System-AI editor now fetches the detected hardware tier and the
hub's curated model database on open. Each model chip's colour
+ tooltip combines both: green star when curated quality is
good/excellent AND min_hw_tier ≤ host; orange when quality is
fine but the host is below min_hw_tier; warning for "basic"
curated entries. Tooltip exposes the curator's note, parameter
count, context window, and licence. Banner above the form
shows "Detected: <summary> · curation reviewed <date>".
Sidebar footer gains a small monospaced live clock between the
theme toggle and the settings icon. Tooltip exposes full local
ISO + UTC for cross-checking against audit-log timestamps.
Signed-off-by: flemming-it <stefan.a.flemming@googlemail.com>
Two operator-felt papercuts on the editor:
* Models only appeared after a manual Refresh click. Now the
editor calls `_refreshModels()` from initState (post-frame)
so chips populate immediately. Provider switch also re-fires
the refresh because endpoint / model set changes too.
* Operator had no signal which model is suitable. Tiny models
fail at instruction following; huge models stall on a
laptop. Adds a heuristic suitability rating (parses
parameter count from common names like `gemma3:4b`,
`llama3.2:70b`; falls back to family hints for OpenAI /
Claude). Each chip now carries:
- a coloured leading dot (green=balanced, orange=small
or large, red=huge, grey=unknown)
- a star icon for the hand-curated "tested with F∆I"
allowlist (gemma3:4b, llama3.2:3b, qwen2.5-coder:7b,
gpt-4o-mini, claude-haiku-4-5, …)
- a tooltip ("balanced", "huge — likely too slow on a
laptop", …)
Sorting also moves recommended → balanced → others to the
front so the operator's eye lands on green first.
* Compact legend below the chip wrap explains the colour
code in place — no lookup needed.
Signed-off-by: flemming-it <stefan.a.flemming@googlemail.com>
Three operator-facing improvements driven by Stefan's testing:
* **Test Connection** runs against the form's *current* values
and no longer requires saving first. Previously we save-then-
test which caused a "system AI is disabled" message when the
form's model field was still empty. Now the test prompt
travels with the request body so unsaved tweaks can be
validated.
* **Model picker** replaces the free-text Model field. A
Refresh button calls HubAdmin.ListSystemAiModels (provider's
`/v1/models`); installed models show up as clickable chips
that fill the field with one tap. The text input stays for
cases where the operator already knows the id.
* **Pull** button (Ollama only) downloads a model via
`/api/pull`. Type `gemma3:4b`, click Pull, spinner, then the
new model lights up in the chip list. No more "switch to a
terminal to ollama-pull" detour.
Helper text updates for the empty-model case ("Required: pick
from the list (Refresh) or type one. Use Pull to download from
Ollama.") so a fresh operator never wonders why Save / Test
errors out.
Signed-off-by: flemming-it <stefan.a.flemming@googlemail.com>
Replaces the read-only System AI block in Settings with a real
editor. Operator never has to touch ~/.fai/config.yaml again
to change provider, endpoint, model, API-key env-var, or
privacy mode.
UX choices that align with the zero-learning-curve rule:
* Provider dropdown lists Ollama / OpenAI / LM Studio / vLLM
/ Custom by friendly name. Each selection auto-fills the
endpoint and api_key_env defaults — but only when the
field is still empty or matches a different preset's
default, so manual overrides are never clobbered.
* Each provider has a one-line description rendered under
the dropdown ("Ollama → Local `ollama serve`. Models stay
on your machine. No API key needed.") and a model-hint
placeholder ("gemma3:4b · llama3.2:3b · qwen2.5-coder:7b")
that goes away when a model is typed.
* Privacy mode is a vertical radio group with each option's
description in place — no doc-lookup needed.
* "Test connection" sends an `ok`-ping and renders the
same error-kind → fix-hint mapping the audit drill-down
uses.
* "Save" persists + hot-reloads the hub. No daemon restart
required. UI status badge flips to "enabled · <privacy>"
immediately.
`fai_dart_sdk` 0.5.0 carries the underlying `updateSystemAi` /
`testSystemAi` RPCs.
Signed-off-by: flemming-it <stefan.a.flemming@googlemail.com>
