opts.onTaskUpdate?.(nextTask);
+                        messages.push({ role: 'assistant', content: finalResponse });
+                        messages.push({
+                            role: 'user',
+                            content: `Good progress! Next task:\n\n${nextTask.id}. ${nextTask.description}\n\nComplete this task now.`
+                        });
+                        // Mark previous task as completed
+                        const prevTask = taskPlan.tasks.find(t => t.status === 'in_progress' && t.id !== nextTask.id);
+                        if (prevTask) {
+                            prevTask.status = 'completed';
+                        }
+                        finalResponse = ''; // Reset for next task
+                        continue;
+                    }
                 }
+                break;
             }
             // Add assistant response to history
             messages.push({ role: 'assistant', content });
@@ -548,12 +539,9 @@ export async function runAgent(prompt, projectContext, options = {}) {
                 }
             }
             // Add tool results to messages
-            const nextStepPrompt = iteration < 5
-                ? `Tool results:\n\n${toolResults.join('\n\n')}\n\nGood progress! Continue working on the task. Use more tools to complete what was requested. Only stop when EVERYTHING is finished and working.`
-                : `Tool results:\n\n${toolResults.join('\n\n')}\n\nContinue with the task. If the task is fully complete, provide a final summary without any tool calls.`;
             messages.push({
                 role: 'user',
-                content: nextStepPrompt,
+                content: `Tool results:\n\n${toolResults.join('\n\n')}\n\nContinue with the task. If this subtask is complete, provide a summary without tool calls.`,
             });
         }
         // Check if we hit max iterations

package/dist/utils/taskPlanner.js

@@ -10,19 +10,21 @@ export async function planTasks(userPrompt, projectContext) {
     const systemPrompt = `You are a task planning expert. Break down user requests into clear, sequential subtasks.
 
 RULES:
-1. Create 3-15 subtasks maximum
-2. Each subtask should be specific and actionable
-3. Order tasks logically (dependencies first)
+1. Create 3-10 subtasks maximum (keep it focused)
+2. Each subtask should be specific and achievable in 2-5 tool calls
+3. Order tasks logically - one file/component per task
 4. Use simple, clear descriptions
-5. Respond ONLY with a JSON object, no other text
+5. For websites: separate HTML, CSS, JS into different tasks
+6. Respond ONLY with a JSON object, no other text
 
-Example response format:
+Example for "create a website":
 {
   "tasks": [
-    {"id": 1, "description": "Create project directory structure", "dependencies": []},
-    {"id": 2, "description": "Create index.html with basic structure", "dependencies": [1]},
-    {"id": 3, "description": "Create styles.css with base styles", "dependencies": [1]},
-    {"id": 4, "description": "Add navigation to all pages", "dependencies": [2, 3]}
+    {"id": 1, "description": "Create directory structure", "dependencies": []},
+    {"id": 2, "description": "Create index.html with page structure", "dependencies": [1]},
+    {"id": 3, "description": "Create styles.css with layout and design", "dependencies": [1]},
+    {"id": 4, "description": "Create script.js with interactive features", "dependencies": [1]},
+    {"id": 5, "description": "Add content and finalize all pages", "dependencies": [2, 3, 4]}
   ]
 }
 
@@ -32,7 +34,7 @@ Project Context:
 
 User Request: ${userPrompt}
 
-Break this down into subtasks. Respond with JSON only.`;
+Break this down into subtasks. Each task = one file or one logical unit. Respond with JSON only.`;
     try {
         const apiKey = await getApiKey();
         if (!apiKey) {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "codeep",
-  "version": "1.0.8",
+  "version": "1.0.9",
   "description": "AI-powered coding assistant built for the terminal. Multiple LLM providers, project-aware context, and a seamless development workflow.",
   "type": "module",
   "main": "dist/index.js",