@elevasis/core 0.7.0 → 0.8.0

package/src/execution/engine/llm/adapters/server/openrouter.ts

@@ -1,235 +1,235 @@
-/**
- * OpenRouter Adapter
- * Stateless translation layer between universal protocol and OpenRouter API
- *
- * OpenRouter provides access to 400+ LLM models (Claude, Gemini, Llama, etc.)
- * via an OpenAI-compatible API with a single API key.
- *
- * Responsibilities:
- * 1. Translate LLMGenerateRequest to OpenRouter format
- * 2. Call OpenRouter API with structured output
- * 3. Parse OpenRouter response
- * 4. Translate back to LLMGenerateResponse
- */
-
-import type OpenAI from 'openai'
-import type { LLMAdapter, LLMGenerateRequest, LLMGenerateResponse } from '../../types'
-import type { LLMModel, OpenRouterOptions } from '../../model-info'
-import { getModelInfo } from '../../model-info'
-import { TokenLimitExceededError, LLMResponseParseError } from '../../errors'
-import { bslogExternalServiceError, isSystemError } from '../../../../../platform/utils/server/betterstack-logger'
-import { composeSignal } from './compose-signal'
-import { DEFAULT_LLM_TIMEOUT } from '../../../../../platform/constants/timeouts'
-
-/**
- * Type guard for errors with HTTP status code
- */
-interface ErrorWithStatus {
-  status: number
-}
-
-function hasStatus(error: unknown): error is ErrorWithStatus {
-  return (
-    typeof error === 'object' &&
-    error !== null &&
-    'status' in error &&
-    typeof (error as ErrorWithStatus).status === 'number'
-  )
-}
-
-/**
- * OpenRouter-specific extra_body options for usage accounting and routing
- */
-interface OpenRouterExtraBody {
-  usage?: { include: boolean }
-  transforms?: string[]
-  route?: 'fallback'
-}
-
-/**
- * Extended params type for OpenRouter with extra_body support
- */
-type OpenRouterCreateParams = OpenAI.Chat.ChatCompletionCreateParamsNonStreaming & {
-  extra_body?: OpenRouterExtraBody
-}
-
-/**
- * OpenRouter adapter configuration
- */
-export interface OpenRouterAdapterConfig {
-  apiKey: string
-  model: LLMModel
-  modelOptions?: OpenRouterOptions
-}
-
-/**
- * Check if the model is an OpenRouter model
- *
- * @param model - Model identifier (e.g., 'openrouter/anthropic/claude-3.5-sonnet')
- * @returns True if this is an OpenRouter model
- */
-export function isOpenRouterModel(model: LLMModel): boolean {
-  return typeof model === 'string' && model.startsWith('openrouter/')
-}
-
-/**
- * OpenRouter Adapter - Implements universal protocol
- * Uses OpenAI SDK with custom baseURL for OpenRouter API
- */
-export class OpenRouterAdapter implements LLMAdapter {
-  private clientPromise: Promise<OpenAI>
-  private model: LLMModel
-  private openRouterModel: string
-  private modelOptions?: OpenRouterOptions
-
-  constructor(config: OpenRouterAdapterConfig) {
-    // Strip 'openrouter/' prefix for API calls
-    // e.g., 'openrouter/anthropic/claude-3.5-sonnet' -> 'anthropic/claude-3.5-sonnet'
-    this.openRouterModel = config.model.replace('openrouter/', '')
-    this.model = config.model
-    this.modelOptions = config.modelOptions
-
-    // Lazy load OpenAI SDK with OpenRouter configuration
-    this.clientPromise = import('openai').then(
-      ({ default: OpenAIClass }) =>
-        new OpenAIClass({
-          baseURL: 'https://openrouter.ai/api/v1',
-          apiKey: config.apiKey,
-          defaultHeaders: {
-            'HTTP-Referer': 'https://elevasis.com',
-            'X-Title': 'Elevasis'
-          }
-        })
-    )
-  }
-
-  /**
-   * Generate structured output from messages using OpenRouter
-   *
-   * @param request - Generic generation request
-   * @returns Structured output with usage metadata
-   */
-  async generate<T>(request: LLMGenerateRequest): Promise<LLMGenerateResponse<T>> {
-    const client = await this.clientPromise
-
-    try {
-      // Build OpenRouter-specific extra_body with usage accounting
-      const extraBody: OpenRouterExtraBody = {
-        usage: { include: true }
-      }
-
-      if (this.modelOptions?.transforms) {
