pi-jev-skill-suggestion

Pi dumps every installed skill into the system prompt. With hundreds of skills that burns context and makes near-duplicates hard to tell apart.

This extension strips <available_skills>, then asks TypeSafe Jev which skill — if any — to load. At most one skill per turn, or quiet.

Install

pi install npm:pi-jev-skill-suggestion

Or from git:

pi install git:github.com/iamdin/pi-jev-skill-suggestion

Or clone and load once without installing:

git clone https://github.com/iamdin/pi-jev-skill-suggestion.git
cd pi-jev-skill-suggestion
bun install
pi -e ./index.ts

Setup

1. Get a TypeSafe key: https://console.typesafe.ai/settings/keys
2. Export it before starting Pi (shell profile, direnv, or process env):

export TYPESAFE_API_KEY=ts_...

3. Start Pi as usual. First session with a key set asks you to pick a mode (tool or auto).

No key → extension no-ops. Pi keeps its normal skill listing; nothing is stripped.

How to use

tool mode (default)

You chat normally. The agent no longer sees the skill roster in the system prompt. When a task looks skill-shaped, it should call:

skill_suggest({ task: "<what the user asked>" })

You do not call the tool yourself. Switch mode anytime with /jev-skill-mode.

auto mode

You chat normally. After each user prompt, the extension runs Jev itself:

• Fit found → a visible message is injected, e.g.
  Skill recommendation for this turn: skill / location / reason
  The agent is told to read that file.
• No fit / gate says quiet / API error → nothing injected; the turn continues.

Cost / latency: every user prompt triggers at least one Jev call before the agent starts — including quiet turns like what is 2+2?. Prefer tool if you only want routing when the model decides a skill might help.

In auto, skill_suggest is deactivated so the model does not double-route.

What to try

create a short pitch deck as pptx

review this diff against the repo standards

what is 2+2?

Skill-shaped asks should route to a skill (or recommend one in auto). Quiet asks like 2+2 should stay quiet.

How it works

In the Pi session

user user prompt
           │
           ▼
   strip <available_skills>
   inject short mode guidance
           │
     ┌─────┴─────┐
     │           │
  tool mode   auto mode
     │           │
     ▼           ▼
 agent may    extension calls
 call         suggest() now
 skill_suggest
     │           │
     └─────┬─────┘
           ▼
      suggest()
           │
     ┌─────┴──────┐
     │            │
  one skill     none
  (+ path)    (quiet / null)

Roster = installed skills that are not disableModelInvocation. Built each turn from Pi's skill list.

Inside suggest()

Same two-stage idea as the skill suggestion cookbook:

1. Gate — three Noul questions (act on user's system? needs a documented procedure? would prose alone suffice?). Mean oriented score; below 0.30 → no skill.
2. Wide rank — roster chunked (≤254 skills + none_of_these per call). Up to 3 concurrent systemOne Choice calls.
3. Shortlist — merge chunk rankings by score; always keep the best chunk; drop other chunks whose none_of_these ≥ 0.50; keep top shortlistSize (default 3).
4. Narrow — read ~700 chars of each shortlisted SKILL.md, Choice + per-candidate fits Noul. Winner must beat fits 0.40; else none.

Timeout / API error → fail open (no skill; turn continues).

Config

Priority:

1. JEV_SKILL_MODE=tool|auto — overrides mode only
2. .pi/jev-skill-suggestion.json
3. ~/.pi/agent/jev-skill-suggestion.json
4. first-session picker → writes global

{ "mode": "tool", "shortlistSize": 3 }

shortlistSize = how many stage-1 candidates enter stage-2 (clamped 1..32). /jev-skill-mode updates global mode and keeps the current shortlistSize.

Privacy

Sent to TypeSafe: current task / user prompt; skill names + descriptions; short SKILL.md excerpt for shortlisted candidates.

Not sent: chat history, workspace files, credentials, tool results, system prompt.

Develop

bun install
bun run check   # tsc + strip/config + router asserts

index.ts               extension entry (strip, modes, tool)
src/config.ts          mode + shortlistSize
src/strip.ts           listing strip + guidance + auto message
src/router.ts          two-stage Jev suggest()
scripts/check-strip.ts
scripts/check-router.ts

See also

• Cookbook
• Hermes router
• Codex router

License

MIT