[!IMPORTANT]
Meta-Architect v0.1.12 is a production-grade skills line.
It is not a lightweight demo branch.
From v0.1.12 onward, the package is expected to ship with stable skill contracts, deterministic packaging, explicit release gates, and honest install and publish surfaces.
Overview
Meta-Architect is a workflow layer for teams that want architecture, evidence, review, and release discipline before build execution.
It adds:
- an architecture-first lane before implementation
- evidence-backed OSS selection through GitMCP-connected sources
- explicit logic, security, and DX/UX review gates
- a singular
$maestro bounded autonomous manager plus non-gating helper skills for alignment, diagnosis, test-first work, and cleanup
- installable skills and a reproducible package surface
[!NOTE]
Meta-Architect does not replace your coding runtime.
It wraps that runtime with architecture, evidence, gate enforcement, and release-sensitive workflow control.
Support
| Linux-native packages | .deb for Debian-family distros, .pkg.tar.xz for Arch-family distros, and .rpm for Fedora/openSUSE-style distros |
| npm package | @jstn-sdk/ma (fallback install path) |
| Helper command | ma (secondary support surface) |
| Runtime | Node.js >=20, npm @10 |
| Release line | v0.1.12 |
| License | MIT |
Screenshots
Prerequisites
- Node.js
>=20
- npm
>=10
- Git
- an MCP-capable coding runtime
- Codex for the recommended package-first path
- macOS, Linux, or WSL2 recommended
[!TIP]
The most reliable default environment is a Unix-like shell with Git, Node.js, and an MCP-capable runtime already configured.
Default Install Surfaces
Meta-Architect is a Codex-native session workflow. On Linux, the primary install surface is now distro-style package installation, not a standalone binary shell.
Debian, Ubuntu, Linux Mint, Pop!_OS
Download the GitHub release .deb asset and install it with your normal package command:
sudo apt install ./meta-architect_<version>_all.deb
Arch, Manjaro, EndeavourOS
Download the GitHub release pacman package asset and install it with:
sudo pacman -U ./meta-architect-<version>-1-any.pkg.tar.xz
Fedora, RHEL-family, openSUSE
Download the GitHub release RPM asset and install it with your distro-native command:
sudo dnf install ./meta-architect-<version>-1.noarch.rpm
sudo zypper install ./meta-architect-<version>-1.noarch.rpm
These packages install the Meta-Architect payload and expose ma / meta-architect, but the product still runs inside a Codex-native session. They do not create a separate desktop or terminal product.
npm fallback
npm i -g @openai/codex@latest @jstn-sdk/ma@latest
ma --madmax --high
npm uninstall -g @jstn-sdk/ma
npm uninstall -g @jstn-sdk/ma @openai/codex
What this assumes:
- Codex is installed globally
- Meta-Architect is installed globally as the skills/plugin package
- Meta-Architect installs its published skill surface into the active Codex home
- the product experience happens through the skill workflow inside Codex
[!IMPORTANT]
On Linux, distro-native package install is the default surface.
npm remains available as a fallback path.
Repository Branch Strategy
Meta-Architect’s repository workflow follows a stricter release posture focused on gated promotion:
main = release-facing protected branch
development = normal integration branch
feature/* = short-lived contribution branches
- contributors branch from
development
- normal PRs target
development
- only curated promotions move
development into main
[!CAUTION]
main is intended to be protected and exceptional.
Maintainers should stop bypass-pushing to main except for genuine emergency or admin recovery cases.
Setup
Package setup
Debian-family install:
sudo apt install ./meta-architect_<version>_all.deb
Arch-family install:
sudo pacman -U ./meta-architect-<version>-1-any.pkg.tar.xz
Fedora/openSUSE install:
sudo dnf install ./meta-architect-<version>-1.noarch.rpm
npm fallback:
npm i -g @openai/codex@latest @jstn-sdk/ma@latest
ma --madmax --high
npm uninstall -g @jstn-sdk/ma
npm uninstall -g @jstn-sdk/ma @openai/codex
This gives you:
- the installed Meta-Architect skill surface
- the canonical Meta-Architect skill entrypoints inside a Codex session
- the optional
ma helper command when a guided start is useful
Contributor setup: source checkout
Use this path only if you want to work on Meta-Architect itself.
git clone https://github.com/JustineDevs/meta-architect.git
cd meta-architect
npm install
npm link
npm link makes ma and meta-architect available from the local checkout.
Quick Start
1. Start Codex context if needed
ma --madmax --high
2. Start with the real usage-workflow prompt
Use the same operator shape defined in example/usage-workflow.md.