-        extraBody.transforms = this.modelOptions.transforms
-      }
-
-      if (this.modelOptions?.route) {
-        extraBody.route = this.modelOptions.route
-      }
-
-      const createParams: OpenRouterCreateParams = {
-        model: this.openRouterModel,
-        messages: request.messages.map((msg) => ({
-          role: msg.role,
-          content: msg.content
-        })),
-        max_tokens: request.maxOutputTokens,
-        response_format: {
-          type: 'json_schema',
-          json_schema: {
-            name: 'llm_response',
-            strict: true,
-            schema: request.responseSchema as Record<string, unknown>
-          }
-        },
-        extra_body: extraBody
-      }
-
-      // Add temperature if specified
-      if (request.temperature !== undefined) {
-        createParams.temperature = request.temperature
-      }
-
-      // Add topP if specified
-      if (request.topP !== undefined) {
-        createParams.top_p = request.topP
-      }
-
-      const response = await client.chat.completions.create(createParams, {
-        signal: composeSignal(DEFAULT_LLM_TIMEOUT, request.signal)
-      })
-
-      const choice = response.choices[0]
-      if (!choice) {
-        throw new Error('No response from OpenRouter')
-      }
-
-      // Detect token limit error
-      if (choice.finish_reason === 'length') {
-        const usage = response.usage
-        const completionTokens = usage?.completion_tokens || 0
-        const modelInfo = getModelInfo(this.model)
-
-        throw new TokenLimitExceededError(
-          `LLM token limit exceeded during generation. ` +
-            `Model consumed all available tokens with no tokens left for output.`,
-          {
-            model: this.model,
-            maxOutputTokens: request.maxOutputTokens,
-            usedTokens: completionTokens,
-            reasoningTokens: 0, // OpenRouter doesn't expose reasoning tokens
-            outputTokens: completionTokens,
-            recommendation: modelInfo
-              ? `Increase maxOutputTokens to at least ${modelInfo.minTokens} for this model.`
-              : 'Increase maxOutputTokens for this model.'
-          }
-        )
-      }
-
-      // Parse response content with error wrapping for retry support
-      const content = choice.message.content || '{}'
-      let output: unknown
-      try {
-        output = JSON.parse(content)
-      } catch (error) {
-        if (error instanceof SyntaxError) {
-          throw new LLMResponseParseError(`Failed to parse LLM response as JSON: ${error.message}`, {
-            rawContent: content.substring(0, 500),
-            parseError: error.message
-          })
-        }
-        throw error
-      }
-
-      // Extract usage data
-      const usage = response.usage
-        ? {
-            inputTokens: response.usage.prompt_tokens,
-            outputTokens: response.usage.completion_tokens,
-            totalTokens: response.usage.total_tokens
-          }
-        : undefined
-
-      // Extract actual cost from OpenRouter (when usage accounting is enabled)
-      // OpenRouter returns cost in the usage object as an extended field
-      const usageWithCost = response.usage as { cost?: number } | undefined
-      const cost = usageWithCost?.cost
-
-      return {
-        output: output as T,
-        usage,
-        cost
-      }
-    } catch (error) {
-      // Add Better Stack logging for system errors (5xx only)
-      if (isSystemError(error)) {
-        bslogExternalServiceError('openrouter', 'chat.completions.create', error as Error, {
-          model: this.model,
-          openRouterModel: this.openRouterModel,
-          statusCode: hasStatus(error) ? error.status : undefined
-        })
-      }
-
-      throw error
-    }
-  }
-}
+/**
+ * OpenRouter Adapter
+ * Stateless translation layer between universal protocol and OpenRouter API
+ *
+ * OpenRouter provides access to 400+ LLM models (Claude, Gemini, Llama, etc.)
