Skip to content

skills.md

Latest commit

 

History

History
87 lines (61 loc) · 2.63 KB

skills.md

File metadata and controls

87 lines (61 loc) · 2.63 KB

Agent Skills

librecode supports Agent Skills as local, reusable instruction bundles.

Discovery

Default roots are checked in priority order:

  1. ./.librecode/skills
  2. ./.agents/skills
  3. ~/.librecode/skills
  4. ~/.agents/skills

A spec skill is a directory containing SKILL.md:

.librecode/skills/my-skill/SKILL.md

Skill discovery uses Agent Skills-compatible directories containing SKILL.md.

Duplicate skill names are resolved by discovery priority. Lower-priority duplicates produce diagnostics and are ignored.

Discovery honors .gitignore, .ignore, and .fdignore files found in scanned skill directories. Symlinked directories are followed through normal filesystem resolution and canonical paths are used for dedupe/cycle prevention.

Frontmatter

Supported fields:

---
name: my-skill
description: Use when this workflow applies.
license: MIT
compatibility: Works with Agent Skills-compatible agents.
allowed-tools: Read Bash(git:*)
user-invocable: true
disable-model-invocation: false
metadata:
  author: me
  version: "0.1.0"
---

Validation rules:

  • name defaults to the parent directory when omitted.
  • name must be lowercase a-z, 0-9, and hyphens only.
  • name must be at most 64 characters.
  • name should match the parent directory name.
  • description is required and must be at most 1024 characters.
  • compatibility must be at most 500 characters.
  • allowed-tools may be either a spec string (Read Bash(git:*)) or a YAML string list.

Prompt behavior

librecode uses progressive disclosure:

  1. The system prompt advertises valid skill names, descriptions, and file paths.
  2. The model may read a skill file when it decides the skill applies.
  3. librecode also performs conservative auto-activation by matching the current user prompt against skill names and descriptions.
  4. Auto-activated skills have their full SKILL.md content injected into the request context, bounded to protect prompt size.

Auto-activation emits a skill_auto_activate lifecycle event for extensions.

Slash command

Inside chat:

  • /skill lists discovered skills.
  • /skill <name> prints the full SKILL.md for that skill.

CLI commands:

librecode skill list
librecode skill show my-skill
librecode skill validate

Remaining compatibility work

The current implementation is intentionally close to the published spec. Future work should focus on:

  • better ignore-pattern parity for advanced .gitignore syntax;
  • explicit skill activation diagnostics in the TUI;
  • optional enforcement/reporting of allowed-tools when tool permissions become first-class.