tuna-agent 0.1.80 → 0.1.81

package/dist/executor/task-runner.js

@@ -1,7 +1,6 @@
 import path from 'path';
 import os from 'os';
 import { spawn } from 'child_process';
-import { StringDecoder } from 'string_decoder';
 import { runClaude } from '../utils/claude-cli.js';
 import { validatePath } from '../utils/validate-path.js';
 const NEEDS_INPUT_MARKER = '"status":"NEEDS_INPUT"';
@@ -23,9 +22,6 @@ export async function runTask(task, onProgress, signal, confirmBeforeEdit) {
     if (confirmBeforeEdit) {
         args.push('--permission-mode', 'default');
     }
-    else {
-        args.push('--permission-mode', 'bypassPermissions');
-    }
     const env = {
         ...process.env,
         HOME: process.env.HOME || '',
@@ -83,9 +79,8 @@ export async function runTask(task, onProgress, signal, confirmBeforeEdit) {
         let stdout = '';
         let stderr = '';
         let buffer = '';
-        const stdoutDecoder = new StringDecoder('utf8');
         proc.stdout.on('data', (chunk) => {
-            const text = stdoutDecoder.write(chunk);
+            const text = chunk.toString();
             stdout += text;
             buffer += text;
             const lines = buffer.split('\n');
@@ -102,9 +97,8 @@ export async function runTask(task, onProgress, signal, confirmBeforeEdit) {
                 }
             }
         });
-        const stderrDecoder = new StringDecoder('utf8');
         proc.stderr.on('data', (chunk) => {
-            stderr += stderrDecoder.write(chunk);
+            stderr += chunk.toString();
         });
         proc.on('close', (code) => {
             clearTimeout(timeoutTimer);
@@ -431,15 +425,8 @@ export async function executeSubtask(subtask, repoPath, contracts, callbacks, si
         }
         else {
             // Fallback: detect if AI just wrote questions as text without using NEEDS_INPUT
-            // Only trigger for very short outputs (< 500 chars) that completed quickly (< 15s)
-            // This avoids false positives when a completed task output contains sentences with "?"
-            const isShortOutput = result.result.length < 500;
-            const isQuickRun = result.durationMs != null && result.durationMs < 15000;
-            const fallbackQuestion = (isShortOutput && isQuickRun)
-                ? parseQuestionFromOutput(result.result, subtask.id)
-                : null;
-            const seemsLikeQuestion = !!fallbackQuestion;
-            if (seemsLikeQuestion && fallbackQuestion && callbacks?.onSubtaskNeedsInput) {
+            const fallbackQuestion = parseQuestionFromOutput(result.result, subtask.id);
+            if (fallbackQuestion && callbacks?.onSubtaskNeedsInput) {
                 info.status = 'waiting_input';
                 info.pendingQuestion = fallbackQuestion;
                 log(info, 'thinking', `Fallback: detected question in output (no NEEDS_INPUT marker): ${fallbackQuestion.question}`);
@@ -490,6 +477,15 @@ export async function executeSubtask(subtask, repoPath, contracts, callbacks, si
                     log(info, 'thinking', 'No answer received — task paused');
                 }
             }
+            else if (looksIncomplete(result.result)) {
+                info.status = 'failed';
+                info.result = result.result;
+                log(info, 'error', `Detected incomplete output — marking as failed`);
+                callbacks?.onSubtaskLog?.(subtask.id, {
+                    type: 'error',
+                    message: `❌ Task did not complete: AI could not fulfill the request`,
+                });
+            }
             else {
                 info.status = 'done';
                 info.result = result.result;
@@ -586,6 +582,26 @@ export async function executeTaskWithPlan(task, plan, onProgress, callbacks, sig
     console.log(`[Executor] Total time: ${(totalTime / 1000).toFixed(1)}s`);
     return { sessions: allSessions, status: 'done' };
 }
+/**
+ * Detect if Claude's output indicates it could NOT complete the task.
+ * Checks the last ~500 chars for common failure/inability patterns.
+ */
+function looksIncomplete(output) {
+    // Check only the tail — that's where the conclusion lives
+    const tail = output.slice(-500).toLowerCase();
+    const patterns = [
+        /i (?:was |am )?(?:unable|not able) to/,
+        /(?:could|can)(?:n't|not) (?:find|locate|access|complete|proceed|continue)/,
+        /(?:doesn't|does not|don't|do not) (?:exist|have access)/,
+        /(?:no (?:such|matching) (?:file|directory|path|repo))/,
+        /(?:not found|file not found|directory not found)/,
+        /(?:i need (?:you to|more information|the .* path|access))/,
+        /(?:please (?:provide|specify|confirm|check))/,
+        /(?:unfortunately|i apologize).{0,50}(?:cannot|couldn't|unable|can't)/,
+        /(?:blocked|stuck).{0,30}(?:because|due to|cannot)/,
+    ];
+    return patterns.some((p) => p.test(tail));
+}
 /**
  * Parse question from Claude output when question.json is missing.
  * Fallback for when Claude outputs NEEDS_INPUT but didn't create the file.

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "tuna-agent",
-  "version": "0.1.80",
+  "version": "0.1.81",
   "description": "Tuna Agent - Run AI coding tasks on your machine",
   "bin": {
     "tuna-agent": "dist/cli/index.js"