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>

src/llm/client.ts

@@ -57,8 +57,33 @@ export type TokenUsage = {
   cachedTokens?: number;
   /** Prompt tokens not served from cache (DeepSeek prompt_cache_miss_tokens) */
   cacheMissTokens?: number;
+  /** 上游若单独返回缓存写入 token（如 cache_creation_input_tokens） */
+  cacheWriteTokens?: number;
 };
 
+/** 用量行上的调度信息。不包含提示词或产物正文。 */
+export type DictateUsageStep = "select" | "fill" | "fill-retry" | "reply" | "play";
+
+export type UsageDetail = {
+  turnId?: string;
+  step?: DictateUsageStep;
+  retryIndex?: number;
+  segmentChars?: {
+    rules: number;
+    history: number;
+    products: number;
+    injected: number;
+    tail: number;
+  };
+  plannedTags?: string[];
+  succeededTags?: string[];
+  failedTags?: string[];
+  finish?: string;
+};
+
+/** complete 返回后由统计层写入 recordId，供补记成功或失败的 tag。 */
+export type UsageSink = { recordId?: string };
+
 export type CompleteResult = {
   content: string;
   /** 推理模型思维链（如 DeepSeek reasoner 的 reasoning_content） */
@@ -84,6 +109,14 @@ export type CompleteOptions = {
   label?: string;
   /** 取消进行中的 LLM 请求（停止并重试） */
   signal?: AbortSignal;
+  /**
+   * 本局会话 id。仅 OpenRouter 会写成请求体 `session_id`，用来粘住同一上游，
+   * 避免开头 message 一变就换缓存节点。其它渠道不发送。
+   */
+  sessionId?: string;
+  /** 统计补充。不进上游请求体。 */
+  usageDetail?: UsageDetail;
+  usageSink?: UsageSink;
 };
 
 export type ToolChoice =
@@ -132,6 +165,33 @@ export type LlmProvider = {
 };
 
 const JSON_OBJECT_HINT = "请用 JSON 对象回复。";
+const OPENROUTER_SESSION_ID_MAX = 256;
+
+/** OpenRouter 用 `session_id` 做粘滞路由；其它 OpenAI 兼容渠道不认识这个字段。 */
+export function isOpenRouterBaseUrl(baseUrl: string): boolean {
+  const trimmed = baseUrl.trim();
+  if (!trimmed) return false;
+  try {
+    const withProtocol = /^https?:\/\//i.test(trimmed)
+      ? trimmed
+      : `https://${trimmed}`;
+    const host = new URL(withProtocol).hostname.toLowerCase();
+    return host === "openrouter.ai" || host.endsWith(".openrouter.ai");
+  } catch {
+    return false;
+  }
+}
+
+/** 合法则返回要写入请求体的 session_id；非 OpenRouter 或空 id 则不写。 */
+export function openRouterSessionId(
+  baseUrl: string,
+  sessionId: string | undefined,
+): string | undefined {
+  if (!isOpenRouterBaseUrl(baseUrl)) return undefined;
+  const id = sessionId?.trim();
+  if (!id) return undefined;
+  return id.slice(0, OPENROUTER_SESSION_ID_MAX);
+}
 
 function messagesMentionJson(messages: ChatMessage[]): boolean {
   for (const message of messages) {
@@ -210,6 +270,9 @@ export function buildRequestBody(
     body.stream_options = { include_usage: true };
   }
 
+  const sessionId = openRouterSessionId(config.baseUrl, options?.sessionId);
+  if (sessionId) body.session_id = sessionId;
+
   return body;
 }
 
@@ -232,6 +295,20 @@ function readCacheMissTokens(u: Record<string, unknown>): number | undefined {
   return readFiniteNumber(u.prompt_cache_miss_tokens ?? u.promptCacheMissTokens);
 }
 
+function readCacheWriteTokens(u: Record<string, unknown>): number | undefined {
+  const details = u.prompt_tokens_details ?? u.promptTokensDetails;
+  if (details && typeof details === "object") {
+    const row = details as Record<string, unknown>;
+    const nested = readFiniteNumber(row.cache_write_tokens ?? row.cacheWriteTokens);
+    if (nested != null) return nested;
+  }
+  return readFiniteNumber(
+    u.cache_creation_input_tokens ??
+      u.cache_write_input_tokens ??
+      u.cacheWriteTokens,
+  );
+}
+
 /** Parse OpenAI-compatible usage object, including provider-specific cache fields. */
 export function parseUsage(raw: unknown): TokenUsage | undefined {
   if (!raw || typeof raw !== "object") return undefined;
@@ -243,6 +320,7 @@ export function parseUsage(raw: unknown): TokenUsage | undefined {
 
   const cachedTokens = readCachedTokens(u);
   const cacheMissTokens = readCacheMissTokens(u);
+  const cacheWriteTokens = readCacheWriteTokens(u);
 
   return {
     promptTokens: Number.isFinite(prompt) ? prompt : 0,
@@ -253,22 +331,66 @@ export function parseUsage(raw: unknown): TokenUsage | undefined {
         (Number.isFinite(completion) ? completion : 0),
     ...(cachedTokens != null ? { cachedTokens } : {}),
     ...(cacheMissTokens != null ? { cacheMissTokens } : {}),
+    ...(cacheWriteTokens != null ? { cacheWriteTokens } : {}),
   };
 }
 
