显示思考内容

2026-08-05 20:04:59 +08:00
parent 98c6664701
commit 64469f9c1d
6 changed files with 76 additions and 5 deletions
@@ -91,6 +91,7 @@ command_code_working_directory: .
 timeout_seconds: 1800
 max_request_bytes: 20971520
 max_turns: 100
+stream_thinking: false
 response_store_directory: .command-code-openai-bridge/responses
 permission_mode: auto-accept
 dangerously_skip_permissions: true
@@ -101,7 +102,7 @@ models:
    effort: max
 ```

-`command_code_working_directory` 决定 Command Code 能看到和操作的项目目录。`response_store_directory` 保存 `store: true` 的 Response；两个相对路径都以配置文件所在目录为基准。默认存储目录是隐藏目录 `.command-code-openai-bridge/responses`。`models.<name>.effort` 会传给 Command Code 的 `--effort`，Responses 请求中的 `reasoning.effort` 优先。可用 effort 由对应底层模型决定。服务强制只监听 `127.0.0.1`。客户端只能选择配置中的模型名，不能注入额外 CLI 参数。
+`command_code_working_directory` 决定 Command Code 能看到和操作的项目目录。`response_store_directory` 保存 `store: true` 的 Response；两个相对路径都以配置文件所在目录为基准。默认存储目录是隐藏目录 `.command-code-openai-bridge/responses`。`stream_thinking` 控制 Chat Completions 流是否把 Command Code 思考事件包装成 `<think>` 块；默认关闭。`models.<name>.effort` 会传给 Command Code 的 `--effort`，Responses 请求中的 `reasoning.effort` 优先。可用 effort 由对应底层模型决定。服务强制只监听 `127.0.0.1`。客户端只能选择配置中的模型名，不能注入额外 CLI 参数。

 ### Obsidian Copilot 流式输出

@@ -112,8 +113,9 @@ models:
 - API Key：任意非空文本
 - CORS：关闭
 - Streaming：开启
+- Reasoning capability：开启

-Bridge 会直接响应浏览器 CORS 预检，包括 Copilot 和 OpenAI SDK 发送的自定义请求头及本机 Private Network Access。关闭 Copilot 的 CORS 绕过后，请求使用原生 `fetch`，可以读取 Bridge 返回的 SSE 流。
+Bridge 会直接响应浏览器 CORS 预检，包括 Copilot 和 OpenAI SDK 发送的自定义请求头及本机 Private Network Access。关闭 Copilot 的 CORS 绕过后，请求使用原生 `fetch`，可以读取 Bridge 返回的 SSE 流。将 `config.yaml` 中的 `stream_thinking` 设为 `true` 后，Bridge 会把实时思考事件作为 `<think>` 块发送，Copilot 将其显示为可折叠思考内容。思考中的 `Sources` 标记会被等效转义，避免 Copilot 把内部草稿误判为最终引用区并丢弃后续正文。

 ## 启动、停止与重启

@@ -148,7 +150,7 @@ Bearer Token 会被忽略。所有生成接口只接受文本。图片、音频

 ### Chat Completions

-支持字符串 `content`，也支持由 `{ "type": "text", "text": "..." }` 组成的数组。`stream` 省略或设为 `false` 时返回普通 JSON。`stream: true` 会立即建立 SSE 连接并发送 assistant role 块，随后在 Command Code 的 `text_delta` 到达时立即发送对应 SSE chunk，最后发送 `finish_reason: "stop"`、可选 usage 块和 `[DONE]`。传入 `stream_options.include_usage: true` 时，结束前返回 usage 块。
+支持字符串 `content`，也支持由 `{ "type": "text", "text": "..." }` 组成的数组。`stream` 省略或设为 `false` 时返回普通 JSON。`stream: true` 会立即建立 SSE 连接并发送 assistant role 块，随后在 Command Code 的 `text_delta` 到达时立即发送对应 SSE chunk，最后发送 `finish_reason: "stop"`、可选 usage 块和 `[DONE]`。`stream_thinking: true` 时，`thinking_delta` 会先作为 `<think>` 内容流发送；该兼容格式不是原生 OpenAI reasoning item。传入 `stream_options.include_usage: true` 时，结束前返回 usage 块。

 为兼容 Obsidian Copilot 和 LangChain OpenAI-format 客户端，Chat 接口还接受 `temperature`、`max_tokens`、`max_completion_tokens`、`top_p`、`frequency_penalty`、`presence_penalty` 和 `n: 1`。这些采样和输出限制参数会按 OpenAI 取值范围严格校验。Command Code CLI 1.12.0 没有对应的 headless 参数，因此 Bridge 不会把它们伪装成已生效：服务终端会输出警告，HTTP 响应带 `X-Command-Code-Ignored-Parameters`。其他未实现字段仍返回 `400 unsupported_parameter`。

@@ -5,6 +5,7 @@ command_code_working_directory: .
 timeout_seconds: 1800
 max_request_bytes: 20971520
 max_turns: 100
+stream_thinking: false
 response_store_directory: .command-code-openai-bridge/responses
 permission_mode: auto-accept
 dangerously_skip_permissions: true
@@ -5,6 +5,7 @@ command_code_working_directory: .
