nex-code 0.3.5 → 0.3.7

package/cli/agent.js

@@ -1,835 +0,0 @@
-/**
- * cli/agent.js — Agentic Loop + Conversation State
- * Hybrid: chat + tool-use in a single conversation.
- */
-
-const { C, Spinner, TaskProgress, formatToolCall, formatResult, formatToolSummary, setActiveTaskProgress } = require('./ui');
-const { callStream } = require('./providers/registry');
-const { parseToolArgs } = require('./ollama');
-const { TOOL_DEFINITIONS, executeTool } = require('./tools');
-const { gatherProjectContext } = require('./context');
-const { fitToContext, forceCompress, getUsage } = require('./context-engine');
-const { autoSave } = require('./session');
-const { getMemoryContext } = require('./memory');
-const { checkPermission, setPermission, savePermissions } = require('./permissions');
-const { confirm, setAllowAlwaysHandler } = require('./safety');
-const { isPlanMode, getPlanModePrompt } = require('./planner');
-const { StreamRenderer } = require('./render');
-const { runHooks } = require('./hooks');
-const { routeMCPCall, getMCPToolDefinitions } = require('./mcp');
-const { getSkillInstructions, getSkillToolDefinitions, routeSkillCall } = require('./skills');
-const { trackUsage } = require('./costs');
-const { validateToolArgs } = require('./tool-validator');
-const { filterToolsForModel, getModelTier, PROVIDER_DEFAULT_TIER } = require('./tool-tiers');
-const { getConfiguredProviders } = require('./providers/registry');
-
-let MAX_ITERATIONS = 50;
-function setMaxIterations(n) { if (Number.isFinite(n) && n > 0) MAX_ITERATIONS = n; }
-
-// Abort signal getter — set by cli/index.js to avoid circular dependency
-let _getAbortSignal = () => null;
-function setAbortSignalGetter(fn) { _getAbortSignal = fn; }
-
-// Tools that can safely run in parallel (read-only, no side effects)
-const PARALLEL_SAFE = new Set([
-  'read_file', 'list_directory', 'search_files', 'glob', 'grep',
-  'web_fetch', 'web_search', 'git_status', 'git_diff', 'git_log',
-]);
-const MAX_RATE_LIMIT_RETRIES = 5;
-const MAX_NETWORK_RETRIES = 3;
-const MAX_STALE_RETRIES = 2;
-const STALE_WARN_MS = 60000;   // Warn after 60s without tokens
-const STALE_ABORT_MS = 120000; // Abort after 120s without tokens
-const CWD = process.cwd();
-
-// Wire up "a" (always allow) from confirm dialog → permission system
-setAllowAlwaysHandler((toolName) => {
-  setPermission(toolName, 'allow');
-  savePermissions();
-  console.log(`${C.green}  ✓ ${toolName}: always allow${C.reset}`);
-});
-
-// ─── Tool Call Helpers ────────────────────────────────────────
-
-/**
- * Prepare a tool call: parse args, validate, check permissions.
- * Returns an object ready for execution.
- */
-async function prepareToolCall(tc) {
-  const fnName = tc.function.name;
-  const args = parseToolArgs(tc.function.arguments);
-  const callId = tc.id || `cli-${Date.now()}-${Math.random().toString(36).slice(2, 6)}`;
-
-  // Malformed args
-  if (!args) {
-    const allToolDefs = [...TOOL_DEFINITIONS, ...getSkillToolDefinitions(), ...getMCPToolDefinitions()];
-    const toolDef = allToolDefs.find(t => t.function.name === fnName);
-    const schema = toolDef ? JSON.stringify(toolDef.function.parameters, null, 2) : 'unknown';
-    console.log(`${C.yellow}  ⚠ ${fnName}: malformed arguments, sending schema hint${C.reset}`);
-    return {
-      callId, fnName, args: null, canExecute: false,
-      errorResult: {
-        role: 'tool',
-        content: `ERROR: Malformed tool arguments. Could not parse your arguments as JSON.\n` +
-          `Raw input: ${typeof tc.function.arguments === 'string' ? tc.function.arguments.substring(0, 200) : 'N/A'}\n\n` +
-          `Expected JSON schema for "${fnName}":\n${schema}\n\n` +
-          `Please retry the tool call with valid JSON arguments matching this schema.`,
-        tool_call_id: callId,
-      },
-    };
-  }
-
-  // Validate
-  const validation = validateToolArgs(fnName, args);
-  if (!validation.valid) {
-    console.log(`${C.yellow}  ⚠ ${fnName}: ${validation.error.split('\n')[0]}${C.reset}`);
-    return {
-      callId, fnName, args, canExecute: false,
-      errorResult: { role: 'tool', content: validation.error, tool_call_id: callId },
-    };
-  }
-
-  const finalArgs = validation.corrected || args;
-
-  // Log validator corrections so user/LLM can see auto-fixes
-  if (validation.corrected) {
-    const orig = Object.keys(args);
-    const fixed = Object.keys(validation.corrected);
-    const renamed = orig.filter(k => !fixed.includes(k));
-    if (renamed.length) {
-      console.log(`${C.dim}  ✓ ${fnName}: corrected args (${renamed.join(', ')})${C.reset}`);
-    }
-  }
-
-  // Permission check
-  const perm = checkPermission(fnName);
-  if (perm === 'deny') {
-    console.log(`${C.red}  ✗ ${fnName}: denied by permissions${C.reset}`);
-    return {
-      callId, fnName, args: finalArgs, canExecute: false,
-      errorResult: { role: 'tool', content: `DENIED: Tool '${fnName}' is blocked by permissions`, tool_call_id: callId },
-    };
-  }
-  if (perm === 'ask') {
-    const ok = await confirm(`  Allow ${fnName}?`, { toolName: fnName });
-    if (!ok) {
-      return {
-        callId, fnName, args: finalArgs, canExecute: false,
-        errorResult: { role: 'tool', content: `CANCELLED: User declined ${fnName}`, tool_call_id: callId },
-      };
-    }
-  }
-
-  return { callId, fnName, args: finalArgs, canExecute: true, errorResult: null };
-}
-
-/**
- * Execute a single prepared tool call through the routing chain.
