补齐缓存写量与请求轮廓记账，便于核对提示缓存是否真的命中。

给用量行挂上轮次步骤和分段字数，统计页能看出缓存断在哪。

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

tests/context-outline.test.ts

@@ -0,0 +1,114 @@
+import { describe, expect, it } from "vitest";
+import {
+  annotateCacheBreaks,
+  firstCacheBreak,
+  outlineRequest,
+} from "../src/llm/context-outline.js";
+
+const sample = [
+  "规则开头",
+  "",
+  "## 写入规则",
+  "固定规则",
+  "",
+  "## 更早的对话",
+  "",
+  "【用户】",
+  "第一句",
+  "",
+  "【助手】",
+  "回过",
+  "",
+  "## 当前产物（已写入黑板）",
+  "",
+  "### 设计.氛围 〔order=0〕",
+  "雨",
+  "",
+  "### 设计.角色 〔order=1〕",
+  "甲",
+  "",
+  "## 本轮注入",
+  "",
+  "### topology @ 3",
+  "#### task",
+  "做一件事",
+  "",
+  "## 本轮",
+  "",
+  "方案进度",
+  "",
+  "最后一轮用户输入",
+  "再改一下",
+].join("\n");
+
+describe("outlineRequest", () => {
+  it("按标题、对话轮次和最后一轮输入切开，不留下正文", () => {
+    const segments = outlineRequest([
+      { role: "system", content: sample.slice(0, sample.indexOf("## 更早的对话")) },
+      { role: "user", content: sample.slice(sample.indexOf("## 更早的对话")) },
+    ]);
+    expect(segments.map((item) => item.tag)).toEqual([
+      "规则",
+      "写入规则",
+      "更早的对话",
+      "更早·用户",
+      "更早·助手",
+      "当前产物",
+      "设计.氛围",
+      "设计.角色",
+      "本轮注入",
+      "topology",
+      "task",
+      "本轮",
+      "最后一轮用户输入",
+    ]);
+    expect(JSON.stringify(segments)).not.toContain("再改一下");
+    expect(JSON.stringify(segments)).not.toContain("雨");
+    const again = outlineRequest([
+      { role: "system", content: sample.slice(0, sample.indexOf("## 更早的对话")) },
+      { role: "user", content: sample.slice(sample.indexOf("## 更早的对话")) },
+    ]);
+    expect(again.map((item) => item.hash)).toEqual(segments.map((item) => item.hash));
+  });
+
+  it("工具定义放在正文之前", () => {
+    const segments = outlineRequest(
+      [{ role: "user", content: "你好" }],
+      [
+        {
+          type: "function",
+          function: { name: "insert", description: "写", parameters: {} },
+        },
+      ],
+    );
+    expect(segments.map((item) => item.tag)).toEqual(["工具", "用户"]);
+  });
+});
+
+describe("firstCacheBreak", () => {
+  it("标出第一处标签或哈希不同的位置", () => {
+    const prev = outlineRequest([{ role: "user", content: sample.slice(sample.indexOf("## 更早的对话")) }]);
+    const changed = sample.replace("雨", "雪");
+    const next = outlineRequest([{ role: "user", content: changed.slice(changed.indexOf("## 更早的对话")) }]);
+    const index = firstCacheBreak(prev, next);
+    expect(next[index]?.tag).toBe("设计.氛围");
+    expect(firstCacheBreak(prev, prev)).toBe(-1);
+    expect(firstCacheBreak(prev, prev.slice(0, 3))).toBe(-1);
+    expect(firstCacheBreak(prev.slice(0, 3), prev)).toBe(3);
+  });
+});
+
+describe("annotateCacheBreaks", () => {
+  it("按同会话时间比较，窗口里第一条算冷启动", () => {
+    const a = outlineRequest([{ role: "user", content: "第一" }]);
+    const b = outlineRequest([{ role: "user", content: "第二" }]);
+    const marks = annotateCacheBreaks([
+      { id: "2", at: "2026-10-02T02:00:00.000Z", sessionId: "s", segments: b },
+      { id: "1", at: "2026-10-02T01:00:00.000Z", sessionId: "s", segments: a },
+      { id: "9", at: "2026-10-02T03:00:00.000Z", sessionId: "other", segments: a },
+    ]);
+    expect(marks.get("1")).toMatchObject({ cold: true, matched: 0 });
+    expect(marks.get("2")).toMatchObject({ cold: false, breakAt: 0, tag: "用户", matched: 0 });
+    expect(marks.get("9")?.cold).toBe(true);
+  });
+});

tests/token-usage.test.ts

@@ -44,6 +44,16 @@ describe("parseUsage", () => {
     expect(usage?.cacheMissTokens).toBe(100);
   });
 
+  it("parses cache write tokens when the provider returns them", () => {
+    const usage = parseUsage({
+      prompt_tokens: 100,
+      completion_tokens: 10,
+      total_tokens: 110,
+      cache_creation_input_tokens: 40,
+    });
+    expect(usage?.cacheWriteTokens).toBe(40);
+  });
+
   it("returns undefined for empty usage", () => {
     expect(parseUsage(null)).toBeUndefined();
     expect(parseUsage({})).toBeUndefined();
@@ -52,8 +62,10 @@ describe("parseUsage", () => {
 
 describe("TokenTrackingProvider", () => {
   it("stores pending usage with cache fields on context", async () => {
+    let seenSession: string | undefined;
     const inner: LlmProvider = {
-      async complete(): Promise<CompleteResult> {
+      async complete(_messages, options): Promise<CompleteResult> {
+        seenSession = options?.sessionId;
         return {
           content: "ok",
           model: "test-model",
@@ -71,6 +83,7 @@ describe("TokenTrackingProvider", () => {
     const ctx: LlmTrackingContext = { sessionId: "sess-1" };
     const provider = new TokenTrackingProvider(inner, () => ctx);
     await provider.complete([], { caller: "worker:write-rules" });
+    expect(seenSession).toBe("sess-1");
 
     expect(ctx.pendingUsage).toMatchObject({
       totalTokens: 15,