feat: учёт token usage в checks с агрегатами и UI

- миграция checks: nullable input_tokens/output_tokens/total_tokens/usage_json
  (идемпотентно, старые строки остаются NULL; probe мигрирует сам — cron стартует раньше uvicorn)
- probe.py и /api/check фиксируют usage ДО валидации контента: токены
  сохраняются при пустом/сломанном ответе; таймауты и ответы без usage = NULL, не нули
- usage_json хранит сырой usage провайдера (OpenAI/Anthropic/camelCase форматы)
- /api/checks и сегментная история несут токены; новый /api/token_stats:
  rolling 24h + all time, покрытие known/unknown, разбивка по провайдерам
  (провайдер из models с fallback на префикс model_id)
- UI: статы «Токены 24h» и «Покрытие usage» в шапке, токены в тултипах
  сегментов и логе проверок
- статика через __file__-относительный путь (работает и в /app/app, и локально)
- тесты (30, реальные HTTP против мока OmniRoute): миграция/идемпотентность,
  NULL при missing/zero usage, usage при упавшем контенте, таймаут, агрегаты,
  оба пути записи (probe.run + /api/check)

.gitignore

@@ -0,0 +1,4 @@
+.venv/
+.pytest_cache/
+__pycache__/
+.smoke/

app/main.py

@@ -17,6 +17,8 @@ from fastapi import FastAPI, Query
 from fastapi.responses import HTMLResponse, JSONResponse
 from fastapi.staticfiles import StaticFiles
 
+from usage_utils import extract_usage
+
 DB_PATH = os.environ.get("DB_PATH", "/data/arena.db")
 OMNIROUTE_URL = os.environ.get("OMNIROUTE_URL", "http://192.168.1.101:20128")
 OMNIROUTE_KEY = os.environ.get("OMNIROUTE_API_KEY", "")
@@ -25,6 +27,10 @@ ARENA_CACHE_AGENT = os.environ.get("ARENA_CACHE_AGENT", "/data/arena_agent.json"
 # Сколько последних проверок показывать сегментами в колонке Uptime.
 HISTORY_LIMIT = int(os.environ.get("HISTORY_LIMIT", "40"))
 
+# Статика — относительно этого файла: работает и в контейнере (/app/app),
+# и при локальном запуске/тестах из чекаута репозитория.
+STATIC_DIR = os.path.join(os.path.dirname(os.path.abspath(__file__)), "static")
+
 
 # ─── DB setup ───────────────────────────────────────────────────────────────
 
@@ -98,7 +104,11 @@ def init_db():
             model_id TEXT NOT NULL,
             checked_at TEXT NOT NULL,
             ok INTEGER NOT NULL,
-            latency_ms INTEGER
+            latency_ms INTEGER,
+            input_tokens INTEGER,
+            output_tokens INTEGER,
+            total_tokens INTEGER,
+            usage_json TEXT
         );
 
         CREATE INDEX IF NOT EXISTS idx_checks_model ON checks(model_id);
@@ -112,6 +122,17 @@ def init_db():
     if "error" not in cols:
         db.execute("ALTER TABLE checks ADD COLUMN error TEXT")
 
+    # Идемпотентная миграция: учёт token usage. Старые строки остаются NULL
+    # (неизвестное потребление не подменяется нулями), повторный запуск — no-op.
+    for col, decl in (
+        ("input_tokens", "INTEGER"),
+        ("output_tokens", "INTEGER"),
+        ("total_tokens", "INTEGER"),
+        ("usage_json", "TEXT"),
+    ):
+        if col not in cols:
+            db.execute(f"ALTER TABLE checks ADD COLUMN {col} {decl}")
+
     db.commit()
     db.close()
 
@@ -360,7 +381,7 @@ def model_stats(db: sqlite3.Connection, model_id: str, hours: int = 24) -> dict:
     ).fetchone()
 
     rows = db.execute(
-        "SELECT ok, latency_ms, checked_at, error FROM checks "
+        "SELECT ok, latency_ms, checked_at, error, input_tokens, output_tokens, total_tokens FROM checks "
         "WHERE model_id=? AND checked_at > ? ORDER BY checked_at ASC",
         (model_id, cutoff)
     ).fetchall()
@@ -370,13 +391,16 @@ def model_stats(db: sqlite3.Connection, model_id: str, hours: int = 24) -> dict:
     pct = round(100 * success / total) if total else None
 
     # История последних проверок (старые → новые) для сегментной витрины.
-    # Каждый элемент — детали проверки: ok, latency, время, текст ошибки.
+    # Каждый элемент — детали проверки: ok, latency, время, текст ошибки, токены.
     history = [
         {
             "ok": bool(r["ok"]),
             "latency_ms": r["latency_ms"],
             "at": r["checked_at"],
             "error": r["error"],
+            "input_tokens": r["input_tokens"],
+            "output_tokens": r["output_tokens"],
+            "total_tokens": r["total_tokens"],
         }
         for r in rows[-HISTORY_LIMIT:]
     ]
@@ -391,6 +415,93 @@ def model_stats(db: sqlite3.Connection, model_id: str, hours: int = 24) -> dict:
     }
 
 
+# ─── Token usage aggregates ───────────────────────────────────────────────────
+
+def _token_agg_rows(db: sqlite3.Connection, where: str, params: list) -> list[dict]:
+    """GROUP BY провайдер по checks с суммами токенов и покрытием known/unknown."""
+    # Провайдер — из таблицы models; если модели уже нет, выводим из model_id.
+    prov_expr = (
+        "COALESCE(m.provider, "
+        "CASE WHEN instr(c.model_id, '/') > 0 "
+        "THEN substr(c.model_id, 1, instr(c.model_id, '/') - 1) "
+        "ELSE 'unknown' END)"
+    )
+    return [
+        {
+            "provider": r["provider"],
+            "checks": r["checks"],
+            "known": r["known"],
+            "unknown": r["unknown"],
+            "coverage_pct": round(100 * r["known"] / r["checks"]) if r["checks"] else None,
+            "input_tokens": r["input_tokens"],
+            "output_tokens": r["output_tokens"],
+            "total_tokens": r["total_tokens"],
+        }
+        for r in db.execute(
+            f"""
+            SELECT {prov_expr} AS provider,
+                   COUNT(*) AS checks,
+                   SUM(CASE WHEN c.total_tokens IS NOT NULL THEN 1 ELSE 0 END) AS known,
+                   SUM(CASE WHEN c.total_tokens IS NULL THEN 1 ELSE 0 END) AS unknown,
+                   SUM(c.input_tokens) AS input_tokens,
+                   SUM(c.output_tokens) AS output_tokens,
+                   SUM(c.total_tokens) AS total_tokens
+            FROM checks c
+            LEFT JOIN models m ON c.model_id = m.id
+            WHERE {where}
+            GROUP BY provider
+            ORDER BY total_tokens DESC
+            """,
+            params,
+        ).fetchall()
+    ]
+
+
+def _merge_aggs(rows: list[dict]) -> dict:
+    """Свернуть группы провайдеров в одну сводку."""
+    checks = sum(r["checks"] for r in rows)
+    known = sum(r["known"] for r in rows)
+    return {
+        "checks": checks,
+        "known": known,
+        "unknown": checks - known,
+        "coverage_pct": round(100 * known / checks) if checks else None,
+        "input_tokens": sum(r["input_tokens"] or 0 for r in rows),
+        "output_tokens": sum(r["output_tokens"] or 0 for r in rows),
+        "total_tokens": sum(r["total_tokens"] or 0 for r in rows),
+    }
+
+
+def token_stats(db: sqlite3.Connection) -> dict:
+    """Сводка token usage: скользящие 24h и всё время + разбивка по провайдерам."""
+    cutoff24 = datetime.fromtimestamp(time.time() - 24 * 3600, timezone.utc).isoformat()
+    now = datetime.now(timezone.utc).isoformat()
+    prov24 = _token_agg_rows(db, "c.checked_at > ?", [cutoff24])
+    prov_all = _token_agg_rows(db, "1=1", [])
+    all_map = {r["provider"]: r for r in prov_all}
+    return {
+        "generated_at": now,
+        "window_24h": _merge_aggs(prov24),
+        "all_time": _merge_aggs(prov_all),
+        "providers": [
+            {
+                "provider": r["provider"],
+                "window_24h": r,
+                "all_time": all_map.get(r["provider"], _merge_aggs([])),
+            }
+            for r in prov24
+        ] + [
+            {
+                "provider": r["provider"],
+                "window_24h": _merge_aggs([]),
+                "all_time": r,
+            }
+            for r in prov_all if r["provider"] not in {p["provider"] for p in prov24}
+        ],
+    }
+
+
+
 # ─── Lifespan ────────────────────────────────────────────────────────────────
 
