Skip to content

Dodaios is an open-source agent OS for your desktop: it runs your coding agents, browser and apps by voice or chat, on your own machine. Formerly Personal Jarvis.

fix(local-models): speech gets a card, embeddings moves to the side row

rubenluetke10-beepbe024bbe4ed3
diff --git a/jarvis/ui/web/frontend/src/i18n/locales/local_models/de.json b/jarvis/ui/web/frontend/src/i18n/locales/local_models/de.json
index ed41c8f918..1f87fb9f65 100644
--- a/jarvis/ui/web/frontend/src/i18n/locales/local_models/de.json
+++ b/jarvis/ui/web/frontend/src/i18n/locales/local_models/de.json
@@ -4,9 +4,8 @@
"chat_purpose": "Antwortet, wenn du tippst.",
"tools_screen_purpose": "Liest den Bildschirm und bedient die Werkzeuge.",
"deep_purpose": "Übernimmt die langen Aufgaben: Code, Planung, Recherche.",
- "embedding_purpose": "Macht deine Notizen durchsuchbar.",
- "voice_purpose": "Antwortet im Gespräch. Wird im Sprachserver gesetzt.",
- "voice_absent": "Noch kein Sprachserver — ein Anruf wird vom Chat-Modell beantwortet.",
+ "embedding_purpose": "Indexiert dein Wiki im Hintergrund. Einmal gewählt, dann vergessen.",
+ "voice_purpose": "Antwortet, wenn du sprichst. Ein Anruf will ein kleines, schnelles Modell.",
"empty": "Noch kein Modell",
"not_on_disk": "Gesetzt, aber nicht heruntergeladen",
"no_facts": "Größe unbekannt",
diff --git a/jarvis/ui/web/frontend/src/i18n/locales/local_models/en.json b/jarvis/ui/web/frontend/src/i18n/locales/local_models/en.json
index 8a9f960f8c..714c9a5ac8 100644
--- a/jarvis/ui/web/frontend/src/i18n/locales/local_models/en.json
+++ b/jarvis/ui/web/frontend/src/i18n/locales/local_models/en.json
@@ -4,9 +4,8 @@
"chat_purpose": "Answers when you type.",
"tools_screen_purpose": "Reads the screen and works the tools.",
"deep_purpose": "Takes the long jobs: coding, planning, research.",
- "embedding_purpose": "Turns your notes into something searchable.",
- "voice_purpose": "Answers a call. Set in the voice server.",
- "voice_absent": "No voice server yet — a call is answered by the chat model.",
+ "embedding_purpose": "Indexes your wiki in the background. Picked once, then forgotten.",
+ "voice_purpose": "Answers when you talk. A call wants a small, quick model.",
"empty": "No model yet",
"not_on_disk": "Set, but not downloaded",
"no_facts": "Size unknown",
diff --git a/jarvis/ui/web/frontend/src/i18n/locales/local_models/es.json b/jarvis/ui/web/frontend/src/i18n/locales/local_models/es.json
index 49ce18242c..4098550ced 100644
--- a/jarvis/ui/web/frontend/src/i18n/locales/local_models/es.json
+++ b/jarvis/ui/web/frontend/src/i18n/locales/local_models/es.json
@@ -4,9 +4,8 @@
"chat_purpose": "Responde cuando escribes.",
"tools_screen_purpose": "Lee la pantalla y maneja las herramientas.",
"deep_purpose": "Se ocupa de las tareas largas: código, planificación, investigación.",
- "embedding_purpose": "Convierte tus notas en algo que se puede buscar.",
- "voice_purpose": "Responde en una llamada. Se define en el servidor de voz.",
- "voice_absent": "Aún no hay servidor de voz: la llamada la responde el modelo de chat.",
+ "embedding_purpose": "Indexa tu wiki en segundo plano. Se elige una vez y se olvida.",
+ "voice_purpose": "Responde cuando hablas. Una llamada quiere un modelo pequeño y rápido.",
"empty": "Aún sin modelo",
"not_on_disk": "Definido, pero no descargado",
"no_facts": "Tamaño desconocido",
diff --git a/jarvis/ui/web/frontend/src/views/local-models/ModelCard.tsx b/jarvis/ui/web/frontend/src/views/local-models/ModelCard.tsx
index c2eb11e8f4..9cba8e417c 100644
--- a/jarvis/ui/web/frontend/src/views/local-models/ModelCard.tsx
+++ b/jarvis/ui/web/frontend/src/views/local-models/ModelCard.tsx
@@ -47,6 +47,11 @@ export interface ModelCardProps {
advanced?: boolean;
}
+/** Human window size: 16384 -> "16k". */
+function contextLabel(numCtx: number): string {
+ return numCtx >= 1024 ? `${Math.round(numCtx / 1024)}k` : String(numCtx);
+}
+
/** The capabilities a job is gated on, as one short phrase. */
function capabilityLine(row: LocalModelRow | null): string {
if (!row || !row.probed) return "";
@@ -54,8 +59,10 @@ function capabilityLine(row: LocalModelRow | null): string {
}
export function cardState(row: RoleRow): ModelCardState {
- // A job served by something other than Ollama says so and offers nothing.
