@chenlongapps/opencode-token-usage 0.3.4 → 0.4.0

package/dist/usage.d.ts

@@ -91,16 +91,48 @@ export declare function summarize(messages: Iterable<UsageMessage>, catalog?: Pr
     total: number;
     cacheRate: number;
     steps: number;
+    calls: number;
     cost: number;
     costStatus: CostStatus;
 };
 export type Summary = ReturnType<typeof summarize>;
+export interface ModelCost {
+    model: string;
+    calls: number;
+    tokens: number;
+    cost: number;
+    costStatus: CostStatus;
+}
+/** Groups the same per-message estimates used by the tree total; unlabelled calls stay visible. */
+export declare function summarizeModels(messages: Iterable<UsageMessage>, catalog?: PriceCatalog): ModelCost[];
 export interface ContextUsage {
     used: number;
     limit: number;
     percent: number;
 }
+export interface ContextDetails {
+    usage: ContextUsage;
+    tokens: Tokens;
+}
+export declare function contextDetails(messages: readonly UsageMessage[], limit?: number): ContextDetails | undefined;
 export declare function contextUsage(messages: readonly UsageMessage[], limit?: number): ContextUsage | undefined;
 export declare function formatTokens(value: number): string;
 export declare const formatCost: (value: number) => string;
+export declare function formatEstimatedCost(cost: number, status: CostStatus): string | undefined;
+/** Compact display for scan-first views: 812, 139.4K, 3.70M. Detailed views keep formatTokens. */
+export declare function formatCompact(value: number): string;
+/** Splits a percentage into filled and empty cell counts for a plain-text bar. */
+export declare function bar(percent: number, width: number): {
+    filled: number;
+    empty: number;
+};
+/**
+ * One model request: the five categories of the last assistant call plus its cache rate.
+ * Compact mode hides zero rows and drops percentages; detailed mode keeps both.
+ */
+export declare function requestRows(details: ContextDetails, compact: boolean): readonly (readonly [string, string])[];
+/** Tree summary rows for the dialog's detailed Session view, calls included. */
+export declare function summaryRows(summary: Summary): readonly (readonly [string, string])[];
+/** `1 step` reads better than `1 steps` in the dialog's single-line summaries. */
+export declare const countLabel: (value: number, noun: string) => string;
 export declare function usageRows(summary?: Summary, context?: ContextUsage, performance?: PerformanceSummary): readonly (readonly [string, string])[];