-function extractMessageParts(message: Record<string, unknown> | undefined): {
+function stringifyToolArguments(value: unknown): string {
+  if (typeof value === "string" && value.trim()) return value;
+  if (value && typeof value === "object") return JSON.stringify(value);
+  return "{}";
+}
+
+function toolCallFromRow(row: Record<string, unknown>, index: number): ParsedToolCall | null {
+  const fn = row.function;
+  const nested = fn && typeof fn === "object" ? (fn as Record<string, unknown>) : null;
+  const source = nested ?? row;
+  const name = typeof source.name === "string" ? source.name.trim() : "";
+  if (!name) return null;
+  const rawId = typeof row.id === "string" ? row.id.trim() : "";
+  return {
+    id: rawId || `call-${index}`,
+    name,
+    arguments: stringifyToolArguments(
+      source.arguments ?? source.args ?? row.arguments ?? row.args ?? row.input,
+    ),
+  };
+}
+
+function embeddedContent(raw: unknown): { text: string | null; calls: ParsedToolCall[] } {
+  if (typeof raw === "string") return { text: raw.trim() || null, calls: [] };
+  if (!Array.isArray(raw)) return { text: raw == null ? null : "", calls: [] };
+  const texts: string[] = [];
+  const calls: ParsedToolCall[] = [];
+  raw.forEach((part, index) => {
+    if (!part || typeof part !== "object") return;
+    const row = part as Record<string, unknown>;
+    if (typeof row.text === "string" && row.text.trim()) texts.push(row.text.trim());
+    const fn = row.functionCall ?? row.function_call;
+    if (row.type === "tool_use" || (fn && typeof fn === "object")) {
+      const nested = fn && typeof fn === "object" ? (fn as Record<string, unknown>) : row;
+      const parsed = toolCallFromRow(
+        {
+          id: row.id,
+          name: nested.name ?? row.name,
+          arguments: nested.args ?? nested.arguments ?? row.input ?? row.arguments,
+        },
+        index,
+      );
+      if (parsed) calls.push(parsed);
+    }
+  });
+  return { text: texts.join("\n").trim() || null, calls };
+}
+
+/** 从一条 assistant message 取出正文和工具调用。对象参数、缺 id、content 片段里的调用都收。 */
+export function extractMessageParts(message: Record<string, unknown> | undefined): {
   content: string | null;
   reasoning?: string;
   toolCalls: ParsedToolCall[];
 } {
   if (!message) return { content: "", toolCalls: [] };
-  const rawContent = message.content;
+  const embedded = embeddedContent(message.content);
-  const content =
-    typeof rawContent === "string"
-      ? rawContent.trim() || null
-      : rawContent == null
-        ? null
-        : "";
   const reasoningRaw =
     (typeof message.reasoning_content === "string" &&
       message.reasoning_content.trim()) ||
@@ -280,21 +402,18 @@ function extractMessageParts(message: Record<string, unknown> | undefined): {
   const toolCalls: ParsedToolCall[] = [];
   const rawCalls = message.tool_calls;
   if (Array.isArray(rawCalls)) {
-    for (const call of rawCalls) {
+    rawCalls.forEach((call, index) => {
-      if (!call || typeof call !== "object") continue;
+      if (!call || typeof call !== "object") return;
-      const c = call as Record<string, unknown>;
+      const parsed = toolCallFromRow(call as Record<string, unknown>, index);
-      const fn = c.function;
+      if (parsed) toolCalls.push(parsed);
-      if (!fn || typeof fn !== "object") continue;
+    });
-      const f = fn as Record<string, unknown>;
-      const name = typeof f.name === "string" ? f.name : "";
-      const id = typeof c.id === "string" ? c.id : "";
-      const args =
-        typeof f.arguments === "string" ? f.arguments : "{}";
-      if (name && id) {
-        toolCalls.push({ id, name, arguments: args });
   }
+  for (const call of embedded.calls) {
+    if (!toolCalls.some((item) => item.id === call.id && item.name === call.name)) {
+      toolCalls.push(call);
     }
   }
+  const content = embedded.text;
 
   if (toolCalls.length > 0) {
     return { content, reasoning: reasoning || undefined, toolCalls };

