From 6c5fa1b51f424cd786e48196fffdff240bdde9c5 Mon Sep 17 00:00:00 2001
From: warkcod <warkcod@gmail.com>
Date: Wed, 20 May 2026 16:52:35 +0800
Subject: [PATCH] Preserve Responses streaming usage for compaction

Claude Code relies on assistant usage metadata as a token-count anchor when deciding whether to auto-compact resumed sessions. The Responses streaming transformer dropped usage from response.completed, leaving transcript input/output usage at zero and allowing long sessions to overrun provider context windows before auto-compact could trigger.

Constraint: OpenAI Responses emits final usage on response.completed rather than ordinary text/tool deltas

Rejected: Lower the local auto-compact window only | masks the bad usage metadata and still misleads session accounting

Confidence: high

Scope-risk: narrow

Directive: Keep streaming usage mapped when changing Responses SSE conversion; Claude Code context management depends on it

Tested: pnpm --filter @musistudio/llms tsx --test scripts/openaiResponsesTransformer.test.ts scripts/providerError.test.ts

Tested: pnpm build
---
 .../openaiResponsesTransformer.test.ts        | 67 +++++++++++++++++++
 .../openai.responses.transformer.ts           | 15 +++++
 2 files changed, 82 insertions(+)
 create mode 100644 packages/core/scripts/openaiResponsesTransformer.test.ts

diff --git a/packages/core/scripts/openaiResponsesTransformer.test.ts b/packages/core/scripts/openaiResponsesTransformer.test.ts
new file mode 100644
index 0000000..6e12cb4
--- /dev/null
+++ b/packages/core/scripts/openaiResponsesTransformer.test.ts
@@ -0,0 +1,67 @@
+import assert from "node:assert/strict";
+import { describe, it } from "node:test";
+import { OpenAIResponsesTransformer } from "../src/transformer/openai.responses.transformer";
+
+const readStream = async (stream: ReadableStream<Uint8Array>) => {
+  const reader = stream.getReader();
+  const decoder = new TextDecoder();
+  let output = "";
+
+  while (true) {
+    const { done, value } = await reader.read();
+    if (done) break;
+    output += decoder.decode(value, { stream: true });
+  }
+
+  output += decoder.decode();
+  return output;
+};
+
+const parseSseData = (output: string) =>
+  output
+    .split(/\r?\n/)
+    .filter((line) => line.startsWith("data: "))
+    .map((line) => line.slice("data: ".length).trim())
+    .filter((line) => line && line !== "[DONE]")
+    .map((line) => JSON.parse(line));
+
+describe("OpenAIResponsesTransformer", () => {
+  it("preserves streaming usage from response.completed", async () => {
+    const upstream = [
+      {
+        type: "response.completed",
+        response: {
+          id: "resp_123",
+          model: "gpt-5.5",
+          output: [{ type: "message" }],
+          usage: {
+            input_tokens: 12345,
+            output_tokens: 67,
+            total_tokens: 12412,
+          },
+        },
+      },
+    ]
+      .map((event) => `data: ${JSON.stringify(event)}\n\n`)
+      .join("");
+
+    const response = new Response(upstream, {
+      headers: { "Content-Type": "text/event-stream" },
+    });
+
+    const transformed =
+      await new OpenAIResponsesTransformer().transformResponseOut(response);
+    assert.ok(transformed.body);
+
+    const chunks = parseSseData(await readStream(transformed.body));
+    const doneChunk = chunks.find(
+      (chunk) => chunk.choices?.[0]?.finish_reason === "stop"
+    );
+
+    assert.deepEqual(doneChunk?.usage, {
+      prompt_tokens: 12345,
+      completion_tokens: 67,
+      total_tokens: 12412,
+    });
+  });
+});
diff --git a/packages/core/src/transformer/openai.responses.transformer.ts b/packages/core/src/transformer/openai.responses.transformer.ts
index a0a3857..6be9cd0 100644
--- a/packages/core/src/transformer/openai.responses.transformer.ts
+++ b/packages/core/src/transformer/openai.responses.transformer.ts
@@ -61,6 +61,11 @@ interface ResponsesStreamEvent {
     output?: Array<{
       type: string;
     }>;
+    usage?: {
+      input_tokens?: number;
+      output_tokens?: number;
+      total_tokens?: number;
+    };
   };
   arguments?: string;
   reasoning_summary?: string; // 添加推理摘要支持
@@ -550,6 +555,16 @@ export class OpenAIResponsesTransformer implements Transformer {
                               finish_reason: finishReason,
                             },
                           ],
+                          usage: data.response?.usage
+                            ? {
+                                prompt_tokens:
+                                  data.response.usage.input_tokens || 0,
+                                completion_tokens:
+                                  data.response.usage.output_tokens || 0,
+                                total_tokens:
+                                  data.response.usage.total_tokens || 0,
+                              }
+                            : undefined,
                         };
 
                         controller.enqueue(