Low-RAM (4GB) hosts: core-only control plane, route everything else to the cluster

[jaylfc]

Policy

A 4GB system should run only core taOS (the FastAPI controller + SPA + lightweight stores/config). Everything resource-heavy — models/qmd, inference backends, the taosmd memory pipeline, userspace container apps, and agent containers — must deploy on the cluster (or a cloud device), not the local host.

Install must never hard-block on 4GB: install the core, then require + guide the user to add a cluster node or cloud device before any model/agent/app/memory workload can run. Until a device is added, those features are gated with clear "add a device" guidance — not silent failure.

This is the control-plane / compute-plane split that fits the cluster + remote-desktop direction.

Current state

• Agent deploy is local-only today (deployer.deploy_agent, incus on the controller host). Deploy-to-remote-target does not exist yet — so "add a cluster/cloud device to deploy" is aspirational until that lands. See E2E integration test: model aggregation + deploy routing on Fedora LXC worker #58 (deploy routing to LXC worker).
• bug(deploy): low-RAM hosts (≤4GB) silently fail at agent container launch with no actionable error #384 (closed via fix(deploy): RAM pre-flight check prevents silent failure on low-RAM hosts (fixes #384) #483) added an interim safety-net: a deploy-time RAM preflight that rejects local deploy on low-RAM hosts with an actionable error instead of a silent incus-launch hang. It checks local host RAM only.

Work required

1. Deploy-to-remote-target — route agent-container deploy (and userspace container apps) to a cluster worker / cloud device instead of the local host.
2. Low-RAM gating — on 4GB hosts, gate local qmd/model downloads + local deploy in the installer and Store, with "add a device" guidance (ties into hardware-tier templates Store: hardware-tier install templates (curated bundles per device class) #332 and store compatibility borders Store: green/amber/red compatibility borders on model cards #331).
3. Make the fix(deploy): RAM pre-flight check prevents silent failure on low-RAM hosts (fixes #384) #483 RAM preflight target-aware — block only when no capable target (local/cluster/cloud) exists, instead of always checking local RAM.
4. Verify the installer never hard-blocks on 4GB — core install must succeed; only workloads are gated.

Acceptance

A 4GB box installs core taOS, runs the desktop/controller, and — after the user adds a cluster/cloud device — can deploy models, agents, memory, and apps to that device, with clear guidance at every gated step and no silent failures.

[kilo-code-bot]

This issue appears to be a duplicate of #384.

bug(deploy): low-RAM hosts (≤4GB) silently fail at agent container launch with no actionable error (#384)

Similarity score: 92%

This comment was generated by Kilo Auto-Triage.