Mycodo Hermes Skill Cron Configuration Guide

May 24, 2026 ยท View on GitHub

This document covers the cron job setup for the Mycodo Hermes Skill decision engine, adaptive follow-up system, and optional contamination pipeline.

Core Cron Jobs

1. Decision Engine (30-minute cycle)

The primary environmental control loop.

FieldValue
NameMycodo Decision Engine -- 30min
Schedule*/30 * * * *
TypeLLM agent (reasoned)
DeliveryTelegram DM (silent on no-action)
*/30 * * * * python3 ~/scripts/mycodo-decision.py --species lions_mane --phase fruiting --auto-phase --execute --camera --html --html-dir ~/reports/

What it does each run:

  1. Fetches live sensor snapshot
  2. Detects or accepts phase
  3. Classifies metrics vs species thresholds
  4. Queries fan + humidifier relay state
  5. Generates and executes actions
  6. Writes follow-up flag if actions fired
  7. Generates HTML report with embedded camera image

Critical flags:

  • --execute -- REQUIRED for relay commands to actually fire (see pitfall P-01)
  • --camera -- REQUIRED for camera image in reports (see pitfall P-11)
  • --html --html-dir <path> -- generates self-contained HTML report
  • --species <id> -- loads species-specific thresholds
  • --auto-phase -- enables auto-transition between phases

2. Adaptive Follow-Up (10-minute cycle)

Zero-cost verification that actuator actions produced the expected environmental change.

FieldValue
NameMycodo Adaptive Follow-up -- 10min
Schedule*/10 * * * *
Typeno_agent (script, zero tokens)
DeliveryOrigin chat (on failure only)
*/10 * * * * bash ~/scripts/mycodo-followup-wrapper.sh

What it does each run:

  1. Checks for ~/.mycodo-skill-followup
  2. If absent: silent exit (no output, no cost)
  3. If present + 10min elapsed: re-reads sensors
  4. Verifies environmental improvement (humidity rose after humidifier ON, CO2 dropped after fan burst, etc.)
  5. Clears flag on success, alerts on failure

Critical: The flag path in the wrapper script MUST match the path used by the decision engine ($HOME/.mycodo-skill-followup). See pitfall P-05 for what happens when they diverge.

3. Scheduled Sensor Checks (Optional)

These provide baseline readings for trend analysis and evening audits.

# Morning sensor fetch
0 6 * * * bash ~/scripts/sensor-query.sh snapshot > ~/sensor-data/raw/sensor-overnight.md

# Morning brief
15 6 * * * # LLM agent run: morning brief with sensor + cron summary

# PM sensor fetch
0 15 * * * bash ~/scripts/sensor-query.sh snapshot > ~/sensor-data/raw/sensor-pm-fetch.md

# PM sensor analysis
15 15 * * * # LLM agent run: PM analysis with phase comparison

# Evening security check
0 20 * * * # LLM agent run: evening audit (sensor health, fan state, visual check)

Evening Audit Pipeline

The evening audit combines sensor health verification, security inspection, and memory accuracy cross-checks.

Protocol

  1. Current sensor readings -- sensor-query.sh quick compared against current phase targets
  2. Fan status check -- verify fan state matches protocol expectations
  3. Visual security scan -- capture camera snapshot for operator review
  4. Memory accuracy cross-check -- validate recent claims against raw sensor logs
  5. Outstanding risk summary -- numbered list of flagged issues

Robustness rules

  • If the decision engine fails for any reason, the audit must NOT abort. Proceed with manual threshold comparison.
  • If raw fetch files are absent (first run or stale cron), query live data instead using sensor-query.sh trend <metric> 6h.
  • Use trend for routine status checks; use history only when investigating specific anomalies (trend produces ~1 KB, history produces 2+ MB).

Token Economics

ComponentPer-Run CostDaily RunsEst. Daily Cost
Decision Engine (30min)Medium (reasoning + actuator execution)48~48 mid-tier calls
Follow-up (10min)Zero1440
Morning/PM/Evening checksLow-to-medium5~5 calls
Total197~53 LLM calls/day

The adaptive follow-up saves approximately 144 LLM calls per day by handling verification in pure Python.

Frequency Tuning

GoalDecision ScheduleFollow-up ScheduleToken Impact
Default*/30 * * * **/10 * * * *53 calls/day
Faster decisions*/15 * * * **/10 * * * *101 calls/day
Slower decisions0 */2 * * **/15 * * * *~17 calls/day

Adjust follow-up independently: keep 10min or increase to 15min for slower-responding environments.

Why 30 minutes?

  • Faster than 60m: Responsive species react to environmental swings quickly. 30m catches drift before it becomes a problem.
  • Not too fast: 15m would over-fire during natural fluctuation. 30m gives the environment time to stabilize after actions.
  • Backed by follow-up: Even if a decision misses something, the 10m follow-up catches it.

Mycodo Restart Handling

On Pi reboot, Mycodo resets ALL outputs to OFF (both fan and humidifier). The next decision engine run:

  1. Detects both actuators are OFF
  2. Reads current conditions
  3. Fires appropriate commands to restore target state
  4. No manual intervention required

Optional: Contamination Pipeline Schedule

If using the contamination monitoring pipeline, add Reddit fetch cron jobs:

# Morning contamination fetch (AM)
0 10 * * * python3 ~/scripts/reddit-fetch-parser.py --out ~/reddit-feeds/raw/contamination-fetch.md --days 7

# Evening contamination fetch (PM)
0 18 * * * python3 ~/scripts/reddit-fetch-parser.py --out ~/reddit-feeds/raw/contamination-fetch-pm.md --days 7

Check fetch freshness: if the latest contamination-fetch.md is older than 24 hours, the pipeline may be stalled. See the contamination pipeline documentation for manual fallback procedures.

Common Cron Mistakes

  1. Missing --execute: Decision engine defaults to dry-run mode. Reports show "Action Taken" but no relays fire. Always include --execute for live control.

  2. Missing --camera: HTML reports will have no camera image. Camera fetching is NOT automatic.

  3. Dropping flags when editing: When adding new flags to a cron prompt, append rather than replace. Always diff old vs new command strings.

  4. Relative paths in prompts: Cron jobs run in isolated sessions. Use absolute paths or $HOME-relative paths, never workspace/... relative paths.

  5. Follow-up wrapper path mismatch: The wrapper and decision engine MUST use the same flag path ($HOME/.mycodo-skill-followup). Verify both after any change.

Directory Requirements

Ensure these directories exist before cron jobs run:

mkdir -p ~/sensor-data/raw \
         ~/sensor-data/snapshots \
         ~/reports/ \
         ~/mycodo-skill-logs/

Camera snapshots should save to ~/sensor-data/snapshots/ (not /tmp/) so they survive reboot and are available for later analysis.