@gonrocca/skillautoforge-pi

The skill auto-learning loop pi is missing. pi already stores, surfaces, and loads skills — this adds the behavioral loop: distill a reusable skill after a hard task, refine one that proved wrong.

💬 Join the community

Questions, ideas, or want to debate skills & agent design? Come hang out:

skillautoforge-pi is a pi (pi.dev) package. It does not reimplement pi's skill system — pi already does the hard parts:

• Storage: scans ~/.pi/agent/skills/, ~/.agents/skills/, and project skill dirs.
• Surfacing (progressive disclosure): only skill descriptions sit in the system prompt; the full SKILL.md loads on demand.
• Invocation: /skill:name loads and runs a skill.

What pi lacks is the behavioral loop — the nudge to create a skill after a hard task and to refine one when it was wrong. This package adds exactly that, and nothing else. Inspired by Hermes Agent's "after difficult/iterative tasks, offer to save as a skill" pattern.

📦 Install

pi install npm:@gonrocca/skillautoforge-pi

Needs Node ≥ 20.6.0. Restart pi after install.

🛠 What it adds

skill_save tool (LLM-callable)

A deterministic writer. The model supplies a kebab-case name, a specific description (what it does + when to use it), and a Markdown body (pattern, ordered steps, exact commands, gotchas). The extension validates the name, builds valid frontmatter, and writes ~/.pi/agent/skills/<name>/SKILL.md — the library pi already scans. The model never hand-serializes frontmatter or guesses the path.

• Invalid names are normalized (slugified) rather than rejected outright.
• Missing description/body are rejected (pi won't load a description-less skill).
• Saving over an existing name without overwrite:true is refused with guidance to read-then-merge — so refinement never blind-clobbers a skill.

Prompt guidelines (the loop)

While skill_save is active, pi surfaces these guidelines:

• After a difficult, multi-step, or iterative task with a reusable pattern, call skill_save. Routine one-off work is skipped.
• If a skill you loaded was missing steps or had a wrong command, read it and call skill_save with overwrite:true and the merged content — refine, don't duplicate.

These ride pi's promptGuidelines mechanism, so they appear only when the tool is active — no manual system-prompt surgery, no always-on token tax beyond the guideline bullets.

/skill-save command (manual)

Force distillation of the current session:

/skill-save                       # distill whatever is reusable here
/skill-save focus on the deploy   # optional hint

🔄 How the loop closes

1. Distill — after a hard task, the model calls skill_save.
2. Store — the extension writes SKILL.md into pi's user skill library.
3. Surface — pi indexes the new skill's description into the system prompt (next session, or now after /reload).
4. Refine — when the model re-applies a skill and finds it lacking, it reads and re-saves the merged version with overwrite:true.

🔗 Composition with zero-pi / devteam-pi

No code coupling. Because skill_save writes into pi's shared skill library, any agent pi runs — a plain session, a /forge run, a /devteam role — that has skills enabled will see distilled skills surface automatically. To let the loop run inside zero's SDD phases, enable skills on the relevant phase agents (they ship with inheritSkills: false).

🧪 Development

npm test   # node --test --experimental-strip-types

The deterministic core (extensions/skill-write.ts) is pure and unit-tested. Dependency-free, no build step — pi loads the TypeScript directly via jiti.

License

MIT © Gonzalo Rocca