math - Multi-Agent Task Harness

A light meta agent orchestration harness designed to coordinate multiple AI agents working together to accomplish tasks managed by dex.

Core Concept

The primary responsibility of this harness is to reduce context bloat by digesting a project plan into focused documents:

Document	Purpose
dex	Task tracking with status, dependencies, and context
LEARNINGS.md	Accumulated insights from completed tasks
PROMPT.md	System prompt with guardrails ("signs")

The harness consists of a simple for-loop, executing a new coding agent with a mandate from PROMPT.md to complete a single task from dex, while reading and recording any insight gained during the work into LEARNINGS.md.

Directory Structure

Path	Description
.dex/	Dex task storage (tasks.jsonl)
.math/todo/	Active sprint files (PROMPT.md, LEARNINGS.md)
.math/backups/	Archived learnings from math iterate

Requirements

Bun is required to run this tool. Node.js is not supported.

# Install Bun (macOS, Linux, WSL)
curl -fsSL https://bun.sh/install | bash

dex is required for task management.

# Install dex
bun add -g @zeeg/dex

OpenCode is required to run the agent loop.

# Install OpenCode
curl -fsSL https://opencode.ai/install | bash

Installation

From npm (recommended)

# Global install (recommended)
bun install -g @cephalization/math

# One-off usage 
bunx @cephalization/math <command>

From source (for development)

git clone https://github.com/cephalization/math.git
cd math
bun install
bun link

Usage

Initialize a project

math init

Initializes dex and creates a .math/todo/ directory with template files. Offers to run planning mode to help you break down your goal into tasks.

Options:

• --no-plan - Skip the planning prompt
• --model <model> - Model to use (default: anthropic/claude-opus-4-5)

Plan your tasks

math plan

Options:

• --model <model> - Model to use (default: anthropic/claude-opus-4-5)
• --quick - Skip clarifying questions and generate plan immediately

Interactively plan your tasks with AI assistance. The planner uses a two-phase approach:

1. Clarification phase: The AI analyzes your goal and asks 3-5 clarifying questions
2. Planning phase: Using your answers, it creates tasks in dex with proper dependencies

Use --quick to skip the clarification phase if you want a faster, assumption-based plan.

Run the agent loop

math run

Options:

• --model <model> - Model to use (default: anthropic/claude-opus-4-5)
• --max-iterations <n> - Safety limit (default: 100)
• --pause <seconds> - Pause between iterations (default: 3)

Iteratively run the agent loop until all tasks are complete. Each iteration will:

• Query dex to find the next ready task
• Invoke the agent with PROMPT.md and task context
• The agent will complete the task and mark it done in dex
• The agent will log learnings to LEARNINGS.md
• The agent will commit the changes to the repository
• The agent will exit

Check status

math status

Shows task progress with a visual progress bar and next task info.

Start a new sprint

math iterate

Archives completed tasks and resets for a new goal:

• Completed dex tasks are archived
• LEARNINGS.md is backed up and reset
• PROMPT.md is preserved (keeping your accumulated "signs")
• Offers to run planning mode for your new goal

Options:

• --no-plan - Skip the planning prompt

Task Management

Tasks are managed by dex. Common commands:

# List ready tasks
dex list --ready

# Create a new task
dex create "Task description" --description "Detailed context"

# View task details
dex show <task-id>

# Mark task complete
dex complete <task-id> --result "What was done"

# View overall status
dex status

See the dex documentation for full CLI reference.

The Loop

┌─────────────────────────────────────────────────────────────┐
│                      math run (loop)                        │
│  ┌───────────────────────────────────────────────────────┐  │
│  │  1. Query dex for ready tasks                         │  │
│  │  2. If all complete → EXIT SUCCESS                    │  │
│  │  3. Invoke agent with PROMPT.md + task context        │  │
│  │  4. Agent: start task → implement → test → commit     │  │
│  │  5. Agent: complete task in dex → log learnings       │  │
│  │  6. Loop back to step 1                               │  │
│  └───────────────────────────────────────────────────────┘  │
└─────────────────────────────────────────────────────────────┘

Planning Mode

When you run math init, math iterate, or math plan, the harness can invoke OpenCode to help you plan:

1. You describe your high-level goal
2. OpenCode asks clarifying questions to understand your requirements
3. You answer the questions interactively
4. OpenCode breaks your goal into discrete, implementable tasks
5. Tasks are created in dex with proper dependencies
6. You're ready to run math run

This bridges the gap between "I want to build X" and a structured task list.

Signs (Guardrails)

"Signs" are explicit guardrails in PROMPT.md that prevent common agent mistakes. Add new signs whenever you discover a failure mode:

### SIGN: Descriptive Name

Clear explanation of what to do or avoid.

❌ WRONG: Example of the mistake
✅ RIGHT: Example of correct behavior

Signs accumulate over time, making the agent increasingly reliable.

Environment Variables

Variable	Default	Description
MODEL	anthropic/claude-opus-4-5	Model to use

Credits

• Ralph Methodology: Geoffrey Huntley
• Task Management: dex
• Agent Runtime: OpenCode

License

MIT