package/dist/usage.js

@@ -22,6 +22,22 @@ export function estimate(tokens, prices = []) {
             && (rate === undefined || !Number.isFinite(rate) || rate < 0))),
     };
 }
+function priceMessage(tokens, model, catalog) {
+    if (!model)
+        return undefined;
+    let value = estimate(tokens, catalog.get(modelKey(model)));
+    if (value.defaultPrice || value.cost === 0) {
+        const snapshot = estimate(tokens, officialPrice(model)?.prices);
+        if (!snapshot.defaultPrice)
+            value = snapshot;
+    }
+    return value.defaultPrice ? undefined : value.cost;
+}
+function costStatus(known, missing) {
+    if (missing > 0)
+        return known > 0 ? "partial" : "unavailable";
+    return known > 0 ? "complete" : "empty";
+}
 /** Prices every call with its recorded model. A complete non-zero OpenCode
  * price wins; a complete snapshot price can replace a complete OpenCode zero.
  * Incomplete OpenCode prices fall back to the snapshot as a whole. */
@@ -29,6 +45,7 @@ export function summarize(messages, catalog = new Map()) {
     const tokens = normalize();
     let cost = 0;
     let steps = 0;
+    let calls = 0;
     let knownCosts = 0;
     let missingCosts = 0;
     for (const message of messages) {
@@ -46,37 +63,55 @@ export function summarize(messages, catalog = new Map()) {
         tokens.cache.write += t.cache.write;
         if (total(t) === 0)
             continue;
-        if (!message.model) {
-            missingCosts++;
-            continue;
-        }
-        let value = estimate(t, catalog.get(modelKey(message.model)));
-        if (value.defaultPrice || value.cost === 0) {
-            const snapshot = estimate(t, officialPrice(message.model)?.prices);
-            if (!snapshot.defaultPrice)
-                value = snapshot;
-        }
-        if (value.defaultPrice)
+        // Same population the per-model rows price, so the dialog's call count reconciles with them.
+        calls++;
+        const value = priceMessage(t, message.model, catalog);
+        if (value === undefined)
             missingCosts++;
         else {
-            cost += value.cost;
+            cost += value;
             knownCosts++;
         }
     }
-    const costStatus = missingCosts > 0
-        ? knownCosts > 0 ? "partial" : "unavailable"
-        : knownCosts > 0 ? "complete" : "empty";
     return {
         tokens, total: total(tokens), cacheRate: incoming(tokens) ? tokens.cache.read / incoming(tokens) : 0,
-        steps, cost, costStatus,
+        steps, calls, cost, costStatus: costStatus(knownCosts, missingCosts),
     };
 }
+/** Groups the same per-message estimates used by the tree total; unlabelled calls stay visible. */
+export function summarizeModels(messages, catalog = new Map()) {
+    const groups = new Map();
+    for (const message of messages) {
+        if (message.type !== "assistant" && message.type !== "compaction")
+            continue;
+        const tokens = normalize(message.tokens);
+        const count = total(tokens);
+        if (count === 0)
+            continue;
+        const key = message.model && modelKey(message.model);
+        const group = groups.get(key) ?? { calls: 0, tokens: 0, cost: 0, known: 0, missing: 0 };
+        group.calls++;
+        group.tokens += count;
+        const value = priceMessage(tokens, message.model, catalog);
+        if (value === undefined)
+            group.missing++;
+        else {
+            group.cost += value;
+            group.known++;
+        }
+        groups.set(key, group);
+    }
+    return [...groups].map(([model, group]) => ({
+        model: model ?? "Unknown model", calls: group.calls, tokens: group.tokens,
+        cost: group.cost, costStatus: costStatus(group.known, group.missing),
+    })).sort((a, b) => b.cost - a.cost || a.model.localeCompare(b.model));
+}
 const usableLimit = (value) => value !== undefined && Number.isFinite(value) && value > 0;
 // Matches the OpenCode 2.0.10 sidebar: the last assistant message with tokens that follows
 // the last completed compaction carries the current context size. Running or failed
 // compactions do not reset it, and a last message without usage hides the row entirely
 // instead of falling back to pre-compaction history.
-export function contextUsage(messages, limit) {