- */
-async function executeToolRouted(fnName, args, options = {}) {
-  const skillResult = await routeSkillCall(fnName, args);
-  if (skillResult !== null) return skillResult;
-  const mcpResult = await routeMCPCall(fnName, args);
-  if (mcpResult !== null) return mcpResult;
-  return executeTool(fnName, args, options);
-}
-
-/**
- * Short arg preview for spinner labels.
- */
-function _argPreview(name, args) {
-  switch (name) {
-    case 'read_file': case 'write_file': case 'edit_file':
-    case 'patch_file': case 'list_directory':
-      return args.path || '';
-    case 'bash':
-      return (args.command || '').substring(0, 60);
-    case 'grep': case 'search_files': case 'glob':
-      return args.pattern || '';
-    case 'web_fetch':
-      return (args.url || '').substring(0, 50);
-    case 'web_search':
-      return (args.query || '').substring(0, 40);
-    default:
-      return '';
-  }
-}
-
-/**
- * Execute a single prepared tool and return { msg, summary }.
- * @param {boolean} quiet - suppress formatToolCall/formatResult output
- */
-async function executeSingleTool(prep, quiet = false) {
-  if (!quiet) {
-    console.log(formatToolCall(prep.fnName, prep.args));
-  }
-
-  runHooks('pre-tool', { tool_name: prep.fnName });
-
-  const toolResult = await executeToolRouted(prep.fnName, prep.args, { silent: true });
-  const safeResult = String(toolResult ?? '');
-  const truncated = safeResult.length > 50000
-    ? safeResult.substring(0, 50000) + `\n...(truncated ${safeResult.length - 50000} chars)`
-    : safeResult;
-
-  if (!quiet) {
-    console.log(formatResult(truncated));
-  }
-
-  runHooks('post-tool', { tool_name: prep.fnName });
-
-  const firstLine = truncated.split('\n')[0];
-  const isError = firstLine.startsWith('ERROR') || firstLine.includes('CANCELLED') || firstLine.includes('BLOCKED')
-    || (prep.fnName === 'spawn_agents' && !/✓ Agent/.test(truncated) && /✗ Agent/.test(truncated));
-  const summary = formatToolSummary(prep.fnName, prep.args, truncated, isError);
-  const msg = { role: 'tool', content: truncated, tool_call_id: prep.callId };
-  return { msg, summary };
-}
-
-/**
- * Execute prepared tool calls with parallel batching.
- * Consecutive PARALLEL_SAFE tools run via Promise.all.