Quick-start prompt:
$maestro
Or start directly with:
$arch I want to build: [PROJECT IDEA]
Context:
- Product type: [web app / mobile app / API / marketplace / agent system / internal tool]
- Users: [who will use it]
- Core problem: [what problem it solves]
- Main features:
1. [feature one]
2. [feature two]
3. [feature three]
- Constraints:
- Budget: [low / medium / high]
- Team size: [solo / small / medium]
- Timeline: [e.g. 2 weeks MVP, 3 months beta]
- Preferred stack: [optional]
- Avoid: [optional]
- Quality priorities:
- [e.g. speed, low cost, security, DX, maintainability, scalability]
- Deployment target:
- [Vercel / Docker / VPS / AWS / GCP / local-first / hybrid]
Required output:
1. Problem framing
2. Recommended architecture
3. Stack decision with justification
4. System components and responsibilities
5. Data model and storage choices
6. Auth/security considerations
7. DX/UX considerations
8. Delivery plan for v0.1.12
9. Risks and trade-offs
10. Decision log
11. Exact next trigger to run after this
3. Run the full trigger sequence inside Codex
The singular umbrella in-session entry point is $maestro. It is the bounded autonomous manager for the in-session workflow. The package does not ship a separate $meta-architect skill.
The release-gated sequence stays fixed:
$arch
$sage
$flow
$vet
$vibe
$build
Optional publishable non-gating helper skills available around that sequence:
$align
$diagnose
$tdd
$cleanup
See example/usage-workflow.md for the full prompt templates for each step.
4. Secondary helper path
If you are working from a repository directly and need scaffolded local support files, use:
ma bootstrap
ma bootstrap --init-mcp
ma doctor
ma setup
ma
Recommended lazy-user path:
- run
ma bootstrap first to repair packaged assets, scaffold local runtime files, and verify the environment
- add
--init-mcp if you want starter GitMCP source files written into the local mcp/ folder when it is empty or invalid
- use
ma doctor later when you want a check-only readiness report without changing files
Expected output for ma setup:
meta-architect setup
====================
ready: .codex/agents
ready: .codex/prompts
ready: .ma/skills
ready: .ma/evidence
ready: .ma/context
ready: .ma/specs
ready: .ma/plans
ready: mcp
ready: docs
ready: docs/qa
ready: sprint
5. Configure GitMCP sources
Add real repository-backed endpoints in mcp/servers.json.
Example:
{
"category": "candidate",
"repo": "owner/repo",
"endpoint": "https://gitmcp.io/owner/repo"
}
Recommended source-selection posture:
- use packaged native references to narrow candidate families first
- map serious candidates to exact upstream GitMCP endpoints
- verify final choices against upstream repos and official docs before approval
Core discovery standard:
https://ossium.live/home
- use Ossium to discover trending OSS, curated repos, YC-backed repos, GSoC orgs, and contribution opportunities faster
https://trendshift.io/
- use Trendshift for rising GitHub engagement and topic-driven trend discovery
https://devhunt.org/
- use Dev Hunt for recently launched developer tools and current dev-tool discovery
https://libraries.io/
- use Libraries.io for package and dependency metadata, with caution because its public data is scraped and not validated/curated for accuracy
https://openhub.net/
- use Open Hub for project activity, contributor, popularity, and comparison signals
https://www.opensourceprojects.dev/
- use Open-source Projects for curated OSS discovery and detailed project writeups
- treat all of these as discovery acceleration, then convert promising finds into exact upstream GitMCP mappings and official-doc checks for
$sage
Useful native references:
skills/maestro/references/native-ingest-map.md
skills/sage/references/source-selection.md
skills/vet/references/security-playbooks.md
Canonical $sage order:
- Start with the upstream repo and official docs if you already know them.
- Use discovery accelerators only when you need help finding or narrowing candidates.
- Map selected candidates to exact upstream GitMCP endpoints.
- Verify against upstream repos and official docs before treating anything as approved evidence.
[!IMPORTANT]
Verified release evidence must come from repository-form GitMCP endpoints such as https://gitmcp.io/{owner}/{repo}.
A generic documentation endpoint such as https://gitmcp.io/docs does not count as VERIFIED evidence for build unlocking.
Discovery surfaces such as Ossium, Trendshift, Dev Hunt, Libraries.io, Open Hub, and Open-source Projects are not substitutes for upstream repo or official-doc verification.
