nodecode-dogmatic-skill-router

DoGMaTiiC's Skill Router: names the one skill of yours that fits the turn, before the model call. A Nodecode rewrite of DoGMaTiiC/hermes-jev by DoGMaTiiC.

You keep procedures as prose skills — (define-skill name "one sentence" ;; the steps ...). Past a handful of them the index is a list, and the model has to guess which one this turn wanted. This cell asks a judgment model first: one choice over your whole roster, a second look at the best three, two thresholds, and at most one name back. The turn then carries one line:

<harness key="skill_relevance">
Relevant to the current request: xlsx. Load it with (help 'xlsx) when it fits; ignore this if it does not fit what the user actually asked for.
</harness>

It says nothing when nothing fits, and the line tells the model to ignore it when it is wrong: the model stays in charge.

Get it

Tell the agent "add the dogmatic-skill-router cell", or run nodecode add nc://dogmatic-skill-router.

Set it up

It needs one key, from any of three routes — TypeSafe's own API, OpenRouter's Decisions API, or the Vercel AI Gateway:

Route	Key	Where it is read
TypeSafe (default)	`TYPESAFE_API_KEY`	the environment, or `api_keys.typesafe` in `auth.json`
OpenRouter	`OPENROUTER_API_KEY`	the environment, or `api_keys.openrouter` in `auth.json`
Vercel AI Gateway	`AI_GATEWAY_API_KEY`	the environment

Make a TypeSafe key at https://console.typesafe.ai/. Set it in the cell's section or export it; backend picks a route, and auto takes the first that has a key. With no key it loads, starts and stays silent — /dogmatic-skill-router check says which key it is missing.

Everything else is optional and documented in the section: mode (auto by default: speak when a key is present; on; off), gate and fits (the two thresholds, in percent), shortlist, excerpt, suggest_chars, timeout, and the model and base URL of each route.

Use it

Just talk. Every prompt is routed before its turn reads its context, and the turn is given one skill name at most.
/dogmatic-skill-router — where it stands: mode, backend, key, roster size, thresholds; its panel lists the roster, runs a live suggestion, and checks the key.
/dogmatic-skill-router skills — the roster it routes over.
/dogmatic-skill-router suggest <text> — one decision on that text, now.
From eval, (dogmatic-skill-router:route "text") answers the same line, and (dogmatic-skill-router:skills) lists the roster.

What it does

Two judgment calls per eligible turn, and none when the turn is left alone:

Rank. One choice question over every skill you keep, each option its one-sentence description, plus three yes/no questions about the turn (does it act on your machine, does it want a documented procedure, would prose do — the last inverted). Their mean is the gate.
Re-read. The best three come back with their own text, 700 characters each, and one "does it fit" question each. Below gate, or below fits, nothing is named.

Rosters past the API's 255-option cap are chunked at 240, each chunk with a none_of_these option. A slash command, an empty message, a message longer than suggest_chars, and a turn that already carries a line are left alone.

It is advisory by construction: no key, no roster, a slow API, a wrong answer — every path ends in silence, and three failures in a row open a breaker for two minutes so a turn never waits on a router that is down.

What it does not do

It does not load the skill, or run it: it names one, and (help 'name) reads it.
It does not rank the cells you have installed, only the prose skills you keep.
It does not log decisions to a file: /dogmatic-skill-router and (dogmatic-skill-router:route ...) answer with what it decided.