- * @param {boolean} quiet - use spinner + compact summaries instead of verbose output
- */
-async function executeBatch(prepared, quiet = false, options = {}) {
-  const results = new Array(prepared.length);
-  const summaries = [];
-  let batch = [];
-
-  // Quiet mode: show a single spinner for all tools
-  let spinner = null;
-  if (quiet && !options.skipSpinner) {
-    const execTools = prepared.filter(p => p.canExecute);
-    if (execTools.length > 0) {
-      let label;
-      if (execTools.length === 1) {
-        const p = execTools[0];
-        label = `▸ ${p.fnName} ${_argPreview(p.fnName, p.args)}`;
-      } else {
-        const names = execTools.map(p => p.fnName).join(', ');
-        label = `▸ ${execTools.length} tools: ${names.length > 60 ? names.substring(0, 57) + '...' : names}`;
-      }
-      spinner = new Spinner(label);
-      spinner.start();
-    }
-  }
-
-  async function flushBatch() {
-    if (batch.length === 0) return;
-    if (batch.length === 1) {
-      const idx = batch[0];
-      const { msg, summary } = await executeSingleTool(prepared[idx], quiet);
-      results[idx] = msg;
-      summaries.push(summary);
-    } else {
-      const promises = batch.map(idx => executeSingleTool(prepared[idx], quiet));
-      const batchResults = await Promise.all(promises);
-      for (let j = 0; j < batch.length; j++) {
-        results[batch[j]] = batchResults[j].msg;
-        summaries.push(batchResults[j].summary);
-      }
-    }
-    batch = [];
-  }
-
-  for (let i = 0; i < prepared.length; i++) {
-    const prep = prepared[i];
-
-    if (!prep.canExecute) {
-      await flushBatch();
-      results[i] = prep.errorResult;
-      summaries.push(formatToolSummary(prep.fnName, prep.args || {}, prep.errorResult.content, true));
-      continue;
-    }
-
-    if (PARALLEL_SAFE.has(prep.fnName)) {
-      batch.push(i);
-    } else {
-      await flushBatch();
-      // spawn_agents manages its own display (MultiProgress) — stop outer spinner
-      if (prep.fnName === 'spawn_agents' && spinner) {
-        spinner.stop();
-        spinner = null;
-      }
-      const { msg, summary } = await executeSingleTool(prep, quiet);
-      results[i] = msg;
-      summaries.push(summary);
-    }
-  }
-
-  await flushBatch();
-
-  // Stop spinner and print compact summaries
-  if (spinner) spinner.stop();
-  if (quiet && summaries.length > 0 && !options.skipSummaries) {
-    for (const s of summaries) console.log(s);
-  }
-
-  return results;
-}
-
-// Persistent conversation state
-let conversationMessages = [];
-
-/**
- * Build dynamic model routing guide for spawn_agents.
- * Only shown when 2+ models are available across configured providers.
- */
-function _buildModelRoutingGuide() {
-  try {
-    const configured = getConfiguredProviders();
-    const allModels = configured.flatMap(p =>
-      p.models.map(m => ({
-        spec: `${p.name}:${m.id}`,
-        tier: getModelTier(m.id, p.name),
-        name: m.name,
-      }))
-    );
-
-    if (allModels.length < 2) return '';
-
-    const tierLabels = {
-      full: 'complex tasks (refactor, implement, generate)',
-      standard: 'regular tasks (edit, fix, analyze)',
-      essential: 'simple tasks (read, search, list)',
-    };
-
-    let guide = '\n# Sub-Agent Model Routing\n\n';
-    guide += 'Sub-agents auto-select models by task complexity. Override with `model: "provider:model"` in agent definition.\n\n';
-    guide += '| Model | Tier | Auto-assigned for |\n|---|---|---|\n';
-    for (const m of allModels) {
-      guide += `| ${m.spec} | ${m.tier} | ${tierLabels[m.tier] || m.tier} |\n`;
-    }
-    return guide;
-  } catch {
-    return '';
-  }
-}
-
-function buildSystemPrompt() {
-  const projectContext = gatherProjectContext(CWD);
-
-  const memoryContext = getMemoryContext();
-  const skillInstructions = getSkillInstructions();
-  const planPrompt = isPlanMode() ? getPlanModePrompt() : '';
-
-  return `You are Nex Code, an expert coding assistant. You help with programming tasks by reading, writing, and editing files, running commands, and answering questions.
-
-WORKING DIRECTORY: ${CWD}
-All relative paths resolve from this directory.
-
-PROJECT CONTEXT:
-${projectContext}
-${memoryContext ? `\n${memoryContext}\n` : ''}${skillInstructions ? `\n${skillInstructions}\n` : ''}${planPrompt ? `\n${planPrompt}\n` : ''}
-# Core Behavior
-
-- You can use tools OR respond with text. For simple questions, answer directly.
-- For coding tasks, use tools to read files, make changes, run tests, etc.
-- Be concise but complete. Keep responses focused while ensuring the user gets the information they asked for.
-- When referencing code, include file:line (e.g. src/app.js:42) so the user can navigate.
-- Do not make up file paths or URLs. Use tools to discover them.
-
-# Response Quality (Critical)
-
-⚠ CRITICAL: The user CANNOT see tool output. They only see your text + 1-line summaries like "✓ bash ssh ... → ok".
-If you run tools but write NO text → the user sees NOTHING useful. This is the #1 quality failure.
-
-MANDATORY RULE: After ANY tool call that gathers information (bash, read_file, grep, ssh commands, etc.), you MUST write a text response summarizing the findings. NEVER end your response with only tool calls and no text.
-
-- Use markdown formatting: **bold** for key points, headers for sections, bullet lists for multiple items, \`code\` for identifiers. The terminal renders markdown with syntax highlighting.
-- Structure longer responses with headers (## Section) so the user can scan quickly.
