feat(llm): migrate runtime to ai sdk

docs/specs/ai-sdk-runtime/plan.md

@@ -0,0 +1,9 @@
+# AI SDK Runtime Plan
+
+1. Add hidden runtime resolution and keep `legacy` available for rollback.
+2. Introduce shared AI SDK runtime modules without changing upper-layer interfaces.
+3. Migrate OpenAI-compatible and OpenAI responses providers first.
+4. Migrate Anthropic / Gemini / Vertex / Bedrock / Ollama to the shared runtime.
+5. Keep routing providers (`new-api`, `zenmux`) as thin delegates over migrated providers.
+6. Freeze `LLMCoreStreamEvent` behavior with adapter-focused tests.
+7. Remove legacy state machines only after the rollback window closes.

docs/specs/ai-sdk-runtime/spec.md

@@ -0,0 +1,90 @@
+# AI SDK Runtime Spec
+
+## Goal
+
+Unify DeepChat's low-level LLM request pipeline on Vercel AI SDK while keeping the upper-layer contracts unchanged:
+
+- `BaseLLMProvider`
+- `LLMProviderPresenter`
+- `LLMCoreStreamEvent`
+- existing provider IDs, model configs, and conversation history
+
+The AI SDK runtime becomes the default implementation. A hidden runtime switch keeps `legacy` available as a rollback path.
+
+## Non-Negotiable Compatibility
+
+- No functional regression in text streaming, reasoning streaming, tool call streaming, image output, prompt cache, proxy handling, request tracing, routing, and embeddings.
+- `LLMCoreStreamEvent` event names, field names, and stop reasons remain unchanged.
+- Existing `function_call_record` history must stay reusable across providers and across runtime switches.
+- Existing provider list / model list / provider check / key status responsibilities remain in provider classes.
+
+## Runtime Modes
+
+- Default: `ai-sdk`
+- Hidden fallback: `legacy`
+- Resolution order:
+  1. `DEEPCHAT_LLM_RUNTIME`
+  2. config setting `llmRuntimeMode`
+  3. default `ai-sdk`
+
+## Scope
+
+Shared runtime under `src/main/presenter/llmProviderPresenter/aiSdk/` provides:
+
+- provider factory
+- model / message mapper
+- MCP tool mapper
+- streaming adapter
+- image runtime
+- embedding runtime
+- provider-options mapper
+- reasoning middleware
+- legacy function-call compatibility middleware
+
+## Provider Rollout
+
+Phase 1:
+
+- `OpenAICompatibleProvider`
+- `OpenAIResponsesProvider`
+- all `extends OpenAICompatibleProvider` providers
+
+Phase 2:
+
+- `AnthropicProvider`
+- `GeminiProvider`
+- `VertexProvider`
+- `AwsBedrockProvider`
+- `OllamaProvider`
+
+Phase 3:
+
+- `NewApiProvider`
+- `ZenmuxProvider`
+
+Out of scope for first unification pass:
+
+- `AcpProvider`
+- `VoiceAIProvider`
+
+## Validation Matrix
+
+- pure text
+- reasoning native
+- reasoning via `<think>`
+- native tool streaming
+- legacy `<function_call>` fallback
+- multi-tool history replay
+- image input
+- image output
+- usage mapping
+- prompt cache mapping
+- proxy / trace / abort
+- embeddings
+- hidden rollback path
+
+## Legacy Removal Exit Criteria
+
+- AI SDK runtime passes the provider regression matrix
+- rollback path remains unused for at least one release cycle
+- duplicated legacy stream parsers / tool parsers have no remaining callers

docs/specs/ai-sdk-runtime/tasks.md

@@ -0,0 +1,14 @@
+# AI SDK Runtime Tasks
+
+- [x] Add hidden `llmRuntimeMode` default and env override contract.
+- [x] Freeze migration scope in SDD docs.
+- [x] Add shared AI SDK runtime modules. (implemented in 4c8345a7)
+- [x] Integrate OpenAI-compatible runtime path. (implemented in 4c8345a7)
+- [x] Integrate OpenAI responses runtime path. (implemented in 4c8345a7)
+- [x] Integrate Anthropic runtime path. (implemented in 4c8345a7)
+- [x] Integrate Gemini runtime path. (implemented in 4c8345a7)
+- [x] Integrate Vertex runtime path. (implemented in 4c8345a7)
+- [x] Integrate Bedrock runtime path. (implemented in 4c8345a7)
+- [x] Integrate Ollama runtime path. (implemented in 4c8345a7)
+- [x] Add regression tests for runtime adapter behavior. (implemented in 4c8345a7)
+- [ ] Run format, i18n, lint, and targeted tests.

