引入游玩回合与开场收口，抽出可替换 agent 司机并支持运行中止；完善预设洞位、世界书拼装与试跑，并打磨会话运行时与 Web 交互。

Co-authored-by: Cursor <cursoragent@cursor.com>
2026-08-24 01:13:48 +08:00
parent ec474cb946
commit b82da4c5a5
95 changed files with 6781 additions and 1344 deletions
@@ -1,6 +1,10 @@
 import type { LlmConfig } from "../config/env.js";
 import type { GenerationParameters } from "../types/preset.js";
 import { consumeOpenAiToolStream } from "./stream-complete.js";
+import {
+  currentAbortSignal,
+  throwIfAborted,
+} from "./run-abort.js";

 export type ToolCallPayload = {
  id: string;
@@ -58,6 +62,8 @@ export type CompleteOptions = {
  generation?: GenerationParameters;
  /** 统计用途，如 main_agent / worker:write-rules */
  caller?: string;
+  /** 取消进行中的 LLM 请求（停止并重试） */
+  signal?: AbortSignal;
 };

 export type CompleteWithToolsOptions = CompleteOptions & {
@@ -237,6 +243,24 @@ function extractMessageParts(message: Record<string, unknown> | undefined): {
  return { content: "", toolCalls };
 }

+async function llmFetch(
+  url: string,
+  apiKey: string,
+  body: unknown,
+  options?: CompleteOptions,
+): Promise<Response> {
+  throwIfAborted(options?.signal);
+  return fetch(url, {
+    method: "POST",
+    headers: {
+      "Content-Type": "application/json",
+      Authorization: `Bearer ${apiKey}`,
+    },
+    body: JSON.stringify(body),
+    signal: options?.signal ?? currentAbortSignal(),
+  });
+}
+
 export class OpenAiCompatibleProvider implements LlmProvider {
  constructor(private readonly config: LlmConfig) {}

@@ -245,14 +269,12 @@ export class OpenAiCompatibleProvider implements LlmProvider {
    options?: CompleteOptions,
  ): Promise<CompleteResult> {
    const url = `${this.config.baseUrl.replace(/\/$/, "")}/chat/completions`;
-    const response = await fetch(url, {
-      method: "POST",
-      headers: {
-        "Content-Type": "application/json",
-        Authorization: `Bearer ${this.config.apiKey}`,
-      },
-      body: JSON.stringify(buildRequestBody(this.config, messages, options)),
-    });
+    const response = await llmFetch(
+      url,
+      this.config.apiKey,
+      buildRequestBody(this.config, messages, options),
+      options,
+    );

    if (!response.ok) {
      const body = await response.text();
@@ -282,16 +304,12 @@ export class OpenAiCompatibleProvider implements LlmProvider {
    callbacks: StreamCallbacks = {},
  ): Promise<CompleteResult> {
    const url = `${this.config.baseUrl.replace(/\/$/, "")}/chat/completions`;
-    const response = await fetch(url, {
-      method: "POST",
-      headers: {
-        "Content-Type": "application/json",
-        Authorization: `Bearer ${this.config.apiKey}`,
-      },
-      body: JSON.stringify(
-        buildRequestBody(this.config, messages, { ...options, stream: true }),
-      ),
-    });
+    const response = await llmFetch(
+      url,
+      this.config.apiKey,
+      buildRequestBody(this.config, messages, { ...options, stream: true }),
+      options,
+    );

    if (!response.ok) {
      const body = await response.text();
@@ -319,19 +337,15 @@ export class OpenAiCompatibleProvider implements LlmProvider {
    options: CompleteWithToolsOptions,
  ): Promise<CompleteWithToolsResult> {
    const url = `${this.config.baseUrl.replace(/\/$/, "")}/chat/completions`;
-    const response = await fetch(url, {
-      method: "POST",
-      headers: {
-        "Content-Type": "application/json",
-        Authorization: `Bearer ${this.config.apiKey}`,
-      },
-      body: JSON.stringify(
-        buildRequestBody(this.config, messages, {
-          ...options,
-          tools: options.tools,
-        }),
-      ),
-    });
+    const response = await llmFetch(
+      url,
+      this.config.apiKey,
+      buildRequestBody(this.config, messages, {
+        ...options,
+        tools: options.tools,
+      }),
+      options,
+    );

