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): an offline sweep no longer poisons the cache, and every job keeps a picker

rubenluetke10-beepdde30c43e18b
diff --git a/docs/BUGS.md b/docs/BUGS.md
index 4cf2c19522..f97b193070 100644
--- a/docs/BUGS.md
+++ b/docs/BUGS.md
@@ -13124,3 +13124,75 @@ only what is true in it.
about the OUTCOME. Every tool that acts without being able to look owes the
model the difference in a field it cannot miss — a note in prose is a caveat
the model will politely paraphrase after the good news.
+## BUG-188: Local models says every model is "not installed" and a job cannot be changed — one sweep taken while Ollama was still booting became the cache both ends read from (HIGH, FIXED 2026-08-27)
+
+**Symptom (maintainer, 2026-08-27 10:09).** The Local models section showed
+`ornith:9b` with a trailing `(not installed)`, and under it the sentence
+"Nothing installed fits this job yet; this download is the pick." Picking a
+different model in the dropdown appeared to do nothing — "it always switches
+back to that odd one, even when you choose something else." Fifteen models
+were on the server at that moment, `ornith:9b` among them.
+
+**What was actually true.** Nothing was broken in the write path. Read live
+during the investigation, the backend was correct on every count:
+
+```
+GET …/local-models/overview?fresh=1 source: live
+ inventory error: None models: 11
+ chat current='ornith:9b' installed=True qualifying=[10 tags]
+PUT …/local-models/roles/chat {"model":"qwen3.5:4b"} -> ok
+ jarvis.toml brain.providers.ollama.model = "qwen3.5:4b"
+ scripts/config-soll.json brain.providers.ollama.model = "qwen3.5:4b"
+```
+
+The pick was written, the drift baseline followed it, and a re-read answered
+the new value. The screen was simply not showing the server.
+
+**Cause — a truthful answer cached as a lasting one.** `build_overview` runs
+its sweep whatever the server's state; when Ollama does not answer it returns
+a payload with an empty inventory and `inventory.error` set. That is the right
+answer to "what can I see right now" and a ruinous one to keep, and it was
+kept twice over:
+
+* `ollama_overview._refresh` handed every payload to `save_snapshot`, so the
+ offline sweep replaced the good file in `data/local_models_snapshot.json`.
+ `get_overview` paints a disk snapshot for `STALE_AFTER_S` — 24 hours.
+* `useOverview` mirrored every payload into `localStorage`, and reads it back
+ as react-query's `initialData` — which survives a failed refetch, so it
+ outlives the outage that produced it.
+
+Opening the section while Ollama is still booting is enough. Everything
+downstream then follows from an empty inventory: `RoleState.qualifying` is
+empty, so each role reads `installed: false` and its shortlist falls back to
+"nothing installed fits this job yet".
+
+**Why the dropdown looked stuck.** `RoleLedgerRow` built its options from
+`row.qualifying` plus the configured tag. With `qualifying` empty that list
+holds exactly ONE entry — the tag already configured. There was nothing else
+to select; the "reset" the maintainer saw was a picker that had never offered
+an alternative. An unprobed download (`probed: false`, no capabilities) fell
+out of the list the same silent way. The Simple page compounded it: the
+picker lived behind a "Change" button, two clicks from the row.
+
+**Fix.**
+
+1. `ollama_overview.is_paintable()` — a payload whose inventory carries an
+ error is never written to the snapshot, and one already on disk is dropped
+ on read instead of painted. The previous, honest snapshot stays.
+2. `localModelsSnapshot.isPaintable()` — the same guard on the browser side,
+ on both write and read.
+3. `RolePicker` offers EVERY installed download, grouped "Fits this job" /
+ "Other installed models", so there is always something to pick and an
+ unprobed model is visible rather than dropped.
+4. The picker is on screen for every job at every detail level; `advanced`
+ adds capability badges, Tune and the read-only jobs, it never removes a
+ control.
+5. A "Refresh" button in the section header clears the painted snapshot and
+ every cached answer, then reads the server again — the recovery a user can
+ reach without knowing what a cache is.
+
+**Lesson.** A cache entry is a claim that outlives the moment it was taken.
+"The server did not answer" is true when measured and false ten seconds
+later, so it may be returned but never stored. And a control whose options
+come from a filtered list needs an answer for the empty case: offering one
+option is indistinguishable, to the person using it, from ignoring the click.
diff --git a/jarvis/brain/ollama_overview.py b/jarvis/brain/ollama_overview.py
index 7bc5f443f1..260d4a2c84 100644
--- a/jarvis/brain/ollama_overview.py
+++ b/jarvis/brain/ollama_overview.py
@@ -53,6 +53,7 @@ __all__ = [
"forget",
"get_overview",
"inventory_payload",
+ "is_paintable",
"load_snapshot",
"model_row",
"role_row",
@@ -325,13 +326,40 @@ def load_snapshot(path: Path | None = None) -> dict[str, Any] | None:
return data
+def is_paintable(payload: dict[str, Any]) -> bool:
+ """Whether ``payload`` may be cached as the next open's head start.
+
+ A sweep taken while the server was unreachable — Ollama still booting,
+ a stopped service, a remote host off the network — is a truthful answer
+ to "what can I see right now", but a ruinous thing to paint later: its
+ inventory is empty, so every role reads "not installed", every shortlist
+ reads "nothing installed fits this job yet", and each role's picker
+ offers only the tag already configured. Keeping the previous, honest
+ snapshot instead means the next open paints stale facts rather than a
+ machine that looks wiped (BUG-188).
+
+ Only that one condition disqualifies a payload: this is a guard against a
+ known lie, not a schema check, so anything else is cached as before.
+ """
+ inv = payload.get("inventory")
+ return not (isinstance(inv, dict) and inv.get("error"))
+
+
def save_snapshot(root: str, payload: dict[str, Any], path: Path | None = None) -> None:
"""Persist ``payload`` atomically (``.part`` + ``os.replace``).
A crash mid-write leaves the previous snapshot intact; a write failure
is logged and swallowed on purpose — the overview was already answered
- live, the cache is only the next open's head start.
+ live, the cache is only the next open's head start. A payload that fails
+ :func:`is_paintable` is not written at all.
"""
+ if not is_paintable(payload):
+ log.info(
+ "local-models: not caching the overview for %s — the server was "
+ "unreachable, so its inventory is empty; the previous snapshot stays",
+ root,
+ )
+ return
target = path or snapshot_path()
part = target.with_suffix(target.suffix + ".part")
data = {"root": root, "fetched_at": float(payload.get("fetched_at") or time.time())}
@@ -351,6 +379,9 @@ def save_snapshot(root: str, payload: dict[str, Any], path: Path | None = None)
# ── Stale-while-revalidate ───────────────────────────────────────────────
_memo: dict[str, tuple[float, dict[str, Any]]] = {} # root -> (monotonic, payload)
+#: Roots whose memo was dropped by a write and whose disk snapshot is
+#: therefore known-stale: the next open builds live instead of painting it.
+_dirty: set[str] = set()
#: The one background refresh per root; the reference is kept so the task
#: cannot be garbage-collected mid-flight, and a second open joins it.
_refresh_tasks: dict[str, asyncio.Task[dict[str, Any]]] = {}
@@ -360,6 +391,7 @@ def _reset_for_tests() -> None:
"""Drop the in-memory overview memo and refresh bookkeeping (tests only)."""
_memo.clear()
_refresh_tasks.clear()
+ _dirty.clear()
def forget(root: str) -> None:
@@ -368,10 +400,13 @@ def forget(root: str) -> None:
A role pick, a tune, an unload or a delete changes what the next open
must show; without this the memo answered the OLD payload for up to
:data:`DEFAULT_MAX_AGE_S` after the write and the row the user had just
- changed stayed where it was. The disk snapshot is left alone — the next
- live build overwrites it.
+ changed stayed where it was. The disk snapshot is known-stale from here
+ on too, so the next plain read builds live rather than painting it —
+ every client sees the write, not only one that asks for ``fresh``.
"""
- _memo.pop(normalize_server_root(root), None)
+ key = normalize_server_root(root)
+ _memo.pop(key, None)
+ _dirty.add(key)
def _fresh(root: str, max_age_s: float) -> dict[str, Any] | None:
@@ -385,6 +420,7 @@ async def _refresh(root: str, cfg: Any, provider_id: str) -> dict[str, Any]:
"""Build live, memoise, save to disk (off-loop); returns the payload."""
payload = await build_overview(root, cfg, provider_id=provider_id)
_memo[root] = (time.monotonic(), payload)
+ _dirty.discard(root)
await asyncio.to_thread(save_snapshot, root, payload)
return payload
@@ -431,12 +467,22 @@ async def get_overview(
"""
root = normalize_server_root(root)
disk: dict[str, Any] | None = None
- if not fresh:
+ # After a write (`forget`) both the memo and the disk snapshot are known
+ # stale, so a plain read goes straight to the live build as well.
+ if not fresh and root not in _dirty:
if (hit := _fresh(root, max_age_s)) is not None:
return hit, "live"
disk = await asyncio.to_thread(load_snapshot)
if disk is not None and disk.get("root") != root:
disk = None
+ # A snapshot written before `is_paintable` existed can still hold an
+ # offline sweep; painting it would show an empty machine for up to
+ # STALE_AFTER_S. Build live instead and let that answer replace it.
+ if disk is not None and not is_paintable(disk["payload"]):
+ log.info(
+ "local-models: the saved overview for %s is an offline sweep; rebuilding", root
+ )
+ disk = None
if disk is not None and time.time() - float(disk["fetched_at"]) < STALE_AFTER_S:
_schedule_refresh(root, cfg, provider_id)
return dict(disk["payload"]), "cache"
diff --git a/jarvis/ui/web/frontend/src/hooks/useLocalModels.ts b/jarvis/ui/web/frontend/src/hooks/useLocalModels.ts
index 0035186545..9785a6fced 100644
--- a/jarvis/ui/web/frontend/src/hooks/useLocalModels.ts
+++ b/jarvis/ui/web/frontend/src/hooks/useLocalModels.ts
@@ -19,6 +19,7 @@ import {
import { useEffect, useMemo } from "react";
import {
+ clearOverviewSnapshot,
readOverviewSnapshot,
writeOverviewSnapshot,
} from "../lib/localModelsSnapshot";
@@ -1081,22 +1082,57 @@ export function useEnvGuide(providerId: string | undefined, os?: EnvGuideOs) {
* next poll. `?fresh=1` skips the cache; the answer replaces the cached
* payload under the same key, so every panel repaints from it.
*/
-export function refetchOverviewFresh(
+export async function refetchOverviewFresh(
qc: QueryClient,
providerId: string,
): Promise<OverviewResponse | undefined> {
- return qc
- .fetchQuery({
- queryKey: localModelsKeys.overview(providerId),
+ const queryKey = localModelsKeys.overview(providerId);
+ try {
+ // A poll may be in flight with the cache-first answer; `fetchQuery`
+ // would join it instead of running the fresh read and the stale payload
+ // would land AFTER the write. Cancel it first, then read live.
+ await qc.cancelQueries({ queryKey });
+ return await qc.fetchQuery({
+ queryKey,
queryFn: () => getOverview(providerId, true),
staleTime: 0,
- })
- .catch((err: unknown) => {
- // The write already landed; a failed re-read only delays the repaint
- // until the next poll, so it is logged rather than surfaced.
- console.warn("[local-models] fresh overview failed", err);
- return undefined;
});
+ } catch (err) {
+ // The write already landed; a failed re-read only delays the repaint
+ // until the next poll, so it is logged rather than surfaced.
+ console.warn("[local-models] fresh overview failed", err);
+ return undefined;
+ }
+}
+
+/** "qwen3.5" and "qwen3.5:latest" are the same download. */
+export function canonicalModelName(name: string): string {
+ return name.endsWith(":latest") ? name.slice(0, -":latest".length) : name;
+}
+
+/**
+ * The overview as it will read once `role` is on `model` — the optimistic
+ * shape a pick shows while the write and the fresh read are underway.
+ */
+export function withRolePick(
+ overview: OverviewResponse,
+ role: LocalModelRole,
+ model: string,
+): OverviewResponse {
+ const installed =
+ model !== "" &&
+ overview.inventory.models.some(
+ (m) => canonicalModelName(m.name) === canonicalModelName(model),
+ );
+ return {
+ ...overview,
+ roles: {
+ ...overview.roles,
+ roles: overview.roles.roles.map((r) =>
+ r.id === role ? { ...r, current: model, installed } : r,
+ ),
+ },
+ };
}
export function useInvalidateLocalModels(providerId: string | undefined) {
@@ -1114,13 +1150,51 @@ export function useInvalidateLocalModels(providerId: string | undefined) {
};
}
+/**
+ * Throw away everything the section has cached and read it again from the
+ * server — the "Refresh" button behind the section header.
+ *
+ * `useInvalidateLocalModels` refetches, but it cannot help against the one
+ * state a user actually gets stuck in: a painted snapshot taken while the
+ * server was silent. That snapshot is `initialData`, so it survives a
+ * refetch failure and keeps a wiped-looking machine on screen. Clearing it
+ * first makes the button mean what it says (BUG-188).
