merge: BOM non-stream synthesis v6

furniture-3d-system/src/server/services/bom-costing/bom-costing-multimodal-gateway.test.ts

@@ -11,6 +11,49 @@ describe('BOM 多模态网关', () => {
     jest.restoreAllMocks()
   })
 
+  it('结构汇总默认请求一次性 JSON，并读取标准非流式响应', async () => {
+    const fetchSpy = jest.spyOn(global, 'fetch').mockResolvedValue(
+      new Response(JSON.stringify({
+        id: 'call-json-response',
+        model: 'gpt-5.6-terra',
+        choices: [{
+          message: { content: '{"ok":true}' },
+          finish_reason: 'stop',
+        }],
+        usage: { prompt_tokens: 12, completion_tokens: 5 },
+      }), {
+        status: 200,
+        headers: { 'Content-Type': 'application/json' },
+      }),
+    )
+
+    const result = await callBomCostingMultimodalModel({
+      config: {
+        baseUrl: 'https://ai.example.com/v1',
+        apiKey: 'test-key',
+        model: 'gpt-5.6-terra',
+      },
+      operation: 'SYNTHESIS',
+      systemPrompt: 'test',
+      userContent: 'test',
+      maxTokens: 100,
+      timeoutMs: 5_000,
+    })
+
+    const requestBody = JSON.parse(String(fetchSpy.mock.calls[0]?.[1]?.body)) as {
+      stream?: boolean
+    }
+    expect(requestBody.stream).toBe(false)
+    expect(result).toMatchObject({
+      id: 'call-json-response',
+      model: 'gpt-5.6-terra',
+      content: '{"ok":true}',
+      finishReason: 'stop',
+      promptTokens: 12,
+      completionTokens: 5,
+    })
+  })
+
   it('流式响应空转超过时限时取消读取并返回可恢复的超时错误', async () => {
     let cancelReason: unknown
     const stalledStream = new ReadableStream({

furniture-3d-system/src/server/services/bom-costing/bom-costing-multimodal-gateway.ts

@@ -30,6 +30,16 @@ export interface BomCostingMultimodalCallResult {
   completedFromJsonPayload?: boolean
 }
 
+interface BomCostingChatCompletionResponse {
+  id: string | null
+  model: string | null
+  content: string
+  promptTokens: number | null
+  completionTokens: number | null
+  finishReason: string | null
+  completedFromJsonPayload: boolean
+}
+
 export type BomCostingMultimodalOperation = 'IMAGE_FACTS' | 'SYNTHESIS'
 
 const BOM_COSTING_REASONING_EFFORTS = new Set([
@@ -285,7 +295,7 @@ function isCompleteJsonObjectPayload(content: string) {
 async function readStreamingChatCompletion(
   response: Response,
   options: { signal?: AbortSignal } = {},
-) {
+): Promise<BomCostingChatCompletionResponse> {
   if (!response.body) {
     throw new AppError(
       ErrorCode.BOM_COSTING_INVALID_PROVIDER,
@@ -466,6 +476,41 @@ async function readStreamingChatCompletion(
   }
 }
 
+async function readNonStreamingChatCompletion(
+  response: Response,
+): Promise<BomCostingChatCompletionResponse> {
+  const payload = await response.json() as {
+    id?: string
+    model?: string
+    choices?: Array<{
+      message?: { content?: string }
+      finish_reason?: string | null
+    }>
+    usage?: {
+      prompt_tokens?: number
+      completion_tokens?: number
+    }
+  }
+  const choice = payload.choices?.[0]
+  return {
+    id: payload.id?.trim() || null,
+    model: payload.model?.trim() || null,
+    content: typeof choice?.message?.content === 'string'
+      ? choice.message.content.trim()
+      : '',
+    promptTokens: Number.isFinite(payload.usage?.prompt_tokens)
+      ? payload.usage!.prompt_tokens!
+      : null,
+    completionTokens: Number.isFinite(payload.usage?.completion_tokens)
+      ? payload.usage!.completion_tokens!
+      : null,
+    finishReason: typeof choice?.finish_reason === 'string' && choice.finish_reason.trim()
+      ? choice.finish_reason.trim()
+      : null,
+    completedFromJsonPayload: false,
+  }
+}
+
 function readTransportErrorCode(error: unknown): string | null {
   let current = error
   for (let depth = 0; depth < 4; depth += 1) {
@@ -577,6 +622,12 @@ export async function callBomCostingMultimodalModel(input: {
 
