统一在预设包装层做提示词夹心，并修正探针实际发信路径。

让所有模型请求与界面预览共用同一套夹心，避免探针与运行时行为分叉。

Co-authored-by: Cursor <cursoragent@cursor.com>

docs/preset-format.md

@@ -303,7 +303,7 @@ Runtime Session
 
 因此，同一个 preset 可以用于多个创作流程；同一个创作流程也可以切换不同 preset。二者是正交关系。
 
-生成参数可全局合并进 LLM 请求。**prompt 条目 + marker 填洞只在 play worker 上用**：创作步 / 总管 tool loop 不得把角色卡骨架套到 design-step 上，也不得在填洞后再叠一份相同的 user 上下文。
+生成参数与 **prompt 条目 + marker 填洞** 均在 `wrapLlmForSession`（`PresetLlmProvider`）统一做：凡 LLM 请求先夹心再发。调用方只交任务 messages；任务 system / 创作上下文进 `worldBookBefore` 等洞，**禁止**在 `prompt_order` 之外再前置一条 system。无选用预设时原样转发。
 
 ## 6. 第一版导入策略
 

src/llm/preset-wrapper.ts

@@ -1,4 +1,5 @@
 import { loadAppSettings } from "../config/settings.js";
+import { frameMessagesWithPreset } from "../preset/play-frame.js";
 import { resolveActivePreset } from "../preset/store.js";
 import type { PresetPackage } from "../types/preset.js";
 import type {
@@ -23,7 +24,7 @@ function readActivePreset(): PresetPackage | null {
 
 /**
  * 统计在内、预设在外：痕迹里的 messages / generation 就是发给模型的那一份。
- * 当前没有选用预设时仍包一层，请求原样转发。
+ * 有选用预设时：在此按 prompt_order 夹心包裹，并合并 generation。
  */
 export function wrapLlmForSession(
   inner: LlmProvider,
@@ -36,10 +37,7 @@ export function wrapLlmForSession(
   return new PresetLlmProvider(tracked, getPreset);
 }
 
-/**
- * 在所有 LLM 请求上合并当前 preset 的生成参数。
- * prompt / marker 不在此前置，由 play worker 按扩展 prompt_order 填洞。
- */
+/** 所有 LLM 请求：preset 夹心 + 生成参数。无选用预设则原样转发。 */
 export class PresetLlmProvider implements LlmProvider {
   constructor(
     private readonly inner: LlmProvider,
@@ -52,9 +50,8 @@ export class PresetLlmProvider implements LlmProvider {
   ): { messages: ChatMessage[]; options: CompleteOptions | undefined } {
     const preset = this.getPreset();
     if (!preset) return { messages, options };
-    // 生成参数全局合并；prompt/marker 只在 play worker 按 prompt_order 填洞，这里不再前置。
     return {
-      messages,
+      messages: frameMessagesWithPreset(preset, messages),
       options: {
         ...options,
         generation: options?.generation ?? preset.generation,
@@ -97,6 +94,8 @@ export class PresetLlmProvider implements LlmProvider {
     return this.inner.completeWithTools(prepared.messages, {
       ...options,
       ...prepared.options,
+      tools: options.tools,
+      toolChoice: options.toolChoice,
     });
   }
 
@@ -111,7 +110,12 @@ export class PresetLlmProvider implements LlmProvider {
     const prepared = this.prepare(messages, options);
     return this.inner.completeWithToolsStream(
       prepared.messages,
-      { ...options, ...prepared.options },
+      {
+        ...options,
+        ...prepared.options,
+        tools: options.tools,
+        toolChoice: options.toolChoice,
+      },
       callbacks,
     );
   }

src/preset/play-frame.ts

@@ -4,12 +4,32 @@ import { assemblePresetMessages } from "./assembler.js";
 import { applyExtendedMarkers } from "./markers.js";
 import {
   resolveWorldInfoMarker,
+  type WorldInfoEntry,
   type WorldInfoPack,
 } from "./world-info-pack.js";
 