-
-Response patterns by request type:
-- **Questions / analysis / "status" / "explain" / "what is"**: Gather data with tools, then respond with a clear, structured summary. NEVER just run tools and stop.
-- **Coding tasks (implement, fix, refactor)**: Brief confirmation of what you'll do, then use tools. After changes, summarize what you did and any important details.
-- **Simple questions ("what does X do?")**: Answer directly without tools when you have enough context.
-- **Ambiguous requests**: When a request is vague or could be interpreted multiple ways (e.g. "optimize this", "improve performance", "fix the issues", "refactor this"), ALWAYS ask clarifying questions first using ask_user. Do NOT guess scope or intent. Ask about: which specific area, what the expected outcome is, any constraints. Only proceed after the user clarifies.
-- **Server/SSH commands**: After running remote commands, ALWAYS present the results: service status, log errors, findings.
-
-After completing multi-step tasks, suggest logical next steps (e.g. "You can run npm test to verify" or "Consider committing with /commit").
-
-# Doing Tasks
-
-- For non-trivial tasks, briefly state your approach before starting (1 sentence). This helps the user know what to expect.
-- ALWAYS read code before modifying it. Never propose changes to code you haven't read.
-- Prefer edit_file for targeted changes over write_file for full rewrites.
-- Do not create new files unless absolutely necessary. Edit existing files instead.
-- Use relative paths when possible.
-- When blocked, try alternative approaches rather than retrying the same thing.
-- Keep solutions simple. Only change what's directly requested or clearly necessary.
-  - Don't add features, refactoring, or "improvements" beyond what was asked.
-  - Don't add error handling for impossible scenarios. Only validate at system boundaries.
-  - Don't add docstrings/comments to code you didn't change.
-  - Don't create helpers or abstractions for one-time operations.
-  - Three similar lines of code is better than a premature abstraction.
-- After completing work, give a brief summary of what was done and any important details. Don't just silently finish.
-
-# Tool Strategy
-
-- Use the RIGHT tool for the job:
-  - read_file to read files (not bash cat/head/tail)
-  - edit_file or patch_file to modify files (not bash sed/awk)
-  - glob to find files by name pattern (not bash find/ls)
-  - grep or search_files to search file contents (not bash grep)
-  - list_directory for directory structure (not bash ls/tree)
-  - Only use bash for actual shell operations: running tests, installing packages, git commands, build tools.
-- Call multiple tools in parallel when they're independent (e.g. reading multiple files at once).
-- For complex tasks with 3+ steps, create a task list with task_list first.
-- Use spawn_agents for 2+ independent tasks that can run simultaneously.
-  - Good for: reading multiple files, analyzing separate modules.
-  - Bad for: tasks that depend on each other or modify the same file.
-  - Max 5 parallel agents.
-${_buildModelRoutingGuide()}
-
-# Edit Reliability (Critical)
-
-- edit_file's old_text must match the file content EXACTLY — including whitespace, indentation, and newlines.
-- Always read the file first (read_file) before editing to see the exact current content.
-- If old_text is not found, the edit fails. Common causes:
-  - Indentation mismatch (tabs vs spaces, wrong level)
-  - Invisible characters or trailing whitespace
-  - Content changed since last read — read again before retrying.
-- For multiple changes to the same file, prefer patch_file (single operation, atomic).
-- Never guess file content. Always read first, then edit with the exact text you saw.
-
-# Error Recovery
-
-When a tool call returns ERROR:
-- edit_file/patch_file "old_text not found": Read the file again with read_file. Compare your old_text with the actual content. The most common cause is stale content — the file changed since you last read it.
-- bash non-zero exit: Read the error output. Fix the root cause (missing dependency, wrong path, syntax error) rather than retrying the same command.
-- "File not found": Use glob or list_directory to find the correct path. Do not guess.
-- After 2 failed attempts at the same operation, stop and explain the issue to the user.
-
-# Git Workflow
-
-- Before committing, review changes with git_diff. Write messages that explain WHY, not WHAT.
-- Stage specific files rather than git add -A to avoid committing unrelated changes.
-- Use conventional commits: type(scope): description (feat, fix, refactor, docs, test, chore).
-- Branch naming: feat/, fix/, refactor/, docs/ prefixes with kebab-case.
-- NEVER force-push, skip hooks (--no-verify), or amend published commits without explicit permission.
-- When asked to commit: review diff, propose message, wait for approval, then execute.
-
-# Safety & Reversibility
-
-- Consider reversibility before acting. File reads and searches are safe. File writes and bash commands may not be.
-- For hard-to-reverse actions (deleting files, force-pushing, dropping tables), confirm with the user first.
-- NEVER read .env files, credentials, or SSH keys.
-- NEVER run destructive commands (rm -rf /, mkfs, dd, etc.).
-- Dangerous commands (git push, npm publish, sudo, rm -rf) require user confirmation.
