vibehacker 4.1.0 → 4.2.0

package/src/approve.js

@@ -25,6 +25,7 @@ const NEEDS_APPROVAL = new Set([
 function describeToolCall(tc) {
   const { name, args } = tc;
   if (name === 'write_file')       return args.path || '(unknown path)';
+  if (name === 'edit_file')        return args.path || '(unknown path)';
   if (name === 'execute_command')  return args.command || '(unknown command)';
   if (name === 'delete_file')      return args.path || '(unknown path)';
   if (name === 'create_directory') return args.path || '(unknown path)';
@@ -77,40 +78,34 @@ function showApproval(screen, tc, context) {
 
     const W = Math.min(Math.max(40, screen.width - 6), 70);
 
-    // Fixed chrome: border(2) + blank(1) + header(1) + detail(1) + blank(1)
-    // + "Do you want to proceed?"(1) + blank(1) + buttons(items.length)
-    // + blank(1) + hint(1) = 10 + items.length
-    // Everything else (context line + preview) must fit in remaining space.
-    const maxH    = Math.max(10, screen.height - 4);
-    const fixedH  = 10 + items.length;
-    const budget  = Math.max(0, maxH - fixedH);
+    // Fixed chrome lines:
+    //   border-top(1) + blank(1) + header(1) + detail(1) + blank(1)
+    //   + "Do you want to proceed?"(1) + blank(1) + buttons(3)
+    //   + blank(1) + hint(1) + border-bottom(1) = 13
+    const CHROME = 13;
+    const maxH   = Math.max(CHROME, screen.height - 2);
 
     // Context eats 1 line if present
     const ctxLine = context ? `{#444444-fg}${esc(context)}{/#444444-fg}` : '';
     const ctxCost = ctxLine ? 1 : 0;
 
-    // Preview: clamp to remaining budget (separator + lines)
+    // Preview: hard-cap to 4 lines max — buttons MUST always be visible
     let preview = '';
     if (tc.name === 'write_file' && tc.args.content) {
       preview = tc.args.content;
     } else if (tc.name === 'execute_command') {
       preview = tc.args.command || '';
     }
-    const previewBudget = Math.max(0, budget - ctxCost - (preview ? 1 : 0)); // -1 for separator
     if (preview) {
       const allLines = preview.split('\n');
-      if (allLines.length > previewBudget) {
-        preview = allLines.slice(0, Math.max(1, previewBudget - 1)).join('\n') + '\n…';
-      }
-      // Also hard-cap to 6 lines regardless, to keep dialog compact
-      const cappedLines = preview.split('\n');
-      if (cappedLines.length > 6) {
-        preview = cappedLines.slice(0, 5).join('\n') + '\n…';
+      const maxPreviewLines = Math.min(4, Math.max(0, maxH - CHROME - ctxCost - 1));
+      if (allLines.length > maxPreviewLines) {
+        preview = allLines.slice(0, Math.max(1, maxPreviewLines - 1)).join('\n') + '\n…';
       }
     }
     const previewCost = preview ? preview.split('\n').length + 1 : 0; // +1 separator
 
-    const H = Math.min(maxH, fixedH + ctxCost + previewCost);
+    const H = Math.min(maxH, CHROME + ctxCost + previewCost);
 
     const box = blessed.box({
       parent:  screen,

package/src/providers.js

@@ -10,7 +10,7 @@ const PROVIDERS = {
     name:      'Vibe Hacker',
     shortName: 'Vibe Hacker',
     baseURL:   VH_GATEWAY_URL,
-    freeNote:  '100 req/day free',
+    freeNote:  '50 req/day free',
     tier:      'free',
     detectKey: k => k.startsWith('vh_'),
     getKey:    () => 'Sign up at https://vibsecurity.com',
@@ -29,7 +29,7 @@ const PROVIDERS = {
     name:      'Vibe Hacker',
     shortName: 'Vibe Hacker',
     baseURL:   'https://openrouter.ai/api/v1',
-    freeNote:  '100 req/day free',
+    freeNote:  '50 req/day free',
     tier:      'free',
     detectKey: k => k.startsWith('sk-or-v1-'),
     getKey:    () => 'vibsecurity.com/login',