+ */
+export function useReloadLocalModels(providerId: string | undefined) {
+ const qc = useQueryClient();
+ return async () => {
+ const id = providerId ?? "";
+ if (!id) return;
+ clearOverviewSnapshot(id);
+ await qc.cancelQueries({ queryKey: ["local-models", id] });
+ qc.removeQueries({ queryKey: ["local-models", id] });
+ await refetchOverviewFresh(qc, id);
+ };
+}
+
export function useSetRole(providerId: string | undefined) {
const qc = useQueryClient();
+ const id = providerId ?? "";
+ const overviewKey = localModelsKeys.overview(id);
return useMutation({
mutationFn: ({ role, model }: { role: LocalModelRole; model: string }) =>
setRole(providerId as string, role, model),
- onSuccess: () => {
- const id = providerId ?? "";
+ // The picker is a controlled input bound to the overview. Without this
+ // it snapped back to the OLD model the instant a pick was made and
+ // stayed there until the fresh read landed (one to three seconds on a
+ // live build) — which reads as "my pick was thrown away". The cache
+ // shows the pick at once; a refused write rolls it back.
+ onMutate: async ({ role, model }) => {
+ await qc.cancelQueries({ queryKey: overviewKey });
+ const previous = qc.getQueryData<OverviewResponse>(overviewKey);
+ if (previous)
+ qc.setQueryData(overviewKey, withRolePick(previous, role, model));
+ return { previous };
+ },
+ onError: (_err, _vars, context) => {
+ if (context?.previous) qc.setQueryData(overviewKey, context.previous);
+ },
+ onSettled: () => {
void qc.invalidateQueries({ queryKey: localModelsKeys.roles(id) });
void qc.invalidateQueries({ queryKey: localModelsKeys.inventory(id) });
void refetchOverviewFresh(qc, id);
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 264c5579b6..581b702d10 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
@@ -6,6 +6,7 @@
"mode_label": "Detailgrad",
"mode_simple": "Einfach",
"mode_advanced": "Erweitert",
+ "reload": "Aktualisieren",
"tabs_label": "Bereiche der lokalen Modelle",
"tab_overview": "Überblick",
"tab_models": "Modelle",
@@ -196,11 +197,20 @@
"status_gpu_unknown": "Grafikspeicher unbekannt",
"status_brain_other": "{server} ist nicht dein aktives Gehirn ({brain} antwortet)",
"status_brain_active": "{server} ist dein aktives Gehirn",
- "actions_label": "Nächste Schritte",
+ "step_server_title": "Server",
+ "step_server_subtitle": "Das Programm, das die Modelle auf diesem Rechner ausführt.",
+ "step_models_title": "Modelle",
+ "step_models_subtitle": "Die Downloads auf dem Server. Jeder Job wählt aus diesen aus.",
+ "step_jobs_title": "Jobs",
+ "step_jobs_subtitle": "Welches Modell welchen Job übernimmt. Hier änderst du jeden davon.",
+ "models_count": "{count} Modelle · {size} auf der Festplatte",
+ "models_one": "1 Modell · {size} auf der Festplatte",
+ "models_unknown": "Der Server hat nicht geantwortet; was installiert ist, ist gerade unbekannt.",
+ "jobs_filled": "{filled} von {total} Jobs laufen auf einem installierten Modell",
+ "fact_version": "Version",
+ "fact_address": "Adresse",
+ "fact_brain": "Gehirn",
"action_browse": "Modelle durchsuchen",
- "action_browse_hint": "Einen Download aus dem Katalog wählen.",
- "disk_models": "{count} Modelle",
- "disk_model_one": "1 Modell",
"gpu_in_use": "{gb} belegt",
"gpu_in_use_none": "nichts belegt",
"loaded_expires": "läuft in {when} ab",
@@ -274,6 +284,9 @@
"pick_discovery": "Jarvis wählen lassen",
"pick_none": "Modell wählen",
"pick_missing_suffix": "(nicht installiert)",
+ "pick_group_fits": "Passt zu diesem Job",
+ "pick_group_others": "Weitere installierte Modelle",
+ "server_silent": "Der Server hat nicht geantwortet, daher kann diese Liste unvollständig sein. Aktualisiere, sobald er läuft.",
"use_recommended": "Empfehlung nutzen",
"download_recommended": "{model} herunterladen",
"tune": "Einstellen",
@@ -290,10 +303,7 @@
"readback_defaults": "Läuft mit den eigenen Vorgaben des Modells.",
"tune_skipped": "Zugewiesen; die Einstellungen bleiben bei den Vorgaben des Modells.",
"voice_context_auto": "Läuft mit {context} Kontext — nach dem Speicher dieses Rechners bemessen. Über „Abstimmen“ selbst festlegen.",
- "voice_context_manual": "Läuft mit {context} Kontext — von dir unter „Abstimmen“ festgelegt.",
- "checklist_not_set": "nicht gesetzt",
- "checklist_change": "Ändern",
- "checklist_done": "Fertig"
+ "voice_context_manual": "Läuft mit {context} Kontext — von dir unter „Abstimmen“ festgelegt."
},
"server": {
"host_eyebrow": "Host",
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 380a3a0f5d..c4f850da98 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
@@ -6,6 +6,7 @@
"mode_label": "Detail level",
"mode_simple": "Simple",
"mode_advanced": "Advanced",
+ "reload": "Refresh",
"tabs_label": "Local model areas",
"tab_overview": "Overview",
"tab_models": "Models",
@@ -196,11 +197,20 @@
"status_gpu_unknown": "graphics memory unknown",
"status_brain_other": "{server} is not your active brain ({brain} answers)",
"status_brain_active": "{server} is your active brain",
- "actions_label": "Next steps",
+ "step_server_title": "Server",
+ "step_server_subtitle": "The program that runs the models on this machine.",
+ "step_models_title": "Models",
+ "step_models_subtitle": "The downloads on the server. Every job picks from these.",
+ "step_jobs_title": "Jobs",
+ "step_jobs_subtitle": "Which model answers for which job. Change any of them here.",
+ "models_count": "{count} models · {size} on disk",
+ "models_one": "1 model · {size} on disk",
+ "models_unknown": "The server did not answer, so what is installed is unknown right now.",
+ "jobs_filled": "{filled} of {total} jobs are on an installed model",
+ "fact_version": "Version",
+ "fact_address": "Address",
+ "fact_brain": "Brain",
"action_browse": "Browse models",
- "action_browse_hint": "Pick a download from the catalogue.",
- "disk_models": "{count} models",
- "disk_model_one": "1 model",
"gpu_in_use": "{gb} in use",
"gpu_in_use_none": "nothing in use",
"loaded_expires": "expires in {when}",
@@ -274,6 +284,9 @@
"pick_discovery": "Let Jarvis pick",
"pick_none": "Choose a model",
"pick_missing_suffix": "(not installed)",
+ "pick_group_fits": "Fits this job",
+ "pick_group_others": "Other installed models",
+ "server_silent": "The server did not answer, so this list may be incomplete. Refresh once it is running.",
"use_recommended": "Use recommended",
"download_recommended": "Download {model}",
"tune": "Tune",
@@ -290,10 +303,7 @@
"readback_defaults": "Runs with the model's own defaults.",
"tune_skipped": "Assigned; the settings stay at the model's defaults.",
"voice_context_auto": "Runs with a {context} context — sized from this machine's memory. Tune to set it yourself.",
- "voice_context_manual": "Runs with a {context} context — set by you in Tune.",
- "checklist_not_set": "not set",
- "checklist_change": "Change",
- "checklist_done": "Done"
+ "voice_context_manual": "Runs with a {context} context — set by you in Tune."
},
"server": {
"host_eyebrow": "Host",
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 dec5d1d2e2..81b2b678d2 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
@@ -6,6 +6,7 @@
"mode_label": "Nivel de detalle",
"mode_simple": "Sencillo",
"mode_advanced": "Avanzado",
+ "reload": "Actualizar",
"tabs_label": "Áreas de modelos locales",
"tab_overview": "Resumen",
"tab_models": "Modelos",
@@ -196,11 +197,20 @@
"status_gpu_unknown": "memoria gráfica desconocida",
"status_brain_other": "{server} no es tu cerebro activo (responde {brain})",
"status_brain_active": "{server} es tu cerebro activo",
- "actions_label": "Siguientes pasos",
+ "step_server_title": "Servidor",
+ "step_server_subtitle": "El programa que ejecuta los modelos en este equipo.",
+ "step_models_title": "Modelos",
+ "step_models_subtitle": "Las descargas en el servidor. Cada tarea elige entre estas.",
+ "step_jobs_title": "Tareas",
+ "step_jobs_subtitle": "Qué modelo responde a cada tarea. Cambia cualquiera aquí.",
+ "models_count": "{count} modelos · {size} en disco",
+ "models_one": "1 modelo · {size} en disco",
+ "models_unknown": "El servidor no respondió, así que ahora se desconoce qué hay instalado.",
+ "jobs_filled": "{filled} de {total} tareas usan un modelo instalado",
+ "fact_version": "Versión",
+ "fact_address": "Dirección",
+ "fact_brain": "Cerebro",
"action_browse": "Explorar modelos",
- "action_browse_hint": "Elige una descarga del catálogo.",
- "disk_models": "{count} modelos",
- "disk_model_one": "1 modelo",
"gpu_in_use": "{gb} en uso",
"gpu_in_use_none": "nada en uso",
"loaded_expires": "caduca en {when}",
@@ -274,6 +284,9 @@
"pick_discovery": "Dejar que Jarvis elija",
"pick_none": "Elige un modelo",
"pick_missing_suffix": "(no instalado)",
+ "pick_group_fits": "Encaja con esta tarea",
+ "pick_group_others": "Otros modelos instalados",
+ "server_silent": "El servidor no respondió, así que esta lista puede estar incompleta. Actualiza cuando esté en marcha.",
"use_recommended": "Usar el recomendado",
"download_recommended": "Descargar {model}",
"tune": "Ajustar",
@@ -290,10 +303,7 @@
"readback_defaults": "Corre con los valores propios del modelo.",
"tune_skipped": "Asignado; los ajustes se quedan en los valores del modelo.",
"voice_context_auto": "Funciona con un contexto de {context}, dimensionado según la memoria de este equipo. Ajústalo tú en «Ajustar».",
- "voice_context_manual": "Funciona con un contexto de {context}, fijado por ti en «Ajustar».",
- "checklist_not_set": "sin definir",
- "checklist_change": "Cambiar",
- "checklist_done": "Listo"
+ "voice_context_manual": "Funciona con un contexto de {context}, fijado por ti en «Ajustar»."
},
"server": {
"host_eyebrow": "Host",
diff --git a/jarvis/ui/web/frontend/src/lib/localModelsSnapshot.test.ts b/jarvis/ui/web/frontend/src/lib/localModelsSnapshot.test.ts
index 32a5d596d2..5a162d07d4 100644
--- a/jarvis/ui/web/frontend/src/lib/localModelsSnapshot.test.ts
+++ b/jarvis/ui/web/frontend/src/lib/localModelsSnapshot.test.ts
@@ -61,6 +61,32 @@ describe("localModelsSnapshot", () => {
expect(readOverviewSnapshot("ollama")).toBeNull();
});
+ it("neither stores nor reads a sweep taken while the server was silent", () => {
+ // Ollama still booting: the payload is truthful now and ruinous later —
+ // an empty inventory paints a machine that looks wiped, and each role's
+ // picker then offers only the tag already configured (BUG-188).
+ const offline = payload({
+ inventory: {
+ models: [],
+ error: "Ollama is not answering on http://localhost:11434.",
+ } as unknown as OverviewResponse["inventory"],
+ });
+ expect(writeOverviewSnapshot("ollama", offline)).toBe(false);
+ expect(readOverviewSnapshot("ollama")).toBeNull();
+
+ // A good snapshot already on disk survives the offline sweep.
+ writeOverviewSnapshot("ollama", payload());
+ expect(writeOverviewSnapshot("ollama", offline)).toBe(false);
+ expect(readOverviewSnapshot("ollama")?.fetched_at).toBe(1_700_000_000);
+
+ // One written by an older build is dropped on read rather than painted.
+ window.localStorage.setItem(
+ snapshotKey("ollama"),
+ JSON.stringify({ v: SNAPSHOT_VERSION, data: offline }),
+ );
+ expect(readOverviewSnapshot("ollama")).toBeNull();
+ });
+
it("keeps one entry per provider", () => {
writeOverviewSnapshot("ollama", payload());
expect(readOverviewSnapshot("other")).toBeNull();
diff --git a/jarvis/ui/web/frontend/src/lib/localModelsSnapshot.ts b/jarvis/ui/web/frontend/src/lib/localModelsSnapshot.ts
index 6320469b48..49b1647ca1 100644
--- a/jarvis/ui/web/frontend/src/lib/localModelsSnapshot.ts
+++ b/jarvis/ui/web/frontend/src/lib/localModelsSnapshot.ts
@@ -30,6 +30,21 @@ export function snapshotKey(providerId: string): string {
return `${KEY_PREFIX}${providerId}`;
}
+/**
+ * Whether a payload is safe to paint from later.
