plato-core

Plato Engine Block — a sub-400-line room runtime for agent-space interaction.

Foundation types and mesh registry for the PLATO (Protocol for Layered Agent Tile Orchestration) ecosystem. Zero external dependencies — everything else builds on top.

Install

pip install plato-core

Quick Start

from plato_core import TrainingTile, TileType, content_hash, LamportClock

# Create a tile — the atomic unit of work in PLATO
tile = TrainingTile(
    tile_id="chk-001",
    tile_type=TileType.CHECKPOINT,
    name="my-model-v1",
)

# Lamport clock for causal ordering across distributed rooms
clock = LamportClock()
tile.lamport = clock.tick()

# Content-addressable hash (SHA-256)
h = content_hash(b"some weights data")
print(h)  # 64-char hex string

The Mesh Registry

PLATO rooms auto-discover each other through the mesh registry. Any installed package that declares an entry point is found automatically:

from plato_core import MeshRegistry

registry = MeshRegistry()
matchers = registry.get_matchers()      # from plato-matcher
compressors = registry.get_compressors()  # from plato-compress
trainers = registry.get_trainers()       # from plato-training

Registering a Plugin

1. Add plato-core as a dependency
2. Declare an entry point in pyproject.toml:

[project.entry-points."superinstance.plugins"]
my-plugin = "my_package._mesh:register"

3. Implement the register function:

# my_package/_mesh.py
def register(registry):
    registry.register("matchers", "my_matcher", lambda: MyMatcher)

That's it. The mesh finds it automatically.

Key Concepts

Tiles are the atomic unit of work — signed, content-addressed artifacts that flow between rooms. Every tile has a lifecycle (Active → Superseded → Retracted) and a Lamport timestamp for ordering.

Rooms are agents. Each agent is a PLATO room that produces and consumes tiles. Rooms discover each other through the mesh registry.

Lamport clocks provide causal ordering without a central coordinator. When two rooms exchange tiles, they merge clocks — enabling distributed ordering.

What's Provided

• TrainingTile, TileType — tile data model
• TileLifecycle — Active / Superseded / Retracted
• LamportClock — distributed causal ordering
• TrainingConfig, AdapterConfig, TrainingMetrics — ML config types
• content_hash() — SHA-256 content addressing
• MeshRegistry — auto-discovery of ecosystem packages

Architecture

plato-core (this)       ← base types + mesh registry
  ├── plato-types       ← protocol-level type definitions
  ├── plato-training    ← training rooms, micro models
  ├── plato-mcp         ← MCP server (expose rooms as tools)
  ├── plato-matcher     ← semantic matching
  ├── plato-compress    ← compression / quantization
  └── ...               ← any future PLATO package

Documentation

• Mesh Architecture — how the registry and auto-discovery work
• docs/README.md — full documentation index

PLATO Engine Block Family

Plato Core is the Python foundation layer of the broader PLATO ecosystem:

Specs & Guides:

• 📜 PLATO Wire Protocol
• 📖 PLATO Master Guide
• 🗺️ PLATO Ecosystem Map

Related Repos

• plato-types — Core tile protocol types (this package re-exports them)
• plato-training — Training rooms and micro models
• plato-mcp — Expose PLATO rooms as MCP tools
• cocapn-plato — Full Cocapn PLATO integration (SDK + server)
• plato-room-musician — Sonify fleet activity via MIDI
• cocapn-glue-core — Binary wire protocol for fleet communication
• penrose-memory — Aperiodic memory palace for AI agents