+export function contextDetails(messages, limit) {
     if (!usableLimit(limit))
         return undefined;
     let boundary = -1;
@@ -91,15 +126,68 @@ export function contextUsage(messages, limit) {
         const message = messages[index];
         if (message.type !== "assistant" || !message.tokens)
             continue;
-        const used = total(normalize(message.tokens));
-        return used > 0 ? { used, limit, percent: used / limit * 100 } : undefined;
+        const tokens = normalize(message.tokens);
+        const used = total(tokens);
+        return used > 0 ? { usage: { used, limit, percent: used / limit * 100 }, tokens } : undefined;
     }
     return undefined;
 }
+export function contextUsage(messages, limit) {
+    return contextDetails(messages, limit)?.usage;
+}
 export function formatTokens(value) {
     return Math.round(safe(value)).toLocaleString("en-US");
 }
 export const formatCost = (value) => value > 0 && value < 0.01 ? "<$0.01" : `$${safe(value).toFixed(2)}`;
+export function formatEstimatedCost(cost, status) {
+    if (status === "empty")
+        return undefined;
+    if (status === "unavailable")
+        return "—";
+    return `${formatCost(cost)}${status === "partial" ? " · partial" : ""}`;
+}
+/** Compact display for scan-first views: 812, 139.4K, 3.70M. Detailed views keep formatTokens. */
+export function formatCompact(value) {
+    const n = safe(value);
+    if (n < 1_000)
+        return String(Math.round(n));
+    const k = n / 1_000;
+    return k >= 999.95 ? `${(n / 1_000_000).toFixed(2)}M` : `${k.toFixed(1)}K`;
+}
+/** Splits a percentage into filled and empty cell counts for a plain-text bar. */
+export function bar(percent, width) {
+    const size = Math.max(0, Math.floor(width));
+    const filled = Math.max(0, Math.min(size, Math.round(percent / 100 * size)));
+    return { filled, empty: size - filled };
+}
+/**
+ * One model request: the five categories of the last assistant call plus its cache rate.
+ * Compact mode hides zero rows and drops percentages; detailed mode keeps both.
+ */
+export function requestRows(details, compact) {
+    const t = details.tokens;
+    const rows = [
+        ["Input", t.input], ["Output", t.output], ["Reasoning", t.reasoning],
+        ["Cache Read", t.cache.read], ["Cache Write", t.cache.write],
+    ].filter(([, count]) => !compact || count > 0);
+    const incomingTokens = incoming(t);
+    return [
+        ...rows.map(([label, count]) => [
+            label,
+            compact ? formatCompact(count) : `${formatTokens(count)} (${(count / details.usage.used * 100).toFixed(1)}%)`,
+        ]),
+        ["Cache Rate", `${(incomingTokens ? t.cache.read / incomingTokens * 100 : 0).toFixed(1)}%`],
+    ];
+}
+/** Tree summary rows for the dialog's detailed Session view, calls included. */
+export function summaryRows(summary) {
+    return [
+        ["Steps", formatTokens(summary.steps)], ["Calls", formatTokens(summary.calls)],
+        ...usageRows(summary).filter(([label]) => label !== "Steps"),
+    ];
+}
+/** `1 step` reads better than `1 steps` in the dialog's single-line summaries. */
+export const countLabel = (value, noun) => `${formatTokens(value)} ${noun}${value === 1 ? "" : "s"}`;
 export function usageRows(summary, context, performance) {
     const t = summary?.tokens;
     const number = (value) => value === undefined ? "—" : formatTokens(value);
@@ -110,12 +198,9 @@ export function usageRows(summary, context, performance) {
     if (t && t.cache.write > 0)
         rows.push(["Cache Write", number(t.cache.write)]);
     rows.push(["Cache Rate", summary ? `${(summary.cacheRate * 100).toFixed(1)}%` : "—"], ["Total", number(summary?.total)]);
-    if (summary?.costStatus === "complete")
-        rows.push(["Cost", formatCost(summary.cost)]);
-    if (summary?.costStatus === "partial")
-        rows.push(["Cost", `${formatCost(summary.cost)} · partial`]);
-    if (summary?.costStatus === "unavailable")
-        rows.push(["Cost", "—"]);
+    const cost = summary && formatEstimatedCost(summary.cost, summary.costStatus);
+    if (cost !== undefined)
+        rows.push(["Cost", cost]);
     if (performance?.tps !== undefined && Number.isFinite(performance.tps)) {
         rows.push(["TPS", `${performance.tpsEstimated ? "~" : ""}${Math.max(0, performance.tps).toFixed(1)} tok/s`]);
     }

