feat(core,runtime,desktop): add Fast mode for OpenAI relays

apps/desktop/src/renderer/locales/settings-provider-copy.ts

@@ -23,6 +23,10 @@ const zhCapabilitiesCopy = {
   contextWindow: '上下文窗口（tokens）',
   contextWindowHelp: '声明后压缩与预算按此值计算；留空跟随内置元数据。',
   saveCapabilities: '保存能力声明',
+  fastMode: 'Fast 模式',
+  fastModeHelp: '使用 OpenAI 的 fast service tier；留空跟随服务商默认值。',
+  fastAuto: '自动',
+  fastEnabled: 'Fast',
 };
 const enCapabilitiesCopy = {
   capabilities: 'Capabilities',
@@ -38,6 +42,10 @@ const enCapabilitiesCopy = {
   contextWindow: 'Context window (tokens)',
   contextWindowHelp: 'When set, compaction and budgets use this value; when empty, built-in metadata decides.',
   saveCapabilities: 'Save capability declarations',
+  fastMode: 'Fast mode',
+  fastModeHelp: "Use OpenAI's fast service tier; empty follows the provider default.",
+  fastAuto: 'Auto',
+  fastEnabled: 'Fast',
 };
 
 const zhCopy = {

apps/desktop/src/renderer/settings/provider-connection-detail.tsx

@@ -15,6 +15,7 @@ import { PROVIDER_DEFAULTS } from '@maka/core/llm-connections';
 import {
   DECLARABLE_RELAY_THINKING_LEVELS,
   THINKING_LEVELS,
+  isRelayProviderType,
   type RelayModelProfile,
   type ThinkingLevel,
 } from '@maka/core/model-thinking';
@@ -152,18 +153,19 @@ function ConnectionDetailInner(props: ConnectionDetailProps) {
     setDraftThinkingLevels,
     setDraftVision,
     setDraftContextWindow,
+    setDraftServiceTier,
     saveRelayProfiles,
     runTest,
     refreshModels,
     remove,
     refreshAfterRelogin,
   } = useConnectionDetail(props);
-  // Capability switches only exist for openai-compatible relays: built-in
+  // Capability switches only exist for custom OpenAI relays: built-in
   // providers declare their thinking support in model metadata, a custom
   // relay's backing model is unknown until the user says what it can do. The
   // declaration is per model — a relay can front both a reasoner and a plain
   // instruct model.
-  const showsCapabilities = connection.providerType === 'openai-compatible';
+  const showsCapabilities = isRelayProviderType(connection.providerType);
   // Rows are the enabled models, exactly — the store prunes a model's profile
   // the moment it is disabled, so no declaration can ever belong to a row
   // this list does not show. The editor edits the per-model draft; 保存
@@ -526,6 +528,7 @@ function ConnectionDetailInner(props: ConnectionDetailProps) {
                       ? 'disabled'
                       : 'auto';
                 const draftLevels = declared?.thinkingLevels ?? [];
+                const serviceTierValue = declared?.serviceTier ?? 'auto';
                 // The menu offers the five declarable levels PLUS anything
                 // the stored table already claims — a level saved while it
                 // was still declarable (or hand-written into the document)
@@ -617,6 +620,23 @@ function ConnectionDetailInner(props: ConnectionDetailProps) {
                         }
                       />
                     </CapabilityRow>
+                    <CapabilityRow label={copy.fastMode} description={copy.fastModeHelp}>
+                      <Selector
+                        label={`${copy.fastMode} — ${modelId}`}
+                        isLabelHidden
+                        size="sm"
+                        width={132}
+                        options={[
+                          { value: 'auto', label: copy.fastAuto },
+                          { value: 'fast', label: copy.fastEnabled },
+                        ]}
+                        value={serviceTierValue}
+                        onChange={(value) =>
+                          setDraftServiceTier(modelId, value === 'fast' ? 'fast' : undefined)
+                        }
+                        isDisabled={allActionsBusy}
+                      />
+                    </CapabilityRow>
                   </VStack>
                 );
               })}

apps/desktop/src/renderer/settings/use-connection-detail.ts

@@ -336,7 +336,7 @@ export function useConnectionDetail(props: ConnectionDetailProps) {
     }
   }
 
-  // Per-model profile declarations for openai-compatible relays, edited as a
+  // Per-model profile declarations for custom OpenAI relays, edited as a
   // LOCAL DRAFT and committed by an explicit 保存 button — never keystroke by
   // keystroke. A draft is `Record<modelId, RelayModelProfile>` seeded from the
   // saved table; entries a user empties fully drop out of the map, and the
@@ -407,6 +407,17 @@ export function useConnectionDetail(props: ConnectionDetailProps) {
     });
   }
 