-- Prefer creating new git commits over amending. Never force-push without explicit permission.
-- If you encounter unexpected state (unfamiliar files, branches), investigate before modifying.`;
-}
-
-function clearConversation() {
-  conversationMessages = [];
-}
-
-function getConversationLength() {
-  return conversationMessages.length;
-}
-
-function getConversationMessages() {
-  return conversationMessages;
-}
-
-function setConversationMessages(messages) {
-  conversationMessages = messages;
-}
-
-/**
- * Print résumé + follow-up suggestions after the agent loop.
- * Only shown for multi-step responses (totalSteps >= 1).
- */
-function _printResume(totalSteps, toolCounts, filesModified, filesRead, startTime) {
-  if (totalSteps < 1) return;
-
-  const totalTools = [...toolCounts.values()].reduce((a, b) => a + b, 0);
-  let resume = `── ${totalSteps} ${totalSteps === 1 ? 'step' : 'steps'} · ${totalTools} ${totalTools === 1 ? 'tool' : 'tools'}`;
-  if (filesModified.size > 0) {
-    resume += ` · ${filesModified.size} ${filesModified.size === 1 ? 'file' : 'files'} modified`;
-  }
-  if (startTime) {
-    const elapsed = Date.now() - startTime;
-    const secs = Math.round(elapsed / 1000);
-    resume += secs >= 60 ? ` · ${Math.floor(secs / 60)}m ${secs % 60}s` : ` · ${secs}s`;
-  }
-  resume += ' ──';
-  console.log(`\n${C.dim}  ${resume}${C.reset}`);
-
-  // Follow-up suggestions based on what happened
-  if (filesModified.size > 0) {
-    console.log(`${C.dim}  💡 /diff · /commit · /undo${C.reset}`);
-  } else if (filesRead.size > 0 && totalSteps >= 2) {
-    console.log(`${C.dim}  💡 /save · /clear${C.reset}`);
-  }
-}
-
-/**
- * Process a single user input through the agentic loop.
- * Maintains conversation state across calls.
- */
-async function processInput(userInput) {
-  conversationMessages.push({ role: 'user', content: userInput });
-
-  const { setOnChange } = require('./tasks');
-  let taskProgress = null;
-  let cumulativeTokens = 0;
-
-  // Wire task onChange to create/update live task display
-  setOnChange((event, data) => {
-    if (event === 'create') {
-      if (taskProgress) taskProgress.stop();
-      taskProgress = new TaskProgress(data.name, data.tasks);
-      taskProgress.setStats({ tokens: cumulativeTokens });
-      taskProgress.start();
-    } else if (event === 'update' && taskProgress) {
-      taskProgress.updateTask(data.id, data.status);
-    } else if (event === 'clear') {
-      if (taskProgress) {
-        taskProgress.stop();
-        taskProgress = null;
-      }
-    }
-  });
-
-  const systemPrompt = buildSystemPrompt();
-  const fullMessages = [{ role: 'system', content: systemPrompt }, ...conversationMessages];
-
-  // Pre-spinner: visible activity during fitToContext + getUsage (can take 50–5000ms with LLM compacting)
-  const preSpinner = new Spinner('Thinking...');
-  preSpinner.start();
-
-  // Context-aware compression: fit messages into context window
-  const { messages: fittedMessages, compressed, compacted, tokensRemoved } = await fitToContext(
-    fullMessages,
-    TOOL_DEFINITIONS
-  );
-
-  // Context budget warning
-  const usage = getUsage(fullMessages, TOOL_DEFINITIONS);
-
-  preSpinner.stop();
-
-  if (compacted) {
-    console.log(`${C.dim}  [context compacted — summary (~${tokensRemoved} tokens freed)]${C.reset}`);
-  } else if (compressed) {
-    const pct = usage.limit > 0 ? Math.round((tokensRemoved / usage.limit) * 100) : 0;
-    console.log(`${C.dim}  [context compressed — ~${tokensRemoved} tokens freed (${pct}%)]${C.reset}`);
-  }
-  if (usage.percentage > 85) {
-    console.log(`${C.yellow}  ⚠ Context ${Math.round(usage.percentage)}% full — consider /clear or /save + start fresh${C.reset}`);
-  }
-
-  // Use fitted messages for the API call, but keep fullMessages reference for appending
-  let apiMessages = fittedMessages;
-  let rateLimitRetries = 0;
-  let networkRetries = 0;
-  let staleRetries = 0;
-  let contextRetries = 0;
-
-  // ─── Stats tracking for résumé ───
-  let totalSteps = 0;
-  const toolCounts = new Map();
-  const filesModified = new Set();
-  const filesRead = new Set();
-  const startTime = Date.now();
-
-  let i;
-  for (i = 0; i < MAX_ITERATIONS; i++) {
-    // Check if aborted (Ctrl+C) at start of each iteration
-    const loopSignal = _getAbortSignal();
-    if (loopSignal?.aborted) break;
-
-    // Step indicator — deferred, only shown for tool iterations (matches résumé count)
-    let stepPrinted = true; // default: no marker (text-only iterations stay silent)
-    function printStepIfNeeded() {
-      if (!stepPrinted) {