 timeout_seconds: 1800
 max_request_bytes: 20971520
 max_turns: 100
+stream_thinking: true
 response_store_directory: .command-code-openai-bridge/responses
 permission_mode: auto-accept
 dangerously_skip_permissions: true
@@ -9,10 +9,13 @@ export class ChatCompletionSseWriter {
  private readonly created = Math.floor(Date.now() / 1000);
  private readonly accumulator = new FinalTurnAccumulator();
  private outputText = "";
+  private thinkingOpen = false;
+  private thinkingBuffer = "";

  constructor(
    private readonly response: ServerResponse,
    private readonly model: string,
+    private readonly streamThinking: boolean,
  ) {}

  begin(): void {
@@ -21,12 +24,31 @@ export class ChatCompletionSseWriter {

  commandEvent(event: Record<string, unknown>): void {
    const chunks = this.accumulator.event(event);
+
+    if (this.streamThinking) {
+      if (event.type === "thinking_start") {
+        this.openThinking();
+        return;
+      }
+      if (event.type === "thinking_delta" && typeof event.delta === "string") {
+        this.openThinking();
+        this.addThinking(event.delta);
+        return;
+      }
+      if (event.type === "thinking_end") {
+        this.closeThinking();
+        return;
+      }
+    }
+
    if (event.type === "text_delta" && typeof event.delta === "string") {
+      this.closeThinking();
      this.addText([event.delta]);
      return;
    }

    if (event.type !== "turn_end") return;
+    this.closeThinking();
    if (event.hadToolCalls === true) {
      this.outputText = "";
      return;
@@ -35,6 +57,7 @@ export class ChatCompletionSseWriter {
  }

  finish(finalText: string, usage?: CommandUsage, includeUsage = false): void {
+    this.closeThinking();
    this.ensureFinalText(finalText);
    this.chunk({}, "stop");

@@ -56,10 +79,48 @@ export class ChatCompletionSseWriter {
  }

  error(error: object): void {
+    this.closeThinking();
    this.write(error);
    this.writeDone();
  }

+  private openThinking(): void {
+    if (this.thinkingOpen) return;
+    this.thinkingOpen = true;
+    this.chunk({ content: "<think>" }, null);
+  }
+
+  private closeThinking(): void {
+    if (!this.thinkingOpen) return;
+    if (this.thinkingBuffer !== "") {
+      this.chunk({ content: escapeThinkingSources(this.thinkingBuffer) }, null);
+      this.thinkingBuffer = "";
+    }
+    this.thinkingOpen = false;
+    this.chunk({ content: "</think>\n\n" }, null);
+  }
+
+  private addThinking(delta: string): void {
+    this.thinkingBuffer += delta;
+    let content = "";
+
+    while (this.thinkingBuffer.length > 6) {
+      const match = /sources/i.exec(this.thinkingBuffer);
+      if (match) {
+        content += this.thinkingBuffer.slice(0, match.index);
+        content += escapeThinkingSources(match[0]);
+        this.thinkingBuffer = this.thinkingBuffer.slice(match.index + match[0].length);
+        continue;
+      }
+
+      const safeLength = this.thinkingBuffer.length - 6;
+      content += this.thinkingBuffer.slice(0, safeLength);
+      this.thinkingBuffer = this.thinkingBuffer.slice(safeLength);
+    }
+
+    if (content !== "") this.chunk({ content }, null);
+  }
+
  private addText(chunks: string[]): void {
    for (const delta of chunks) {
      if (delta === "") continue;
@@ -104,9 +165,14 @@ export class ChatCompletionSseWriter {
  }
 }

+function escapeThinkingSources(text: string): string {
+  return text.replace(/sources/gi, (word) => `${word.slice(0, 4)}\u200B${word.slice(4)}`);
+}
+
 export function startChatCompletionSse(
  reply: FastifyReply,
  model: string,
+  streamThinking: boolean,
 ): ChatCompletionSseWriter {
  reply.hijack();
  reply.raw.writeHead(200, {
@@ -116,7 +182,7 @@ export function startChatCompletionSse(
    "X-Accel-Buffering": "no",
  });
  reply.raw.flushHeaders();
-  const writer = new ChatCompletionSseWriter(reply.raw, model);
+  const writer = new ChatCompletionSseWriter(reply.raw, model, streamThinking);
  writer.begin();
  return writer;
 }
@@ -16,6 +16,7 @@ const configSchema = z.object({
  timeout_seconds: z.number().int().positive().default(1800),
  max_request_bytes: z.number().int().positive().default(20 * 1024 * 1024),
  max_turns: z.number().int().positive().default(100),
+  stream_thinking: z.boolean().default(false),
  response_store_directory: z.string().min(1).default(".command-code-openai-bridge/responses"),
  permission_mode: z.enum(["default", "standard", "plan", "auto-accept", "dont-ask"]).default("auto-accept"),
  dangerously_skip_permissions: z.boolean().default(false),
@@ -173,7 +173,7 @@ export async function createServer(config: BridgeConfig): Promise<BridgeServer>

    let writer: ChatCompletionSseWriter | undefined;
    try {
-      if (parsed.stream) writer = startChatCompletionSse(reply, parsed.model);
+      if (parsed.stream) writer = startChatCompletionSse(reply, parsed.model, config.stream_thinking);
      const result = await runCommandCode(
        config,
        model.cli_model,