OKF Workbench
0002 โ Keep the core deterministic and AI-provider-free
0002 โ Keep the core deterministic and AI-provider-free
- Status: Accepted
- Date: 2026-07-22
Context
OKF is intended for both humans and agents, but initialization, parsing, validation, indexing, and graph construction do not require a language model. Adding a provider would introduce authentication, cost, network access, privacy concerns, and non-deterministic behavior before the core workflow is proven.
Coding agents can already consume repository instructions through AGENTS.md and portable Agent Skills.
Decision
- Do not embed an AI provider in the MVP.
- Do not require an account, API key, or network request for core features.
- Implement parsing, validation, generation, and graph construction deterministically.
- Support external agents by generating a concise
AGENTS.mdsection and an optional project-local Agent Skill.
Consequences
Positive:
- The extension can operate on sensitive bundles without uploading content.
- Tests can assert exact outcomes.
- Open VSX users are not tied to a specific model vendor.
- Installation and onboarding remain lightweight.
Costs:
- Semantic enrichment and relationship inference are not included.
- Users who want generated knowledge must use a separate agent or tool.
- Agent templates need careful wording and safe merge behavior.