- if (!row.writable) return "blocked";
+ // A job served by something other than Ollama, or whose own server is not
+ // installed yet, says so and offers nothing: a picker whose write is going
+ // to be refused is worse than a sentence naming what is missing.
+ if (!row.writable || (row.note !== "" && row.current === "")) return "blocked";
if (!row.current) return "empty";
return row.installed ? "ready" : "missing";
}
@@ -148,15 +155,30 @@ export function ModelCard({
</>
) : (
<p className="text-sm text-muted-foreground">
- {row.writable
- ? t("local_models.jobs.empty")
- : row.note || t("local_models.roles.read_only")}
+ {state === "blocked"
+ ? row.note || t("local_models.roles.read_only")
+ : t("local_models.jobs.empty")}
</p>
)}
</div>
+ {/* Speech runs with a window sized for this machine; a call is judged
+ by its first word, so that number belongs on the card. */}
+ {row.id === "voice" && row.context_tokens ? (
+ <p className="mt-1.5 text-[11px] text-muted-foreground" data-testid="voice-context">
+ {fill(
+ t(
+ row.context_source === "manual"
+ ? "local_models.roles.voice_context_manual"
+ : "local_models.roles.voice_context_auto",
+ ),
+ { context: contextLabel(row.context_tokens) },
+ )}
+ </p>
+ ) : null}
+
{/* The memory bar: this model against this machine's graphics memory. */}
- {row.writable && (
+ {state !== "blocked" && (
<div className="mt-2.5" data-testid={`model-card-memory-${row.id}`}>
<div
className="h-1.5 overflow-hidden rounded-full bg-sheen/[0.08]"
@@ -195,7 +217,7 @@ export function ModelCard({
{/* The control. Always present, always the same shape. */}
<div className="mt-3.5 space-y-2">
- {row.writable ? (
+ {state !== "blocked" ? (
<RolePicker
row={row}
models={models}
diff --git a/jarvis/ui/web/frontend/src/views/local-models/OverviewPanel.test.tsx b/jarvis/ui/web/frontend/src/views/local-models/OverviewPanel.test.tsx
index dfe165d0fe..cbe6ec56e5 100644
--- a/jarvis/ui/web/frontend/src/views/local-models/OverviewPanel.test.tsx
+++ b/jarvis/ui/web/frontend/src/views/local-models/OverviewPanel.test.tsx
@@ -431,11 +431,12 @@ describe("OverviewPanel grid", () => {
);
expect(cards.map((c) => c.getAttribute("data-testid"))).toEqual([
"model-card-chat",
+ "model-card-voice",
"model-card-tools_screen",
"model-card-deep",
- "model-card-embedding",
]);
- for (const job of ["chat", "tools_screen", "deep", "embedding"])
+ // Every job on the page can be changed, side row included.
