A top-level OpenCode agent that you can switch to directly, such as build, plan, build-deep, or plan-deep.
The single workflow authority responsible for the current task phase.
In this stack, ownership is usually expressed by the active primary agent or an explicit command such as /plan or /implement.
A narrower helper agent used for a specific kind of work, such as exploration or review.
Subagents assist the owner. They do not co-own the task.
A stronger, slower, or more expensive route intentionally reserved for harder work.
Examples in this repository include build-deep and plan-deep.
A repo-local planning artifact created during /plan.
It captures goals, constraints, acceptance criteria, test planning, and ordered implementation steps.
Natural-language conditions that make the work testable and constrain implementation drift.
These are the bridge between planning and /spec.
The command that turns PLAN.md acceptance criteria into Vitest-oriented test skeletons.
A helper layer that adds value without taking ownership.
Examples include optional skills and Sentry MCP.
Framework- or domain-specific guidance loaded only when the task clearly matches a specific area.
Skills are intentionally not the baseline in this repository.
A runtime-observation layer used to inspect production issues, validate changes against real signals, and reduce false confidence.
An optional lifecycle system that can still be useful for persistent specs, proposals, or larger tracked changes.
It is not required for the default stack documented in this repository.
The earlier convergence explored in this repository: Oh My OpenCode, OpenSpec, Superpowers, Gentleman skills, and Sentry MCP.
It is still important historically, but it is no longer the default recommendation.