 @asynccontextmanager
@@ -493,6 +604,8 @@ def api_check(model_id: str):
     ok = False
     latency = None
     error = None
+    usage = {"input_tokens": None, "output_tokens": None,
+             "total_tokens": None, "usage_json": None}
     try:
         body = json.dumps({
             "model": model_id,
@@ -509,6 +622,9 @@ def api_check(model_id: str):
         )
         with urllib.request.urlopen(req, timeout=30) as r:
             resp = json.loads(r.read())
+            # Usage фиксируем ДО валидации контента: даже если парсинг choices
+            # упадёт, реально потраченные токены сохраняются в checks.
+            usage = extract_usage(resp)
             content = resp["choices"][0]["message"]["content"]
             ok = bool(content.strip())
     except Exception as e:
@@ -520,8 +636,12 @@ def api_check(model_id: str):
     now = datetime.now(timezone.utc).isoformat()
     db = get_db()
     db.execute(
-        "INSERT INTO checks (model_id, checked_at, ok, latency_ms, error) VALUES (?,?,?,?,?)",
+        "INSERT INTO checks (model_id, checked_at, ok, latency_ms, error, "
-        (model_id, now, int(ok), latency, error),
+        "input_tokens, output_tokens, total_tokens, usage_json) "
+        "VALUES (?,?,?,?,?,?,?,?,?)",
+        (model_id, now, int(ok), latency, error,
+         usage["input_tokens"], usage["output_tokens"],
+         usage["total_tokens"], usage["usage_json"]),
     )
     db.commit()
     stats = model_stats(db, model_id)
@@ -531,6 +651,9 @@ def api_check(model_id: str):
         "ok": ok,
         "latency_ms": latency,
         "checked_at": now,
+        "input_tokens": usage["input_tokens"],
+        "output_tokens": usage["output_tokens"],
+        "total_tokens": usage["total_tokens"],
         "history": stats["history"],
         "uptime_pct": stats["uptime_pct"],
     }
@@ -542,8 +665,8 @@ def api_checks(model_id: str):
     limit = int(os.environ.get("CHECKS_LIMIT", "100"))
     db = get_db()
     rows = db.execute(
-        "SELECT ok, latency_ms, checked_at, error FROM checks "
+        "SELECT ok, latency_ms, checked_at, error, input_tokens, output_tokens, total_tokens, usage_json "
-        "WHERE model_id=? ORDER BY checked_at DESC LIMIT ?",
+        "FROM checks WHERE model_id=? ORDER BY checked_at DESC LIMIT ?",
         (model_id, limit),
     ).fetchall()
     db.close()
@@ -555,17 +678,31 @@ def api_checks(model_id: str):
                 "latency_ms": r["latency_ms"],
                 "checked_at": r["checked_at"],
                 "error": r["error"],
+                "input_tokens": r["input_tokens"],
+                "output_tokens": r["output_tokens"],
+                "total_tokens": r["total_tokens"],
+                "usage_json": r["usage_json"],
             }
             for r in rows
         ],
     }
 
 
+@app.get("/api/token_stats")
+def api_token_stats():
+    """Агрегат token usage: скользящие 24h и всё время, покрытие known/unknown,
+    разбивка по провайдерам."""
+    db = get_db()
+    result = token_stats(db)
+    db.close()
+    return JSONResponse(result)
+
+
 @app.get("/", response_class=HTMLResponse)
 def index():
-    with open("/app/app/static/index.html") as f:
+    with open(os.path.join(STATIC_DIR, "index.html")) as f:
         return f.read()
 
 
-app.mount("/static", StaticFiles(directory="/app/app/static"), name="static")
+app.mount("/static", StaticFiles(directory=STATIC_DIR), name="static")
 

app/probe.py

@@ -12,6 +12,8 @@ import time
 import urllib.request
 from datetime import datetime, timezone
 
+from usage_utils import extract_usage
+
 DB_PATH = os.environ.get("DB_PATH", "/data/arena.db")
 OMNIROUTE_URL = os.environ.get("OMNIROUTE_URL", "http://192.168.1.101:20128")
 OMNIROUTE_KEY = os.environ.get("OMNIROUTE_API_KEY", "")
@@ -19,9 +21,17 @@ TIMEOUT = int(os.environ.get("CHECK_TIMEOUT", "30"))
 CONCURRENCY = int(os.environ.get("CHECK_CONCURRENCY", "5"))
 
