Full Stack HQ

One engineering core. Every AI coding agent on your team.

Permission-first engineering rules, agents, skills, and workflows for
Claude Code, OpenAI Codex, and Google Antigravity IDE, plus Cursor,
GitHub Copilot, and Gemini CLI through a repository's AGENTS.md.

Quick start · Architecture · What's inside · Setup guide · FAQ · Comparison · Türkçe · Launch playbook

The short version: keep the engineering policy in one readable source, then render only the host-specific details each AI tool needs.

Full Stack HQ is an open-source, permission-first engineering configuration kit for AI coding agents. One shared set of rules, 10 specialist agents, 28 Agent Skills, and 10 workflows is rendered into the native formats of Claude Code (CLAUDE.md and a plugin), OpenAI Codex (AGENTS.md and TOML agents), and Google Antigravity IDE (GEMINI.md and workflows).

Use it when you want an AI coding agent to plan before it edits, wait for an explicit approval phrase such as PLAN APPROVED, and report what it actually verified, and when you want the same rules in every coding host you use without maintaining hand-written copies. Common questions are answered in the FAQ.

It is not a SaaS product, agent runtime, application framework, or hard security boundary. The host still controls tools, approvals, sandboxing, network access, and filesystem behavior.

Start here

Claude Code: two commands, no clone

Run these inside Claude Code:

/plugin marketplace add sabahattink/antigravity-fullstack-hq
/plugin install full-stack-hq@full-stack-hq

Start a new session. The shared rules load at session start, the 10 specialist agents and 28 skills become available, and the 10 workflows appear as namespaced commands such as /full-stack-hq:plan. Upgrade with claude plugin update full-stack-hq@full-stack-hq and remove it with claude plugin uninstall full-stack-hq@full-stack-hq.

Use either the plugin or the Claude installer (--only-claude), not both; together they load the same rules, agents, and skills twice.

Any host: one line, no clone

Preview first. --dry-run shows every file that would be written and changes nothing:

# macOS / Linux
curl -fsSL https://raw.githubusercontent.com/sabahattink/antigravity-fullstack-hq/main/bootstrap.sh | bash -s -- --dry-run

# Windows PowerShell
& ([scriptblock]::Create((irm https://raw.githubusercontent.com/sabahattink/antigravity-fullstack-hq/main/bootstrap.ps1))) -DryRun

Drop --dry-run / -DryRun to install, or add any installer option such as --only-codex / -OnlyCodex and --backup / -Backup. The bootstrap makes a temporary shallow checkout, runs the regular installer with your options, and deletes the checkout. Pin a release with FULL_STACK_HQ_REF=v1.3.0 (Bash) or -Ref v1.3.0 (PowerShell). Read bootstrap.sh or bootstrap.ps1 before piping either into a shell.

Your repository: shared rules for the whole team

Add the rules to a project instead of your home directory. Everyone who opens the repository gets them, whatever coding agent they use:

# macOS / Linux, from a checkout of this repository
bash install.sh --project ../your-repo

# Windows PowerShell
.\install.ps1 -Project ..\your-repo

The installer only manages the block between <!-- full-stack-hq:start --> and <!-- full-stack-hq:end -->, so instructions your repository already has are kept. Re-run the command to update the rules; add --uninstall / -Uninstall to remove the block and get the original files back. The one-line bootstrap accepts the same options.

Try it from a checkout

Preview the Codex adapter without changing your real host configuration:

git clone https://github.com/sabahattink/antigravity-fullstack-hq.git
Set-Location antigravity-fullstack-hq
.\install.ps1 -OnlyCodex -DryRun -TargetRoot .\.demo-home

macOS/Linux:

git clone https://github.com/sabahattink/antigravity-fullstack-hq.git
cd antigravity-fullstack-hq
bash install.sh --only-codex --dry-run --target-root ./.demo-home

This preview validates the source and shows the host files that would be generated while keeping the real home directory untouched. When the result looks right, continue with the full setup guide, or test a different host with --only-claude / --only-antigravity.

For the three reproducible first-run recipes, see First-run recipes.

Report an installation result · Join the discussion · Record the 45-second demo

The experience

The repository is designed to feel like a small engineering control plane:

$ bash install.sh --only-codex --backup

FULL STACK HQ / CODEX ADAPTER
────────────────────────────────────────────────────────
 source      ✓  validated
 rules       →  ~/.codex/AGENTS.md
 agents      →  ~/.codex/agents/*.toml       10 specialists
 skills      →  ~/.agents/skills/            28 + 10 workflow bridges
 result      ✓  ready for a new Codex session

$ .\install.ps1 -OnlyClaude -Backup

The terminal above is a representative flow. Run the validator and installer from a local checkout to see the exact output for your machine.

Why this exists

Architecture

The core is the deep module: callers learn one stable engineering policy, and each adapter carries only the details that genuinely vary. The existing top-level agents/, skills/, and workflows/ paths remain public so current customizations and links continue to work.

The source-to-host contract

The Codex adapter uses its current native formats: layered AGENTS.md instructions, Agent Skills for reusable procedures, and TOML custom agents with name, description, and developer_instructions.

Portable package

The repository root is also a Claude Code plugin marketplace. .claude-plugin/ describes the plugin, hooks/hooks.json loads the shared rules at session start, and the plugin reads the canonical agents/, skills/, and workflows/ directly, so there is no second copy to drift.

The repository root also contains a portable plugin.json manifest beside the canonical skills/ directory. Plugin-aware OpenAI hosts can consume that package boundary without a second copy of the shared guidance. The regular installers remain the complete path when you also want global rules, native agents, Antigravity workflows, Codex TOML agents, backups, or dry-runs.