+ * via an OpenAI-compatible API with a single API key.
+ *
+ * Responsibilities:
+ * 1. Translate LLMGenerateRequest to OpenRouter format
+ * 2. Call OpenRouter API with structured output
+ * 3. Parse OpenRouter response
+ * 4. Translate back to LLMGenerateResponse
+ */
+
+import type OpenAI from 'openai'
+import type { LLMAdapter, LLMGenerateRequest, LLMGenerateResponse } from '../../types'
+import type { LLMModel, OpenRouterOptions } from '../../model-info'
+import { getModelInfo } from '../../model-info'
+import { TokenLimitExceededError, LLMResponseParseError } from '../../errors'
+import { bslogExternalServiceError, isSystemError } from '../../../../../platform/utils/server/betterstack-logger'
+import { composeSignal } from './compose-signal'
+import { DEFAULT_LLM_TIMEOUT } from '../../../../../platform/constants/timeouts'
+
+/**
+ * Type guard for errors with HTTP status code
+ */
+interface ErrorWithStatus {
+  status: number
+}
+
+function hasStatus(error: unknown): error is ErrorWithStatus {
+  return (
+    typeof error === 'object' &&
+    error !== null &&
+    'status' in error &&
+    typeof (error as ErrorWithStatus).status === 'number'
+  )
+}
+
+/**
+ * OpenRouter-specific extra_body options for usage accounting and routing
+ */
+interface OpenRouterExtraBody {
+  usage?: { include: boolean }
+  transforms?: string[]
+  route?: 'fallback'
+}
+
+/**
+ * Extended params type for OpenRouter with extra_body support
+ */
+type OpenRouterCreateParams = OpenAI.Chat.ChatCompletionCreateParamsNonStreaming & {
+  extra_body?: OpenRouterExtraBody
+}
+
+/**
+ * OpenRouter adapter configuration
+ */
+export interface OpenRouterAdapterConfig {
+  apiKey: string
+  model: LLMModel
+  modelOptions?: OpenRouterOptions
+}
+
+/**
+ * Check if the model is an OpenRouter model
+ *
+ * @param model - Model identifier (e.g., 'openrouter/anthropic/claude-3.5-sonnet')
+ * @returns True if this is an OpenRouter model
+ */
+export function isOpenRouterModel(model: LLMModel): boolean {
+  return typeof model === 'string' && model.startsWith('openrouter/')
+}
+
+/**
+ * OpenRouter Adapter - Implements universal protocol
+ * Uses OpenAI SDK with custom baseURL for OpenRouter API
+ */
+export class OpenRouterAdapter implements LLMAdapter {
+  private clientPromise: Promise<OpenAI>
+  private model: LLMModel
+  private openRouterModel: string
+  private modelOptions?: OpenRouterOptions
+
+  constructor(config: OpenRouterAdapterConfig) {
+    // Strip 'openrouter/' prefix for API calls
+    // e.g., 'openrouter/anthropic/claude-3.5-sonnet' -> 'anthropic/claude-3.5-sonnet'
+    this.openRouterModel = config.model.replace('openrouter/', '')
+    this.model = config.model
+    this.modelOptions = config.modelOptions
+
+    // Lazy load OpenAI SDK with OpenRouter configuration
+    this.clientPromise = import('openai').then(
+      ({ default: OpenAIClass }) =>
+        new OpenAIClass({
+          baseURL: 'https://openrouter.ai/api/v1',
+          apiKey: config.apiKey,
+          defaultHeaders: {
+            'HTTP-Referer': 'https://elevasis.com',
+            'X-Title': 'Elevasis'
+          }
+        })
+    )
+  }
+
+  /**
+   * Generate structured output from messages using OpenRouter
+   *
+   * @param request - Generic generation request
+   * @returns Structured output with usage metadata
+   */
+  async generate<T>(request: LLMGenerateRequest): Promise<LLMGenerateResponse<T>> {
+    const client = await this.clientPromise
+
+    try {
+      // Build OpenRouter-specific extra_body with usage accounting
+      const extraBody: OpenRouterExtraBody = {
+        usage: { include: true }
+      }
+
+      if (this.modelOptions?.transforms) {
+        extraBody.transforms = this.modelOptions.transforms