+const TASK_SYSTEM_ENTRY_ID = "task-system";
+const TASK_SYSTEM_ENTRY_NAME = "任务契约";
+
+function withTaskSystem(
+  pack: WorldInfoPack,
+  systemPrompt: string,
+): WorldInfoPack {
+  const text = systemPrompt.trim();
+  if (!text) return pack;
+  const entry: WorldInfoEntry = {
+    id: TASK_SYSTEM_ENTRY_ID,
+    name: TASK_SYSTEM_ENTRY_NAME,
+    content: text,
+  };
+  return {
+    ...pack,
+    worldBookBefore: [entry, ...pack.worldBookBefore],
+  };
+}
+
 /**
- * Play worker：system 契约 + 按扩展 prompt_order 填洞。
- * 不要再把同一份 context-order 正文叠成第二包 user。
+ * 严格按扩展 prompt_order 填洞。任务契约进 worldBookBefore，不在预设外另挂 system。
  */
 export function assemblePlayWorkerMessages(params: {
   systemPrompt: string;
@@ -17,8 +37,82 @@ export function assemblePlayWorkerMessages(params: {
   pack: WorldInfoPack;
 }): ChatMessage[] {
   const framed = applyExtendedMarkers(params.preset);
-  const filled = assemblePresetMessages(framed, (id) =>
-    resolveWorldInfoMarker(params.pack, id),
+  const pack = withTaskSystem(params.pack, params.systemPrompt);
+  return assemblePresetMessages(framed, (id) =>
+    resolveWorldInfoMarker(pack, id),
   );
-  return [{ role: "system", content: params.systemPrompt }, ...filled];
+}
+
+function isToolLoopBoundary(m: ChatMessage): boolean {
+  if (m.role === "tool") return true;
+  return Boolean(
+    m.role === "assistant" && m.tool_calls && m.tool_calls.length > 0,
+  );
+}
+
+function formatHistoryMessages(msgs: ChatMessage[]): string {
+  const parts: string[] = [];
+  for (const m of msgs) {
+    if (m.role === "user" && m.content.trim()) {
+      parts.push(`【用户】\n${m.content.trim()}`);
+    } else if (m.role === "assistant") {
+      const text = (m.content ?? "").trim();
+      if (text) parts.push(`【助手】\n${text}`);
+    } else if (m.role === "system" && m.content.trim()) {
+      parts.push(`【系统】\n${m.content.trim()}`);
+    }
+  }
+  return parts.join("\n\n");
+}
+
+/**
+ * 任务 messages → 只填四袋 + chatHistory；发出序列的首尾由预设 prompt_order 决定。
+ * tool 循环尾接在夹心后。
+ */
+export function frameMessagesWithPreset(
+  preset: PresetPackage,
+  messages: ChatMessage[],
+): ChatMessage[] {
+  let splitAt = messages.length;
+  for (let i = 0; i < messages.length; i++) {
+    if (isToolLoopBoundary(messages[i]!)) {
+      splitAt = i;
+      break;
+    }
+  }
+  const prefix = messages.slice(0, splitAt);
+  const continuation = messages.slice(splitAt);
+
+  const systems: string[] = [];
+  const body: ChatMessage[] = [];
+  for (const m of prefix) {
+    if (m.role === "system") systems.push(m.content);
+    else body.push(m);
+  }
+
+  let turn = "";
+  let history = "";
+  if (body.length > 0) {
+    const last = body[body.length - 1]!;
+    if (last.role === "user") {
+      turn = last.content.trim();
+      history = formatHistoryMessages(body.slice(0, -1));
+    } else {
+      history = formatHistoryMessages(body);
+    }
+  }
+
+  const framed = assemblePlayWorkerMessages({
+    systemPrompt: systems.join("\n\n"),
+    preset,
+    pack: {
+      worldBookBefore: [],
+      history,
+      worldBookAfter: [],
+      turn,
+      postTurn: [],
+    },
+  });
+
+  return continuation.length > 0 ? [...framed, ...continuation] : framed;
 }

src/preset/probe.ts

@@ -62,6 +62,7 @@ export function worldInfoPackFromProbeInput(
   };
 }
 
+/** 已夹心的试跑 messages（仅 UI / 对照）；真正发模型走任务消息 + PresetLlmProvider。 */
 export function assemblePresetProbe(
   preset: PresetPackage,
   input: PresetProbeInput,
@@ -73,6 +74,32 @@ export function assemblePresetProbe(
   });
 }
 
