@lanlance/pi-multi-skill

Stack multiple /skill: references in one message — one turn, one chip per skill.

中文说明

What Is This

pi invokes one skill per message, and only at the very start of the first line. A newline after /skill:name even makes the core parser silently degrade the whole message to plain text.

This extension makes the natural multi-line form just work — no new syntax to remember:

/skill:grill-me
/skill:human-writing

Explain this repo to me.

Submit once: each /skill: reference becomes its own collapsible [skill] chip, followed by the remaining text as your message — and the whole stack reaches the model in a single turn, no intermediate "skill loaded" rounds.

References work anywhere in the message — a standalone line (leading, middle, or trailing) or a trailing token after prose:

Add an AGENTS.md to this project please /skill:writing-for-agents

This is a multi-repo working directory and needs an index

Behavior

Input	Result
leading /skill:a / /skill:b lines + body (multi-line)	one [skill] chip each + body message — handled by this extension
single /skill:a + body on following lines	skill chip + body — also repairs the core newline gap
/skill:a trailing prose (… please /skill:a)	chip + body with the reference stripped — handled by this extension
standalone /skill:a line in the middle/at the end	chip + remaining body — handled by this extension
/skill:a args on one line	untouched core behavior (chip + args)
bare /skill:a on one line	untouched core behavior
reference mid-line (text after it)	treated as plain text, no interception
unknown skill name	reference stays in the text (no interception, no partial sends)
messages with images	passed through to core

Install

pi install npm:@lanlance/pi-multi-skill

or via git:

pi install git:https://github.com/L2ncE/pi-multi-skill

or try without installing:

pi -e /path/to/pi-multi-skill/extensions/multi-skill.ts

How It Works

1. An input event handler extracts line-ending skill references (standalone lines or trailing tokens after prose) and takes over the submission ({ action: "handled" }).
2. Skill names resolve against api.getCommands() (skill:* entries) — user, project, and package sources are all covered.
3. Blocks are built with the exact format the core /skill: expansion produces, so the standard renderer draws the collapsible chips.
4. Every block plus the body is sent as ONE custom message (triggerTurn fires a single turn); the custom renderer reuses the core parser and chip component for each block.

Pure public extension APIs (on, sendMessage, getCommands, registerMessageRenderer, and the officially exported parseSkillBlock / stripFrontmatter rendering components). No editor wrapping, no private seams, no conflict with other editor extensions.

License

Apache-2.0