预设兼容：按真实抓包对齐装配与包装层，补上声明式工具与预设正则。

- 装配只认 prompt_order.enabled；squash_system_messages 只合并相邻 system（新增 system 档，连接符按酒馆用单换行）
- 包装层：new_chat_prompt、示例对话（new_example_chat_prompt + <START> 切段 + name）、角色卡三块与 personality_format/scenario_format、{{user}}/{{char}} 取值、任务契约并进预设条目
- 工具协议：按 ToolBindings / 伪工具协议数据声明真 tools（presetToolPolicy=preset），工具描述用到的条目不再重复作为消息发出
- 正则：promptOnly 作用于对话历史（含 depth 过滤）、写入时脚本改写回复；{{//}} 注释按配平扫描
- tools/st-request-diff：真假上游抓包对比工具、三份预设的机制结论与实测报告

src/llm/client.ts

@@ -1,5 +1,7 @@
 import type { LlmConfig } from "../config/env.js";
+import { ensureNetworkProxy } from "../config/network.js";
 import type { GenerationParameters } from "../types/preset.js";
+import type { WorldInfoPack as PresetPackForFrame } from "../preset/world-info-pack.js";
 import {
   applyOpenAiGeneration,
   fetchWithGenerationCompat,
@@ -24,13 +26,14 @@ export type ToolCallPayload = {
 };
 
 export type ChatMessage =
-  | { role: "system" | "user"; content: string }
+  | { role: "system" | "user"; content: string; name?: string }
   | {
       role: "assistant";
       content: string | null;
       tool_calls?: ToolCallPayload[];
+      name?: string;
     }
-  | { role: "tool"; content: string; tool_call_id: string };
+  | { role: "tool"; content: string; tool_call_id: string; name?: string };
 
 export type ToolDefinition = {
   type: "function";
@@ -117,6 +120,11 @@ export type CompleteOptions = {
   /** 统计补充。不进上游请求体。 */
   usageDetail?: UsageDetail;
   usageSink?: UsageSink;
+  /**
+   * 内部：play 真多轮装配的分袋上下文（由 worker 从 contextSegments 切成四袋 + 历史轮次）。
+   * 不进上游请求体；由 PresetLlmProvider 取用后丢弃。
+   */
+  presetPack?: PresetPackForFrame;
 };
 
 export type ToolChoice =
@@ -223,6 +231,33 @@ export function ensureJsonObjectPrompt(messages: ChatMessage[]): ChatMessage[] {
   return [{ role: "system", content: JSON_OBJECT_HINT }, ...messages];
 }
 
+/** 空正文时报得能定位：是限额吃光了，还是模型真没给。 */
+function emptyToolResponseMessage(params: {
+  stream: boolean;
+  finishReason?: string;
+  reasoningChars?: number;
+  usage?: TokenUsage;
+}): string {
+  const bits: string[] = [];
+  if (params.finishReason) bits.push(`finish_reason=${params.finishReason}`);
+  if (params.reasoningChars) bits.push(`思维链=${params.reasoningChars}字`);
+  if (params.usage) {
+    bits.push(
+      `prompt=${params.usage.promptTokens}/completion=${params.usage.completionTokens}`,
+    );
+  }
+  const hint =
+    params.finishReason === "length"
+      ? "（输出上限被思考/正文吃光，调大 maxOutputTokens）"
+      : params.reasoningChars
+        ? "（只产出思维链没有正文，多半是输出上限不足）"
+        : "";
+  const head = params.stream
+    ? "LLM 流式返回空正文且没有工具调用"
+    : "LLM 返回空正文且没有工具调用";
+  return `${head}${bits.length ? `：${bits.join(" ")}` : ""}${hint}`;
+}
+
 export function buildRequestBody(
   config: LlmConfig,
   messages: ChatMessage[],
@@ -431,6 +466,7 @@ async function llmFetch(
 ): Promise<Response> {
   return fetchWithGenerationCompat(async (payload) => {
     throwIfAborted(options?.signal);
+    await ensureNetworkProxy();
     return fetch(url, {
       method: "POST",
       headers: {
@@ -533,17 +569,29 @@ export class OpenAiCompatibleProvider implements LlmProvider {
     const data = (await response.json()) as {
       model?: string;
       usage?: unknown;
-      choices?: Array<{ message?: Record<string, unknown> }>;
+      choices?: Array<{
+        message?: Record<string, unknown>;
+        finish_reason?: string;
+      }>;
     };
-    const parts = extractMessageParts(data.choices?.[0]?.message);
+    const choice = data.choices?.[0];
+    const parts = extractMessageParts(choice?.message);
+    const usage = parseUsage(data.usage);
     if (!parts.content && parts.toolCalls.length === 0) {
-      throw new Error("LLM returned empty content and no tool calls");
+      throw new Error(
+        emptyToolResponseMessage({
+          stream: false,
+          finishReason: choice?.finish_reason,
+          reasoningChars: parts.reasoning?.length ?? 0,
+          usage,
+        }),
+      );
     }
     return {
       content: parts.content,
       toolCalls: parts.toolCalls,
       reasoning: parts.reasoning,
-      usage: parseUsage(data.usage),
+      usage,
       model: data.model ?? this.config.model,
     };
   }
@@ -576,7 +624,12 @@ export class OpenAiCompatibleProvider implements LlmProvider {
 
     const parts = await consumeOpenAiToolStream(response.body, callbacks);
     if (!parts.content && parts.toolCalls.length === 0 && !parts.reasoning) {
-      throw new Error("LLM stream returned empty content and no tool calls");
+      throw new Error(
+        emptyToolResponseMessage({
+          stream: true,
+          usage: parts.usage,
+        }),
+      );
     }
     const content =
       parts.content ??