6. Secondary helper flow outside Codex
If you need scripted repo-local validation rather than the interactive runtime workflow:
ma idea "Build a real-time collaborative whiteboard for product teams"
ma run '$arch'
ma run '$sage'
ma run '$flow'
ma run '$vet'
ma run '$vibe'
ma status
ma run '$build'
Expected status before the helper-path $build:
Meta-Architect Status
=====================
Idea: CLEAR
Architecture: APPROVED
Evidence: VERIFIED
Logic: GREEN
Security: GREEN
Experience: GREEN
Build: LOCKED
Next allowed triggers:
$build
Expected helper-path build output:
Build gate is green.
Suggested branches:
- feature/implementation
- feature/verification
Optional worktree commands:
git worktree add ../implementation feature/implementation
git worktree add ../verification feature/verification
7. Simple command guide
Meta-Architect has two surfaces.
- terminal helper commands
- in-session skills
The umbrella in-session entry point is $maestro. There is no separate shipped $meta-architect skill surface.
Terminal commands are normal shell commands you run in the terminal:
ma setup
ma init
ma idea "Build a product"
ma status
ma run '$arch'
In-session skills are prompts you use inside the Codex conversation after launch:
$maestro
$arch
$sage
$flow
$vet
$vibe
$build
$align
$diagnose
$tdd
$cleanup
Plain-language difference:
ma ... = helper commands in the terminal
$... = the product experience inside Codex
Autonomous-manager contract:
$maestro is the only umbrella in-session surface
- it manages the next allowed step and lane handoff, but it does not replace the gated outputs owned by
$arch -> $sage -> $flow -> $vet -> $vibe -> $build
$align, $diagnose, $tdd, and $cleanup are publishable helper skills that do not move release gates
What ma setup and ma init do:
- both currently do the same thing
- they create the local support files and folders
- they prepare
.ma/ runtime files such as context, specs, plans, evidence, and runbook files
- they do not run the skill workflow by themselves
What ma bootstrap does:
- checks whether
codex is callable
- repairs installed skills and support-bundle assets when possible
- runs local scaffold setup
- can seed starter MCP files with
--init-mcp when the local MCP config is empty or invalid
- reports
READY, READY_WITH_WARNINGS, or BLOCKED
What ma doctor does:
- runs the same environment checks without changing files
- prints the current readiness state and exact next step
What to use when:
- use Codex and run the skills in-session
- use
$maestro when you want Meta-Architect to choose the best next step for you or act as the bounded autonomous manager for the umbrella workflow
- use
$arch -> $sage -> $flow -> $vet -> $vibe -> $build inside the Codex session
- use
$align, $diagnose, $tdd, or $cleanup when a helper is enough and the release gate should stay where it is
- use
ma bootstrap when you want the lazy-user setup path
- use
ma doctor when you want a check-only environment report
- use
ma setup or ma init only when you want local scaffolding or scripted helper automation from the terminal
- use
ma sdk-path when you need the exact installed support-bundle path for packaged prompts, MCP files, sprint files, scripts, plugin metadata, or templates
Skill Surface
Meta-Architect’s in-session surface has three layers:
- umbrella autonomous manager:
$maestro
- fixed gated lanes:
$arch, $sage, $flow, $vet, $vibe, $build
- non-gating helper skills:
$align, $diagnose, $tdd, $cleanup
Helper skills are publishable surfaces, but they do not own release-state transitions.
Core Maintainers
Gated Lanes
$arch | Produce the first-pass architecture blueprint | decision entry | architecture_status = APPROVED |
$sage | Ground major choices in configured GitMCP evidence | evidence records | `evidence_status = VERIFIED |
$flow | Review baseline logic and state transitions | logic review entry | `logic_status = GREEN |
$vet | Run baseline security and dependency review | audit and CVE records | `security_status = GREEN |
$vibe | Review developer and user experience implications | DX/UX outcome record | `experience_status = GREEN |
$build | Unlock bounded build planning | build-ready decision + .ma/plans/build.md | build_status = READY |
Helper Skills
$align | Normalize terminology, tighten scope, and improve prompt or docs clarity | none |
$diagnose | Decompose blocked-lane symptoms into hypotheses and next probes | none |
$tdd | Lock behavior with regression-first or test-first scaffolding | none |
$cleanup | Simplify noisy output and run a final-pass anti-slop cleanup | none |
Gate Model
Meta-Architect is intentionally fail-closed.
