fix: default shadow eval cost_actual to 0 so judge spend accumulates

The column was declared Float? with no default, so every row started NULL.
The verdict writer increments it, and NULL + x is NULL in Postgres, meaning
judge spend never accumulated and the UI always showed no spend.

Makes the column non-null with a default of 0 across all three schema copies
and the migration, and tightens the response model to a plain float.

litellm-proxy-extras/litellm_proxy_extras/migrations/20260807000000_add_shadow_eval_job/migration.sql

@@ -12,7 +12,7 @@ CREATE TABLE IF NOT EXISTS "LiteLLM_ShadowEvalJob" (
     "failed_count" INTEGER NOT NULL DEFAULT 0,
     "result_json" JSONB,
     "cost_estimate" DOUBLE PRECISION,
-    "cost_actual" DOUBLE PRECISION,
+    "cost_actual" DOUBLE PRECISION NOT NULL DEFAULT 0,
     "created_at" TIMESTAMP(3) NOT NULL DEFAULT CURRENT_TIMESTAMP,
     "created_by" TEXT,
     "completed_at" TIMESTAMP(3),

litellm-proxy-extras/litellm_proxy_extras/schema.prisma

@@ -1469,7 +1469,7 @@ model LiteLLM_ShadowEvalJob {
   result_json         Json?
 
   cost_estimate       Float?   // estimated cost of running the judge
-  cost_actual         Float?   // actual cost (judge calls * price)
+  cost_actual         Float    @default(0) // actual cost (judge calls * price)
 
   created_at          DateTime @default(now())
   created_by          String?  // user_id of the key that created this job

litellm/proxy/schema.prisma

@@ -1469,7 +1469,7 @@ model LiteLLM_ShadowEvalJob {
   result_json         Json?
 
   cost_estimate       Float?   // estimated cost of running the judge
-  cost_actual         Float?   // actual cost (judge calls * price)
+  cost_actual         Float    @default(0) // actual cost (judge calls * price)
 
   created_at          DateTime @default(now())
   created_by          String?  // user_id of the key that created this job

litellm/types/management_endpoints/auto_router_endpoints.py

@@ -221,6 +221,6 @@ class GetShadowEvalJobResponse(BaseModel):
         default=None, description="Present once at least one verdict has been recorded"
     )
     cost_estimate: float | None = None
-    cost_actual: float | None = Field(default=None, description="Running total of judge-call spend for this job")
+    cost_actual: float = Field(default=0.0, description="Running total of judge-call spend for this job")
     created_at: str
     completed_at: str | None = None

schema.prisma

@@ -1469,7 +1469,7 @@ model LiteLLM_ShadowEvalJob {
   result_json         Json?
 
   cost_estimate       Float?   // estimated cost of running the judge
-  cost_actual         Float?   // actual cost (judge calls * price)
+  cost_actual         Float    @default(0) // actual cost (judge calls * price)
 
   created_at          DateTime @default(now())
   created_by          String?  // user_id of the key that created this job

tests/test_litellm/integrations/test_shadow_eval_logger.py

@@ -261,6 +261,38 @@ class TestStoppedJobCannotBeReactivated:
         assert call_kwargs["data"]["status"] == "running"
 
 
+@pytest.mark.asyncio
+class TestVerdictWriteAccumulatesCost:
+    async def test_cost_actual_uses_increment_not_a_raw_set(self):
+        """Regression: cost_actual must accumulate across verdicts, not overwrite.
+
+        The DB column defaults to 0 (not NULL) precisely so this increment lands
+        on a real number instead of NULL + x = NULL. If this update ever
+        regresses to a flat `"cost_actual": judge_cost` assignment, only the
+        last verdict's cost would survive instead of the running total.
+        """
+        prisma = MagicMock()
+        prisma.db.litellm_shadowevalverdict.create = AsyncMock()
+        prisma.db.litellm_shadowevaljob.update_many = AsyncMock()
+
+        logger = ShadowEvalLogger(router_provider=lambda: MagicMock(), prisma_provider=lambda: prisma)
+        logger._call_router_shadow = AsyncMock(return_value=("shadow text", "shadow-model", "SIMPLE", 10))
+        logger._call_judge = AsyncMock(return_value=("real", 0.9, "clearer", 0.05))
+
+        job = {"id": "j1", "router_name": "r", "judge_model": "m", "shadow_percentage": 100.0, "status": "running"}
+        await logger._run_shadow_eval(
+            job=job,
+            request_id="req-1",
+            messages=[{"role": "user", "content": "hi"}],
+            response_obj={"choices": [{"message": {"content": "real text"}}]},
+            real_model="gpt-4o",
+            model_parameters={},
+        )
+
+        _, call_kwargs = prisma.db.litellm_shadowevaljob.update_many.call_args
+        assert call_kwargs["data"]["cost_actual"] == {"increment": 0.05}
+
+
 class TestExtractResponseText:
     def test_dict_response(self):
         resp = {"choices": [{"message": {"content": "hello"}}]}

ui/litellm-dashboard/src/lib/http/schema.d.ts

@@ -25339,8 +25339,9 @@ export interface components {
             /**
              * Cost Actual
              * @description Running total of judge-call spend for this job
+             * @default 0
              */
-            cost_actual?: number | null;
+            cost_actual: number;
             /** Cost Estimate */
             cost_estimate?: number | null;
             /** Created At */