AI config migration to app_settings + ai_usage logging

* config/index.js: добавлен reloadAi() — асинхронно подтягивает значения
  из app_settings (категория ai_providers), мутирует config.ai in-place.
  Сервисы продолжают использовать config.ai.* синхронно. 3-уровневый fallback:
  app_settings.value → process.env (новое имя) → process.env (старое имя) → дефолт.

* index.js: добавлен await config.reloadAi() в startup после migrate().
  Добавлен middleware AsyncLocalStorage для проброса service/userId в
  AI-сервисы. Сервис определяется по URL-префиксу (zeropost-blog vs
  zeropost-tool). Подмонтирован роут /api/usage.

* routes/settings.js: PUT и invalidate вызывают config.reloadAi() после
  изменения настройки категории ai_providers — горячая перезагрузка
  без рестарта PM2.

* routes/usage.js (новый): GET /api/usage/summary?range&group_by — сводка
  расходов с разбивкой по сервису/провайдеру/модели. GET /api/usage/recent
  — последние вызовы.

* lib/aiContext.js (новый): обёртка над AsyncLocalStorage для service/userId.

* services/aiUsage.js (новый): log() с расчётом cost_rub по справочнику цен
  Anthropic/OpenAI (USD/1M токенов или USD/картинку) × markup × usd_rub.
  Никогда не бросает наружу — ошибки логирования не валят генерацию.

* services/ai.js: chat() и image() обёрнуты в try/catch с aiUsage.log().
  Все вышестоящие функции (generatePost, transformPost, generateTopics,
  generateArticle) идут через chat() — покрытие 100%.

* services/covers.js: 3 call sites хукнуты (generateCoverViaResponses,
  generateCoverViaImagesEndpoint, generateCoverViaImageGenerations). Заодно
  поправлен process.env.AI_MODEL_IMAGE_VIA_RESPONSES → config.ai.imageModelViaResponses.
  Удалён мёртвый код после throw.

* services/postImages.js: 1 call site (/responses + image_generation tool).
  Тот же fix env-ref.

* services/articleAutoSeries.js: 1 call site (/messages, Anthropic-формат).
  usage.input_tokens / output_tokens корректно парсится.

Verify:
* Все 11 строк app_settings ai_providers заполнены значениями из .env.
* Горячая перезагрузка: PUT /api/settings/admin/AI_TEXT_MODEL_POST -> runtime
  обновился без рестарта PM2.
* Live test: ai.chat() через aiprimetech, 118+49 токенов, 0.0414 ₽ — запись
  в ai_usage с правильным service/provider/model.
* GET /api/usage/summary возвращает корректные totals + breakdown.

src/services/ai.js

@@ -1,6 +1,7 @@
 const axios = require('axios');
 const config = require('../config');
 const pb = require('./promptBuilder');
+const aiUsage = require('./aiUsage');
 
 const client = axios.create({
   baseURL: config.ai.baseUrl,
@@ -25,16 +26,46 @@ async function chat(model, systemPrompt, userPrompt, options = {}) {
   };
   if (temperature !== undefined) body.temperature = temperature;
 
-  const res = await client.post('/chat/completions', body, {
-    headers: { Authorization: `Bearer ${config.ai.apiKey}` },
-  });
+  const started = Date.now();
+  let res;
+  try {
+    res = await client.post('/chat/completions', body, {
+      headers: { Authorization: `Bearer ${config.ai.apiKey}` },
+      baseURL: config.ai.baseUrl, // подхватить актуальный URL (горячая перезагрузка)
+    });
+  } catch (err) {
+    aiUsage.log({
+      provider: aiUsage.providerFromBaseUrl(config.ai.baseUrl),
+      requestType: 'chat',
+      model,
+      durationMs: Date.now() - started,
+      succeeded: false,
+      errorMessage: (err.response?.data?.error?.message || err.message || '').slice(0, 500),
+    }).catch(() => {});
+    throw err;
+  }
 