src/llm/context-outline.ts

@@ -0,0 +1,181 @@
+import { createHash } from "node:crypto";
+import type { ToolDefinition } from "./client.js";
+
+/** 一段请求的位置标签。只存标签、字数和哈希，不存正文。 */
+export type ContextSegment = {
+  tag: string;
+  chars: number;
+  hash: string;
+};
+
+/** 挂在一条回复上的各次请求轮廓，选题和填写都保留。 */
+export type RequestOutline = {
+  at: string;
+  caller: string;
+  step?: string;
+  turnId?: string;
+  model?: string;
+  promptTokens?: number;
+  completionTokens?: number;
+  cachedTokens?: number;
+  cacheMissTokens?: number;
+  cacheWriteTokens?: number;
+  segments: ContextSegment[];
+};
+
+/** 和同会话上一条请求比，前缀从哪里开始对不上。 */
+export type CacheBreakInfo = {
+  /** 本窗口里没有更早的同会话请求 */
+  cold: boolean;
+  /** -1：本条在上一请求的前缀里整段对上 */
+  breakAt: number;
+  tag?: string;
+  /** 断点之前、和上一请求逐段相同的段数 */
+  matched: number;
+};
+
+const SPLIT =
+  /(?:^|\n)(#{2,4} [^\n]+|最后一轮用户输入|【用户】|【助手】)/g;
+
+function hashText(text: string): string {
+  return createHash("sha256").update(text).digest("hex").slice(0, 12);
+}
+
+function segment(tag: string, text: string): ContextSegment {
+  return { tag, chars: text.length, hash: hashText(text) };
+}
+
+function headingTag(title: string): string {
+  const beforeOrder = title.split("〔")[0]?.trim() ?? title;
+  const beforeParen = beforeOrder.split(/（|\(/)[0]?.trim() ?? beforeOrder;
+  const beforeAt = beforeParen.split(" @ ")[0]?.trim() ?? beforeParen;
+  const tag = (beforeAt || title).trim();
+  return tag.length > 48 ? `${tag.slice(0, 48)}…` : tag;
+}
+
+function collectMarks(content: string): Array<{ index: number; tag: string }> {
+  const marks: Array<{ index: number; tag: string }> = [];
+  let section = "";
+  for (const match of content.matchAll(SPLIT)) {
+    const raw = match[1] ?? "";
+    const start = (match.index ?? 0) + (match[0].startsWith("\n") ? 1 : 0);
+    if (raw.startsWith("#")) {
+      const level = /^#+/.exec(raw)?.[0].length ?? 2;
+      const title = raw.replace(/^#+\s*/, "");
+      const tag = headingTag(title);
+      if (level === 2) section = tag;
+      marks.push({ index: start, tag });
+      continue;
+    }
+    if (raw === "最后一轮用户输入" && section.startsWith("本轮")) {
+      marks.push({ index: start, tag: "最后一轮用户输入" });
+      continue;
+    }
+    if (
+      (raw === "【用户】" || raw === "【助手】") &&
+      section.startsWith("更早的对话")
+    ) {
+      marks.push({
+        index: start,
+        tag: raw === "【用户】" ? "更早·用户" : "更早·助手",
+      });
+    }
+  }
+  return marks;
+}
+
+/** 把一块正文切成有序标签。前缀里的标题与更早对话轮次分开，方便对上断点。 */
+export function outlineBlock(fallbackTag: string, content: string): ContextSegment[] {
+  if (!content) return [];
+  const marks = collectMarks(content);
+  if (marks.length === 0) return [segment(fallbackTag, content)];
+  const out: ContextSegment[] = [];
+  if (marks[0]!.index > 0) {
+    const head = content.slice(0, marks[0]!.index);
+    if (head.trim()) out.push(segment(fallbackTag, head));
+  }
+  for (let i = 0; i < marks.length; i++) {
+    const end = i + 1 < marks.length ? marks[i + 1]!.index : content.length;
+    out.push(segment(marks[i]!.tag, content.slice(marks[i]!.index, end)));
+  }
+  return out;
+}
+
+function roleTag(role: string): string {
+  if (role === "system") return "规则";
+  if (role === "user") return "用户";
+  if (role === "assistant") return "助手";
+  if (role === "tool") return "工具结果";
+  return role || "消息";
+}
+
+/** 按实际送出的消息切段。工具定义若有，放在正文之前。 */
+export function outlineRequest(
+  messages: Array<{ role: string; content?: string | null }>,
+  tools?: ToolDefinition[],
+): ContextSegment[] {
+  const out: ContextSegment[] = [];
+  if (tools?.length) {
+    out.push(segment("工具", JSON.stringify(tools)));
+  }
+  for (const message of messages) {
+    const content = typeof message.content === "string" ? message.content : "";
+    out.push(...outlineBlock(roleTag(message.role), content));
+  }
+  return out;
+}
+
+/**
+ * 返回下一请求里第一段对不上的下标。
+ * 下一请求比上一请求更短、但已有段全部相同，视为整段接上（-1）。
+ */
+export function firstCacheBreak(
+  prev: Array<Pick<ContextSegment, "tag" | "hash">>,
+  next: Array<Pick<ContextSegment, "tag" | "hash">>,
+): number {
+  const shared = Math.min(prev.length, next.length);
+  for (let i = 0; i < shared; i++) {
+    if (prev[i]!.tag !== next[i]!.tag || prev[i]!.hash !== next[i]!.hash) return i;
+  }
+  if (next.length > prev.length) return prev.length;
+  return -1;
+}
+
+export function annotateCacheBreaks<
+  T extends { id: string; at: string; sessionId?: string; segments?: ContextSegment[] },
+>(records: T[]): Map<string, CacheBreakInfo> {
+  const bySession = new Map<string, T[]>();
+  for (const record of records) {
+    const key = record.sessionId?.trim() || `__none__:${record.id}`;
+    const list = bySession.get(key) ?? [];
+    list.push(record);
+    bySession.set(key, list);
+  }
+  const marks = new Map<string, CacheBreakInfo>();
+  for (const list of bySession.values()) {
+    list.sort((a, b) => a.at.localeCompare(b.at) || a.id.localeCompare(b.id));
+    let prev: ContextSegment[] | undefined;
+    for (const record of list) {
+      const segments = record.segments ?? [];
+      if (!prev) {
+        marks.set(record.id, {
+          cold: true,
+          breakAt: segments.length ? 0 : -1,
+          tag: segments[0]?.tag,
+          matched: 0,
+        });
+        prev = segments;
+        continue;
+      }
+      const breakAt = firstCacheBreak(prev, segments);
+      marks.set(record.id, {
+        cold: false,
+        breakAt,
+        tag: breakAt >= 0 ? segments[breakAt]?.tag : undefined,
+        matched: breakAt < 0 ? segments.length : breakAt,
+      });
+      prev = segments;
+    }
+  }
+  return marks;
+}