 
-def check_model(model_id: str) -> tuple[bool, int | None, str | None]:
+def check_model(model_id: str) -> tuple[bool, int | None, str | None, dict]:
+    """Возвращает (ok, latency_ms, error, usage).
+
+    usage извлекается ДО валидации контента: если парсинг choices падает,
+    потраченные токены всё равно сохраняются. Таймаут/ошибка сети без тела
+    ответа дают usage=None по всем полям (NULL, а не нули).
+    """
     t0 = time.time()
     error = None
+    usage = {"input_tokens": None, "output_tokens": None,
+             "total_tokens": None, "usage_json": None}
     try:
         body = json.dumps({
             "model": model_id,
@@ -38,6 +48,8 @@ def check_model(model_id: str) -> tuple[bool, int | None, str | None]:
         )
         with urllib.request.urlopen(req, timeout=TIMEOUT) as r:
             resp = json.loads(r.read())
+            # Usage — до любой валидации контента.
+            usage = extract_usage(resp)
             content = resp["choices"][0]["message"]["content"]
             ok = bool(content.strip())
     except Exception as e:
@@ -45,11 +57,30 @@ def check_model(model_id: str) -> tuple[bool, int | None, str | None]:
         print(f"  ✗ {model_id}: {error}", file=sys.stderr)
         ok = False
     latency = int((time.time() - t0) * 1000)
-    return ok, latency, error
+    return ok, latency, error, usage
+
+
+def _ensure_usage_columns(db: sqlite3.Connection):
+    """Идемпотентно добавить колонки token usage, если БД ещё со старой схемой.
+
+    Cron стартует раньше uvicorn, поэтому probe не может полагаться на то,
+    что init_db из main.py уже прогнал миграцию.
+    """
+    cols = [r[1] for r in db.execute("PRAGMA table_info(checks)").fetchall()]
+    for col, decl in (
+        ("input_tokens", "INTEGER"),
+        ("output_tokens", "INTEGER"),
+        ("total_tokens", "INTEGER"),
+        ("usage_json", "TEXT"),
+    ):
+        if col not in cols:
+            db.execute(f"ALTER TABLE checks ADD COLUMN {col} {decl}")
+    db.commit()
 
 
 def run():
     db = sqlite3.connect(DB_PATH, check_same_thread=False)
+    _ensure_usage_columns(db)
     models = [r[0] for r in db.execute("SELECT id FROM models").fetchall()]
     db.close()
 
@@ -58,8 +89,8 @@ def run():
     from concurrent.futures import ThreadPoolExecutor, as_completed
 
     def do(mid):
-        ok, lat, err = check_model(mid)
+        ok, lat, err, usage = check_model(mid)
-        return mid, ok, lat, err
+        return mid, ok, lat, err, usage
 
     db = sqlite3.connect(DB_PATH, check_same_thread=False)
     now = datetime.now(timezone.utc).isoformat()
@@ -68,10 +99,14 @@ def run():
         futs = {ex.submit(do, m): m for m in models}
         done = ok_count = 0
         for f in as_completed(futs):
-            mid, ok, lat, err = f.result()
+            mid, ok, lat, err, usage = f.result()
             db.execute(
-                "INSERT INTO checks (model_id, checked_at, ok, latency_ms, error) VALUES (?,?,?,?,?)",
+                "INSERT INTO checks (model_id, checked_at, ok, latency_ms, error, "
-                (mid, now, int(ok), lat, err),
+                "input_tokens, output_tokens, total_tokens, usage_json) "
+                "VALUES (?,?,?,?,?,?,?,?,?)",
+                (mid, now, int(ok), lat, err,
+                 usage["input_tokens"], usage["output_tokens"],
+                 usage["total_tokens"], usage["usage_json"]),
             )
             done += 1
             if ok:

app/static/index.html

@@ -245,6 +245,8 @@
   <div class="stat"><div class="val" id="s-ranked">—</div><div class="lbl">В рейтинге</div></div>
   <div class="stat"><div class="val" id="s-alive">—</div><div class="lbl">Живых</div></div>
   <div class="stat"><div class="val" id="s-checked">—</div><div class="lbl">Проверено сегодня</div></div>
+  <div class="stat"><div class="val" id="s-tokens-24h">—</div><div class="lbl">Токены 24ч</div></div>
+  <div class="stat"><div class="val" id="s-tokens-cover">—</div><div class="lbl">Покрытие usage</div></div>
 </div>
 
 <div class="search-wrap">
@@ -351,9 +353,10 @@ function showSegTip(el) {
   const tip = segTip();
   const ok = data.ok;
   const lat = data.latency_ms !== null && data.latency_ms !== undefined ? `${data.latency_ms} мс` : '—';
+  const toks = usageStr(data);
   tip.innerHTML = `
     <div class="tip-head ${ok ? 'tip-ok' : 'tip-fail'}">${ok ? '✓ Успех' : '✗ Провал'}</div>
-    <div class="tip-meta">${fmtTime(data.at)} · ${lat}</div>
+    <div class="tip-meta">${fmtTime(data.at)} · ${lat}${toks ? ` · ${escapeHtml(toks)}` : ''}</div>
     ${data.error ? `<div class="tip-err">${escapeHtml(data.error)}</div>` : ''}
   `;
   tip.style.display = 'block';
@@ -413,6 +416,37 @@ function renderRow(m) {
   </tr>`;
 }
 
+function fmtTokens(n) {
+  if (n === null || n === undefined) return null;
+  if (n >= 1e9) return (n / 1e9).toFixed(1) + 'B';
+  if (n >= 1e6) return (n / 1e6).toFixed(1) + 'M';
+  if (n >= 1e3) return (n / 1e3).toFixed(1) + 'k';
+  return String(n);
+}
+
+// Строка токенов для тултипа/лога: "in 12 → out 5" или null если usage неизвестен.
+function usageStr(u) {
+  if (!u || u.total_tokens === null || u.total_tokens === undefined) return null;
+  const inp = u.input_tokens !== null && u.input_tokens !== undefined ? u.input_tokens : '?';
+  const out = u.output_tokens !== null && u.output_tokens !== undefined ? u.output_tokens : '?';
+  return `in ${inp} → out ${out} (${u.total_tokens})`;
+}
+
+async function loadTokenStats() {
+  try {
+    const res = await fetch('/api/token_stats');
+    const ts = await res.json();
+    const w = ts.window_24h || {};
+    const t24 = document.getElementById('s-tokens-24h');
+    const cov = document.getElementById('s-tokens-cover');
+    const tot = fmtTokens(w.total_tokens);
+    t24.textContent = tot === null ? '—' : tot;
+    t24.title = `in ${fmtTokens(w.input_tokens) ?? '—'} / out ${fmtTokens(w.output_tokens) ?? '—'} · всё время: ${fmtTokens((ts.all_time || {}).total_tokens) ?? '—'}`;
+    cov.textContent = w.coverage_pct === null || w.coverage_pct === undefined ? '—' : w.coverage_pct + '%';
+    cov.title = `24ч: известно ${w.known ?? 0} из ${w.checks ?? 0} проверок (неизвестно ${w.unknown ?? 0})`;
+  } catch (e) { /* статы токенов не критичны */ }
+}
+
 function updateStats(data) {
   document.getElementById('s-total').textContent = data.length;
   document.getElementById('s-ranked').textContent = data.filter(m => m.arena_text_score).length;
@@ -498,6 +532,7 @@ async function loadData() {
   document.getElementById('last-updated').textContent = `Обновлено: ${now.toLocaleTimeString('ru')}`;
 