package.json

@@ -63,6 +63,13 @@
   },
   "dependencies": {
     "@agentclientprotocol/sdk": "^0.16.1",
+    "@ai-sdk/amazon-bedrock": "^4.0.92",
+    "@ai-sdk/anthropic": "^3.0.68",
+    "@ai-sdk/google": "^3.0.61",
+    "@ai-sdk/google-vertex": "^4.0.106",
+    "@ai-sdk/openai": "^3.0.52",
+    "@ai-sdk/openai-compatible": "^2.0.41",
+    "@ai-sdk/provider": "^3.0.8",
     "@anthropic-ai/sdk": "^0.53.0",
     "@aws-sdk/client-bedrock": "^3.958.0",
     "@aws-sdk/client-bedrock-runtime": "^3.958.0",
@@ -74,6 +81,7 @@
     "@jxa/run": "^1.4.0",
     "@larksuiteoapi/node-sdk": "^1.60.0",
     "@modelcontextprotocol/sdk": "^1.28.0",
+    "ai": "^6.0.157",
     "axios": "^1.13.6",
     "better-sqlite3-multiple-ciphers": "12.8.0",
     "cheerio": "^1.2.0",
@@ -96,6 +104,7 @@
     "nanoid": "^5.1.7",
     "node-pty": "^1.1.0",
     "ollama": "^0.5.18",
+    "ollama-ai-provider": "^1.2.0",
     "openai": "^6.33.0",
     "pdf-parse-new": "^1.4.1",
     "run-applescript": "^7.1.0",
@@ -110,6 +119,7 @@
     "zod": "^3.25.76"
   },
   "devDependencies": {
+    "@antv/infographic": "^0.2.7",
     "@electron-toolkit/tsconfig": "^1.0.1",
     "@electron/notarize": "^3.1.1",
     "@iconify-json/lucide": "^1.2.99",
@@ -189,7 +199,6 @@
     "vue-virtual-scroller": "^2.0.0-beta.10",
     "vuedraggable": "^4.1.0",
     "yaml": "^2.8.3",
-    "@antv/infographic": "^0.2.7",
     "zod-to-json-schema": "^3.25.1"
   },
   "simple-git-hooks": {

src/main/presenter/agentRuntimePresenter/accumulator.ts

@@ -1,5 +1,6 @@
 import type { AssistantMessageBlock } from '@shared/types/agent-interface'
 import type { LLMCoreStreamEvent } from '@shared/types/core/llm-events'
+import type { ChatMessageProviderOptions } from '@shared/types/core/chat-message'
 import type { StreamState } from './types'
 
 export function finalizeTrailingPendingNarrativeBlocks(blocks: AssistantMessageBlock[]): void {
@@ -17,32 +18,49 @@ export function finalizeTrailingPendingNarrativeBlocks(blocks: AssistantMessageB
 
 function getCurrentBlock(
   blocks: AssistantMessageBlock[],
-  type: 'content' | 'reasoning_content'
+  type: 'content' | 'reasoning_content',
+  providerOptions?: ChatMessageProviderOptions
 ): AssistantMessageBlock {
+  const providerOptionsJson = serializeProviderOptions(providerOptions)
   const last = blocks[blocks.length - 1]
   if (
     last &&
     last.status === 'pending' &&
-    (last.type === 'content' || last.type === 'reasoning_content') &&
-    last.type !== type
+    (last.type === 'content' || last.type === 'reasoning_content')
   ) {
+    const lastProviderOptionsJson =
+      typeof last.extra?.providerOptionsJson === 'string'
+        ? last.extra.providerOptionsJson
+        : undefined
+
+    if (last.type === type && lastProviderOptionsJson === providerOptionsJson) {
+      return last
+    }
+
     last.status = 'success'
   }
 
