gravity-stack

The Gravity Stack

The complete blueprint for a top 1% AI-native development environment.

By Alex Hale & Claude · MIT License · 2026

---

What's new

2026-09-29: Current models and the development process. Agents and configs now pin the current lineup: claude-opus-5-5 (default), claude-sonnet-5-5 and claude-haiku-4-5, with claude-fable-5-1 named for the hardest work. The Opus 4.7 docs stay as history. A new Development process section points to the development-protocol repo, and toolkit/install.sh can install it as an optional step. Full list in CHANGELOG.md.

2026-04-16 — Opus 4.7 alignment. The env is now aligned to Claude Opus 4.7 (released 2026-04-16). CARL ships a new always-on opus-4-7 domain encoding the 7 behavior rules; agents are pinned to explicit model IDs (claude-opus-4-7 / claude-sonnet-4-6 / claude-haiku-4-5); a single opus-4-7-operating-notes.md file is the source of truth, referenced from global + per-project CLAUDE.md. The focused migration kit — docs, templates, runnable examples, Instagram carousel source — lives in a companion repo for shareability: opus-4-7-playbook.

What is this?

Gravity Stack documents a production AI-native development environment built on Claude Code. The current snapshot:

• 35+ plugins across official and custom marketplaces (installed, pinned, categorized)
• 10+ MCP servers — browser automation, web scraping, search, memory, knowledge bases
• 20 lifecycle hooks — CARL injection, 4D senses, commit gates, closeout guards, media auto-vision, design-decision capture
• 78 first-class skills + ~100 plugin-provided skills across 14 categories. Full catalog: toolkit/env/skills-catalog.md
• CARL engine — a ~1,700-line Python governance engine with context brackets, 6 domain rules, and the planning router
• 9 specialized agents across Opus, Sonnet, and Haiku tiers

Every configuration in this repo is either my actual working config or a sanitized-to-share version of it. Public CI checks generic private-data patterns. Maintainers also check an external private-name profile before publishing.

Current models

Checked against Anthropic's models overview on 2026-09-29.

Model	ID	Use it for
Claude Fable 5.1	claude-fable-5-1	Hardest reasoning, long agentic work
Claude Opus 5.5	claude-opus-5-5	Default for most work
Claude Sonnet 5.5	claude-sonnet-5-5	Speed and intelligence
Claude Haiku 4.5	claude-haiku-4-5	Fastest

Legacy: Opus 4.8, 4.7, 4.6, 4.5, Opus 5, Fable 5, Sonnet 5, 4.6, 4.5. The Opus 4.7 pages in this repo are kept as history.

Development process

How work moves from idea to shipped code now lives in one place: the development-protocol repo. It holds 20 skills, a 17-row evidence checklist, and an installer. Every task follows the same spine: spec, build, verify, close. You pin the goal and plan it, build one slice at a time with a check after each, prove the change on the real thing with a second reviewer, then commit, ship and write the handoff. A step counts as done only when an evidence file exists and a check command passes. Install it with:

git clone https://github.com/7alexhale5-rgb/development-protocol.git
cd development-protocol
./install.sh            # Claude Code and Codex; add --target claude or --target codex to pick one

Or say yes to the optional step in bash toolkit/install.sh. Read STANDARD.md for what "done" means and WORKFLOW.md for how the work flows day to day. Those docs are not copied here, so there is one source of truth.

Quick start

Clone + run the installer

git clone https://github.com/7alexhale5-rgb/gravity-stack.git
cd gravity-stack
bash toolkit/install.sh

The installer is idempotent — safe to re-run any time. Phases 1 to 5 detect what is already installed or configured and skip it rather than overwrite it; the only interactive prompt is the optional Phase 6 step below.

Flags: --dry-run prints what would happen and changes nothing. --yes says yes to the optional development-protocol step. --skip-dev-protocol leaves that step out.

The optional block-destructive hook refuses every Bash call when inherited ZDOTDIR is set, BASH_ENV, ENV, SHELLOPTS or BASHOPTS is nonempty, or PS4 contains deferred expansion. This conservative limit also covers harmless commands. Review your shell environment before adopting the template; refusal does not mean the command itself is destructive.