package/src/src/agent.js

@@ -0,0 +1,311 @@
+'use strict';
+
+const fs = require('fs');
+const path = require('path');
+const { streamChat } = require('./api');
+const { parseToolCalls, executeTool, TOOL_DOCS } = require('./tools');
+const config = require('./config');
+
+// ── Modes ────────────────────────────────────────────────────────────────────
+
+const MODES = [
+  { id: 'chat', name: 'Chat', description: 'Security Q&A and threat intel' },
+  { id: 'hunt', name: 'Hunt', description: 'Autonomous coding, security ops & tool use' },
+];
+
+// ── Project Memory ───────────────────────────────────────────────────────────
+// Cached — only reads filesystem once per cwd, refreshes when cwd changes.
+
+let _memCache = { cwd: null, content: null };
+
+function loadProjectMemory(cwd) {
+  if (_memCache.cwd === cwd) return _memCache.content;
+
+  const candidates = [
+    path.join(cwd, 'VIBEHACKER.md'),
+    path.join(cwd, '.vibehacker', 'context.md'),
+    path.join(cwd, '.vibehacker', 'instructions.md'),
+  ];
+  let result = null;
+  for (const f of candidates) {
+    try {
+      const c = fs.readFileSync(f, 'utf8');
+      if (c.trim()) { result = { file: path.relative(cwd, f), content: c.trim() }; break; }
+    } catch (_) {}
+  }
+  _memCache = { cwd, content: result };
+  return result;
+}
+
+// ── System Prompt — Cached, Only Rebuilt When cwd/mode Changes ───────────────
+
+let _promptCache = { mode: null, cwd: null, prompt: null };
+
+function buildSystemPrompt(mode, cwd) {
+  if (_promptCache.mode === mode && _promptCache.cwd === cwd && _promptCache.prompt) {
+    return _promptCache.prompt;
+  }
+
+  const os    = process.platform === 'win32' ? 'Windows' : process.platform === 'darwin' ? 'macOS' : 'Linux';
+  const shell = process.platform === 'win32' ? 'powershell' : 'bash';
+  const date  = new Date().toISOString().split('T')[0];
+
+  const projectMem = loadProjectMemory(cwd);
+  const projectSection = projectMem
+    ? `\n\n# Project Instructions (${projectMem.file})\n${projectMem.content}\n`
+    : '';
+
+  let prompt;
+
+  if (mode === 'hunt') {
+    prompt = `You are Vibe Hacker v${config.version} — an expert autonomous AI agent.
+
+# Environment
+- CWD: ${cwd}
+- OS: ${os} | Shell: ${shell} | Date: ${date}
+${projectSection}
+# HUNT MODE — Autonomous Agent with Tool Access
+
+You have filesystem + shell access via XML tool blocks. DO the work — don't describe it.
+
+${TOOL_DOCS}
+
+# Rules (MANDATORY)
+
+1. USE TOOLS. Don't explain what you'd do — DO IT with tool calls.
+   ✗ "You could run npm install" → ✓ <execute_command><command>npm install</command></execute_command>
+
+2. READ BEFORE EDIT. Always read_file before edit_file. The tool rejects edits on unread files.
+
+3. EDIT > WRITE. Modify existing files with edit_file (surgical replacement). Only write_file for NEW files.
+
+4. EXACT MATCHING. edit_file old_string must match the file exactly — whitespace, indentation, everything.
+   If it fails: read the file again, the content changed. Add more surrounding context if not unique.
+
+5. MULTIPLE TOOLS OK. Use several tools in one response when they're independent.
+
+6. GREP > MANUAL SEARCH. Use grep/glob to find code. Don't read every file looking for something.
+
+7. NON-INTERACTIVE COMMANDS ONLY. No vim, nano, interactive prompts. Use -y/--yes flags. 2 min timeout.
+
+8. COMPLETE CODE. Never write "// ...", "// TODO", "// rest of code". Write the full implementation.
+
+9. VERIFY. After changes: read back files, run tests, fix issues. Don't declare done without checking.
+
+# Error Recovery
+
+- edit_file "not found" → Read file again. Check whitespace. Content may have changed.
+- edit_file "multiple matches" → Add more surrounding lines to old_string.
+- Command failed → Read error. Check dependencies. Try different approach.
+- File not found → Use glob/list_files to find correct path.
