fableforge-agent-skills 0.1.0

Skill registry, decomposition, and LoRA building for coding agents

AgentSkills.org

Skill registry, decomposition, and LoRA building for coding agents.

Overview

AgentSkills.org provides a framework for extracting, publishing, and building LoRA adapters from recurring skill patterns in agent traces. Think of it as a package manager for agent capabilities.

Installation

pip install agent-skills

Quick Start

Install a Skill

askill install debug
askill install edit
askill list

Extract Skills from Traces

askill decompose traces.jsonl --min-occurrences 3

Build LoRA Adapters

askill build traces.jsonl --base-model Qwen/Qwen2.5-14B --output-dir output/lora

Programmatic Usage

from agent_skills import SkillRegistry, SkillDecomposer
from agent_skills.lora_builder import build_lora

# Install and list skills
registry = SkillRegistry()
registry.install("debug")
skills = registry.list_skills()

# Extract skills from traces
decomposer = SkillDecomposer(min_occurrences=3)
skills = decomposer.extract_skills_from_trace("traces.jsonl")
clusters = decomposer.cluster_skills()

# Build LoRA adapters from skill clusters
for cluster in clusters:
    adapter = build_lora(cluster, base_model="Qwen/Qwen2.5-14B")
    print(f"Built: {adapter.name} with {adapter.num_examples} examples")

Built-in Skills

Skill	Tools	Description
debug	read, bash, grep	Diagnose and fix errors in code
edit	edit, write	Make targeted edits to code files
verify	bash, read, grep	Run tests and verify correctness
recover	bash, edit, read, grep	Recover from errors and retry
plan	question, glob, read	Plan and coordinate multi-step tasks
bash	bash	Execute shell commands and scripts

Skill YAML Format

name: debug
version: "1.0.0"
description: "Diagnose and fix errors in code"
category: "core"
tools:
  - read
  - bash
  - grep
triggers:
  - "fix the bug"
  - "debug this error"
author: "your-name"
license: "MIT"
tags:
  - debugging
  - error-recovery

License

MIT

Ecosystem

Part of the FableForge ecosystem — 21 open-source projects built from 210K real agent traces:

Project	Description
Anvil	Self-verified coding agent
VerifyLoop	Plan→Execute→Verify→Recover framework
ErrorRecovery	Self-healing middleware (3,725 error patterns)
FableForge-14B	The fine-tuned 14B model (4-stage training)
ShellWhisperer	1.5B edge agent (phone/RPi, 50ms)
ReasonCritic	Verification model (130 benchmark tasks)
TraceCompiler	Compile traces → LoRA skills
AgentRuntime	Persistent agent daemon (systemd for AI)
AgentSwarm	Multi-agent from real trace transitions
AgentTelemetry	Datadog for agents (token tracking, costs)
BenchAgent	HumanEval for tool-use (107 tasks)
AgentDev	VSCode extension with verification
TraceViz	Trace replay visualizer (Next.js)
AgentSkills	npm for agent behaviors
AgentCurriculum	5-stage progressive training
AgentFuzzer	Adversarial testing for agents
AgentConstitution	Safety guardrails from traces
CostOptimizer	Token cost reduction (50-80%)
AgentProfiler	Behavioral fingerprinting
TrajectoryDistiller	Trace→training data pipeline
Fable5-Dataset	HuggingFace dataset release