nc://fleet-policy

Fleet Policy: wake a scheduled turn only when an MCP read tool has something new. A Nodecode rewrite of Julientalbot/hermes-fleet-policy by Julientalbot.

The original asks one question of a Hermes agent: what did it actually send to its provider, what does each cron job cost, and is there anything new to look at? It answered from Hermes's own hooks, its managed scope and its cron store. Nodecode already reports its own turns and its own cron spend, so this rewrite keeps the piece no cell had: the wake gate. fleet_news calls one read tool on one MCP server, remembers the ids it has answered before, and reports only the items nobody has seen -- {"wakeAgent": false} when there are none. A scheduled prompt that begins with that answer spends no model call on "no".

Get it

Tell the agent "add the fleet-policy cell", or run

nodecode add nc://fleet-policy

Set it up

One thing is required: the MCP endpoint every watch calls.

{
  "fleet-policy": {
    "server_url": "http://127.0.0.1:8787/mcp",
    "token_env": "MY_MCP_TOKEN"          // optional: the NAME of an env var, never the token
  }
}

Set it in /settings, on the page's Settings tab, or by hand in ~/.nodecode/config.jsonc, then (restart-cells "nodecode-fleet-policy"). An endpoint that is not an http:// or https:// address is refused at the start, in the loader's own words.

Use it

The model calls the tool; you can call the same verb from eval:

(fleet-policy:watch-new-items :tool "LIST_MAIL" :id-field "id"
                               :show-fields '("subject" "from.name"))

• The first call for a watch records a baseline and reports nothing.
• Every later call reports the items whose id was never seen, one short line each, then marks them seen. Nothing is reported twice.
• Nothing new answers {"wakeAgent": false} -- the stock wake gate.
• :dry-run t prints what a call would answer and records nothing.
• The state is one small file per watch under fleet-policy/ in the Nodecode home (~/.nodecode/fleet-policy/ by default, state_dir moves it). A corrupt state file is a refusal, not a silent fresh start, and a read that failed records nothing, so a broken endpoint never marks items as seen.

/fleet-policy lists what each watch remembers; /fleet-policy LIST_MAIL prints what that tool would report now, as a dry run.

What counts as an item

Every outermost object in the answer that carries a scalar id_field, at any depth -- including a JSON document a relay wrapped in a string. Inside an item nothing is searched, so a nested id (an item's author, say) is an attribute, not an item of its own. An answer that says successful: false, or carries a non-empty error, is a failed read: the ids are left alone and the call refuses in words.

In a scheduled prompt

First run (fleet-policy:watch-new-items :tool "LIST_MAIL")
through eval. If it answers {"wakeAgent": false}, stop here and say nothing.
Otherwise summarise the new items and what they need.

That is the whole gate: no new items, no model work.

Credit

A rewrite of Julientalbot's hermes-fleet-policy; see NOTICE. MIT, the original's licence; see LICENSE.