Skip to content

README.md

Latest commit

 

History

History
97 lines (74 loc) · 4.39 KB

README.md

File metadata and controls

97 lines (74 loc) · 4.39 KB
Raw

VS Code Copilot Agents Pack (Architect + Test Architect)

Created: 2026-02-06

This pack defines a multi-agent workflow for VS Code GitHub Copilot Chat custom agents ("custom agents", previously called custom chat modes). It is optimized for an experienced IT architect + test architect workflow: discovery → architecture & plan → implementation → review → quality gates → documentation.

These agents are meant to be explicitly invoked (e.g. @Orchestrator …, @Researcher …) and to delegate to each other using VS Code’s agent handoffs and subagent tooling.

What you get

Orchestration

  • Orchestrator - Orchestrator / Conductor
    Runs an iterative lifecycle: Discover → Plan → Implement → Review → Commit → Verify, with explicit "STOP" gates so you stay in control.

Planning & Architecture

  • Planner - Autonomous Planner
    Produces a phased plan (TDD-first, incremental), plus risks, non-functional requirements, and acceptance criteria. Hands off to Orchestrator.
  • Researcher-subagent - Context & Research
    Reads code/docs, extracts conventions, finds similar examples.
  • Architect-subagent - Architecture & NFRs
    Produces ADR-ready decisions, trade-offs, and quality attributes checklist.
  • QA-Strategy-subagent - Test Strategy Designer
    Builds a test plan and test pyramid mapping (unit/integration/e2e/contract/visual/perf), plus coverage goals.

Execution

  • Implementer-subagent - Implementer
    Implements only the delegated phase (tests-first), runs targeted test commands, keeps changes small.
  • Explorer-subagent - Scout / Codebase Cartographer
    Fast discovery across repositories: "what files exist, where is X used, what patterns are present".

Assurance & Delivery

  • Code-Review-subagent - Reviewer
    Reviews changed files, enforces quality gates, flags risks, suggests improvements.
  • Security-Auditor-subagent - Security / Threat Model
    Threat-model-light review (auth, secrets, deps, OWASP-ish checks), red-flags only.
  • Docs-Writer-subagent - Documentation & CHANGELOG
    Produces docs/README sections, usage examples, and release notes from the diff.

Installation options

Option A - User-level agents (quickest, per machine)

Copy the *.agent.md files from this folder into your VS Code prompts directory:

  • Windows: %APPDATA%\Code\User\prompts\ (or Code - Insiders)
  • macOS: ~/Library/Application Support/Code/User/prompts/ (or Code - Insiders)
  • Linux: ~/.config/Code/User/prompts/ (or Code - Insiders)

Reload VS Code.

Option B - Repo-level agents (share with team)

Place files under: .github/agents/

VS Code settings

Agent mode is generally available in current VS Code. If your version still gates custom agents or subagents behind settings, check the current VS Code Copilot documentation for the relevant flags.

How to use (typical workflow)

  1. Kick off:
    @Orchestrator Build a solution for <problem>. Constraints: <…>.

  2. Plan-first alternative:
    @Planner Plan implementing <feature> with TDD, include risks and test strategy.
    Then click the handoff to Orchestrator.

  3. Docs/release notes:
    @Docs-Writer Draft release notes and README updates for the last change set.

  4. Analyze tech debt:
    @Code-Review Analyze the last PR for tech debt and maintainability issues.

  5. Analyze tech debt (full lifecycle):
    @Orchestrator Analyze project for tech debt and maintainability issues.

Files

  • Orchestrator.agent.md
  • Planner.agent.md
  • Researcher-subagent.agent.md
  • Explorer-subagent.agent.md
  • Implementer-subagent.agent.md
  • Code-Review-subagent.agent.md
  • Architect-subagent.agent.md
  • QA-Strategy-subagent.agent.md
  • Security-Auditor-subagent.agent.md
  • Docs-Writer-subagent.agent.md

Best practices

  • Use Explorer and Researcher in parallel for large or unfamiliar areas.
  • Keep phases small and TDD-first.
  • Avoid manual testing unless explicitly requested.
  • Stop at commit gates and wait for user approval.

Design principles

  • Context conservation: delegate heavy reading, summarize findings.
  • Parallelism: run independent subagent tasks together.
  • Structured outputs: each agent returns consistent, strict formats.
  • Tool minimalism: only declare tools actually needed.