nc://understudy

Understudy: auditions cheaper models on your own past decisions and reports how often each could have taken yours. A Nodecode rewrite of stffinfcti/understudy by stffinfcti.

Get it

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

Set it up

The cell needs an OpenAI-compatible endpoint and one model to judge with:

• base_url — https://openrouter.ai/api/v1 unless you name another.
• api_key or api_key_env — the endpoint's key (OPENROUTER_API_KEY is read either way).
• judge_model — the model id that compares a candidate's answer with the one you actually got. It is required: a folder without it refuses to start and says so.

Optional: sample (past decisions per candidate, 8), days (how far back to look, 14), budget_usd (a hard cap for one run, "1.50"), and prices ("model=input/output", USD per 1M tokens) so the scorecard can say what a run cost. /understudy shows what the cell is connected to.

Use it

Ask the model to audition a model, or type it yourself:

/understudy moonshotai/kimi-k3 openai/gpt-6-mini
/understudy last
/understudy runs

The cell reads your recent turns from this home's session history, replays a sample of the questions on each candidate, has the judge compare each answer with the one the agent actually gave, and answers a scorecard: the takeover rate with its sample size, the misses, the tokens, and the cost when the endpoint or prices says it. It costs real tokens: one call per decision per candidate, plus one judging call each, stopped at budget_usd.

The model reaches the same two verbs as understudy_audition and understudy_report, and (help :understudy) is the primer it reads.

What the port leaves out

The original was written against Hermes's own model facade and its session database, drew a PNG and an HTML report, and priced models from a local models.dev cache. This cell speaks one OpenAI-compatible endpoint over nlk:http, replays questions and answers (not full tool-call transcripts), and answers a text scorecard. See NOTICE.