src/llm/stream-complete.ts

@@ -29,6 +29,9 @@ function applyToolCallDelta(
       const f = fn as Record<string, unknown>;
       if (typeof f.name === "string") entry.name = f.name;
       if (typeof f.arguments === "string") entry.arguments += f.arguments;
+      else if (f.arguments && typeof f.arguments === "object" && !entry.arguments) {
+        entry.arguments = JSON.stringify(f.arguments);
+      }
     }
     acc.set(index, entry);
   }
@@ -36,10 +39,10 @@ function applyToolCallDelta(
 
 function toolCallsFromAccumulator(acc: ToolCallAccumulator): ParsedToolCall[] {
   const out: ParsedToolCall[] = [];
-  for (const [, entry] of [...acc.entries()].sort((a, b) => a[0] - b[0])) {
+  for (const [index, entry] of [...acc.entries()].sort((a, b) => a[0] - b[0])) {
     const name = entry.name?.trim();
-    const id = entry.id?.trim();
+    if (!name) continue;
-    if (!name || !id) continue;
+    const id = entry.id?.trim() || `call-${index}`;
     out.push({ id, name, arguments: entry.arguments || "{}" });
   }
   return out;

src/llm/token-tracker.ts

@@ -6,7 +6,15 @@ import type {
   LlmProvider,
   StreamCallbacks,
   TokenUsage,
+  ToolDefinition,
+  UsageDetail,
+  UsageSink,
 } from "./client.js";
+import {
+  outlineRequest,
+  type ContextSegment,
+  type RequestOutline,
+} from "./context-outline.js";
 import {
   recordTokenUsage,
   toMessageTokenUsage,
@@ -28,6 +36,9 @@ export type LlmTrackingContext = {
   pendingReasoning?: string;
   /** 全量请求上下文；挂到下一条「结果向」系统消息 */
   pendingContextTrace?: LlmContextTrace;
+  /** 本条回复之前的各次请求轮廓；查看时选题和填写都在 */
+  pendingOutlines?: RequestOutline[];
+  lastSegments?: ContextSegment[];
 };
 
 function capturePendingTrace(
@@ -41,12 +52,16 @@ function capturePendingTrace(
   caller: string | undefined,
   model?: string,
   generation?: CompleteOptions["generation"],
+  tools?: ToolDefinition[],
 ): void {
+  const segments = outlineRequest(messages, tools);
+  ctx.lastSegments = segments;
   ctx.pendingContextTrace = buildContextTrace({
     caller: caller ?? "unknown",
     messages,
     model,
     generation,
+    segments,
   });
 }
 
