About ADRs
What Architecture Decision Records are, how they're organized here, and the MADR convention in use.
0000 — MADR template
Blank MADR 3.0 template to copy for each new ADR.
0001 — Site generator choice: Docusaurus
Decision to choose Docusaurus over Starlight and MkDocs, driven by the CALM, LikeC4, and Slidev integrations.
0002 — VMM choice: Cloud Hypervisor
Decision to choose Cloud Hypervisor as the virtual machine monitor for Kata isolation, driven by its neutral Linux Foundation governance and virtio-fs support.
0003 — Container runtime: Kata by default, runc for GPU
Choice of OCI runtime for agent sandboxing (Kata Containers by default) and a per-workload runc escape hatch for GPU-accelerated local AI model inference, orchestrated by containerd/nerdctl.
0004 — Local model executor: vLLM
Choice of vLLM as the local AI model inference engine, driven by speed (PagedAttention, continuous batching) and open-source governance under the PyTorch Foundation / Linux Foundation.
0005 — Exception: vLLM on classic containerization (GPU constraint)
Decision to run vLLM on classic containerization (docker/runc) rather than the default Kata sandbox, because GPU access is not available under the Cloud Hypervisor VMM.