+
+# Workflow: EXPLORE → PLAN (1-2 sentences) → EXECUTE → VERIFY
+
+# Style: Direct. No filler. Brief explanations between tools. Summary when done.`;
+
+  } else {
+    prompt = `You are Vibe Hacker v${config.version} — expert AI for cybersecurity and engineering.
+
+# Environment
+- OS: ${os} | Date: ${date}
+${projectSection}
+# CHAT MODE — Expert Answers
+
+Direct, accurate, actionable. No filler. Markdown with language-tagged code blocks.
+For security: include attack vectors + mitigations. For code: working examples, not pseudocode.`;
+  }
+
+  _promptCache = { mode, cwd, prompt };
+  return prompt;
+}
+
+// ── Thinking Extraction ──────────────────────────────────────────────────────
+
+function extractThinking(text) {
+  let visible = text;
+  let thinking = '';
+  // Remove <think>...</think> and <thinking>...</thinking> blocks
+  visible = visible.replace(/<think(?:ing)?>([\s\S]*?)<\/think(?:ing)?>/g, (_, t) => {
+    thinking += t.trim() + '\n';
+    return '';
+  });
+  return { visible: visible.replace(/\n{3,}/g, '\n\n').trim(), thinking: thinking.trim() };
+}
+
+// ── Token Estimation (fast, cached per string length) ────────────────────────
+
+function estimateTokens(text) {
+  if (!text) return 0;
+  // Empirical: ~3.5 chars per token for mixed code/text
+  return Math.ceil(text.length / 3.5);
+}
+
+// ── Context Trimming — Proactive, Prioritized ────────────────────────────────
+
+function trimHistory(messages, maxContextTokens) {
+  // Fast total estimate
+  let total = 0;
+  for (let i = 0; i < messages.length; i++) total += estimateTokens(messages[i].content);
+
+  const budget = Math.floor(maxContextTokens * 0.55); // 45% headroom for response + tool results
+  if (total <= budget) return messages;
+
+  // Phase 1: Strip thinking blocks from old assistant messages
+  let trimmed = messages.map((m, i) => {
+    if (i === 0 || i >= messages.length - 6) return m; // keep system + recent
+    if (m.role === 'assistant' && (m.content.includes('<think>') || m.content.includes('<thinking>'))) {
+      const { visible } = extractThinking(m.content);
+      return { ...m, content: visible };
+    }
+    return m;
+  });
+
+  total = 0;
+  for (const m of trimmed) total += estimateTokens(m.content);
+  if (total <= budget) return trimmed;
+
+  // Phase 2: Compress old tool results to headers only
+  trimmed = trimmed.map((m, i) => {
+    if (i === 0 || i >= trimmed.length - 6) return m;
+    if (m.role === 'user' && m.content.startsWith('[Tool Result:') && m.content.length > 800) {
+      return { ...m, content: m.content.split('\n')[0] + '\n[output trimmed]' };
+    }
+    if (m.role === 'assistant' && m.content.length > 1500 && i < trimmed.length - 8) {
+      return { ...m, content: m.content.substring(0, 400) + '\n[...]\n' + m.content.slice(-300) };
+    }
+    return m;
+  });
+
+  total = 0;
+  for (const m of trimmed) total += estimateTokens(m.content);
+  if (total <= budget) return trimmed;
+
+  // Phase 3: Drop middle messages
+  const keep = Math.min(8, trimmed.length - 2);
+  if (trimmed.length <= keep + 2) return trimmed;
+  const dropped = trimmed.length - keep - 1;
+  return [
+    trimmed[0],
+    { role: 'user', content: `[${dropped} earlier messages trimmed for context]` },
+    ...trimmed.slice(-keep),
+  ];
+}
+
+// ── Agent ────────────────────────────────────────────────────────────────────
+
+class Agent {
+  constructor() {
+    this.history = [];
+    this.mode    = 'chat';
+    this.cwd     = process.cwd();
+  }
+
+  setMode(mode) { this.mode = mode; this.history = []; _promptCache.mode = null; }
+  setCwd(dir)   { this.cwd = dir; _promptCache.cwd = null; _memCache.cwd = null; }
+  clearHistory(){ this.history = []; }
+
+  async run({ userMessage, model, signal, onToken, onDone, onError, onToolCall, onToolResult, beforeToolCall }) {
