feat: 使用正式用量完善上下文与会话统计

src/eval/gaia/__tests__/pricing.test.ts

@@ -107,6 +107,30 @@ describe("calculateCost", () => {
       calculateCost("claude-sonnet-4-6", { input: 0, output: 0, total: 0 }),
     ).toBe(0);
   });
+
+  it("prices provider-reported cache reads and writes separately", () => {
+    const cost = calculateCost("claude-sonnet-4-6", {
+      input: 100_000,
+      output: 20_000,
+      cacheRead: 300_000,
+      cacheWrite: 10_000,
+      total: 430_000,
+    });
+
+    expect(cost).toBeCloseTo(0.7275, 6);
+  });
+
+  it("uses the official long-context tier for GPT-5.5", () => {
+    const cost = calculateCost("gpt-5.5", {
+      input: 100_000,
+      output: 10_000,
+      cacheRead: 200_000,
+      cacheWrite: 0,
+      total: 310_000,
+    });
+
+    expect(cost).toBeCloseTo(1.65, 6);
+  });
 });
 
 // ─── formatCost ──────────────────────────────────────────────────────

src/eval/gaia/pricing.ts

@@ -1,127 +1,26 @@
-/**
- * 用于 GAIA 成本核算的模型价格表。
- *
- * 价格单位为每百万 token 的美元数,取自各服务商在 PR 创建时的
- * 官方定价页。它们一定会漂移 —— 当某服务商重新调整价格档位时,
- * 需更新此表,或在汇总中将成本字段置为 `null`。
- *
- * 未知 / 未定价的模型会让 `calculateCost` 返回 `null`,以便
- * 下游调用方能区分"无法定价"与"$0"。
- */
-
+import { calculateModelUsageCost } from "../../shared/model-pricing";
 import type { TokenUsage } from "./types";
 
-export interface ModelPrice {
-  /** 每百万输入 token 的美元价(未命中缓存)。 */
-  inputUsdPerMTok: number;
-  /** 每百万输出 token 的美元价。 */
-  outputUsdPerMTok: number;
-  /** 每百万缓存读取 token 的美元价。Anthropic / OpenAI 提示缓存。 */
-  cacheReadUsdPerMTok?: number;
-  /** 每百万缓存写入 token 的美元价。 */
-  cacheWriteUsdPerMTok?: number;
-}
-
-/**
- * 剥除服务商特定的日期 / 区域后缀,使诸如
- * `claude-sonnet-4-6-20251022` 的配置映射到规范的 `claude-sonnet-4-6`
- * 条目,无需为每个带日期的快照维护一行。
- */
-export const normalizeModelId = (id: string): string =>
-  id
-    // `-YYYYMMDD` 后缀
-    .replace(/-(2\d{3}\d{4})$/i, "")
-    // Anthropic `-latest` 别名
-    .replace(/-latest$/i, "");
-
-/**
- * 静态价格表。键为规范化(剥除日期)后的模型 id。
- * 当有新模型加入项目的适配器列表时,在此新增一行。
- */
-export const MODEL_PRICES: Readonly<Record<string, ModelPrice>> = Object.freeze(
-  {
-    // Anthropic Claude 4.x 系列
-    "claude-opus-4-7": {
-      inputUsdPerMTok: 15,
-      outputUsdPerMTok: 75,
-      cacheReadUsdPerMTok: 1.5,
-      cacheWriteUsdPerMTok: 18.75,
-    },
-    "claude-sonnet-4-6": {
-      inputUsdPerMTok: 3,
-      outputUsdPerMTok: 15,
-      cacheReadUsdPerMTok: 0.3,
-      cacheWriteUsdPerMTok: 3.75,
-    },
-    "claude-sonnet-4-7": {
-      inputUsdPerMTok: 3,
-      outputUsdPerMTok: 15,
-      cacheReadUsdPerMTok: 0.3,
-      cacheWriteUsdPerMTok: 3.75,
-    },
-    "claude-haiku-4-5": {
-      inputUsdPerMTok: 1,
-      outputUsdPerMTok: 5,
-      cacheReadUsdPerMTok: 0.1,
-      cacheWriteUsdPerMTok: 1.25,
-    },
-    // OpenAI
-    "gpt-4o": { inputUsdPerMTok: 2.5, outputUsdPerMTok: 10 },
-    "gpt-4o-mini": { inputUsdPerMTok: 0.15, outputUsdPerMTok: 0.6 },
-    // DeepSeek
-    "deepseek-chat": { inputUsdPerMTok: 0.14, outputUsdPerMTok: 0.28 },
-    "deepseek-reasoner": { inputUsdPerMTok: 0.55, outputUsdPerMTok: 2.19 },
-    // MiniMax
-    "abab6.5s-chat": { inputUsdPerMTok: 1, outputUsdPerMTok: 1 },
-    "minimax-text-01": { inputUsdPerMTok: 1, outputUsdPerMTok: 1 },
-    // Xiaomi MiMo V2.5 系列（2026-05-27 调价公告生效）。官方报价为 CNY，
-    // 按 $1 ≈ ¥7 折算入表；漂移以后只需在此处调汇率：
-    //   - V2.5-Pro：输入 ¥3 / 输出 ¥6 / 命中缓存 ¥0.025（per MTok）
-    //   - V2.5：    输入 ¥1 / 输出 ¥2 / 命中缓存 ¥0.02 （per MTok）
-    // V2.1-TTS 限时免费、V2 系列官方建议迁移到 V2.5，均未录入。
-    // calculateCost() 只用 input/output 两个字段；cache 字段记录在此供
-    // 未来按 cache 命中拆账时复用。MiMo 不区分 cache write（首次写入按
-    // 普通 input 计费），故 cacheWriteUsdPerMTok 留空。
-    "mimo-v2.5-pro": {
-      inputUsdPerMTok: 0.43,
-      outputUsdPerMTok: 0.86,
-      cacheReadUsdPerMTok: 0.0036,
-    },
-    "mimo-v2.5": {
-      inputUsdPerMTok: 0.14,
-      outputUsdPerMTok: 0.29,
-      cacheReadUsdPerMTok: 0.0029,
-    },
-  },
-);
-
-/**
- * 返回 `model` 对应的规范价格行,无对应条目时返回 `null`。
- * 适用于希望提示"该模型未定价 —— 请更新价格表"告警的工具。
- */
-export const getModelPrice = (model: string): ModelPrice | null =>
-  MODEL_PRICES[normalizeModelId(model)] ?? null;
+export {
+  getModelPrice,
+  MODEL_PRICES,
+  type ModelPrice,
+  normalizeModelId,
+} from "../../shared/model-pricing";
 