@@ -55,9 +70,13 @@ function recordUsage(
   caller: string | undefined,
   model: string | undefined,
   usage: TokenUsage | undefined,
+  meta?: { detail?: UsageDetail; latencyMs?: number; sink?: UsageSink },
 ): void {
-  if (!usage) return;
+  const segments = ctx.lastSegments;
-  const record = recordTokenUsage({
+  ctx.lastSegments = undefined;
+  const detail = meta?.detail;
+  const record = usage
+    ? recordTokenUsage({
     sessionId: ctx.sessionId,
     bookId: ctx.bookId,
     bookTitle: ctx.bookTitle,
@@ -69,8 +88,36 @@ function recordUsage(
     totalTokens: usage.totalTokens,
     cachedTokens: usage.cachedTokens,
     cacheMissTokens: usage.cacheMissTokens,
-  });
+    cacheWriteTokens: usage.cacheWriteTokens,
-  ctx.pendingUsage = toMessageTokenUsage(record);
+    turnId: detail?.turnId,
+    step: detail?.step,
+    retryIndex: detail?.retryIndex,
+    segmentChars: detail?.segmentChars,
+    plannedTags: detail?.plannedTags,
+    succeededTags: detail?.succeededTags,
+    failedTags: detail?.failedTags,
+    finish: detail?.finish,
+    latencyMs: meta?.latencyMs,
+    segments,
+  })
+    : undefined;
+  if (record && meta?.sink) meta.sink.recordId = record.id;
+  if (record) ctx.pendingUsage = toMessageTokenUsage(record);
+  if (!segments?.length && !record) return;
+  const outline: RequestOutline = {
+    at: record?.at ?? new Date().toISOString(),
+    caller: record?.caller ?? caller ?? "unknown",
+    step: detail?.step,
+    turnId: detail?.turnId,
+    model: record?.model ?? model,
+    promptTokens: record?.promptTokens,
+    completionTokens: record?.completionTokens,
+    cachedTokens: record?.cachedTokens,
+    cacheMissTokens: record?.cacheMissTokens,
+    cacheWriteTokens: record?.cacheWriteTokens,
+    segments: segments ?? [],
+  };
+  ctx.pendingOutlines = [...(ctx.pendingOutlines ?? []), outline];
 }
 
 export class TokenTrackingProvider implements LlmProvider {
@@ -79,6 +126,14 @@ export class TokenTrackingProvider implements LlmProvider {
     private readonly getContext: () => LlmTrackingContext,
   ) {}
 
+  /** 调用方未指定时，用本局会话 id，供 OpenRouter 粘滞路由。 */
+  private withSession<T extends CompleteOptions>(options?: T): T {
+    const sessionId = this.getContext().sessionId?.trim();
+    if (!sessionId) return (options ?? {}) as T;
+    if (options?.sessionId?.trim()) return options;
+    return { ...options, sessionId } as T;
+  }
+
   async complete(
     messages: Parameters<LlmProvider["complete"]>[0],
     options?: CompleteOptions,
@@ -91,7 +146,8 @@ export class TokenTrackingProvider implements LlmProvider {
         messages,
       },
       async () => {
-        const result = await this.inner.complete(messages, options);
+        const started = Date.now();
+        const result = await this.inner.complete(messages, this.withSession(options));
         const ctx = this.getContext();
         capturePendingTrace(
           ctx,
@@ -100,7 +156,11 @@ export class TokenTrackingProvider implements LlmProvider {
           result.model,
           options?.generation,
         );
-        recordUsage(ctx, options?.caller, result.model, result.usage);
+        recordUsage(ctx, options?.caller, result.model, result.usage, {
+          detail: options?.usageDetail,
+          latencyMs: Date.now() - started,
+          sink: options?.usageSink,
+        });
         if (result.reasoning?.trim()) {
           ctx.pendingReasoning = result.reasoning.trim();
         }
@@ -130,7 +190,8 @@ export class TokenTrackingProvider implements LlmProvider {
       },
       async () => {
         let reasoningBuf = "";
-        const result = await inner.completeStream!(messages, options, {
+        const started = Date.now();
+        const result = await inner.completeStream!(messages, this.withSession(options), {
           onReasoningDelta: (delta) => {
             reasoningBuf += delta;
             const ctx = this.getContext();
@@ -147,7 +208,11 @@ export class TokenTrackingProvider implements LlmProvider {
           result.model,
           options?.generation,
         );
-        recordUsage(ctx, options?.caller, result.model, result.usage);
+        recordUsage(ctx, options?.caller, result.model, result.usage, {
+          detail: options?.usageDetail,
+          latencyMs: Date.now() - started,
+          sink: options?.usageSink,
+        });
         if (result.reasoning?.trim()) {
           ctx.pendingReasoning = result.reasoning.trim();
         }
@@ -169,7 +234,11 @@ export class TokenTrackingProvider implements LlmProvider {
         tools: options.tools,
       },
       async () => {
-        const result = await this.inner.completeWithTools(messages, options);
+        const started = Date.now();
+        const result = await this.inner.completeWithTools(
+          messages,
+          this.withSession(options),
+        );
         const ctx = this.getContext();
         capturePendingTrace(
           ctx,
@@ -177,8 +246,13 @@ export class TokenTrackingProvider implements LlmProvider {
           options.caller,
           result.model,
           options.generation,
+          options.tools,
         );
-        recordUsage(ctx, options.caller, result.model, result.usage);
+        recordUsage(ctx, options.caller, result.model, result.usage, {
+          detail: options.usageDetail,
+          latencyMs: Date.now() - started,
+          sink: options.usageSink,
+        });
         if (result.reasoning?.trim()) {
           ctx.pendingReasoning = result.reasoning.trim();
         }
@@ -206,7 +280,11 @@ export class TokenTrackingProvider implements LlmProvider {
       },
       async () => {
         let reasoningBuf = "";
-        const result = await inner.completeWithToolsStream!(messages, options, {
+        const started = Date.now();
+        const result = await inner.completeWithToolsStream!(
+          messages,
+          this.withSession(options),
+          {
             onReasoningDelta: (delta) => {
               reasoningBuf += delta;
               const ctx = this.getContext();
@@ -214,7 +292,8 @@ export class TokenTrackingProvider implements LlmProvider {
               callbacks.onReasoningDelta?.(delta);
             },
             onContentDelta: callbacks.onContentDelta,
-        });
+          },
+        );
         const ctx = this.getContext();
         capturePendingTrace(
           ctx,
@@ -222,8 +301,13 @@ export class TokenTrackingProvider implements LlmProvider {
           options.caller,
           result.model,
           options.generation,
+          options.tools,
         );
-        recordUsage(ctx, options.caller, result.model, result.usage);
+        recordUsage(ctx, options.caller, result.model, result.usage, {
+          detail: options.usageDetail,
+          latencyMs: Date.now() - started,
+          sink: options.usageSink,
+        });
         if (result.reasoning?.trim()) {
           ctx.pendingReasoning = result.reasoning.trim();
         }