   const text = res.data.choices?.[0]?.message?.content;
+  const usage = res.data.usage || {};
+
+  // лог успешного вызова — fire-and-forget, не блокируем основной поток
+  aiUsage.log({
+    provider: aiUsage.providerFromBaseUrl(config.ai.baseUrl),
+    requestType: 'chat',
+    model,
+    promptTokens: usage.prompt_tokens,
+    completionTokens: usage.completion_tokens,
+    totalTokens: usage.total_tokens,
+    requestId: res.data.id,
+    durationMs: Date.now() - started,
+    succeeded: true,
+  }).catch(() => {});
+
   if (!text) throw new Error('Empty response from AI gateway');
 
   return {
     text: text.trim(),
-    usage: res.data.usage || {},
+    usage,
   };
 }
 
@@ -43,14 +74,40 @@ async function chat(model, systemPrompt, userPrompt, options = {}) {
  */
 async function image(prompt, options = {}) {
   const { size = '1024x1024' } = options;
-  const res = await client.post('/images/generations', {
+  const started = Date.now();
+  let res;
+  try {
+    res = await client.post('/images/generations', {
+      model: config.ai.models.image,
+      prompt,
+      n: 1,
+      size,
+    }, {
+      headers: { Authorization: `Bearer ${config.ai.imageApiKey}` },
+      baseURL: config.ai.imageBaseUrl,
+    });
+  } catch (err) {
+    aiUsage.log({
+      provider: aiUsage.providerFromBaseUrl(config.ai.imageBaseUrl),
+      requestType: 'image',
+      model: config.ai.models.image,
+      imageCount: 1,
+      durationMs: Date.now() - started,
+      succeeded: false,
+      errorMessage: (err.response?.data?.error?.message || err.message || '').slice(0, 500),
+    }).catch(() => {});
+    throw err;
+  }
+
+  aiUsage.log({
+    provider: aiUsage.providerFromBaseUrl(config.ai.imageBaseUrl),
+    requestType: 'image',
     model: config.ai.models.image,
-    prompt,
-    n: 1,
-    size,
-  }, {
-    headers: { Authorization: `Bearer ${config.ai.imageApiKey}` },
-  });
+    imageCount: 1,
+    durationMs: Date.now() - started,
+    succeeded: true,
+  }).catch(() => {});
+
   return res.data.data?.[0]?.url || res.data.data?.[0]?.b64_json;
 }
 