-/**
- * 计算单次 agent 运行的美元成本,使用基础的输入/输出定价
- * (不区分缓存差额 —— 调用方在 `TokenUsage` 中并未拆分出这些)。
- *
- * 当模型不在 `MODEL_PRICES` 中时返回 `null`,使 runner 可以记录
- * "未定价",而非错误地按 $0 计算。
- */
 export const calculateCost = (
   model: string,
   usage: TokenUsage | undefined,
 ): number | null => {
   if (!usage) return null;
-  const price = getModelPrice(model);
-  if (!price) return null;
-  const inputCost = (usage.input * price.inputUsdPerMTok) / 1_000_000;
-  const outputCost = (usage.output * price.outputUsdPerMTok) / 1_000_000;
-  return inputCost + outputCost;
+  return calculateModelUsageCost(model, {
+    inputTokens: usage.input,
+    outputTokens: usage.output,
+    cacheReadTokens: usage.cacheRead,
+    cacheWriteTokens: usage.cacheWrite,
+  });
 };
 
-/** 将成本数值格式化为便于人读的形式。将 `null` 视为 "—"。 */
 export const formatCost = (usd: number | null | undefined): string => {
   if (usd === null || usd === undefined) return "—";
   if (usd === 0) return "$0.00";

src/eval/gaia/types.ts

@@ -71,6 +71,8 @@ export interface TokenUsage {
   input: number;
   output: number;
   total: number;
+  cacheRead?: number;
+  cacheWrite?: number;
 }
 
 export interface QuestionResult {

src/main/core/__tests__/agent-loop.test.ts

@@ -277,8 +277,8 @@ describe("AgentLoop", () => {
 
     expect(end.totalUsage?.inputTokens).toBe(1_000_000);
     expect(end.contextUsage).toEqual({
-      latestStepContextTokens: 48_000,
-      maxStepContextTokens: 50_000,
+      latestStepContextTokens: 49_000,
+      maxStepContextTokens: 51_000,
     });
   });
 

