Key Design Decisions
- Autonomy model: Binary per-repo, with CODEOWNERS enforcing human approval on specific paths
- Problem structure: Problem-oriented documents (not ADRs or RFCs) that can evolve independently, with ADRs spun off later when decisions crystallize
- Threat priority order: External prompt injection > insider/compromised creds > agent drift > supply chain
- Code generation is considered a solved problem. The hard problems are review, intent, governance, and security.
- Trust derives from repository permissions, not agent identity. No agent trusts another based on who produced the output.
- CODEOWNERS files are always human-owned. Agents cannot modify their own guardrails.
- The repo is the coordinator. No coordinator agent — branch protection, CODEOWNERS, and status checks are the coordination layer.
- Organization-specific content is cordoned. Core problem docs are general; applied considerations live in
docs/problems/applied/.