src/log.ts

@@ -98,6 +98,8 @@ const TOOL_LABEL: Record<string, string> = {
   review_blackboard: "审黑板",
   finish: "结束",
   insert: "写入产物",
+  reply_user: "回复用户",
+  select_nodes: "选定节点",
   write_product: "写入产物",
   delete: "删除产物",
   declare_variable: "声明变量",
@@ -214,6 +216,7 @@ export function labelCaller(caller: string | undefined): string {
   if (who === "main_agent" || who.startsWith("main_agent:")) return "总管";
   if (who === "intake_extract") return "需求抽取";
   if (who === "dictate_agent") return "对话落盘";
+  if (who.startsWith("listen_serial_")) return "听书连载";
   if (who === "preset-probe") return "预设探测";
   if (who.startsWith("model-compare:")) {
     return `模型对比 ${who.slice("model-compare:".length)}`;

src/server/stats-handlers.ts

@@ -14,6 +14,7 @@ import {
   SUGGESTED_PRICING_DOCS,
 } from "../stats/pricing.js";
 import { attachRecordCost, buildCostReport } from "../stats/pricing-cost.js";
+import { annotateCacheBreaks } from "../llm/context-outline.js";
 
 async function readBody(req: IncomingMessage): Promise<string> {
   const chunks: Buffer[] = [];
@@ -77,6 +78,7 @@ export async function handleStatsApi(
     const records = readRecords({ ...query, limit: query.limit ?? 100 });
     const catalog = resolveRateCatalog();
     const cost = buildCostReport(summaryRecords, catalog);
+    const breaks = annotateCacheBreaks(summaryRecords);
     json(res, 200, {
       summary: {
         ...summarizeRecords(summaryRecords),
@@ -92,6 +94,7 @@ export async function handleStatsApi(
           ...r,
           estimatedCost: priced.cost,
           unmatchedRate: priced.unmatched,
+          cacheBreak: breaks.get(r.id) ?? null,
         };
       }),
     });

src/stats/token-store.ts

@@ -1,7 +1,8 @@
-import { appendFileSync, existsSync, mkdirSync, readFileSync, readdirSync } from "node:fs";
+import { appendFileSync, existsSync, mkdirSync, readFileSync, readdirSync, writeFileSync } from "node:fs";
 import path from "node:path";
 import { randomUUID } from "node:crypto";
 import { ensureUserDataDirs, getUserDataDir } from "../config/user-data-dir.js";
+import type { ContextSegment } from "../llm/context-outline.js";
 
 export type TokenUsageRecord = {
   id: string;
@@ -17,7 +18,25 @@ export type TokenUsageRecord = {
   totalTokens: number;
   cachedTokens?: number;
   cacheMissTokens?: number;
+  cacheWriteTokens?: number;
   messageId?: string;
+  turnId?: string;
+  step?: string;
+  retryIndex?: number;
+  segmentChars?: {
+    rules: number;
+    history: number;
+    products: number;
+    injected: number;
+    tail: number;
+  };
+  /** 有序标签与哈希，不含正文 */
+  segments?: ContextSegment[];
+  plannedTags?: string[];
+  succeededTags?: string[];
+  failedTags?: string[];
+  finish?: string;
+  latencyMs?: number;
 };
 
