Sloop's reason to exist is multi-AI-provider support, so the codebase is provider-aware by
design: all per-provider knowledge lives in the adapter manifest (internal/adapter/builtin/ <tool>.yaml, user-overridable in ~/.sloop/adapters/), and every provider-aware feature reads it.
The rule: never hardcode a tool name in a feature. Add capability to the manifest instead.
Detect, launch, context delivery, skills, hooks, and completion are all manifest-driven. To support a
new CLI, drop in one <tool>.yaml; no Go changes unless it needs a brand-new mechanism.
- Provider contract + manifest schema:
docs/reference/ADAPTERS.md - Config layering (unified user config vs per-provider manifests):
docs/reference/CONFIG.md - Runtime capability matrix:
sloop tools
User config is unified (one local .sloop/config.yaml + one global ~/.sloop/config.yaml +
the global DB), never split per provider.
- Always run
make lint,make test, andgo test -v -tags e2e ./e2e/...before committing. This project requires strict adherence to linting rules and 100% passing tests (including E2E tests which require tmux).