src/services/aiUsage.js

@@ -0,0 +1,143 @@
+// Логирование AI-расхода в таблицу ai_usage.
+//
+// Цены прошиты в коде (USD за 1M токенов или за 1 картинку — это прямые ставки
+// OpenAI/Anthropic). Реальная стоимость считается так:
+//   cost_usd = (prompt_tokens * input_rate + completion_tokens * output_rate) / 1e6
+//   cost_rub = cost_usd * provider_markup * usd_rub_rate
+// provider_markup и usd_rub_rate берутся из app_settings, можно править из админки.
+//
+// Если модель неизвестна — лог пишется, но cost_rub = null.
+
+const { query } = require('../config/db');
+const aiContext = require('../lib/aiContext');
+const settings = require('./settings');
+
+// USD за 1 МИЛЛИОН токенов (прямые цены Anthropic/OpenAI; aiprimetech как реселлер
+// добавляет наценку — она учитывается через AI_PROVIDER_MARKUP в app_settings).
+const TEXT_PRICES_USD_PER_1M = {
+  'claude-haiku-4-5-20251001':    { input: 1.0,  output: 5.0  },
+  'claude-haiku-4-5':              { input: 1.0,  output: 5.0  },
+  'claude-haiku-4.5':              { input: 1.0,  output: 5.0  },
+  'claude-sonnet-4-6':             { input: 3.0,  output: 15.0 },
+  'claude-sonnet-4.6':             { input: 3.0,  output: 15.0 },
+  'claude-sonnet-4-5':             { input: 3.0,  output: 15.0 },
+  'claude-sonnet-4':               { input: 3.0,  output: 15.0 },
+  'gpt-5':                          { input: 1.25, output: 10.0 },
+  'gpt-5.5':                        { input: 1.25, output: 10.0 },
+  'gpt-4o':                         { input: 2.5,  output: 10.0 },
+  'gpt-4o-mini':                    { input: 0.15, output: 0.6  },
+};
+
+// USD за 1 картинку (приблизительно — aiguoguo не публикует точно).
+const IMAGE_PRICES_USD = {
+  'gpt-image-1':       0.04,
+  'gpt-image-1-mini':  0.01,   // оценка, поправь после первого месяца использования
+  'dall-e-3':          0.04,
+  'dall-e-2':          0.02,
+};
+
+function providerFromBaseUrl(url) {
+  if (!url) return 'unknown';
+  if (url.includes('aiprimetech')) return 'aiprimetech';
+  if (url.includes('aiguoguo'))    return 'aiguoguo';
+  if (url.includes('vsegpt'))      return 'vsegpt';
+  try { return new URL(url).hostname; } catch (_) { return 'unknown'; }
+}
+
+async function computeCostRub({ requestType, model, promptTokens, completionTokens, imageCount }) {
+  let usdRubRate = 95;
+  let markup = 1.2;
+  try {
+    const cfg = await settings.getMany(['AI_USD_RUB_RATE', 'AI_PROVIDER_MARKUP']);
+    if (cfg.AI_USD_RUB_RATE) usdRubRate = parseFloat(cfg.AI_USD_RUB_RATE) || 95;
+    if (cfg.AI_PROVIDER_MARKUP) markup = parseFloat(cfg.AI_PROVIDER_MARKUP) || 1.2;
+  } catch (_) { /* fall back to defaults */ }
+
+  if (requestType === 'chat') {
+    const p = TEXT_PRICES_USD_PER_1M[model];
+    if (!p) return null;
+    const costUsd = ((promptTokens || 0) * p.input + (completionTokens || 0) * p.output) / 1e6;
+    return +(costUsd * markup * usdRubRate).toFixed(4);
+  }
+  if (requestType === 'image' || requestType === 'image_via_responses') {
+    const perImage = IMAGE_PRICES_USD[model];
+    if (perImage === undefined) return null;
+    return +((imageCount || 1) * perImage * markup * usdRubRate).toFixed(4);
+  }
+  return null;
+}
+
+/**
+ * Логирует один AI-вызов. Никогда не бросает наружу — ошибки логируются в stderr,
+ * чтобы баг логирования не ронял production-генерацию.
+ */
+async function log(o) {
+  try {
+    const ctx = aiContext.get();
+    const costRub = await computeCostRub({
+      requestType: o.requestType,
+      model: o.model,
+      promptTokens: o.promptTokens,
+      completionTokens: o.completionTokens,
+      imageCount: o.imageCount,
+    });
+    await query(
+      `INSERT INTO ai_usage (
+        service, provider, request_type, model, user_id,
+        prompt_tokens, completion_tokens, total_tokens, image_count,
+        cost_rub, request_id, meta, duration_ms, succeeded, error_message
+      ) VALUES ($1,$2,$3,$4,$5,$6,$7,$8,$9,$10,$11,$12,$13,$14,$15)`,
+      [
+        ctx.service || 'zeropost-unknown',
+        o.provider || 'unknown',
+        o.requestType,
+        o.model || 'unknown',
+        ctx.userId ?? null,
+        o.promptTokens ?? null,
+        o.completionTokens ?? null,
+        o.totalTokens ?? (((o.promptTokens || 0) + (o.completionTokens || 0)) || null),
+        o.imageCount ?? null,
+        costRub,
+        o.requestId || null,
+        o.meta ? JSON.stringify(o.meta) : null,
+        o.durationMs ?? null,
+        o.succeeded !== false,
+        o.errorMessage || null,
+      ]
+    );
+  } catch (err) {
+    console.error('[aiUsage.log] failed:', err.message);
+  }
+}
+
+/**
+ * Обёртка-измеритель: запускает async-функцию, замеряет длительность,
+ * вытаскивает usage из axios-ответа, логирует. Возвращает оригинальный ответ.
+ */
+async function wrap(meta, fn) {
+  const start = Date.now();
+  try {
+    const resp = await fn();
+    const usage = resp?.data?.usage || resp?.usage || {};
+    await log({
+      ...meta,
+      promptTokens: usage.prompt_tokens ?? usage.input_tokens,
+      completionTokens: usage.completion_tokens ?? usage.output_tokens,
+      totalTokens: usage.total_tokens,
+      requestId: resp?.data?.id || resp?.id,
+      durationMs: Date.now() - start,
+      succeeded: true,
+    });
+    return resp;
+  } catch (err) {
+    await log({
+      ...meta,
+      durationMs: Date.now() - start,
+      succeeded: false,
+      errorMessage: (err.response?.data?.error?.message || err.message || '').slice(0, 500),
+    });
+    throw err;
+  }
+}
+
+module.exports = { log, wrap, providerFromBaseUrl, computeCostRub };