    if (!response.ok) {
      const body = await response.text();
@@ -362,20 +376,16 @@ export class OpenAiCompatibleProvider implements LlmProvider {
    callbacks: StreamCallbacks,
  ): Promise<CompleteWithToolsResult> {
    const url = `${this.config.baseUrl.replace(/\/$/, "")}/chat/completions`;
-    const response = await fetch(url, {
-      method: "POST",
-      headers: {
-        "Content-Type": "application/json",
-        Authorization: `Bearer ${this.config.apiKey}`,
-      },
-      body: JSON.stringify(
-        buildRequestBody(this.config, messages, {
-          ...options,
-          tools: options.tools,
-          stream: true,
-        }),
-      ),
-    });
+    const response = await llmFetch(
+      url,
+      this.config.apiKey,
+      buildRequestBody(this.config, messages, {
+        ...options,
+        tools: options.tools,
+        stream: true,
+      }),
+      options,
+    );

    if (!response.ok) {
      const body = await response.text();
@@ -425,6 +435,7 @@ export class MockLlmProvider implements LlmProvider {
    _messages: ChatMessage[],
    options?: CompleteOptions,
  ): Promise<CompleteResult> {
+    throwIfAborted(options?.signal);
    const step = this.nextStep();
    const response =
      typeof step === "string"
@@ -448,6 +459,7 @@ export class MockLlmProvider implements LlmProvider {
    options?: CompleteOptions,
    callbacks: StreamCallbacks = {},
  ): Promise<CompleteResult> {
+    throwIfAborted(options?.signal);
    const step = this.nextStep();
    const response =
      typeof step === "string"
@@ -455,10 +467,12 @@ export class MockLlmProvider implements LlmProvider {
        : (step.content ?? JSON.stringify({ action: "ask_user", reason: "mock" }));
    const reasoning = "用户需要明确分工 → 调用 design-intake 产出 worker 集。";
    for (const ch of reasoning) {
+      throwIfAborted(options?.signal);
      callbacks.onReasoningDelta?.(ch);
      await new Promise((r) => setTimeout(r, 0));
    }
    for (const ch of response) {
+      throwIfAborted(options?.signal);
      callbacks.onContentDelta?.(ch);
      await new Promise((r) => setTimeout(r, 0));
    }
@@ -480,6 +494,7 @@ export class MockLlmProvider implements LlmProvider {
    _messages: ChatMessage[],
    _options: CompleteWithToolsOptions,
  ): Promise<CompleteWithToolsResult> {
+    throwIfAborted(_options?.signal);
    const step = this.nextStep();
    if (typeof step === "string") {
      return {
@@ -507,9 +522,11 @@ export class MockLlmProvider implements LlmProvider {
    _options: CompleteWithToolsOptions,
    callbacks: StreamCallbacks,
  ): Promise<CompleteWithToolsResult> {
+    throwIfAborted(_options?.signal);
    const reasoning =
      "用户需要明确 Worker 分工 → 先读取黑板与 worker 列表 → 调用 design-intake 产出 worker 集。";
    for (const ch of reasoning) {
+      throwIfAborted(_options?.signal);
      callbacks.onReasoningDelta?.(ch);
      await new Promise((r) => setTimeout(r, 0));
    }
@@ -1,4 +1,5 @@
-import { assemblePresetMessages, mergeMessages } from "../preset/assembler.js";
+import { loadAppSettings } from "../config/settings.js";
+import { resolveActivePreset } from "../preset/store.js";
 import type { PresetPackage } from "../types/preset.js";
 import type {
  ChatMessage,
@@ -9,9 +10,35 @@ import type {
  LlmProvider,
  StreamCallbacks,
 } from "./client.js";
+import {
+  TokenTrackingProvider,
+  type LlmTrackingContext,
+} from "./token-tracker.js";
+
+export type LlmTrackingRef = { current: LlmTrackingContext };
+
+function readActivePreset(): PresetPackage | null {
+  return resolveActivePreset(loadAppSettings().activePresetId);
+}

 /**
- * 在所有 LLM 请求前注入当前 preset 的 prompt 片段与生成参数。
+ * 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。
+ */
+export function wrapLlmForSession(
+  inner: LlmProvider,
+  trackingRef?: LlmTrackingRef,
+  getPreset: () => PresetPackage | null = readActivePreset,
+): LlmProvider {
+  const tracked = trackingRef
+    ? new TokenTrackingProvider(inner, () => trackingRef.current)
+    : inner;
+  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。
 */
 export class PresetLlmProvider implements LlmProvider {
  constructor(
@@ -19,41 +46,57 @@ export class PresetLlmProvider implements LlmProvider {
    private readonly getPreset: () => PresetPackage | null,
  ) {}

+  private prepare(
+    messages: ChatMessage[],
+    options?: CompleteOptions,
+  ): { 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,
+      options: {
+        ...options,
+        generation: options?.generation ?? preset.generation,
+      },
+    };
+  }
+
  async complete(
    messages: ChatMessage[],
    options?: CompleteOptions,
  ): Promise<CompleteResult> {
-    const preset = this.getPreset();
-    if (!preset) {
-      return this.inner.complete(messages, options);
+    const prepared = this.prepare(messages, options);
+    return this.inner.complete(prepared.messages, prepared.options);
+  }
+
+  async completeStream(
+    messages: ChatMessage[],
+    options?: CompleteOptions,
+    callbacks: StreamCallbacks = {},
+  ): Promise<CompleteResult> {
+    if (!this.inner.completeStream) {
+      const result = await this.complete(messages, options);
+      if (result.reasoning) callbacks.onReasoningDelta?.(result.reasoning);
+      if (result.content) callbacks.onContentDelta?.(result.content);
+      return result;
    }
-
-    const presetMessages = assemblePresetMessages(preset);
-    const merged = mergeMessages(presetMessages, messages);
-    const generation = options?.generation ?? preset.generation;
-
-    return this.inner.complete(merged, {
-      ...options,
-      generation,
-    });
+    const prepared = this.prepare(messages, options);
+    return this.inner.completeStream(
+      prepared.messages,
+      prepared.options,
+      callbacks,
+    );
  }

  async completeWithTools(
    messages: ChatMessage[],
    options: CompleteWithToolsOptions,
  ): Promise<CompleteWithToolsResult> {
-    const preset = this.getPreset();
-    if (!preset) {
-      return this.inner.completeWithTools(messages, options);
-    }
-
-    const presetMessages = assemblePresetMessages(preset);
-    const merged = mergeMessages(presetMessages, messages);
-    const generation = options.generation ?? preset.generation;
-
-    return this.inner.completeWithTools(merged, {
+    const prepared = this.prepare(messages, options);
+    return this.inner.completeWithTools(prepared.messages, {
      ...options,
-      generation,
+      ...prepared.options,
    });
  }

@@ -62,17 +105,14 @@ export class PresetLlmProvider implements LlmProvider {
    options: CompleteWithToolsOptions,
    callbacks: StreamCallbacks,
  ): Promise<CompleteWithToolsResult> {
-    const preset = this.getPreset();
    if (!this.inner.completeWithToolsStream) {
      return this.completeWithTools(messages, options);
    }
-    const presetMessages = preset ? assemblePresetMessages(preset) : [];
-    const merged = preset ? mergeMessages(presetMessages, messages) : messages;
-    const generation = options.generation ?? preset?.generation;
-
-    return this.inner.completeWithToolsStream(merged, {
-      ...options,
-      generation,
-    }, callbacks);
+    const prepared = this.prepare(messages, options);
+    return this.inner.completeWithToolsStream(
+      prepared.messages,
+      { ...options, ...prepared.options },
+      callbacks,
+    );
  }
 }
@@ -0,0 +1,30 @@
+import { AsyncLocalStorage } from "node:async_hooks";
+
+const storage = new AsyncLocalStorage<AbortSignal>();
+
+export function runWithAbortSignal<T>(
+  signal: AbortSignal,
+  fn: () => Promise<T>,
+): Promise<T> {
+  return storage.run(signal, fn);
+}
+
+export function currentAbortSignal(): AbortSignal | undefined {
+  return storage.getStore();
+}
+
+export function isAbortError(err: unknown): boolean {
+  if (!err || typeof err !== "object") return false;
+  const name = "name" in err ? String(err.name) : "";
+  if (name === "AbortError") return true;
+  const message = "message" in err ? String(err.message) : "";
+  return /aborted|AbortError/i.test(message);
+}
+
+export function throwIfAborted(signal?: AbortSignal): void {
+  const active = signal ?? currentAbortSignal();
+  if (!active?.aborted) return;
+  const err = new Error("This operation was aborted");
+  err.name = "AbortError";
+  throw err;
+}
@@ -32,14 +32,21 @@ export type LlmTrackingContext = {

 function capturePendingTrace(
  ctx: LlmTrackingContext,
-  messages: Array<{ role: string; content: string }>,
+  messages: Array<{
+    role: string;
+    content?: string | null;
+    tool_calls?: unknown;
+    tool_call_id?: string;
+  }>,
  caller: string | undefined,
  model?: string,
+  generation?: CompleteOptions["generation"],
 ): void {
  ctx.pendingContextTrace = buildContextTrace({
    caller: caller ?? "unknown",
    messages,
    model,
+    generation,
  });
 }

@@ -130,7 +137,13 @@ export class TokenTrackingProvider implements LlmProvider {
      async () => {
        const result = await this.inner.complete(messages, options);
        const ctx = this.getContext();
-        capturePendingTrace(ctx, messages, options?.caller, result.model);
+        capturePendingTrace(
+          ctx,
+          messages,
+          options?.caller,
+          result.model,
+          options?.generation,
+        );
        recordUsage(ctx, options?.caller, result.model, result.usage);
        if (result.reasoning?.trim()) {
          ctx.pendingReasoning = result.reasoning.trim();
@@ -168,7 +181,13 @@ export class TokenTrackingProvider implements LlmProvider {
          onContentDelta: callbacks.onContentDelta,
        });
        const ctx = this.getContext();
-        capturePendingTrace(ctx, messages, options?.caller, result.model);
+        capturePendingTrace(
+          ctx,
+          messages,
+          options?.caller,
+          result.model,
+          options?.generation,
+        );
        recordUsage(ctx, options?.caller, result.model, result.usage);
        if (result.reasoning?.trim()) {
          ctx.pendingReasoning = result.reasoning.trim();
@@ -189,7 +208,13 @@ export class TokenTrackingProvider implements LlmProvider {
      async () => {
        const result = await this.inner.completeWithTools(messages, options);
        const ctx = this.getContext();
-        capturePendingTrace(ctx, messages, options.caller, result.model);
+        capturePendingTrace(
+          ctx,
+          messages,
+          options.caller,
+          result.model,
+          options.generation,
+        );
        recordUsage(ctx, options.caller, result.model, result.usage);
        if (result.reasoning?.trim()) {
          ctx.pendingReasoning = result.reasoning.trim();
@@ -224,7 +249,13 @@ export class TokenTrackingProvider implements LlmProvider {
          onContentDelta: callbacks.onContentDelta,
        });
        const ctx = this.getContext();
-        capturePendingTrace(ctx, messages, options.caller, result.model);
+        capturePendingTrace(
+          ctx,
+          messages,
+          options.caller,
+          result.model,
+          options.generation,
+        );
        recordUsage(ctx, options.caller, result.model, result.usage);
        if (result.reasoning?.trim()) {
          ctx.pendingReasoning = result.reasoning.trim();