From 49e31cf90cc2fc00330d5fd0d5e3b7407b23de5c Mon Sep 17 00:00:00 2001
From: NeuralEmpowerment <129192050+NeuralEmpowerment@users.noreply.github.com>
Date: Tue, 14 Jul 2026 14:11:56 -0700
Subject: [PATCH] fix: normalize cached and reasoning usage details

---
 plugins/tracing/dist/index.mjs     | 10 ++++++----
 plugins/tracing/src/trace.ts       | 23 ++++++++++++++++-------
 plugins/tracing/test/trace.test.ts | 21 ++++++++++++++++-----
 3 files changed, 38 insertions(+), 16 deletions(-)

diff --git a/plugins/tracing/dist/index.mjs b/plugins/tracing/dist/index.mjs
index cdc80c5..dbd553b 100644
--- a/plugins/tracing/dist/index.mjs
+++ b/plugins/tracing/dist/index.mjs
@@ -46984,11 +46984,13 @@ async function seededTraceParent(config$1, sessionMeta, turnNumber) {
 function toUsageDetails(usage) {
 	if (!usage) return void 0;
 	const details = {};
-	if (typeof usage.input_tokens === "number") details.input = usage.input_tokens;
-	if (typeof usage.output_tokens === "number") details.output = usage.output_tokens;
+	const cachedInputTokens = usage.cached_input_tokens ?? 0;
+	const reasoningOutputTokens = usage.reasoning_output_tokens ?? 0;
+	if (typeof usage.input_tokens === "number") details.input = Math.max(usage.input_tokens - cachedInputTokens, 0);
+	if (cachedInputTokens > 0) details.cache_read_input_tokens = cachedInputTokens;
+	if (typeof usage.output_tokens === "number") details.output = Math.max(usage.output_tokens - reasoningOutputTokens, 0);
+	if (reasoningOutputTokens > 0) details.output_reasoning_tokens = reasoningOutputTokens;
 	if (typeof usage.total_tokens === "number") details.total = usage.total_tokens;
-	if (typeof usage.cached_input_tokens === "number") details.cache_read_input_tokens = usage.cached_input_tokens;
-	if (typeof usage.reasoning_output_tokens === "number") details.reasoning_tokens = usage.reasoning_output_tokens;
 	return Object.keys(details).length > 0 ? details : void 0;
 }
 /** Build a clip() that truncates long strings to `maxChars`. */
diff --git a/plugins/tracing/src/trace.ts b/plugins/tracing/src/trace.ts
index 3f5e2d0..ef55868 100644
--- a/plugins/tracing/src/trace.ts
+++ b/plugins/tracing/src/trace.ts
@@ -112,15 +112,24 @@ async function seededTraceParent(
 function toUsageDetails(usage: TokenUsage | undefined): Record<string, number> | undefined {
   if (!usage) return undefined;
   const details: Record<string, number> = {};
-  if (typeof usage.input_tokens === "number") details.input = usage.input_tokens;
-  if (typeof usage.output_tokens === "number") details.output = usage.output_tokens;
-  if (typeof usage.total_tokens === "number") details.total = usage.total_tokens;
-  if (typeof usage.cached_input_tokens === "number") {
-    details.cache_read_input_tokens = usage.cached_input_tokens;
+  const cachedInputTokens = usage.cached_input_tokens ?? 0;
+  const reasoningOutputTokens = usage.reasoning_output_tokens ?? 0;
+
+  // Codex reports cached and reasoning tokens as subsets of input/output. Langfuse
+  // pricing treats usage-detail keys as additive buckets, so make them exclusive.
+  if (typeof usage.input_tokens === "number") {
+    details.input = Math.max(usage.input_tokens - cachedInputTokens, 0);
+  }
+  if (cachedInputTokens > 0) {
+    details.cache_read_input_tokens = cachedInputTokens;
   }
-  if (typeof usage.reasoning_output_tokens === "number") {
-    details.reasoning_tokens = usage.reasoning_output_tokens;
+  if (typeof usage.output_tokens === "number") {
+    details.output = Math.max(usage.output_tokens - reasoningOutputTokens, 0);
   }
+  if (reasoningOutputTokens > 0) {
+    details.output_reasoning_tokens = reasoningOutputTokens;
+  }
+  if (typeof usage.total_tokens === "number") details.total = usage.total_tokens;
   return Object.keys(details).length > 0 ? details : undefined;
 }
 
diff --git a/plugins/tracing/test/trace.test.ts b/plugins/tracing/test/trace.test.ts
index e3c0c95..473d9c2 100644
--- a/plugins/tracing/test/trace.test.ts
+++ b/plugins/tracing/test/trace.test.ts
@@ -92,11 +92,22 @@ describe("convertRollout", () => {
       expect(parentId(gen)).toBe(root!.spanContext().spanId);
       expect(attr(gen, "langfuse.observation.model.name")).toBe("gpt-5.4");
     }
-    // First generation carries token usage.
-    const usage = generations
-      .map((g) => attr(g, "langfuse.observation.usage_details"))
-      .find((u) => u.includes("120"));
-    expect(usage, "expected usage details with 120 total tokens").toBeTruthy();
+    // Usage buckets must be mutually exclusive for Langfuse cost inference.
+    const usageDetails = generations.map(
+      (g) => JSON.parse(attr(g, "langfuse.observation.usage_details")) as Record<string, number>,
+    );
+    expect(usageDetails).toContainEqual({
+      input: 100,
+      output: 15,
+      output_reasoning_tokens: 5,
+      total: 120,
+    });
+    expect(usageDetails).toContainEqual({
+      input: 100,
+      cache_read_input_tokens: 50,
+      output: 30,
+      total: 180,
+    });
 
     // One tool span, nested under a generation, with the captured command output.
     const tools = spans.filter((s) => obsType(s) === "tool");