Plugin Packages¶

G6 distributes its components as cluster-based Poetry packages downloadable from the G6 server. This keeps the core install small (~15 MB) while letting users add capability clusters on demand.

Design Principles¶

• Core under 200 MB — the base install includes only essential components
• Cluster granularity — each package groups related components, not individual bricks
• Poetry-managed — standard poetry add g6-<cluster> workflow
• Server-hosted — packages distributed from the G6 package server, not PyPI
• Tier enforcement at import — components check the user's license tier before exposing tools

Package Catalog¶

Dependency Graph¶

g6-core (required by all)
  ├── g6-safety
  │     ├── g6-planning
  │     ├── g6-jobs
  │     └── g6-optim
  ├── g6-formal
  ├── g6-ml
  ├── g6-dl
  ├── g6-context
  ├── g6-synthesis
  ├── g6-agents
  ├── g6-data
  ├── g6-media
  ├── g6-cogarch
  ├── g6-infra
  ├── g6-viz
  └── g6-grounding

All packages depend on g6-core. Some have additional inter-package dependencies:

Download URL Scheme¶

Packages are hosted on the G6 package server:

https://packages.g6solver.com/simple/g6-<cluster>/

Poetry configuration:

[[tool.poetry.source]]
name = "g6"
url = "https://packages.g6solver.com/simple/"
priority = "supplemental"

Install example:

poetry source add g6 https://packages.g6solver.com/simple/
poetry add g6-core g6-ml g6-context --source g6

Tier Enforcement¶

Each package checks the user's license tier at import time. If a component requires a higher tier than the user's license allows, the import succeeds but tool invocation returns a tier-upgrade message.

# Enforcement happens in the MCP gateway layer
from mvp.core.tier import check_tier

@server.tool()
async def decompose_goal(goal: str) -> dict:
    check_tier("goal_engine", required_tier=1)  # raises TierError if insufficient
    ...

Server-Side-Only Assets¶

Grounding Corpus¶

The grounding knowledge base corpus is over 4 GB and contains proprietary copyrighted material. It is:

• Never distributed in any package
• Never downloadable by end users
• Hosted exclusively on the G6 server
• Accessed only via the grounding MCP API

The g6-grounding package contains the retrieval and embedding code but not the corpus data.

Client (g6-grounding installed)
  └── MCP call: grounding_search(query="...")
        └── G6 Server
              └── FAISS index + corpus (4GB+, server-only)
              └── Returns: search results via MCP response

Future Server-Side Assets¶

Any future proprietary data assets (model weights, licensed datasets, curated knowledge bases) follow the same pattern: retrieval code in the package, data on the server.

Poetry Project Structure¶

Each package is a Poetry project under projects/:

projects/
  g6_core/
    pyproject.toml      # [tool.poetry] name = "g6-core"
    g6_core/__init__.py
  g6_ml/
    pyproject.toml      # [tool.poetry] name = "g6-ml"
    g6_ml/__init__.py
  ...

Each pyproject.toml declares its brick dependencies via the Polylith plugin and its inter-package dependencies:

[tool.poetry]
name = "g6-ml"
version = "1.0.0"
packages = [{include = "mvp/adapt_sklearn", from = "../../components"}]

[tool.poetry.dependencies]
python = "^3.12"
g6-core = "^1.0"
scikit-learn = "^1.5"
pygad = "^3.0"

External Dependencies¶

Some packages require external binaries or services not bundled in the package:

Users must install and configure these independently. Each component degrades gracefully when its external dependency is unavailable.

See Also¶

• Deployment -- server deployment architecture
• Component Model -- how components are structured
• Polylith -- workspace layout and brick management
• Tools by Tier -- tier access matrix