plugin-manager

Universal Agent Plugins & Skills Ecosystem

Project Overview

A strictly cross-platform (Windows, Mac, Ubuntu) library that serves as the universal upstream source for reusable AI agent plugins and skills across multiple IDEs and agent frameworks:

• Claude Code, GitHub Copilot, Gemini CLI, Antigravity, Roo Code, Windsurf, Cursor, and other compliant integrations.
• Now universally supporting the single .agents/ folder standard (no duplicate copies needed for .github, .gemini, .agent, etc).

120 skills across 29 plugins — all maintained from a single hub-and-spoke source tree.

---

Core Philosophy: Transitional Architectures & Decoupled Skills

This repository is built on a pragmatic acceptance of the current AI engineering landscape: the ecosystem changes weekly, and workflows that were revolutionary six months ago are obsolete today.

Frameworks like agent-agentic-os, spec-kitty, and agent-execution-disciplines are treated as Transitional Architectures — bridges between what agents need to do today and what native SDKs will eventually handle. When Anthropic, Google, and GitHub harden native memory persistence, execution safety, and multi-agent orchestration, large swaths of this tooling will be happily discarded.

Skills are Applications; the SDK is the OS. Individual skills must function in complete isolation — no hard dependencies on sibling plugins, no assumptions about which framework is running.

---

Installation

[!IMPORTANT] Start here — fresh clone or first-time setup. The single .agents/ environment directory is not committed to your repo. It will be empty by default.

All installation methods (uvx, bootstrap.py, npx skills, and Claude Marketplace) are now consolidated in a single authoritative guide:

👉 Go to INSTALL.md

---

Architecture Highlights

Triple-Loop Autonomous Skill Improvement

The agent-agentic-os plugin implements a Triple-Loop architecture for continuous, autonomous skill improvement:

Layer	Agent	Role
L0	triple-loop-architect (Claude)	Interactive setup: scaffolds isolated sibling lab, seeds all files, launches L1
L1	Gemini CLI (gemini --yolo --model gemini-3-flash-preview)	Headless orchestrator: reads eval-instructions.md, runs the loop, gates via evaluate.py
L2	Copilot CLI (gpt-5-mini)	Cheap mutation proposer: proposes SKILL.md edits using free Copilot quota

The loop is autonomous and cost-effective: L2 uses GitHub Copilot's gpt-5-mini (free quota), enabling 20–80 mutation proposals per run at near-zero cost. L1 (Gemini Flash) orchestrates unattended overnight. evaluate.py is the absolute gate — exit 0 = KEEP, exit 1 = DISCARD + auto-revert.

Not all skills are good candidates — the best targets have clear, objective routing criteria and adversarial eval cases. Use eval-autoresearch-fit to score a skill before running a loop.

To start a loop on any skill:

@triple-loop-architect

Kick off a 10-iteration Triple-Loop optimization run targeting the `<skill-name>` skill
inside the `<plugin-folder>` plugin. Use gemini-3-flash-preview as L1 and gpt-5-mini as L2.

See the full sample prompt: references/sample-prompts/triple-loop-architect-prompt.md

Live example — convert-mermaid skill, 26 iterations across 2 rounds: 0.61 → 1.00

Each blue diamond is a baseline anchor (one per session). Green = new best score. Amber = kept but not a record. The two-segment shape shows a fresh re-baseline for round 2 — the plotter handles this automatically.

Monitor a live run: python3 plugins/agent-agentic-os/scripts/plot_eval_progress.py --tsv <lab>/evals/ --live

Flywheel layers:

• OUTER flywheel (os-improvement-loop): improves OS-level protocols and session ledgers between sessions
• INNER flywheel (os-eval-runner + os-skill-improvement): improves individual skill routing accuracy within a session
• Overnight (os-nightly-evolver): runs the INNER flywheel unattended — see agents/os-nightly-evolver.md

Karpathy Autoresearch Loop

Skills that score HIGH on the autoresearch viability rubric (objectivity + speed + frequency + utility) can run fully autonomous self-improvement loops:

mutate SKILL.md → evaluate.py → exit 0 (KEEP) or exit 1 (DISCARD) → repeat

Ecosystem Fitness Sweep v1 is complete — all 116/120 production skills scored for autoresearch viability. Results: