GStack — AI Engineering Workflow

Date Compiled: 2026-04-27
Summary: GStack is a collection of structured skill files that give AI agents persistent roles for software development — CEO reviewer, engineering manager, designer, QA lead, release engineer, debugger. The core insight: an AI agent interacting with a browser needs sub-second latency and persistent state, so gstack runs a long-lived Chromium daemon over localhost HTTP with CDP (Chrome DevTools Protocol), keeping cookies and login sessions alive across commands.

Key Concepts

Persistent Browser Architecture

  • Chromium daemon runs continuously (~3s startup, ~100ms per command after)
  • CLI binary compiled with Bun — single ~58MB executable, no node_modules at runtime
  • Cookie decryption reads Chromium's SQLite cookie database directly (native Bun SQLite)
  • 30-minute idle timeout before browser shuts down

Skill System

Skills are Markdown files invoked by name (e.g., /office-hours). Each skill is a specialist:

Skill Role
/office-hours Start here — reframes product idea before writing code
/plan-ceo-review CEO-level review — find the 10-star product
/plan-eng-review Lock architecture, data flow, edge cases, tests
/plan-design-review Rate design dimensions 0-10
/design-consultation Build complete design system from scratch
/review Pre-landing PR review — catches bugs that pass CI
/debug Systematic root-cause debugging
/qa Open real browser, find bugs, fix, re-verify
/ship Run tests, review, push, open PR — one command
/browse Headless Chromium — real clicks, ~100ms/command
/careful Warn before destructive commands
/freeze Lock edits to one directory — hard block
/guard Activate careful + freeze simultaneously

Why Bun

  1. Compiled binaries — single executable, no node_modules, no npx
  2. Native SQLite — cookie decryption without better-sqlite3
  3. Native TypeScript — no compilation step in development

Key Conventions

  • SKILL.md files are generated from .tmpl templates
  • Safety skills (careful, freeze, guard) use inline advisory prose — always confirm before destructive operations
  • Run bun run gen:skill-docs --host codex to regenerate Codex-specific output

bmad-agents, kelly-handbook-multi-agent, kelly-tweets-agents