+ *
+ * A sweep taken while the server was unreachable — still booting, stopped,
+ * a remote host off the network — carries an empty inventory. Painting that
+ * on the next open shows a machine that looks wiped: every role reads "not
+ * installed", every shortlist reads "nothing installed fits this job yet",
+ * and each role's picker offers only the tag already configured, so no other
+ * model can be chosen at all. Such a payload is neither stored nor read back;
+ * the previous, honest snapshot stays (BUG-188).
+ */
+export function isPaintable(data: OverviewResponse): boolean {
+ return !data.inventory?.error;
+}
+
/**
* Synchronously read the last overview for `providerId`; `null` when there
* is none, the version does not match, the JSON is broken or storage is
@@ -53,6 +68,7 @@ export function readOverviewSnapshot(
!data.recommended
)
return null;
+ if (!isPaintable(data)) return null;
return { ...data, source: "cache" };
} catch {
return null;
@@ -60,14 +76,16 @@ export function readOverviewSnapshot(
}
/**
- * Persist an overview for the next paint. Oversized payloads and blocked
- * storage are silent no-ops — the section still works, it just paints a
- * beat later next time.
+ * Persist an overview for the next paint. Oversized payloads, a sweep taken
+ * while the server was unreachable (see `isPaintable`) and blocked storage
+ * are silent no-ops — the section still works, it just paints a beat later
+ * next time.
*/
export function writeOverviewSnapshot(
providerId: string,
data: OverviewResponse,
): boolean {
+ if (!isPaintable(data)) return false;
try {
const env: Envelope = { v: SNAPSHOT_VERSION, data };
const json = JSON.stringify(env);
diff --git a/jarvis/ui/web/frontend/src/views/LocalModelsView.test.tsx b/jarvis/ui/web/frontend/src/views/LocalModelsView.test.tsx
index bcfbdc64fd..4ef1d187ce 100644
--- a/jarvis/ui/web/frontend/src/views/LocalModelsView.test.tsx
+++ b/jarvis/ui/web/frontend/src/views/LocalModelsView.test.tsx
@@ -82,11 +82,14 @@ vi.mock("@/views/local-models/TuneSheet", () => ({
</div>
),
}));
+const mockReload = vi.fn(async () => {});
+
vi.mock("@/hooks/useLocalModels", () => ({
useInventory: () => ({
isLoading: false,
data: { models: [{ name: "qwen3.5:4b" }] },
}),
+ useReloadLocalModels: () => mockReload,
}));
import {
@@ -246,4 +249,12 @@ describe("LocalModelsView", () => {
expect(screen.getByTestId("catalogue-panel")).toBeDefined();
expect(screen.queryByTestId("overview-panel")).toBeNull();
});
+
+ it("Refresh throws the section's cached answers away and reads again", () => {
+ render(<LocalModelsView />);
+ fireEvent.click(
+ screen.getByRole("button", { name: "local_models.reload" }),
+ );
+ expect(mockReload).toHaveBeenCalledTimes(1);
+ });
});
diff --git a/jarvis/ui/web/frontend/src/views/LocalModelsView.tsx b/jarvis/ui/web/frontend/src/views/LocalModelsView.tsx
index a49e87b99a..2bf5e93a9f 100644
--- a/jarvis/ui/web/frontend/src/views/LocalModelsView.tsx
+++ b/jarvis/ui/web/frontend/src/views/LocalModelsView.tsx
@@ -17,14 +17,16 @@
* for `/api/providers`; the list still resolves and corrects the seed.
*/
import { useCallback, useEffect, useMemo, useState } from "react";
+import { Loader2, RefreshCw } from "lucide-react";
import { ScrollArea } from "@/components/ui/scroll-area";
import {
BackLink,
Panel,
PanelHeader,
SegmentedFilter,
+ SoftButton,
} from "@/components/extensions/primitives";
-import { useInventory } from "@/hooks/useLocalModels";
+import { useInventory, useReloadLocalModels } from "@/hooks/useLocalModels";
import { useProviders } from "@/hooks/useProviders";
import {
LOCAL_MODELS_MARK_MOUNT,
@@ -153,6 +155,17 @@ export function LocalModelsView() {
setMode(next);
storeMode(next);
}, []);
+
+ // Refresh: drops the painted snapshot and every cached answer, then reads
+ // the server again. The one recovery from a snapshot taken while the
+ // server was silent, which otherwise paints a wiped machine (BUG-188).
+ const reload = useReloadLocalModels(providerId ?? undefined);
+ const [reloading, setReloading] = useState(false);
+ const onReload = useCallback(() => {
+ if (reloading) return;
+ setReloading(true);
+ void reload().finally(() => setReloading(false));
+ }, [reload, reloading]);
// "Manage" on the overview's installed list: the full ledger lives on the
// Advanced-only Models tab, so the switch flips along with the tab.
const openManage = useCallback(() => {
@@ -202,6 +215,19 @@ export function LocalModelsView() {
options={tabs}
/>
<div className="flex items-center gap-2">
+ <SoftButton
+ onClick={onReload}
+ disabled={!providerId || reloading}
+ ariaLabel={t("local_models.reload")}
+ className="h-8"
+ >
+ {reloading ? (
+ <Loader2 className="h-3.5 w-3.5 animate-spin" />
+ ) : (
+ <RefreshCw className="h-3.5 w-3.5" />
+ )}
+ {t("local_models.reload")}
+ </SoftButton>
<span className="text-[11px] uppercase tracking-[0.14em] text-muted-foreground">
{t("local_models.mode_label")}
</span>
diff --git a/jarvis/ui/web/frontend/src/views/local-models/InstalledPanel.tsx b/jarvis/ui/web/frontend/src/views/local-models/InstalledPanel.tsx
index 6b88135532..dcac88d3ea 100644
--- a/jarvis/ui/web/frontend/src/views/local-models/InstalledPanel.tsx
+++ b/jarvis/ui/web/frontend/src/views/local-models/InstalledPanel.tsx
@@ -35,6 +35,11 @@ export interface InstalledPanelProps {
error?: string | null;
/** Opens the full ledger (Advanced → Models); the button hides without it. */
onManage?: () => void;
+ /**
+ * Drop the card frame and the title row: the caller already draws both
+ * (the overview's `StepCard`), and a card inside a card reads as clutter.
+ */
+ bare?: boolean;
}
/** The capabilities a role is gated on; the rest is noise on a one-liner. */
@@ -74,6 +79,7 @@ export function InstalledPanel({
loading = false,
error = null,
onManage,
+ bare = false,
}: InstalledPanelProps) {
const t = useT();
const k = (key: string) => t(`local_models.installed.${key}`);
@@ -113,18 +119,19 @@ export function InstalledPanel({
size: formatGb(diskBytes),
});
- return (
- <Panel className="p-4">
- <div className="space-y-3" data-testid="local-models-installed">
- <div className="flex flex-wrap items-start justify-between gap-3">
- <PanelHeader title={k("title")} subtitle={subtitle} />
- {onManage && models.length > 0 && (
- <SoftButton onClick={onManage} ariaLabel={k("manage")}>
- <Database className="h-3.5 w-3.5" />
- {k("manage")}
- </SoftButton>
- )}
- </div>
+ const body = (
+ <div className="space-y-3" data-testid="local-models-installed">
+ {!bare && (
+ <div className="flex flex-wrap items-start justify-between gap-3">
+ <PanelHeader title={k("title")} subtitle={subtitle} />
+ {onManage && models.length > 0 && (
+ <SoftButton onClick={onManage} ariaLabel={k("manage")}>
+ <Database className="h-3.5 w-3.5" />
+ {k("manage")}
+ </SoftButton>
+ )}
+ </div>
+ )}
{error && (
<p
@@ -206,7 +213,8 @@ export function InstalledPanel({
})}
</ul>
)}
- </div>
- </Panel>
+ </div>
);
+ if (bare) return body;
+ return <Panel className="p-4">{body}</Panel>;
}
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 b3bb5e70c5..997cd463e5 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
@@ -345,7 +345,7 @@ describe("OverviewPanel paint-first", () => {
name: "local_models.overview.disk",
});
expect(disk.textContent).toContain("80 GB");
- expect(screen.getByTestId("overview-status").textContent).toContain(
+ expect(screen.getByTestId("step-server").textContent).toContain(
"local_models.overview.status_runningOllama|0.32.15",
);
expect(screen.getByTestId("role-row-chat")).toBeDefined();
@@ -397,22 +397,22 @@ describe("OverviewPanel paint-first", () => {
});
describe("OverviewPanel", () => {
- it("opens with one status sentence: server, graphics memory, active brain", async () => {
+ it("step 1 says whether the server runs, on what hardware, and who the brain is", async () => {
installFetchMock();
renderPanel();
- const status = await screen.findByTestId("overview-status");
+ const status = await screen.findByTestId("step-server");
await waitFor(() =>
expect(status.textContent).toContain(
"local_models.overview.status_runningOllama|0.32.15",
),
);
- expect(status.textContent).toContain(
- "local_models.overview.status_gpu16.0",
- );
+ expect(status.textContent).toContain("16.0 GB");
expect(status.textContent).toContain(
"local_models.overview.status_brain_activeOllama",
);
+ // A running server satisfies step 1: the marker is a check, not a "1".
+ expect(status.getAttribute("data-state")).toBe("done");
});
it("names the other brain in the status sentence when Ollama is not active", async () => {
@@ -429,35 +429,42 @@ describe("OverviewPanel", () => {
);
});
- it("offers one action: browse the catalogue", async () => {
+ it("browsing the catalogue is step 2's action", async () => {
installFetchMock();
const onBrowse = vi.fn();
renderPanel({ onBrowse });
- await screen.findByTestId("overview-actions");
+ const models = await screen.findByTestId("step-models");
fireEvent.click(
- screen.getByRole("button", { name: "local_models.overview.action_browse" }),
+ within(models).getByRole("button", {
+ name: "local_models.overview.action_browse",
+ }),
);
expect(onBrowse).toHaveBeenCalledTimes(1);
});
- it("the detail level changes the roles on this very screen", async () => {
+ it("the detail level adds detail to the jobs but never takes the picker away", async () => {
installFetchMock();
- const { unmount } = renderPanel();
- const simple = await screen.findByTestId("local-models-roles");
- expect(simple.getAttribute("data-variant")).toBe("checklist");
+ const { unmount } = renderPanel({ onTune: vi.fn() });
+ await screen.findByTestId("role-row-chat");
+ // Simple: the picker is there, the capability badges and Tune are not.
+ expect(screen.getByTestId("role-picker-chat")).toBeDefined();
+ expect(screen.queryByText("tools")).toBeNull();
+ expect(screen.queryByText("local_models.roles.tune")).toBeNull();
unmount();
installFetchMock();
- renderPanel({ advanced: true });
- const advanced = await screen.findByTestId("local-models-roles");
- expect(advanced.getAttribute("data-variant")).not.toBe("checklist");
+ renderPanel({ advanced: true, onTune: vi.fn() });
+ await screen.findByTestId("role-row-chat");
+ expect(screen.getByTestId("role-picker-chat")).toBeDefined();
+ expect(screen.getAllByText("tools").length).toBeGreaterThan(0);
+ expect(screen.getByText("local_models.roles.tune")).toBeDefined();
});
it("keeps 'Set up everything' even when no browse handler is wired", async () => {
installFetchMock();
renderPanel();
- await screen.findByTestId("overview-status");
+ await screen.findByTestId("step-server");
expect(
screen.queryByRole("button", {
name: "local_models.overview.action_browse",
@@ -487,10 +494,11 @@ describe("OverviewPanel", () => {
name: "local_models.overview.disk",
});
await waitFor(() => expect(disk.textContent).toContain("12 GB"));
- await waitFor(() =>
- expect(disk.textContent).toContain("local_models.overview.disk_model_one"),
- );
expect(disk.textContent).toContain("/home/x/.ollama/models");
+ // How many downloads there are is step 2's own status line.
+ expect(
+ (await screen.findByTestId("step-models")).textContent,
+ ).toContain("local_models.overview.models_one");
const gpu = screen.getByRole("group", {
name: "local_models.overview.gpu",
@@ -512,67 +520,53 @@ describe("OverviewPanel", () => {
});
renderPanel();
- const status = await screen.findByTestId("overview-status");
+ const status = await screen.findByTestId("step-server");
await waitFor(() =>
expect(status.textContent).toContain(
"local_models.overview.status_stoppedOllama",
),
);
- expect(screen.getByText("local_models.overview.gpu_unknown")).toBeDefined();
+ // Named in step 1's facts and again on step 2's tile.
+ expect(
+ screen.getAllByText(/local_models\.overview\.gpu_unknown/).length,
+ ).toBeGreaterThan(0);
});
- it("renders the roles as a checklist, expands one row to the full ledger row and folds it back", async () => {
+ it("gives every job its picker straight away, listing more than the qualifying models", async () => {
installFetchMock();
- renderPanel({ onTune: vi.fn() });
+ renderPanel({ advanced: true, onTune: vi.fn() });
- const chat = await screen.findByTestId("role-row-chat");
- expect(chat.getAttribute("data-variant")).toBe("checklist");
+ await screen.findByTestId("role-row-chat");
expect(screen.getByTestId("role-row-tools_screen")).toBeDefined();
expect(screen.getByTestId("role-row-deep")).toBeDefined();
expect(screen.getByTestId("role-row-embedding")).toBeDefined();
expect(screen.queryByText("local_models.role_voice")).toBeNull();
- // Compact rows: the model, "not set", no badges, no picker, no Tune.
