nc://jev-skill-router

Jev Skill Router: names at most one kept skill that fits a request, before anything is loaded. A Nodecode rewrite of ydmw74/jev-skill-router by ydmw74.

Nodecode shows the model one line about every definition it keeps, cut at 60 characters. With a hundred of them the two skills that look alike — the .pptx editing one and the authoring one — read the same, and on a turn where nothing fits the roster still invites a load. This cell answers the question the index cannot:

user request ──▶ skill_select ──▶ Jev (TypeSafe's Decisions API)
                                  1. rank the whole roster, and gate the turn
                                  2. rerank the top 3 with their own text
                 ◀── one skill name, or an abstain

It answers one name or abstains, and an abstain is a real answer: on a prose question the model keeps its own judgement and loads nothing.

Get it

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

Set it up

Two backends, one payload. Either make a key at https://console.typesafe.ai/keys and let JEV_TYPESAFE_API_KEY hold it, or point the cell at an endpoint of your own that forwards {state, questions} to api.typesafe.ai/v1/systemone and returns the raw answer, with its bearer token in JEV_SKILL_ADAPTER_TOKEN. An adapter, when one is named, wins.

The section lives in ~/.nodecode/config.jsonc; every member is optional:

{
  "jev-skill-router": {
    "api_key_env": "JEV_TYPESAFE_API_KEY",
    "adapter_url": "",                   // an endpoint of your own; wins over base_url
    "adapter_token_env": "JEV_SKILL_ADAPTER_TOKEN",
    "base_url": "",                      // empty: https://api.typesafe.ai
    "model": "",                         // empty: jev-latest
    "gate": "0.30",                      // mean of the three request judgments
    "fits": "0.30",                      // the shortlist's best fit judgment
    "shortlist": 3,
    "excerpt": 700,                      // characters of a candidate's own text
    "timeout": 60
  }
}

With no key and no token the cell still starts: a call answers what to set and spends nothing.

Use it

The model calls one tool:

skill_select {request: "the user's request, as they typed it"}
→ {decision: "suggest" | "abstain", suggested: "queue-records" | null,
   roster_size, gate, gate_values, shortlist, fits,
   suggestion_block, seconds}

suggestion_block is the <skill_relevance> paragraph to put after the roster, so the roster itself stays byte-stable and prefix caching holds. gate_values and fits say why the decision came out as it did.

The roster is every skill this image keeps — the same definitions (index) lists, (help 'name) reads, including skills another harness keeps and a SKILL.md a folder links in.

/jev-skill-router answers status: where the credential comes from, how many skills are kept, the endpoint and the thresholds. /jev-skill-router <text> routes one request now and says what it named — the way to try a threshold against real requests before a turn asks a model.

(help :jev-skill-router) carries the same manual into the conversation.

How it works

Two calls, thresholds in the section:

1. Rank and gate. One choice over every skill name, on the same one-line descriptions the index shows, plus three judgments on the request: does it act on the user's own material, does it follow a documented procedure, and (inverted) would prose alone satisfy it. The gate is the mean of the three; below gate nothing is suggested.
2. Rerank. The top shortlist candidates come back with their full description and excerpt characters of their own text, one does-this-skill-do-the-specific-thing judgment each, and one choice with a none_of_these option. Nothing is suggested when the best judgment is below fits.

The answer is at most one name. Candidates rank by probability then name, so a replay is stable.

Test

From the repository root:

.github/test-cell.sh tools/nodecode-jev-skill-router

The slice is offline: every exchange is a stubbed POST answering canned JSON, so both calls, both thresholds and the answer object are exercised with no key and no network.

What the rewrite does not carry

The stdio MCP server loop and its uv launcher — Nodecode dispatches the tool itself, so the loop had nothing left to do. The per-session roster cache is gone too: the roster is read per call, which is what a live index is. The companion SKILL.md is this README and (help :jev-skill-router). See NOTICE.

MIT — see LICENSE.