兼容responses

2026-08-20 14:32:30 +08:00
parent d84b6f863d
commit b03d47c891
3 changed files with 64 additions and 9 deletions
@@ -231,7 +231,9 @@ Chat Completions 和 Responses 共用 Command Code NDJSON 执行层。服务终
 - `tool_choice`
 - `parallel_tool_calls`

-`input` 可以是字符串，也可以是 item 数组。message 支持 `system`、`developer`、`user`、`assistant` 角色、字符串 content，以及 `input_text`、`output_text`、`input_image`、`input_file` part。工具轮次还支持 `function_call` 和 `function_call_output` item，分别携带 `call_id`、工具名、`arguments` 和工具执行结果 `output`。
+为兼容 DeepSeek Harness、OpenAI Node.js SDK 等 Responses 客户端，接口还会接受但不实际应用这些字段：`prompt_cache_key`、`prompt_cache_retention`、`prompt_cache_options`、`max_output_tokens`、`temperature`、`top_p`、`service_tier`、`stream_options`、`include`、`reasoning.generate_summary` 和 `reasoning.summary`。请求显式提供这些字段时，服务终端会输出警告，HTTP 响应的 `X-Command-Code-Ignored-Parameters` 会列出未实际应用的顶层字段。
+
+`input` 可以是字符串，也可以是 item 数组。message 支持 `system`、`developer`、`user`、`assistant` 角色、字符串 content，以及 `input_text`、`output_text`、`input_image`、`input_file` part。历史 message 常见的 `id`、`status`、`phase` 和 output_text 的 `annotations` 也会被接受，但只用于兼容客户端回放，不会写入本地响应存储。工具轮次还支持 `function_call` 和 `function_call_output` item，分别携带 `call_id`、工具名、`arguments` 和工具执行结果 `output`。

 `input_image.image_url` 与 Chat 一样只接受图片 data URL；`file_id`、远程 URL、`file://`、本地路径和 low/high detail 不支持。`input_file` 只接受 `file_data` 加 `filename`，其中 `file_data` 可以是严格 base64 或带受支持文本 MIME 的 base64 data URL；解码结果必须是无 NUL 的 UTF-8 文本。`file_url`、`file_id`、PDF、Office 和其他二进制会返回带准确字段路径的 400。

@@ -14,15 +14,15 @@ permission_mode: auto-accept
 dangerously_skip_permissions: true

 models:
-  deepseek-v4-flash:
+  command-flash:
    cli_model: deepseek/deepseek-v4-flash
    effort: max
    supports_image_input: false
-  deepseek-v4-pro:
+  command-pro:
    cli_model: deepseek/deepseek-v4-pro
    effort: max
    supports_image_input: false
-  gpt-5.6-luna:
+  command-luna:
    cli_model: gpt-5.6-luna
    effort: max
    supports_image_input: true
@@ -52,6 +52,7 @@ const inputTextPartSchema = z.object({
 const outputTextPartSchema = z.object({
  type: z.literal("output_text"),
  text: z.string(),
+  annotations: z.array(z.unknown()).optional(),
 }).strict();

 const inputImagePartSchema = z.object({
@@ -94,6 +95,7 @@ const inputFilePartSchema = z.object({

 const inputMessageSchema = z.object({
  type: z.literal("message").optional(),
+  id: responseItemIdSchema.optional(),
  role: roleSchema,
  content: z.union([
    z.string(),
@@ -104,6 +106,8 @@ const inputMessageSchema = z.object({
      inputFilePartSchema,
    ])).min(1),
  ]),
+  status: z.enum(["in_progress", "completed", "incomplete"]).optional(),
+  phase: z.string().min(1).optional(),
 }).strict();

 const inputFunctionCallSchema = z.object({
@@ -175,17 +179,38 @@ const metadataSchema = z.record(z.string().max(512)).superRefine((metadata, cont
  }
 });

