feat: implement reasoning cache for improved performance and error handling in AI SDK integration

packages/shared/adapters/AiSdkToAnthropicSSE.ts

@@ -36,7 +36,8 @@ import type {
   Usage
 } from '@anthropic-ai/sdk/resources/messages'
 import { loggerService } from '@logger'
-import type { FinishReason, LanguageModelUsage, TextStreamPart, ToolSet } from 'ai'
+import { reasoningCache } from '@main/apiServer/services/cache'
+import { type FinishReason, type LanguageModelUsage, type TextStreamPart, type ToolSet } from 'ai'
 
 const logger = loggerService.withContext('AiSdkToAnthropicSSE')
 
@@ -125,6 +126,9 @@ export class AiSdkToAnthropicSSE {
 
       // Ensure all blocks are closed and emit final events
       this.finalize()
+    } catch (error) {
+      await reader.cancel()
+      throw error
     } finally {
       reader.releaseLock()
     }
@@ -188,8 +192,13 @@ export class AiSdkToAnthropicSSE {
         // })
         break
 
-      // === Completion Events ===
       case 'finish-step':
+        if (
+          chunk.providerMetadata?.openrouter?.reasoning_details &&
+          Array.isArray(chunk.providerMetadata.openrouter.reasoning_details)
+        ) {
+          reasoningCache.set('openrouter', chunk.providerMetadata?.openrouter?.reasoning_details)
+        }
         if (chunk.finishReason === 'tool-calls') {
           this.state.stopReason = 'tool_use'
         }
@@ -199,10 +208,8 @@ export class AiSdkToAnthropicSSE {
         this.handleFinish(chunk)
         break
 
-      // === Error Events ===
       case 'error':
-        this.handleError(chunk.error)
+        throw chunk.error
-        break
 
       // Ignore other event types
       default:
@@ -496,33 +503,6 @@ export class AiSdkToAnthropicSSE {
     }
   }
 
-  private handleError(error: unknown): void {
-    // Log the error for debugging
-    logger.warn('AiSdkToAnthropicSSE - Provider error received:', { error })
-
-    // Extract error message
-    let errorMessage = 'Unknown error from provider'
-    if (error && typeof error === 'object') {
-      const err = error as { message?: string; metadata?: { raw?: string } }
-      if (err.metadata?.raw) {
-        errorMessage = `Provider error: ${err.metadata.raw}`
-      } else if (err.message) {
-        errorMessage = err.message
-      }
-    } else if (typeof error === 'string') {
-      errorMessage = error
-    }
-
-    // Emit error as a text block so the user can see it
-    // First close any open thinking blocks to maintain proper event order
-    for (const reasoningId of Array.from(this.state.thinkingBlocks.keys())) {
-      this.stopThinkingBlock(reasoningId)
-    }
-
-    // Emit the error as text
-    this.emitTextDelta(`\n\n[Error: ${errorMessage}]\n`)
-  }
-
   private finalize(): void {
     // Close any open blocks
     if (this.state.textBlockIndex !== null) {

src/main/apiServer/routes/messages.ts

@@ -245,7 +245,6 @@ async function handleUnifiedProcessing({
       res.json(response)
     }
   } catch (error: any) {
-    logger.error('Unified processing error', { error })
     const { statusCode, errorResponse } = messagesService.transformError(error)
     res.status(statusCode).json(errorResponse)
   }