-  const current = blocks[blocks.length - 1]
-  if (current && current.type === type && current.status === 'pending') {
-    return current
-  }
   const block: AssistantMessageBlock = {
     type,
     content: '',
     status: 'pending',
-    timestamp: Date.now()
+    timestamp: Date.now(),
+    ...(providerOptionsJson ? { extra: { providerOptionsJson } } : {})
   }
   blocks.push(block)
   return block
 }
 
+function serializeProviderOptions(
+  providerOptions?: ChatMessageProviderOptions
+): string | undefined {
+  if (!providerOptions) {
+    return undefined
+  }
+
+  return JSON.stringify(providerOptions)
+}
+
 function updateReasoningMetadata(state: StreamState, start: number, end: number): void {
   const relativeStart = Math.max(0, start - state.startTime)
   const relativeEnd = Math.max(0, end - state.startTime)
@@ -61,15 +79,15 @@ export function accumulate(state: StreamState, event: LLMCoreStreamEvent): void
   switch (event.type) {
     case 'text': {
       if (state.firstTokenTime === null) state.firstTokenTime = Date.now()
-      const block = getCurrentBlock(state.blocks, 'content')
+      const block = getCurrentBlock(state.blocks, 'content', event.provider_options)
       block.content += event.content
       state.dirty = true
       break
     }
     case 'reasoning': {
       const currentTime = Date.now()
       if (state.firstTokenTime === null) state.firstTokenTime = currentTime
-      const block = getCurrentBlock(state.blocks, 'reasoning_content')
+      const block = getCurrentBlock(state.blocks, 'reasoning_content', event.provider_options)
       block.content += event.reasoning_content
       if (
         typeof block.reasoning_time !== 'object' ||
@@ -91,6 +109,7 @@ export function accumulate(state: StreamState, event: LLMCoreStreamEvent): void
     }
     case 'tool_call_start': {
       finalizeTrailingPendingNarrativeBlocks(state.blocks)
+      const providerOptionsJson = serializeProviderOptions(event.provider_options)
       const toolBlock: AssistantMessageBlock = {
         type: 'tool_call',
         content: '',
@@ -101,13 +120,15 @@ export function accumulate(state: StreamState, event: LLMCoreStreamEvent): void
           name: event.tool_call_name,
           params: '',
           response: ''
-        }
+        },
+        ...(providerOptionsJson ? { extra: { providerOptionsJson } } : {})
       }
       state.blocks.push(toolBlock)
       state.pendingToolCalls.set(event.tool_call_id, {
         name: event.tool_call_name,
         arguments: '',
-        blockIndex: state.blocks.length - 1
+        blockIndex: state.blocks.length - 1,
+        providerOptions: event.provider_options
       })
       state.dirty = true
       break
@@ -116,9 +137,18 @@ export function accumulate(state: StreamState, event: LLMCoreStreamEvent): void
       const pending = state.pendingToolCalls.get(event.tool_call_id)
       if (pending) {
         pending.arguments += event.tool_call_arguments_chunk
+        if (!pending.providerOptions && event.provider_options) {
+          pending.providerOptions = event.provider_options
+        }
         const block = state.blocks[pending.blockIndex]
         if (block?.tool_call) {
           block.tool_call.params = pending.arguments
+          if (event.provider_options) {
+            block.extra = {
+              ...block.extra,
+              providerOptionsJson: serializeProviderOptions(event.provider_options)
+            }
+          }
         }
         state.dirty = true
       }
@@ -128,19 +158,24 @@ export function accumulate(state: StreamState, event: LLMCoreStreamEvent): void
       const pending = state.pendingToolCalls.get(event.tool_call_id)
       if (pending) {
         const finalArgs = event.tool_call_arguments_complete ?? pending.arguments
+        const providerOptions = event.provider_options ?? pending.providerOptions
         pending.arguments = finalArgs
         const block = state.blocks[pending.blockIndex]
         if (block?.tool_call) {
           block.tool_call.params = finalArgs
           block.extra = {
             ...block.extra,
-            toolCallArgsComplete: true
+            toolCallArgsComplete: true,
+            ...(providerOptions
+              ? { providerOptionsJson: serializeProviderOptions(providerOptions) }
+              : {})
           }
         }
         state.completedToolCalls.push({
           id: event.tool_call_id,
           name: pending.name,
-          arguments: finalArgs
+          arguments: finalArgs,
+          ...(providerOptions ? { providerOptions } : {})
         })
         state.pendingToolCalls.delete(event.tool_call_id)
         state.dirty = true