src/services/articleAutoSeries.js

@@ -13,6 +13,7 @@
 const axios = require('axios');
 const { query } = require('../config/db');
 const config = require('../config');
+const aiUsage = require('./aiUsage');
 
 const SERIES_DESCRIPTIONS = [
   {
@@ -59,11 +60,13 @@ ${seriesList}
 Отвечай ТОЛЬКО одним словом — slug серии (prompts / mcp-agents / cases) или "none" если статья ни к одной не подходит достаточно хорошо.
 Выбирай серию только если уверен на 80%+. Лучше "none" чем неточное попадание.`;
 
+  const model = config.ai.models?.post || 'claude-haiku-4-5-20251001';
+  const started = Date.now();
   try {
     const res = await axios.post(
       `${config.ai.baseUrl}/messages`,
       {
-        model: config.ai.models?.post || 'claude-haiku-4-5-20251001',
+        model,
         max_tokens: 10,
         messages: [{ role: 'user', content: prompt }],
       },
@@ -73,12 +76,33 @@ ${seriesList}
       }
     );
 
+    const usage = res.data?.usage || {};
+    aiUsage.log({
+      provider: aiUsage.providerFromBaseUrl(config.ai.baseUrl),
+      requestType: 'chat',
+      model,
+      promptTokens: usage.input_tokens,
+      completionTokens: usage.output_tokens,
+      requestId: res.data?.id,
+      meta: { purpose: 'auto-series-detection' },
+      durationMs: Date.now() - started,
+      succeeded: true,
+    }).catch(() => {});
+
     const raw = res.data?.content?.[0]?.text?.trim().toLowerCase() || 'none';
     // Извлекаем только slug без лишнего текста
     const valid = SERIES_DESCRIPTIONS.map(s => s.slug);
     const found = valid.find(s => raw.includes(s));
     return found || null;
   } catch (err) {
+    aiUsage.log({
+      provider: aiUsage.providerFromBaseUrl(config.ai.baseUrl),
+      requestType: 'chat', model,
+      meta: { purpose: 'auto-series-detection' },
+      durationMs: Date.now() - started,
+      succeeded: false,
+      errorMessage: (err.response?.data?.error?.message || err.message || '').slice(0, 500),
+    }).catch(() => {});
     console.warn('[AutoSeries] Claude detection failed:', err.message.slice(0, 100));
     return null;
   }

src/services/covers.js

@@ -4,6 +4,7 @@ const axios = require('axios');
 const config = require('../config');
 const { query } = require('../config/db');
 const localGen = require('./localCoverGenerator');
+const aiUsage = require('./aiUsage');
 
 const UPLOADS_DIR = process.env.UPLOADS_DIR || '/var/www/zeropost-uploads';
 