+/** 交给 wrapLlmForSession 的任务消息（未夹心）。 */
+export function buildPresetProbeTaskMessages(
+  input: PresetProbeInput,
+): ChatMessage[] {
+  const pack = worldInfoPackFromProbeInput(input);
+  const parts: string[] = [];
+  for (const e of pack.worldBookBefore) {
+    parts.push(`### ${e.name}\n\n${e.content}`);
+  }
+  if (pack.history.trim()) parts.push(pack.history.trim());
+  for (const e of pack.worldBookAfter) {
+    parts.push(`### ${e.name}\n\n${e.content}`);
+  }
+  for (const e of pack.postTurn) {
+    parts.push(`### ${e.name}\n\n${e.content}`);
+  }
+  const ctx = parts.join("\n\n").trim();
+  const user = ctx
+    ? `${ctx}\n\n## 本轮输入\n${pack.turn}`
+    : pack.turn;
+  return [
+    { role: "system", content: PROBE_SYSTEM_PROMPT },
+    { role: "user", content: user },
+  ];
+}
+
 export function serializeProbeMessages(
   messages: ChatMessage[],
 ): PresetProbeMessage[] {
@@ -94,11 +121,11 @@ export async function runPresetProbe(params: {
     throw new Error("缺少本轮输入");
   }
 
-  const messages = assemblePresetProbe(params.preset, {
-    ...params.input,
-    message,
-  });
-  const serialized = serializeProbeMessages(messages);
+  const input = { ...params.input, message };
+  // UI 仍展示「理想分袋夹心」；发模型只交任务消息，由统一包装层夹心
+  const serialized = serializeProbeMessages(
+    assemblePresetProbe(params.preset, input),
+  );
 
   if (!params.complete) {
     return {
@@ -111,7 +138,7 @@ export async function runPresetProbe(params: {
   }
 
   try {
-    const result = await params.complete(messages);
+    const result = await params.complete(buildPresetProbeTaskMessages(input));
     return {
       messages: serialized,
       reply: result.content,

tests/play-frame.test.ts

@@ -51,9 +51,11 @@ describe("assemblePlayWorkerMessages", () => {
     });
 
     expect(messages.map((m) => [m.role, m.content])).toEqual([
-      ["system", "SKILL"],
       ["system", "MAIN"],
-      ["system", "### 舞台\n\n古镇夜雨"],
+      [
+        "system",
+        "### 任务契约\n\nSKILL\n\n### 舞台\n\n古镇夜雨",
+      ],
       ["system", "用户：昨天来过"],
       ["system", "### 变量\n\nHP 3"],
       ["user", "我推门"],
@@ -62,6 +64,57 @@ describe("assemblePlayWorkerMessages", () => {
     ]);
   });
 
+  it("does not emit any message before the first preset order entry", () => {
+    const report = importSillyTavernPreset({
+      prompts: [
+        {
+          identifier: "wb-head",
+          name: "世界书首",
+          role: "system",
+          content: "<世界书>",
+        },
+        {
+          identifier: "chatHistory",
+          name: "History",
+          role: "system",
+          marker: true,
+        },
+        {
+          identifier: "wb-tail",
+          name: "世界书尾",
+          role: "system",
+          content: "</世界书>",
+        },
+      ],
+      prompt_order: [
+        {
+          character_id: 1,
+          order: [
+            { identifier: "wb-head", enabled: true },
+            { identifier: "chatHistory", enabled: true },
+            { identifier: "wb-tail", enabled: true },
+          ],
+        },
+      ],
+    });
+    const messages = assemblePlayWorkerMessages({
+      systemPrompt: "创作步骤契约",
+      preset: report.preset,
+      pack: {
+        worldBookBefore: [{ id: "lore", name: "设定", content: "港口" }],
+        history: "昨日对话",
+        worldBookAfter: [],
+        turn: "本轮",
+        postTurn: [],
+      },
+    });
+    expect(messages[0]).toEqual({ role: "system", content: "<世界书>" });
+    expect(messages.some((m) => m.content === "创作步骤契约")).toBe(false);
+    expect(
+      messages.some((m) => m.content.includes("### 任务契约\n\n创作步骤契约")),
+    ).toBe(true);
+  });
+
   it("does not fill ST charDescription even if resolver returns content", () => {
     const report = importSillyTavernPreset({
       prompts: [

tests/preset-probe.test.ts

@@ -46,9 +46,11 @@ describe("assemblePresetProbe", () => {
     });
 
