dev-workflow-agents

Agentic development workflow for Claude Code, VS Code Copilot, OpenCode, and GitHub Copilot CLI.

Orchestrates a Coordinator → Executor → Verifier pipeline with an adaptive fix loop. Quality-first: the Verifier reviews code with fresh eyes, without knowledge of implementation decisions.

---

Requirements

Platform	Requirement
Claude Code	Claude Code CLI installed
VS Code Copilot	VS Code + active GitHub Copilot subscription
OpenCode	OpenCode + any LLM provider
GitHub Copilot CLI	Copilot CLI installed

---

Quick Install

Claude Code

curl -fsSL https://raw.githubusercontent.com/kzzalews/dev-workflow-agents/main/install-claude-code.sh | bash

Installs agents to ~/.claude/agents/ and registers the /dev-workflow skill via claude plugins.

VS Code Copilot

curl -fsSL https://raw.githubusercontent.com/kzzalews/dev-workflow-agents/main/install-vscode.sh | bash

Installs agents to the VS Code user data agents/ directory and ~/.claude/agents/.

OpenCode

curl -fsSL https://raw.githubusercontent.com/kzzalews/dev-workflow-agents/main/install-opencode.sh | bash

Installs agents to ~/.config/opencode/agents/.

GitHub Copilot CLI

curl -fsSL https://raw.githubusercontent.com/kzzalews/dev-workflow-agents/main/install-github-copilot-cli.sh | bash

Installs agents to ~/.copilot/agents/ and registers the /dev-workflow skill via copilot plugin.

---

Usage

Claude Code

Type /dev-workflow in Claude Code to start. You will see:

╔══════════════════════════════════════════╗
║  dev-workflow — configuration            ║
╠══════════════════════════════════════════╣
║  Coordinator : claude-sonnet-latest      ║
║  Executor    : claude-haiku-latest       ║
║  Verifier    : claude-sonnet-latest      ║
╠══════════════════════════════════════════╣
║  Project complexity?                     ║
║    simple  — skip Verifier               ║
║    complex — full pipeline (default)     ║
╠══════════════════════════════════════════╣
║  Override a model? (optional)            ║
║  Syntax: role=model                      ║
║  Example: verifier=claude-opus-latest    ║
╚══════════════════════════════════════════╝

VS Code Copilot

Select dev-workflow from the agent dropdown to start. It guides you through the entire pipeline — tells you which agent to switch to at each step and what to send.

Agent	Role
dev-workflow	Entry point — collects requirements, guides the pipeline
dev-coordinator	Planning, pre-check approval, fix routing
dev-executor	Code analysis and implementation
dev-verifier	Fresh-eyes code review

Custom agents use the dropdown selector — they are NOT invoked via @mention. The @mention syntax only works for built-in chat participants (like @github or @terminal).

OpenCode

Press Tab (or use your switch_agent keybind) to switch to the dev-workflow agent. Describe your task — the guide walks you through the pipeline using @mention for subagents within the same session.

Agent	Mode	Role
dev-workflow	primary	Entry point — guides the full pipeline
dev-coordinator	subagent	Planning, pre-check approval, fix routing
dev-executor	subagent	Code analysis and implementation
dev-verifier	subagent	Fresh-eyes code review

Optional: configure per-agent models in ~/.config/opencode/opencode.json:

{
  "agent": {
    "dev-coordinator": { "model": "openrouter/anthropic/claude-sonnet-4-5" },
    "dev-executor":    { "model": "openrouter/anthropic/claude-haiku-4-5" },
    "dev-verifier":    { "model": "openrouter/anthropic/claude-sonnet-4-5" }
  }
}

GitHub Copilot CLI

Type /dev-workflow to start the skill. Or switch to the dev-workflow custom agent with copilot --agent dev-workflow.

The pipeline agents (dev-coordinator, dev-executor, dev-verifier, dev-workflow) are loaded automatically from ~/.copilot/agents/.

---

Default Models

Role	Claude Code	VS Code Copilot	OpenCode	GitHub Copilot CLI
Coordinator	claude-sonnet-latest	claude-sonnet-latest	global config	claude-sonnet-latest
Executor	claude-haiku-latest	claude-haiku-latest	global config	claude-haiku-latest
Verifier	claude-sonnet-latest	claude-sonnet-latest	global config	claude-sonnet-latest

Claude Code, VS Code Copilot, and GitHub Copilot CLI use *-latest aliases — automatically upgrades to the newest model version.

OpenCode agents inherit the globally configured model by default. Override per-agent via opencode.json (see Usage above).

