Accepted
The earlier public version of this repository documented a five-tool stack that combined:
- Oh My OpenCode
- OpenSpec
- Superpowers
- Gentleman skills
- Sentry MCP
That stack encoded valuable ideas, but it also created overlap:
- multiple systems trying to shape planning discipline
- duplicated lifecycle or routing concepts
- higher always-loaded context cost
- slower reasoning because too much process lived in prompts instead of native runtime features
During the 2026-04 convergence work, the stack was re-audited from source, checked against current OpenCode documentation, and rebuilt around the capabilities OpenCode already provides natively.
The baseline stack documented by this repository is now native-first.
That means:
- OpenCode agents, subagents, commands, variants, and plugins form the default baseline
- strong models are routed on demand instead of being used as the default for every turn
- planning writes
PLAN.mdwith explicit acceptance criteria /specconverts those criteria into test anchors- implementation proceeds against the plan and tests instead of free-form execution
- Sentry MCP stays as the runtime evidence layer
- OpenSpec and specialist skills become optional extensions instead of mandatory baseline components
- lower always-loaded context cost
- fewer overlapping ownership systems
- better mapping between planning and testing
- simpler public documentation
- easier cost control for model usage
The native-first stack is less "magical" out of the box. It requires more deliberate command design and clearer repo-local rules, because discipline is encoded closer to the runtime instead of being delegated to a separate all-in-one overlay.