     expect(messages.map((m) => [m.role, m.content])).toEqual([
-      ["system", PROBE_SYSTEM_PROMPT],
       ["system", "MAIN"],
-      ["system", `### 试跑设定\n\n${DEFAULT_PROBE_CONTEXT.loreBefore}`],
+      [
+        "system",
+        `### 任务契约\n\n${PROBE_SYSTEM_PROMPT}\n\n### 试跑设定\n\n${DEFAULT_PROBE_CONTEXT.loreBefore}`,
+      ],
       ["system", DEFAULT_PROBE_CONTEXT.history],
       ["system", `### 本轮状态\n\n${DEFAULT_PROBE_CONTEXT.loreAfter}`],
       ["user", "我推开门。"],
@@ -66,8 +68,8 @@ describe("assemblePresetProbe", () => {
     });
 
     expect(messages.map((m) => m.content)).toEqual([
-      PROBE_SYSTEM_PROMPT,
       "MAIN",
+      `### 任务契约\n\n${PROBE_SYSTEM_PROMPT}`,
       "你好",
       "PHI",
     ]);

tests/preset-wrapper.test.ts

@@ -53,7 +53,7 @@ function recordingProvider(): {
 }
 
 describe("wrapLlmForSession", () => {
-  it("context trace matches the exact messages and generation sent to the inner LLM", async () => {
+  it("frames task messages with preset prompt_order and merges generation", async () => {
     const { provider, received } = recordingProvider();
     const ctx: LlmTrackingContext = {};
     const llm = wrapLlmForSession(provider, { current: ctx }, () => samplePreset);
@@ -64,14 +64,15 @@ describe("wrapLlmForSession", () => {
     );
 
     expect(received).toHaveLength(1);
-    expect(received[0].messages).toEqual([{ role: "user", content: "TASK" }]);
+    expect(received[0].messages.map((m) => [m.role, m.content])).toEqual([
+      ["system", "PRESET_RULES"],
+      ["user", "TASK"],
+    ]);
     expect(received[0].generation).toMatchObject({
       temperature: 0.91,
       maxOutputTokens: 1234,
     });
-    expect(ctx.pendingContextTrace?.messages).toEqual([
-      { role: "user", content: "TASK" },
-    ]);
+    expect(ctx.pendingContextTrace?.messages).toEqual(received[0].messages);
     expect(ctx.pendingContextTrace?.generation).toMatchObject({
       temperature: 0.91,
       maxOutputTokens: 1234,
@@ -80,7 +81,44 @@ describe("wrapLlmForSession", () => {
     expect(ctx.pendingContextTrace?.model).toBe("rec-model");
   });
 
-  it("completeStream also records the merged request", async () => {
+  it("puts task system inside worldBookBefore, not before the preset shell", async () => {
+    const { provider, received } = recordingProvider();
+    const llm = wrapLlmForSession(provider, undefined, () => samplePreset);
+
+    await llm.completeWithTools(
+      [
+        { role: "system", content: "SKILL" },
+        { role: "user", content: "do it" },
+        {
+          role: "assistant",
+          content: null,
+          tool_calls: [
+            {
+              id: "1",
+              type: "function",
+              function: { name: "t", arguments: "{}" },
+            },
+          ],
+        },
+        { role: "tool", content: "{}", tool_call_id: "1" },
+      ],
+      { tools: [], caller: "dictate_agent" },
+    );
+
+    expect(received[0].messages[0].content).toBe("PRESET_RULES");
+    expect(
+      received[0].messages.some(
+        (m) => m.content === "### 任务契约\n\nSKILL",
+      ),
+    ).toBe(true);
+    expect(received[0].messages.some((m) => m.content === "do it")).toBe(true);
+    expect(received[0].messages.map((m) => m.role).slice(-2)).toEqual([
+      "assistant",
+      "tool",
+    ]);
+  });
+
+  it("completeStream also records the framed request", async () => {
     const { provider, received } = recordingProvider();
     const ctx: LlmTrackingContext = {};
     const llm = wrapLlmForSession(provider, { current: ctx }, () => samplePreset);
@@ -93,7 +131,9 @@ describe("wrapLlmForSession", () => {
     );
 
     expect(chunks.join("")).toBe("ok");
-    expect(received[0].messages).toEqual([{ role: "user", content: "TASK" }]);
+    expect(received[0].messages.some((m) => m.content === "PRESET_RULES")).toBe(
+      true,
+    );
     expect(ctx.pendingContextTrace?.messages).toEqual(received[0].messages);
   });
 });