nc://skill-router

TypeSafe Skill Router: names the one kept skill that fits the request, before the model call. A Nodecode rewrite of DECRUX9812/typesafe-skill-router by DECRUX9812.

Nodecode shows the model one line about every definition it keeps. With a few hundred of them the model reads past the one that would have done the job. This cell sends the request to TypeSafe first and, when a skill genuinely fits, adds one line to the round about to be asked:

<skill_relevance>
Relevant to the current request: queue-records. Ignore this if it does not fit what the user
actually asked for.
</skill_relevance>

It injects nothing when nothing fits, and every failure path — no key, no roster, a timeout, a transport error — stays silent: a routing problem can never break a turn.

Get it

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

Set it up

Make a key at https://console.typesafe.ai/ and put it where this home keeps secrets — TYPESAFE_API_KEY in the environment of the process, or the same variable named in the section. The section lives in ~/.nodecode/config.jsonc; every member is optional:

{
  "skill-router": {
    "api_key_env": "TYPESAFE_API_KEY",
    "base_url": "",          // empty: https://api.typesafe.ai
    "model": "",             // empty: jev-latest
    "gate": "0.30",          // mean of the three request judgments; below it, nothing is named
    "fits": "0.40",          // the winner's own fit judgment; below it, nothing is injected
    "shortlist": 3,
    "chunk": 240,            // skills per request; the API caps one choice at 255 options
    "excerpt": 700,          // characters of a candidate's own text stage two shows
    "timeout": 10,
    "suggest_chars": 4000    // longer requests are left alone
  }
}

A base_url of your own speaking the same POST /v1/systemone takes no key: with none configured the request goes without an Authorization header. With no key and no base of your own the cell still starts and says nothing.

Use it

Nothing to type. Put the cell in and each turn's first round is routed:

• the roster is the skills this image keeps — what (index) lists, including skills another harness keeps and a folder links in. Memories are not routed; only define-skill definitions are.
• the request is that turn's own committed text. Slash commands, empty requests and anything longer than suggest_chars are left alone.
• the answer is at most one name, and the line rides as that round's reminder — never the system prompt, so the cached prefix stays put.

/skill-router answers status: where the key comes from, how many skills are kept, the endpoint and the two thresholds. /skill-router <text> routes one request now and says what it named, without asking a model — the way to try a threshold against real requests before a turn spends anything:

/skill-router queue the new record on the kitchen speaker
queue-records — queue-records fits 0.88 (gate 0.78)

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

How it works

Two requests, thresholds in the section:

1. Wide. One choice per chunk of chunk skills over the same one-line descriptions the index shows, each chunk also offering none_of_these; plus three noul judgments on the request (does it act on the user's own material, does it follow a documented procedure, would prose alone satisfy it). The gate is the mean of those three, the third inverted.
2. Narrow. One choice over the top shortlist plus none_of_these, and one noul per candidate: does this skill do the specific thing the request asks for?

A chunk whose P(none_of_these) is at least a half nominates nobody, except the best chunk, so a shortlist is never empty when something fits. Candidates rank by probability then name, so a replay is bit-stable.

Test

From the repository root:

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

The slice is offline: every exchange is a stubbed POST answering canned JSON, so the chunking, the two thresholds and the recorded line are exercised with no key and no network.

What the rewrite does not carry

The original's on-disk answer cache and its per-turn log file, and its on/off/status/suggest/check command family: a section is the switch here, and the answers are cheap enough to ask again. The thread-pool budget around a decision is gone too — the HTTP call takes a timeout of its own. See NOTICE.

MIT — see LICENSE.