 /** Token usage attached to a chat message in SessionView */
@@ -97,7 +116,18 @@ export function recordTokenUsage(
     totalTokens: partial.totalTokens,
     cachedTokens: partial.cachedTokens,
     cacheMissTokens: partial.cacheMissTokens,
+    cacheWriteTokens: partial.cacheWriteTokens,
     messageId: partial.messageId,
+    turnId: partial.turnId,
+    step: partial.step,
+    retryIndex: partial.retryIndex,
+    segmentChars: partial.segmentChars,
+    segments: partial.segments,
+    plannedTags: partial.plannedTags,
+    succeededTags: partial.succeededTags,
+    failedTags: partial.failedTags,
+    finish: partial.finish,
+    latencyMs: partial.latencyMs,
   };
   appendFileSync(dayFile(new Date(record.at)), `${JSON.stringify(record)}\n`, "utf8");
   return record;
@@ -238,6 +268,39 @@ export function getSessionTokenTotals(sessionId: string): {
   };
 }
 
+export function patchTokenUsage(
+  id: string,
+  patch: Partial<
+    Pick<
+      TokenUsageRecord,
+      "plannedTags" | "succeededTags" | "failedTags" | "finish"
+    >
+  >,
+): void {
+  if (!id) return;
+  const files = listStatFiles();
+  for (let i = files.length - 1; i >= 0; i--) {
+    const file = files[i]!;
+    const lines = readFileSync(file, "utf8").split(/\r?\n/);
+    let changed = false;
+    const next = lines.map((line) => {
+      if (!line.trim()) return line;
+      try {
+        const record = JSON.parse(line) as TokenUsageRecord;
+        if (record.id !== id) return line;
+        changed = true;
+        return JSON.stringify({ ...record, ...patch });
+      } catch {
+        return line;
+      }
+    });
+    if (changed) {
+      writeFileSync(file, next.join("\n"), "utf8");
+      return;
+    }
+  }
+}
+
 export function toMessageTokenUsage(record: TokenUsageRecord): MessageTokenUsage {
   return {
     totalTokens: record.totalTokens,

src/types/context-trace.ts

@@ -1,3 +1,4 @@
+import type { ContextSegment } from "../llm/context-outline.js";
 import type { GenerationParameters } from "./preset.js";
 
 /** 一次 LLM 调用实际发送的上下文（全量，供右键观察） */
@@ -11,6 +12,8 @@ export type LlmContextTrace = {
   caller: string;
   createdAt: string;
   messages: LlmContextMessage[];
+  /** 有序标签，供查看排列和缓存断点；不含正文 */
+  segments?: ContextSegment[];
   /** 便于列表展示 */
   charCount: number;
   model?: string;
@@ -57,6 +60,7 @@ export function buildContextTrace(params: {
   }>;
   model?: string;
   generation?: GenerationParameters;
+  segments?: ContextSegment[];
 }): LlmContextTrace {
   const messages = snapshotRequestMessages(params.messages);
   const charCount = messages.reduce((n, m) => n + m.content.length, 0);
@@ -64,6 +68,7 @@ export function buildContextTrace(params: {
     caller: params.caller,
     createdAt: new Date().toISOString(),
     messages,
+    ...(params.segments?.length ? { segments: params.segments } : {}),
     charCount,
     model: params.model,
     ...(params.generation && Object.keys(params.generation).length

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,

web/stats.css

@@ -396,6 +396,26 @@
   border-bottom: none;
 }
 
+.stats-table tbody tr.stats-call {
+  cursor: pointer;
+}
+.stats-table tbody tr.stats-detail td {
+  padding-top: 8px;
+  padding-bottom: 12px;
+  white-space: normal;
+}
+.stats-table .stats-system-note {
+  margin: 0 0 6px;
+  color: var(--st-muted, var(--muted, #6b7280));
+  font-size: 12px;
+}
+.stats-table .stats-segments {
+  margin: 8px 0 0;
+  font-size: 12px;
+  line-height: 1.45;
+  white-space: pre-wrap;
+  word-break: break-word;
+}
 .stats-table tbody tr:hover td {
   background: color-mix(in srgb, var(--st-surface) 58%, var(--st-bg));
 }

web/stats.html

@@ -117,12 +117,13 @@
             <thead>
               <tr>
                 <th scope="col">时间</th>
+                <th scope="col">步骤</th>
                 <th scope="col">caller</th>
                 <th scope="col">model</th>
                 <th scope="col" class="num">prompt</th>
-                <th scope="col" class="num">completion</th>
+                <th scope="col" class="num">未缓存</th>
-                <th scope="col" class="num">total</th>
                 <th scope="col" class="num">cached</th>
+                <th scope="col">断开</th>
                 <th scope="col" class="num">花销</th>
               </tr>
             </thead>

