agentify

A project-agnostic agent-governance framework for GitHub Copilot: guardrails and skills that keep a hub-and-spoke loop in-lane, shipping reviewable slices. Self-learning, with a hands-free 4-pack — conductor, coder, verifier, architect — or a solo generalist.

The human designs the lanes, guardrails, and constraints and stays final decision-maker.

You design and continuously enhance the loops and guardrails; the agents do the work for you. The system retrospects and self-learns.

Packs

agentify asks for a pack and persona at install; neither has a default.

Pack	Makeup	Separation of duties	Tokens	Use when
1-pack	One generalist designs, implements, verifies, reviews, and owns git.	Waived	Lightest	Small, low-risk work
4-pack	Conductor + Anders (architect), Dave (coder), Bhaskar (verifier)	Strict	Heavy	Independent review matters

The loops

① Hands-free loop — WIP mode

  ① HANDS-FREE LOOP · one task at a time

      ┌───────────────────────────────────────────────────┐
      │  Human — decides · E2E-tests · merges · deploys   │
      └───────────────────────────────────────────────────┘
          │ requests              ▲ escalate anytime
          ▼                       │
      ┌───────────────────────────────────────────────────┐
   ┌─►│  Assistant · conductor — owns git + task file     │
   │  └───────────────────────────────────────────────────┘
   │      │ hands off one task
   │      ▼
   │  ┌───────┐  fail ↔ fix  ┌─────────┐  green  ┌────────┐
   │  │  Dave │◄────────────►│ Bhaskar │────────►│ Anders │
   │  │  code │              │  verify │         │ review │
   │  └───────┘              └─────────┘         └────────┘
   │  uncommitted       build + tests + gates     design
   │                                               │ pass
   │                                               ▼
   │  ┌───────────────────────────────────────────────────┐
   └──┤ commit vibe/<nnn> · push · PR                     │
      └───────────────────────────────────────────────────┘
   ↺ next task

  slice end → pause only if sign-off needed · never trunk · never deploy

② Design session — new-feature mode

  ② DESIGN SESSION · Anders leads with the human

   ┌────────┐   requirements   ┌───────────┐   writes   ┌──────────────────────────┐
   │ Human  │─────────────────►│   Anders  │───────────►│ docs/features/<nnn>.md   │
   │        │◄─────────────────│ architect │            ├──────────────────────────┤
   └────────┘   options · recs └───────────┘            │ O  Options  → pick one   │
  approves the design                                   │ S  Slices                │
                                                        │ T  Tasks                 │
                                                        │ R  Risks                 │
                                                        │ A  Assumptions           │
                                                        │ D  Deferrals             │
                                                        └──────────────────────────┘

  Feature ─► Slices ─► Tasks ─► loop ①

③ Retrospective loop — self-learning

  ③ RETROSPECTIVE LOOP · every ≥ 5 features

   ┌───────────┐    ┌──────────┐    ┌────────┐    ┌────────┐
   │ Assistant │───►│  Anders  │───►│ Human  │───►│  Dave  │
   │  reminds  │    │ distills │    │ okays  │    │applies │
   └───────────┘    └──────────┘    └────────┘    └────────┘
        ▲                                                │
        └───────────── updated guardrails ↺ ─────────────┘
          (1-pack: the solo assistant fills all roles)

Install

1. Run .github/skills/agentify.md from this checkout against a target repository on a non-trunk branch.
2. Choose a pack, persona, and form of address. Choose a model profile or accept mix-1.
3. Review Agentify's repository scan: generated docs/design.md, CI-derived Commands table, gate recipes, test classification, and preflight gates. If CI evidence is absent, supply how to obtain or run the required commands.
4. Answer whether the project has local run/liveness. A “no” removes those duties.
5. Invoke the installed assistant.

Installation is one-shot. The target owns every installed file. If temporarily staged in the target, the installer deletes itself, source templates, version stamps, update markers, and bootstrap references after generating the active governance.

Model profiles

Every agent uses maximum reasoning.

Profile	Designs + codes	Verifies + drives
mix-1 (default)	Claude Opus 5	GPT-5.6 Sol
mix-2	GPT-5.6 Sol	Claude Opus 5
anthropic	Claude Opus 5	Claude Opus 5
openai	GPT-5.6 Sol	GPT-5.6 Sol

Source layout

• .github/agent-templates/ — source-only role and persona inputs.
• .github/agents/ — 4-pack sub-agent sources.
• .github/skills/agentify.md — one-shot installer; never copied.
• .github/skills/ — installed Markdown, preflight, retrospective, and gate recipes.
• .github/instructions/ — path-scoped language rules, including C#.
• docs/ — design templates and feature method.

Installed layout

• .github/copilot-instructions.md — shared guardrails and commands.
• .github/agents/ — one complete file per installed agent.
• .github/instructions/ — path-scoped language rules.
• .github/skills/ — project-owned Markdown, preflight, retrospective, and gate recipes.
• docs/design.md — project architecture, operations, and conventions.
• docs/meta-design.md — feature and test taxonomy.
• docs/features/ — in-flight feature source of truth.