   applyFilters();
+  loadTokenStats();
 }
 
 async function loadProviders() {
@@ -527,6 +562,7 @@ async function checkOne(modelId, btn) {
       if (data.uptime_pct !== undefined) allData[idx].uptime_pct = data.uptime_pct;
     }
     applyFilters();
+    loadTokenStats();
   } finally {
     btn.textContent = '▶ Тест';
     btn.classList.remove('checking');
@@ -562,15 +598,17 @@ async function openCheckHistory(modelId, btn) {
   if (!checks.length) {
     body.innerHTML = '<div class="no-data" style="text-align:center;padding:20px;">Нет данных о проверках</div>';
   } else {
-    body.innerHTML = checks.map(c => `
+    body.innerHTML = checks.map(c => {
+      const toks = usageStr(c);
+      return `
       <div class="check-log-item">
         <div class="log-line1">
           <span class="log-status ${c.ok ? 'ok' : 'fail'}">${c.ok ? '✓ Успех' : '✗ Провал'}</span>
-          <span class="log-meta">${fmtTime(c.checked_at)} · ${c.latency_ms !== null && c.latency_ms !== undefined ? c.latency_ms + ' мс' : '—'}</span>
+          <span class="log-meta">${fmtTime(c.checked_at)} · ${c.latency_ms !== null && c.latency_ms !== undefined ? c.latency_ms + ' мс' : '—'}${toks ? ` · ${escapeHtml(toks)} т.` : ''}</span>
         </div>
         ${c.error ? `<div class="log-err">${escapeHtml(c.error)}</div>` : ''}
-      </div>
+      </div>`;
-    `).join('');
+    }).join('');
   }
 
   modal.classList.add('open');

app/usage_utils.py

@@ -0,0 +1,83 @@
+"""
+Общий модуль извлечения token usage из ответа OmniRoute.
+
+Используется и cron-пробой (probe.py), и ручной проверкой (main.py).
+Правила:
+  * usage извлекается ДО валидации контента — чтобы провал content-парсинга
+    не терял реально потраченные токены;
+  * если провайдер не отдал usage (или отдал полностью нулевой) — все поля
+    остаются NULL: неизвестное потребление не подменяется нулями;
+  * usage_json хранит сырой словарь usage как пришел от провайдера.
+"""
+from __future__ import annotations
+
+import json
+from typing import Optional
+
+# Возможные имена полей в разных диалектах (OpenAI / Anthropic / camelCase).
+_INPUT_KEYS = ("prompt_tokens", "input_tokens", "promptTokens", "inputTokens")
+_OUTPUT_KEYS = ("completion_tokens", "output_tokens", "completionTokens", "outputTokens")
+_TOTAL_KEYS = ("total_tokens", "totalTokens")
+
+
+def _first_int(usage: dict, keys: tuple[str, ...]) -> Optional[int]:
+    """Первое числовое значение из ключей; не-числа и None игнорируются."""
+    for k in keys:
+        v = usage.get(k)
+        if isinstance(v, bool):  # bool — подкласс int, отсекаем явно
+            continue
+        if isinstance(v, (int, float)) and v >= 0:
+            return int(v)
+    return None
+
+
+def extract_usage(resp) -> dict:
+    """Достать usage из тела ответа chat/completions.
+
+    Возвращает dict с ключами input_tokens / output_tokens / total_tokens /
+    usage_json. Отсутствующий или вырожденный (полностью нулевой) usage дает
+    None во всех полях — NULL, а не нулевая оценка.
+    """
+    empty = {"input_tokens": None, "output_tokens": None,
+             "total_tokens": None, "usage_json": None}
+
+    usage = None
+    if isinstance(resp, dict):
+        u = resp.get("usage")
+        if isinstance(u, dict):
+            usage = u
+        elif isinstance(u, str):
+            # редко: usage пришел строкой-JSON
+            try:
+                parsed = json.loads(u)
+                if isinstance(parsed, dict):
+                    usage = parsed
+            except (ValueError, TypeError):
+                usage = None
+    if not isinstance(usage, dict):
+        return empty
+
+    inp = _first_int(usage, _INPUT_KEYS)
+    out = _first_int(usage, _OUTPUT_KEYS)
+    tot = _first_int(usage, _TOTAL_KEYS)
+
+    # Полностью нулевой usage (0/0/0) трактуем как «не отдал» — провайдеры
+    # так отвечают в ошибочных путях; реальные токены всегда > 0.
+    if not any(v for v in (inp, out, tot) if v is not None):
+        return empty
+
+    # total — точная арифметика по отчетным значениям, не оценка.
+    if tot is None and inp is not None and out is not None:
+        tot = inp + out
+
+    try:
+        usage_json = json.dumps(usage, ensure_ascii=False, sort_keys=True)
+    except (TypeError, ValueError):
+        usage_json = None
+
+    return {
+        "input_tokens": inp,
+        "output_tokens": out,
+        "total_tokens": tot,
+        "usage_json": usage_json,
+    }

