@pi-unipi/skill-registry

Skill registry for pi — per-project skill on/off, a skill vault, and jev-judged skill exposure

Packages

Package details

extension

Install @pi-unipi/skill-registry from npm and Pi will load the resources declared by the package manifest.

npm repohome report
$ pi install npm:@pi-unipi/skill-registry
Package
@pi-unipi/skill-registry
Version
3.0.3
Published
Nov 4, 2026
Downloads
3,086/mo · 496/wk
Author
neuron-mr-white
License
MIT
Types
extension
Size
54 KB
Dependencies
1 dependency · 2 peers
Pi manifest JSON
{
  "themes": [],
  "prompts": [],
  "extensions": [
    "./index.ts"
  ]
}

Security note

Pi packages can execute code and influence agent behavior. Review the source before installing third-party packages.

README

Skill Registry

Skill Registry controls which skills the agent sees, per project, and keeps the system prompt stable for the prefix cache.

@pi-unipi/skill-registry · part of UniPi

What it does

  • Turns each skill on or off, in the global scope or the project scope.
  • Marks a skill as "must show", so the agent always sees it.
  • Keeps a vault of extra skills in ~/.unipi/skill-vault/. Vault skills stay off until a scope turns them on.
  • When you have many skills, lists only the skills that matter to the first real request.
  • Names each hidden skill in a short system prompt section, so the agent can read it when necessary.
  • Keeps the listed set the same for all of the session, so the prefix cache stays valid.

Quick start

Skill Registry ships in @pi-unipi/unipi. To install it alone:

pi install npm:@pi-unipi/skill-registry
  1. Type /unipi:skills. The settings hub opens on the Skills group.
  2. Set Skill proxy to on. Without the proxy, Skill Registry saves your per-skill choices but does not apply them.
  3. Select Skill settings… to open the skill grid.
  4. Set the cells, then press Enter to save.

Changes apply from the next prompt. You do not need /reload.

Commands

Command What it does
/unipi:skills Opens /unipi:settings on the Skills group.

Skill grid

Skill settings… shows each skill with its source (vault, project, user, unipi or package) and two columns:

  • E (enabled): the agent can see the skill. Off removes it from the session, and blocks /skill:name.
  • M (must show): the agent always sees the skill, also when exposure judging would hide it. The bundled summarize skill has M on by default.
Key What it does
↑ ↓, PgUp, PgDn Moves between skills.
← →, Tab Moves between the E and M columns.
Space Toggles the cell.
d Clears the cell in the current scope. The cell then inherits again.
g Changes the scope: global or project.
p Turns the skill proxy on or off.
Other letters Filter by name, source or description. Esc clears the filter.
Enter Saves.
Esc Cancels.

A project value wins over a global value. With no value, vault skills are off and all other skills are on. Skill Registry writes only the cells that you change.

Settings

Open /unipi:settings → Skills. The namespace is skills.

Key Default What it does
proxy false Applies the per-skill states and includes the vault.
exposure.mode judged judged: jev picks the listed skills. all: list all skills. off: remove the UniPi bundled skills from the list.
exposure.threshold 0.8 Minimum jev relevance score for a skill to stay listed. Range 0 to 1.
exposure.maxSkills 12 Maximum number of listed skills. With this number of skills or fewer, no judging occurs.
exposure.recheck true On later prompts, tells the agent about hidden skills that jev finds relevant.
states {} Per-skill enabled and mustShow. Edit it in Skill settings….

The group also has a Decision model section. Set decisionModel.source to inherit or custom.

Set UNIPI_SKILL_VAULT to use a different vault folder.

How it works

In judged mode, Skill Registry picks the listed set on the first real prompt of the session. Greetings do not count. jev is the UniPi Decision Model.

  1. A skill that the prompt names is always kept. A distinctive word of the name is enough: "ssh into coffee" keeps coffee-sandbox.
  2. Must-show skills are always kept.
  3. jev scores the other skills against the prompt. Skills at or above exposure.threshold stay, best first, up to exposure.maxSkills.
  4. A system prompt section names each hidden skill and its folder.

After this, the system prompt does not change for the session. On a later prompt, a message tells the agent about hidden skills that the prompt names. With exposure.recheck, it also tells the agent about hidden skills that jev finds relevant. Each message names 5 skills at most.

If jev does not answer, Skill Registry lists all skills.

The vault holds any folder with a SKILL.md, up to three levels deep: vault/<skill>/ or vault/<pack>/<skill>/.

Other packages can emit the unipi:skills:reveal event with { names: [...] } to tell the agent about hidden skills.

See also