- // (Scoped to the ledger: the installed list below names the model too.)
- expect(
- within(screen.getByTestId("roles-ledger")).getByText("qwen3.5:4b", {
- selector: "span",
- }),
- ).toBeDefined();
- expect(
- screen.getAllByText("local_models.roles.checklist_not_set").length,
- ).toBe(3);
- expect(screen.queryByText("tools")).toBeNull();
- expect(screen.queryByText("local_models.roles.tune")).toBeNull();
- // Embedding has no recommendation applied yet -> "Download ..."; the
- // trailing action is exactly one per row.
- expect(
- screen.getByText("local_models.roles.download_recommendedqwen3-embedding:4b"),
- ).toBeDefined();
- // Expand chat via its line -> the full row with badges, picker, Tune, Done.
- fireEvent.click(
- screen.getByRole("button", { name: /local_models.role_chat/ }),
- );
- expect(screen.getAllByText("tools").length).toBeGreaterThan(0);
- expect(screen.getByText("local_models.roles.tune")).toBeDefined();
+ // No expanding: the control that changes a job is on screen for each one.
+ for (const role of ["chat", "tools_screen", "deep", "embedding"])
+ expect(screen.getByTestId(`role-picker-${role}`)).toBeDefined();
+
+ // Embedding qualifies for nothing installed, yet its picker still offers
+ // the installed downloads — the state that used to leave a user with a
+ // single option and no way to change it (BUG-188).
+ const embedding = screen.getByTestId(
+ "role-picker-embedding",
+ ) as HTMLSelectElement;
+ const offered = [...embedding.options].map((o) => o.value);
+ expect(offered).toContain("qwen3.5:4b");
expect(
- screen.getByLabelText("local_models.roles.pick_labellocal_models.role_chat"),
+ screen.getByText("local_models.roles.download_recommendedqwen3-embedding:4b"),
).toBeDefined();
- fireEvent.click(screen.getByText("local_models.roles.checklist_done"));
- expect(screen.queryByText("local_models.roles.tune")).toBeNull();
fireEvent.click(screen.getByText("local_models.roles.more_roles"));
expect(screen.getByText("local_models.role_voice")).toBeDefined();
});
- it("writes the pick through PUT roles/{role} from the expanded row", async () => {
+ it("writes the pick through PUT roles/{role}", async () => {
const fetchMock = installFetchMock();
renderPanel();
await screen.findByTestId("role-row-deep");
- fireEvent.click(
- screen.getByRole("button", { name: /local_models.role_deep/ }),
- );
const select = screen.getByLabelText(
"local_models.roles.pick_labellocal_models.role_deep",
);
@@ -670,7 +664,7 @@ describe("OverviewPanel", () => {
];
renderPanel();
- await screen.findByTestId("overview-actions");
+ await screen.findByTestId("step-server");
fireEvent.click(
screen.getByRole("button", { name: "local_models.overview.action_setup" }),
);
@@ -735,9 +729,13 @@ describe("OverviewPanel", () => {
const onManage = vi.fn();
renderPanel({ onManage });
- const list = await screen.findByTestId("local-models-installed");
+ await screen.findByTestId("local-models-installed");
await screen.findByTestId("installed-qwen3.5:4b");
- expect(list.textContent).toContain("local_models.installed.subtitle2|");
+ // How many downloads and how much disk is step 2's status line; the list
+ // itself no longer repeats it inside the card.
+ expect(
+ (await screen.findByTestId("step-models")).textContent,
+ ).toContain("local_models.overview.models_count2|");
// Each line says which roles it is the pick for (the fixture's rows
// carry no used_by, so the markers are the recommendations).
expect(
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 af5682b49d..d59713d461 100644
--- a/jarvis/ui/web/frontend/src/views/local-models/OverviewPanel.tsx
+++ b/jarvis/ui/web/frontend/src/views/local-models/OverviewPanel.tsx
@@ -1,16 +1,20 @@
/**
- * Overview — the Simple-mode front page of the Local models section, written
- * so it reads like a plan rather than a dashboard.
+ * Overview — the front page of the Local models section, written as three
+ * steps rather than a dashboard.
*
- * Top to bottom: one status sentence ("Ollama 0.32 running · 16 GB graphics
- * memory · Ollama is not your active brain (OpenRouter answers)"), the
- * actions — "Set up everything" first (the one click that starts the server,
- * fills every role with the best installed download and fetches only what
- * is missing; `useLocalSetup`), then "Browse models" — with the flow's
- * progress line underneath, the headline tiles (downloads on disk, graphics
- * memory, loaded now), the roles as a checklist, and every installed model
- * as one line (`InstalledPanel`), so "what do I have?" never needs the
- * Advanced switch.
+ * Running a model on your own machine is three questions in a fixed order:
+ * is there a server, are there models on it, and which model answers for
+ * which job. The page is those three questions, one numbered `StepCard`
+ * each, whose marker turns into a check once the step is satisfied — so
+ * "what is still missing" is a glance. Each card carries at most one primary
+ * action, and step 1 carries the one that does everything (`useLocalSetup`:
+ * install or start the server, fill every job with the best installed
+ * download, fetch only what is missing, verify, and remember the autostart).
+ *
+ * Every job's picker is on screen at every detail level. The Simple page used
+ * to fold them away behind a "Change" button, which is half of why the
+ * section read as "it keeps resetting my model" (BUG-188). `advanced` adds
+ * capability badges, Tune and the read-only jobs; it never removes a control.
*
* Props take `providerId` — the id of the card that declares
* `supports_model_pull` — never a provider name.
@@ -19,6 +23,7 @@ import { useCallback, useEffect, useRef, useState, type ReactNode } from "react"
import { Cpu, HardDrive, Layers, Loader2, Search, Sparkles } from "lucide-react";
import {
+ FactRows,
SoftButton,
StatTile,
StatusDot,
@@ -36,6 +41,7 @@ import { InstalledPanel } from "./InstalledPanel";
import { formatExpiry } from "./localModelsFormat";
import type { SetupStep, SetupSummary } from "./localSetup";
import { RolesPanel, type RolesPanelProps } from "./RolesPanel";
+import { StepCard, type StepState } from "./StepCard";
import { useLocalSetup } from "./useLocalSetup";
import { verifyLines } from "./verifyLines";
@@ -46,8 +52,8 @@ export interface OverviewPanelProps extends RolesPanelProps {
onManage?: () => void;
/**
* Detail level. The switch beside the rail must change what is on screen
- * HERE too, or it reads as a dead control: Simple keeps the roles a
- * checklist, Advanced opens every row to its picker, capabilities and Tune.
+ * HERE too, or it reads as a dead control: Advanced adds the capability
+ * badges, the Tune buttons and the read-only jobs.
*/
advanced?: boolean;
}
@@ -73,11 +79,13 @@ export function OverviewPanel({
const { providers, refetch: refetchProviders } = useProviders();
const setup = useLocalSetup(providerId);
- // One round-trip (or the painted snapshot) feeds every tile.
+ // One round-trip (or the painted snapshot) feeds every card.
const s = overview.data?.server;
const shortlist = { data: overview.data?.recommended };
const inventory = { data: overview.data?.inventory };
const server = { isLoading: overview.isLoading, isError: overview.isError };
+ // Painted from the snapshot while the live read runs: say so rather than
+ // present a cached machine as the current one.
const checking = overview.isFetching && overview.data?.source === "cache";
const marked = useRef(false);
@@ -91,7 +99,7 @@ export function OverviewPanel({
providers.find((p) => p.id === providerId)?.label ||
t("local_models.overview.status_generic_server");
- // --- status sentence --------------------------------------------------
+ // --- step 1: the server ----------------------------------------------
let serverClause = t("local_models.overview.status_checking");
let serverTone: "ok" | "warn" | "error" | "busy" = "busy";
if (s) {
@@ -135,32 +143,17 @@ export function OverviewPanel({
const acceleratorGb = shortlist.data?.accelerator_gb ?? 0;
const acceleratorSource = shortlist.data?.accelerator_source ?? "";
- const gpuClause =
- acceleratorGb > 0
- ? fill(t("local_models.overview.status_gpu"), {
- gb: acceleratorGb.toFixed(1),
- })
- : shortlist.data
- ? t("local_models.overview.status_gpu_unknown")
- : "";
- // The third clause only renders once the provider list is here — no gate.
+ // The brain clause only renders once the provider list is here — no gate.
const activeBrain =
providers.find((p) => p.tier === "brain" && p.active) ?? null;
const otherBrainActive =
activeBrain !== null && activeBrain.id !== providerId;
- const brainClause = activeBrain
- ? activeBrain.id === providerId
- ? fill(t("local_models.overview.status_brain_active"), {
- server: serverLabel,
- })
- : fill(t("local_models.overview.status_brain_other"), {
- server: serverLabel,
- brain: activeBrain.label,
- })
- : "";
- // --- tiles ------------------------------------------------------------
+ // --- step 2: the models ----------------------------------------------
+ const models = inventory.data?.models ?? [];
+ const modelCount = models.length;
+ const serverSilent = Boolean(inventory.data?.error);
const gpuValue =
acceleratorGb > 0
? `${acceleratorGb.toFixed(1)} GB`
@@ -179,18 +172,6 @@ export function OverviewPanel({
})
: t("local_models.overview.gpu_unknown_hint");
- const modelCount = inventory.data?.models.length ?? null;
- const diskHint = [
- modelCount === null
- ? ""
- : modelCount === 1
- ? t("local_models.overview.disk_model_one")
- : fill(t("local_models.overview.disk_models"), { count: modelCount }),
- s?.models_dir ?? "",
- ]
- .filter(Boolean)
- .join(" · ");
-
const loaded = s?.running_models ?? [];
const firstExpiry = loaded.length > 0 ? formatExpiry(loaded[0].expires_at) : "";
const loadedHint =
@@ -207,14 +188,41 @@ export function OverviewPanel({
.join(" · ")
: t("local_models.overview.loaded_none");
- const dotClass = {
- ok: "bg-emerald-500",
- warn: "bg-amber-500",
- error: "bg-destructive",
- busy: "bg-primary animate-pulse",
- }[serverTone];
+ // --- step 3: the jobs -------------------------------------------------
+ const roleRows = overview.data?.roles.roles ?? [];
+ const writableRows = roleRows.filter((r) => !r.advanced && r.writable);
+ const filledRows = writableRows.filter((r) => r.current !== "" && r.installed);
+ const roleLabel = useCallback(
+ (role: LocalModelRole) => {
+ const row = roleRows.find((r) => r.id === role);
+ return t(row?.label_key ?? `local_models.role_${role}`);
+ },
+ [roleRows, t],
+ );
+
+ // --- step states ------------------------------------------------------
+ // "Checking" is busy, not a verdict; a step only claims done on facts.
+ const serverState: StepState = !s
+ ? server.isError
+ ? "attention"
+ : "busy"
+ : s.running
+ ? "done"
+ : "attention";
+ const modelsState: StepState = !inventory.data
+ ? "busy"
+ : serverSilent
+ ? "attention"
+ : modelCount > 0
+ ? "done"
+ : "attention";
+ const jobsState: StepState =
+ writableRows.length === 0
+ ? "busy"
+ : filledRows.length === writableRows.length
+ ? "done"
+ : "attention";
- // --- actions ----------------------------------------------------------
// A server that is not installed gets a button that NAMES the install:
// the click is the consent the dangerous-flagged install route asks for.