   for (let attempt = 1; attempt <= maxAttempts; attempt += 1) {
     let response: Response
+    // The UI streams durable stage events, not raw model tokens. Synthesis only
+    // needs one strict JSON object, so asking the provider for a regular JSON
+    // response avoids OpenAI-compatible gateways that emit a partial SSE body
+    // and then keep the socket open for minutes. Image analysis keeps streaming
+    // because those responses can be substantially larger.
+    const requestStreamingResponse = input.operation === 'IMAGE_FACTS'
     const timeoutSignal = AbortSignal.timeout(input.timeoutMs)
     const requestSignal = input.signal
       ? AbortSignal.any([timeoutSignal, input.signal])
@@ -597,7 +648,7 @@ export async function callBomCostingMultimodalModel(input: {
           response_format: { type: 'json_object' },
           temperature: 0.1,
           max_tokens: input.maxTokens,
-          stream: true,
+          stream: requestStreamingResponse,
           ...(reasoningEffort ? { reasoning_effort: reasoningEffort } : {}),
         }),
         cache: 'no-store',
@@ -673,11 +724,15 @@ export async function callBomCostingMultimodalModel(input: {
       )
     }
 
-    let result: Awaited<ReturnType<typeof readStreamingChatCompletion>>
+    let result: BomCostingChatCompletionResponse
     try {
-      result = await readStreamingChatCompletion(response, {
-        signal: requestSignal,
-      })
+      const responseContentType = response.headers.get('content-type')?.toLowerCase() ?? ''
+      const responseUsesEventStream = responseContentType.includes('text/event-stream') || (
+        requestStreamingResponse && !responseContentType.includes('application/json')
+      )
+      result = responseUsesEventStream
+        ? await readStreamingChatCompletion(response, { signal: requestSignal })
+        : await readNonStreamingChatCompletion(response)
     } catch (error) {
       const taskDeadlineError = readTaskDeadlineError(requestSignal)
       if (taskDeadlineError) throw taskDeadlineError

furniture-3d-system/src/server/services/bom-costing/bom-costing-structure-blueprint.ts

@@ -360,13 +360,14 @@ function sanitizeBlueprintFactIds(
   }
 }
 
-function getTimeoutMs() {
-  return readBomCostingBoundedInteger(
+function getTimeoutMs(maximumMs = 60_000) {
+  const configured = readBomCostingBoundedInteger(
     process.env.BOM_COSTING_STRUCTURE_TIMEOUT_MS,
-    100_000,
+    60_000,
     30_000,
     180_000,
   )
+  return Math.min(configured, maximumMs)
 }
 
 function recordResponse(trace: BomCostingStructureTrace, response: BomCostingMultimodalCallResult) {
@@ -463,8 +464,8 @@ async function generateBlueprint(input: {
         reviewInstruction: input.analysisInstruction?.trim() || undefined,
       }),
       maxTokens: 2_400,
-      timeoutMs: getTimeoutMs(),
-      retrySynthesisTimeout: true,
+      timeoutMs: getTimeoutMs(60_000),
+      retrySynthesisTimeout: false,
       signal: input.signal,
     })
     recordResponse(input.trace, response)
@@ -557,8 +558,8 @@ async function verifyBlueprintGroup(input: {
         systemPrompt: VERIFICATION_SYSTEM_PROMPT,
         userContent: content,
         maxTokens: 4_000,
-        timeoutMs: getTimeoutMs(),
-        retrySynthesisTimeout: true,
+        timeoutMs: getTimeoutMs(60_000),
+        retrySynthesisTimeout: false,
         signal: input.signal,
       })
       recordResponse(input.trace, response)
@@ -933,8 +934,8 @@ async function instantiateBlueprint(input: {
               reviewInstruction: input.analysisInstruction?.trim() || undefined,
             }),
             maxTokens: 2_200,
-            timeoutMs: getTimeoutMs(),
-            retrySynthesisTimeout: true,
+            timeoutMs: getTimeoutMs(45_000),
+            retrySynthesisTimeout: false,
             signal: input.signal,
           })
           recordResponse(input.trace, response)