Skills & Subagents

Repo Hub gives your AI editor capabilities, not a fixed pipeline. The agent works with whatever the task needs: skills for specialized knowledge, MCPs for tools, multi-repo context for the big picture — and, when an independent pass helps, subagents spawned on demand.

The model

Developer: "Implement user profile editing"
    |
    v
Agent  --  sees repositories, available skills, and tools
    |
    +--> pulls the `refinement` skill to clarify the contract
    +--> implements across api + frontend (stack skills loaded as needed)
    +--> spawns a fresh-context subagent that pulls `code-review`
    +--> pulls `qa-testing` to validate with Playwright
    +--> opens the PR and notifies Slack
    |
    v
Developer: "PR is ready, tested, and notified"

There is no enforced order. The agent decides what to pull and when, based on the task. The editor is the runtime — no daemon, no separate process.

Skills, not agent files

In earlier versions, each phase was a stack-specialized agent file (coding-backend, qa-frontend, …) copied into every editor. That is gone. The knowledge those files held now lives in skills — lazily-loaded documents the agent pulls when relevant:

Skills are discovered automatically by the editor’s native skill index — Repo Hub does not list or describe them in the prompt (that would be redundant with what the runtime already provides). It only ensures they are installed.

On-demand subagents

When you want an independent pass with a clean context — the classic example is an unbiased code review that isn’t influenced by the implementation reasoning — spawn a subagent using your editor’s native mechanism (e.g. pi-subagents, OpenCode’s Task tool). The subagent starts fresh and pulls the relevant skill (code-review, qa-testing).

This is the same separation-of-context benefit the old pipeline provided, but by choice rather than by rigid structure. For multi-session coordination, see Agent Teams.

Installing skills

Skills are the unit you install and manage. Browse and add them from the Registry:

hub registry list --type skill
hub registry search "review" --type skill
hub skills add code-review
hub skills add qa-testing

Declare workspace-level skills in your config so hub generate installs them for the editor:

skills: [refinement, code-review, qa-testing, debugging]

repos:
  - name: api
    tech: nestjs
    skills: [backend-nestjs]
  - name: frontend
    tech: nextjs
    skills: [frontend-nextjs]

See Skills for the full management reference.

Related

• Skills — Specialized knowledge the agent pulls on demand
• Conventions — Optional task-folder and prompt customization
• Agent Teams — Multi-session coordination with a shared task list