+  function setDraftServiceTier(modelId: string, serviceTier: 'fast' | undefined): void {
+    updateRelayProfileDraft(modelId, (current) => {
+      if (serviceTier === undefined) {
+        if (!current) return current;
+        const { serviceTier: _dropped, ...rest } = current;
+        return Object.keys(rest).length > 0 ? rest : undefined;
+      }
+      return { ...(current ?? {}), serviceTier };
+    });
+  }
+
   // Compare against what persistence would store: drafts pruned to the
   // current selection and order-normalized by the same sanitizer the write
   // path applies, so a reordered-but-equal draft doesn't keep 保存 lit.
@@ -641,6 +652,7 @@ export function useConnectionDetail(props: ConnectionDetailProps) {
     setDraftThinkingLevels,
     setDraftVision,
     setDraftContextWindow,
+    setDraftServiceTier,
     saveRelayProfiles,
     runTest,
     refreshModels,

packages/core/src/__tests__/model-thinking.test.ts

@@ -10,6 +10,7 @@ import {
   thinkingVariantsForConnection,
   thinkingVariantsForModel,
 } from '../model-thinking.js';
+import { isRelayProviderType } from '../llm-connections.js';
 
 test('declarable relay levels are every intensity tier but off', () => {
   // `off` is a disable-wire encoding (reasoning_effort 'none'), not an
@@ -25,6 +26,13 @@ test('declarable relay levels are every intensity tier but off', () => {
   assert.deepEqual([...thinkingVariantsForConnection(declaredOff, 'm')], ['low']);
 });
 
+test('relay profiles preserve the fast service tier declaration', () => {
+  assert.deepEqual(normalizeRelayModelProfiles({ m: { serviceTier: 'fast' } }), {
+    m: { serviceTier: 'fast' },
+  });
+  assert.deepEqual(normalizeRelayModelProfiles({ m: { serviceTier: 'unknown' } }), undefined);
+});
+
 test('relayModelProfile returns undefined without a usable declaration', () => {
   const connection = {
     providerType: 'openai-compatible',
@@ -78,18 +86,32 @@ test('relayModelProfile normalizes order, keeps explicit vision:false, and bound
   }
 });
 
-test('relayModelProfile gates declarations to openai-compatible relays', () => {
+test('relayModelProfile gates declarations to custom OpenAI relays', () => {
   const profiles = { m: { vision: true, contextWindow: 64_000 } };
   assert.deepEqual(
     relayModelProfile({ providerType: 'openai-compatible', relayModelProfiles: profiles }, 'm'),
     { vision: true, contextWindow: 64_000 },
   );
+  assert.deepEqual(
+    relayModelProfile(
+      { providerType: 'openai-responses-compatible', relayModelProfiles: profiles },
+      'm',
+    ),
+    { vision: true, contextWindow: 64_000 },
+  );
   // The same table on a non-relay connection is inert: metadata rules.
   for (const providerType of ['anthropic', 'openai'] as const) {
     assert.equal(relayModelProfile({ providerType, relayModelProfiles: profiles }, 'm'), undefined);
   }
 });
 
+test('isRelayProviderType only accepts the two custom OpenAI relay providers', () => {
+  assert.equal(isRelayProviderType('openai-compatible'), true);
+  assert.equal(isRelayProviderType('openai-responses-compatible'), true);
+  assert.equal(isRelayProviderType('openai'), false);
+  assert.equal(isRelayProviderType('anthropic'), false);
+});
+
 test('normalizeRelayModelProfiles sanitizes write-side tables', () => {
   const sanitized = normalizeRelayModelProfiles({
     reasoner: { thinkingLevels: ['high', 'low', 'turbo'], vision: true, contextWindow: 200_000 },

packages/core/src/__tests__/runtime-policy-codec.test.ts

@@ -222,6 +222,7 @@ test('relay model profiles round-trip canonical entries and drafts, strictly', (
       thinkingLevels: ['minimal', 'low'],
       vision: true,
       contextWindow: 128_000,
+      serviceTier: 'fast',
     },
   };
   const draft = normalizeCreateCatalogConnectionInput({
@@ -237,6 +238,19 @@ test('relay model profiles round-trip canonical entries and drafts, strictly', (
     },
   });
   assert.deepEqual(draft.connection.relayModelProfiles, table);
+  const responsesDraft = normalizeCreateCatalogConnectionInput({
+    expectedCatalogRevision: 0,
+    connection: {
+      slug: 'responses-relay',
+      name: 'Responses Relay',
+      providerType: 'openai-responses-compatible',
+      baseUrl: 'https://responses.example/v1',
+      enabled: true,
+      enabledModelIds: ['relay-reasoner'],
+      relayModelProfiles: table,
+    },
+  });
+  assert.deepEqual(responsesDraft.connection.relayModelProfiles, table);
   // The canonical path re-decodes the same table (entry = draft + identity).
   const entry = decodeCanonicalConnectionCatalogEntry({
     ...draft.connection,

packages/core/src/llm-connections.ts

@@ -49,6 +49,12 @@ export type {
   ProviderType,
 };
 