Read the portable package guide for the boundary, release rules, and local validation checklist.

Quick start

Choose one or more hosts. The default installs all three integrations.

macOS / Linux

$ git clone https://github.com/sabahattink/antigravity-fullstack-hq.git
$ cd antigravity-fullstack-hq
$ bash install.sh

Windows PowerShell

PS> git clone https://github.com/sabahattink/antigravity-fullstack-hq.git
PS> Set-Location antigravity-fullstack-hq
PS> .\install.ps1

Install one host only

# macOS / Linux
$ bash install.sh --only-codex
$ bash install.sh --only-claude
$ bash install.sh --only-antigravity

# Windows PowerShell
PS> .\install.ps1 -OnlyCodex
PS> .\install.ps1 -OnlyClaude
PS> .\install.ps1 -OnlyAntigravity

Safe upgrade controls

For a conservative upgrade:

# macOS / Linux
$ bash install.sh --dry-run
$ bash install.sh --force --backup

# Windows PowerShell
PS> .\install.ps1 -DryRun
PS> .\install.ps1 -Force -Backup

Existing global instruction files are confirmed interactively. Existing agent and skill files are kept unless Force is selected. Restart the host application and start a new conversation after installation.

Diagnose before installing

The read-only doctor checks the repository, adapter validator, host commands, target paths, and Codex override precedence. Missing host files are reported as warnings so it is useful both before a first install and after an upgrade.

# Windows PowerShell
PS> .\scripts\doctor.ps1
PS> .\scripts\doctor.ps1 -OnlyCodex -TargetRoot .\tmp\host-home

# macOS / Linux
$ bash scripts/doctor.sh
$ bash scripts/doctor.sh --only-codex --target-root ./tmp/host-home

Use -Strict / --strict when warnings should make the diagnostic fail.

First run

Use a small, reversible prompt to check that the configuration is loaded:

Create a React component called UserCard

The intended engineering loop is:

inspect → scope → approval → implement → verify → report

This is instruction-level guidance. It does not intercept shell commands or guarantee runtime permissions.

Help us test the host adapters

The most useful contribution is a real installation result, including a successful dry-run or a small host-specific difference. Open the installation feedback form with:

• the host and operating system;
• the command and Full Stack HQ version or commit;
• the expected and actual generated paths;
• a redacted log excerpt or screenshot, with tokens and private paths removed.

If you are comparing more than one AI coding host, continue the conversation in GitHub Discussions so repeated questions can become documentation improvements.

What's inside

Specialist agents

Canonical skills

The 28 skills are grouped around the work they support:

Workflow catalog

The canonical workflows remain in workflows/. They are also rendered as skills so the same procedures can be used across all three hosts:

The exact invocation is host-native: Antigravity and Claude commonly expose slash commands, while Codex exposes skill selection and $skill-name invocation.

Compatibility and migration

• Existing top-level agents/, skills/, and workflows/ paths remain intact.
• New Antigravity installations use documented ~/.gemini/config/ locations.
• An existing ~/.gemini/antigravity/ tree is refreshed as a compatibility bridge unless --no-legacy-paths / -NoLegacyPaths is selected.
• New installations do not create the legacy Antigravity tree.
• Codex user skills are installed under ~/.agents/skills/.
• Codex global rules and custom agents use CODEX_HOME when it is set, otherwise the active ~/.codex/ directory.
• If ~/.codex/AGENTS.override.md exists, Codex prioritizes it. The installer warns about the override and never overwrites it.
• The root plugin.json packages the existing skills/ tree for plugin-aware hosts; it does not create a parallel source tree.

There is intentionally no separate Codex workflow directory. The canonical workflow bodies are converted to Agent Skills because skills are the current shareable workflow format.

Verification

Validate the repository before installing or proposing a release:

# Windows PowerShell
PS> .\scripts\validate.ps1
PS> .\install.ps1 -Check

# macOS / Linux
$ bash scripts/validate.sh
$ bash install.sh --check
$ bash scripts/smoke-test.sh

# Windows PowerShell
PS> .\scripts\smoke-test.ps1

The validators check:

• YAML frontmatter and name/path consistency.
• Skill and workflow metadata.
• Canonical content for accidental host-specific leakage.
• Codex TOML agent generation.
• Workflow-to-skill conversion counts.
• Antigravity rule-size limits and workflow/skill name collisions.
• Cross-platform renderer parity.

The smoke tests then install all three adapters into a disposable temporary home and verify the generated rules, agents, skills, and workflows without touching the real user configuration.

The repository also runs least-privilege validation on Windows and Ubuntu for every push to main and every pull request.

Documentation

• Setup guide — installation, upgrades, migration, and troubleshooting.
• FAQ — short answers to the most common questions.
• Comparison — how Full Stack HQ differs from rule-sync tools and skill frameworks.
• Türkçe özet — Turkish summary.
• First-run recipes — safe previews, isolated tests, and feedback.
• Customization guide — extend the core without creating drift.
• Contributing guide — change discipline and verification.
• Core and adapter design — the source-of-truth model.
• Portable plugin package — package boundary and release rules.
• Launch announcement pack — reviewable copy for GitHub and social channels.
• Launch playbook — channel sequence, demo brief, and measurement plan.
• Installation feedback — structured host-specific reports.
• Security policy — reporting and security boundaries.
• Changelog — release history.

Contributing

Please read docs/CONTRIBUTING.md before opening a pull request. Changes should remain focused on the core, adapters, installers, validators, skills, agents, workflows, and documentation.

License

MIT License. Maintained by Sabahattin Kalkan.