nc://model-picker

Model Picker: each turn runs the cheapest of three model tiers that can do the job. A Nodecode rewrite of aean0x/hermes-model-picker by aean0x.

low, default and high are three model ids on one provider, labelled Quick, Standard and Expert. Every prompt is classified into one of them and the turn that follows runs on it; high is meant for money, an irreversible or published action, or a security decision, and a classifier that cannot answer falls back to low rather than failing the turn. Repeated failed tool calls unlock escalate_model, which climbs one tier and no further than high.

Get it

Tell the agent "add the model-picker cell", or run nodecode add nc://model-picker.

Set it up

Give each tier a model id in the model-picker section of ~/.nodecode/config.jsonc (or in /settings), three ids on the same provider, plus that provider's id:

{
  "model-picker": {
    "provider": "openai",
    "low_model": "gpt-5-mini",
    "default_model": "gpt-5",
    "high_model": "gpt-5-pro",
    "escalate_after": 3
  }
}

A tier with no model refuses to start, and the refusal names the key. The other settings are optional: classifier_model sends the classification to a cheaper model (the provider must be named too, since a side call may not move hosts on its own), keep_previous (default true) keeps the last tier when the classifier cannot answer, and each tier's best_for list is what the classifier is told the tier is for — that list, not the code, is the routing policy. config.example.jsonc shows the whole section.

Use it

• Say what you want. A prompt that reads like a lookup runs on the low tier; ordinary work on the default one; anything about money, a publish or security on the high tier.
• /model-picker says where the session stands and offers the picks; /model-picker low|default|high pins a tier for the session, and /model-picker auto resumes classifying. A pinned session is never classified.
• Ask the model to escalate_model after a call has failed escalate_after times in a row: the next turn runs one tier higher.

What it writes is one durable session fact, the same one /models writes: the tier's model becomes the session's own selection, so the next turn reads it with nothing patched. A new fact lands only when the tier changes.

Gaps against the original

The original compacts the transcript before a classifier-driven switch, through a subclass of Hermes's context compressor; Nodecode's context is the model's own to edit (see (help :context)), so nothing is compacted here. escalate_model counts consecutive failed tool calls from the :TOOL point and clears the count on a call that worked, rather than staging a checkpoint. The /medium alias of the original is carried; its MODEL_PICKER_* environment variables and its adjacent config.json are not — settings live in the section, where the setup panel reads them. A tier on another provider is still refused by construction: the section names one provider.

Test

NODECODE_SRC=<a nodecode checkout> .github/test-cell.sh tools/nodecode-model-picker

The slice is offline: the classifier call is stubbed, and a session's selection is the real fact, in a temp store. 16 tests, 0 failed.