feat(runtime): restore DeepSeek native WebSearch

docs/web-search-provider-capability.md

@@ -19,10 +19,11 @@
 
 # Provider-hosted web search capability
 
-Status: implemented for OpenAI Responses and Anthropic Messages
-`web_search_20250305`; other provider-native wires remain explicit follow-ups.
+Status: implemented for OpenAI Responses, Alibaba Token Plan Open Responses,
+and Anthropic Messages `web_search_20250305`; other provider-native wires
+remain explicit follow-ups.
 Verified against public provider documentation and shipped client/SDK behavior
-on 2026-08-04.
+through 2026-08-24.
 
 ## Problem
 
@@ -59,8 +60,15 @@ set `capabilities.webSearch`; otherwise narrow provider/model rules apply.
 The implemented native adapters are:
 
 - `openai-responses` for Codex-style `web_search`;
+- `openai-responses` over the Alibaba Token Plan Open Responses profile for
+  `qwen3.8-max` Harness `web_search`;
 - `anthropic-messages` for Claude Code-compatible `web_search_20250305`.
 
+Alibaba Token Plan stays deliberately narrower than the provider's published
+model matrix: the current Runtime routes only `qwen3.8-max` through its verified
+Responses profile. Qwen 3.7 models remain on Chat in Maka and therefore do not
+receive a provider-native search tool yet.
+
 ## Execution surfaces
 
 All production `AiSdkBackend` composition roots use the same
@@ -105,8 +113,11 @@ turn-start tool surface
 The search request uses the existing model credential. It does not duplicate the
 secret into web-search settings, start a nested model request, or expose the
 secret to the renderer. Provider search call/results are marked
-`providerExecuted`, preserved separately from local ToolRuntime execution, and
-replayed through the native provider-tool shape.
+`providerExecuted` and preserved separately from local ToolRuntime execution.
+Adapters that can round-trip the pair replay its native provider-tool shape.
+Alibaba Open Responses currently keeps the durable episode and grounded text but
+omits the pair from the next provider request, because that adapter cannot yet
+round-trip provider-executed results without producing a dangling output.
 
 The durable event keeps normalized `result` data for the canonical read model,
 UI, and exports, plus opaque `providerOutput` data for provider-protocol replay.
@@ -125,6 +136,15 @@ stream. DeepSeek returned search actions but no structured source rows or URL
 annotations in that response, so Maka preserves citations when supplied but
 does not synthesize or invent them.
 
+After DeepSeek moved to the plaintext `@ai-sdk/open-responses` adapter, native
+search was temporarily disabled because that adapter dropped provider-defined
+tools. Live revalidation on 2026-08-24 against first-party
+`deepseek-v4-flash` completed six provider-executed actions (`search`,
+`open_page`, and `find_in_page`) and a grounded final answer. The shared Open
+Responses mapper preserves those actions while the existing replay boundary
+continues to omit the provider-owned pair from the next request instead of
+emitting a dangling function output.
+
 The same key was also verified through DeepSeek's Anthropic-compatible endpoint:
 one real Maka turn sent `web_search_20250305`, received one provider-executed
 WebSearch result containing ten source rows, and completed the final answer
@@ -187,7 +207,8 @@ search-heavy workflows that value source visibility over cache economics.
 | OpenAI API | Responses `web_search` tool | Maka currently enables the native path for GPT-5 families, whose runtime wire is already Responses | Integrated through `openai-responses` |
 | Custom Responses relay | Responses `web_search` tool when explicitly declared by model metadata | `openai-responses-compatible` connections with `apiProtocol=openai-responses` and `capabilities.webSearch=true` | Integrated through `openai-responses` |
 | xAI API / OAuth | Responses Agent Tools `web_search` | Maka currently enables the verified Grok 4.5 Responses route | Integrated through `openai-responses` |
-| Alibaba Model Studio | Responses `web_search` | Qwen 3.5 Plus/Flash provider support is recorded | Provider supports it; Maka Responses adapter pending |
+| Alibaba Token Plan | Responses Harness `web_search` | `qwen3.8-max` on the Token Plan China/Singapore access paths | Integrated through the Alibaba Open Responses profile; live smoke pending |
+| Alibaba Model Studio pay-as-you-go | Responses `web_search` | Provider model and deployment support varies | Provider supports it; Maka pay-as-you-go Responses adapter pending |
 | Anthropic / Claude subscription | Messages `web_search_20250305` | Current Claude Opus/Sonnet/Haiku/Fable families | Integrated through `anthropic-messages` |
 | MiniMax API / Coding Plan | Anthropic-compatible `web_search_20250305` | MiniMax M2.7/M3 families | Integrated through `anthropic-messages`; live provider verification pending |
 | Google Gemini | Gemini API grounding with Google Search | Supported Gemini 2.0+ model families vary by release | Provider supports it; Maka adapter pending |