@@ -102,32 +103,65 @@ Strictly: no text, no letters, no logos, no human faces, no robots, no brains, n
  * Это работает даже когда /v1/images/generations отдаёт unavailable.
  */
 async function generateCoverViaResponses({ prompt }) {
-  const model = process.env.AI_MODEL_IMAGE_VIA_RESPONSES || 'gpt-5.2';
+  const model = config.ai.imageModelViaResponses || 'gpt-5.2';
   // GPT-5 иногда не вызывает инструмент сама — даём явную инструкцию
   const wrappedInput = `Use the image_generation tool to create the following illustration. Do not write any text response, only call the tool.\n\n${prompt}`;
 
-  const res = await axios.post(
-    `${config.ai.baseUrl}/responses`,
-    {
-      model,
-      input: wrappedInput,
-      tools: [{ type: 'image_generation' }],
-      tool_choice: { type: 'image_generation' },
-    },
-    {
-      headers: { Authorization: `Bearer ${config.ai.imageApiKey}` },
-      timeout: 300_000, // GPT-5 reasoning + image — медленно, до 5 минут
-    }
-  );
+  const started = Date.now();
+  let res;
+  try {
+    res = await axios.post(
+      `${config.ai.baseUrl}/responses`,
+      {
+        model,
+        input: wrappedInput,
+        tools: [{ type: 'image_generation' }],
+        tool_choice: { type: 'image_generation' },
+      },
+      {
+        headers: { Authorization: `Bearer ${config.ai.imageApiKey}` },
+        timeout: 300_000, // GPT-5 reasoning + image — медленно, до 5 минут
+      }
+    );
+  } catch (err) {
+    aiUsage.log({
+      provider: aiUsage.providerFromBaseUrl(config.ai.baseUrl),
+      requestType: 'image_via_responses',
+      model, imageCount: 1,
+      durationMs: Date.now() - started, succeeded: false,
+      errorMessage: (err.response?.data?.error?.message || err.message || '').slice(0, 500),
+    }).catch(() => {});
+    throw err;
+  }
 
