fix

2026-08-23 22:52:13 +02:00
parent 8983cfc8a1
commit 296dc3c841
32 changed files with 2025 additions and 52 deletions
@@ -0,0 +1,81 @@
+package aiaudit
+
+import (
+	"testing"
+)
+
+// gpt_response nests the usage block differently across pipeline versions, so the
+// walker looks for any object carrying "usage" rather than one fixed path. A rigid
+// path would silently recover nothing from older rows.
+func TestExtractUsageCalls_findsNestedUsage(t *testing.T) {
+	t.Parallel()
+	raw := []byte(`{
+		"steps": [
+			{"step": "normalize", "raw": {"keys": 21}},
+			{"step": "categorize", "raw": {"model": "gpt-5.6-luna",
+				"usage": {"prompt_tokens": 700, "completion_tokens": 5, "total_tokens": 705}}},
+			{"step": "ai_enhance", "raw": {"languages": [
+				{"language": "sl", "raw": {"raw": {"model": "gpt-5.6-luna", "status": 200,
+					"usage": {"prompt_tokens": 4000, "completion_tokens": 800, "total_tokens": 4800}}}}
+			]}}
+		]
+	}`)
+	calls := extractUsageCalls(raw)
+	if len(calls) != 2 {
+		t.Fatalf("expected 2 recovered calls, got %d: %+v", len(calls), calls)
+	}
+	byRole := map[string]recoveredCall{}
+	for _, c := range calls {
+		byRole[c.Role] = c
+	}
+	cat, ok := byRole[RoleCategorize]
+	if !ok {
+		t.Fatalf("categorize call not recovered: %+v", calls)
+	}
+	if cat.PromptTokens != 700 || cat.Model != "gpt-5.6-luna" {
+		t.Fatalf("categorize call wrong: %+v", cat)
+	}
+	enh, ok := byRole[RoleProcessing]
+	if !ok {
+		t.Fatalf("enhance call not recovered: %+v", calls)
+	}
+	if enh.PromptTokens != 4000 || enh.CompletionTokens != 800 {
+		t.Fatalf("enhance call wrong: %+v", enh)
+	}
+}
+
+func TestExtractUsageCalls_ignoresRowsWithoutUsage(t *testing.T) {
+	t.Parallel()
+	for name, raw := range map[string]string{
+		"empty":          `{}`,
+		"steps only":     `{"steps":[{"step":"normalize","raw":{"keys":3}}]}`,
+		"zero tokens":    `{"raw":{"usage":{"prompt_tokens":0,"completion_tokens":0}}}`,
+		"not json":       `not json at all`,
+		"unchanged skip": `{"steps":[{"step":"ai_enhance","raw":{"languages":[{"raw":{"status":"unchanged"}}]}}]}`,
+	} {
+		t.Run(name, func(t *testing.T) {
+			if got := extractUsageCalls([]byte(raw)); len(got) != 0 {
+				t.Fatalf("expected no recovered calls, got %+v", got)
+			}
+		})
+	}
+}
+
+// Backfilled history has no cached-token detail, so input is priced as fully
+// fresh. That must over-estimate, never under-estimate.
+func TestBackfillPricing_treatsInputAsUncached(t *testing.T) {
+	t.Parallel()
+	luna, _ := DefaultModelPrice("gpt-5.6-luna")
+	backfilled := CostMicros(luna, 10_000, 0, 1_000)
+	ifHalfCached := CostMicros(luna, 10_000, 5_000, 1_000)
+	if backfilled <= ifHalfCached {
+		t.Fatalf("backfill cost %d should exceed the cached-aware cost %d", backfilled, ifHalfCached)
+	}
+}
+
+func TestBackfillSourcesAreDistinct(t *testing.T) {
+	t.Parallel()
+	if SourceLive == SourceBackfill {
+		t.Fatal("live and backfill rows must be distinguishable")
+	}
+}