+ for (const job of ["chat", "voice", "tools_screen", "deep", "embedding"])
expect(screen.getByTestId(`role-picker-${job}`)).toBeDefined();
});
@@ -450,7 +451,7 @@ describe("OverviewPanel grid", () => {
await screen.findByTestId("model-card-chat");
expect(screen.getByTestId("model-card-chat").dataset.state).toBe("ready");
expect(screen.getByTestId("model-card-deep").dataset.state).toBe("missing");
- expect(screen.getByTestId("model-card-embedding").dataset.state).toBe(
+ expect(screen.getByTestId("model-card-tools_screen").dataset.state).toBe(
"empty",
);
expect(screen.getByTestId("model-card-deep").textContent).toContain(
@@ -535,15 +536,58 @@ describe("OverviewPanel grid", () => {
expect(urls).toContain(`PUT ${BASE}/models/qwen3.8%3A27b/options`);
});
- it("keeps speech out of the grid and says where it is set", async () => {
+ it("gives speech a card of its own, with the window it runs on", async () => {
+ installFetchMock({
+ roles: ROLES.map((r) =>
+ r.id === "voice"
+ ? { ...r, context_tokens: 32768, context_source: "automatic" }
+ : r,
+ ),
+ });
+ renderPanel();
+
+ const card = await screen.findByTestId("model-card-voice");
+ expect(card.textContent).toContain("local_models.jobs.voice_purpose");
+ expect(screen.getByTestId("role-picker-voice")).toBeDefined();
+ expect(screen.getByTestId("voice-context").textContent).toContain(
+ "local_models.roles.voice_context_auto32k",
+ );
+ });
+
+ it("blocks the card, rather than offering a doomed picker, when a job has no server", async () => {
+ installFetchMock({
+ roles: ROLES.map((r) =>
+ r.id === "voice"
+ ? {
+ ...r,
+ current: "",
+ installed: false,
+ note: "The managed voice server is not installed.",
+ }
+ : r,
+ ),
+ });
+ renderPanel();
+
+ const card = await screen.findByTestId("model-card-voice");
+ expect(card.dataset.state).toBe("blocked");
+ expect(card.textContent).toContain(
+ "The managed voice server is not installed.",
+ );
+ expect(screen.queryByTestId("role-picker-voice")).toBeNull();
+ });
+
+ it("keeps embeddings settable in the side row instead of the grid", async () => {
installFetchMock();
renderPanel();
await screen.findByTestId("model-card-chat");
- expect(screen.queryByTestId("model-card-voice")).toBeNull();
- const voice = await screen.findByTestId("voice-row");
- expect(voice.textContent).toContain("local_models.role_voice");
- expect(voice.textContent).toContain("local_models.jobs.voice_purpose");
+ expect(screen.queryByTestId("model-card-embedding")).toBeNull();
+ const side = await screen.findByTestId("side-job-embedding");
+ expect(side.textContent).toContain("local_models.role_embedding");
+ expect(side.textContent).toContain("local_models.jobs.embedding_purpose");
+ // The setting has nowhere else to live, so the row keeps the picker.
+ expect(within(side).getByTestId("role-picker-embedding")).toBeDefined();
});
it("the detail level adds detail, never takes the picker away", async () => {
@@ -580,7 +624,7 @@ describe("OverviewPanel launch bar", () => {
],
accelerator_gb: 16,
roles: ROLES.map((r) =>
- r.id === "embedding"
+ r.id === "voice"
? { ...r, current: "qwen3-embedding:4b", installed: true }
: r.id === "tools_screen" || r.id === "deep"
? { ...r, current: "qwen3.5:4b", installed: true }
@@ -590,7 +634,7 @@ describe("OverviewPanel launch bar", () => {
renderPanel();
const facts = await screen.findByTestId("launch-facts");
- // chat + tools + deep share qwen3.5:4b, embedding adds 2 GB -> 5 GB.
+ // chat + tools + deep share qwen3.5:4b, voice adds 2 GB -> 5 GB.
await waitFor(() =>
expect(facts.textContent).toContain(
"local_models.launch.total_on_disk5.0 GB",
@@ -609,10 +653,10 @@ describe("OverviewPanel launch bar", () => {
renderPanel();
const launch = await screen.findByTestId("server-launch");
- // Only chat is set in the fixture: 1 of 4.