   const output = res.data?.output || [];
   const imgCall = output.find(o => o.type === 'image_generation_call');
   if (!imgCall) {
+    aiUsage.log({
+      provider: aiUsage.providerFromBaseUrl(config.ai.baseUrl),
+      requestType: 'image_via_responses', model, imageCount: 1,
+      durationMs: Date.now() - started, succeeded: false,
+      errorMessage: 'No image_generation_call in response output',
+    }).catch(() => {});
     throw new Error('No image_generation_call in response output');
   }
   if (!imgCall.result) {
+    aiUsage.log({
+      provider: aiUsage.providerFromBaseUrl(config.ai.baseUrl),
+      requestType: 'image_via_responses', model, imageCount: 1,
+      durationMs: Date.now() - started, succeeded: false,
+      errorMessage: `image_generation_call without result, status=${imgCall.status}`,
+    }).catch(() => {});
     throw new Error(`image_generation_call without result, status=${imgCall.status}`);
   }
+
+  aiUsage.log({
+    provider: aiUsage.providerFromBaseUrl(config.ai.baseUrl),
+    requestType: 'image_via_responses', model, imageCount: 1,
+    requestId: res.data?.id,
+    durationMs: Date.now() - started, succeeded: true,
+  }).catch(() => {});
+
   const bytes = Buffer.from(imgCall.result, 'base64');
   return {
     bytes,
@@ -142,16 +176,41 @@ async function generateCoverViaResponses({ prompt }) {
  */
 async function generateCoverViaImagesEndpoint({ prompt }) {
   const model = config.ai.models?.image || 'gpt-image-1';
-  const res = await axios.post(
-    `${config.ai.baseUrl}/images/generations`,
-    { model, prompt, n: 1, size: '1536x1024' },
-    {
-      headers: { Authorization: `Bearer ${config.ai.imageApiKey}` },
-      timeout: 120_000,
-    }
-  );
+  const started = Date.now();
+  let res;
+  try {
+    res = await axios.post(
+      `${config.ai.baseUrl}/images/generations`,
+      { model, prompt, n: 1, size: '1536x1024' },
+      {
+        headers: { Authorization: `Bearer ${config.ai.imageApiKey}` },
+        timeout: 120_000,
+      }
+    );
+  } catch (err) {
+    aiUsage.log({
+      provider: aiUsage.providerFromBaseUrl(config.ai.baseUrl),
+      requestType: 'image', model, imageCount: 1,
+      durationMs: Date.now() - started, succeeded: false,
+      errorMessage: (err.response?.data?.error?.message || err.message || '').slice(0, 500),
+    }).catch(() => {});
+    throw err;
+  }
   const item = res.data?.data?.[0];
-  if (!item) throw new Error('Empty image response');
+  if (!item) {
+    aiUsage.log({
+      provider: aiUsage.providerFromBaseUrl(config.ai.baseUrl),
+      requestType: 'image', model, imageCount: 1,
+      durationMs: Date.now() - started, succeeded: false,
+      errorMessage: 'Empty image response',
+    }).catch(() => {});
+    throw new Error('Empty image response');
+  }
+  aiUsage.log({
+    provider: aiUsage.providerFromBaseUrl(config.ai.baseUrl),
+    requestType: 'image', model, imageCount: 1,
+    durationMs: Date.now() - started, succeeded: true,
+  }).catch(() => {});
   if (item.b64_json) return { bytes: Buffer.from(item.b64_json, 'base64'), format: 'png' };
   if (item.url) {
     const r = await axios.get(item.url, { responseType: 'arraybuffer', timeout: 60_000 });
@@ -167,21 +226,46 @@ async function generateCoverViaImagesEndpoint({ prompt }) {
 async function generateCoverViaImageGenerations({ prompt }) {
   const model = config.ai.imageModel || 'gpt-image-1-mini';
   const baseUrl = config.ai.imageBaseUrl || config.ai.baseUrl;
-  const res = await axios.post(
-    `${baseUrl}/images/generations`,
-    {
-      model,
-      prompt: prompt.slice(0, 4000),
-      n: 1,
-      size: '1024x1024',
-    },
-    {
-      headers: { Authorization: `Bearer ${config.ai.imageApiKey}` },
-      timeout: 120_000,
-    }
-  );
+  const started = Date.now();
+  let res;
+  try {
+    res = await axios.post(
+      `${baseUrl}/images/generations`,
+      {
+        model,
+        prompt: prompt.slice(0, 4000),
+        n: 1,
+        size: '1024x1024',
+      },
+      {
+        headers: { Authorization: `Bearer ${config.ai.imageApiKey}` },
+        timeout: 120_000,
+      }
+    );
+  } catch (err) {
+    aiUsage.log({
+      provider: aiUsage.providerFromBaseUrl(baseUrl),
+      requestType: 'image', model, imageCount: 1,
+      durationMs: Date.now() - started, succeeded: false,
+      errorMessage: (err.response?.data?.error?.message || err.message || '').slice(0, 500),
+    }).catch(() => {});
+    throw err;
+  }
   const item = res.data?.data?.[0];
-  if (!item) throw new Error('No image data in response');
+  if (!item) {
+    aiUsage.log({
+      provider: aiUsage.providerFromBaseUrl(baseUrl),
+      requestType: 'image', model, imageCount: 1,
+      durationMs: Date.now() - started, succeeded: false,
+      errorMessage: 'No image data in response',
+    }).catch(() => {});
+    throw new Error('No image data in response');
+  }
+  aiUsage.log({
+    provider: aiUsage.providerFromBaseUrl(baseUrl),
+    requestType: 'image', model, imageCount: 1,
+    durationMs: Date.now() - started, succeeded: true,
+  }).catch(() => {});
   // b64_json или url
   const b64 = item.b64_json;
   if (b64) return { bytes: Buffer.from(b64, 'base64'), format: 'png' };
@@ -190,7 +274,6 @@ async function generateCoverViaImageGenerations({ prompt }) {
     return { bytes: Buffer.from(r.data), format: 'png' };
   }
   throw new Error('No b64_json or url in response');
-  return { bytes: Buffer.from(b64, 'base64'), format: 'png' };
 }
 