src/main/core/agent/__tests__/agent-loop-reasoning.test.ts

@@ -66,4 +66,24 @@ describe("mapUsage — reasoning token extraction", () => {
     });
     expect(u?.cacheReadTokens).toBe(64);
   });
+
+  it("separates provider cache usage from uncached input tokens", () => {
+    const u = mapUsage({
+      inputTokens: 100,
+      outputTokens: 10,
+      inputTokenDetails: {
+        noCacheTokens: 20,
+        cacheReadTokens: 64,
+        cacheWriteTokens: 16,
+      },
+    });
+
+    expect(u).toMatchObject({
+      inputTokens: 20,
+      outputTokens: 10,
+      totalTokens: 110,
+      cacheReadTokens: 64,
+      cacheWriteTokens: 16,
+    });
+  });
 });

src/main/core/agent/agent-loop.ts

@@ -728,16 +728,16 @@ export class AgentLoop {
           }
           case "finish-step": {
             const stepUsage = mapUsage(part.usage);
-            const stepInputTokens = stepUsage?.inputTokens;
+            const stepContextTokens = stepUsage?.totalTokens;
             if (
-              typeof stepInputTokens === "number" &&
-              Number.isFinite(stepInputTokens) &&
-              stepInputTokens >= 0
+              typeof stepContextTokens === "number" &&
+              Number.isFinite(stepContextTokens) &&
+              stepContextTokens >= 0
             ) {
-              latestStepContextTokens = stepInputTokens;
+              latestStepContextTokens = stepContextTokens;
               maxStepContextTokens = Math.max(
                 maxStepContextTokens,
-                stepInputTokens,
+                stepContextTokens,
               );
             }
             if (messageOpen) {
@@ -1039,7 +1039,9 @@ interface RawUsage {
   reasoningTokens?: number | null;
   /** AI SDK v7 的输入 token 细分。 */
   inputTokenDetails?: {
+    noCacheTokens?: number | null;
     cacheReadTokens?: number | null;
+    cacheWriteTokens?: number | null;
   } | null;
   /** AI SDK v6+ 的嵌套形式(优先于已废弃的扁平字段)。 */
   outputTokenDetails?: {
@@ -1051,11 +1053,23 @@ interface RawUsage {
 export function mapUsage(raw: unknown): TokenUsage | undefined {
   if (!raw || typeof raw !== "object") return undefined;
   const u = raw as RawUsage;
-  const input = u.inputTokens ?? null;
+  const cacheRead =
+    u.inputTokenDetails?.cacheReadTokens ?? u.cachedInputTokens ?? null;
+  const cacheWrite = u.inputTokenDetails?.cacheWriteTokens ?? null;
+  const input =
+    u.inputTokenDetails?.noCacheTokens ??
+    (u.inputTokens != null && (cacheRead != null || cacheWrite != null)
+      ? Math.max(0, u.inputTokens - (cacheRead ?? 0) - (cacheWrite ?? 0))
+      : (u.inputTokens ?? null));
   const output = u.outputTokens ?? null;
   const total =
     u.totalTokens ??
-    (input !== null || output !== null ? (input ?? 0) + (output ?? 0) : null);
+    (input !== null ||
+    output !== null ||
+    cacheRead !== null ||
+    cacheWrite !== null
+      ? (input ?? 0) + (output ?? 0) + (cacheRead ?? 0) + (cacheWrite ?? 0)
+      : null);
   // 优先使用嵌套的 `outputTokenDetails.reasoningTokens`;对较旧的 SDK
   // 响应则回退到已废弃的扁平 `reasoningTokens` 字段。
   const reasoning =
@@ -1064,8 +1078,8 @@ export function mapUsage(raw: unknown): TokenUsage | undefined {
     inputTokens: input,
     outputTokens: output,
     totalTokens: total,
-    cacheReadTokens:
-      u.inputTokenDetails?.cacheReadTokens ?? u.cachedInputTokens ?? null,
+    cacheReadTokens: cacheRead,
+    cacheWriteTokens: cacheWrite,
     reasoningTokens: reasoning,
   };
 }

