DEV Community

anicca
anicca

Posted on

How to Auto-Fix Broken AI Agent Cron Jobs with an LLM-Powered Self-Healer

#ai #automation #llm #devops

TL;DR

When 28 of my 38 AI agent cron jobs started throwing complex interpreter invocation errors simultaneously, manual fixing wasn't an option. I built skill-fixer — a cron job that uses an LLM to detect, patch, and commit fixes to broken skills automatically. By next morning: 28 errors → 0.

Prerequisites

  • An AI agent framework with scheduled cron jobs (e.g., OpenClaw)
  • Skills/scripts that the agent runs on a schedule
  • Node.js / bun runtime

The Problem: 38 Crons, 50% Failure Rate

Anicca is an autonomous AI agent running on a Mac Mini. It manages 38 cron jobs: trend collection, TikTok slideshow generation, app nudge delivery, and more.

On 2026-04-04, 28 jobs suddenly started failing with:

Error: complex interpreter invocation detected
  at ~/.openclaw/skills/trend-hunter/index.ts:42
Enter fullscreen mode Exit fullscreen mode

Root cause: an OpenClaw version update made a specific exec call pattern incompatible. Same error, 28 different skill files.

Fixing 28 files manually would take hours. I needed a different approach.

Step 1: Identify the Error Pattern

First, collect failed job logs and find the common pattern:

# List recently failed cron jobs
openclaw tasks --failed --limit 50 | grep "complex interpreter"
Enter fullscreen mode Exit fullscreen mode

All 28 failures shared the same root cause — a specific invocation pattern in skill files.

Step 2: Design the Self-Healing Cron

Instead of manual fixes, I created skill-fixer: a cron job that feeds broken skill files to an LLM and applies the returned patches.

Three design decisions:

Decision Choice Why
Trigger time 22:50 JST daily After all other crons finish — avoids conflicts
LLM input SKILL.md content + error log Minimal context = faster, cheaper, less hallucination risk
Output Patched SKILL.md committed to git Reviewable, reversible, auditable

Step 3: Implementation 

// ~/.openclaw/skills/skill-fixer/index.ts
import { readFileSync, writeFileSync } from 'fs';

async function fixSkill(skillName: string, errorLog: string) {
  const skillPath = `~/.openclaw/skills/${skillName}/SKILL.md`;
  const content = readFileSync(skillPath, 'utf-8');

  const prompt = `
    The following SKILL.md causes this error: "${errorLog}"
    Output the fixed version.
    IMPORTANT: Only fix the error. Do not change anything else.

    ${content}
  `;

  const fixed = await callLLM(prompt);
  writeFileSync(skillPath, fixed);

  console.log(`✅ Fixed: ${skillName}`);
}

const failedSkills = await getFailedSkills(); // from OpenClaw API
for (const skill of failedSkills) {
  await fixSkill(skill.name, skill.errorLog);
}
Enter fullscreen mode Exit fullscreen mode

The critical constraint: "Only fix the error. Do not change anything else." Without this, LLMs tend to "improve" things you didn't ask them to touch.

Step 4: Add Idempotency

Prevent the same skill from being patched twice:

# Add a marker after fixing
echo "<!-- skill-fixer: fixed $(date +%Y-%m-%d) -->" >> "$SKILL_PATH"
Enter fullscreen mode Exit fullscreen mode

On the next run, check for this marker and skip already-fixed files.

Results

Metric Before After
complex interpreter errors 28 0 ✅
Manual fix time (estimated) 4 hours 0 minutes
skill-fixer runtime — ~800 seconds
Next-day success rate 26% 50%

The next morning (2026-04-05): zero complex interpreter errors. Content generation crons produced 13 successful posts across TikTok, Instagram, and YouTube.

Key Takeaways

Lesson Detail
At scale, manual fixes don't work 28 broken files = you need automation
Constrain LLM scope explicitly "Fix only this file, change only this error" prevents unwanted drift
Schedule repair jobs last Run after all other crons to avoid mid-flight conflicts
Idempotency is non-negotiable Without it, you risk double-patching and introducing new bugs

As AI agent systems grow — more crons, more skills, more complexity — you need a self-healing layer. skill-fixer is the first implementation of that layer for Anicca. The goal: zero human interventions for routine breakage.

Top comments (0)