English · Русский · 简体中文 · Español · हिन्दी

EmbrAIon

AI-First Engineering System by GORYNED

Where sparks become AI-built products

EmbrAIon is a portable AI-First Engineering System for organizing AI-assisted software engineering around explicit roles, reusable skills, workflow orchestration, model routing, access control, validation, review, security, learning, and project overlays.

It is designed to sit above individual languages and frameworks. A Unity/C# project, a Python service, a web application, or another software repository can use the same Core and add only project-specific knowledge and rules.

What EmbrAIon does

EmbrAIon separates the engineering system into independent dimensions:

• Agent — who is responsible: Lead, Worker, Reviewer, Architect, Analyst, Validator, Researcher, Steward.
• Skill — how a repeatable class of work is performed.
• Rule — what is required, forbidden, or protected.
• Workflow — in what order capabilities are composed.
• Routing — which access profile, model tier, host, and provider may execute the work.
• Adapter — how canonical capabilities are represented in Codex, Copilot, Claude Code, providers, or a host-neutral Portable bundle.
• Tool — deterministic executable behavior such as validation, security scanning, worktree management, synchronization, and diagnostics.
• Eval — whether AI behavior actually follows the intended engineering contract.

core/catalog.yaml is the discovery index. Instead of loading the entire framework for every task, EmbrAIon can load only capabilities whose triggers match the current work.

Data classes

EmbrAIon uses three canonical data classes:

Unknown or ambiguous classification fails closed. Model choice never expands access or privacy eligibility.

Supported host adapters

Current adapters:

• Codex — native model catalog, model/effort route mapping, generated project agents and config.
• GitHub Copilot — advisory model catalog and generated custom-agent projection.
• Claude Code — Claude model catalog and generated subagent projection.
• Portable — host-neutral installable capability bundle.
• Providers — direct API model catalogs for OpenAI, Anthropic, Google, and DeepSeek, plus transport metadata.

Core remains model-neutral. Current model identities and host selectors live only in adapters.

Installation

Install once per machine

EmbrAIon is distributed through PyPI. For normal use, install the CLI once on each Windows or macOS computer with pipx. A source checkout or git clone is not required.

A regular pip installation is also supported when you intentionally manage the Python environment yourself, but pipx is the recommended CLI installation path.

Windows

EmbrAIon requires Python 3.11 or newer.

1. Check whether Python is already available:

py --version

If py is unavailable or reports a version older than 3.11, install a current Python 3 release from the official Python downloads for Windows, then reopen PowerShell.

2. Install pipx and add its command directory to PATH:

py -m pip install --user pipx
py -m pipx ensurepath

3. Close and reopen PowerShell so the updated PATH is loaded, then install EmbrAIon:

pipx install embraion

If you deliberately use a managed Python environment instead of pipx, this is also supported:

py -m pip install embraion

macOS

If Homebrew is already installed, the simplest path is:

brew install pipx
pipx ensurepath

Open a new Terminal window, then install EmbrAIon:

pipx install embraion

Without Homebrew, first check Python:

python3 --version

If Python is missing or older than 3.11, install a current Python 3 release from the official Python downloads for macOS. Then install pipx:

python3 -m pip install --user pipx
python3 -m pipx ensurepath

Open a new Terminal window and run:

pipx install embraion

A regular python3 -m pip install embraion is also supported when you manage the Python environment yourself.

Verify

embraion --version
embraion validate
embraion doctor
embraion help
embraion status
embraion cache list

embraion validate validates the framework data bundled with the active EmbrAIon installation. embraion doctor prints a human-readable diagnostic report by default and automatically detects whether the current directory is inside a Git or EmbrAIon project. Outside a project it performs installation/framework diagnostics only; it does not recursively scan your home directory or another arbitrary folder. Use embraion doctor --json when structured machine-readable output is required.

