feat(moa): add reference model toggles

agent/moa_loop.py

@@ -334,7 +334,7 @@ def _run_reference(
 
 
 def _run_references_parallel(
-    reference_models: list[dict[str, str]],
+    reference_models: list[dict[str, Any]],
     ref_messages: list[dict[str, Any]],
     *,
     temperature: float | None = None,
@@ -594,6 +594,7 @@ def aggregate_moa_context(
     provider default applies — matching single-model agent behavior. Presets
     may still pin explicit values.
     """
+    reference_models = [slot for slot in reference_models if slot.get("enabled", True)]
     reference_outputs: list[tuple[str, str, Any]] = []
     ref_messages = _reference_messages(api_messages)
     reference_outputs = _run_references_parallel(
@@ -803,7 +804,10 @@ def create(self, **api_kwargs: Any) -> Any:
 
         preset = resolve_moa_preset(load_config().get("moa") or {}, self.preset_name)
         messages = list(api_kwargs.get("messages") or [])
-        reference_models = preset.get("reference_models") or []
+        reference_models = [
+            slot for slot in (preset.get("reference_models") or [])
+            if slot.get("enabled", True)
+        ]
         aggregator = preset.get("aggregator") or {}
         # Expose the resolved aggregator slot so session cost accounting can
         # price the aggregator's acting turn at its REAL model/provider. The

apps/desktop/src/app/settings/model-settings.test.tsx

@@ -1,3 +1,4 @@
+import { QueryClient, QueryClientProvider } from '@tanstack/react-query'
 import { cleanup, fireEvent, render, screen, waitFor } from '@testing-library/react'
 import { afterEach, beforeAll, beforeEach, describe, expect, it, vi } from 'vitest'
 
@@ -16,6 +17,9 @@ const getAuxiliaryModels = vi.fn()
 const setModelAssignment = vi.fn()
 const getRecommendedDefaultModel = vi.fn()
 const setEnvVar = vi.fn()
+const getMoaModels = vi.fn()
+const saveMoaModels = vi.fn()
+const setApiRequestProfile = vi.fn()
 const getHermesConfigRecord = vi.fn()
 const saveHermesConfig = vi.fn()
 const startManualProviderOAuth = vi.fn()
@@ -27,6 +31,9 @@ vi.mock('@/hermes', () => ({
   setModelAssignment: (body: unknown) => setModelAssignment(body),
   getRecommendedDefaultModel: (slug: string) => getRecommendedDefaultModel(slug),
   setEnvVar: (key: string, value: string) => setEnvVar(key, value),
+  getMoaModels: () => getMoaModels(),
+  saveMoaModels: (body: unknown) => saveMoaModels(body),
+  setApiRequestProfile: (profile: string | null) => setApiRequestProfile(profile),
   getHermesConfigRecord: () => getHermesConfigRecord(),
   saveHermesConfig: (config: unknown) => saveHermesConfig(config)
 }))
@@ -46,6 +53,12 @@ beforeEach(() => {
         authenticated: true,
         capabilities: { 'hermes-4': { reasoning: true, fast: true } }
       },
+      {
+        name: 'OpenRouter',
+        slug: 'openrouter',
+        models: ['deepseek/deepseek-v4-pro', 'anthropic/claude-opus-4.8'],
+        authenticated: true
+      },
       // An unconfigured api_key provider — surfaced by the full-universe payload.
       {
         name: 'DeepSeek',
@@ -64,6 +77,33 @@ beforeEach(() => {
   setModelAssignment.mockResolvedValue({ provider: 'nous', model: 'hermes-4', gateway_tools: [] })
   getRecommendedDefaultModel.mockResolvedValue({ provider: 'deepseek', model: 'deepseek-chat', free_tier: null })
   setEnvVar.mockResolvedValue({ ok: true })
+  getMoaModels.mockResolvedValue({
+    default_preset: 'review',
+    active_preset: '',
+    presets: {
+      review: {
+        reference_models: [
+          { provider: 'nous', model: 'hermes-4', enabled: true },
+          { provider: 'openrouter', model: 'deepseek/deepseek-v4-pro', enabled: true }
+        ],
+        aggregator: { provider: 'openrouter', model: 'anthropic/claude-opus-4.8' },
+        reference_temperature: null,
+        aggregator_temperature: null,
+        max_tokens: 4096,
+        enabled: true
+      }
+    },
+    reference_models: [
+      { provider: 'nous', model: 'hermes-4', enabled: true },
+      { provider: 'openrouter', model: 'deepseek/deepseek-v4-pro', enabled: true }
+    ],
+    aggregator: { provider: 'openrouter', model: 'anthropic/claude-opus-4.8' },
+    reference_temperature: null,
+    aggregator_temperature: null,
+    max_tokens: 4096,
+    enabled: true
+  })
+  saveMoaModels.mockImplementation(async body => ({ ...(body as object), ok: true }))
   getHermesConfigRecord.mockResolvedValue({ agent: { reasoning_effort: 'medium', service_tier: 'normal' } })
   saveHermesConfig.mockResolvedValue({ ok: true })
 })
@@ -75,8 +115,13 @@ afterEach(() => {
 