+      }
+
+      if (this.modelOptions?.route) {
+        extraBody.route = this.modelOptions.route
+      }
+
+      const createParams: OpenRouterCreateParams = {
+        model: this.openRouterModel,
+        messages: request.messages.map((msg) => ({
+          role: msg.role,
+          content: msg.content
+        })),
+        max_tokens: request.maxOutputTokens,
+        response_format: {
+          type: 'json_schema',
+          json_schema: {
+            name: 'llm_response',
+            strict: true,
+            schema: request.responseSchema as Record<string, unknown>
+          }
+        },
+        extra_body: extraBody
+      }
+
+      // Add temperature if specified
+      if (request.temperature !== undefined) {
+        createParams.temperature = request.temperature
+      }
+
+      // Add topP if specified
+      if (request.topP !== undefined) {
+        createParams.top_p = request.topP
+      }
+
+      const response = await client.chat.completions.create(createParams, {
+        signal: composeSignal(DEFAULT_LLM_TIMEOUT, request.signal)
+      })
+
+      const choice = response.choices[0]
+      if (!choice) {
+        throw new Error('No response from OpenRouter')
+      }
+
+      // Detect token limit error
+      if (choice.finish_reason === 'length') {
+        const usage = response.usage
+        const completionTokens = usage?.completion_tokens || 0
+        const modelInfo = getModelInfo(this.model)
+
+        throw new TokenLimitExceededError(
+          `LLM token limit exceeded during generation. ` +
+            `Model consumed all available tokens with no tokens left for output.`,
+          {
+            model: this.model,
+            maxOutputTokens: request.maxOutputTokens,
+            usedTokens: completionTokens,
+            reasoningTokens: 0, // OpenRouter doesn't expose reasoning tokens
+            outputTokens: completionTokens,
+            recommendation: modelInfo
+              ? `Increase maxOutputTokens to at least ${modelInfo.minTokens} for this model.`
+              : 'Increase maxOutputTokens for this model.'
+          }
+        )
+      }
+
+      // Parse response content with error wrapping for retry support
+      const content = choice.message.content || '{}'
+      let output: unknown
+      try {
+        output = JSON.parse(content)
+      } catch (error) {
+        if (error instanceof SyntaxError) {
+          throw new LLMResponseParseError(`Failed to parse LLM response as JSON: ${error.message}`, {
+            rawContent: content.substring(0, 500),
+            parseError: error.message
+          })
+        }
+        throw error
+      }
+
+      // Extract usage data
+      const usage = response.usage
+        ? {
+            inputTokens: response.usage.prompt_tokens,
+            outputTokens: response.usage.completion_tokens,
+            totalTokens: response.usage.total_tokens
+          }
+        : undefined
+
+      // Extract actual cost from OpenRouter (when usage accounting is enabled)
+      // OpenRouter returns cost in the usage object as an extended field
+      const usageWithCost = response.usage as { cost?: number } | undefined
+      const cost = usageWithCost?.cost
+
+      return {
+        output: output as T,
+        usage,
+        cost
+      }
+    } catch (error) {
+      // Add Better Stack logging for system errors (5xx only)
+      if (isSystemError(error)) {
+        bslogExternalServiceError('openrouter', 'chat.completions.create', error as Error, {
+          model: this.model,
+          openRouterModel: this.openRouterModel,
+          statusCode: hasStatus(error) ? error.status : undefined
+        })
+      }
+
+      throw error
+    }
+  }
+}