web/stats.js

@@ -187,6 +187,57 @@ function renderModels(rows) {
     .join("");
 }
 
+function uncachedTokens(record) {
+  if (record.cacheMissTokens != null) return record.cacheMissTokens;
+  if (record.promptTokens != null && record.cachedTokens != null) {
+    return Math.max(0, record.promptTokens - record.cachedTokens);
+  }
+  return null;
+}
+
+function breakLabel(info) {
+  if (!info) return "—";
+  if (info.cold) return "冷启动";
+  if (info.breakAt < 0) return "整段接上";
+  return info.tag || "断开";
+}
+
+function breakNote(record) {
+  const info = record.cacheBreak;
+  const segments = record.segments || [];
+  if (!segments.length) return "这次没有分段标签。";
+  if (!info || info.cold) {
+    return "本页没有更早的同会话请求可比。";
+  }
+  if (info.breakAt < 0) {
+    let note = "与上一条同会话请求的前缀整段相同。";
+    if (!record.cachedTokens) note += " 账单缓存是 0。";
+    return note;
+  }
+  const tag = segments[info.breakAt]?.tag || info.tag || "（上一条更长）";
+  let note = `前 ${info.matched ?? info.breakAt} 段与上一条相同，从「${tag}」起前缀不再一致。`;
+  if (!record.cachedTokens && info.breakAt > 0) {
+    note += " 账单缓存是 0。";
+  }
+  return note;
+}
+
+function segmentLines(record) {
+  const info = record.cacheBreak;
+  const segments = record.segments || [];
+  return segments
+    .map((seg, i) => {
+      let mark = "    ";
+      if (info && !info.cold) {
+        if (info.breakAt < 0 || i < info.breakAt) mark = "已接上";
+        else if (i === info.breakAt) mark = "断开  ";
+        else mark = "之后  ";
+      }
+      return `${String(i + 1).padStart(2, "0")}  ${mark}  ${escapeHtml(seg.tag)}  ${formatNum(seg.chars)}字  ${escapeHtml(seg.hash)}`;
+    })
+    .join("\n");
+}
+
 function renderRows(records) {
   const tbody = document.getElementById("rows");
   const empty = document.getElementById("empty");
@@ -206,15 +257,35 @@ function renderRows(records) {
       const cost = r.unmatchedRate || !r.estimatedCost
         ? `<span class="stats-unpriced">—</span>`
         : formatMoney(r.estimatedCost);
-      return `<tr>
+      const chars = r.segmentChars;
+      const charLine = chars
+        ? `规则 ${formatNum(chars.rules)} · 更早对话 ${formatNum(chars.history)} · 产物 ${formatNum(chars.products)} · 注入 ${formatNum(chars.injected)} · 末尾 ${formatNum(chars.tail)}`
+        : "";
+      const tags = [
+        r.turnId ? `turn ${escapeHtml(r.turnId)}` : "",
+        r.plannedTags?.length ? `计划 ${escapeHtml(r.plannedTags.join("、"))}` : "",
+        r.succeededTags?.length ? `已写入 ${escapeHtml(r.succeededTags.join("、"))}` : "",
+        r.failedTags?.length ? `未完成 ${escapeHtml(r.failedTags.join("、"))}` : "",
+      ].filter(Boolean).join(" · ");
+      return `<tr class="stats-call" data-call="${escapeHtml(r.id)}">
       <td class="col-time">${formatAt(r.at)}</td>
+      <td>${escapeHtml(r.step || "—")}</td>
       <td class="col-caller" title="${escapeHtml(r.caller || "")}">${escapeHtml(r.caller || "—")}</td>
       <td class="col-model" title="${escapeHtml(r.model || "")}">${escapeHtml(r.model || "—")}</td>
       <td class="num">${formatNum(r.promptTokens ?? 0)}</td>
-      <td class="num">${formatNum(r.completionTokens ?? 0)}</td>
+      <td class="num">${uncachedTokens(r) == null ? "—" : formatNum(uncachedTokens(r))}</td>
-      <td class="num col-total">${formatNum(r.totalTokens ?? 0)}</td>
       <td class="num">${r.cachedTokens == null ? "—" : formatNum(r.cachedTokens)}</td>
+      <td class="col-break" title="相对本页更早的同会话请求">${escapeHtml(breakLabel(r.cacheBreak))}</td>
       <td class="num">${cost}</td>
+    </tr>
+    <tr class="stats-detail" hidden>
+      <td colspan="9">
+        <p class="stats-system-note">系统信息：断点、分段标签和哈希只给你对照，不会发给模型。「工具」是 tools 字段。</p>
+        <p>${escapeHtml(breakNote(r))} 写出 ${formatNum(r.completionTokens ?? 0)}。</p>
+        ${charLine ? `<p>${charLine}</p>` : ""}
+        ${tags ? `<p>${tags}</p>` : ""}
+        <pre class="stats-segments">${segmentLines(r) || "（无分段）"}</pre>
+      </td>
     </tr>`;
     })
     .join("");
@@ -394,6 +465,14 @@ document.getElementById("pricing-sources").addEventListener("click", async (ev)
   }
 });
 
+document.getElementById("rows")?.addEventListener("click", (event) => {
+  const row = event.target.closest("tr.stats-call");
+  if (!row) return;
+  const detail = row.nextElementSibling;
+  if (!detail?.classList.contains("stats-detail")) return;
+  detail.hidden = !detail.hidden;
+});
+
 (() => {
   const now = new Date();
   const start = new Date(now.getFullYear(), now.getMonth(), now.getDate(), 0, 0, 0);