 /**

src/services/postImages.js

@@ -2,6 +2,7 @@ const fs = require('fs');
 const path = require('path');
 const axios = require('axios');
 const config = require('../config');
+const aiUsage = require('./aiUsage');
 
 const UPLOADS_DIR = process.env.UPLOADS_DIR || '/var/www/zeropost-uploads';
 
@@ -79,26 +80,57 @@ Channel context: ${channel.niche || channel.name}.
 Composition: 16:9 wide format, balanced, suitable for social media.
 Strictly: no text, no letters, no logos, no faces of real people.`;
 
-  const model = process.env.AI_MODEL_IMAGE_VIA_RESPONSES || 'gpt-5.2';
+  const model = config.ai.imageModelViaResponses || 'gpt-5.2';
   const wrappedInput = `Use the image_generation tool to create the following illustration. Do not write any text response, only call the tool.\n\n${prompt}`;
 
-  const res = await axios.post(
-    `${config.ai.baseUrl}/responses`,
-    {
-      model,
-      input: wrappedInput,
-      tools: [{ type: 'image_generation' }],
-      tool_choice: { type: 'image_generation' },
-    },
-    {
-      headers: { Authorization: `Bearer ${config.ai.imageApiKey}` },
-      timeout: 300_000,
-    }
-  );
+  const started = Date.now();
+  let res;
+  try {
+    res = await axios.post(
+      `${config.ai.baseUrl}/responses`,
+      {
+        model,
+        input: wrappedInput,
+        tools: [{ type: 'image_generation' }],
+        tool_choice: { type: 'image_generation' },
+      },
+      {
+        headers: { Authorization: `Bearer ${config.ai.imageApiKey}` },
+        timeout: 300_000,
+      }
+    );
+  } catch (err) {
+    aiUsage.log({
+      provider: aiUsage.providerFromBaseUrl(config.ai.baseUrl),
+      requestType: 'image_via_responses',
+      model, imageCount: 1,
+      meta: { channel_id: channel.id },
+      durationMs: Date.now() - started, succeeded: false,
+      errorMessage: (err.response?.data?.error?.message || err.message || '').slice(0, 500),
+    }).catch(() => {});
+    throw err;
+  }
 
   const output = res.data?.output || [];
   const imgCall = output.find(o => o.type === 'image_generation_call');
-  if (!imgCall?.result) throw new Error('No image generated');
+  if (!imgCall?.result) {
+    aiUsage.log({
+      provider: aiUsage.providerFromBaseUrl(config.ai.baseUrl),
+      requestType: 'image_via_responses', model, imageCount: 1,
+      meta: { channel_id: channel.id },
+      durationMs: Date.now() - started, succeeded: false,
+      errorMessage: 'No image generated',
+    }).catch(() => {});
+    throw new Error('No image generated');
+  }
+
+  aiUsage.log({
+    provider: aiUsage.providerFromBaseUrl(config.ai.baseUrl),
+    requestType: 'image_via_responses', model, imageCount: 1,
+    requestId: res.data?.id,
+    meta: { channel_id: channel.id },
+    durationMs: Date.now() - started, succeeded: true,
+  }).catch(() => {});
 
   const bytes = Buffer.from(imgCall.result, 'base64');
   const tsKey = `post-${channel.id}-${Date.now()}`;