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

@@ -104,4 +104,12 @@ describe('openAiAdapterApiProtocol', () => {
     assert.equal(openAiAdapterApiProtocol('muse-spark-1.2-contributor', 'opencode'), 'openai-chat');
     assert.equal(openAiAdapterApiProtocol('minimax-m3', 'opencode-go'), 'openai-chat');
   });
+
+  it('routes only Qwen3.8 Max through Alibaba Token Plan Responses', () => {
+    for (const providerType of ['alibaba-token-plan-cn', 'alibaba-token-plan'] as const) {
+      assert.equal(openAiAdapterApiProtocol('qwen3.8-max', providerType), 'openai-responses');
+      assert.equal(openAiAdapterApiProtocol('qwen3.7-max', providerType), 'openai-chat');
+    }
+    assert.equal(openAiAdapterApiProtocol('qwen3.8-max', 'alibaba-cn'), 'openai-chat');
+  });
 });

packages/core/src/__tests__/model-web-search.test.ts

@@ -26,11 +26,11 @@ describe('hosted web search capability', () => {
   it('enables the implemented Responses path only for declared model families', () => {
     assert.deepEqual(resolveHostedWebSearchCapability('deepseek', undefined, 'deepseek-v4-flash'), {
       adapter: 'openai-responses',
-      implemented: false,
+      implemented: true,
     });
     assert.deepEqual(resolveHostedWebSearchCapability('deepseek', undefined, 'deepseek-v4-pro'), {
       adapter: 'openai-responses',
-      implemented: false,
+      implemented: true,
     });
     assert.equal(resolveHostedWebSearchCapability('deepseek', undefined, 'deepseek-chat'), null);
     assert.deepEqual(resolveHostedWebSearchCapability('openai', undefined, 'gpt-5.5'), {
@@ -49,6 +49,19 @@ describe('hosted web search capability', () => {
       adapter: 'openai-responses',
       implemented: false,
     });
+    for (const providerType of ['alibaba-token-plan-cn', 'alibaba-token-plan'] as const) {
+      assert.deepEqual(resolveHostedWebSearchCapability(providerType, undefined, 'qwen3.8-max'), {
+        adapter: 'openai-responses',
+        implemented: true,
+      });
+      assert.equal(resolveHostedWebSearchCapability(providerType, undefined, 'qwen3.7-max'), null);
+      assert.equal(resolveHostedWebSearchCapability(providerType, undefined, 'qwen3.7-plus'), null);
+      assert.equal(
+        resolveHostedWebSearchCapability(providerType, undefined, 'qwen3.8-max-preview'),
+        null,
+      );
+      assert.equal(resolveHostedWebSearchCapability(providerType, undefined, 'qwen3.6-plus'), null);
+    }
     assert.equal(resolveHostedWebSearchCapability('openai', undefined, 'gpt-4.1'), null);
   });
 
@@ -124,7 +137,7 @@ describe('hosted web search capability', () => {
   it('keeps dual-wire providers on the configured connection protocol', () => {
     assert.deepEqual(resolveHostedWebSearchCapability('deepseek', undefined, 'deepseek-v4-flash'), {
       adapter: 'openai-responses',
-      implemented: false,
+      implemented: true,
     });
     assert.deepEqual(
       resolveHostedWebSearchCapability(
@@ -155,6 +168,14 @@ describe('hosted web search capability', () => {
       ),
       null,
     );
+    assert.equal(
+      resolveHostedWebSearchCapability(
+        'alibaba-token-plan-cn',
+        [{ id: 'qwen3.8-max', apiProtocol: 'openai-chat' }],
+        'qwen3.8-max',
+      ),
+      null,
+    );
   });
 
