Ralph Wiggum loop via crons feasibility
I have an autonomous code improvement system running on a VPS. I want you to evaluate whether this architecture is sound or if there are risks I'm missing.
Setup:
• Two cron jobs running 24/7 via OpenClaw (an AI agent orchestration platform)
• Both run as isolated sessions with free LLM models (Groq llama-3.3-70b, Cerebras gpt-oss-120b)
• Both deliver results to Telegram
Cron 1 — "Reviewer" (every 6 hours):
• Reads one Python source file from a list of 12 trading bot monitors
• Hunts for real bugs: crashes, unhandled exceptions, logic errors, values contradicting a config file (DECISIONS.json)
• Adds found bugs to a BACKLOG.json file with severity ratings
• Never modifies source code — read-only
• Rotates through files sequentially, loops back when all reviewed
Cron 2 — "Fixer" (every 2 hours):
• Picks the first unfixed bug from BACKLOG.json
• Uses sed/python one-liners to make surgical edits (never rewrites full files)
• Safety: backs up file before edit, checks line count after (must be within ±5), verifies Python import passes, auto-reverts from backup if any check fails
• Updates backlog and logs results
Rate limit protection:
• 3-model fallback chain (minimax → cerebras → groq) with exponential backoff cooldowns
• Jobs spaced well below API rate limits (16 requests/day vs 1000+/day allowed)
File safety:
• Git committed before loop started — full rollback available
• Backup-before-edit + auto-revert on failure
• No parallel agents editing the same file
My concerns:
1. Is the fixer prompt robust enough to prevent bad edits even with auto-revert?
2. Could the reviewer generate false positive bugs that the fixer then "fixes" incorrectly?
3. Am I missing any failure mode?
4. What would you change?
Setup:
• Two cron jobs running 24/7 via OpenClaw (an AI agent orchestration platform)
• Both run as isolated sessions with free LLM models (Groq llama-3.3-70b, Cerebras gpt-oss-120b)
• Both deliver results to Telegram
Cron 1 — "Reviewer" (every 6 hours):
• Reads one Python source file from a list of 12 trading bot monitors
• Hunts for real bugs: crashes, unhandled exceptions, logic errors, values contradicting a config file (DECISIONS.json)
• Adds found bugs to a BACKLOG.json file with severity ratings
• Never modifies source code — read-only
• Rotates through files sequentially, loops back when all reviewed
Cron 2 — "Fixer" (every 2 hours):
• Picks the first unfixed bug from BACKLOG.json
• Uses sed/python one-liners to make surgical edits (never rewrites full files)
• Safety: backs up file before edit, checks line count after (must be within ±5), verifies Python import passes, auto-reverts from backup if any check fails
• Updates backlog and logs results
Rate limit protection:
• 3-model fallback chain (minimax → cerebras → groq) with exponential backoff cooldowns
• Jobs spaced well below API rate limits (16 requests/day vs 1000+/day allowed)
File safety:
• Git committed before loop started — full rollback available
• Backup-before-edit + auto-revert on failure
• No parallel agents editing the same file
My concerns:
1. Is the fixer prompt robust enough to prevent bad edits even with auto-revert?
2. Could the reviewer generate false positive bugs that the fixer then "fixes" incorrectly?
3. Am I missing any failure mode?
4. What would you change?
The AI that actually does things. Emails, calendar, home automation - all from your favorite chat app. New shell, ready to help. The lobster way.
134,827Members
Resources
Similar Threads
Was this page helpful?
