jev-cron-gate

Skip Hermes agent cron runs that would do nothing.

Agent cron jobs often wake up, spend a full LLM run, and report "still waiting." Hermes cron already supports a wake gate: if a job's pre-run script prints {"wakeAgent": false} as its last line, the agent run is skipped. This plugin fills that slot with a TypeSafe Jev check. Jev scores one yes/no question ("would running now do useful work?") over a small snapshot of the job's state, in about 0.5s and for a fraction of a cent.

Install

hermes plugins install jev-cron-gate --enable
# TypeSafe key in your profile .env (JEV_API_KEY also works):
echo 'TYPESAFE_API_KEY=...' >> ~/.hermes/.env

Use

Ask the agent to gate a job, or call the tool yourself:

jev_cron_gate(action="setup", name="heartbeat", job_id="3ce7f1ea833a",
  question="Running the agent now would make real new progress, rather than re-reporting that it is still waiting on approval.",
  sources='[{"file": "~/proj/APPROVALS.md"}, {"file": "~/proj/log.jsonl", "tail": 900},
            {"changed": ["~/proj/src"]}]')
jev_cron_gate(action="test", name="heartbeat")      # dry run: score + state preview
hermes cron edit 3ce7f1ea833a --script jev_gate_heartbeat.py
jev_cron_gate(action="status")                      # runs, skips, skip rate per gate

Sources are what Jev sees: {"file", "tail"}, {"changed": [paths]} (modified since the agent last ran?), {"url"} (localhost only), {"text"}. The combined state is capped at 6,000 characters.

Safety

• Fail-open. A missing key, bad config, network error or Jev error wakes the agent.
• No silent stalls. After max_consecutive_skips skips in a row (default 4), the agent is woken anyway.
• Data leaves the machine. The state snapshot is sent to TypeSafe. Don't point sources at secrets.
• Stdlib only, no hooks. Each decision is logged to <HERMES_HOME>/jev-cron-gates/log.jsonl.

Config

<HERMES_HOME>/jev-cron-gates/<name>.json: question, sources, skip_below (default 0.2), max_consecutive_skips (default 4). The tool writes it for you.

MIT licensed.