+ // chat and voice are set in the fixture, tools & deep are not: 2 of 4.
await waitFor(() =>
expect(launch.textContent).toContain(
- "local_models.launch.partly_picked1|4",
+ "local_models.launch.partly_picked2|4",
),
);
const button = within(screen.getByTestId("launch-run")).getByRole("button");
diff --git a/jarvis/ui/web/frontend/src/views/local-models/OverviewPanel.tsx b/jarvis/ui/web/frontend/src/views/local-models/OverviewPanel.tsx
index ae12a1cec7..bb0ae223bc 100644
--- a/jarvis/ui/web/frontend/src/views/local-models/OverviewPanel.tsx
+++ b/jarvis/ui/web/frontend/src/views/local-models/OverviewPanel.tsx
@@ -2,13 +2,16 @@
* Local models, front page: a grid of jobs, then the server that runs them.
*
* The page answers one question — "which model does what on my machine" —
- * and it answers it the way the machine is actually shaped. Chat, tools &
- * screen, deep & coding and embeddings are four PEERS, not four steps: you
- * can fill them in any order, and a number beside each would claim a
- * sequence that does not exist. So they are a grid of equal cards
- * (`ModelCard`), each carrying its own model, its own memory bar and its own
- * picker. Speech sits below them, on its own, because it is the one job
- * served by a different server and cannot be filled from here.
+ * and it answers it the way the machine is actually shaped. Chat, speech,
+ * tools & screen and deep & coding are four PEERS, not four steps: you can
+ * fill them in any order, and a number beside each would claim a sequence
+ * that does not exist. So they are a grid of equal cards (`ModelCard`), each
+ * carrying its own model, its own memory bar and its own picker.
+ *
+ * All four answer the user directly, which is what earns a card. Embeddings
+ * does not — it indexes the wiki in the background, is picked once and then
+ * forgotten — so it keeps its picker in the side row below instead of taking
+ * a quarter of the page.
*
* Under the grid, `ServerLaunch` adds up what the picks cost and offers the
* one button that puts a working server behind them.
@@ -27,6 +30,7 @@ import { SoftButton, StatusDot } from "@/components/extensions/primitives";
import {
useOverview,
type LocalModelRole,
+ type LocalModelRow,
type RoleRow,
} from "@/hooks/useLocalModels";
import { switchBrainProvider, useProviders } from "@/hooks/useProviders";
@@ -41,6 +45,7 @@ import { InstalledPanel } from "./InstalledPanel";
import { canonical } from "./localSetup";
import type { SetupStep, SetupSummary } from "./localSetup";
import { ModelCard } from "./ModelCard";
+import { RolePicker } from "./RolePicker";
import { ServerLaunch } from "./ServerLaunch";
import { useLocalSetup } from "./useLocalSetup";
import { useRoleActions, IDLE } from "./useRoleActions";
@@ -62,17 +67,24 @@ export interface OverviewPanelProps {
}
/**
- * The jobs the grid shows, in reading order. Four, because these are the
- * four a local server can fill on its own — speech needs the managed voice
- * server and is handled apart, and `ack`/`polish` follow other cards.
+ * The jobs the grid shows, in the order you meet them: type, talk, let it
+ * work the screen, hand it something long.
+ *
+ * All four answer YOU. Embeddings does not — it indexes the wiki in the
+ * background, is picked once and then forgotten — so it sits in the side row
+ * below with its own picker rather than taking a quarter of the page.
+ * `ack`/`polish` follow other cards and are read-only.
*/
export const GRID_JOBS: readonly LocalModelRole[] = [
"chat",
+ "voice",
"tools_screen",
"deep",
- "embedding",
];
+/** The jobs below the grid: set here too, just not part of the conversation. */
+export const SIDE_JOBS: readonly LocalModelRole[] = ["embedding"];
+
/** Bytes to a short gigabyte figure ("12.4 GB"); "0 GB" below a megabyte. */
export function formatGigabytes(bytes: number): string {
const gb = bytes / 1024 ** 3;
@@ -124,10 +136,13 @@ export function OverviewPanel({
() => GRID_JOBS.map((id) => byId.get(id)).filter((r): r is RoleRow => !!r),
[byId],
);
- const voiceRow = byId.get("voice") ?? null;
- // Speech has its own row above; whatever else is advanced follows another
- // card's pick and belongs in the quiet list at the bottom.