const needsInstall = Boolean(s && !s.installed && !remote);
@@ -223,146 +231,216 @@ export function OverviewPanel({
server: serverLabel,
})
: t("local_models.overview.action_setup");
- const setupHint = needsInstall
- ? t("local_models.overview.action_setup_install_hint")
- : t("local_models.overview.action_setup_hint");
-
- const roleRows = overview.data?.roles.roles ?? [];
- const roleLabel = useCallback(
- (role: LocalModelRole) => {
- const row = roleRows.find((r) => r.id === role);
- return t(row?.label_key ?? `local_models.role_${role}`);
- },
- [roleRows, t],
- );
return (
<div className="space-y-4" data-testid="local-models-overview-panel">
- <p
- className="flex flex-wrap items-center gap-x-2 gap-y-1 text-sm text-foreground"
- data-testid="overview-status"
- >
- <span className={cn("h-2 w-2 shrink-0 rounded-full", dotClass)} />
- <span>{serverClause}</span>
- {gpuClause && (
- <>
- <span className="text-muted-foreground">·</span>
- <span className="text-muted-foreground">{gpuClause}</span>
- </>
- )}
- {brainClause && (
- <>
- <span className="text-muted-foreground">·</span>
- <span
- className={cn(
- activeBrain?.id === providerId
- ? "text-muted-foreground"
- : "text-amber-700 dark:text-amber-400",
- )}
- data-testid="overview-brain-clause"
- >
- {brainClause}
- </span>
- </>
- )}
- {checking && (
- <span className="ml-1" data-testid="overview-checking">
- <StatusDot
- tone="busy"
- pulse
- label={t("local_models.overview.checking")}
- />
+ {/* ── 1 · Server ────────────────────────────────────────────────── */}
+ <StepCard
+ step={1}
+ title={t("local_models.overview.step_server_title")}
+ subtitle={t("local_models.overview.step_server_subtitle")}
+ state={serverState}
+ testId="step-server"
+ status={
+ <span className="flex flex-wrap items-center gap-x-3 gap-y-1">
+ <StatusDot tone={serverTone} pulse={serverTone === "busy"} label={serverClause} />
+ {checking && (
+ <span data-testid="overview-checking">
+ <StatusDot tone="busy" pulse label={t("local_models.overview.checking")} />
+ </span>
+ )}
</span>
- )}
- </p>
-
- <div
- className="flex flex-wrap gap-2"
- role="group"
- aria-label={t("local_models.overview.actions_label")}
- data-testid="overview-actions"
- >
- <ActionButton
- primary
- icon={
- setup.busy ? (
+ }
+ action={
+ <SoftButton
+ primary
+ onClick={setup.run}
+ disabled={setup.busy}
+ className="h-9"
+ >
+ {setup.busy ? (
<Loader2 className="h-4 w-4 animate-spin" />
) : (
<Sparkles className="h-4 w-4" />
- )
- }
- label={setupLabel}
- hint={setupHint}
- onClick={setup.run}
- disabled={setup.busy}
- testId="overview-setup"
- />
- {onBrowse && (
- <ActionButton
- icon={<Search className="h-4 w-4" />}
- label={t("local_models.overview.action_browse")}
- hint={t("local_models.overview.action_browse_hint")}
- onClick={onBrowse}
- />
- )}
- </div>
-
- <SetupProgress
- step={setup.step}
- serverLabel={serverLabel}
- roleLabel={roleLabel}
- otherBrain={otherBrainActive ? (activeBrain?.label ?? "") : ""}
- onSwitchBrain={async () => {
- await switchBrainProvider(providerId);
- await refetchProviders();
- }}
- t={t}
- />
-
- <div className="grid grid-cols-1 gap-3 sm:grid-cols-3">
- <StatTile
- icon={<HardDrive className="h-3.5 w-3.5" />}
- label={t("local_models.overview.disk")}
- value={formatGigabytes(s?.disk_bytes ?? 0)}
- hint={diskHint}
- tone="primary"
- loading={server.isLoading}
- />
- <StatTile
- icon={<Cpu className="h-3.5 w-3.5" />}
- label={t("local_models.overview.gpu")}
- value={gpuValue}
- hint={gpuHint}
- tone="primary"
- loading={server.isLoading}
+ )}
+ {setupLabel}
+ </SoftButton>
+ }
+ >
+ <p className="text-xs text-muted-foreground" data-testid="setup-hint">
+ {needsInstall
+ ? t("local_models.overview.action_setup_install_hint")
+ : t("local_models.overview.action_setup_hint")}
+ </p>
+ <SetupProgress
+ step={setup.step}
+ serverLabel={serverLabel}
+ roleLabel={roleLabel}
+ otherBrain={otherBrainActive ? (activeBrain?.label ?? "") : ""}
+ onSwitchBrain={async () => {
+ await switchBrainProvider(providerId);
+ await refetchProviders();
+ }}
+ t={t}
/>
- <StatTile
- icon={<Layers className="h-3.5 w-3.5" />}
- label={t("local_models.overview.loaded")}
- value={String(loaded.length)}
- hint={loadedHint}
- tone="primary"
- loading={server.isLoading}
+ <FactRows
+ className="mt-3 !text-sm"
+ rows={[
+ {
+ label: t("local_models.overview.fact_version"),
+ value: s?.version || t("local_models.overview.server_unknown"),
+ },
+ {
+ label: t("local_models.overview.fact_address"),
+ value: s?.base_url ? (
+ <span className="font-mono text-xs">{s.base_url}</span>
+ ) : (
+ ""
+ ),
+ },
+ {
+ label: t("local_models.overview.gpu"),
+ value: (
+ <span>
+ {gpuValue}
+ <span className="ml-2 text-xs text-muted-foreground">{gpuHint}</span>
+ </span>
+ ),
+ },
+ {
+ label: t("local_models.overview.fact_brain"),
+ value: activeBrain ? (
+ <span
+ className={cn(
+ otherBrainActive && "text-amber-700 dark:text-amber-400",
+ )}
+ data-testid="overview-brain-clause"
+ >
+ {otherBrainActive
+ ? fill(t("local_models.overview.status_brain_other"), {
+ server: serverLabel,
+ brain: activeBrain.label,
+ })
+ : fill(t("local_models.overview.status_brain_active"), {
+ server: serverLabel,
+ })}
+ </span>
+ ) : (
+ ""
+ ),
+ },
+ ]}
/>
- </div>
+ </StepCard>
- <RolesPanel
- providerId={providerId}
- onTune={onTune}
- onOpenApiKeys={onOpenApiKeys}
- variant={advanced ? "ledger" : "checklist"}
- onSetup={setup.run}
- setupBusy={setup.busy}
- />
+ {/* ── 2 · Models ───────────────────────────────────────────────── */}
+ <StepCard
+ step={2}
+ title={t("local_models.overview.step_models_title")}
+ subtitle={t("local_models.overview.step_models_subtitle")}
+ state={modelsState}
+ testId="step-models"
+ status={
+ <StatusDot
+ tone={modelsState === "done" ? "ok" : modelsState === "busy" ? "busy" : "warn"}
+ pulse={modelsState === "busy"}
+ label={
+ serverSilent
+ ? t("local_models.overview.models_unknown")
+ : modelCount === 1
+ ? fill(t("local_models.overview.models_one"), {
+ size: formatGigabytes(inventory.data?.disk_bytes ?? 0),
+ })
+ : fill(t("local_models.overview.models_count"), {
+ count: modelCount,
+ size: formatGigabytes(inventory.data?.disk_bytes ?? 0),
+ })
+ }
+ />
+ }
+ action={
+ <>
+ {onBrowse && (
+ <SoftButton onClick={onBrowse} className="h-9">
+ <Search className="h-4 w-4" />
+ {t("local_models.overview.action_browse")}
+ </SoftButton>
+ )}
+ {onManage && modelCount > 0 && (
+ <SoftButton onClick={onManage} className="h-9">
+ {t("local_models.installed.manage")}
+ </SoftButton>
+ )}
+ </>
+ }
+ >
+ <div className="grid grid-cols-1 gap-3 sm:grid-cols-3">
+ <StatTile
+ icon={<HardDrive className="h-3.5 w-3.5" />}
+ label={t("local_models.overview.disk")}
+ value={formatGigabytes(s?.disk_bytes ?? 0)}
+ hint={s?.models_dir ?? ""}
+ tone="primary"
+ loading={server.isLoading}
+ />
+ <StatTile
+ icon={<Cpu className="h-3.5 w-3.5" />}
+ label={t("local_models.overview.gpu")}
+ value={gpuValue}
+ hint={gpuHint}
+ tone="primary"
+ loading={server.isLoading}
+ />
+ <StatTile
+ icon={<Layers className="h-3.5 w-3.5" />}
+ label={t("local_models.overview.loaded")}
+ value={String(loaded.length)}
+ hint={loadedHint}
+ tone="primary"
+ loading={server.isLoading}
+ />
+ </div>
+ <div className="mt-3">
+ <InstalledPanel
+ bare
+ models={models}
+ roles={roleRows}
+ diskBytes={inventory.data?.disk_bytes ?? s?.disk_bytes ?? 0}
+ loading={overview.isLoading}
+ error={inventory.data?.error ?? null}
+ onManage={onManage}
+ />
+ </div>
+ </StepCard>
- <InstalledPanel
- models={inventory.data?.models ?? []}
- roles={roleRows}
- diskBytes={inventory.data?.disk_bytes ?? s?.disk_bytes ?? 0}
- loading={overview.isLoading}
- error={inventory.data?.error ?? null}
- onManage={onManage}
- />
+ {/* ── 3 · Jobs ─────────────────────────────────────────────────── */}
+ <StepCard
+ step={3}
+ title={t("local_models.overview.step_jobs_title")}
+ subtitle={t("local_models.overview.step_jobs_subtitle")}
+ state={jobsState}
+ testId="step-jobs"
+ status={
+ writableRows.length > 0 ? (
+ <StatusDot
+ tone={jobsState === "done" ? "ok" : "warn"}
+ label={fill(t("local_models.overview.jobs_filled"), {
+ filled: filledRows.length,
+ total: writableRows.length,
+ })}
+ />
+ ) : null
+ }
+ >
+ <RolesPanel
+ providerId={providerId}
+ onTune={onTune}
+ onOpenApiKeys={onOpenApiKeys}
+ advanced={advanced}
+ onSetup={setup.run}
+ setupBusy={setup.busy}
+ />
+ </StepCard>
</div>
);
}
@@ -498,7 +576,7 @@ function SetupProgress({
}
return (
- <div className="space-y-1 text-xs" data-testid="setup-progress">
+ <div className="mt-2 space-y-1 text-xs" data-testid="setup-progress">
<StatusDot tone={tone} pulse={tone === "busy"} label={text} />
{lines.map((line, i) => (
<p key={i} className="ml-4 text-muted-foreground">
@@ -553,7 +631,7 @@ function SetupProgress({
}
/** A 44 px SoftButton with icon, label and a one-line hint underneath. */
-function ActionButton({
+export function ActionButton({
icon,
label,
hint,
diff --git a/jarvis/ui/web/frontend/src/views/local-models/RolePicker.test.tsx b/jarvis/ui/web/frontend/src/views/local-models/RolePicker.test.tsx
new file mode 100644
index 0000000000..4d90d65d95
--- /dev/null
+++ b/jarvis/ui/web/frontend/src/views/local-models/RolePicker.test.tsx
@@ -0,0 +1,138 @@
+/**
+ * The role picker's job is to always have something to pick.
+ *
+ * `qualifying` alone leaves a user stranded twice over — an unreachable
+ * server empties it, and an unprobed download reports no capabilities — and
+ * both read as "my pick jumped back" (BUG-188). These tests pin the two
+ * groups and the fact that neither list can swallow an installed model.
+ */
+import { fireEvent, render, screen } from "@testing-library/react";
+import { describe, expect, it, vi } from "vitest";
+
+import type { LocalModelRow, RoleRow } from "@/hooks/useLocalModels";
+
+import { RolePicker, splitChoices } from "./RolePicker";
+
+vi.mock("@/i18n", () => ({
+ useT: () => (key: string) => key,
+ fill: (template: string, values: Record<string, string>) =>
+ `${template}${Object.values(values).join("|")}`,
+}));
+
+function role(extra: Partial<RoleRow> = {}): RoleRow {
+ return {
+ id: "chat",
+ label_key: "local_models.role_chat",
+ config_key: "brain.providers.ollama.model",
+ current: "",
+ installed: false,
+ required: [],
+ recommended_capabilities: [],
+ qualifying: [],
+ recommended: "",
+ writable: true,
+ advanced: false,
+ note: "",
+ ...extra,
+ } as RoleRow;
+}
+
+const model = (name: string): LocalModelRow =>
+ ({ name, capabilities: [], probed: true, used_by: [] }) as unknown as LocalModelRow;
+
+describe("splitChoices", () => {
+ it("puts the qualifying models first and every other download after them", () => {
+ const { fits, others } = splitChoices(
+ role({ qualifying: ["gemma4:12b", "qwen3.5:4b"] }),
+ [model("gemma4:12b"), model("qwen3.5:4b"), model("deepseek-r1:14b")],
+ );
+ expect(fits).toEqual(["gemma4:12b", "qwen3.5:4b"]);
+ expect(others).toEqual(["deepseek-r1:14b"]);
+ });
+
+ it("still offers the installed models when nothing qualifies", () => {
+ // The state the section used to get stuck in: no shortlist, so the only
+ // option was the tag already configured.
