feat(assist): сбор usage-токенов и отдача в meta (время этапов уже было)

assist.js

@@ -272,6 +272,7 @@ function createSemaphore(max) {
  *   model?:string,
  *   temperature?:number,
  *   responseFormat?:{type:string} // напр. {type:'json_object'}
+ *   usageRef?:object           // мутабельный объект для сбора токенов {prompt,completion,total}
  * }} params
  * @returns {Promise<string>} — содержимое choices[0].message.content
  */
@@ -284,6 +285,7 @@ async function llmChat({
     model = ENV.LLM_MODEL,
     temperature,
     responseFormat,
+    usageRef,
 }) {
     const base = String(baseUrl).replace(/\/+$/, '');
     const url = `${base}/v1/chat/completions`;
@@ -349,6 +351,14 @@ async function llmChat({
     }
 
     const data = await response.json();
+
+    // Собираем токены (usage) для отображения пользователю.
+    if (usageRef && data && data.usage) {
+        usageRef.prompt = (usageRef.prompt || 0) + (data.usage.prompt_tokens || 0);
+        usageRef.completion = (usageRef.completion || 0) + (data.usage.completion_tokens || 0);
+        usageRef.total = (usageRef.total || 0) + (data.usage.total_tokens || 0);
+    }
+
     const content = data && data.choices && data.choices[0] &&
         data.choices[0].message && data.choices[0].message.content;
 

orchestrator.js

@@ -241,11 +241,13 @@ async function runPipeline(question, stand, vmDocsBaseUrl, controller, start) {
     // Этап 1 (resource_type).
     let resourceType = null;
     let stage1Ok = false;
+    const usage = { prompt: 0, completion: 0, total: 0 };
     const budget = TOTAL_DEADLINE_MS - (Date.now() - start);
     if (budget > 0 && resourceTypes.size > 0) {
         const r1 = await stage1ExtractResourceType(question, {
             resourceTypesSet: resourceTypes,
             controller,
+            usageRef: usage,
         });
         stage1Ok = r1.ok;
         if (r1.ok && r1.resourceType) resourceType = r1.resourceType;
@@ -267,6 +269,7 @@ async function runPipeline(question, stand, vmDocsBaseUrl, controller, start) {
         model: ENV.LLM_MODEL,
         maxTokens: 2048,
         controller,
+        usageRef: usage,
     });
     const tStage2 = Date.now();
 
@@ -288,6 +291,9 @@ async function runPipeline(question, stand, vmDocsBaseUrl, controller, start) {
             stage1_ms: tStage1 - tIndex,
             search_ms: tSearch - tStage1,
             stage2_ms: tStage2 - tSearch,
+            tokens: usage.total,
+            prompt_tokens: usage.prompt,
+            completion_tokens: usage.completion,
         },
     };
 }

stage1.js

@@ -109,6 +109,7 @@ function parseResourceType(content) {
  * @param {string} [opts.model]
  * @param {number} [opts.timeoutMs]
  * @param {AbortController} [opts.controller]
+ * @param {object} [opts.usageRef]
  * @returns {Promise<{ok:boolean, resourceType?:string|null, error?:string}>}
  */
 async function stage1ExtractResourceType(question, {
@@ -118,6 +119,7 @@ async function stage1ExtractResourceType(question, {
     model = process.env.LLM_MODEL || DEFAULT_MODEL,
     timeoutMs,
     controller,
+    usageRef,
 }) {
     const types = new Set(
         [...new Set(resourceTypesSet || [])].map((t) => String(t).toLowerCase().trim())
@@ -139,6 +141,7 @@ async function stage1ExtractResourceType(question, {
             baseUrl,
             apiKey,
             model,
+            usageRef,
         };
         let timer = null;
         let ctrl = controller;

stage2.js

@@ -229,6 +229,7 @@ function hasHeader(text, header) {
  * @param {string} [params.model]
  * @param {number} [params.maxTokens]
  * @param {AbortController} [params.controller]
+ * @param {object} [params.usageRef]
  * @returns {Promise<{ok:true, answer:string} | {ok:false, error:string}>}
  */
 async function stage2Generate({
@@ -239,6 +240,7 @@ async function stage2Generate({
     model = process.env.LLM_MODEL || DEFAULT_MODEL,
     maxTokens = 2048,
     controller,
+    usageRef,
 }) {
     const docsBlock = context
         .map((c, i) => `[${i + 1}] ${c.title}\n${c.text}`)
@@ -270,6 +272,7 @@ async function stage2Generate({
                 baseUrl,
                 apiKey,
                 model,
+                usageRef,
             });
         } catch (err) {
             const code = err && err.code ? err.code : 'stage2_failed';