Override models at runtime (Claude Code only)

At the startup screen, type: coordinator=claude-opus-latest

Available roles: coordinator, executor, verifier

---

Pipeline Overview

User → /dev-workflow
  │
  ▼
[Startup screen: mode + model config]
  │
  ▼
Phase 1 — Coordinator plans, Executor analyzes code
  │        User approves plan
  ▼
Phase 2 — Coordinator oversees, Executor implements
  │        (pre-check → approval → implementation per task)
  ▼
Phase 3 — Verifier reviews with fresh eyes [complex only]
  │        (receives: requirements + task titles + git diff only)
  ▼
Phase 4 — Fix loop (max 3 iterations)
  │   MINOR findings    → Executor fixes directly
  │   ARCHITECTURAL     → Coordinator plans → Executor fixes
  │   UNCERTAIN         → User decides
  ▼
Final report + cleanup

---

Uninstall

# Claude Code
curl -fsSL https://raw.githubusercontent.com/kzzalews/dev-workflow-agents/main/uninstall-claude-code.sh | bash

# VS Code Copilot
curl -fsSL https://raw.githubusercontent.com/kzzalews/dev-workflow-agents/main/uninstall-vscode.sh | bash

# OpenCode
curl -fsSL https://raw.githubusercontent.com/kzzalews/dev-workflow-agents/main/uninstall-opencode.sh | bash

# GitHub Copilot CLI
curl -fsSL https://raw.githubusercontent.com/kzzalews/dev-workflow-agents/main/uninstall-github-copilot-cli.sh | bash

---

Cost Considerations

The pipeline uses Haiku (cheapest) for the Executor and Sonnet for the Coordinator and Verifier. In simple mode, the Verifier is skipped entirely.

The fix loop runs a maximum of 3 iterations. If no progress is detected (same findings repeated), the loop terminates early to avoid wasting tokens.

Tip: For small changes, use simple mode to skip verification and reduce cost.

---

Package structure for agents

Installed files after install-claude-code.sh:

• ~/.claude/agents/dev-coordinator.md — Coordinator agent (claude-sonnet-latest)
• ~/.claude/agents/dev-executor.md — Executor agent (claude-haiku-latest)
• ~/.claude/agents/dev-verifier.md — Verifier agent (claude-sonnet-latest)
• ~/.claude/plugins/cache/kzzalews-dev-workflow-agents/dev-workflow-agents/1.0.0/skills/dev-workflow/SKILL.md — Orchestrating skill (invoked via /dev-workflow)

Installed files after install-vscode.sh:

• <vscode-user-data>/agents/dev-workflow.agent.md — Workflow guide (entry point)
• <vscode-user-data>/agents/dev-coordinator.agent.md — Coordinator agent (claude-sonnet-latest)
• <vscode-user-data>/agents/dev-executor.agent.md — Executor agent (claude-haiku-latest)
• <vscode-user-data>/agents/dev-verifier.agent.md — Verifier agent (claude-sonnet-latest)

Installed files after install-opencode.sh:

• ~/.config/opencode/agents/dev-workflow.md — Workflow guide (primary agent, entry point)
• ~/.config/opencode/agents/dev-coordinator.md — Coordinator agent (subagent)
• ~/.config/opencode/agents/dev-executor.md — Executor agent (subagent)
• ~/.config/opencode/agents/dev-verifier.md — Verifier agent (subagent)

Installed files after install-github-copilot-cli.sh:

• ~/.copilot/agents/dev-coordinator.md — Coordinator agent (claude-sonnet-latest)
• ~/.copilot/agents/dev-executor.md — Executor agent (claude-haiku-latest)
• ~/.copilot/agents/dev-verifier.md — Verifier agent (claude-sonnet-latest)
• ~/.copilot/agents/dev-workflow.md — Workflow guide (entry point)
• plugin dev-workflow-agents@kzzalews-dev-workflow-agents — skill invoked via /dev-workflow

<vscode-user-data> per OS: macOS ~/Library/Application Support/Code/User, Linux ~/.config/Code/User, Windows %APPDATA%\Code\User

State file created at runtime (deleted on completion):

• .dev-workflow-state.md — pipeline memory: plan, pre-checks, results, verification findings

Pipeline agent call sequence:

1. dev-coordinator (Phase 1: analyze + plan, checkpoint with user)
2. dev-executor (Phase 1: code analysis, read-only)
3. dev-coordinator (Phase 2: approve pre-checks)
4. dev-executor (Phase 2: pre-check + implement, per task)
5. dev-verifier (Phase 3: fresh-eyes review, complex mode only)
6. Fix loop: MINOR→dev-executor, ARCHITECTURAL→dev-coordinator→dev-executor, UNCERTAIN→user
7. dev-coordinator (final report)

Model info: each agent outputs [Model: <id>] as the first line of every response.