Agent Tools

Reusable skills and agents for AI coding assistants — a harness-agnostic plugin ecosystem for codebase analysis, feature development, debugging, documentation, code review, and spec-driven development.

Architecture

The project is a three-domain system unified by the filesystem as the sole integration contract:

flowchart TD
    subgraph plugins["Plugin Ecosystem (30 skills, 12 agents)"]
        CS[create-spec]:::primary --> CT[create-tasks]:::primary
        CT --> ET[execute-tasks]:::primary
        ET --> TE[task-executor agents]:::secondary
    end

    subgraph filesystem["Filesystem Contract"]
        TF[".agents/tasks/{status}/{group}/task-N.json"]:::warning
    end

    subgraph webapp["Task Manager Web App"]
        W[chokidar watcher]:::success --> WS[WebSocket broadcast]:::success
        WS --> KB[Kanban Board UI]:::success
    end

    subgraph cli["agt CLI"]
        AGT[agt start/stop/status]:::neutral
    end

    TE -->|writes JSON| TF
    TF -->|watches| W
    AGT -.->|manages lifecycle| webapp

    classDef primary fill:#dbeafe,stroke:#2563eb,color:#000
    classDef secondary fill:#f3e8ff,stroke:#7c3aed,color:#000
    classDef success fill:#dcfce7,stroke:#16a34a,color:#000
    classDef warning fill:#fef3c7,stroke:#d97706,color:#000
    classDef neutral fill:#f3f4f6,stroke:#6b7280,color:#000

Loading

Skills are organized into a 4-type taxonomy:

Type	Count	Purpose
Workflow	14	Multi-phase orchestrations with agent coordination
Dispatcher	3	Thin wrappers for shared agent dispatch
Reference	9	Knowledge bases loaded on demand by other skills
Utility	3	Standalone single-purpose tools

Agents are nested inside their owning skills. When a second skill needs the same agent, it's promoted to a dispatcher skill.

Skills

Core (20 skills)

Workflows: deep-analysis, feature-dev, bug-killer, codebase-analysis, mr-reviewer, docs-manager, release-python-package

Dispatchers: code-exploration, code-architecture, research

Knowledge: language-patterns, architecture-patterns, technical-diagrams, code-quality, project-conventions, changelog-format, glab

Utilities: git-commit, document-changes, project-learnings

SDD Pipeline (8 skills)

Four-stage spec-driven development: create-spec → [analyze-spec] → create-tasks → execute-tasks

analyze-spec is an optional quality gate that scores specs across 4 dimensions before task decomposition.

Supported by sdd-specs (templates) and sdd-tasks (task schema) reference skills. Supplemented by inverted-spec — a special-use-case skill that reverse-engineers specs from existing codebases.

Meta (2 skills)

Workflows: create-skill (GAS-only portable), create-skill-opencode (multi-platform: GAS/OpenCode/Codex)

Key Design Decisions

• Harness-agnostic: Every skill includes a dual Execution Strategy — subagent dispatch when available, sequential inline fallback when not
• Mostly read-only agents: Most agents have only Read/Glob/Grep/Bash access. Two exceptions: task-executor (Write+Edit for code implementation) and changelog-manager (Edit for CHANGELOG.md updates). All other file modifications are handled by the orchestrating lead.
• Hub-and-spoke coordination: Workers explore independently; all coordination flows through the lead
• File-based task management: SDD tasks use directory position as state (pending/ → in-progress/ → completed/)

Structure

plugins/
├── core/skills/       # 20 general-purpose skills
├── meta/skills/       # 2 skill-authoring skills
├── sdd/skills/        # 8 spec-driven development skills
└── manifest.json      # Skill registry
apps/
├── task-manager/      # React + Node.js web app for SDD task management
└── cli/               # agt CLI for app lifecycle management
internal/
├── reports/           # Architecture decision reports
├── docs/              # Analysis documents
└── specs/             # SDD spec inputs
scripts/
└── installers/        # Cross-platform install scripts

Installation

Cross-platform installers are available in scripts/installers/ for Bash, PowerShell, and CMD.