+    this.history.push({ role: 'user', content: userMessage });
+
+    const maxCtx  = (model && model.contextWindow) || 32768;
+    const maxIter = config.maxToolIterations || 25;
+    let iterations = 0;
+
+    // ── Iterative agent loop ──────────────────────────────────────────
+    while (true) {
+      if (signal && signal.aborted) {
+        onError(Object.assign(new Error('aborted'), { type: 'ABORTED' }));
+        return;
+      }
+
+      iterations++;
+      if (iterations > maxIter) {
+        onDone('[Tool iteration limit reached. Use /retry to continue.]');
+        return;
+      }
+
+      // Build messages with proactive trimming (BEFORE sending)
+      const sysPrompt = buildSystemPrompt(this.mode, this.cwd);
+      let messages = [{ role: 'system', content: sysPrompt }, ...this.history];
+      messages = trimHistory(messages, maxCtx);
+
+      // ── Stream response ─────────────────────────────────────────────
+      let fullResponse = '';
+      let streamError  = null;
+
+      await new Promise((resolve) => {
+        let resolved = false;
+        const done = () => { if (!resolved) { resolved = true; resolve(); } };
+        streamChat({
+          messages, model: model.id, signal,
+          maxTokens: model.maxTokens || config.maxTokens || 8192,
+          onToken: (token, full) => { fullResponse = full; if (onToken) onToken(token, full); },
+          onDone:  (content) => { fullResponse = content || fullResponse; done(); },
+          onError: (errObj) => { streamError = errObj; done(); },
+        });
+      });
+
+      // Rate limit backoff inside tool loop — rotate-first strategy
+      if (streamError && streamError.type === 'RATE_LIMIT' && iterations > 1) {
+        // Don't retry same model. Surface error to app.js for provider rotation.
+        onError(Object.assign(new Error(streamError.msg || 'Rate limited'), streamError));
+        return;
+      }
+      if (streamError) {
+        onError(Object.assign(new Error(streamError.msg || 'error'), streamError));
+        return;
+      }
+      if (signal && signal.aborted) return;
+
+      // Store full response in history (including thinking for continuity)
+      this.history.push({ role: 'assistant', content: fullResponse });
+
+      // Chat mode — done after single response
+      if (this.mode !== 'hunt') {
+        const { visible } = extractThinking(fullResponse);
+        onDone(visible || fullResponse);
+        return;
+      }
+
+      // Hunt mode — parse tool calls
+      const toolCalls = parseToolCalls(fullResponse);
+      if (toolCalls.length === 0) {
+        const { visible } = extractThinking(fullResponse);
+        onDone(visible || fullResponse);
+        return;
+      }
+
+      // Execute all tools
+      for (const tc of toolCalls) {
+        if (signal && signal.aborted) return;
+
+        if (onToolCall) onToolCall(tc);
+
+        if (beforeToolCall) {
+          const decision = await beforeToolCall(tc);
+          if (decision === 'no') {
+            this.history.push({ role: 'user', content: `[Tool Denied: ${tc.name}] User rejected. Try a different approach.` });
+            continue;
+          }
+        }
+
+        let result;
+        const toolStart = Date.now();
+        try {
+          result = await executeTool(tc, this.cwd);
+        } catch (err) {
+          result = `[Error: ${tc.name}] ${err.message}`;
+        }
+
+        if (onToolResult) onToolResult(tc, result);
+        this.history.push({ role: 'user', content: `[Tool Result: ${tc.name}]\n${result}` });
+      }
+
+      // Adaptive throttle based on tool types — minimal delay
+      const hasWrite = toolCalls.some(tc => ['write_file', 'edit_file', 'execute_command', 'delete_file'].includes(tc.name));
+      await new Promise(r => setTimeout(r, hasWrite ? 300 : 100));
+    }
+  }
+}
+
+module.exports = { Agent, MODES };