src/main/apiServer/services/cache.ts

@@ -0,0 +1,116 @@
+import { loggerService } from '@logger'
+
+const logger = loggerService.withContext('Cache')
+/**
+ * Cache entry with TTL support
+ */
+interface CacheEntry<T> {
+  details: T[]
+  timestamp: number
+}
+
+/**
+ * In-memory cache for reasoning details
+ * Key: signature
+ * Value: reasoning array with timestamp
+ */
+export class ReasoningCache<T> {
+  private cache = new Map<string, CacheEntry<T>>()
+  private readonly ttlMs: number
+  private cleanupInterval: ReturnType<typeof setInterval> | null = null
+
+  constructor(ttlMs: number = 30 * 60 * 1000) {
+    // Default 30 minutes TTL
+    this.ttlMs = ttlMs
+    this.startCleanup()
+  }
+
+  /**
+   * Store reasoning details by signature
+   */
+  set(signature: string, details: T[]): void {
+    if (!signature || !details.length) return
+
+    this.cache.set(signature, {
+      details,
+      timestamp: Date.now()
+    })
+
+    logger.debug('Cached reasoning details', {
+      signature: signature.substring(0, 20) + '...',
+      detailsCount: details.length
+    })
+  }
+
+  /**
+   * Retrieve reasoning details by signature
+   */
+  get(signature: string): T[] | undefined {
+    const entry = this.cache.get(signature)
+    if (!entry) return undefined
+
+    // Check TTL
+    if (Date.now() - entry.timestamp > this.ttlMs) {
+      this.cache.delete(signature)
+      return undefined
+    }
+
+    logger.debug('Retrieved reasoning details from cache', {
+      signature: signature.substring(0, 20) + '...',
+      detailsCount: entry.details.length
+    })
+
+    return entry.details
+  }
+
+  /**
+   * Clear expired entries
+   */
+  cleanup(): void {
+    const now = Date.now()
+    let cleaned = 0
+
+    for (const [key, entry] of this.cache) {
+      if (now - entry.timestamp > this.ttlMs) {
+        this.cache.delete(key)
+        cleaned++
+      }
+    }
+
+    if (cleaned > 0) {
+      logger.debug('Cleaned up expired reasoning cache entries', { cleaned, remaining: this.cache.size })
+    }
+  }
+
+  /**
+   * Start periodic cleanup
+   */
+  private startCleanup(): void {
+    // Cleanup every 5 minutes
+    this.cleanupInterval = setInterval(() => this.cleanup(), 5 * 60 * 1000)
+  }
+
+  /**
+   * Stop cleanup and clear cache
+   */
+  destroy(): void {
+    if (this.cleanupInterval) {
+      clearInterval(this.cleanupInterval)
+      this.cleanupInterval = null
+    }
+    this.cache.clear()
+  }
+
+  /**
+   * Get cache stats for debugging
+   */
+  stats(): { size: number; ttlMs: number } {
+    return {
+      size: this.cache.size,
+      ttlMs: this.ttlMs
+    }
+  }
+}
+
+// Singleton cache instance
+export const reasoningCache = new ReasoningCache()

src/main/apiServer/services/messages.ts

@@ -4,6 +4,7 @@ import { loggerService } from '@logger'
 import anthropicService from '@main/services/AnthropicService'
 import { buildClaudeCodeSystemMessage, getSdkClient } from '@shared/anthropic'
 import type { Provider } from '@types'
+import { APICallError } from 'ai'
 import { net } from 'electron'
 import type { Response } from 'express'
 
@@ -253,9 +254,36 @@ export class MessagesService {
   }
 
   transformError(error: any): { statusCode: number; errorResponse: ErrorResponse } {
-    let statusCode = 500
+    let statusCode: number | undefined = undefined
-    let errorType = 'api_error'
+    let errorType: string | undefined = undefined
-    let errorMessage = 'Internal server error'
+    let errorMessage: string | undefined = undefined
+
+    const errorMap: Record<number, string> = {
+      400: 'invalid_request_error',
+      401: 'authentication_error',
+      403: 'forbidden_error',
+      404: 'not_found_error',
+      429: 'rate_limit_error',
+      500: 'internal_server_error'
+    }
+
+    if (APICallError.isInstance(error)) {
+      statusCode = error.statusCode
+      errorMessage = error.message
+      if (statusCode) {
+        return {
+          statusCode,
+          errorResponse: {
+            type: 'error',
+            error: {
+              type: errorMap[statusCode] || 'api_error',
+              message: errorMessage,
+              requestId: error.name
+            }
+          }
+        }
+      }
+    }
 
