CheckMate Agent System

Chess-governed multi-agent planning for Claude Code, Codex, npm/npx, and Python.

CheckMate is an open source multi-agent orchestration framework and prompt-engineering system built around chess-piece roles. It ships as a Claude Code skill, Codex skill, npm CLI, npx package, and Python runtime for structured, constraint-based planning.

A chess-based system for multi-agent planning.

Quick Start • npx CLI • Skill Package • Runtime • Architecture • npx Reference

---

Status & Scope

CheckMate is alpha (0.1.0). Read this before you wire it into anything.

• The Python runtime is a deterministic governance scaffold. It does not call an LLM. Bring your own model integration in your own agent loop.
• The npm CLI emits a planning skeleton (text or JSON) with a naive hard-constraint contradiction heuristic. It is not a full runtime; the Python package is where the conflict resolver and message-bus topology live.
• The skill bundle is a prompt-engineering frame meant to be loaded into Claude Code or Codex sessions. It shapes how a model reasons; it is not itself a model.
• "Agent" here means a typed role with permissions and message routing, not an LLM-backed actor.

---

Quick Start

Step 1: Run with npx

npx checkmate-agent-system@latest plan \
  --goal "Build a migration plan" \
  --constraint "must preserve uptime" \
  --constraint "deadline is 2 weeks"

JSON output:

npx checkmate-agent-system@latest plan \
  --goal "Build a migration plan" \
  --constraint "must preserve uptime" \
  --format json

Step 2: Install the skill

Codex:

npx checkmate-agent-system@latest install codex

Claude Code user scope:

npx checkmate-agent-system@latest install claude

Claude Code project scope:

npx checkmate-agent-system@latest install claude-project

Step 3: Use it in-session

In Claude Code or Codex, mention CheckMate or the chess pieces in your prompt — the skill auto-loads on relevance based on its SKILL.md description. There is no $name invocation syntax.

Use CheckMate to plan this feature with King, Queen, Rooks, Bishops, Knights, and Pawns.

That gives you a governed loop:

King -> Queen -> Bishops -> Rooks -> Knights -> Queen -> King -> Pawns

---

What Is CheckMate?

CheckMate is a general-purpose agent system where each chess piece is a constraint-based reasoning primitive:

Piece	Runtime role	Authority
King	Governance	Objective, constraints, success criteria, final approval
Queen	Orchestration	Decomposition, assignment, synthesis, sequencing
Rooks	Validation	Logic, dependencies, resources, timeline, vetoes
Bishops	Analysis	Causality, patterns, implications, tradeoffs
Knights	Exploration	Alternatives, workarounds, recovery paths
Pawns	Execution	Atomic tasks, observations, evidence reports

The core idea:

• who thinks
• how they think
• what they are allowed to influence

npx CLI

CheckMate exposes a dependency-free Node CLI.

# One-shot planning
npx checkmate-agent-system@latest plan --goal "Design an auth system" -c "must support SSO"

# Install portable skill bundle
npx checkmate-agent-system@latest install codex
npx checkmate-agent-system@latest install claude
npx checkmate-agent-system@latest install claude-project

# Replace existing skill install
npx checkmate-agent-system@latest install codex --force

# Inspect package paths
npx checkmate-agent-system@latest doctor
npx checkmate-agent-system@latest skill-path
npx checkmate-agent-system@latest zip-path

Install globally if you prefer a persistent command:

npm i -g checkmate-agent-system@latest
checkmate plan --goal "Ship the release" -c "no downtime"

CLI Commands vs Skill Usage

Feature	Terminal CLI	In-session skill	Notes
Plan skeleton	checkmate plan ...	mention CheckMate in prompt	CLI emits text or JSON; skill guides the agent's reasoning.
Codex install	checkmate install codex	—	Copies the skill into ${CODEX_HOME:-~/.codex}/skills.
Claude install	checkmate install claude	—	Copies the skill into ~/.claude/skills.
Project install	checkmate install claude-project	—	Copies the skill into .claude/skills for the current project.
Diagnostics	checkmate doctor	—	Checks bundled skill and default target paths.
Runtime execution	Python API	—	Use the Python runtime for programmatic board/agent execution.

The CLI emits a planning skeleton and catches simple direct contradictions such as must use external API plus cannot use external API. The full conflict resolver and message-bus runtime live in Python.

Skill Package

The portable skill lives here:

skills/checkmate-agent-system/
  SKILL.md
  agents/openai.yaml
  references/piece-system.md
  references/schemas.md
  scripts/checkmate_plan.py

For Claude.ai custom skill upload, use:

dist/checkmate-agent-system-skill.zip

The ZIP root contains checkmate-agent-system/, so it is ready for upload/import.

Manual install is still supported:

cp -R skills/checkmate-agent-system "${CODEX_HOME:-$HOME/.codex}/skills/"
cp -R skills/checkmate-agent-system ~/.claude/skills/
mkdir -p .claude/skills && cp -R skills/checkmate-agent-system .claude/skills/

Python Runtime

The repo also includes a Python runtime that models the board, pieces, message bus, tool permissions, conflict resolver, memory layer, and execution engine.

python3 examples/basic_runtime.py
python3 -m unittest discover -s tests

from checkmate import ExecutionEngine

engine = ExecutionEngine.default()
board = engine.run(
    goal="Build a project plan for an API migration",
    constraints=[
        "must preserve uptime",
        "deadline is 2 weeks",
        "use existing team capacity",
    ],
)

print(board.status)
print(board.current_plan)

Features

• Governance-first planning: King owns objective, constraints, approval, check, and checkmate.
• Explicit orchestration: Queen decomposes and synthesizes without bypassing validators.
• Rook validation pattern (logic + resources, currently keyword-heuristic).
• Strategic analysis: Bishops surface causal effects, patterns, and tradeoffs.
• Nonlinear recovery: Knights challenge assumptions and generate fallback paths.
• Pawn role for atomic execution and evidence reporting (you wire the actual execution to your tools).
• Portable skills: One skill folder works for Codex and Claude-style skill imports.
• npx-ready CLI: Install skills and emit planning skeletons from any terminal.

Check And Checkmate

check means the goal is threatened but recovery may still be possible: deadline risk, resource shortage, missing information, contradictory requirements, or execution failure.

checkmate means no valid plan satisfies the current goal and hard constraints. The system must stop and ask the King to relax the objective, relax constraints, or redefine success.

Requirements

Surface	Requirement
npx CLI	Node.js 18+
Codex skill install	Codex with a skills folder, usually ~/.codex/skills
Claude Code skill install	Claude Code with ~/.claude/skills or project .claude/skills
Python runtime	Python 3.9+

Development

# Run Node and Python tests
npm run check

# Try the CLI locally
node bin/checkmate.js plan --goal "Package CheckMate" -c "must work with npx"
node bin/checkmate.js doctor

# Inspect npm package contents before publishing
npm pack --dry-run

Publish when ready:

npm login
npm publish

Package note: the project is branded CheckMate Agent System and the npm package is checkmate-agent-system.

Documentation

• Architecture
• npx Reference
• Skill entrypoint
• Piece system reference
• Schemas reference

License

MIT

---

Governance, synthesis, validation, analysis, exploration, execution.

A chess side for thinking.