Upgrade

pipx upgrade embraion

The global pipx installation makes the embraion command available from any project on that computer. It does not automatically enable EmbrAIon in every repository or modify repositories in the background.

Add EmbrAIon to each project

Each repository opts in explicitly. From the project root:

cd /path/to/your/project
embraion init

By default, init uses the directory name as the project name. Use --name MyProject only when you want to override it.

This creates:

.embraion/
└── project.yaml

The Project Overlay records the EmbrAIon repository/version declaration and project-specific configuration in version control.

Install a host projection

Install the projection for each AI client used by the repository.

Codex:

embraion install --host codex --destination .

GitHub Copilot:

embraion install --host copilot --destination .

Claude Code:

embraion install --host claude-code --destination .

Portable bundle:

embraion install --host portable --destination ./vendor/embraion

If a repository uses multiple clients, run the corresponding install command once for each one. Existing generated files are protected by default; use --force only when intentionally replacing them.

Automatic project runtime resolution

v0.2.0 adds automatic per-project version resolution.

For ordinary commands, the global embraion launcher walks upward from the current directory until it finds the nearest .embraion/project.yaml. It then reads framework.version:

global embraion launcher
        ↓
nearest .embraion/project.yaml
        ↓
framework.version
        ↓
exact cached runtime for that project

If the pinned version differs from the launcher version, EmbrAIon installs the exact PyPI distribution into an isolated cache under:

~/.embraion/versions/<version>/

The first command for a version may download it from PyPI. Later commands reuse the cached runtime. Projects can therefore use different EmbrAIon releases on the same Windows PC or Mac without separate global installations.

Legacy project overlays created by v0.1.0 may contain 0.1.0-dev; the new resolver treats that legacy pin as the published 0.1.0 distribution.

embraion init and embraion update intentionally run in the globally installed launcher instead of delegating to the old project runtime:

• embraion init opts a repository into the current launcher version.
• after pipx upgrade embraion, embraion update moves only the current project to the new launcher version.
• embraion update --framework-version X.Y.Z explicitly pins a chosen release; the next ordinary command resolves it automatically.

For framework development, setting EMBRAION_HOME keeps using the explicitly selected framework checkout. Automatic resolution can also be disabled explicitly with EMBRAION_DISABLE_VERSION_RESOLUTION=1.

Inspect launcher, project pin, and runtime cache

embraion status

The status report shows the globally installed launcher version, nearest project, project pin, resolved runtime, whether the pinned runtime is already cached, the cache location, and detected host projections. Use embraion status --json for automation.

Inspect cached project runtimes:

embraion cache list

Clean invalid or stale cache entries with a dry run first:

embraion cache prune

Optionally include valid runtimes not used for a chosen number of days:

embraion cache prune --older-than 90
embraion cache prune --older-than 90 --apply

The current launcher version and the current project's resolved version are protected from age-based pruning.

How a task flows through EmbrAIon

A substantial task conceptually follows this path:

User goal
  ↓
Project overlay + Core catalog
  ↓
Relevant rules / agents / skills / workflows
  ↓
Data class + access profile + complexity
  ↓
Host adapter + model route
  ↓
Implementation
  ↓
Validation
  ↓
Independent review
  ↓
Verification
  ↓
Delivery / human merge gate

For substantial specification-driven work, Spec Kit is recommended as an independent companion. It refines planning and specification but does not replace Core rules, project truth, compatibility contracts, or validation evidence.

CLI

Framework and project

embraion init
embraion install
embraion update
embraion sync
embraion validate
embraion doctor

Use embraion help for the categorized command catalog. Use embraion help <command> (including nested paths such as embraion help cache prune) or embraion <command> --help for detailed command-specific help.

Routing and runtime state

Resolve a route:

embraion route --host codex --route-class strong --data PRIVATE

Create a bounded dispatch plan:

embraion dispatch \
  --task "Implement feature" \
  --role worker \
  --host codex \
  --route-class economy-write \
  --data PRIVATE \
  --access write \
  --owned-path "src/**"

Writable dispatch planning requires explicit owned paths and refuses the stable main/master branch. The plan and privacy-safe routing telemetry are written under .embraion/state/.

