model usage: Requests counts only successful calls

The Requests column summed every spend-log row (COUNT(*)), so a model's
call count included refused/failed attempts — e.g. Metis showed 26 calls
when only 16 succeeded (the other 10 were 401s on /v1/systemone).

Count only rows whose status is not 'failure', and report the refused
count separately as 'refused' (per model and in totals) so a failure
cannot inflate use without disappearing from view. spend and tokens stay
summed over all rows: a call that failed after consuming tokens still
cost money.

Fixes the panel and dashboard model tables, which both read this
aggregation (panel via /admin/model-usage, dashboard via
/broker/model-leaderboard).

app/main.py

@@ -402,6 +402,13 @@ def aggregate_model_usage(days: int = 0, team_id: str | None = None) -> dict:
     Prefers the `model_group` column, which carries our provider/model alias
     (e.g. greenpt/glm-5.3-flash); falls back to `model` for pre-alias rows — matching
     the logic the HTTP path used, expressed in SQL.
+
+    `requests` counts only metered calls that SUCCEEDED. Refused/failed attempts
+    (rows with status = 'failure', e.g. a 401 on the /v1/systemone decision
+    route) are excluded from that count and reported separately as `refused`,
+    so a failure cannot inflate a model's call count. `spend` and `tokens` stay
+    summed over every row: a call that failed after tokens were consumed still
+    cost money and still processed those tokens.
     """
     if not litellm_db_ready():
         raise RuntimeError("LiteLLM DB credentials not configured")
@@ -425,7 +432,8 @@ def aggregate_model_usage(days: int = 0, team_id: str | None = None) -> dict:
             COALESCE(NULLIF(model_group, ''), NULLIF(model, ''), 'unknown') AS model,
             COALESCE(SUM(spend), 0) AS spend,
             COALESCE(SUM(total_tokens), 0) AS tokens,
-            COUNT(*) AS requests
+            COUNT(*) FILTER (WHERE status IS DISTINCT FROM 'failure') AS requests,
+            COUNT(*) FILTER (WHERE status = 'failure') AS refused
         FROM "LiteLLM_SpendLogs"
         {where_sql}
         GROUP BY 1
@@ -442,7 +450,7 @@ def aggregate_model_usage(days: int = 0, team_id: str | None = None) -> dict:
 
     models = [
         {"model": r[0], "spend": float(r[1] or 0.0), "tokens": int(r[2] or 0),
-         "requests": int(r[3] or 0)}
+         "requests": int(r[3] or 0), "refused": int(r[4] or 0)}
         for r in rows
     ]
     return {
@@ -451,6 +459,7 @@ def aggregate_model_usage(days: int = 0, team_id: str | None = None) -> dict:
             "total_spend": sum(m["spend"] for m in models),
             "total_tokens": sum(m["tokens"] for m in models),
             "total_requests": sum(m["requests"] for m in models),
+            "total_refused": sum(m["refused"] for m in models),
         },
     }
 
@@ -3345,8 +3354,10 @@ async def admin_model_usage(request: Request, days: int = 0):
     """Co-op-wide per-model usage from LiteLLM spend logs.
 
     Authenticated by the X-Admin-Token header. Aggregates spend-log rows
-    (across ALL members) by model: total spend, tokens, and request count,
-    sorted by spend descending. Optional ?days=N limits to the last N days
+    (across ALL members) by model: total spend, tokens, and the count of
+    successful requests (refused/failed attempts are excluded and reported
+    separately as `refused`), sorted by request count descending. Optional
+    ?days=N limits to the last N days
     (0 or omitted = all-time). No per-member breakdown — this is co-op-level
     so admins can see which models the membership actually uses.
 
@@ -3374,6 +3385,7 @@ async def admin_model_usage(request: Request, days: int = 0):
     total_spend = 0.0
     total_tokens = 0
     total_requests = 0
+    total_refused = 0
     try:
         async with httpx.AsyncClient(timeout=30.0) as client:
             r = await client.get(
@@ -3395,13 +3407,19 @@ async def admin_model_usage(request: Request, days: int = 0):
                     if started and started < cutoff:
                         continue
                 model = row.get("model_group") or row.get("model") or "unknown"
-                m = models.setdefault(model, {"spend": 0.0, "tokens": 0, "requests": 0})
+                m = models.setdefault(model, {"spend": 0.0, "tokens": 0, "requests": 0, "refused": 0})
                 m["spend"] += float(row.get("spend") or 0.0)
                 m["tokens"] += int(row.get("total_tokens") or 0)
-                m["requests"] += 1
+                # Refused/failed attempts are not served requests: keep them out
+                # of the request count (reported separately as `refused`).
+                if str(row.get("status") or "").lower() == "failure":
+                    m["refused"] += 1
+                    total_refused += 1
+                else:
+                    m["requests"] += 1
+                    total_requests += 1
                 total_spend += float(row.get("spend") or 0.0)
                 total_tokens += int(row.get("total_tokens") or 0)
-                total_requests += 1
     except Exception as e:
         logger.warning("admin model-usage: spend/logs fetch failed: %s", e)
 
@@ -3416,6 +3434,7 @@ async def admin_model_usage(request: Request, days: int = 0):
             "total_spend": total_spend,
             "total_tokens": total_tokens,
             "total_requests": total_requests,
+            "total_refused": total_refused,
         },
     }
 
@@ -3753,7 +3772,9 @@ async def broker_get_usage(email: str, days: int = 0) -> dict:
                 m = models.setdefault(model, {"spend": 0.0, "tokens": 0, "calls": 0})
                 m["spend"] += float(row.get("spend") or 0.0)
                 m["tokens"] += int(row.get("total_tokens") or 0)
-                m["calls"] += 1
+                # Refused/failed attempts are not served requests.
+                if str(row.get("status") or "").lower() != "failure":
+                    m["calls"] += 1
                 total_tokens += int(row.get("total_tokens") or 0)
         except Exception as e:  # spend logs are best-effort; don't fail the whole view
             logger.warning("spend/logs fetch failed: %s", e)