tests/conftest.py

@@ -0,0 +1,180 @@
+"""Общие фикстуры: реальный мок-сервер OmniRoute на localhost + чистые модули."""
+from __future__ import annotations
+
+import json
+import os
+import sys
+import threading
+import time
+from http.server import BaseHTTPRequestHandler, ThreadingHTTPServer
+from types import SimpleNamespace
+
+import pytest
+from fastapi.testclient import TestClient
+
+REPO = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
+APP_DIR = os.path.join(REPO, "app")
+if APP_DIR not in sys.path:
+    sys.path.insert(0, APP_DIR)
+
+MODULES = ("main", "probe", "usage_utils")
+SAVED_ENV = {}
+
+
+class OmniMock:
+    """Управляемый мок OmniRoute: реальные HTTP-запросы на 127.0.0.1."""
+
+    def __init__(self):
+        self.state = {
+            # поведение /v1/chat/completions по умолчанию; переопределяется
+            # per-model: status, usage, no_usage, omit_choices, content, delay
+            "completions": {"status": 200, "usage": None, "content": "4"},
+            "models": {"data": [
+                {"id": "zai/test-model-a"},
+                {"id": "agy/test-model-b"},
+            ]},
+            "global_delay": 0.0,
+            "per_model": {},
+        }
+        self.requests = []
+        self.lock = threading.Lock()
+
+        outer = self
+
+        class Handler(BaseHTTPRequestHandler):
+            def log_message(self, *a):
+                pass
+
+            def _send_json(self, status, payload):
+                data = json.dumps(payload).encode()
+                try:
+                    self.send_response(status)
+                    self.send_header("Content-Type", "application/json")
+                    self.send_header("Content-Length", str(len(data)))
+                    self.end_headers()
+                    self.wfile.write(data)
+                except (BrokenPipeError, ConnectionResetError):
+                    pass  # клиент отвалился по таймауту — это ожидаемо
+
+            def _read_body(self):
+                n = int(self.headers.get("Content-Length") or 0)
+                raw = self.rfile.read(n) if n else b""
+                try:
+                    return json.loads(raw)
+                except ValueError:
+                    return {}
+
+            def do_POST(self):
+                body = self._read_body()
+                with outer.lock:
+                    outer.requests.append({"path": self.path, "body": body})
+                    beh = dict(outer.state["completions"])
+                    beh.update(outer.state["per_model"].get(body.get("model"), {}))
+                    delay = beh.get("delay", 0.0) or outer.state["global_delay"]
+                if delay:
+                    time.sleep(delay)
+                if beh.get("status", 200) != 200:
+                    self._send_json(beh["status"], {"error": {"message": "mock error"}})
+                    return
+                resp = {"id": "chatcmpl-mock", "object": "chat.completion",
+                        "created": int(time.time()), "model": body.get("model", "?")}
+                if not beh.get("no_usage"):
+                    resp["usage"] = beh.get("usage")
+                if not beh.get("omit_choices"):
+                    resp["choices"] = [{"index": 0, "finish_reason": "stop",
+                                        "message": {"role": "assistant",
+                                                    "content": beh.get("content", "4")}}]
+                self._send_json(200, resp)
+
+            def do_GET(self):
+                if self.path == "/v1/models":
+                    with outer.lock:
+                        payload = dict(outer.state["models"])
+                    self._send_json(200, payload)
+                    return
+                self._send_json(404, {"detail": "not found"})
+
+        self.server = ThreadingHTTPServer(("127.0.0.1", 0), Handler)
+        self.thread = threading.Thread(target=self.server.serve_forever, daemon=True)
+        self.thread.start()
+        self.url = f"http://127.0.0.1:{self.server.server_address[1]}"
+
+    def set_completions(self, **kw):
+        with self.lock:
+            self.state["completions"].update(kw)
+
+    def set_per_model(self, model_id, beh):
+        with self.lock:
+            self.state["per_model"][model_id] = beh
+
+    def close(self):
+        self.server.shutdown()
+        self.server.server_close()
+        self.thread.join(timeout=5)
+
+
+@pytest.fixture(scope="session")
+def omni_server():
+    mock = OmniMock()
+    yield mock
+    mock.close()
+
+
+@pytest.fixture(autouse=True)
+def _reset_omni(omni_server):
+    """Изолировать тесты: состояние мока не должно протекать между тестами."""
+    with omni_server.lock:
+        omni_server.state["per_model"] = {}
+        omni_server.state["global_delay"] = 0.0
+        omni_server.state["completions"] = {"status": 200, "usage": None, "content": "4"}
+        omni_server.requests.clear()
+    yield
+
+
+def _set_env(db_path: str, omni_url: str):
+    env = {
+        "DB_PATH": db_path,
+        "OMNIROUTE_URL": omni_url,
+        "OMNIROUTE_API_KEY": "test-key",
+        "CHECK_TIMEOUT": "2",
+        "CHECK_CONCURRENCY": "4",
+        "ARENA_CACHE_TEXT": "/nonexistent/arena_text.json",
+        "ARENA_CACHE_AGENT": "/nonexistent/arena_agent.json",
+    }
+    for k, v in env.items():
+        SAVED_ENV[k] = os.environ.get(k)
+        os.environ[k] = v
+
+
+def restore_env():
+    for k, v in SAVED_ENV.items():
+        if v is None:
+            os.environ.pop(k, None)
+        else:
+            os.environ[k] = v
+
+
+@pytest.fixture
+def fresh_main(tmp_path, omni_server):
+    """Модуль main с env, привязанным к tmp-БД и мок-серверу (без запуска app).
+
+    init_db НЕ вызывается — тест миграции сам решает, что лежит в БД до него.
+    """
+    db_path = str(tmp_path / "arena.db")
+    _set_env(db_path, omni_server.url)
+    for mod in MODULES:
+        sys.modules.pop(mod, None)
+    import main as m
+    try:
+        yield m
+    finally:
+        restore_env()
+
+
+@pytest.fixture
+def core(fresh_main):
+    """Приложение с запущенным lifespan (init_db + sync_models) + модуль probe."""
+    import probe as probe_mod
+    with TestClient(fresh_main.app) as client:
+        yield SimpleNamespace(main=fresh_main, probe=probe_mod,
+                              client=client, db_path=fresh_main.DB_PATH)

tests/smoke_e2e.py

@@ -0,0 +1,69 @@
+"""Реальный smoke: uvicorn + мок OmniRoute, проверка живых HTTP-запросов."""
+import json
+import os
+import signal
+import subprocess
+import sys
+import time
+import urllib.request
+
+sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))
+from conftest import OmniMock
+
+REPO = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
+TMP = os.path.join(REPO, ".smoke")
+os.makedirs(TMP, exist_ok=True)
+
+mock = OmniMock()
+mock.set_completions(usage={"prompt_tokens": 9, "completion_tokens": 4, "total_tokens": 13})
+
+env = dict(os.environ)
+env.update({
+    "DB_PATH": os.path.join(TMP, "arena.db"),
+    "OMNIROUTE_URL": mock.url,
+    "OMNIROUTE_API_KEY": "smoke",
+    "ARENA_CACHE_TEXT": "/nonexistent/a.json",
+    "ARENA_CACHE_AGENT": "/nonexistent/b.json",
+})
+
+proc = subprocess.Popen(
+    [os.path.join(REPO, ".venv/bin/uvicorn"), "main:app", "--host", "127.0.0.1", "--port", "18700"],
+    cwd=os.path.join(REPO, "app"), env=env,
+    stdout=subprocess.PIPE, stderr=subprocess.STDOUT, text=True,
+)
+
+# pip install uvicorn into venv if missing handled outside
+try:
+    base = "http://127.0.0.1:18700"
+    for _ in range(50):
+        try:
+            urllib.request.urlopen(f"{base}/api/providers", timeout=1)
+            break
+        except Exception:
+            time.sleep(0.2)
+
+    # 1. UI отдаётся (статика через __file__-путь)
+    ui = urllib.request.urlopen(f"{base}/", timeout=5).read().decode()
+    assert "Arena Dashboard" in ui and "s-tokens-24h" in ui and "loadTokenStats" in ui, "UI не содержит token-элементов"
+
+    # 2. Ручная проверка с usage
+    chk = json.loads(urllib.request.urlopen(f"{base}/api/check/zai/test-model-a", timeout=10).read())
+    assert chk["input_tokens"] == 9 and chk["total_tokens"] == 13, chk
+
+    # 3. Агрегат
+    ts = json.loads(urllib.request.urlopen(f"{base}/api/token_stats", timeout=5).read())
+    assert ts["window_24h"]["known"] >= 1 and ts["window_24h"]["total_tokens"] >= 13, ts
+    print("SMOKE_OK", json.dumps({
+        "check_usage": [chk["input_tokens"], chk["output_tokens"], chk["total_tokens"]],
+        "token_stats_24h": ts["window_24h"],
+        "providers": [(p["provider"], p["window_24h"]["total_tokens"]) for p in ts["providers"]],
+    }, ensure_ascii=False))
+finally:
+    proc.send_signal(signal.SIGTERM)
+    try:
+        out, _ = proc.communicate(timeout=10)
+        print("--- uvicorn log tail ---")
+        print("\n".join(out.splitlines()[-8:]))
+    except subprocess.TimeoutExpired:
+        proc.kill()
+    mock.close()