 async function renderModelSettings() {
   const { ModelSettings } = await import('./model-settings')
+  const client = new QueryClient({ defaultOptions: { queries: { retry: false } } })
 
-  return render(<ModelSettings />)
+  return render(
+    <QueryClientProvider client={client}>
+      <ModelSettings />
+    </QueryClientProvider>
+  )
 }
 
 describe('ModelSettings', () => {
@@ -91,11 +136,10 @@ describe('ModelSettings', () => {
     const triggers = await screen.findAllByRole('combobox')
     fireEvent.click(triggers[0])
 
-    // "Nous" shows in both the trigger and the open list; the unconfigured
-    // provider + its setup hint are the unique signal of the full universe.
+    // "Nous" shows in both the trigger and the open list; DeepSeek is the
+    // unique signal that the unconfigured provider universe is present.
     expect((await screen.findAllByText('Nous')).length).toBeGreaterThan(0)
     expect(await screen.findByText(/DeepSeek/)).toBeTruthy()
-    expect(await screen.findByText(/set up/)).toBeTruthy()
   })
 
   it('activates an unconfigured api_key provider inline by saving its key', async () => {
@@ -123,7 +167,7 @@ describe('ModelSettings', () => {
     await renderModelSettings()
     await waitFor(() => expect(getHermesConfigRecord).toHaveBeenCalled())
 
-    const fastSwitch = await screen.findByRole('switch')
+    const fastSwitch = await screen.findByText('Fast')
     fireEvent.click(fastSwitch)
 
     await waitFor(() =>
@@ -149,7 +193,7 @@ describe('ModelSettings', () => {
     await renderModelSettings()
     await waitFor(() => expect(getHermesConfigRecord).toHaveBeenCalled())
 
-    expect(screen.queryByRole('switch')).toBeNull()
+    expect(screen.queryByText('Fast')).toBeNull()
   })
 
   it('renders the auxiliary task rows', async () => {
@@ -159,6 +203,31 @@ describe('ModelSettings', () => {
     expect(screen.getAllByText('auto · use main model').length).toBeGreaterThan(0)
   })
 
