cron-no-agent-conversion

name	cron-no-agent-conversion
version	1.0.0
category	devops
description	Convert LLM-driven Hermes agent crons to no_agent scripts with targeted API calls. Maximizes deterministic Python work, uses LLM only for the single creative task the model uniquely provides.
author	Hermes Cortex
license	MIT
metadata	{"hermes":{"tags":["cron","no-agent","automation","script","conversion","deepseek","ollama"]}}

Cron → no_agent Script Conversion

A class-level pattern: replace full agent-loop crons (which run an LLM agent with tools each tick) with lightweight no_agent=true scripts that handle deterministic orchestration in Python and make a single API call for creative work.

When to Convert

Convert when the cron's LLM produces 5-50 token useless output (1 API call, no real work). This is the signature of a small model struggling with a multi-step agentic workflow.

Small local models like qwen2.5:3b (3B params) are reliable for:

• ✅ Single-shot tasks (code gen, classification, short summaries)
• ✅ Deterministic transformation of structured input
• ❌ Multi-step tool loops (search → read → decide → act) — they lose the thread

If your cron's agent calls 3+ tools per tick but the output is boilerplate, it's a candidate. Convert it.

The Conversion Pattern

Before (agent cron)

LLM-driven cron: the prompt asks the agent to gather data, reason, and deliver — and the agent burns tokens on tool calls every tick.

After (no_agent script)

#!/usr/bin/env python3
"""Cron script: gather data deterministically, ONE LLM call for the summary."""
import json, subprocess, urllib.request

# 1. Deterministic data gathering (no LLM)
logs = subprocess.check_output(["grep", "-E", "ERROR", "/var/log/nginx/error.log"])
lines = logs.decode().splitlines()[-50:]

# 2. Single creative LLM call
payload = {
    "model": "deepseek-v4-flash",
    "messages": [{"role": "user", "content": f"Summarize these nginx errors:\n{chr(10).join(lines)}"}],
    "max_tokens": 300,
}
req = urllib.request.Request(API_URL, json.dumps(payload).encode(), headers={"Authorization": f"Bearer {KEY}"})
summary = json.loads(urllib.request.urlopen(req).read())["choices"][0]["message"]["content"]

# 3. Deterministic delivery
print(summary)  # cron delivers stdout verbatim

Cron registration

# Script lives at ~/.hermes/scripts/<name>.py
cronjob action=create
  name=local-nginx-error-summary
  schedule="0 6 * * *"
  no_agent=true
  script=~/.hermes/scripts/local-nginx-error-summary.py

Name it per the fleet convention: agent- (all agents) or local- (this machine only) — see cron-job-management.

Decision Rules

Condition	Keep agent cron	Convert to no_agent
Needs tool access (terminal, web, files) mid-loop	✅ keep	❌
Needs skills / multi-step reasoning	✅ keep	❌
Deterministic pipeline + one summary/decision at the end	❌	✅
Output is always same shape (status, digest, watchdog)	❌	✅
Small model + many tool calls + boilerplate output	❌	✅

Verification After Conversion

# Run the script once manually
python3 ~/.hermes/scripts/local-nginx-error-summary.py

# Confirm output shape is stable
# Then in Hermes: cronjob action=run <job_id> → confirm delivery

Pitfalls

• ❌ Keeping LLM in the hot path — if the deterministic part can compute the answer, don't pay for an LLM call every tick.
• ❌ Converting a cron that genuinely needs tools — tool loops belong in agent crons; forcing them into scripts creates brittle subprocess chains.
• ❌ Secrets in the script — read API keys from a file (KEY = open("~/.secrets/key").read().strip()), never hardcode.
• ❌ Manual test ≠ scheduler test — after fixing a cron, run cronjob action=run job_id=<id> so the scheduler's recorded status refreshes, then run the doctor to confirm it clears.
• no_agent scripts must be silent on success — empty stdout + exit 0 delivers nothing; any stdout or non-zero exit delivers an alert.

Related

• cron-job-management — naming, installers, doctor truth source
• cron-format-standard — output format for LLM cron deliverables
• cron-quality-gate — preventing garbage cron output