+export function isRelayProviderType(
+  providerType: ProviderType,
+): providerType is 'openai-compatible' | 'openai-responses-compatible' {
+  return PROVIDER_REGISTRY[providerType].relayModelProfiles === true;
+}
+
 export type ConnectionAuth =
   | { kind: 'api_key'; apiKey: string }
   | { kind: 'optional_api_key'; apiKey?: string }
@@ -107,13 +113,13 @@ export interface RuntimeExecutionConnection {
   defaultModel: string;
   models?: ModelInfo[];
   /**
-   * Per-model user declarations for an `openai-compatible` relay: the facts
+   * Per-model user declarations for a custom OpenAI relay: the facts
    * (offered thinking levels, vision enable/disable, context window) that
    * neither the relay's /models report nor built-in metadata can decide
    * (see `RelayModelProfile` in `model-thinking.ts`). First-class and typed —
    * relay models are unknown to metadata and a catalog refresh rewrites
    * `models[]` rows, so declarations live next to the user-edited fields.
-   * Invariants enforced at store boundaries: only `openai-compatible`
+   * Invariants enforced at store boundaries: only custom OpenAI relay
    * connections carry profiles, and only for ids in `enabledModelIds`
    * (disabling a model deletes its profile).
    */
@@ -598,7 +604,7 @@ export interface UpdateConnectionInput {
   /**
    * Replace the whole relay profiles table: absent leaves it untouched,
    * `null` clears it outright, a table replaces it (with the usual rules —
-   * only `openai-compatible`, only for `enabledModelIds`).
+   * only custom OpenAI relays, only for `enabledModelIds`).
    */
   relayModelProfiles?: RelayModelProfiles | null;
   requestBodyOverlay?: JsonObject | null;

packages/core/src/model-thinking.ts

@@ -17,6 +17,7 @@
  * per-model supported set, so the UI and runtime share one source of truth.
  */
 
+import { isRelayProviderType } from './llm-connections.js';
 import type { ProviderType } from './llm-connections.js';
 import { lookupModelMetadata } from './model-metadata.js';
 
@@ -101,7 +102,8 @@ export function deriveThinkingChoices(
 }
 
 /**
- * One model behind an `openai-compatible` relay, as declared by the user:
+ * One model behind an OpenAI-compatible relay (Chat Completions or Responses),
+ * as declared by the user:
  * the facts neither the relay's /models report nor built-in metadata can be
  * trusted to know. Every field is independent, and every ABSENT field means
  * "Auto" — the /models report and the metadata chain decide. The single
@@ -113,14 +115,16 @@ export function deriveThinkingChoices(
  * (`relayModelProfiles`, keyed by model id): relay models are unknown to
  * `model-metadata.ts` and a catalog refresh rewrites `models[]` rows, so
  * declarations sit next to the user-edited connection fields. Two invariants
- * are enforced at the store boundaries — profiles exist only for
- * `openai-compatible` connections, and only for models in `enabledModelIds`
+ * are enforced at the store boundaries — profiles exist only for custom OpenAI
+ * relay connections, and only for models in `enabledModelIds`
  * (disabling a model deletes its profile).
  */
 export interface RelayModelProfile {
   readonly thinkingLevels?: readonly ThinkingLevel[];
   readonly vision?: boolean;
   readonly contextWindow?: number;
+  /** Use OpenAI's low-latency service tier for this relay model. */
+  readonly serviceTier?: 'fast';
 }
 
 export type RelayModelProfiles = Readonly<Record<string, RelayModelProfile>>;
@@ -135,6 +139,7 @@ function normalizeRelayModelProfile(entry: unknown): RelayModelProfile | undefin
     thinkingLevels?: readonly ThinkingLevel[];
     vision?: boolean;
     contextWindow?: number;
+    serviceTier?: 'fast';
   } = {};
   if (Array.isArray(entry.thinkingLevels)) {
     // Declared levels are filtered to the declarable vocabulary, not merely
@@ -166,6 +171,7 @@ function normalizeRelayModelProfile(entry: unknown): RelayModelProfile | undefin
   ) {
     declared.contextWindow = entry.contextWindow;
   }
+  if (entry.serviceTier === 'fast') declared.serviceTier = 'fast';
   return Object.keys(declared).length > 0 ? (declared as RelayModelProfile) : undefined;
 }
 