Start normalized session state:

embraion session start \
  --session-id task-001 \
  --task "Implement feature" \
  --role lead \
  --host codex \
  --access plan

Inspect or update it:

embraion session show
embraion session set --state review --validation passed

The host adapter performs actual AI execution; EmbrAIon owns canonical routing, generated agent definitions, access/ownership boundaries, dispatch plans, normalized state, and privacy-safe operational telemetry.

Security

embraion security scan --path .

The scanner checks for high-risk configuration problems such as possible embedded credentials and policy drift.

MCP inventory

embraion mcp inventory

This creates a normalized privacy-safe inventory under .embraion/state/. Environment-variable names may be recorded; secret values are never intentionally persisted.

Worktrees

embraion worktree list
embraion worktree create ai/my-task
embraion worktree gc
embraion worktree salvage /path/to/worktree

gc is dry-run by default and only considers directly proven, clean, unlocked worktrees. Use --apply explicitly to remove safe candidates.

Learning

Record repeated evidence:

embraion learning observe \
  --id repeated-review-gap \
  --kind repeated-failure \
  --target-type skill \
  --target-id review \
  --summary "Repeated review gap"

Promotion is gated:

observe → accumulate evidence → propose → approve → promote

Promotion never edits Core automatically. A promoted candidate still requires an ordinary reviewed engineering change.

Behavioral evals

Run an eval case:

embraion eval run \
  --case reviewer-readonly \
  --record path/to/execution-record.json

Create and compare baselines:

embraion eval baseline --reports build/evals --output baseline.json
embraion eval compare --baseline baseline.json --reports build/evals

Generate adapter projections

Generate all supported projections without installing them into a project:

embraion sync --host all --output build/generated --force

Generated outputs are disposable projections. Canonical policy always remains in core/.

Repository layout

brand/         Brand specification and README assets
core/          Canonical rules, agents, skills, workflows, routing, knowledge
adapters/      Codex, Copilot, Claude Code, Portable, provider integrations
tools/         CLI, runtime, learning, security, MCP, worktree, validation, sync
schemas/       Machine-readable contracts
templates/     Project overlay templates
docs/          Canonical English documentation
localization/  Russian, Simplified Chinese, Hindi, and Spanish translations
tests/         Deterministic unit/integration tests
evals/         Behavioral cases, baselines, graders, fixtures, reports
examples/      Reference integrations

Validation and CI

Every push and pull request is intended to run:

• schema and catalog validation;
• localization parity;
• unit and integration tests;
• Linux, Windows, and macOS compatibility on Python 3.11 and 3.14;
• project-version resolver E2E checks on each operating system;
• security scanning;
• generation of all host projections;
• behavioral eval smoke tests.

Tagged releases generate source, Codex, Copilot, Claude Code, and Portable archives.

Documentation

Canonical documentation: docs/

Translations:

• Русский
• 简体中文
• Español
• हिन्दी

License and brand

EmbrAIon source code and documentation are licensed under the MIT License, except where a file or directory explicitly states otherwise.

The EmbrAIon and GORYNED names, logos, wordmarks, visual marks, and the files under brand/assets/ are not licensed under MIT. The MIT License does not grant trademark or brand-identity rights. See TRADEMARKS.md for the canonical policy.

Current status

EmbrAIon is pre-stable. The architecture and first executable CLI are in place, but the public compatibility contract is not frozen yet.

Before the first stable release, model catalogs, generated host projections, validation coverage, security rules, installation behavior, and release packaging may still evolve.

EmbrAIon · AI-FIRST ENGINEERING SYSTEM · by GORYNED

_{Last updated: 2026-09-23 22:00 UTC}