22 agents · 3 model tiers · full handoff map
Why junai?
Most AI coding tools are chat assistants. You ask, they answer. There is no memory, no process, no structure — and when your session ends, context dies with it.
junai is different. It brings Agentic Engineering to your codebase: a structured software delivery lifecycle orchestrated by 23 role-scoped AI agents, driven by a deterministic state machine, and tracked in a plain-text pipeline-state.json that survives every session restart.
Idea → Intent → PRD → Architecture → Plan → Implement → Test → Review → Shipped
↑ every stage has a specialist agent, every transition is logged, every gate is explicit
|
junai |
Generic AI Chat |
Other Agent Tools |
| Persistent state across sessions |
✅ |
❌ |
❌ |
| Deterministic routing state machine |
✅ |
❌ |
❌ |
| 23 role-scoped specialist agents |
✅ |
❌ |
4–6 generic agents |
| Full SDLC: Intent → PRD → Arch → Plan → Impl → Test → Review |
✅ |
❌ |
Partial |
| Three pipeline modes: supervised / assisted / autopilot |
✅ |
❌ |
❌ |
| Autopilot watcher — auto-opens next agent, zero clicks |
✅ |
❌ |
❌ |
| Works with GitHub Copilot you already pay for |
✅ |
❌ |
No — requires new API keys |
Repo-portable — travels with your .github/ folder |
✅ |
❌ |
❌ |
| Auditable — all state in git-committed plain text |
✅ |
❌ |
❌ |
| MCP tools — pipeline ops callable from Copilot chat |
✅ |
❌ |
Varies |
🚀 Three Pipeline Modes
| Mode |
How it works |
Best for |
| 🎛️ supervised |
Every gate requires your explicit approval before advancing |
Learning the pipeline, high-stakes changes |
| 🤝 assisted |
AI generates guidance and gate recommendations — you approve |
Most day-to-day feature work |
| 🤖 autopilot |
All gates auto-satisfied after intent sign-off. The extension watches pipeline-state.json and automatically opens the next agent and sends its routing prompt — no clicks needed |
Trusted, well-scoped work |
Switch mode anytime from Copilot chat:
"Switch pipeline to autopilot mode"
🤖 Autopilot Watcher
In autopilot mode, the junai extension watches your pipeline-state.json in real-time. The moment a stage completes and the routing decision is written, the extension:
- Reads
_routing_decision.target_agent from the state file
- Automatically opens the correct specialist agent in Copilot chat
- Sends the handoff prompt — the agent starts working immediately
You sign off the intent once. The pipeline does the rest.
🧠 23 Specialist Agents
Each agent is a deeply crafted instruction file in .github/agents/ — scoped to a single responsibility, model-matched for its task, and wired with handoff buttons to the next stage.
| Tier |
Agents |
Model |
| 🔵 Deep Reasoning |
Architect, PRD, Plan, SQL Expert, Security Analyst, Data Engineer, Debug, UX Designer, UI/UX Designer |
Claude Opus 4.6 |
| ⚡ Multi-file Coding |
Implement, Frontend Developer, Streamlit Developer, DevOps, Tester |
Claude Sonnet 4.6 |
| 🎯 Orchestration |
Orchestrator, Code Reviewer, Mentor, Janitor, Prompt Engineer, Mermaid Specialist, SVG Diagram, Accessibility |
Claude Sonnet 4.6 |
All agents share a handoff protocol — each completion writes artefact paths and routing context into pipeline-state.json, so the Orchestrator can cold-start a new session from state alone.
The junai MCP server provides callable pipeline operations from Copilot chat:
| Tool |
What it does |
pipeline_init |
Initialise a new pipeline (active-pipeline guard built-in) |
pipeline_reset |
Force-clear and restart (bypasses guard) |
notify_orchestrator |
Record stage completion + trigger routing decision |
set_pipeline_mode |
Switch supervised / assisted / autopilot |
satisfy_gate |
Manually satisfy a supervision gate |
get_pipeline_status |
Read current stage, mode, routing decision |
validate_deferred_paths |
Verify deferred item file paths before close |
run_command |
Execute CLI commands from chat context |
⚡ Quick Start
- Open any project in VS Code
- Press
Ctrl+Shift+P → junai: Initialize Agent Pipeline
- The full agent pool installs into
.github/ and the MCP server is configured
Then in Copilot chat, open @Orchestrator and say:
New feature: <describe what you want to build>
I want to run this in autopilot mode.
The pipeline takes it from there.
📁 What Gets Installed
One command. Everything below lands in your .github/ folder and travels with your repo.
| Folder |
Contents |
agents/ |
23 role-scoped agent files — deeply crafted system prompts per specialist |
skills/ |
Reusable skill modules agents can load on demand |
prompts/ |
Workflow-level prompt templates (ADR, conventional commit, handoff, etc.) |
instructions/ |
VS Code .instructions.md files for Copilot context (Python, FastAPI, SQL, Docker...) |
plans/ |
Plan templates and backlog scaffold |
agent-docs/ |
ARTIFACTS hub, architecture, PRD, UX, security docs |
handoffs/ |
Cross-session context handoff protocol |
Plus at the root level:
pipeline-state.json — live pipeline state (stage, mode, gates, routing decisions, artefact refs).vscode/mcp.json — MCP server registered automatically (requires uv)
🛠️ Commands
| Command |
What it does |
junai: Initialize Agent Pipeline |
Install the full agent pool and configure the MCP server |
junai: Update Agent Pool |
Pull latest agent/skill files from the extension — preserves your pipeline state |
junai: Show Pipeline Status |
Inline status: current stage, mode, gate states, last routing decision |
junai: Set Pipeline Mode |
Switch pipeline mode without re-initialising |
junai: Remove from this project |
Clean uninstall — removes agent pool and MCP config |
⚙️ Extension Settings
| Setting |
Default |
Description |
junai.defaultMode |
supervised |
Pipeline mode applied on Initialize |
📋 Requirements
- VS Code 1.101+
- GitHub Copilot subscription (the agents are Copilot chat participants)
uv on PATH for MCP server — install uv (one command, 30 seconds)
📚 Learn More
The future of software engineering is agentic.
junai makes it structured, auditable, and yours.
MIT © junai Labs
| |
