63 lines
2.0 KiB
Markdown
63 lines
2.0 KiB
Markdown
# No-Agent Cron Job Pattern
|
|
|
|
## Problem
|
|
|
|
LLM-driven cron jobs (even simple ones like "ping router and report") use an LLM on every tick — wasting tokens, adding latency, and most importantly causing the **phantom typing indicator** in Telegram. The user sees "agent is typing..." every few minutes even for trivial watchdog checks.
|
|
|
|
## Solution: no_agent=true
|
|
|
|
When the cron job's task is a deterministic script check (ping, port test, service status, file check, API test), set `no_agent=true`. The script runs standalone — no LLM invoked, no typing indicator, zero tokens.
|
|
|
|
### Design pattern
|
|
|
|
The script must:
|
|
1. Exit 0 (silent) when everything is fine — **no output at all** means no delivery
|
|
2. Print an error message and exit non-zero when something is wrong — `stdout` becomes the delivered alert
|
|
3. Need no LLM reasoning to compose the output
|
|
|
|
### Implementation
|
|
|
|
```bash
|
|
cronjob action=update job_id=<id> no_agent=true prompt="" script=<script_name>
|
|
```
|
|
|
|
The `prompt` is **ignored** when `no_agent=true`. Only the script's `stdout` (on non-zero exit) is delivered.
|
|
|
|
### Watchdog script template
|
|
|
|
```bash
|
|
#!/bin/bash
|
|
# watchdog.sh — Silent on success, alert on failure
|
|
# Runs via no_agent cron
|
|
|
|
TARGET="10.77.0.2"
|
|
SUCCESS=0
|
|
|
|
for i in 1 2 3; do
|
|
if ping -c 1 -W 3 "$TARGET" >/dev/null 2>&1; then
|
|
SUCCESS=$((SUCCESS + 1))
|
|
[ "$SUCCESS" -ge 2 ] && exit 0 # silent
|
|
fi
|
|
sleep 5
|
|
done
|
|
|
|
echo "[-] $TARGET unreachable after 3 attempts"
|
|
exit 1
|
|
```
|
|
|
|
### When NOT to use no_agent
|
|
|
|
- The cron produces raw data that needs narrative synthesis (daily digest, tech briefing)
|
|
- Multiple data sources need correlation before alerting
|
|
- The output format must be platform-specific (e.g. HTML email)
|
|
- A decision must be made from the data (e.g. "is this spike routine maintenance or a problem?")
|
|
|
|
### Converting existing LLM-driven crons
|
|
|
|
Best candidates:
|
|
- Watchdog/ping checks
|
|
- Backup status checks
|
|
- Bounce detection (scan inbox for bounces, output findings)
|
|
- Security scans (Lynis, etc.)
|
|
- Any cron where the output is "up/down" or "ok/fail"
|