+const responseReasoningSchema = z.object({
+  effort: reasoningEffortSchema,
+  generate_summary: z.enum(["auto", "concise", "detailed"]).nullable().optional(),
+  summary: z.enum(["auto", "concise", "detailed"]).nullable().optional(),
+}).strict();
+
+const responseStreamOptionsSchema = z.object({
+  include_usage: z.boolean().optional(),
+}).strict();
+
+const responsePromptCacheOptionsSchema = z.object({
+  mode: z.enum(["explicit"]),
+}).strict();
+
 export const responseRequestSchema = z.object({
  model: z.string().min(1),
  input: z.union([z.string(), z.array(responseInputItemSchema).min(1)]),
  instructions: z.string().nullable().optional().default(null),
  stream: z.boolean().optional().default(false),
+  stream_options: responseStreamOptionsSchema.nullable().optional(),
  store: z.boolean().optional().default(true),
  previous_response_id: responseIdSchema.nullable().optional().default(null),
  metadata: metadataSchema.optional().default({}),
-  reasoning: z.object({
-    effort: reasoningEffortSchema,
-  }).strict().optional(),
+  prompt_cache_key: z.string().min(1).optional(),
+  prompt_cache_retention: z.enum(["in-memory", "24h"]).nullable().optional(),
+  prompt_cache_options: responsePromptCacheOptionsSchema.optional(),
+  max_output_tokens: z.number().int().positive().nullable().optional(),
+  temperature: z.number().min(0).max(2).nullable().optional(),
+  top_p: z.number().min(0).max(1).nullable().optional(),
+  service_tier: z.enum(["auto", "default", "flex", "scale", "priority"]).nullable().optional(),
+  include: z.array(z.string()).min(1).nullable().optional(),
+  reasoning: responseReasoningSchema.optional(),
  text: z.object({
    format: responseTextFormatSchema,
  }).strict().optional().default({ format: { type: "text" } }),
@@ -227,6 +252,22 @@ export type ResponseTextFormat = z.infer<typeof responseTextFormatSchema>;
 type ResponseRole = z.infer<typeof roleSchema>;
 type ResponseStatus = "in_progress" | "completed" | "incomplete" | "failed" | "cancelled";

+const ignoredResponseCompatibilityParameters = [
+  "prompt_cache_key",
+  "prompt_cache_retention",
+  "prompt_cache_options",
+  "max_output_tokens",
+  "temperature",
+  "top_p",
+  "service_tier",
+  "stream_options",
+  "include",
+] as const;
+
+function ignoredResponseCompatibilityParameterList(request: ResponseRequest): string[] {
+  return ignoredResponseCompatibilityParameters.filter((key) => request[key] !== undefined);
+}
+
 interface ResponseInputPart {
  type: "input_text" | "output_text";
  text: string;
@@ -844,6 +885,16 @@ export async function registerResponseRoutes(
        "model",
      ));
    }
+
+    const ignoredParameters = ignoredResponseCompatibilityParameterList(parsed);
+    if (ignoredParameters.length > 0) {
+      const parameterList = ignoredParameters.join(",");
+      reply.raw.setHeader("X-Command-Code-Ignored-Parameters", parameterList);
+      process.stderr.write(
+        `警告：Responses 参数 ${parameterList} 已为客户端兼容而接受；Command Code CLI 不支持映射，实际生成不会应用这些参数。\n`,
+      );
+    }
+
    const firstImage = pendingAttachments.find((attachment) => attachment.kind === "image");
    if (firstImage && !model.supports_image_input) {
      return sendRouteError(reply, new AttachmentInputError(
@@ -1042,8 +1093,10 @@ function parseResponseRequest(body: unknown): ResponseRequest {
 function findUnsupportedParameter(body: unknown): string | undefined {
  if (!isRecord(body)) return undefined;
  const supported = new Set([
-    "model", "input", "instructions", "stream", "store",
-    "previous_response_id", "metadata", "reasoning", "text",
+    "model", "input", "instructions", "stream", "stream_options", "store",
+    "previous_response_id", "metadata", "prompt_cache_key",
+    "prompt_cache_retention", "prompt_cache_options", "max_output_tokens",
+    "temperature", "top_p", "service_tier", "include", "reasoning", "text",
    "tools", "tool_choice", "parallel_tool_calls",
  ]);
  const unknownTopLevel = Object.keys(body).find((key) => !supported.has(key));