Cluster app: capability map + live placement + move things across nodes + agent-assisted auto-organise

[jaylfc]

Vision (Jay 2026-06-14)

Make the Cluster app the place to SEE and ORGANISE the whole cluster's capabilities and placement.

Must-haves

• See what capabilities you have across the cluster (image-gen, chat/LLM, embeddings, rerank, vision, TTS/STT, etc.), aggregated from all nodes.
• See where things are installed / live (which model/backend runs on which node, loaded vs installed-but-stopped, VRAM/RAM use, health).
• Move things around the cluster: drag a model/backend from one node to another (relocate where a capability runs). Builds on the Visual Resource Manager idea (RPG-inventory drag-drop).
• Auto-organise with the taOS agent: the agent reads the cluster state + the optimiser's suggestions and proposes (or applies) a better layout for the user's hardware and goals (e.g. move the image model to the GPU box, keep small chat on the NPU). One-click apply with a clear before/after.

Build on what exists

• desktop/src/apps/ClusterApp.tsx (the app to evolve).
• tinyagentos/cluster/capabilities.py (capability model), cluster/optimiser.py (placement optimisation -> the auto-organise engine), cluster/manager.py (ClusterManager), cluster/router.py (capability routing).
• The describe_image_capabilities agent tool already gives read-only capability awareness for one domain; generalise that view.

Tie-ins

• Scheduler Phase 2: VRAM-accounted admission + queue + eviction (the GPU arbiter both agents use) #894 (scheduler admission decides feasibility of a placement); the lifecycle/decay (load/unload) executes moves; Worker/cluster-node service: auto-update with graceful pause, install, restart #890/Cluster worker: hybrid deployment (containerised default + native/bare-metal option) under one transport-agnostic contract #892 (worker contract carries remote start/stop for a relocation).
• The capability map is the same data the universal control plane (Universal control plane: live queue + full-trace observability/audit that every interaction flows through #896) surfaces.

Notes

End-users never need the terminal: moves/auto-organise are UI actions with clear guidance. Hardware is heterogeneous, so the view adapts to what the user actually has. Brainstorm -> spec; do not block the storybook demo.