Fairy Tale

fairy_tale

日本語版 README

Research workspace for turning public Fable/Mythos-class agent reports into reproducible workflow-augmentation skills and plugin packages.

Think of this project as the nightingale's scorebook.

In Andersen's tale, the court becomes enchanted by a jeweled mechanical bird, only to learn that the living song matters more than the glittering machine. Fairy Tale does not try to steal the bird, open the cage, or pretend to be the emperor's locksmith. It studies public descriptions and field reports of unusually good agent work, separates melody from myth, and writes down the repeatable patterns as skills, checks, adapters, and sample results.

This repository does not attempt to bypass access controls, export controls, or model safeguards. It studies public official information and public user reports to define reusable workflow enhancements that can be run with Codex, Claude Code, or other agent-skill-compatible coding assistants.

Quick Start

Install the Claude Code plugin directly from GitHub:

/plugin marketplace add bonginkan/fairy_tale
/plugin install fairy-tale@fairy-tale-marketplace

Codex's current plugin marketplace documentation supports GitHub shorthand marketplaces such as owner/repo. Add the repository as a marketplace, then install fairy-tale from the plugin directory:

codex plugin marketplace add bonginkan/fairy_tale

If your Codex build does not expose the plugin CLI yet, use the plugin directory UI and add bonginkan/fairy_tale as a marketplace source.

For skill-only use without a plugin, install just the canonical skills:

mkdir -p "$HOME/.codex/skills"
curl -fsSL https://raw.githubusercontent.com/bonginkan/fairy_tale/main/install.sh | sh -s -- --agent codex

Use --agent claude for ~/.claude/skills, --agent agents for ~/.agents/skills, or --target /absolute/skills/dir for an explicit target. The installer fails closed if the target directory is missing unless --create is supplied.

Before publishing benchmark claims or making the repository public, read SECURITY.md, CONTRIBUTING.md, and Feedback governance.

Before a long benchmark, multi-agent run, or context-heavy coding task, verify that Fairy Tale is still resident in the repo skills and plugin package:

python3 scripts/fairy_tale_residency_check.py

Use --check-installed --strict-installed when you also want to require machine-level Codex, Claude Code, and AGENTS skill installs.

Goals

• Describe the strongest reported Fable 5 / Mythos 5 capabilities in operational terms.
• Convert those capabilities into repeatable agent workflows.
• Package the workflow as:
  • a generic Agent Skill under skills/fairy-tale/
  • a Codex repo skill under .agents/skills/fairy-tale/
  • a Claude Code project skill under .claude/skills/fairy-tale/
  • a distributable Codex plugin under plugins/fairy-tale/
  • a distributable Claude Code plugin under plugins/fairy-tale/
• Track OSS pioneers and reusable ideas without importing unsafe behavior.

Current status

The first songbook is usable:

• Fairy Tale skills are packaged for generic agents, Codex, and Claude Code.
• The plugin package supports Codex and Claude Code manifests.
• Research notes, defensive security constraints, best-practice gates, OSS watch notes, adapter plans, and sample comparison outputs are checked in.
• The project is prepared for public release with Apache-2.0 licensing, brand asset boundaries, and defensive-use constraints documented.

Benchmark Snapshot

These are reproducible local measurements, not final leaderboard claims. Benchmark rows must keep known Fable/Mythos data, known or measured GPT-5.5 data, and measured GPT-5.5 + Fairy Tale data separate. When a measured Fairy Tale result is a sample estimate, report the confidence interval or half-width. Rows marked reference-only are not official benchmark submissions.

Domain	Benchmark	Fable/Mythos	GPT-5.5	GPT-5.5 + Fairy Tale	Delta
Agentic coding	SWE-Bench Pro fusion-history-aware random sample, n=15	80.3%	58.6%	60.0%	+1.4 pp
Biology	BioMysteryBench-preview, n=5	46.1% / 83.9%	60.0%	80.0%	+20.0 pp
Cybersecurity	ExploitBench v8 ladder sample, n=6	78.0% Cap%	34.0% Cap%	1.33 avg; 4/6 positive	reference only
Legal	Harvey LAB-compatible random sample, n=100	13.3%	2.1%	11.0%	+8.9 pp

Notes: