Add gm harness with batch chance tools for play.

Main world layer can loop on the chance tool (requests array) until settlement JSON. Driver forwards output deltas; standalone chance worker accepts batch requests. play_slots.chance gates the harness.

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

src/worker/executor.ts

@@ -1,5 +1,5 @@
 import type { Blackboard } from "../blackboard/blackboard.js";
-import type { LlmProvider, StreamCallbacks } from "../llm/client.js";
+import type { LlmProvider, StreamCallbacks, ChatMessage } from "../llm/client.js";
 import { supportsContentStream } from "../llm/stream-complete.js";
 import { loadWorkerSkillWithContext } from "../skills/loader.js";
 import { filterInputsForRolePerspective } from "../skills/worker-llm.js";
@@ -21,6 +21,11 @@ import { resolveActivePreset } from "../preset/store.js";
 import { assemblePlayWorkerMessages } from "../preset/play-frame.js";
 import { worldInfoPackFromSegments } from "../preset/world-info-pack.js";
 import { PRESENT_TAG } from "../skills/present-packet.js";
+import {
+  gmChanceToolsEnabled,
+  GM_CHANCE_HARNESS_INSTRUCTION,
+} from "../skills/gm-tools.js";
+import { runGmHarness } from "./gm-harness.js";
 import {
   mergeQuestionsPreferFragment,
   normalizeQuestions,
@@ -38,6 +43,11 @@ import {
   serializeContextOrder,
 } from "../skills/context-order.js";
 
+export type WorkerStreamCallbacks = {
+  onThinkingDelta?: (delta: string) => void;
+  onOutputDelta?: (delta: string) => void;
+};
+
 export type WorkerRunParams = {
   skillName: string;
   workerId: string;
@@ -49,11 +59,6 @@ export type WorkerRunParams = {
   declared?: { worker: ParsedWorkerSkill; promptBody: string };
 };
 
-export type WorkerStreamCallbacks = {
-  onThinkingDelta?: (delta: string) => void;
-  onOutputDelta?: (delta: string) => void;
-};
-
 export type WorkerRunResult = {
   outputs: Record<string, string>;
   summary: string;
@@ -557,7 +562,19 @@ export async function runWorkerSkill(params: WorkerRunParams): Promise<WorkerRun
     inputs[CONTEXT_BRIEF_TAG] = priorBrief;
   }
 
-  const systemContent = promptBody + WORKER_OUTPUT_INSTRUCTION;
+  const useGmHarness =
+    isPlayLayerActive(params.slots) &&
+    worker.id === "world-simulator" &&
+    gmChanceToolsEnabled(
+      parseWorkerSetYaml(
+        params.blackboard.getContentByTag("设计.worker集") ?? "",
+      )?.play_slots,
+    );
+
+  const systemContent =
+    promptBody +
+    WORKER_OUTPUT_INSTRUCTION +
+    (useGmHarness ? GM_CHANCE_HARNESS_INSTRUCTION : "");
   const playPreset = isPlayLayerActive(params.slots)
     ? resolveActivePreset(loadAppSettings().activePresetId)
     : null;
@@ -589,6 +606,19 @@ export async function runWorkerSkill(params: WorkerRunParams): Promise<WorkerRun
         },
       ];
 
+  if (useGmHarness) {
+    const systemMsg = messages.find((m) => m.role === "system");
+    const rest = messages.filter((m) => m.role !== "system") as ChatMessage[];
+    const harness = await runGmHarness({
+      llm: params.llm,
+      system: systemMsg?.content ?? systemContent,
+      messages: rest,
+      stream: params.stream,
+      caller: `worker:${worker.id}`,
+    });
+    return parseWorkerResponse(harness.content, worker.outputTags);
+  }
+
   const result = await completeWorkerPreferStream(
     params.llm,
     messages,

src/worker/gm-harness.ts

@@ -0,0 +1,84 @@
+/**
+ * 主世界层（world-simulator）游玩期 tool harness：
+ * 反复 completeWithTools 直到不再调用工具，最后一次正文视为裁决 JSON。
+ */
+import type { ChatMessage, LlmProvider } from "../llm/client.js";
+import {
+  createLocalLlmDriver,
+  type DriverToolCall,
+} from "../runtime/driver.js";
+import {
+  GM_CHANCE_TOOL_NAME,
+  GM_TOOL_DEFINITIONS,
+  handleGmToolCall,
+} from "../skills/gm-tools.js";
+
+export type GmHarnessStreamCallbacks = {
+  onThinkingDelta?: (delta: string) => void;
+  onOutputDelta?: (delta: string) => void;
+};
+
+export const DEFAULT_GM_HARNESS_MAX_ITERATIONS = 6;
+
+export type GmHarnessParams = {
+  llm: LlmProvider;
+  system: string;
+  messages: ChatMessage[];
+  stream?: GmHarnessStreamCallbacks;
+  maxIterations?: number;
+  caller?: string;
+};
+
+export type GmHarnessResult = {
+  content: string;
+  iterations: number;
+};
+
+export async function runGmHarness(
+  params: GmHarnessParams,
+): Promise<GmHarnessResult> {
+  const driver = createLocalLlmDriver(params.llm);
+  const label = "主世界层 harness";
+
+  const result = await driver.run({
+    system: params.system,
+    messages: params.messages,
+    tools: GM_TOOL_DEFINITIONS,
+    caller: params.caller ?? "worker:world-simulator",
+    label,
+    maxIterations: params.maxIterations ?? DEFAULT_GM_HARNESS_MAX_ITERATIONS,
+    onThinkingDelta: (delta) => params.stream?.onThinkingDelta?.(delta),
+    onOutputDelta: (delta) => params.stream?.onOutputDelta?.(delta),
+    onThinkingDone: () => {
+      /* 思维链由 stream 增量展示 */
+    },
+    handleStep: (calls: DriverToolCall[]) => {
+      const results: Array<{ callId: string; content: string }> = [];
+      for (const call of calls) {
+        if (call.name !== GM_CHANCE_TOOL_NAME) {
+          results.push({
+            callId: call.id,
+            content: JSON.stringify({
+              ok: false,
+              error: `不允许的工具：${call.name}`,
+            }),
+          });
+          continue;
+        }
+        const content =
+          handleGmToolCall(call.name, call.arguments) ??
+          JSON.stringify({ ok: false, error: "未知工具" });
+        results.push({ callId: call.id, content });
+      }
+      return { kind: "continue", results };
+    },
+  });
+
+  if (result.stop.kind === "text") {
+    return { content: result.stop.content, iterations: result.iterations };
+  }
+
+  throw new Error(
+    `${label} 以边界 tool 结束（${result.stop.call.name}），主世界层仅允许 ${GM_CHANCE_TOOL_NAME}`,
+  );
+}