src/main/core/session/message-parts.ts

@@ -157,6 +157,9 @@ export interface UsagePart {
   inputTokens: number | null;
   outputTokens: number | null;
   totalTokens: number | null;
+  cacheReadTokens?: number | null;
+  cacheWriteTokens?: number | null;
+  costUsd?: number | null;
   modelId: string | null;
   provider: string | null;
   latestStepContextTokens?: number | null;

src/main/db/__tests__/automation-runs-migration.test.ts

@@ -200,6 +200,15 @@ describe("automation run migrations", () => {
     expect(sqliteMock.state.execCalls.join("\n")).toContain(
       "ALTER TABLE tasks ADD COLUMN updated_at TEXT",
     );
+    expect(sqliteMock.state.execCalls.join("\n")).toContain(
+      "ALTER TABLE tasks ADD COLUMN cache_read_tokens INTEGER",
+    );
+    expect(sqliteMock.state.execCalls.join("\n")).toContain(
+      "ALTER TABLE tasks ADD COLUMN cache_write_tokens INTEGER",
+    );
+    expect(sqliteMock.state.execCalls.join("\n")).toContain(
+      "ALTER TABLE tasks ADD COLUMN cost_usd REAL",
+    );
   });
 
   it("adds context checkpoint columns to existing task summaries", async () => {

src/main/db/index.ts

@@ -52,7 +52,15 @@ export const initDatabase = async () => {
       completed_at TEXT,
       session_id TEXT,
       assistant_message_id TEXT,
-      updated_at TEXT
+      updated_at TEXT,
+      input_tokens INTEGER,
+      output_tokens INTEGER,
+      total_tokens INTEGER,
+      cache_read_tokens INTEGER,
+      cache_write_tokens INTEGER,
+      cost_usd REAL,
+      model_id TEXT,
+      provider TEXT
     );
     CREATE TABLE IF NOT EXISTS task_trace_events (
       id TEXT PRIMARY KEY,
@@ -567,6 +575,15 @@ export const initDatabase = async () => {
   if (!taskColumnNames.has("updated_at")) {
     sqlite.exec("ALTER TABLE tasks ADD COLUMN updated_at TEXT");
   }
+  if (!taskColumnNames.has("cache_read_tokens")) {
+    sqlite.exec("ALTER TABLE tasks ADD COLUMN cache_read_tokens INTEGER");
+  }
+  if (!taskColumnNames.has("cache_write_tokens")) {
+    sqlite.exec("ALTER TABLE tasks ADD COLUMN cache_write_tokens INTEGER");
+  }
+  if (!taskColumnNames.has("cost_usd")) {
+    sqlite.exec("ALTER TABLE tasks ADD COLUMN cost_usd REAL");
+  }
 
   // 迁移:task_trace_events 表缺失时创建(旧版数据库)
   sqlite.exec(`
@@ -890,6 +907,9 @@ interface Task {
   inputTokens?: number | null;
   outputTokens?: number | null;
   totalTokens?: number | null;
+  cacheReadTokens?: number | null;
+  cacheWriteTokens?: number | null;
+  costUsd?: number | null;
   modelId?: string | null;
   provider?: string | null;
 }
@@ -1115,6 +1135,11 @@ export const updateTask = (id: string, updates: Partial<Task>) => {
     mapped.outputTokens = updates.outputTokens;
   if (updates.totalTokens !== undefined)
     mapped.totalTokens = updates.totalTokens;
+  if (updates.cacheReadTokens !== undefined)
+    mapped.cacheReadTokens = updates.cacheReadTokens;
+  if (updates.cacheWriteTokens !== undefined)
+    mapped.cacheWriteTokens = updates.cacheWriteTokens;
+  if (updates.costUsd !== undefined) mapped.costUsd = updates.costUsd;
   if (updates.modelId !== undefined) mapped.modelId = updates.modelId;
   if (updates.provider !== undefined) mapped.provider = updates.provider;
   if (updates.updatedAt !== undefined) {

src/main/db/schema.ts

@@ -26,6 +26,9 @@ export const tasks = sqliteTable("tasks", {
   inputTokens: integer("input_tokens"),
   outputTokens: integer("output_tokens"),
   totalTokens: integer("total_tokens"),
+  cacheReadTokens: integer("cache_read_tokens"),
+  cacheWriteTokens: integer("cache_write_tokens"),
+  costUsd: real("cost_usd"),
   modelId: text("model_id"),
   provider: text("provider"),
 });