elasticdash-test 0.1.18-alpha-24 → 0.1.18-alpha-26

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "elasticdash-test",
-  "version": "0.1.18-alpha-24",
+  "version": "0.1.18-alpha-26",
   "description": "AI-native test runner for ElasticDash workflow testing",
   "type": "module",
   "bin": {

package/src/interceptors/ai-interceptor.ts

@@ -364,6 +364,35 @@ async function bufferSSEStream(
   return completion
 }
 
+/** Extract usage from buffered raw SSE text */
+function extractStreamUsage(provider: string, rawSSE: string): UsageInfo | undefined {
+  const lines = rawSSE.split('\n')
+  // Walk backwards to find usage in the final events
+  for (let i = lines.length - 1; i >= 0; i--) {
+    const line = lines[i]
+    if (!line.startsWith('data: ')) continue
+    const data = line.slice(6).trim()
+    if (data === '[DONE]') continue
+    try {
+      const obj = JSON.parse(data) as Record<string, unknown>
+      // OpenAI / Grok / Kimi: usage in the final chunk
+      const usage = extractUsage(provider, obj)
+      if (usage) return usage
+      // Anthropic: usage in message_delta event
+      if (obj.type === 'message_delta') {
+        const u = (obj as Record<string, unknown>).usage as Record<string, number> | undefined
+        if (u) return extractUsage('anthropic', { usage: u })
+      }
+      // Anthropic: usage in message_start event
+      if (obj.type === 'message_start') {
+        const msg = (obj as Record<string, unknown>).message as Record<string, unknown> | undefined
+        if (msg?.usage) return extractUsage('anthropic', { usage: msg.usage })
+      }
+    } catch { /* skip */ }
+  }
+  return undefined
+}
+
 /** Build a minimal non-streaming JSON response body from a completion string (for replay) */
 function synthesizeCompletionJSON(
   provider: string,
@@ -474,12 +503,30 @@ export function installAIInterceptor(): void {
           url, provider, model: capturedModel, messages: capturedMessages,
           body: capturedReq, promptSnippet: capturedSnippet,
         }
+        const isStreaming = capturedReq.stream === true
         try {
           const cloned = response.clone()
-          const responseBody = await cloned.json() as Record<string, unknown>
-          captured.usage = extractUsage(provider, responseBody)
+          if (!isStreaming) {
+            // Non-streaming: parse JSON response for usage
+            const responseBody = await cloned.json() as Record<string, unknown>
+            captured.usage = extractUsage(provider, responseBody)
+          } else if (cloned.body) {
+            // Streaming: read the raw SSE text to extract usage from final events
+            try {
+              const decoder = new TextDecoder()
+              const reader = cloned.body.getReader()
+              let rawSSE = ''
+              for (;;) {
+                const { done, value } = await reader.read()
+                if (done) break
+                rawSSE += decoder.decode(value, { stream: true })
+              }
+              reader.releaseLock()
+              captured.usage = extractStreamUsage(provider, rawSSE)
+            } catch { /* stream read failed */ }
+          }
         } catch {
-          // Streaming or non-JSON response — usage not available from response
+          // Response body not available — usage won't be captured
         }
         ;(globalThis as Record<string, unknown>)[LLM_REQUEST_KEY] = captured
       }

package/src/interceptors/http.ts

@@ -7,6 +7,7 @@ import { rawDateNow } from './side-effects.js'
 // Skip them here to avoid duplicate observations.
 const AI_URL_PATTERNS = [
   /https?:\/\/api\.openai\.com\/v1\/((chat\/)?completions|embeddings)/,
+  /https?:\/\/api\.anthropic\.com\/v1\/messages/,
   /https?:\/\/generativelanguage\.googleapis\.com\/.*\/models\/[^/:]+:(generateContent|streamGenerateContent)/,
   /https?:\/\/api\.x\.ai\/v1\/(chat\/)?completions/,
   /https?:\/\/api\.moonshot\.ai\/v1\/(chat\/)?completions/,

package/src/interceptors/workflow-ai.ts

@@ -45,6 +45,14 @@ function extractUsage(output: unknown): UsageInfo | undefined {
         totalTokens: u.total_tokens,
       }
     }
+    // Vercel AI SDK: { inputTokens, outputTokens }
+    if (u.inputTokens != null || u.outputTokens != null) {
+      return {
+        inputTokens: u.inputTokens,
+        outputTokens: u.outputTokens,
+        totalTokens: (u.inputTokens ?? 0) + (u.outputTokens ?? 0),
+      }
+    }
   }
   // Gemini SDK: usageMetadata
   if (o.usageMetadata && typeof o.usageMetadata === 'object') {
@@ -55,6 +63,14 @@ function extractUsage(output: unknown): UsageInfo | undefined {
       totalTokens: u.totalTokenCount,
     }
   }
+  // Flat token fields (e.g. { tokens: N } or { outputTokens: N, inputTokens: N })
+  if (typeof o.tokens === 'number' || typeof o.outputTokens === 'number') {
+    return {
+      inputTokens: typeof o.inputTokens === 'number' ? o.inputTokens : undefined,
+      outputTokens: typeof o.outputTokens === 'number' ? o.outputTokens : (typeof o.tokens === 'number' ? o.tokens : undefined),
+      totalTokens: (typeof o.inputTokens === 'number' ? o.inputTokens : 0) + (typeof o.outputTokens === 'number' ? o.outputTokens : (typeof o.tokens === 'number' ? o.tokens : 0)),
+    }
+  }
   return undefined
 }