+  it('saves disabled MoA reference models without removing them', async () => {
+    await renderModelSettings()
+
+    const referenceSwitch = await screen.findByRole('switch', { name: 'Disable reference 1' })
+    fireEvent.click(referenceSwitch)
+
+    const saveButton = await screen.findByRole('button', { name: 'Save' })
+    fireEvent.click(saveButton)
+
+    await waitFor(() =>
+      expect(saveMoaModels).toHaveBeenCalledWith(
+        expect.objectContaining({
+          presets: expect.objectContaining({
+            review: expect.objectContaining({
+              reference_models: [
+                expect.objectContaining({ provider: 'nous', model: 'hermes-4', enabled: false }),
+                expect.objectContaining({ provider: 'openrouter', model: 'deepseek/deepseek-v4-pro', enabled: true })
+              ]
+            })
+          })
+        })
+      )
+    )
+  })
+
   it('assigns an auxiliary task to the main model via setModelAssignment', async () => {
     await renderModelSettings()
 

apps/desktop/src/app/settings/model-settings.tsx

@@ -940,6 +940,21 @@ export function ModelSettings({ onMainModelChanged }: ModelSettingsProps) {
           <div className="grid gap-1">
             {currentMoaPreset.reference_models.map((slot, index) => (
               <ListRow
+                action={
+                  <Switch
+                    aria-label={`${slot.enabled !== false ? 'Disable' : 'Enable'} reference ${index + 1}`}
+                    checked={slot.enabled !== false}
+                    disabled={applying}
+                    onCheckedChange={checked =>
+                      updateMoaPreset(prev => ({
+                        ...prev,
+                        reference_models: prev.reference_models.map((s, i) =>
+                          i === index ? { ...s, enabled: checked === true } : s
+                        )
+                      }))
+                    }
+                  />
+                }
                 below={
                   <div className="mt-2 flex flex-wrap items-center gap-2 pt-1">
                     <Select
@@ -1001,6 +1016,7 @@ export function ModelSettings({ onMainModelChanged }: ModelSettingsProps) {
                     </Button>
                   </div>
                 }
+                className={cn(slot.enabled === false && 'opacity-60')}
                 description={
                   <span className="font-mono text-[0.68rem]">
                     {slot.provider} · {slot.model}
@@ -1013,7 +1029,10 @@ export function ModelSettings({ onMainModelChanged }: ModelSettingsProps) {
             <Button
               disabled={applying}
               onClick={() =>
-                updateMoaPreset(prev => ({ ...prev, reference_models: [...prev.reference_models, prev.aggregator] }))
+                updateMoaPreset(prev => ({
+                  ...prev,
+                  reference_models: [...prev.reference_models, { ...prev.aggregator, enabled: true }]
+                }))
               }
               size="sm"
               variant="textStrong"

apps/desktop/src/app/settings/primitives.tsx

@@ -66,20 +66,22 @@ export function ListRow({
   hint,
   action,
   below,
-  wide = false
+  wide = false,
+  className
 }: {
   title: ReactNode
   description?: ReactNode
   hint?: ReactNode
   action?: ReactNode
   below?: ReactNode
   wide?: boolean
+  className?: string
 }) {
   return (
     // Container-queried, not viewport-queried: the label/control split keys on
     // the row's own pane width, so a narrow detail column (messaging, split
     // views) stacks instead of squishing the label against minmax(15rem,…).
-    <div className="@container">
+    <div className={cn('@container', className)}>
       <div
         className={cn(
           'grid gap-3 py-3',

apps/desktop/src/types/hermes.ts

@@ -853,6 +853,7 @@ export interface AuxiliaryModelsResponse {
 export interface MoaModelSlot {
   provider: string
   model: string
+  enabled?: boolean
 }
 
 export interface MoaConfigResponse {

hermes_cli/moa_cmd.py

@@ -96,7 +96,9 @@ def cmd_moa(args) -> None:
         idx = 0
         while True:
             base = existing[idx] if idx < len(existing) else None
-            refs.append(_pick_slot(base))
+            picked = _pick_slot(base)
+            picked["enabled"] = bool((base or {}).get("enabled", True))
+            refs.append(picked)
             idx += 1
             choice = _prompt_choice("Add another reference model?", ["Add another", "Done"], 1)
             if choice == 1:

hermes_cli/moa_config.py

@@ -21,6 +21,10 @@
 }
 
 
+def _default_reference_models() -> list[dict[str, Any]]:
+    return [{**slot, "enabled": True} for slot in deepcopy(DEFAULT_MOA_REFERENCE_MODELS)]
+
+
 def _coerce_float_or_none(value: Any) -> float | None:
     """Coerce to a float, or None when unset/blank/invalid.
 