+ const { fits, others } = splitChoices(
+ role({ current: "ornith:9b", qualifying: [] }),
+ [model("gemma4:12b"), model("qwen3.5:4b"), model("ornith:9b")],
+ );
+ expect(fits).toEqual([]);
+ expect(others).toEqual(["gemma4:12b", "qwen3.5:4b", "ornith:9b"]);
+ });
+
+ it("keeps a configured tag that is gone from disk, exactly once", () => {
+ const { fits, others } = splitChoices(
+ role({ current: "removed:7b", qualifying: ["qwen3.5:4b"] }),
+ [model("qwen3.5:4b")],
+ );
+ expect(fits).toEqual(["qwen3.5:4b"]);
+ expect(others).toEqual(["removed:7b"]);
+ });
+
+ it("treats a bare tag and its :latest as one download", () => {
+ const { fits, others } = splitChoices(
+ role({ current: "bge-m3", qualifying: ["bge-m3:latest"] }),
+ [model("bge-m3:latest")],
+ );
+ expect(fits).toEqual(["bge-m3:latest"]);
+ expect(others).toEqual([]);
+ });
+});
+
+describe("RolePicker", () => {
+ it("groups the options and reports the pick", () => {
+ const onPick = vi.fn();
+ render(
+ <RolePicker
+ row={role({ current: "gemma4:12b", qualifying: ["gemma4:12b"] })}
+ models={[model("gemma4:12b"), model("deepseek-r1:14b")]}
+ onPick={onPick}
+ />,
+ );
+ const select = screen.getByTestId("role-picker-chat") as HTMLSelectElement;
+ expect([...select.options].map((o) => o.value)).toEqual([
+ "",
+ "gemma4:12b",
+ "deepseek-r1:14b",
+ ]);
+ expect(
+ [...select.querySelectorAll("optgroup")].map((g) => g.label),
+ ).toEqual([
+ "local_models.roles.pick_group_fits",
+ "local_models.roles.pick_group_others",
+ ]);
+
+ fireEvent.change(select, { target: { value: "deepseek-r1:14b" } });
+ expect(onPick).toHaveBeenCalledWith("deepseek-r1:14b");
+ });
+
+ it("marks a configured model that is no longer on the server", () => {
+ render(
+ <RolePicker
+ row={role({ current: "removed:7b" })}
+ models={[model("qwen3.5:4b")]}
+ onPick={vi.fn()}
+ />,
+ );
+ const select = screen.getByTestId("role-picker-chat") as HTMLSelectElement;
+ const gone = [...select.options].find((o) => o.value === "removed:7b");
+ expect(gone?.textContent).toContain(
+ "local_models.roles.pick_missing_suffix",
+ );
+ });
+
+ it("offers no discovery option for the embedding role once one is set", () => {
+ render(
+ <RolePicker
+ row={role({ id: "embedding", current: "bge-m3:latest" })}
+ models={[model("bge-m3:latest")]}
+ onPick={vi.fn()}
+ />,
+ );
+ const select = screen.getByTestId(
+ "role-picker-embedding",
+ ) as HTMLSelectElement;
+ expect([...select.options].map((o) => o.value)).toEqual(["bge-m3:latest"]);
+ });
+});
diff --git a/jarvis/ui/web/frontend/src/views/local-models/RolePicker.tsx b/jarvis/ui/web/frontend/src/views/local-models/RolePicker.tsx
new file mode 100644
index 0000000000..761b3e0d1b
--- /dev/null
+++ b/jarvis/ui/web/frontend/src/views/local-models/RolePicker.tsx
@@ -0,0 +1,121 @@
+/**
+ * The model picker of one role row.
+ *
+ * It used to list `row.qualifying` only — the installed downloads that
+ * declare every capability the job requires — plus the configured tag when
+ * that had gone missing. Two ways that leaves a user stuck with a value they
+ * did not choose (BUG-188):
+ *
+ * * the sweep behind the payload found no server, so `qualifying` is empty
+ * and the list holds exactly one entry: the tag already configured;
+ * * a download Jarvis could not probe (`probed: false`) reports no
+ * capabilities, so it silently never appears — even though it runs.
+ *
+ * Both read as "I picked something else and it jumped back". The picker
+ * therefore offers EVERY installed download, in two groups: the ones that
+ * fit the job first, the rest under "Other installed models" with the
+ * consequence spelled out rather than hidden. Choosing one is a real choice
+ * a user is allowed to make; the section says what it costs, it does not
+ * decide for them.
+ */
+import { useMemo } from "react";
+
+import { canonicalModelName, type LocalModelRow, type RoleRow } from "@/hooks/useLocalModels";
+import { fill, useT } from "@/i18n";
+import { cn } from "@/lib/utils";
+
+export interface RolePickerProps {
+ row: RoleRow;
+ /** Every download the inventory knows; empty while the server is silent. */
+ models: LocalModelRow[];
+ /** True while a write for this row is in flight. */
+ disabled?: boolean;
+ onPick: (model: string) => void;
+ className?: string;
+}
+
+/** `qualifying`, then every other installed tag, each list de-duplicated. */
+export function splitChoices(
+ row: RoleRow,
+ models: LocalModelRow[],
+): { fits: string[]; others: string[] } {
+ const seen = new Set<string>();
+ const fits: string[] = [];
+ const others: string[] = [];
+ const take = (name: string, into: string[]) => {
+ const key = canonicalModelName(name);
+ if (!name || seen.has(key)) return;
+ seen.add(key);
+ into.push(name);
+ };
+ for (const name of row.qualifying) take(name, fits);
+ for (const m of models) take(m.name, others);
+ // A configured tag that is gone from disk still belongs in the list, or
+ // the row would silently show a value the picker cannot express.
+ if (row.current) take(row.current, others);
+ return { fits, others };
+}
+
+export function RolePicker({
+ row,
+ models,
+ disabled,
+ onPick,
+ className,
+}: RolePickerProps) {
+ const t = useT();
+ const { fits, others } = useMemo(() => splitChoices(row, models), [row, models]);
+ const installed = useMemo(() => {
+ const names = new Set<string>();
+ for (const m of models) names.add(canonicalModelName(m.name));
+ return names;
+ }, [models]);
+
+ const label = (name: string) =>
+ installed.has(canonicalModelName(name))
+ ? name
+ : `${name} ${t("local_models.roles.pick_missing_suffix")}`;
+
+ return (
+ <select
+ aria-label={fill(t("local_models.roles.pick_label"), {
+ role: t(row.label_key),
+ })}
+ value={row.current}
+ disabled={disabled}
+ onChange={(e) => onPick(e.target.value)}
+ data-testid={`role-picker-${row.id}`}
+ className={cn(
+ "h-9 w-full rounded-lg border border-border bg-background/60 px-2.5 text-sm text-foreground",
+ "transition-colors hover:border-border/80 focus:outline-none focus:ring-1 focus:ring-primary",
+ "disabled:cursor-not-allowed disabled:opacity-50",
+ className,
+ )}
+ >
+ {row.id !== "embedding" && (
+ <option value="">{t("local_models.roles.pick_discovery")}</option>
+ )}
+ {row.id === "embedding" && !row.current && (
+ <option value="">{t("local_models.roles.pick_none")}</option>
+ )}
+ {fits.length > 0 && (
+ <optgroup label={t("local_models.roles.pick_group_fits")}>
+ {fits.map((name) => (
+ <option key={name} value={name}>
+ {label(name)}
+ </option>
+ ))}
+ </optgroup>
+ )}
+ {others.length > 0 && (
+ <optgroup label={t("local_models.roles.pick_group_others")}>
+ {others.map((name) => (
+ <option key={name} value={name}>
+ {label(name)}
+ </option>
+ ))}
+ </optgroup>
+ )}
+ </select>
+ );
+}
diff --git a/jarvis/ui/web/frontend/src/views/local-models/RolesPanel.tsx b/jarvis/ui/web/frontend/src/views/local-models/RolesPanel.tsx
index 668f2fd62a..1431db8a79 100644
--- a/jarvis/ui/web/frontend/src/views/local-models/RolesPanel.tsx
+++ b/jarvis/ui/web/frontend/src/views/local-models/RolesPanel.tsx
@@ -1,48 +1,37 @@
/**
- * Roles — the four model slots Jarvis fills from the local server, one row
- * each, one primary action each.
+ * Jobs — which download answers for which job, one row each.
*
* Chat & voice, Tools & screen, Deep & coding and Embeddings come from
* `GET roles`; every row shows the configured pick (or the discovery
- * sentence), the capabilities the slot requires, a picker over the installed
- * models that qualify, and "Use recommended" — which downloads the pick when
- * it is missing, assigns it, then writes the suggested options silently and
- * reads back one sentence. The read-only advanced roles hide under
- * "More roles". When nothing is installed at all, the ledger shows one
- * "Set up the recommended models" button that runs the overview's set-up
- * flow (`useLocalSetup`).
+ * sentence), a picker over the installed models (`RolePicker`) and, when the
+ * shortlist has a better one, "Use recommended" — which downloads that pick
+ * when it is missing, assigns it, then writes the suggested options silently
+ * and reads back one sentence. The read-only advanced roles hide under
+ * "More roles".
*
- * Two variants: `ledger` (the full rows: picker, badges, Tune) and
- * `checklist` — one compact line per writable role (dot, label, model or
- * "not set", ONE trailing action: "Use recommended" when the shortlist has a
- * better pick, otherwise "Change"), where Change expands that line into the
- * full ledger row in place. The Overview uses the checklist.
+ * The picker is on screen for every row at every detail level. It used to be
+ * folded away behind a "Change" button on the Simple page, which — together
+ * with a shortlist that could come back empty — is what made the section
+ * read as "it keeps resetting my model" (BUG-188): the one control that
+ * changes a role was two clicks away and then offered nothing to pick.
+ * `advanced` now adds detail (capability badges, Tune, the read-only roles),
+ * it never takes the control away.
*
+ * The panel draws no frame of its own — the caller's `StepCard` is the frame.
* Props take `providerId` (the pull-capable card's id) rather than the whole
- * descriptor so the panel can be mounted from anywhere the id is known. The
- * "Tune" button only calls `onTune(model)` — the sheet belongs to the caller.
+ * descriptor so the panel can be mounted from anywhere the id is known.
*/
import { useCallback, useEffect, useMemo, useRef, useState } from "react";
-import {
- ChevronDown,
- ChevronRight,
- Loader2,
- SlidersHorizontal,
- Sparkles,
-} from "lucide-react";
+import { ChevronDown, Loader2, SlidersHorizontal, Sparkles } from "lucide-react";
-import {
- Panel,
- PanelHeader,
- SoftButton,
- StatusDot,
-} from "@/components/extensions/primitives";
+import { SoftButton, StatusDot } from "@/components/extensions/primitives";
import {
setRole,
useInvalidateLocalModels,
useOverview,
useSetRole,
type LocalModelRole,
+ type LocalModelRow,
type RoleRow,
} from "@/hooks/useLocalModels";
import { useProviders } from "@/hooks/useProviders";
@@ -50,6 +39,7 @@ import { fill, useT } from "@/i18n";
import { cn } from "@/lib/utils";
import { canonical, waitForPull } from "./localSetup";
+import { RolePicker } from "./RolePicker";
import { useSilentTune } from "./useLocalSetup";
export interface RolesPanelProps {
@@ -59,8 +49,8 @@ export interface RolesPanelProps {
onTune?: (model: string) => void;
/** Navigates to the API Keys section; shown when another brain is active. */
onOpenApiKeys?: () => void;
- /** `ledger` = full rows (default); `checklist` = compact rows that expand. */
- variant?: "ledger" | "checklist";
+ /** Adds capability badges, Tune and the read-only roles. Never removes the picker. */
+ advanced?: boolean;
/** Runs the overview's "Set up everything" flow (the empty state's button). */
onSetup?: () => void;
/** That flow is underway: the empty state's button waits. */
@@ -94,7 +84,7 @@ export function RolesPanel({
providerId,
onTune,
onOpenApiKeys,
- variant = "ledger",
+ advanced = false,
onSetup,
setupBusy = false,
}: RolesPanelProps) {
@@ -116,16 +106,6 @@ export function RolesPanel({
const [progress, setProgress] = useState<Record<string, Progress>>({});
const [moreOpen, setMoreOpen] = useState(false);
- // Checklist rows the user expanded into the full ledger row.
- const [expanded, setExpanded] = useState<Set<string>>(() => new Set());
- const toggleExpanded = useCallback((roleId: string) => {
- setExpanded((prev) => {
- const next = new Set(prev);
- if (next.has(roleId)) next.delete(roleId);
- else next.add(roleId);
- return next;
- });
- }, []);
// Async flows outlive a tab switch; nothing is written to state afterwards.
const mounted = useRef(true);
@@ -137,11 +117,15 @@ export function RolesPanel({
}, []);
const alive = useCallback(() => mounted.current, []);
+ const models: LocalModelRow[] = useMemo(
+ () => inventory.data?.models ?? [],
+ [inventory.data],
+ );
const installed = useMemo(() => {
const names = new Set<string>();
- for (const m of inventory.data?.models ?? []) names.add(canonical(m.name));
+ for (const m of models) names.add(canonical(m.name));
return names;
- }, [inventory.data]);
+ }, [models]);
const isInstalled = useCallback(
(name: string) => installed.has(canonical(name)),
[installed],
@@ -149,7 +133,7 @@ export function RolesPanel({
const rows = roles.data?.roles ?? [];
const primary = rows.filter((r) => !r.advanced);
- const advanced = rows.filter((r) => r.advanced);
+ const advancedRows = rows.filter((r) => r.advanced);
// The brain switch stays a user-only decision; the panel only says so.
const activeBrain =
@@ -197,206 +181,197 @@ export function RolesPanel({
[alive, invalidate, isInstalled, patch, providerId, tuneSilently],
);
+ // "Nothing is installed" is only true when the server actually answered.
+ // While it is silent the inventory is empty for a different reason, and
+ // saying "download something" there sends the user after the wrong problem.