     const anthropicStatus = typeof error?.status === 'number' ? error.status : undefined
     const anthropicError = error?.error
@@ -297,11 +325,11 @@ export class MessagesService {
       typeof errorMessage === 'string' && errorMessage.length > 0 ? errorMessage : 'Internal server error'
 
     return {
-      statusCode,
+      statusCode: statusCode ?? 500,
       errorResponse: {
         type: 'error',
         error: {
-          type: errorType,
+          type: errorType || 'api_error',
           message: safeErrorMessage,
           requestId: error?.request_id
         }

src/main/apiServer/services/unified-messages.ts

@@ -23,11 +23,13 @@ import {
 } from '@shared/provider'
 import { defaultAppHeaders } from '@shared/utils'
 import type { Provider } from '@types'
-import type { ImagePart, ModelMessage, Provider as AiSdkProvider, TextPart, Tool } from 'ai'
+import type { ImagePart, JSONValue, ModelMessage, Provider as AiSdkProvider, TextPart, Tool } from 'ai'
 import { jsonSchema, simulateStreamingMiddleware, stepCountIs, tool, wrapLanguageModel } from 'ai'
 import { net } from 'electron'
 import type { Response } from 'express'
 
+import { reasoningCache } from './cache'
+
 const logger = loggerService.withContext('UnifiedMessagesService')
 
 const MAGIC_STRING = 'skip_thought_signature_validator'
@@ -154,8 +156,6 @@ function convertAnthropicToAiMessages(params: MessageCreateParams): ModelMessage
     }
   }
 
-  // Build a map of tool_use_id -> toolName from all messages first
-  // This is needed because tool_result references tool_use from previous assistant messages
   const toolCallIdToName = new Map<string, string>()
   for (const msg of params.messages) {
     if (Array.isArray(msg.content)) {
@@ -227,13 +227,16 @@ function convertAnthropicToAiMessages(params: MessageCreateParams): ModelMessage
         const assistantContent = [...reasoningParts, ...textParts, ...toolCallParts]
         if (assistantContent.length > 0) {
           let providerOptions: ProviderOptions | undefined = undefined
-          if (isGemini3ModelId(params.model)) {
+          if (reasoningCache.get('openrouter')) {
+            providerOptions = {
+              openrouter: {
+                reasoning_details: (reasoningCache.get('openrouter') as JSONValue[]) || []
+              }
+            }
+          } else if (isGemini3ModelId(params.model)) {
             providerOptions = {
               google: {
                 thoughtSignature: MAGIC_STRING
-              },
-              openrouter: {
-                reasoning_details: []
               }
             }
           }
@@ -367,6 +370,7 @@ export async function streamUnifiedMessages(config: UnifiedStreamConfig): Promis
       middlewares,
       plugins,
       onEvent: (event) => {
+        logger.silly('Streaming event', { eventType: event.type })
         const sseData = formatSSEEvent(event)
         response.write(sseData)
       }
@@ -380,22 +384,6 @@ export async function streamUnifiedMessages(config: UnifiedStreamConfig): Promis
     onComplete?.()
   } catch (error) {
     logger.error('Error in unified message stream', error as Error, { providerId: provider.id, modelId })
-
-    if (!response.writableEnded) {
-      try {
-        const errorMessage = error instanceof Error ? error.message : 'Unknown error'
-        response.write(
-          `event: error\ndata: ${JSON.stringify({
-            type: 'error',
-            error: { type: 'api_error', message: errorMessage }
-          })}\n\n`
-        )
-        response.end()
-      } catch {
-        // Response already ended
-      }
-    }
-
     onError?.(error)
     throw error
   }