-        console.log(`${C.dim}  ── step ${totalSteps} ──${C.reset}`);
-        stepPrinted = true;
-      }
-    }
-
-    let spinner = null;
-    if (taskProgress && taskProgress.isActive()) {
-      // Resume the live task display instead of a plain spinner
-      if (taskProgress._paused) taskProgress.resume();
-    } else if (!taskProgress) {
-      const spinnerText = totalSteps > 0 ? `Thinking... (step ${totalSteps + 1})` : 'Thinking...';
-      spinner = new Spinner(spinnerText);
-      spinner.start();
-    }
-    let firstToken = true;
-    let streamedText = '';
-    const stream = new StreamRenderer();
-
-    let result;
-    // Stale-stream detection: warn/abort if provider stops sending tokens
-    let lastTokenTime = Date.now();
-    let staleWarned = false;
-    const staleAbort = new AbortController();
-    const staleTimer = setInterval(() => {
-      const elapsed = Date.now() - lastTokenTime;
-      if (elapsed >= STALE_ABORT_MS) {
-        stream._clearCursorLine();
-        console.log(`${C.yellow}  ⚠ Stream stale for ${Math.round(elapsed / 1000)}s — aborting and retrying${C.reset}`);
-        staleAbort.abort();
-      } else if (elapsed >= STALE_WARN_MS && !staleWarned) {
-        staleWarned = true;
-        stream._clearCursorLine();
-        console.log(`${C.yellow}  ⚠ No tokens received for ${Math.round(elapsed / 1000)}s — waiting...${C.reset}`);
-      }
-    }, 5000);
-    try {
-      const allTools = filterToolsForModel([...TOOL_DEFINITIONS, ...getSkillToolDefinitions(), ...getMCPToolDefinitions()]);
-      const userSignal = _getAbortSignal();
-      // Combine user abort (Ctrl+C) and stale abort into one signal
-      const combinedAbort = new AbortController();
-      if (userSignal) userSignal.addEventListener('abort', () => combinedAbort.abort(), { once: true });
-      staleAbort.signal.addEventListener('abort', () => combinedAbort.abort(), { once: true });
-
-      result = await callStream(apiMessages, allTools, {
-        signal: combinedAbort.signal,
-        onToken: (text) => {
-          lastTokenTime = Date.now();
-          staleWarned = false;
-          if (firstToken) {
-            if (taskProgress && !taskProgress._paused) {
-              taskProgress.pause();
-            } else if (spinner) {
-              spinner.stop();
-            }
-            printStepIfNeeded();
-            stream.startCursor();
-            firstToken = false;
-          }
-          streamedText += text;
-          stream.push(text);
-        },
-      });
-    } catch (err) {
-      clearInterval(staleTimer);
-      if (taskProgress && !taskProgress._paused) taskProgress.pause();
-      if (spinner) spinner.stop();
-      stream.stopCursor();
-
-      // Stale abort → retry this iteration (with limit)
-      if (staleAbort.signal.aborted && !_getAbortSignal()?.aborted) {
-        staleRetries++;
-        if (staleRetries > MAX_STALE_RETRIES) {
-          console.log(`${C.red}  ✗ Stream stale: max retries (${MAX_STALE_RETRIES}) exceeded. The model may be overloaded — try again or switch models.${C.reset}`);
-          if (taskProgress) { taskProgress.stop(); taskProgress = null; }
-          setOnChange(null);
-          _printResume(totalSteps, toolCounts, filesModified, filesRead, startTime);
-          autoSave(conversationMessages);
-          break;
-        }
-        i--; // Don't count stale timeouts as iterations
-        continue;
-      }
-
-      // Abort errors (Ctrl+C) — break silently
-      if (err.name === 'AbortError' || err.name === 'CanceledError' ||
-          err.message?.includes('canceled') || err.message?.includes('aborted')) {
-        if (taskProgress) { taskProgress.stop(); taskProgress = null; }
-        setOnChange(null);
-        _printResume(totalSteps, toolCounts, filesModified, filesRead, startTime);
-        autoSave(conversationMessages);
-        break;
-      }
-
-      // User-friendly error message (avoid raw stack traces/cryptic codes)
-      let userMessage = err.message;
-      if (err.code === 'ECONNREFUSED' || err.message.includes('ECONNREFUSED')) {
-        userMessage = 'Connection refused — please check your internet connection or API endpoint';
-      } else if (err.code === 'ENOTFOUND' || err.message.includes('ENOTFOUND')) {
-        userMessage = 'Network error — could not reach the API server. Please check your connection';
-      } else if (err.code === 'ETIMEDOUT' || err.message.includes('timeout')) {
-        userMessage = 'Request timed out — the API server took too long to respond. Please try again';
