nc://effortless

Effortless: judges a message's reasoning effort and sets it on the session. A Nodecode rewrite of ivcained/effortless-hermes by ivcained.

Easy questions run cheap, hard jobs get deeper thinking. A fast judge reads the next user message once — TypeSafe's Jev, a local OpenAI-compatible server, or any custom API you name — and answers one rung of the ladder low, medium, high, xhigh, max. The cell writes that rung to the session with nlk:select-effort, so the turns from the next one on run at it.

The plugin did this automatically, as a middleware on every request. Nodecode has the same knob as a session setting, so here the same judge runs on the model's call, as the pick_effort tool, and on the operator's, as /effortless.

Get it

Tell the agent "add the effortless cell", or run nodecode add nc://effortless.

Set it up

The default judge is TypeSafe's Jev, so it needs a key: put one in /settings under the effortless section as typesafe_key, or set TYPESAFE_API_KEY. A key is made at https://console.typesafe.ai/.

No key wanted? Point it at a model on your own machine:

{
  "effortless": {
    "judge_backend": "local",
    "local_base_url": "http://127.0.0.1:8080/v1",
    "local_model": "qwen3-4b"
  }
}

custom does the same against any OpenAI-compatible API: custom_base_url, custom_model, custom_api_key. fallback_backend names the judge asked when the first is unsure or down (none asks nobody else).

Bounding the verdict: effort_floor, effort_ceiling and effort_bias (-2 cheaper .. +2 smarter) hold a judge inside what you want to pay for, and apply (default true) is whether a verdict is written to the session at all. A model whose own ladder lacks a rung still gets the nearest one below it.

Use it

• The pick_effort tool. The model calls it when a message is plainly a big multi-step job or plainly a quick answer; it fills in the session itself and answers the rung it set and why.
• /effortless — the judge in force.
• /effortless test — one sample prompt through the judge chain.
• /effortless stats — the verdicts this run took, by rung and by judge.
• /effortless <a message> — judge that message and set the session on it.
• In eval: (effortless:pick-effort "review the entire repo" :session-id "s-…").

A level named in the message ("max effort", "effort: high") wins with no judge, and a short follow-up ("go", "ok") keeps the effort already in force.

What leaves the machine: the message (up to 4000 characters) and a short excerpt of the conversation, to the judge you configured, to pick the rung. Nothing else, and no telemetry.