From 59ab6642bedc683bcabef27a7b8ca640e4fa8906 Mon Sep 17 00:00:00 2001
From: cepvor <cepvor@qq.com>
Date: Sat, 16 May 2026 12:02:18 +0800
Subject: [PATCH] =?UTF-8?q?fix:=20OpenAI=20=E8=B7=AF=E5=BE=84=E6=B7=BB?=
 =?UTF-8?q?=E5=8A=A0=E9=98=B2=E5=BE=A1=E6=80=A7=20usage=20=E5=90=88?=
 =?UTF-8?q?=E5=B9=B6=E9=80=BB=E8=BE=91?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

将 message_delta 处理器中的 `{ ...usage, ...deltaUsage }` spread
替换为 updateOpenAIUsage() 函数，与 claude.ts 的 updateUsage() 模式
保持一致。

当前适配器发送的 message_delta 包含完整字段，spread 不会丢失数据。
但若未来适配器在某事件中省略了 cache 字段（传 explicit 0），spread
会无声覆盖有效值。新函数仅在 delta 字段有意义值时更新，否则保留
当前值。

Co-Authored-By: deepseek-v4-pro[1m] <deepseek-ai@claude-code-best.win>
---
 src/services/api/openai/index.ts | 45 +++++++++++++++++++++++++++++++-
 1 file changed, 44 insertions(+), 1 deletion(-)

diff --git a/src/services/api/openai/index.ts b/src/services/api/openai/index.ts
index 520290b189..3e8225f516 100644
--- a/src/services/api/openai/index.ts
+++ b/src/services/api/openai/index.ts
@@ -130,6 +130,46 @@ function isOpenAIConvertibleMessage(
   return msg.type === 'assistant' || msg.type === 'user'
 }
 
+/**
+ * Merge delta usage into accumulated usage, preserving cache fields from
+ * previous values when the delta carries explicit zeroes.
+ *
+ * Mirrors updateUsage() in claude.ts: Anthropic's streaming API may send
+ * explicit 0 for cache fields in message_delta events, which should not
+ * overwrite valid values from message_start. OpenAI-compatible endpoints
+ * don't currently exhibit this behavior, but defensive field-level merging
+ * prevents a future adapter change from silently zeroing cache data.
+ */
+function updateOpenAIUsage(
+  current: {
+    input_tokens: number
+    output_tokens: number
+    cache_creation_input_tokens: number
+    cache_read_input_tokens: number
+  },
+  delta: {
+    input_tokens?: number
+    output_tokens?: number
+    cache_creation_input_tokens?: number
+    cache_read_input_tokens?: number
+  },
+): typeof current {
+  return {
+    input_tokens: delta.input_tokens ?? current.input_tokens,
+    output_tokens: delta.output_tokens ?? current.output_tokens,
+    cache_creation_input_tokens:
+      delta.cache_creation_input_tokens !== undefined &&
+      delta.cache_creation_input_tokens > 0
+        ? delta.cache_creation_input_tokens
+        : current.cache_creation_input_tokens,
+    cache_read_input_tokens:
+      delta.cache_read_input_tokens !== undefined &&
+      delta.cache_read_input_tokens > 0
+        ? delta.cache_read_input_tokens
+        : current.cache_read_input_tokens,
+  }
+}
+
 /**
  * Assemble the final AssistantMessage (and optional max_tokens error) from
  * accumulated stream state. Extracted to avoid duplication between the
@@ -449,7 +489,10 @@ export async function* queryModelOpenAI(
         case 'message_delta': {
           const deltaUsage = (event as any).usage
           if (deltaUsage) {
-            usage = { ...usage, ...deltaUsage }
+            // Defensive merge: only update fields that are present and meaningful.
+            // Matches the pattern in claude.ts updateUsage() — prevents a future
+            // adapter change from silently zeroing cache fields via spread.
+            usage = updateOpenAIUsage(usage, deltaUsage)
           }
           if ((event as any).delta?.stop_reason != null) {
             stopReason = (event as any).delta.stop_reason