-      } else if (err.message.includes('401') || err.message.includes('Unauthorized')) {
-        userMessage = 'Authentication failed — please check your API key in the .env file';
-      } else if (err.message.includes('403') || err.message.includes('Forbidden')) {
-        userMessage = 'Access denied — your API key may not have permission for this model';
-      } else if (err.message.includes('400')) {
-        // Check if this is a context-too-long error before generic 400 handling
-        const errLower = (err.message || '').toLowerCase();
-        const isContextTooLong = errLower.includes('context') || errLower.includes('token') ||
-          errLower.includes('length') || errLower.includes('too long') || errLower.includes('too many');
-        if (isContextTooLong && contextRetries < 1) {
-          contextRetries++;
-          console.log(`${C.yellow}  ⚠ Context too long — force-compressing and retrying...${C.reset}`);
-          const allTools = [...TOOL_DEFINITIONS, ...getSkillToolDefinitions(), ...getMCPToolDefinitions()];
-          const { messages: compressedMsgs, tokensRemoved } = forceCompress(apiMessages, allTools);
-          apiMessages = compressedMsgs;
-          console.log(`${C.dim}  [force-compressed — ~${tokensRemoved} tokens freed]${C.reset}`);
-          i--;
-          continue;
-        }
-        if (isContextTooLong) {
-          userMessage = 'Context too long — force compression exhausted. Use /clear to start fresh';
-        } else {
-          userMessage = 'Bad request — the conversation may be too long or contain unsupported content. Try /clear and retry';
-        }
-      } else if (err.message.includes('500') || err.message.includes('502') || err.message.includes('503') || err.message.includes('504')) {
-        userMessage = 'API server error — the provider is experiencing issues. Please try again in a moment';
-      } else if (err.message.includes('fetch failed') || err.message.includes('fetch')) {
-        userMessage = 'Network request failed — please check your internet connection';
-      }
-      console.log(`${C.red}  ✗ ${userMessage}${C.reset}`);
-
-      if (err.message.includes('429')) {
-        rateLimitRetries++;
-        if (rateLimitRetries > MAX_RATE_LIMIT_RETRIES) {
-          console.log(`${C.red}  Rate limit: max retries (${MAX_RATE_LIMIT_RETRIES}) exceeded. Try again later or use /budget to check your limits.${C.reset}`);
-          if (taskProgress) { taskProgress.stop(); taskProgress = null; }
-          setOnChange(null);
-          _printResume(totalSteps, toolCounts, filesModified, filesRead, startTime);
-          autoSave(conversationMessages);
-          break;
-        }
-        const delay = Math.min(10000 * Math.pow(2, rateLimitRetries - 1), 120000);
-        const waitSpinner = new Spinner(`Rate limit — waiting ${Math.round(delay / 1000)}s (retry ${rateLimitRetries}/${MAX_RATE_LIMIT_RETRIES})`);
-        waitSpinner.start();
-        await new Promise((r) => setTimeout(r, delay));
-        waitSpinner.stop();
-        continue;
-      }
-
-      // Network/TLS errors — retry with backoff (don't burn iterations)
-      const isNetworkError = err.message.includes('socket disconnected') ||
-        err.message.includes('TLS') || err.message.includes('ECONNRESET') ||
-        err.message.includes('ECONNABORTED') || err.message.includes('ETIMEDOUT') ||
-        err.code === 'ECONNRESET' || err.code === 'ECONNABORTED';
-      if (isNetworkError) {
-        networkRetries++;
-        if (networkRetries > MAX_NETWORK_RETRIES) {
-          console.log(`${C.red}  Network error: max retries (${MAX_NETWORK_RETRIES}) exceeded. Check your connection and try again.${C.reset}`);
-          if (taskProgress) { taskProgress.stop(); taskProgress = null; }
-          setOnChange(null);
-          _printResume(totalSteps, toolCounts, filesModified, filesRead, startTime);
-          autoSave(conversationMessages);
-          break;
-        }
-        const delay = Math.min(2000 * Math.pow(2, networkRetries - 1), 30000);
-        const waitSpinner = new Spinner(`Network error — retrying in ${Math.round(delay / 1000)}s (${networkRetries}/${MAX_NETWORK_RETRIES})`);
-        waitSpinner.start();
-        await new Promise((r) => setTimeout(r, delay));
-        waitSpinner.stop();
-        i--; // Don't count network errors as iterations
-        continue;
-      }
-
-      // Auto-save on error so conversation isn't lost
-      if (taskProgress) { taskProgress.stop(); taskProgress = null; }
-      setOnChange(null);
-      _printResume(totalSteps, toolCounts, filesModified, filesRead, startTime);
