How it works
You create tasks. The system builds a dependency graph, fans out independent tasks to separate AI agents working at the same time. When a task depends on another, it waits.
Each task goes through a pipeline: an issuer agent explores the codebase and writes implementation guidance, a worker agent does the code changes, a finisher verifies. The coordinator (singularity) manages the lifecycle, steers agents going off track, restarts stuck ones.
Parallelism
Submit ten tasks, seven are independent, seven agents spin up at once. The other three wait on their dependencies and start when those finish.
Agents share the working tree but coordinate through the task system. If one agent's changes affect another's work, the system can broadcast updates or steer running agents to adapt.
Singularity vs oh-my-pi
The actual coding is done by oh-my-pi, a fork of pi-mono. That's the AI agent that reads files, runs grep, uses LSP, edits code, opens browsers. Runs locally on your machine with your repo.
oh-my-singularity is the orchestration layer on top. It manages the task queue, resolves dependencies, spins up agents, steers them, handles the issuer/worker/finisher lifecycle. oh-my-pi is the hands. Singularity is the project manager.
Why this exists
Existing AI coding tools are chatbots (prompt back and forth) or autocomplete (type and it suggests). Neither handles "here are twelve things, go do them."
We use it daily for our own projects.