coffeemix_all

Language: English | 한국어

---

An OpenCode + OMO workspace with a layered specialist system for workflow, routing, and validation.

Prerequisites

coffeemix_all requires two programs installed on your system:

Program	Description	Install
OpenCode	Terminal-based AI coding agent	npm install -g opencode-ai
oh-my-openagent	OpenCode plugin — multi-agent orchestration layer	npm install -g oh-my-opencode

After installing both, run bunx oh-my-opencode install to configure your model providers.

Quick Start

# Clone or download this package, then:
cd coffeemix_all_0_1
.\install-global.ps1

See docs/installation-guide.md for full installation options.

What this is

coffeemix_all provides:

• 14 gb-* specialists — narrow-scope agents for review, debugging, planning, diagnostics, status, handoff, and more
• 8 local skills — reusable workflow gates for planning, debugging, verification, and adoption intake
• Validation harness — scenario-based testing for both task behavior and specialist routing
• Global installer — PowerShell script with automatic model mapping from your existing OMO config

Structure

Path	Purpose
.opencode/agents/	14 gb-* specialist prompts
.opencode/skills/	8 local workflow skills
.opencode/plugins/	UI badge plugin
scenarios/	Task-based behavior scenarios
routing-scenarios/	Identity-prompt routing validation (28 scenarios × 14 specialists)
tools/	Validation harness and runners
docs/	Public-facing guides and architecture reference
archive/	Internal planning history and evidence docs

Skills

Skill	Purpose
ask-user-question	Gate dangerous or irreversible operations behind explicit confirmation
compact-context	Summarize long-running sessions while preserving critical information
enter-plan-mode	Switch to planning mode before non-trivial implementation
integration-intake	Gate external workflow/tool adoption behind explicit intake questions
systematic-debugging	Root-cause-first investigation discipline
test-driven-development	Define failing checks or acceptance targets before changes
tool-search	Discover and select tools based on current task context
verification-before-completion	Require fresh evidence before declaring work complete

Validation

Run the validation harness to verify specialist routing and scenario behavior:

# Smoke checks (file existence, inventory, subset scenarios)
python tools/workspace_smoke_runner.py

# Full e2e validation (all scenarios, strict checks)
python tools/workspace_e2e_runner.py

# Routing validation (28 identity-prompt scenarios across 14 specialists)
python tools/routing_validation_runner.py

Reports are written to reports/ (gitignored).

Acknowledgments

This project builds on the work of others in the AI agent ecosystem. We are grateful to:

• OpenCode — The terminal-based AI coding agent runtime that powers this entire workspace. OpenCode provides the execution shell, tool system, and plugin architecture that make coffeemix_all agents and skills runnable. Without OpenCode's open, extensible design, this layered specialist system would not be possible.
• oh-my-openagent — The multi-agent orchestration plugin that sits between OpenCode and coffeemix_all. It provides the model routing, fallback chain, agent registry, and category system that enable gb-* specialists to be selected and dispatched automatically. Its install command also served as the reference model for our global installer's automatic model mapping.
• obra/superpowers — The 8 local discipline skills (ask-user-question, compact-context, enter-plan-mode, systematic-debugging, test-driven-development, tool-search, verification-before-completion, integration-intake) were adapted from this project's skill patterns and workflow gate concepts.
• garrytan/gstack — The multi-agent routing architecture and specialist layering concept (prefer narrow-scope specialists over general agents when tasks match) was inspired by this project's agent orchestration patterns.

Origin of Components

Prefix / Source	Origin
gb-* specialists	Original to coffeemix_all ("gb" = 가베, old Korean word for coffee)
Local discipline skills	Adapted from obra/superpowers
integration-intake	Gate for gb-* integration screening
Baseline agents (sisyphus, oracle, librarian, etc.)	Provided by oh-my-openagent
Command and workflow inspiration	Claude Code by Anthropic

All specialist prompts, local skills, validation scripts, and routing scenarios in this repository are independently authored workspace-local artifacts.

License

This project is licensed under the MIT License — see the LICENSE file for details.

This project includes components from:

• oh-my-openagent — Sustainable Use License v1.0
• obra/superpowers — MIT License (Copyright © 2025 Jesse Vincent)
• garrytan/gstack — MIT License (Copyright © 2026 Garry Tan)