nodecode-effort-router
Effort Router: each turn is routed to the model and the reasoning effort a decision model picks for that ask. A Nodecode rewrite of AlphaPerseii3000/jev-effort-router by AlphaPerseii3000.
A conversation is not one kind of work. "what is the capital of Norway" wants a cheap fast model with no thinking; a refactor across four files wants the strongest coding model at high effort. The original plugin put a router between Hermes and its model pick: one decision call per turn, two Choice questions — which model, how hard should it think — and the request went out on the answer. This cell puts the same router at the seam where Nodecode freezes a turn's provider settings, so what a turn runs on is decided once, per turn, and every round of its tool loop shares that one decision.
Get it
Tell the agent "add the effort-router cell", or run nodecode add nc://effort-router.
Set it up
The router asks TypeSafe's Jev, through OpenRouter's Decisions API, for the decision. That needs an OpenRouter key: https://openrouter.ai/settings/keys.
"effort-router": {
"api_key_env": "OPENROUTER_API_KEY"
}
api_key_env is the name of the environment variable holding the key. The other
members, all optional:
| Member | Default | What it does |
|---|---|---|
provider |
ollama-cloud |
the one Nodecode provider whose turns are routed |
grid |
the measured six | the models the decision may choose, one "model: what it is for" per entry |
decision_model |
typesafe/jev-1.13 |
the decision model, pinned so an upstream swap cannot change routing |
endpoint |
OpenRouter's Decisions API | where the decision is asked for |
confidence_percent |
50 |
a decision under this confidence is dropped |
route_per_turn |
on | off, one decision serves a whole session |
timeout_seconds |
8 |
one decision's ceiling |
audit |
on | one line per decision in routes.jsonl under the cell's home |
A section naming no key refuses at start and says so; the tools and the command are
absent until it is set, and /cells shows the line to fix.
Use it
Once it runs, nothing has to be asked for: every turn on the routed provider is decided, and the turn runs on what was decided. Two tools let the model look:
effort_status— the provider, the decision model, the grid, the confidence floor, how many decisions this boot made, and the newest one.effort_route— the decision for a text you hand it, without running a turn on it.
and /effort-router shows the same status with a row to route a text. In eval:
(nodecode-effort-router:decide "prove this lemma about groups") ;; => the route, and why there is none (two values)
What it will not do
The router never costs a turn. No key, a timeout, a connection that dies, an answer that is not the shape it should be, a model off the grid, a confidence under the floor: each of those leaves the request exactly the request it would have been. One decision a turn, not one a round. And only the section's provider is routed: a turn on any other provider is not even asked about, because the grid names models one provider must serve.
The effort rung (low, medium, high) is what the decision picks; which sound each
model's wire takes for it is the provider lane's own clamp, so a model that has no
medium gets the nearest rung it does have, and a model that does not reason gets no
effort field at all.
Credit
Rewritten from AlphaPerseii3000's jev-effort-router; see NOTICE. The original's MIT licence is kept in LICENSE.
