minime

---

"You complete me." - Dr. Evil

Evidence-based agent orchestration for GitHub Copilot.

Four phases. One tiered human gate. The gate stays cheap by handing the inspector an evidence package instead of a verdict, then using ask_user only when human input is truly required.

The core workflow is informed by empirical work (DeepMind 2025, Springer 2026, ClassEval). See REFERENCES.md.

---

🧬 What you get

Five skills plus two agents, distributed as a plugin:

	Skill	Phase	Auto
	/minime:blueprint	Read ranked wiki pages, nudge EARS quality, plan silently, self-challenge	✅
	/minime:replicate	Test-driven generate -> run -> observe -> fix loop	✅
	/minime:inspect	Verify criteria against evidence; route by uncertainty tier	✅
	/minime:extract	Capture lessons into the raw/wiki knowledge base with code citations	✅
	/minime:lab	Bootstrap Virtucon HQ (auto-runs via session-start hook)	🔧

Agent	Role
minime:dr-evil	🦹 Runs the flow end-to-end. claude --agent minime:dr-evil
minime:frau	👓 Evidence-gathering inspector in fresh context. Full tool access, no implementation writes.

Runtime state lives in VIRTUCON_HQ (defaults to $HOME/.minime, overridable via env var). The SessionStart hook resolves paths, injects them into every session, and auto-bootstraps the Karpathy-style knowledge layout.

Knowledge layout

VIRTUCON_HQ now has one shared knowledge root plus per-repo blueprint folders:

• raw/<org>/<repo>/ for immutable source documents such as findings, results, user messages, and hard-won discoveries
• wiki/index.md and wiki/log.md for the shared catalog and ingest log
• wiki/orgs/<org>/<repo>/ for repo topic pages
• wiki/patterns/ for cross-repo guidance
• schema.md for the conventions that shape the wiki
• <org>/_<repo>/blueprints/ for living blueprints

A legacy repo wiki.md becomes raw/<org>/<repo>/legacy-wiki.md when bootstrap sees it. Logs and large outputs stay in execution evidence, not in raw/.

🔐 Subagent policy

• Always use strong, high-reasoning models for subagents.
• Give subagents all tools.

📊 Formal VOI policy (decision hygiene)

• Treat unknowns as decidable-by-data first; only escalate true undecidable-now tradeoffs to user decisions.
• Apply a Value-of-Information gate: run additional research only if it is likely to materially change the choice.
• When a decision is still needed, use ask_user with evidence, suggestions, confidence, reasoning, and a free-text override.

---

The session-start hook auto-bootstraps VIRTUCON_HQ and injects a nudge into every session. No repo-level custom instructions are needed. Describe your task inline and invoke /minime:blueprint to start the flow.

---

💻 Install: GitHub Copilot CLI

Recommended (team/shared): install from marketplace

copilot plugin marketplace add abossard/virtucon
copilot plugin install virtucon@minime

Verify and inspect what loaded:

copilot plugin list
copilot plugin marketplace list

In an interactive session:

/agent
/skills list

Update to latest published version:

copilot plugin update minime

⚗️ Initialize Virtucon HQ

The session-start hook auto-bootstraps VIRTUCON_HQ on every session. Manual bootstrap is rarely needed:

/minime:lab

---

🎮 Use

Two modes: pick the one that fits the task.

🔧 Manual mode (step-by-step)