Glossary
The platform's vocabulary, defined once. Terms are expanded at first use elsewhere. This page is the reference.
Infrastructure
| Term | Meaning |
|---|---|
| tailnet | The private Tailscale mesh network (*.taila659a.ts.net) joining every host; the only admin plane into the fleet |
| A2A | Agent-to-agent protocol — how Hermes agents talk to each other over the :9900 API and :8642 A2A on the tailnet |
| RKE2 | Rancher's hardened Kubernetes distribution; what the nishir cluster nodes run, installed by the knix Nix module |
| knix | Shikanime Studio's NixOS module set that turns a Nix host into an RKE2 node (services.knix.*) |
| comin | Declarative NixOS deployment: each host polls its Git branch and re-activates when it moves |
| Flux | Pull-based, self-healing GitOps operator reconciling the manifests repo onto the cluster |
| sops / sops-nix | Encrypted-secrets tooling: secrets live age-encrypted in Git (*.enc.yaml), decrypted to env at activation time |
| Hermes Agent | The Nous Research AI agent runtime, running as a service on every host; agents operate the platform through the same Git/PR gates as humans |
| KubeVirt | Kubernetes virtualization; runs the automata VM, the A2A hub, as a pod-backed VM |
| Longhorn | Replicated block storage for the cluster; cannot shrink a bound PVC |
| Envoy Gateway / Gateway API | Typed Kubernetes CRs for ingress; one Gateway per exposed app, each holding its own tailnet IP |
| Envoy AI Gateway | Envoy Gateway extension for LLM routing: model-based routing, token auth, provider fallback |
Compute
| Term | Meaning |
|---|---|
| Strix Halo | AMD APU in the two inference minis kushira and sashina — Radeon 8060S iGPU with 128 GiB unified memory |
| GTT | Graphics Translation Table — the AMD mechanism letting the iGPU address host RAM; on Halo it IS the memory budget, with no VRAM split |
| w4b | halogen-flash's weight format, a 4-bit weights checkpoint; the qwen3.8-flash file pins ~68 GiB |
| MTP | Multi-token prediction — generates several tokens per step; ships as a sidecar in the halogen checkpoint |
| TTFT | Time-to-first-token; the latency users feel. halogen's 13× TTFT win is why it replaced the RPC pair |
| prefill | Processing the prompt before generation begins; the bandwidth-bound phase on unified-memory APUs |
| KV cache | Per-conversation attention state; shares the GTT pool with weights, so context length trades against model size |
| RPC pair | Retired: two llama.cpp processes bridged over RPC to pool both Halo nodes for one model. Removed Oct 2026 |
Process
| Term | Meaning |
|---|---|
| jj | Jujutsu, the version-control front end used in all repos; workspaces and bookmarks instead of branches/stash |
| SEP | Shikanime Studio Enhancement Proposal — KEP-style cross-org design doc in shikanime-labs/enhancements |
| ghstack | Stacked-PR tooling; sks-* skill family encodes the org's issue→PR lifecycle for humans and agents |
| gitlint CC1 | Commit-lint rule requiring Signed-off-by, enforced on manifests and skills |
| wharf / niximgs | OCI image supply chain: wharf builds images from Nix flakes; niximgs is the app-image catalog (ghcr.io/shikanime) |
| automata | Overloaded: the KubeVirt VM that is the A2A hub, the Go CLI for repo/cluster hygiene, and the agent signing persona. Context disambiguates |
Products
| Term | Meaning |
|---|---|
| Touhou / Gensokyo | The fan-work universe hakurei is set in, © ZUN / Team Shanghai Alice. "Dofus-like" refers to its tactical, AP/MP-budgeted combat |
| better-auth | Auth library powering the accounts IdP and reiya in the web identity mesh |
| D1 | Cloudflare's SQLite database service; reiya's and accounts' datastore |