CLEAR | enough input exists to proceed |
APPROVED | the architecture lane produced an acceptable first-pass blueprint |
VERIFIED | live evidence was grounded through approved GitMCP sources |
PARTIAL | evidence is configured but live proof is incomplete or unavailable |
GREEN | the current baseline review passed |
RED | the lane is blocked or failed |
WAIVED | the lane was intentionally waived with a recorded reason |
LOCKED | downstream work is not allowed yet |
READY | the next gated step is allowed |
[!CAUTION]
$build must stay locked until the upstream release state in .ma/release.json satisfies the gate contract.
Meta-Architect is designed to stop on blockers rather than silently continue.
Rich runtime artifacts live in .ma/context/, .ma/specs/, .ma/plans/, and .ma/runbook.md.
Release and Packaging
Meta-Architect has three related but different distribution surfaces.
| Linux native packages | distro-managed install assets for Debian-family and Arch-family environments | npm run linux:packages:build and GitHub release assets |
| npm package | public package containing the installable Meta-Architect skills/plugin system, docs, scripts, and canonical skills | npm publish or npm pack |
| skills bundle | narrower tarball containing skills/ only | npm run skills:pack |
Required packaging commands:
npm run skills:manifest
npm run skills:validate
npm run skills:pack
npm run skills:install -- --path ./dist/installed-skills
npm run linux:packages:build
npm run release:assets
npm run pack:inspect
Pre-publish rules:
skills/index.json must be current
npm run skills:validate must pass
dist/meta-architect-skills.tgz must exist
dist/meta-architect_<version>_all.deb must exist for Debian-family installs
dist/meta-architect-<version>-1-any.pkg.tar.xz must exist for Arch-family installs
dist/meta-architect-<version>-1.noarch.rpm must exist for Fedora/openSUSE-style installs
npm pack --dry-run must show only intended public files
- docs must match the real skills/plugin and release behavior
Release lane discipline:
- stable versions publish to npm
latest
- prerelease versions such as
0.2.0-beta.1 must publish with an explicit dist-tag such as beta
- alternate lanes such as
next, beta, and canary must never overwrite latest
Maintainer version-bump flow:
- Bump the package with
npm version <version> --no-git-tag-version
- Update
CHANGELOG.md, RELEASE.md, and docs/qa/release-readiness-<version>.md
- Run
npm run release:verify
- Run
npm run release:check
- Create and push tag
v<version>
- Build and smoke-check the Linux native packages with
npm run linux:packages:build, npm run linux:packages:smoke, and npm run release:assets on Linux
- Preferred publish path: publish from
.github/workflows/npm-publish.yml on a supported cloud runner so provenance can be generated
- Local shell fallback when not publishing from GitHub Actions or GitLab CI/CD:
- Stable publish:
npm publish --access public
- Prerelease publish:
npm publish --access public --tag <lane>
- Verify publish state with
npm view @jstn-sdk/ma version dist-tags time --json
- Confirm the GitHub release includes
dist/meta-architect-skills.tgz, meta-architect_<version>_all.deb, meta-architect-<version>-1-any.pkg.tar.xz, and meta-architect-<version>-1.noarch.rpm
Provenance note:
npm publish --provenance requires a supported cloud CI/CD provider
- a local shell publish will fail with
Automatic provenance generation not supported for provider: null
- use the repository publish workflow when provenance is required
Release automation:
npm run release:sync bumps and synchronizes the active release line only when watched release-relevant files changed
npm run release:advance force-bumps the next patch line and rewrites the same version-bearing files
.github/workflows/release-sync.yml runs the sync path on main pushes that touch watched release-relevant paths
.github/workflows/release-advance.yml runs after a published GitHub release and advances the repo to the next patch line automatically
[!CAUTION]
Do not claim npm, GitHub release, or any other publish channel until that channel has actually succeeded.
Release documentation must match reality, not intent.
Package Surface
| Included | bin/, skills/, docs/, scripts/, index.js, README.md, LICENSE |
| Excluded | .ma/ runtime state, context, specs, plans, logs, caches, and temp install outputs |
Repository Structure
| Path | Responsibility |
.codex/ | runtime prompts, hooks, and repo guidance |
skills/ | canonical public skill contracts |
plugins/meta-architect/ | plugin-oriented distribution surface |
docs/ | installation, publishing, and release documentation |
missions/ | reproducible scenario-driven workflows |
mcp/ | GitMCP endpoint and collection configuration |
scripts/ | validation, packing, and install helpers |
sprint/ | human-readable phased workflow documents |
Documentation
Release Hygiene
[!WARNING]
Runtime .ma logs, state, tmp, and cache files must not be shipped.
Public docs must match actual package behavior.
Publish statements must match reality.
Skill contracts must stay aligned across canonical and plugin-facing copies.
License
MIT