feat(metrics): add per-agent token tracking and cost estimation

Record total_tokens from OpenRouter usage response in MetricsCollector.
Expose in /api/v1/metrics snapshot as total_tokens and estimated_cost_usd.
Cost rate set for llama-3.3-70b; update constant when switching models.

app/core/llm.py

@@ -58,6 +58,12 @@ async def call_llm(prompt: str, system_prompt: Optional[str] = None, max_tokens:
             response.raise_for_status()
 
     data = response.json()
+    usage = data.get("usage", {})
+    total_tokens = usage.get("total_tokens", 0)
+    if total_tokens:
+        from app.core.metrics import metrics
+        metrics.record_tokens("llm_total", total_tokens)
+    return str(data["choices"][0]["message"]["content"])
     return str(data["choices"][0]["message"]["content"])
 
 async def call_llm_for_json(prompt: str, max_tokens: int = 2000) -> dict:

app/core/metrics.py

@@ -11,6 +11,7 @@ class MetricsCollector:
         self._latencies: dict[str, list[float]] = defaultdict(list)
         self._errors: dict[str, int] = defaultdict(int)
         self._started_at = time.time()
+        self._tokens: dict[str, int] = defaultdict(int)
 
     def record_run(self, agent: str, duration_seconds: float, success: bool) -> None:
         with self._lock:
@@ -19,6 +20,10 @@ class MetricsCollector:
             if not success:
                 self._errors[agent] += 1
 
+    def record_tokens(self, agent: str, tokens: int) -> None:
+        with self._lock:
+            self._tokens[agent] += tokens
+
     def snapshot(self) -> dict[str, Any]:
         with self._lock:
             agents: dict[str, Any] = {}
@@ -29,6 +34,8 @@ class MetricsCollector:
                     "errors": self._errors.get(agent, 0),
                     "avg_latency_s": round(sum(lats) / len(lats), 3) if lats else 0.0,
                     "p99_latency_s": round(sorted(lats)[int(len(lats) * 0.99)], 3) if lats else 0.0,
+                    "total_tokens": self._tokens.get(agent, 0),
+                    "estimated_cost_usd": round(self._tokens.get(agent, 0) * 0.0000009, 6),
                 }
             return {
                 "uptime_seconds": round(time.time() - self._started_at, 1),