tests/test_api.py

@@ -0,0 +1,171 @@
+"""Ручной путь (main.py /api/check) + API-агрегаты token usage."""
+import json
+import sqlite3
+from datetime import datetime, timedelta, timezone
+
+
+def _con(db_path):
+    con = sqlite3.connect(db_path)
+    con.row_factory = sqlite3.Row
+    return con
+
+
+def _insert_check(db_path, model_id, at, ok=1, inp=None, out=None, tot=None, ujson=None):
+    con = _con(db_path)
+    con.execute(
+        "INSERT INTO checks (model_id, checked_at, ok, latency_ms, error, "
+        "input_tokens, output_tokens, total_tokens, usage_json) VALUES (?,?,?,?,?,?,?,?,?)",
+        (model_id, at, ok, 50, None, inp, out, tot, ujson),
+    )
+    con.commit()
+    con.close()
+
+
+def _iso(**kw):
+    return (datetime.now(timezone.utc) + timedelta(**kw)).isoformat()
+
+
+# ─── ручная проверка /api/check ───────────────────────────────────────────────
+
+def test_api_check_captures_usage(core, omni_server):
+    omni_server.set_completions(usage={"prompt_tokens": 12, "completion_tokens": 3, "total_tokens": 15})
+    r = core.client.get("/api/check/zai/test-model-a")
+    assert r.status_code == 200
+    body = r.json()
+    assert body["ok"] is True
+    assert body["input_tokens"] == 12
+    assert body["output_tokens"] == 3
+    assert body["total_tokens"] == 15
+
+    row = _con(core.db_path).execute(
+        "SELECT * FROM checks WHERE model_id='zai/test-model-a'").fetchone()
+    assert row["input_tokens"] == 12 and row["total_tokens"] == 15
+    assert json.loads(row["usage_json"])["prompt_tokens"] == 12
+
+
+def test_api_check_missing_usage_null(core, omni_server):
+    omni_server.set_completions(no_usage=True)
+    body = core.client.get("/api/check/zai/test-model-a").json()
+    assert body["ok"] is True
+    assert body["input_tokens"] is None and body["total_tokens"] is None
+
+    row = _con(core.db_path).execute(
+        "SELECT input_tokens, total_tokens, usage_json FROM checks WHERE model_id='zai/test-model-a'").fetchone()
+    assert row["input_tokens"] is None and row["total_tokens"] is None and row["usage_json"] is None
+
+
+def test_api_check_failed_content_preserves_usage(core, omni_server):
+    omni_server.set_completions(usage={"prompt_tokens": 8, "completion_tokens": 1, "total_tokens": 9},
+                                omit_choices=True)
+    body = core.client.get("/api/check/agy/test-model-b").json()
+    assert body["ok"] is False
+    assert body["input_tokens"] == 8  # usage пережил ошибку контента
+
+    row = _con(core.db_path).execute(
+        "SELECT ok, error, total_tokens FROM checks WHERE model_id='agy/test-model-b'").fetchone()
+    assert row["ok"] == 0 and row["error"]
+    assert row["total_tokens"] == 9
+
+
+def test_api_check_no_response_stores_null(core, omni_server):
+    # недоступный OmniRoute: ответа нет — usage остаётся NULL (не нулями)
+    saved = core.main.OMNIROUTE_URL
+    core.main.OMNIROUTE_URL = "http://127.0.0.1:1"  # connection refused, мгновенно
+    try:
+        body = core.client.get("/api/check/zai/test-model-a").json()
+    finally:
+        core.main.OMNIROUTE_URL = saved
+    assert body["ok"] is False
+    assert body["input_tokens"] is None and body["total_tokens"] is None
+
+    row = _con(core.db_path).execute(
+        "SELECT ok, error, total_tokens, usage_json FROM checks WHERE model_id='zai/test-model-a'").fetchone()
+    assert row["ok"] == 0 and row["error"]
+    assert row["total_tokens"] is None and row["usage_json"] is None
+
+
+# ─── истории: usage в /api/checks и /api/models.history ──────────────────────
+
+def test_api_checks_history_includes_usage(core, omni_server):
+    omni_server.set_completions(usage={"prompt_tokens": 5, "completion_tokens": 2, "total_tokens": 7})
+    core.client.get("/api/check/zai/test-model-a")
+    omni_server.set_completions(no_usage=True)
+    core.client.get("/api/check/zai/test-model-a")
+
+    data = core.client.get("/api/checks/zai/test-model-a").json()
+    assert len(data["checks"]) == 2
+    with_usage = [c for c in data["checks"] if c["total_tokens"] == 7]
+    without = [c for c in data["checks"] if c["total_tokens"] is None]
+    assert len(with_usage) == 1 and with_usage[0]["input_tokens"] == 5
+    assert with_usage[0]["usage_json"]
+    assert len(without) == 1 and without[0]["input_tokens"] is None
+
+    # сегментная история тоже несёт токены (для тултипов UI)
+    models = core.client.get("/api/models").json()
+    model = next(m for m in models if m["id"] == "zai/test-model-a")
+    toks = [h["total_tokens"] for h in model["history"]]
+    assert set(toks) == {7, None}
+
+
+# ─── /api/token_stats ─────────────────────────────────────────────────────────
+
+def test_token_stats_empty_db(core):
+    ts = core.client.get("/api/token_stats").json()
+    assert ts["window_24h"]["checks"] == 0
+    assert ts["window_24h"]["coverage_pct"] is None
+    assert ts["all_time"]["checks"] == 0
+    assert ts["providers"] == []
+
+
+def test_token_stats_windows_coverage_providers(core):
+    # zai: 2 свежих с usage + 1 старая (>24h) с usage
+    _insert_check(core.db_path, "zai/test-model-a", _iso(minutes=-10), inp=10, out=5, tot=15)
+    _insert_check(core.db_path, "zai/test-model-a", _iso(minutes=-20), inp=20, out=2, tot=22)
+    _insert_check(core.db_path, "zai/test-model-a", _iso(hours=-30), inp=100, out=50, tot=150)
+    # agy: 1 свежая без usage + 1 старая с usage
+    _insert_check(core.db_path, "agy/test-model-b", _iso(minutes=-5))
+    _insert_check(core.db_path, "agy/test-model-b", _iso(hours=-40), inp=7, out=3, tot=10)
+    # модель, которой нет в таблице models: провайдер выводится из model_id
+    _insert_check(core.db_path, "orphan/x", _iso(minutes=-1), inp=5, out=5, tot=10)
+
+    ts = core.client.get("/api/token_stats").json()
+
+    w = ts["window_24h"]
+    assert w["checks"] == 4 and w["known"] == 3 and w["unknown"] == 1
+    assert w["coverage_pct"] == 75
+    assert (w["input_tokens"], w["output_tokens"], w["total_tokens"]) == (35, 12, 47)
+
+    a = ts["all_time"]
+    assert a["checks"] == 6 and a["known"] == 5 and a["unknown"] == 1
+    assert (a["input_tokens"], a["output_tokens"], a["total_tokens"]) == (142, 65, 207)
+
+    provs = {p["provider"]: p for p in ts["providers"]}
+    assert set(provs) == {"zai", "agy", "orphan"}
+    assert provs["zai"]["window_24h"]["total_tokens"] == 37
+    assert provs["zai"]["all_time"]["total_tokens"] == 187
+    assert provs["agy"]["window_24h"]["known"] == 0
+    assert provs["agy"]["window_24h"]["unknown"] == 1
+    assert provs["agy"]["window_24h"]["total_tokens"] is None  # сумма по NULL-группе
+    assert provs["agy"]["all_time"]["total_tokens"] == 10
+    assert provs["orphan"]["window_24h"]["total_tokens"] == 10  # fallback из model_id
+
+
+def test_token_stats_counts_both_paths(core, omni_server):
+    """probe.run() и ручной /api/check пишут usage, агрегат считает оба пути."""
+    omni_server.set_completions(usage={"prompt_tokens": 10, "completion_tokens": 5, "total_tokens": 15})
+    core.probe.run()  # путь cron: zai 15 + agy 15
+    core.client.get("/api/check/zai/test-model-a")  # путь ручной: +15
+    omni_server.set_completions(no_usage=True)
+    core.client.get("/api/check/agy/test-model-b")  # путь ручной без usage: NULL
+
+    ts = core.client.get("/api/token_stats").json()
+    w = ts["window_24h"]
+    assert w["checks"] == 4          # 2 probe + 2 ручных
+    assert w["known"] == 3           # probe×2 + ручной zai
+    assert w["unknown"] == 1         # ручной agy без usage
+    assert w["total_tokens"] == 45   # 15 × 3 известных
+    assert w["input_tokens"] == 30 and w["output_tokens"] == 15
+
+    provs = {p["provider"]: p for p in ts["providers"]}
+    assert provs["zai"]["window_24h"]["total_tokens"] == 30      # probe + manual
+    assert provs["agy"]["window_24h"]["total_tokens"] == 15       # только probe