   it('publishes native search for both first-party DeepSeek V4 models', () => {

packages/core/src/__tests__/provider-catalog-contract.test.ts

@@ -78,6 +78,19 @@ describe('provider catalog contract — structural invariants over CATALOG_PROVI
       );
     }
   });
+
+  it('delegates Alibaba Token Plan execution through one explicit Runtime profile', () => {
+    const delegated = Object.entries(PROVIDER_REGISTRY).flatMap(([providerType, definition]) => {
+      const adapter = definition.runtimeAdapter;
+      return adapter.kind === 'openai-compatible' && adapter.runtimeProfile
+        ? [{ providerType, runtimeProfile: adapter.runtimeProfile }]
+        : [];
+    });
+    assert.deepEqual(delegated, [
+      { providerType: 'alibaba-token-plan-cn', runtimeProfile: 'alibaba-token-plan' },
+      { providerType: 'alibaba-token-plan', runtimeProfile: 'alibaba-token-plan' },
+    ]);
+  });
 });
 
 describe('retired provider contract', () => {

packages/core/src/llm-connections.ts

@@ -44,6 +44,7 @@ import {
   type ProviderCategory,
   type ProviderDefaults,
   type ProviderRuntimeAdapter,
+  type ProviderRuntimeProfileId,
   type ProviderResponsesContract,
   type ProviderType,
 } from './provider-registry.js';
@@ -64,6 +65,7 @@ export type {
   ProviderCategory,
   ProviderDefaults,
   ProviderRuntimeAdapter,
+  ProviderRuntimeProfileId,
   ProviderResponsesContract,
   ProviderType,
 };

packages/core/src/model-metadata.ts

@@ -125,11 +125,10 @@ export function lookupModelProviderOverride(
 /**
  * The request wire a model served over the OpenAI adapter must use.
  *
- * OpenAI's `gpt-5*` families and xAI's `grok-4.5` are served only over the
- * Responses API; every other model on the native OpenAI adapter uses Chat
- * Completions. This is the single declared source of that protocol split,
- * expressed through the {@link ModelInfo.apiProtocol} seam. It is consumed by
- * the runtime model factory and the conformance matrix.
+ * Provider/model routing facts live here even when the concrete Responses SDK
+ * and replay policy are delegated to a Runtime profile. This is the single
+ * declared source of the default protocol split, expressed through the
+ * {@link ModelInfo.apiProtocol} seam.
  */
 export function openAiAdapterApiProtocol(
   modelId: string,
@@ -138,6 +137,8 @@ export function openAiAdapterApiProtocol(
   const id = modelId.trim();
   return (providerType === 'deepseek' && deepSeekModelSupportsResponses(id)) ||
     (providerType === 'opencode-go' && id === 'muse-spark-1.2-contributor') ||
+    ((providerType === 'alibaba-token-plan-cn' || providerType === 'alibaba-token-plan') &&
+      id === 'qwen3.8-max') ||
     /^gpt-5/i.test(id) ||
     ((providerType === 'xai' || providerType === 'xai-oauth') && id === 'grok-4.5')
     ? 'openai-responses'

packages/core/src/model-web-search.ts

@@ -77,10 +77,7 @@ function providerHostedWebSearchAdapter(
 ): HostedWebSearchCapability | null {
   switch (providerType) {
     case 'deepseek':
-      // @ai-sdk/open-responses currently serializes function tools only.
-      // Mark native search unavailable so routing never hands it a provider
-      // tool that would be silently filtered from the request.
-      return { adapter: 'openai-responses', implemented: false };
+      return { adapter: 'openai-responses', implemented: true };
     case 'openai':
     case 'openai-responses-compatible':
     case 'xai':
@@ -89,6 +86,9 @@ function providerHostedWebSearchAdapter(
     case 'alibaba':
     case 'alibaba-cn':
       return { adapter: 'openai-responses', implemented: false };
+    case 'alibaba-token-plan-cn':
+    case 'alibaba-token-plan':
+      return { adapter: 'openai-responses', implemented: true };
     case 'anthropic':
     case 'MiniMax':
     case 'MiniMax-cn':
@@ -127,6 +127,11 @@ function providerDefaultHostedWebSearchCapability(
     case 'alibaba':
     case 'alibaba-cn':
       return /^qwen3\.5-(?:plus|flash)(?:[.-]|$)/i.test(modelId) ? capability : null;
+    case 'alibaba-token-plan-cn':
+    case 'alibaba-token-plan':
+      // Keep the hosted tool on the exact model that the Token Plan runtime
+      // routes through Responses. Aliases and Qwen 3.7 still use Chat here.
+      return modelId === 'qwen3.8-max' ? capability : null;
     case 'anthropic':
       return /^claude-(?:[\d.]+-)*(?:opus|sonnet|haiku|fable)\b/i.test(modelId) ? capability : null;
     case 'MiniMax':