- const followerRows = allRows.filter((r) => r.advanced && r.id !== "voice");
+ const sideRows = useMemo(
+ () => SIDE_JOBS.map((id) => byId.get(id)).filter((r): r is RoleRow => !!r),
+ [byId],
+ );
+ // Whatever is advanced follows another card's pick and belongs in the quiet
+ // list at the bottom.
+ const followerRows = allRows.filter((r) => r.advanced);
const serverLabel =
providers.find((p) => p.id === providerId)?.label ||
@@ -223,14 +238,18 @@ export function OverviewPanel({
<p className="text-sm text-destructive">{actions.writeError}</p>
)}
- {/* Speech: its own row, because it is served by its own server. */}
- {voiceRow && (
- <VoiceRow
- row={voiceRow}
+ {/* Set here, but not part of the conversation. */}
+ {sideRows.map((row) => (
+ <SideJobRow
+ key={row.id}
+ row={row}
+ models={models}
+ busy={actions.isBusy(row.id)}
+ onPick={(model) => actions.pick(row.id, model)}
onTune={advanced ? onTune : undefined}
t={t}
/>
- )}
+ ))}
<ServerLaunch
rows={gridRows}
@@ -283,55 +302,63 @@ function CardSkeleton() {
}
/**
- * Speech. Read-only from here: the model is baked into the voice server's
- * launch command, so the row says what answers a call and where to change it
- * rather than offering a picker that would fail.
+ * A job that is set here but is not part of the conversation — today just
+ * embeddings, which indexes the wiki in the background.
+ *
+ * One line, not a card: it carries the same picker (dropping it would take a
+ * setting away with nowhere else to reach it) but none of the card's weight,
+ * because it is picked once and then forgotten.
*/
-function VoiceRow({
+function SideJobRow({
row,
+ models,
+ busy,
+ onPick,
onTune,
t,
}: {
row: RoleRow;
+ models: LocalModelRow[];
+ busy: boolean;
+ onPick: (model: string) => void;
onTune?: (model: string) => void;
t: (key: string) => string;
}) {
const missing = row.current !== "" && !row.installed;
+ const blocked = !row.writable || (row.note !== "" && row.current === "");
return (
<div
- className="flex flex-wrap items-center justify-between gap-x-4 gap-y-2 rounded-xl border border-border/70 bg-card/40 px-4 py-3"
- data-testid="voice-row"
+ className="grid gap-3 rounded-xl border border-border/70 bg-card/40 px-4 py-3 lg:grid-cols-[minmax(0,1fr)_minmax(0,20rem)_auto] lg:items-center"
+ data-testid={`side-job-${row.id}`}
>
<div className="min-w-0">
<p className="text-[11px] font-medium uppercase tracking-[0.18em] text-muted-foreground">
{t(row.label_key)}
</p>
- <div className="mt-1">
- {row.current ? (
- <StatusDot
- tone={missing ? "warn" : "ok"}
- label={
- <span className="font-mono text-xs text-foreground/85">
- {row.current}
- {missing && (
- <span className="ml-2 font-sans text-muted-foreground">
- {t("local_models.roles.not_installed")}
- </span>
- )}
- </span>
- }
- />
- ) : (
- <span className="text-sm text-muted-foreground">
- {row.note || t("local_models.jobs.voice_absent")}
- </span>
- )}
- </div>
- </div>
- <div className="flex items-center gap-3">
- <p className="text-xs text-muted-foreground">
- {t("local_models.jobs.voice_purpose")}
+ <p className="mt-0.5 text-xs text-muted-foreground">
+ {row.note || t(`local_models.jobs.${row.id}_purpose`)}
</p>
+ </div>
+ <div className="min-w-0">
+ {blocked ? (
+ <span className="text-xs text-muted-foreground">
+ {t("local_models.roles.read_only")}
+ </span>
+ ) : (
+ <RolePicker
+ row={row}
+ models={models}
+ disabled={busy}
+ onPick={onPick}
+ />
+ )}
+ </div>
+ <div className="flex items-center gap-2 lg:justify-end">
+ {missing && (
+ <span className="text-xs text-amber-700 dark:text-amber-400">
+ {t("local_models.jobs.not_on_disk")}
+ </span>
+ )}
{onTune && row.current && row.installed && (
<SoftButton onClick={() => onTune(row.current)} className="h-8">
{t("local_models.roles.tune")}