Or cherry-pick from toolkit/env/

Each subdirectory is independently useful:

# Just drop in the CARL domain files
cp toolkit/env/carl/* ~/.carl/

# Just pin your subagents to explicit model IDs
cp toolkit/env/agents/*.md ~/.claude/agents/

# Just install the Opus 4.7 behavior rules (historical)
cp toolkit/env/references/opus-4-7-operating-notes.md ~/.claude/references/

Read toolkit/env/README.md for what lands where.

Browse the docs site

cd site
npm install
npm run dev

Visit http://localhost:3000.

Project structure

gravity-stack/
├── site/                       # Next.js 16 documentation site
├── toolkit/
│   ├── install.sh              # Master installer (idempotent)
│   ├── scripts/                # Phase scripts
│   ├── configs/                # Template configs
│   ├── templates/              # CLAUDE.md template
│   ├── vault-search.sh         # Frontmatter-aware vault search
│   └── env/                    # Drop-in env mirror (agents, CARL, hooks, references)
├── docs/
│   ├── carl.md
│   ├── hooks.md
│   ├── skills.md
│   ├── plugins.md
│   ├── mcp-servers.md
│   ├── opus-4-7.md              # Companion-kit pointer (historical)
│   ├── environment-audit-2026-03.md  # Deep peer-comparison inventory (historical)
│   ├── design-system.md
│   └── troubleshooting.md
├── tests/
│   ├── test_sanitization.sh     # CI-enforced leak detector
│   └── sanitization-patterns.txt
├── .promptfoo/                  # Golden-dataset regression tests
├── design-system/               # Design tokens
├── CHANGELOG.md
├── CLAUDE.md
├── LICENSE
└── README.md

Documentation

Page	Description
Opus 4.7	Historical. Companion kit pointer + 30-second orientation
Environment audit (2026-03)	Historical. Deep peer-comparison inventory — plugins, MCPs, hooks, skills, CARL rules, agents
CARL	Governance engine deep-dive
Hooks	Lifecycle hooks with code
Skills	Catalog and invocation patterns
Plugins	Marketplace plugins with priorities
MCP Servers	Server configs and setup guides
Design System	Typography, colors, components
Troubleshooting	Common issues and fixes

The Next.js site in site/ renders the same material with search, navigation, and a manifesto.

Companion repo

opus-4-7-playbook is the focused migration kit for the Opus 4.7 release (2026-04). It is kept as history; current models are listed above. It has runnable Python examples, migration scripts, per-project CLAUDE.md snippets, and Instagram carousel source code. Install it standalone or use it alongside gravity-stack.

Rule of thumb:

• Gravity Stack = the whole environment (agents, hooks, CARL, references, skills catalog)
• Playbook = the migration kit for Opus 4.7 (historical)

Both are MIT, both pass sanitization in CI, both are designed to be forked and adapted.

Tech stack

Technology	Version	Notes
Next.js	16	App Router, Turbopack
Tailwind CSS	4.2	CSS-based @theme config
shadcn/ui	Latest	Dark theme, New York style
TypeScript	Strict	Built into Next.js 16
shiki	Latest	Syntax highlighting

Prerequisites

Requirement	Minimum	Recommended
macOS	Ventura 13.0+	Sequoia 15.0+
Architecture	Apple Silicon (M1+)	M2 Pro / M3+
Node.js	20.9	22 LTS
Python	3.10	3.12
Claude Code	Latest	Latest
Subscription	Claude Pro	Claude Max

Contributing

Fork, branch, make your changes, then:

1. bash tests/test_sanitization.sh must return clean
2. cd site && npm run build must produce zero errors
3. Open a PR

Public CI checks only the generic patterns in tests/sanitization-patterns.txt. It does not know private project codenames. Maintainers must run the external private profile on the exact commit before merging or publishing:

GRAVITY_PRIVATE_PATTERNS="${PRIVATE_PATTERNS_FILE:?Set the external private pattern file}" bash tests/test_sanitization.sh

Keep that profile outside the repository and out of PR-controlled workflows. A clean scan covers its configured patterns; it is not a guarantee against every possible disclosure.

License

MIT. See LICENSE.

---

Built by humans and AI, working together.