packages/core/src/provider-registry.ts

@@ -31,6 +31,13 @@ export type ProviderCatalogGroup = 'recommended' | 'plans' | 'api' | 'aggregator
 
 export type ApplyPatchProtocol = 'openai-structured' | 'codex-v4a-freeform';
 
+/**
+ * Stable reference to provider execution policy implemented by `@maka/runtime`.
+ * Core owns only this protocol-level delegation; SDK selection, replay
+ * carriers, and request mutation remain Runtime implementation details.
+ */
+export type ProviderRuntimeProfileId = 'alibaba-token-plan';
+
 export type ProviderResponsesContract =
   | {
       readonly adapter: 'openai';
@@ -41,6 +48,29 @@ export type ProviderResponsesContract =
       readonly reasoningReplay: 'plaintext-content';
     };
 
+type OpenAiCompatibleRuntimeAdapterBase = {
+  kind: 'openai-compatible';
+  name: 'provider' | 'connection';
+  includeUsage?: boolean;
+  requireBaseUrl?: boolean;
+  replayAssistantReasoningAs?: 'reasoning';
+  replayAssistantReasoningDetails?: true;
+};
+
+type OpenAiCompatibleRuntimeAdapter = OpenAiCompatibleRuntimeAdapterBase &
+  (
+    | {
+        /** Presence enables a complete Core-owned Responses contract. */
+        responses?: ProviderResponsesContract;
+        runtimeProfile?: never;
+      }
+    | {
+        responses?: never;
+        /** Explicitly delegates concrete execution policy to `@maka/runtime`. */
+        runtimeProfile: ProviderRuntimeProfileId;
+      }
+  );
+
 type ProviderRuntimeAdapterDefinition =
   | { kind: 'anthropic'; auth: 'api-key' | 'bearer'; normalizeBaseUrl: boolean }
   /**
@@ -53,16 +83,7 @@ type ProviderRuntimeAdapterDefinition =
   | { kind: 'google'; normalizeBaseUrl?: boolean }
   | { kind: 'github-copilot' }
   | { kind: 'cohere' }
-  | {
-      kind: 'openai-compatible';
-      name: 'provider' | 'connection';
-      includeUsage?: boolean;
-      requireBaseUrl?: boolean;
-      /** Presence enables Responses and fixes the only supported SDK/replay pairing. */
-      responses?: ProviderResponsesContract;
-      replayAssistantReasoningAs?: 'reasoning';
-      replayAssistantReasoningDetails?: true;
-    };
+  | OpenAiCompatibleRuntimeAdapter;
 
 export type ProviderRuntimeAdapter = ProviderRuntimeAdapterDefinition & {
   /** Provider wire contract for ApplyPatch. Model support is resolved separately. */
@@ -1664,7 +1685,11 @@ const providerRegistry = {
     fallbackModels: [...alibabaTokenPlanModelIds],
     status: 'ready',
     protocol: 'openai',
-    runtimeAdapter: { kind: 'openai-compatible', name: 'provider' },
+    runtimeAdapter: {
+      kind: 'openai-compatible',
+      name: 'provider',
+      runtimeProfile: 'alibaba-token-plan',
+    },
     modelDiscovery: { kind: 'protocol' },
     category: 'domestic',
     catalogGroup: 'plans',
@@ -1684,7 +1709,11 @@ const providerRegistry = {
     fallbackModels: [...alibabaTokenPlanModelIds],
     status: 'ready',
     protocol: 'openai',
-    runtimeAdapter: { kind: 'openai-compatible', name: 'provider' },
+    runtimeAdapter: {
+      kind: 'openai-compatible',
+      name: 'provider',
+      runtimeProfile: 'alibaba-token-plan',
+    },
     modelDiscovery: { kind: 'protocol' },
     category: 'overseas',
     catalogGroup: 'plans',