-      autoSave(conversationMessages);
-      break;
-    }
-
-    clearInterval(staleTimer);
-
-    if (firstToken) {
-      if (taskProgress && !taskProgress._paused) taskProgress.pause();
-      if (spinner) spinner.stop();
-    }
-
-    // Reset retry counters on success
-    networkRetries = 0;
-    staleRetries = 0;
-
-    // Flush remaining stream buffer
-    if (streamedText) {
-      stream.flush();
-    }
-
-    // Track token usage for cost dashboard
-    if (result && result.usage) {
-      const { getActiveProviderName, getActiveModelId } = require('./providers/registry');
-      trackUsage(
-        getActiveProviderName(),
-        getActiveModelId(),
-        result.usage.prompt_tokens || 0,
-        result.usage.completion_tokens || 0
-      );
-      cumulativeTokens += (result.usage.prompt_tokens || 0) + (result.usage.completion_tokens || 0);
-      if (taskProgress) taskProgress.setStats({ tokens: cumulativeTokens });
-    }
-
-    const { content, tool_calls } = result;
-
-    // Build assistant message for history
-    const assistantMsg = { role: 'assistant', content: content || '' };
-    if (tool_calls && tool_calls.length > 0) {
-      assistantMsg.tool_calls = tool_calls;
-    }
-    conversationMessages.push(assistantMsg);
-    apiMessages.push(assistantMsg);
-
-    // No tool calls → response complete (or nudge if empty after tools)
-    if (!tool_calls || tool_calls.length === 0) {
-      const hasText = (content || '').trim().length > 0 || streamedText.trim().length > 0;
-      // If we just ran tools but the LLM produced no text → nudge it to summarize
-      if (!hasText && totalSteps > 0 && i < MAX_ITERATIONS - 1) {
-        const nudge = { role: 'user', content: '[SYSTEM] You ran tools but produced no visible output. The user CANNOT see tool results — only your text. Please summarize your findings now.' };
-        apiMessages.push(nudge);
-        continue; // retry — don't count as a new step
-      }
-      if (taskProgress) { taskProgress.stop(); taskProgress = null; }
-      setOnChange(null);
-      _printResume(totalSteps, toolCounts, filesModified, filesRead, startTime);
-      autoSave(conversationMessages);
-      return;
-    }
-
-    // ─── Update stats ───
-    totalSteps++;
-    if (totalSteps > 1) stepPrinted = false; // enable deferred step marker for steps 2+
-    for (const tc of tool_calls) {
-      const name = tc.function.name;
-      toolCounts.set(name, (toolCounts.get(name) || 0) + 1);
-    }
-
-    // ─── Prepare all tool calls (parse, validate, permissions — sequential) ───
-    const prepared = [];
-    for (const tc of tool_calls) {
-      prepared.push(await prepareToolCall(tc));
-    }
-
-    // ─── Execute with parallel batching (quiet mode: spinner + compact summaries) ───
-    const batchOpts = taskProgress ? { skipSpinner: true, skipSummaries: true } : {};
-    if (!batchOpts.skipSummaries) printStepIfNeeded();
-    // Resume TaskProgress animation during tool execution so the UI never looks frozen
-    if (taskProgress && taskProgress._paused) taskProgress.resume();
-    const toolMessages = await executeBatch(prepared, true, batchOpts);
-
-    // Track modified and read files
-    for (let j = 0; j < prepared.length; j++) {
-      const prep = prepared[j];
-      if (!prep.canExecute) continue;
-      const res = toolMessages[j].content;
-      const isOk = !res.startsWith('ERROR') && !res.includes('CANCELLED');
-      if (isOk && ['write_file', 'edit_file', 'patch_file'].includes(prep.fnName)) {
-        if (prep.args && prep.args.path) filesModified.add(prep.args.path);
-      }
-      if (isOk && prep.fnName === 'read_file') {
-        if (prep.args && prep.args.path) filesRead.add(prep.args.path);
-      }
-    }
-
-    for (const toolMsg of toolMessages) {
-      conversationMessages.push(toolMsg);
-      apiMessages.push(toolMsg);
-    }
-  }
-
-  // Only print résumé + max-iterations warning if the loop actually exhausted (not on break)
-  if (i >= MAX_ITERATIONS) {
-    if (taskProgress) { taskProgress.stop(); taskProgress = null; }
-    setOnChange(null);
-    _printResume(totalSteps, toolCounts, filesModified, filesRead, startTime);
-    autoSave(conversationMessages);
-    console.log(`\n${C.yellow}⚠ Max iterations (${MAX_ITERATIONS}) reached. Try ${C.bold}--max-turns ${MAX_ITERATIONS + 20}${C.reset}${C.yellow} or break into smaller steps.${C.reset}`);
-  }
-}
-
-module.exports = { processInput, clearConversation, getConversationLength, getConversationMessages, setConversationMessages, setAbortSignalGetter, setMaxIterations };