+ const serverSilent = Boolean(inventory.data?.error);
const nothingInstalled =
- inventory.data !== undefined &&
- !inventory.data.error &&
- inventory.data.models.length === 0;
+ inventory.data !== undefined && !serverSilent && models.length === 0;
return (
- <Panel className="p-4">
- <div className="space-y-3" data-testid="local-models-roles" data-variant={variant}>
- <PanelHeader
- title={t("local_models.roles.title")}
- subtitle={t("local_models.roles.subtitle")}
- />
+ <div className="space-y-3" data-testid="local-models-roles">
+ {otherBrainActive && (
+ <p
+ className="text-sm text-amber-700 dark:text-amber-400"
+ data-testid="roles-other-brain"
+ >
+ {fill(t("local_models.roles.other_brain"), {
+ brain: activeBrain?.label ?? "",
+ })}{" "}
+ {onOpenApiKeys && (
+ <button
+ type="button"
+ onClick={onOpenApiKeys}
+ className="font-medium text-primary hover:underline"
+ >
+ {t("local_models.roles.other_brain_link")}
+ </button>
+ )}
+ </p>
+ )}
- {otherBrainActive && (
- <p
- className="text-sm text-amber-700 dark:text-amber-400"
- data-testid="roles-other-brain"
- >
- {fill(t("local_models.roles.other_brain"), {
- brain: activeBrain?.label ?? "",
- })}{" "}
- {onOpenApiKeys && (
- <button
- type="button"
- onClick={onOpenApiKeys}
- className="font-medium text-primary hover:underline"
- >
- {t("local_models.roles.other_brain_link")}
- </button>
- )}
- </p>
- )}
+ {serverSilent && (
+ <p
+ className="text-sm text-amber-700 dark:text-amber-400"
+ data-testid="roles-server-silent"
+ >
+ {t("local_models.roles.server_silent")}
+ </p>
+ )}
+ {roles.data?.error && (
+ <p className="text-sm text-amber-700 dark:text-amber-400">
+ {roles.data.error}
+ </p>
+ )}
+ {roles.isError && (
+ <p className="text-sm text-destructive">
+ {roles.error instanceof Error
+ ? roles.error.message
+ : String(roles.error)}
+ </p>
+ )}
+ {roles.isLoading && (
+ <p className="text-sm text-muted-foreground">
+ {t("local_models.roles.loading")}
+ </p>
+ )}
- {roles.data?.error && (
- <p className="text-sm text-amber-700 dark:text-amber-400">
- {roles.data.error}
- </p>
- )}
- {roles.isError && (
- <p className="text-sm text-destructive">
- {roles.error instanceof Error
- ? roles.error.message
- : String(roles.error)}
- </p>
- )}
- {roles.isLoading && (
+ {nothingInstalled && rows.length > 0 && onSetup && (
+ <div
+ className="rounded-lg border border-dashed border-border px-6 py-8 text-center"
+ data-testid="roles-empty"
+ >
<p className="text-sm text-muted-foreground">
- {t("local_models.roles.loading")}
+ {t("local_models.roles.empty_body")}
</p>
- )}
-
- {nothingInstalled && rows.length > 0 && onSetup && (
- <div
- className="rounded-lg border border-dashed border-border px-6 py-8 text-center"
- data-testid="roles-empty"
- >
- <p className="text-sm text-muted-foreground">
- {t("local_models.roles.empty_body")}
- </p>
- <div className="mt-4 flex justify-center">
- <SoftButton primary onClick={onSetup} disabled={setupBusy}>
- {setupBusy ? (
- <Loader2 className="h-3.5 w-3.5 animate-spin" />
- ) : (
- <Sparkles className="h-3.5 w-3.5" />
- )}
- {t("local_models.roles.setup_button")}
- </SoftButton>
- </div>
+ <div className="mt-4 flex justify-center">
+ <SoftButton primary onClick={onSetup} disabled={setupBusy}>
+ {setupBusy ? (
+ <Loader2 className="h-3.5 w-3.5 animate-spin" />
+ ) : (
+ <Sparkles className="h-3.5 w-3.5" />
+ )}
+ {t("local_models.roles.setup_button")}
+ </SoftButton>
</div>
- )}
+ </div>
+ )}
- {primary.length > 0 && (
- <div
- className="divide-y divide-border/70 overflow-hidden rounded-xl border border-border bg-card/60"
- data-testid="roles-ledger"
- >
- {primary.map((row) => {
- const rowProgress = progress[row.id] ?? { phase: "idle" };
- const saving =
+ {primary.length > 0 && (
+ <div
+ className="divide-y divide-border/70 overflow-hidden rounded-lg border border-border/70"
+ data-testid="roles-ledger"
+ >
+ {primary.map((row) => (
+ <RoleRowView
+ key={row.id}
+ row={row}
+ models={models}
+ progress={progress[row.id] ?? { phase: "idle" }}
+ installed={isInstalled}
+ saving={
setRoleMutation.isPending &&
- setRoleMutation.variables?.role === row.id;
- const onPick = (model: string) =>
+ setRoleMutation.variables?.role === row.id
+ }
+ advanced={advanced}
+ onPick={(model) =>
setRoleMutation.mutate({
role: row.id as LocalModelRole,
model,
- });
- if (variant === "checklist" && !expanded.has(row.id)) {
- return (
- <RoleChecklistRow
- key={row.id}
- row={row}
- progress={rowProgress}
- installed={isInstalled}
- saving={saving}
- onUseRecommended={() => void useRecommended(row)}
- onExpand={() => toggleExpanded(row.id)}
- t={t}
- />
- );
+ })
}
- return (
- <RoleLedgerRow
- key={row.id}
- row={row}
- progress={rowProgress}
- installed={isInstalled}
- saving={saving}
- onPick={onPick}
- onUseRecommended={() => void useRecommended(row)}
- onTune={onTune}
- onCollapse={
- variant === "checklist"
- ? () => toggleExpanded(row.id)
- : undefined
- }
- t={t}
- />
- );
- })}
- </div>
- )}
+ onUseRecommended={() => void useRecommended(row)}
+ onTune={onTune}
+ t={t}
+ />
+ ))}
+ </div>
+ )}
- {setRoleMutation.isError && (
- <p className="text-sm text-destructive">
- {setRoleMutation.error instanceof Error
- ? setRoleMutation.error.message
- : String(setRoleMutation.error)}
- </p>
- )}
+ {setRoleMutation.isError && (
+ <p className="text-sm text-destructive">
+ {setRoleMutation.error instanceof Error
+ ? setRoleMutation.error.message
+ : String(setRoleMutation.error)}
+ </p>
+ )}
- {advanced.length > 0 && (
- <div>
- <button
- type="button"
- onClick={() => setMoreOpen((v) => !v)}
- aria-expanded={moreOpen}
- className="inline-flex items-center gap-1.5 text-xs font-medium text-muted-foreground transition-colors hover:text-foreground"
+ {advanced && advancedRows.length > 0 && (
+ <div>
+ <button
+ type="button"
+ onClick={() => setMoreOpen((v) => !v)}
+ aria-expanded={moreOpen}
+ className="inline-flex items-center gap-1.5 text-xs font-medium text-muted-foreground transition-colors hover:text-foreground"
+ >
+ <ChevronDown
+ className={cn(
+ "h-3.5 w-3.5 transition-transform",
+ moreOpen && "rotate-180",
+ )}
+ />
+ {t("local_models.roles.more_roles")}
+ </button>
+ {moreOpen && (
+ <div
+ className="mt-2 divide-y divide-border/70 overflow-hidden rounded-lg border border-border/70"
+ data-testid="roles-more"
>
- <ChevronDown
- className={cn(
- "h-3.5 w-3.5 transition-transform",
- moreOpen && "rotate-180",
- )}
- />
- {t("local_models.roles.more_roles")}
- </button>
- {moreOpen && (
- <div
- className="mt-2 divide-y divide-border/70 overflow-hidden rounded-xl border border-border bg-card/60"
- data-testid="roles-more"
- >
- {advanced.map((row) => (
- <div
- key={row.id}
- className="grid gap-1 px-3.5 py-3 sm:grid-cols-[minmax(0,1fr)_minmax(0,2fr)] sm:items-center"
- >
- <div className="text-sm text-foreground">
- {t(row.label_key)}
- </div>
- <div className="text-sm text-muted-foreground">
- {row.current ? (
- <span className="font-mono text-xs text-foreground/85">
- {row.current}
- </span>
- ) : (
- row.note || t("local_models.roles.follows_chat")
- )}
- </div>
+ {advancedRows.map((row) => (
+ <div
+ key={row.id}
+ className="grid gap-1 px-3.5 py-3 sm:grid-cols-[minmax(0,1fr)_minmax(0,2fr)] sm:items-center"
+ >
+ <div className="text-sm text-foreground">
+ {t(row.label_key)}
</div>
- ))}
- </div>
- )}
- </div>
- )}
- </div>
- </Panel>
+ <div className="text-sm text-muted-foreground">
+ {row.current ? (
+ <span className="font-mono text-xs text-foreground/85">
+ {row.current}
+ </span>
+ ) : (
+ row.note || t("local_models.roles.follows_chat")
+ )}
+ </div>
+ </div>
+ ))}
+ </div>
+ )}
+ </div>
+ )}
+ </div>
);
}
-function RoleLedgerRow({
+/**
+ * One job: what it is, what answers it now, and the picker that changes it.
+ *
+ * Three columns on a wide pane — name and state, picker, actions — so the
+ * pickers line up as a column a user can run down. Below `lg` they stack in
+ * the same order.
+ */
+function RoleRowView({
row,
+ models,
progress,
installed,
saving,
+ advanced,
onPick,
onUseRecommended,
onTune,
- onCollapse,
t,
}: {
row: RoleRow;
+ models: LocalModelRow[];
progress: Progress;
installed: (name: string) => boolean;
saving: boolean;
+ advanced: boolean;
onPick: (model: string) => void;
onUseRecommended: () => void;
onTune?: (model: string) => void;
- /** Checklist only: folds the row back into its compact line. */
- onCollapse?: () => void;
t: (key: string) => string;
}) {
const running =
@@ -406,17 +381,10 @@ function RoleLedgerRow({
const currentMissing = row.current !== "" && !row.installed;
const canUseRecommended =
row.writable && row.recommended !== "" && row.recommended !== row.current;
- // The picker lists every qualifying download, plus the configured one when
- // it is gone, so the gap is visible rather than silently replaced.
- const choices = useMemo(() => {
- const list = [...row.qualifying];
- if (row.current && !list.includes(row.current)) list.unshift(row.current);
- return list;
- }, [row.current, row.qualifying]);
return (
<div
- className="grid gap-3 px-3.5 py-3 lg:grid-cols-[minmax(0,1.2fr)_minmax(0,1fr)_auto] lg:items-center"
+ className="grid gap-3 px-3.5 py-3 lg:grid-cols-[minmax(0,1fr)_minmax(0,1fr)_auto] lg:items-start"
data-testid={`role-row-${row.id}`}
>
<div className="min-w-0">
@@ -424,30 +392,16 @@ function RoleLedgerRow({
<span className="text-sm font-medium text-foreground">
{t(row.label_key)}
</span>
- {row.required.map((cap) => (
- <span
- key={cap}
- className="rounded-md border border-border px-1.5 py-0.5 text-[11px] uppercase tracking-[0.18em] text-muted-foreground"
- >
- {cap}
- </span>
- ))}
+ {advanced &&
+ row.required.map((cap) => (
+ <span
+ key={cap}
+ className="rounded-md border border-border px-1.5 py-0.5 text-[11px] uppercase tracking-[0.18em] text-muted-foreground"
+ >
+ {cap}
+ </span>
+ ))}
</div>
- {row.id === "voice" && row.context_tokens ? (
- <p
- className="mt-1 text-xs 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}
<div className="mt-1 text-sm">
{row.current ? (
<StatusDot
@@ -469,35 +423,32 @@ function RoleLedgerRow({
</span>
)}
</div>
+ {row.id === "voice" && row.context_tokens ? (
+ <p
+ className="mt-1 text-xs 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}
<ProgressLine progress={progress} t={t} className="mt-1.5" />
</div>
<div className="min-w-0">
{row.writable ? (
- <select
- aria-label={fill(t("local_models.roles.pick_label"), {
- role: t(row.label_key),
- })}
- value={row.current}
+ <RolePicker
+ row={row}
+ models={models}
disabled={saving || running}
- onChange={(e) => onPick(e.target.value)}
- className="h-8 w-full rounded-md border border-border bg-background/60 px-2 text-sm text-foreground focus:outline-none focus:ring-1 focus:ring-primary disabled:opacity-50"
- >
- {row.id !== "embedding" && (
- <option value="">{t("local_models.roles.pick_discovery")}</option>
- )}
- {row.id === "embedding" && !row.current && (
- <option value="">{t("local_models.roles.pick_none")}</option>
- )}
- {choices.map((name) => (
- <option key={name} value={name}>
- {name}
- {installed(name)
- ? ""
- : ` ${t("local_models.roles.pick_missing_suffix")}`}
- </option>
- ))}
- </select>
+ onPick={onPick}
+ />
) : (
<span className="text-xs text-muted-foreground">
{t("local_models.roles.read_only")}
@@ -505,7 +456,7 @@ function RoleLedgerRow({
)}
{canUseRecommended && row.recommended_reason && (
<p
- className="mt-1 text-xs text-muted-foreground"
+ className="mt-1.5 text-xs text-muted-foreground"
data-testid="role-recommended-reason"
>
<span className="font-mono text-foreground/85">
@@ -517,7 +468,7 @@ function RoleLedgerRow({
)}
</div>
- <div className="flex items-center gap-1.5 lg:justify-end">
+ <div className="flex flex-wrap items-center gap-1.5 lg:justify-end">
{canUseRecommended && (
<SoftButton
primary
@@ -536,119 +487,13 @@ function RoleLedgerRow({
})}
</SoftButton>
)}
- {onTune && row.current && row.installed && (
+ {advanced && onTune && row.current && row.installed && (
<SoftButton onClick={() => onTune(row.current)} disabled={running}>
<SlidersHorizontal className="h-3.5 w-3.5" />
{t("local_models.roles.tune")}
</SoftButton>
)}
- {onCollapse && (
- <SoftButton onClick={onCollapse} ariaExpanded>
- <ChevronDown className="h-3.5 w-3.5" />
- {t("local_models.roles.checklist_done")}
- </SoftButton>
- )}
- </div>
- </div>
- );
-}
-
-/**
- * The compact checklist line: dot + label + model (or "not set") + ONE
- * trailing action. The line itself is a button that expands to the full row;
- * the trailing action is "Use recommended" when the shortlist has a better
- * pick, otherwise "Change" (which expands too).