@@ -231,12 +237,12 @@ export function relayModelProfile(
   connection: ConnectionThinkingContext,
   modelId: string,
 ): RelayModelProfile | undefined {
-  if (connection.providerType !== 'openai-compatible') return undefined;
+  if (!isRelayProviderType(connection.providerType)) return undefined;
   return normalizeRelayModelProfile(connection.relayModelProfiles?.[modelId]);
 }
 
 /**
- * `openai-compatible` connections declare thinking support **per model** via
+ * OpenAI-compatible relay connections declare thinking support **per model** via
  * `relayModelProfiles[modelId].thinkingLevels` — a relay may front a
  * DeepSeek-family reasoner and a plain instruct model side by side, so the
  * declaration granularity is the model, not the connection. Without a usable

packages/core/src/provider-registry.ts

@@ -83,6 +83,8 @@ export interface ProviderDefaults {
   status: 'ready' | 'phase3-experimental';
   protocol: 'anthropic' | 'openai' | 'google' | 'cohere';
   runtimeAdapter: ProviderRuntimeAdapter;
+  /** User-declared per-model capabilities are authoritative for this provider. */
+  relayModelProfiles?: boolean;
   modelDiscovery: ProviderModelDiscovery;
   category: ProviderCategory;
   catalogGroup?: ProviderCatalogGroup;
@@ -1727,6 +1729,7 @@ const providerRegistry = {
     status: 'ready',
     protocol: 'openai',
     runtimeAdapter: { kind: 'openai-compatible', name: 'connection', requireBaseUrl: true },
+    relayModelProfiles: true,
     modelDiscovery: { kind: 'protocol' },
     category: 'custom',
     catalogGroup: 'aggregators',
@@ -1745,6 +1748,7 @@ const providerRegistry = {
     status: 'ready',
     protocol: 'openai',
     runtimeAdapter: { kind: 'openai', apiProtocol: 'openai-responses' },
+    relayModelProfiles: true,
     modelDiscovery: { kind: 'protocol' },
     category: 'custom',
     catalogGroup: 'aggregators',

packages/core/src/runtime-policy/connection-catalog-codec.ts

@@ -1,4 +1,9 @@
-import { PROVIDER_DEFAULTS, validateSlug, type ProviderType } from '../llm-connections.js';
+import {
+  isRelayProviderType,
+  PROVIDER_DEFAULTS,
+  validateSlug,
+  type ProviderType,
+} from '../llm-connections.js';
 import {
   DECLARABLE_RELAY_THINKING_LEVELS,
   isThinkingLevel,
@@ -205,12 +210,12 @@ export function normalizeConnectionCatalogEntryUpdateForProvider(
   };
 }
 
-// Relay profiles are an openai-compatible feature: on any other provider the
+// Relay profiles are a custom OpenAI relay feature: on any other provider the
 // metadata chain is the truth, and a table here would either sit as dead
 // state or silently shadow metadata for the ungated read seams.
 function rejectForeignProfiles(providerType: ProviderType): void {
-  if (providerType !== 'openai-compatible') {
-    throw domainError('relay model profiles are only supported for openai-compatible connections');
+  if (!isRelayProviderType(providerType)) {
+    throw domainError('relay model profiles are only supported for OpenAI-compatible connections');
   }
 }
 
@@ -244,13 +249,14 @@ export function decodeRelayModelProfilesTable(
     const entry = exactRecord(
       rawEntry,
       `relay model profile for ${modelId}`,
-      ['thinkingLevels', 'vision', 'contextWindow'],
+      ['thinkingLevels', 'vision', 'contextWindow', 'serviceTier'],
       [],
     );
     const declared: {
       thinkingLevels?: readonly ThinkingLevel[];
       vision?: boolean;
       contextWindow?: number;
+      serviceTier?: 'fast';
     } = {};
     if (entry.thinkingLevels !== undefined) {
       if (!Array.isArray(entry.thinkingLevels) || entry.thinkingLevels.length === 0) {
@@ -283,6 +289,12 @@ export function decodeRelayModelProfilesTable(
         Number.MAX_SAFE_INTEGER,
       );
     }
+    if (entry.serviceTier !== undefined) {
+      if (entry.serviceTier !== 'fast') {
+        throw domainError(`declared service tier for ${modelId} must be fast`);
+      }
+      declared.serviceTier = 'fast';
+    }
     if (Object.keys(declared).length === 0) {
       throw domainError(`relay model profile for ${modelId} declares nothing`);
     }