tests/test_migration.py

@@ -0,0 +1,77 @@
+"""Миграция схемы checks: идемпотентность и сохранение старых NULL."""
+import sqlite3
+
+
+def _cols(db_path):
+    con = sqlite3.connect(db_path)
+    cols = [r[1] for r in con.execute("PRAGMA table_info(checks)").fetchall()]
+    con.close()
+    return cols
+
+
+def _make_old_db(db_path, with_error_col=False):
+    """БД со старой схемой checks (до учёта usage) и живыми данными."""
+    con = sqlite3.connect(db_path)
+    con.execute("""CREATE TABLE checks (
+        id INTEGER PRIMARY KEY AUTOINCREMENT,
+        model_id TEXT NOT NULL,
+        checked_at TEXT NOT NULL,
+        ok INTEGER NOT NULL,
+        latency_ms INTEGER)""")
+    if with_error_col:
+        con.execute("ALTER TABLE checks ADD COLUMN error TEXT")
+    con.execute(
+        "INSERT INTO checks (model_id, checked_at, ok, latency_ms, error) VALUES (?,?,?,?,?)"
+        if with_error_col else
+        "INSERT INTO checks (model_id, checked_at, ok, latency_ms) VALUES (?,?,?,?)",
+        ("zai/old-model", "2026-01-01T00:00:00+00:00", 1, 123, None)
+        if with_error_col else ("zai/old-model", "2026-01-01T00:00:00+00:00", 1, 123),
+    )
+    con.commit()
+    con.close()
+
+
+def test_migration_old_schema_preserves_null(fresh_main):
+    _make_old_db(fresh_main.DB_PATH, with_error_col=True)
+    fresh_main.init_db()
+
+    cols = _cols(fresh_main.DB_PATH)
+    for c in ("input_tokens", "output_tokens", "total_tokens", "usage_json"):
+        assert c in cols, f"колонка {c} не добавлена"
+
+    con = sqlite3.connect(fresh_main.DB_PATH)
+    con.row_factory = sqlite3.Row
+    row = con.execute("SELECT * FROM checks WHERE model_id='zai/old-model'").fetchone()
+    con.close()
+    assert row["ok"] == 1 and row["latency_ms"] == 123
+    # старые данные: неизвестное потребление = NULL, не нули
+    assert row["input_tokens"] is None
+    assert row["output_tokens"] is None
+    assert row["total_tokens"] is None
+    assert row["usage_json"] is None
+
+
+def test_migration_from_ancient_schema_without_error(fresh_main):
+    _make_old_db(fresh_main.DB_PATH, with_error_col=False)
+    fresh_main.init_db()  # не должен падать: добавит и error, и usage-колонки
+    cols = _cols(fresh_main.DB_PATH)
+    assert "error" in cols
+    assert "usage_json" in cols
+
+
+def test_migration_idempotent(fresh_main):
+    _make_old_db(fresh_main.DB_PATH, with_error_col=True)
+    fresh_main.init_db()
+    fresh_main.init_db()  # повторный запуск — no-op без ошибок
+
+    con = sqlite3.connect(fresh_main.DB_PATH)
+    n = con.execute("SELECT COUNT(*) FROM checks").fetchone()[0]
+    con.close()
+    assert n == 1  # данные не задвоились и не потерялись
+
+
+def test_fresh_db_has_usage_columns(fresh_main):
+    fresh_main.init_db()
+    cols = _cols(fresh_main.DB_PATH)
+    for c in ("input_tokens", "output_tokens", "total_tokens", "usage_json"):
+        assert c in cols