- */
-function RoleChecklistRow({
- row,
- progress,
- installed,
- saving,
- onUseRecommended,
- onExpand,
- t,
-}: {
- row: RoleRow;
- progress: Progress;
- installed: (name: string) => boolean;
- saving: boolean;
- onUseRecommended: () => void;
- onExpand: () => void;
- t: (key: string) => string;
-}) {
- const running =
- progress.phase === "pulling" ||
- progress.phase === "assigning" ||
- progress.phase === "tuning";
- const currentMissing = row.current !== "" && !row.installed;
- const canUseRecommended =
- row.writable && row.recommended !== "" && row.recommended !== row.current;
- const tone: "ok" | "warn" | "off" = row.current
- ? currentMissing
- ? "warn"
- : "ok"
- : "off";
- return (
- <div
- className="grid gap-2 px-3.5 py-2.5 sm:grid-cols-[minmax(0,1fr)_auto] sm:items-center"
- data-testid={`role-row-${row.id}`}
- data-variant="checklist"
- >
- <button
- type="button"
- onClick={onExpand}
- aria-expanded={false}
- className="flex min-w-0 items-center gap-2 text-left"
- >
- <ChevronRight className="h-3.5 w-3.5 shrink-0 text-muted-foreground" />
- <span className="text-sm font-medium text-foreground">
- {t(row.label_key)}
- </span>
- <StatusDot
- tone={tone}
- label={
- row.current ? (
- <span className="font-mono text-xs text-foreground/85">
- {row.current}
- {currentMissing && (
- <span className="ml-2 font-sans text-muted-foreground">
- {t("local_models.roles.not_installed")}
- </span>
- )}
- </span>
- ) : (
- <span className="text-xs">
- {t("local_models.roles.checklist_not_set")}
- </span>
- )
- }
- />
- </button>
- <div className="flex items-center gap-1.5 sm:justify-end">
- {canUseRecommended ? (
- <SoftButton
- primary
- onClick={onUseRecommended}
- disabled={running || saving}
- >
- {running ? (
- <Loader2 className="h-3.5 w-3.5 animate-spin" />
- ) : (
- <Sparkles className="h-3.5 w-3.5" />
- )}
- {installed(row.recommended)
- ? t("local_models.roles.use_recommended")
- : fill(t("local_models.roles.download_recommended"), {
- model: row.recommended,
- })}
- </SoftButton>
- ) : (
- <SoftButton onClick={onExpand} disabled={running || saving}>
- {t("local_models.roles.checklist_change")}
- </SoftButton>
- )}
</div>
- <ProgressLine progress={progress} t={t} className="sm:col-span-2" />
</div>
);
}
diff --git a/jarvis/ui/web/frontend/src/views/local-models/StepCard.tsx b/jarvis/ui/web/frontend/src/views/local-models/StepCard.tsx
new file mode 100644
index 0000000000..cd02c23eed
--- /dev/null
+++ b/jarvis/ui/web/frontend/src/views/local-models/StepCard.tsx
@@ -0,0 +1,87 @@
+/**
+ * A numbered step card — the shape the Local models overview is built from.
+ *
+ * The section answers one question in three moves: is there a server, are
+ * there models, and which model does which job. Written as a dashboard that
+ * reads as a wall of equal tiles, a newcomer has no idea where to start; so
+ * each move is one card that carries its own ordinal, its own state and at
+ * most one primary action. The ordinal turns into a check once the step is
+ * satisfied, which makes "what is still missing" a glance rather than a read.
+ *
+ * `state` drives the ordinal only. A card is never hidden or disabled by it:
+ * step 3 stays open while step 1 is unsatisfied, because someone reinstalling
+ * a server still wants to see which model each job is on.
+ */
+import type { ReactNode } from "react";
+import { Check } from "lucide-react";
+
+import { cn } from "@/lib/utils";
+
+export type StepState = "done" | "attention" | "todo" | "busy";
+
+export interface StepCardProps {
+ /** 1-based ordinal shown in the marker; replaced by a check when done. */
+ step: number;
+ title: string;
+ /** One plain sentence saying what this step is for. */
+ subtitle: string;
+ state: StepState;
+ /** The step's status line, right of the title on a wide pane. */
+ status?: ReactNode;
+ /** The one primary action, rendered in the header. */
+ action?: ReactNode;
+ children: ReactNode;
+ testId?: string;
+}
+
+const MARKER: Record<StepState, string> = {
+ // Emerald carries "finished cleanly" here exactly as it does in StatTile —
+ // the token set has no name for it, and it needs its own value per theme.
+ done: "border-emerald-500/40 bg-emerald-500/10 text-emerald-600 dark:text-emerald-400",
+ attention: "border-amber-500/40 bg-amber-500/10 text-amber-600 dark:text-amber-400",
+ busy: "border-primary/40 bg-primary/10 text-primary",
+ todo: "border-border bg-sheen/[0.06] text-muted-foreground",
+};
+
+export function StepCard({
+ step,
+ title,
+ subtitle,
+ state,
+ status,
+ action,
+ children,
+ testId,
+}: StepCardProps) {
+ return (
+ <section
+ className="overflow-hidden rounded-xl border border-border bg-card/50"
+ aria-label={title}
+ data-testid={testId}
+ data-step={step}
+ data-state={state}
+ >
+ <header className="flex flex-wrap items-start gap-x-4 gap-y-3 border-b border-border/70 px-4 py-3.5">
+ <span
+ aria-hidden
+ className={cn(
+ "mt-0.5 flex h-7 w-7 shrink-0 items-center justify-center rounded-full border",
+ "font-display text-[13px] font-semibold tabular-nums",
+ MARKER[state],
+ )}
+ >
+ {state === "done" ? <Check className="h-3.5 w-3.5" /> : step}
+ </span>
+ <div className="min-w-0 flex-1">
+ <h3 className="font-display text-[15px] font-semibold tracking-tight text-foreground">
+ {title}
+ </h3>
+ <p className="mt-0.5 text-xs text-muted-foreground">{subtitle}</p>
+ {status ? <div className="mt-1.5 text-sm">{status}</div> : null}
+ </div>
+ {action ? <div className="flex shrink-0 items-center gap-2">{action}</div> : null}
+ </header>
+ <div className="px-4 py-3.5">{children}</div>
+ </section>
+ );
+}
diff --git a/tests/unit/brain/test_ollama_overview.py b/tests/unit/brain/test_ollama_overview.py
index e9c280b234..97e6f19555 100644
--- a/tests/unit/brain/test_ollama_overview.py
+++ b/tests/unit/brain/test_ollama_overview.py
@@ -115,20 +115,51 @@ async def test_the_disk_snapshot_paints_at_once_and_one_refresh_runs(fake, tmp_p
assert time.perf_counter() - started < 0.5
task = ollama_overview._refresh_tasks.get(ROOT)
assert task is not None and not task.done()
- # A second open while the refresh runs joins it — no second task.
- again, source_again = await ollama_overview.get_overview(ROOT, _cfg())
- assert source_again == "cache"
- assert ollama_overview._refresh_tasks.get(ROOT) is task
refreshed = await task
assert refreshed["server"]["running"] is False
assert "did not answer" in refreshed["inventory"]["error"]
assert ROOT not in ollama_overview._refresh_tasks
- # The refreshed payload is now the memo, and it reached the disk.
+ # The refreshed payload is now the memo...
memo, memo_source = await ollama_overview.get_overview(ROOT, _cfg())
assert memo_source == "live" and memo is refreshed
+ # ...but it did NOT reach the disk: a sweep with no server behind it
+ # would paint a wiped machine on the next open (BUG-188), so the good
+ # snapshot from the first open stays.
saved = ollama_overview.load_snapshot()
- assert saved is not None and saved["payload"]["server"]["running"] is False
+ assert saved is not None and saved["payload"]["server"]["running"] is True
+ assert [m["name"] for m in saved["payload"]["inventory"]["models"]]
+
+
+async def test_a_second_open_during_a_refresh_joins_it(fake, monkeypatch) -> None:
+ """One refresh per root: a second open while one runs must not start a
+ second sweep of the same server."""
+ await ollama_overview.get_overview(ROOT, _cfg())
+ ollama_overview._reset_for_tests()
+ ollama_inventory._reset_for_tests()
+
+ # Hold the refresh open, so "while one runs" is a fact and not a race.
+ release = asyncio.Event()
+ real_build = ollama_overview.build_overview
+
+ async def _slow_build(*args: object, **kwargs: object) -> dict:
+ await release.wait()
+ return await real_build(*args, **kwargs) # type: ignore[arg-type]
+
+ monkeypatch.setattr(ollama_overview, "build_overview", _slow_build)
+
+ _payload, source = await ollama_overview.get_overview(ROOT, _cfg())
+ assert source == "cache"
+ task = ollama_overview._refresh_tasks.get(ROOT)
+ assert task is not None and not task.done()
+
+ _again, source_again = await ollama_overview.get_overview(ROOT, _cfg())
+ assert source_again == "cache"
+ assert ollama_overview._refresh_tasks.get(ROOT) is task
+
+ release.set()
+ await task
+ assert ROOT not in ollama_overview._refresh_tasks
async def test_a_day_old_snapshot_is_replaced_by_a_live_build(fake) -> None:
@@ -196,3 +227,57 @@ async def test_a_failing_refresh_is_logged_not_lost(fake, monkeypatch, caplog) -
await asyncio.sleep(0) # let the done-callback run
assert ROOT not in ollama_overview._refresh_tasks
assert any("background refresh" in r.getMessage() for r in caplog.records)
+
+
+async def test_an_offline_sweep_never_becomes_the_next_open_s_snapshot(fake) -> None:
+ """A sweep taken while the server is down must not replace a good one.
+
+ What would lie if it broke: the next open paints an empty inventory, so
+ every role reads "not installed", every shortlist reads "nothing fits
+ this job yet", and each role's picker offers only the tag already
+ configured — the machine looks wiped and nothing can be changed
+ (BUG-188).
+ """
+ good, source = await ollama_overview.get_overview(ROOT, _cfg())
+ assert source == "live"
+ assert good["inventory"]["models"]
+ saved_before = ollama_overview.load_snapshot()
+ assert saved_before is not None
+
+ fake.offline = True
+ ollama_overview._reset_for_tests()
+ ollama_inventory._reset_for_tests()
+ offline, source = await ollama_overview.get_overview(ROOT, _cfg(), fresh=True)
+ assert source == "live"
+ # The live answer is honest about the outage...
+ assert offline["inventory"]["error"]
+ assert offline["inventory"]["models"] == []
+ # ...and the snapshot on disk is still the one that holds real models.
+ saved_after = ollama_overview.load_snapshot()
+ assert saved_after == saved_before
+
+
+async def test_an_offline_snapshot_from_an_older_build_is_rebuilt_not_painted(fake) -> None:
+ """A poisoned snapshot already on disk is dropped on read.
+
+ What would lie if it broke: a file written before the guard existed keeps
+ painting a wiped machine for up to STALE_AFTER_S — a day.
+ """
+ await ollama_overview.get_overview(ROOT, _cfg())
+ poisoned = ollama_overview.load_snapshot()
+ assert poisoned is not None
+ poisoned["payload"]["inventory"] = {"models": [], "error": "Ollama is not answering."}
+ # Straight past the guard, the way an older build wrote it.
+ ollama_overview.snapshot_path().write_text(json.dumps(poisoned), encoding="utf-8")
+ ollama_overview._reset_for_tests()
+ ollama_inventory._reset_for_tests()
+
+ payload, source = await ollama_overview.get_overview(ROOT, _cfg())
+ assert source == "live"
+ assert [m["name"] for m in payload["inventory"]["models"]]
+
+
+def test_is_paintable_only_rejects_an_inventory_error() -> None:
+ assert ollama_overview.is_paintable({"inventory": {"models": [], "error": None}})
+ assert ollama_overview.is_paintable({"fetched_at": 1.0}) # not a schema check
+ assert not ollama_overview.is_paintable({"inventory": {"models": [], "error": "down"}})