src/main/services/agents/services/claudecode/claude-stream-state.ts

@@ -87,6 +87,7 @@ export class ClaudeStreamState {
   private pendingUsage: PendingUsageState = {}
   private pendingToolCalls = new Map<string, PendingToolCall>()
   private stepActive = false
+  private _streamFinished = false
 
   constructor(options: ClaudeStreamStateOptions) {
     this.logger = loggerService.withContext('ClaudeStreamState')
@@ -289,6 +290,16 @@ export class ClaudeStreamState {
   getNamespacedToolCallId(rawToolCallId: string): string {
     return buildNamespacedToolCallId(this.agentSessionId, rawToolCallId)
   }
+
+  /** Marks the stream as finished (either completed or errored). */
+  markFinished(): void {
+    this._streamFinished = true
+  }
+
+  /** Returns true if the stream has already emitted a terminal event. */
+  isFinished(): boolean {
+    return this._streamFinished
+  }
 }
 
 export type { PendingToolCall }

src/main/services/agents/services/claudecode/index.ts

@@ -529,6 +529,19 @@ class ClaudeCodeService implements AgentServiceInterface {
         return
       }
 
+      // Skip emitting error if stream already finished (error was handled via result message)
+      if (streamState.isFinished()) {
+        logger.debug('SDK process exited after stream finished, skipping duplicate error event', {
+          duration,
+          error: errorObj instanceof Error ? { name: errorObj.name, message: errorObj.message } : String(errorObj)
+        })
+        // Still emit complete to signal stream end
+        stream.emit('data', {
+          type: 'complete'
+        })
+        return
+      }
+
       errorChunks.push(errorObj instanceof Error ? errorObj.message : String(errorObj))
       const errorMessage = errorChunks.join('\n\n')
       logger.error('SDK query failed', {

src/main/services/agents/services/claudecode/transform.ts

@@ -121,7 +121,7 @@ export function transformSDKMessageToStreamParts(sdkMessage: SDKMessage, state:
     case 'system':
       return handleSystemMessage(sdkMessage)
     case 'result':
-      return handleResultMessage(sdkMessage)
+      return handleResultMessage(sdkMessage, state)
     default:
       logger.warn('Unknown SDKMessage type', { type: (sdkMessage as any).type })
       return []
@@ -707,7 +707,13 @@ function handleSystemMessage(message: Extract<SDKMessage, { type: 'system' }>):
  * Successful runs yield a `finish` frame with aggregated usage metrics, while
  * failures are surfaced as `error` frames.
  */
-function handleResultMessage(message: Extract<SDKMessage, { type: 'result' }>): AgentStreamPart[] {
+function handleResultMessage(
+  message: Extract<SDKMessage, { type: 'result' }>,
+  state: ClaudeStreamState
+): AgentStreamPart[] {
+  // Mark stream as finished to prevent duplicate error events when SDK process exits
+  state.markFinished()
+
   const chunks: AgentStreamPart[] = []
 
   let usage: LanguageModelUsage | undefined
@@ -719,7 +725,6 @@ function handleResultMessage(message: Extract<SDKMessage, { type: 'result' }>):
     }
   }
 
-  if (message.subtype === 'success') {
   chunks.push({
     type: 'finish',
     totalUsage: usage ?? emptyUsage,
@@ -732,13 +737,21 @@ function handleResultMessage(message: Extract<SDKMessage, { type: 'result' }>):
       raw: message
     }
   } as AgentStreamPart)
-  } else {
+  if (message.subtype !== 'success') {
     chunks.push({
       type: 'error',
       error: {
         message: `${message.subtype}: Process failed after ${message.num_turns} turns`
       }
     } as AgentStreamPart)
+  } else {
+    if (message.is_error) {
+      const errorMatch = message.result.match(/\{.*\}/)
+      if (errorMatch) {
+        const errorDetail = JSON.parse(errorMatch[0])
+        chunks.push(errorDetail)
+      }
+    }
   }
   return chunks
 }