feat: allign with vscode-copilot-chat extractThinkingData, otherwise it will cause miss cache occasionally

Author: caozhiyuan

src/routes/messages/responses-stream-translation.ts

@@ -163,7 +163,7 @@ const handleOutputItemDone = (
 
   const outputIndex = rawEvent.output_index
   const blockIndex = openThinkingBlockIfNeeded(state, outputIndex, events)
-  const signature = item.encrypted_content
+  const signature = (item.encrypted_content ?? "") + "@" + item.id
   if (signature) {
     events.push({
       type: "content_block_delta",
@@ -468,8 +468,9 @@ const openThinkingBlockIfNeeded = (
   outputIndex: number,
   events: Array<AnthropicStreamEventData>,
 ): number => {
-  const contentIndex = 0
-  const key = getBlockKey(outputIndex, contentIndex)
+  //thinking blocks has multiple summary_index, should combine into one block
+  const summaryIndex = 0
+  const key = getBlockKey(outputIndex, summaryIndex)
   let blockIndex = state.blockIndexByKey.get(key)
 
   if (blockIndex === undefined) {

src/routes/messages/responses-translation.ts

@@ -72,7 +72,7 @@ export const translateAnthropicMessagesToResponsesPayload = (
     stream: payload.stream ?? null,
     store: false,
     parallel_tool_calls: true,
-    reasoning: { effort: "high", summary: "auto" },
+    reasoning: { effort: "high", summary: "detailed" },
     include: ["reasoning.encrypted_content"],
   }
 
@@ -232,16 +232,20 @@ const createImageContent = (
 
 const createReasoningContent = (
   block: AnthropicThinkingBlock,
-): ResponseInputReasoning => ({
-  type: "reasoning",
-  summary: [
-    {
-      type: "summary_text",
-      text: block.thinking,
-    },
-  ],
-  encrypted_content: block.signature,
-})
+): ResponseInputReasoning => {
+  // allign with vscode-copilot-chat extractThinkingData, otherwise it will cause miss cache occasionally —— the usage input cached tokens to be 0
+  // https://github.com/microsoft/vscode-copilot-chat/blob/main/src/platform/endpoint/node/responsesApi.ts#L162
+  // when use in codex cli, reasoning id is empty, so it will cause miss cache occasionally
+  const array = block.signature.split("@")
+  const signature = array[0]
+  const id = array[1]
+  return {
+    id,
+    type: "reasoning",
+    summary: [],
+    encrypted_content: signature,
+  }
+}
 
 const createFunctionToolCall = (
   block: AnthropicToolUseBlock,

src/services/copilot/create-responses.ts

@@ -77,6 +77,7 @@ export interface ResponseFunctionCallOutputItem {
 }
 
 export interface ResponseInputReasoning {
+  id?: string
   type: "reasoning"
   summary: Array<{
     type: "summary_text"