pi-jev-skill-picker

# pi-jev-skill-picker

A Pi extension that keeps the Agent Skills catalog out of model requests and replaces it with one ranking tool backed by [TypeSafe's Jev](https://docs.typesafe.ai) System One model.

Skills stay loaded, so `/skill:name` keeps working. Before each agent turn the extension removes Pi's generated `<skills>` catalog from the effective system prompt and gives the model two tools to pull skills in on demand.

`skill_search` takes a plain-language description of the task, rates every enabled skill against it, and returns the complete `SKILL.md` instructions of the skills that apply.

## What it saves

On a 137-skill catalog the generated `<skills>` block runs about 19,000 tokens, which is 87% of Pi's system prompt. Pi resends it on every request.

Both rows below use the same captured prompt and one trivial turn:

| Model | Catalog present | Catalog stripped | Saving |
|---|---|---|---|
| `gpt-6-astra` | 21,074 tokens, $0.2107 | 2,541 tokens, $0.0254 | 87.9% |
| `deepseek-v4.1-flash` | 22,377 tokens, $0.0034 | 3,434 tokens, $0.0005 | 84.6% |

One `skill_search` call costs about 38,600 Jev input tokens, or $0.0016 at $42 per billion. Jev bills input only. Against `gpt-6-astra` that is under 1% of what a single un-stripped request wastes.

## How the ranking works

Each skill becomes its own Score question. The task goes in the shared `state`, and the skill's name and description go in that skill's own `instructions`. Jev judges each skill without seeing the others, so no keyword prefilter can drop one first.

Every skill is rated on the same three ordered levels:

| Level | Meaning |
|---|---|
| 0 | Unrelated. A different domain, tool or workflow. |
| 1 | Adjacent. Same general area, not the specific thing the task needs. |
| 2 | Directly applicable. Cove