From d6d585e20099744d5a67d71a4595d56b9ffc5765 Mon Sep 17 00:00:00 2001
From: Tyler Ford <tylerford012@gmail.com>
Date: Sun, 17 May 2026 20:11:06 -0400
Subject: [PATCH] fix: extract flat cached_tokens/reasoning_tokens from
 OpenAI-compatible usage objects (#2350)
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Integrated into release/v3.8.0 — flat cached_tokens/reasoning_tokens extraction for Xiaomi MiMo and similar providers
---
 open-sse/handlers/usageExtractor.ts |  6 +++--
 open-sse/utils/usageTracking.ts     |  6 +++--
 tests/unit/usage-extractor.test.ts  | 40 +++++++++++++++++++++++++++++
 3 files changed, 48 insertions(+), 4 deletions(-)

diff --git a/open-sse/handlers/usageExtractor.ts b/open-sse/handlers/usageExtractor.ts
index e2e87983a9..92affd1c39 100644
--- a/open-sse/handlers/usageExtractor.ts
+++ b/open-sse/handlers/usageExtractor.ts
@@ -25,10 +25,12 @@ export function extractUsageFromResponse(responseBody, provider) {
       cached_tokens:
         responseBody.usage.prompt_tokens_details?.cached_tokens ??
         responseBody.usage.input_tokens_details?.cached_tokens ??
-        responseBody.usage.prompt_cache_hit_tokens,
+        responseBody.usage.prompt_cache_hit_tokens ??
+        responseBody.usage.cached_tokens,
       reasoning_tokens:
         responseBody.usage.completion_tokens_details?.reasoning_tokens ??
-        responseBody.usage.output_tokens_details?.reasoning_tokens,
+        responseBody.usage.output_tokens_details?.reasoning_tokens ??
+        responseBody.usage.reasoning_tokens,
     };
   }
 
diff --git a/open-sse/utils/usageTracking.ts b/open-sse/utils/usageTracking.ts
index 15e7d681d4..38913a2263 100644
--- a/open-sse/utils/usageTracking.ts
+++ b/open-sse/utils/usageTracking.ts
@@ -366,10 +366,12 @@ export function extractUsage(chunk) {
       cached_tokens:
         chunk.usage.prompt_tokens_details?.cached_tokens ??
         chunk.usage.input_tokens_details?.cached_tokens ??
-        chunk.usage.prompt_cache_hit_tokens,
+        chunk.usage.prompt_cache_hit_tokens ??
+        chunk.usage.cached_tokens,
       reasoning_tokens:
         chunk.usage.completion_tokens_details?.reasoning_tokens ??
-        chunk.usage.output_tokens_details?.reasoning_tokens,
+        chunk.usage.output_tokens_details?.reasoning_tokens ??
+        chunk.usage.reasoning_tokens,
     });
   }
 
diff --git a/tests/unit/usage-extractor.test.ts b/tests/unit/usage-extractor.test.ts
index 4d64bf6eca..6c7bdf5273 100644
--- a/tests/unit/usage-extractor.test.ts
+++ b/tests/unit/usage-extractor.test.ts
@@ -292,3 +292,43 @@ test("extractUsage reads OpenAI streaming chunk with prompt_tokens_details", ()
   assert.equal(usage.cached_tokens, 50);
   assert.equal(usage.reasoning_tokens, 20);
 });
+
+// ── Flat field extraction tests (Xiaomi MiMo-style providers) ──
+
+test("extractUsageFromResponse reads flat cached_tokens and reasoning_tokens from OpenAI-compatible usage", () => {
+  const usage = extractUsageFromResponse(
+    {
+      usage: {
+        prompt_tokens: 258,
+        completion_tokens: 50,
+        total_tokens: 308,
+        cached_tokens: 192,
+        reasoning_tokens: 49,
+      },
+    },
+    "xiaomi-mimo"
+  );
+
+  assert.deepEqual(usage, {
+    prompt_tokens: 258,
+    completion_tokens: 50,
+    cached_tokens: 192,
+    reasoning_tokens: 49,
+  });
+});
+
+test("extractUsage reads flat cached_tokens and reasoning_tokens from streaming chunk", () => {
+  const usage = extractUsage({
+    choices: [{ delta: {}, finish_reason: "stop" }],
+    usage: {
+      prompt_tokens: 258,
+      completion_tokens: 50,
+      total_tokens: 308,
+      cached_tokens: 192,
+      reasoning_tokens: 49,
+    },
+  });
+
+  assert.equal(usage.cached_tokens, 192);
+  assert.equal(usage.reasoning_tokens, 49);
+});