#49 No way to switch back to an external llama.cpp from the LLM tab

[Anonymous]

Originally created by: sachin-detrax

Problem

Once you pick managed llama.cpp at first run, the TUI gives you no way to move the chat route back to a llama-server you run yourself.

The LLM tab (/models, /model, /llm) has a Mode: Local | Cloud switch, but neither pane covers this case: Local browses the managed model catalog and Cloud browses API providers. The panel displays the current mode in its status line —

local daemon: running pid 42 on 127.0.0.1:19091 · mode managed

— but nothing in the UI changes it.

The only route today is the undocumented back-compat form of the slash command:

/models http://192.168.1.50:8080

which is discoverable only by reading dispatchModelsSub in src/tui/commands/slash-command-handler.ts.

Two defects behind it

Both live in persistLlamaUrl (src/tui/tui-command.ts), so they affect /models <base-url> today and would affect any new UI built on it:

1. The saved URL can be inert. persistUserLocalLlmUrl writes localModels.url + mode: "external", but if a cloud provider is the active text provider the chat route never moves. The URL is saved and nothing changes.

2. The LLM tab keeps reporting managed mode. localModelsPanel.configMode is only recomputed by localModels.refresh(), which the persist path never calls, so the status line contradicts the config until something else triggers a refresh.

Related: the managed daemon is left running

Selecting a cloud provider stops the local daemons (stopLocalDaemonsForCloudSelection, src/tui/llm-panel/llm-panel-primary-actions.ts). Switching to an external llama-server does not — the managed llama-server keeps its VRAM for a route nothing uses. On a 16 GB card that is the difference between the external model loading and not.

Expected

• A discoverable way to point the chat route at an external llama-server, and to see whether that route is live.
• Saving a URL should move the chat route and be reflected in the panel.
• Switching away from managed should release the managed chat model's VRAM, without taking hybrid recall down with it.

Environment

atomic-agent 0.1.67, TUI, managed mode.

[Anonymous]

Ticket changed by: Ooooze

• status: open --> closed