@@ -73,7 +77,22 @@ def _coerce_fanout(value: Any) -> str:
     return mode if mode in {"per_iteration", "user_turn"} else "per_iteration"
 
 
-def _clean_slot(slot: Any) -> dict[str, str] | None:
+def _coerce_bool(value: Any, default: bool = True) -> bool:
+    if value is None:
+        return default
+    if isinstance(value, bool):
+        return value
+    if isinstance(value, str):
+        text = value.strip().lower()
+        if text in {"0", "false", "no", "off"}:
+            return False
+        if text in {"1", "true", "yes", "on"}:
+            return True
+        return default
+    return bool(value)
+
+
+def _clean_slot(slot: Any, *, include_enabled: bool = False) -> dict[str, Any] | None:
     if not isinstance(slot, dict):
         return None
     provider = str(slot.get("provider") or "").strip()
@@ -87,12 +106,15 @@ def _clean_slot(slot: Any) -> dict[str, str] | None:
     # an invalid slot is dropped, falling back to the preset's defaults.
     if provider.lower() == "moa":
         return None
-    return {"provider": provider, "model": model}
+    clean: dict[str, Any] = {"provider": provider, "model": model}
+    if include_enabled:
+        clean["enabled"] = _coerce_bool(slot.get("enabled"), True)
+    return clean
 
 
 def _default_preset() -> dict[str, Any]:
     return {
-        "reference_models": deepcopy(DEFAULT_MOA_REFERENCE_MODELS),
+        "reference_models": _default_reference_models(),
         "aggregator": deepcopy(DEFAULT_MOA_AGGREGATOR),
         # None = temperature omitted from API calls (provider default),
         # matching single-model agent behavior.
@@ -115,15 +137,15 @@ def _normalize_preset(raw: Any) -> dict[str, Any]:
         # defaults instead of crashing the iteration, mirroring the tolerance
         # for the scalar fields below (reference_temperature / max_tokens).
         raw_refs = [raw_refs] if isinstance(raw_refs, dict) else []
-    refs = [_clean_slot(item) for item in raw_refs]
+    refs = [_clean_slot(item, include_enabled=True) for item in raw_refs]
     refs = [item for item in refs if item is not None]
     if not refs:
-        refs = deepcopy(DEFAULT_MOA_REFERENCE_MODELS)
+        refs = _default_reference_models()
 
     aggregator = _clean_slot(raw.get("aggregator")) or deepcopy(DEFAULT_MOA_AGGREGATOR)
 
     return {
-        "enabled": bool(raw.get("enabled", True)),
+        "enabled": _coerce_bool(raw.get("enabled"), True),
         "reference_models": refs,
         "aggregator": aggregator,
         "reference_temperature": _coerce_float_or_none(raw.get("reference_temperature")),

hermes_cli/web_server.py

@@ -947,6 +947,7 @@ class ModelAssignment(BaseModel):
 class MoaModelSlot(BaseModel):
     provider: str = ""
     model: str = ""
+    enabled: bool = True
 
 
 class MoaPresetPayload(BaseModel):

tests/cli/test_moa_command.py

@@ -79,4 +79,6 @@ def test_decode_legacy_encoded_moa_turn_still_works():
     encoded = build_moa_turn_prompt("hello", _make_cli().config["moa"], preset="review")
     prompt, cfg = decode_moa_turn(encoded)
     assert prompt == "hello"
-    assert cfg["reference_models"] == [{"provider": "openrouter", "model": "deepseek/deepseek-v4-pro"}]
+    assert cfg["reference_models"] == [
+        {"provider": "openrouter", "model": "deepseek/deepseek-v4-pro", "enabled": True}
+    ]