tests/test_probe.py

@@ -0,0 +1,122 @@
+"""Путь cron-пробы (probe.py): захват usage, NULL при неизвестном, сохранение при ошибках."""
+import json
+import sqlite3
+
+
+def _rows(db_path):
+    con = sqlite3.connect(db_path)
+    con.row_factory = sqlite3.Row
+    rows = con.execute("SELECT * FROM checks ORDER BY id").fetchall()
+    out = [dict(r) for r in rows]
+    con.close()
+    return out
+
+
+def test_probe_captures_usage(core, omni_server):
+    omni_server.set_completions(usage={"prompt_tokens": 11, "completion_tokens": 4, "total_tokens": 15},
+                                no_usage=False)
+    ok, lat, err, usage = core.probe.check_model("zai/test-model-a")
+    assert ok is True and err is None
+    assert usage["input_tokens"] == 11
+    assert usage["output_tokens"] == 4
+    assert usage["total_tokens"] == 15
+    assert json.loads(usage["usage_json"])["prompt_tokens"] == 11
+
+
+def test_probe_missing_usage_stays_null(core, omni_server):
+    omni_server.set_completions(no_usage=True)
+    ok, lat, err, usage = core.probe.check_model("zai/test-model-a")
+    assert ok is True  # контент есть — проверка успешна
+    # но usage неизвестен: NULL, не нули
+    assert usage == {"input_tokens": None, "output_tokens": None,
+                     "total_tokens": None, "usage_json": None}
+
+
+def test_probe_zero_usage_treated_as_unknown(core, omni_server):
+    omni_server.set_completions(usage={"prompt_tokens": 0, "completion_tokens": 0, "total_tokens": 0},
+                                no_usage=False)
+    _, _, _, usage = core.probe.check_model("agy/test-model-b")
+    assert usage["total_tokens"] is None and usage["input_tokens"] is None
+
+
+def test_probe_failed_content_preserves_usage(core, omni_server):
+    # ответ без choices: парсинг контента упадёт, но usage уже зафиксирован
+    omni_server.set_completions(usage={"prompt_tokens": 9, "completion_tokens": 2, "total_tokens": 11},
+                                omit_choices=True)
+    ok, lat, err, usage = core.probe.check_model("zai/test-model-a")
+    assert ok is False
+    assert err and "KeyError" in err
+    assert usage["input_tokens"] == 9  # токены не потеряны
+    assert usage["total_tokens"] == 11
+
+
+def test_probe_timeout_usage_null(core, omni_server):
+    omni_server.set_completions(delay=6)  # CHECK_TIMEOUT=2 в тестовом env
+    ok, lat, err, usage = core.probe.check_model("zai/test-model-a")
+    assert ok is False
+    assert err and ("timed out" in err.lower() or "timeout" in err.lower())
+    assert usage["total_tokens"] is None
+    assert usage["usage_json"] is None
+    assert lat and lat >= 2000  # latency фиксируется всегда
+    omni_server.set_completions(delay=0)
+
+
+def test_probe_http_error_usage_null(core, omni_server):
+    omni_server.set_completions(status=500)
+    ok, lat, err, usage = core.probe.check_model("agy/test-model-b")
+    assert ok is False and err
+    assert usage["input_tokens"] is None
+
+
+def test_probe_run_end_to_end_both_models(core, omni_server):
+    """Полный прогон run(): разные модели — разный исход по usage."""
+    omni_server.set_completions(usage={"prompt_tokens": 10, "completion_tokens": 5, "total_tokens": 15})
+    omni_server.set_per_model("agy/test-model-b", {"no_usage": True})
+    omni_server.set_per_model("zai/test-model-a",
+                              {"usage": {"prompt_tokens": 10, "completion_tokens": 5, "total_tokens": 15}})
+
+    core.probe.run()
+
+    rows = {r["model_id"]: r for r in _rows(core.db_path)}
+    assert set(rows) == {"zai/test-model-a", "agy/test-model-b"}
+    a = rows["zai/test-model-a"]
+    assert a["ok"] == 1 and (a["input_tokens"], a["output_tokens"], a["total_tokens"]) == (10, 5, 15)
+    assert json.loads(a["usage_json"])["total_tokens"] == 15
+    b = rows["agy/test-model-b"]
+    assert b["ok"] == 1  # без usage проверка всё равно успешна
+    assert b["input_tokens"] is None and b["total_tokens"] is None and b["usage_json"] is None
+
+
+def test_probe_run_migrates_old_schema_standalone(tmp_path, omni_server):
+    """Cron стартует раньше uvicorn: probe должен сам уметь мигрировать старую БД."""
+    from conftest import _set_env, restore_env, MODULES
+    import sys as _sys
+
+    db = str(tmp_path / "probe-only.db")
+    con = sqlite3.connect(db)
+    con.execute("CREATE TABLE checks (id INTEGER PRIMARY KEY AUTOINCREMENT, model_id TEXT NOT NULL, "
+                "checked_at TEXT NOT NULL, ok INTEGER NOT NULL, latency_ms INTEGER, error TEXT)")
+    con.execute("CREATE TABLE models (id TEXT PRIMARY KEY, provider TEXT NOT NULL, name TEXT NOT NULL, "
+                "raw_base TEXT NOT NULL, norm_base TEXT NOT NULL, clean_base TEXT NOT NULL, "
+                "text_model_match TEXT, agent_model_match TEXT, updated_at TEXT)")
+    con.execute("INSERT INTO models (id, provider, name, raw_base, norm_base, clean_base) "
+                "VALUES ('zai/test-model-a', 'zai', 'zai/test-model-a', 'test-model-a', 'testmodela', 'testmodela')")
+    con.commit()
+    con.close()
+
+    _set_env(db, omni_server.url)
+    try:
+        for m in MODULES:
+            _sys.modules.pop(m, None)
+        import probe as standalone_probe
+        omni_server.set_per_model("zai/test-model-a",
+                                  {"usage": {"input_tokens": 6, "output_tokens": 2}})
+        standalone_probe.run()
+    finally:
+        restore_env()
+
+    rows = _rows(db)
+    assert len(rows) == 1
+    assert rows[0]["input_tokens"] == 6
+    assert rows[0]["output_tokens"] == 2
+    assert rows[0]["total_tokens"] == 8  # вычислен из отчётных значений

tests/test_usage_utils.py

@@ -0,0 +1,70 @@
+"""extract_usage: разбор форматов usage разных провайдеров."""
+import sys
+import os
+
+sys.path.insert(0, os.path.join(os.path.dirname(os.path.dirname(os.path.abspath(__file__))), "app"))
+
+from usage_utils import extract_usage
+
+EMPTY = {"input_tokens": None, "output_tokens": None, "total_tokens": None, "usage_json": None}
+
+
+def test_openai_shape():
+    u = extract_usage({"choices": [], "usage": {"prompt_tokens": 10, "completion_tokens": 5, "total_tokens": 15}})
+    assert u["input_tokens"] == 10
+    assert u["output_tokens"] == 5
+    assert u["total_tokens"] == 15
+    assert u["usage_json"] == '{"completion_tokens": 5, "prompt_tokens": 10, "total_tokens": 15}'
+
+
+def test_anthropic_shape_total_computed():
+    u = extract_usage({"usage": {"input_tokens": 7, "output_tokens": 3}})
+    assert u["input_tokens"] == 7
+    assert u["output_tokens"] == 3
+    assert u["total_tokens"] == 10  # точная сумма отчётных значений, не оценка
+
+
+def test_camel_case_shape():
+    u = extract_usage({"usage": {"promptTokens": 4, "completionTokens": 2, "totalTokens": 6}})
+    assert (u["input_tokens"], u["output_tokens"], u["total_tokens"]) == (4, 2, 6)
+
+
+def test_missing_usage_all_none():
+    assert extract_usage({"choices": []}) == EMPTY
+    assert extract_usage({"usage": None}) == EMPTY
+    assert extract_usage({}) == EMPTY
+
+
+def test_non_dict_resp():
+    assert extract_usage(None) == EMPTY
+    assert extract_usage("error") == EMPTY
+
+
+def test_zero_usage_treated_as_unknown():
+    # полностью нулевой usage = провайдер не отдал данные → NULL, не нули
+    u = extract_usage({"usage": {"prompt_tokens": 0, "completion_tokens": 0, "total_tokens": 0}})
+    assert u == EMPTY
+
+
+def test_string_encoded_usage():
+    u = extract_usage({"usage": '{"prompt_tokens": 8, "completion_tokens": 1, "total_tokens": 9}'})
+    assert (u["input_tokens"], u["output_tokens"], u["total_tokens"]) == (8, 1, 9)
+
+
+def test_partial_total_only():
+    u = extract_usage({"usage": {"total_tokens": 42}})
+    assert u["input_tokens"] is None
+    assert u["output_tokens"] is None
+    assert u["total_tokens"] == 42
+
+
+def test_garbage_values_ignored():
+    u = extract_usage({"usage": {"prompt_tokens": "много", "completion_tokens": True, "total_tokens": None}})
+    assert u == EMPTY
+
+
+def test_usage_json_roundtrip():
+    raw = {"prompt_tokens": 10, "completion_tokens": 5, "total_tokens": 15, "extra": {"x": 1}}
+    u = extract_usage({"usage": raw})
+    import json
+    assert json.loads(u["usage_json"]) == raw