package/docs/verification.md

@@ -1,5 +1,41 @@
 # 验证记录
 
+## v0.4.0（`/usage` 原生弹窗）
+
+验证日期：2026-09-23。Node.js v22.23.2；`@opencode/plugin`、`@opencode/client` 等 SDK 精确锁定 2.0.11。弹窗布局参考本地 OpenCode 2.0.15 源码中的 `dialog-status.tsx`、`dialog-debug.tsx` 和 `dialog-shell-output.tsx`，通过插件公开的斜杠命令、弹窗、`session.hook("context")` 和 RPC 接口实现，并在真实 OpenCode 2.0.11 宿主中验证。
+
+| 检查 | 结果 |
+| --- | --- |
+| `npm run typecheck` | 通过 |
+| `npm test` | 62 项通过 |
+| `npm run build` | 通过，输出 ESM 与类型声明 |
+| `npm run test:smoke` | 通过，使用 `.tgz` 安装产物、真实 OpenCode 2.0.11 与本地模拟提供商 |
+
+`/usage` 仅在会话页面注册为 CLI 斜杠／命令面板命令，打开原生弹窗而不向模型发送新消息。弹窗按当前查看会话展示最后一次已完成压缩后的上下文用量；五类 token 的占比以该条 assistant 消息的已用 context 为分母，窗口占用率以活动模型上下文上限为分母。会话树明细沿用侧边栏的完整分页、子代理汇总和分叉继承去重口径；按模型汇总的费用与全树 Cost 共用逐消息计价规则，分别处理免费、缺价、`partial` 及内置官方价格回退。
+
+自动化测试新增：五类 context 分类百分比、压缩边界与未知上限；混合模型和 compaction、价格回退、免费与缺价时逐模型费用对全树费用的核对；弹窗详细快照在子会话、刷新失败恢复和切换会话时的生命周期。打包集成测试确认空会话不会伪造上下文占用，真实 `/usage` 可打开与关闭、显示 1,270 token 当前上下文分类和模型费用；100 × 28 终端内滚动可到达模型费用；真实子代理合并后显示全树 5,080 token；切换活动模型后弹窗上限变为 32,000，历史消息费用仍列在原模型下。2.0.11 子代理视图最初处于宿主的子代理覆盖层，不提供斜杠输入；其全树面板验证继续由 `session.composer.top` 覆盖。
+
+新增 **Context Sources · estimated**：服务端只在 `context` 钩子中观察最近一次已组装的模型请求，按可识别的消息文本、系统工具定义／Code Mode 目录、系统提示词、技能文本／目录、MCP 指令／工具以及其他内容分别估算，持久化的仅是六类数值、模型和采样时间；CLI 通过位置限定的 RPC 读取。估算使用 UTF-8 字节数近似，不是提供商真实 token 归因，媒体负载及提供商包装无法准确计量，也不会与含输出／推理的实测 Context 相加一致。未捕获过请求的会话显示不可用；服务端插件或 RPC 不可用时仍显示实测统计。测试覆盖六类独立归因、占比分母、无技能/MCP 时的零值、Code Mode 后续项目指令、工具结果媒体忽略、无效 RPC 数据，以及 RPC 缓慢／失败时不阻塞实测用量或串会话。真实 2.0.11 打包集成验证了请求钩子记录与 CLI 弹窗读取均生效。
+
+## v0.4.0（`/usage` 弹窗信息架构重构）
+
+验证日期：2026-09-23。Node.js v22.23.2；SDK 仍精确锁定 2.0.11。本轮只改弹窗的数据映射与呈现，统计与定价逻辑不变。
+
+| 检查 | 结果 |
+| --- | --- |
+| `npm run typecheck` | 通过 |
+| `npm test` | 62 项通过 |
+| `npm run build` | 通过 |
+| `npm run test:smoke` | 通过，真实 OpenCode 2.0.11 + 本地模拟提供商 |
+
+信息架构由 `Context / Context Sources / Session Tree / Cost by Model` 改为 **Context Window / Last Request / Context Breakdown / Session / By Model**：Context Window 只呈现上下文占用（分母为活动模型上限）并带条形图；原 Context 内的五类 token 迁入 Last Request 并补充该次请求的缓存率；`Context Sources · estimated` 改名 Context Breakdown；`Session Tree` 改名 Session，因为面板本来就是全树汇总而非树结构；`Cost by Model` 改名 By Model。标题下方新增一行显示会话名与活动模型（多模型时显示模型数量），正文不再重复模型名，弹窗改用宿主 `centered` 选项垂直居中。
+
+默认紧凑模式使用 K/M 缩写（< 1,000 精确、< 1,000,000 一位小数 K、≥ 1,000,000 两位小数 M，四舍五入不会出现 `1000.0K`），隐藏零值行，并把 Context Breakdown 的两类工具合并为 `Tools` 行按占比降序；按 `d` 切换详细模式，恢复精确数字、零值行、`System Tools` / `MCP Tools` 拆分以及 Session 逐类明细（含 Calls）。Session 单行汇总为 `steps · calls · tokens · cached% · cost`，其中 calls 与 By Model 各行同源，可与逐模型行核对。条形图宽度、Last Request 双列与内容宽度随宿主 `large` 档（88 列）收敛，窄终端自动改为单列并收窄条长。
+
+自动化测试新增：`formatCompact` 边界、`bar` 填充与越界、`requestRows` 的紧凑／详细两态、`summarize().calls` 与 `summaryRows`、`countLabel` 单复数、`breakdownRows` 的工具合并、降序与零值策略。打包集成测试确认紧凑与详细两种模式可切换、`Used / Limit` 显示 `1.3K / 128.0K (1.0%)`、Context Breakdown 显示 `Tools` 合并行与条形图、Session 单行汇总、By Model 逐模型费用，以及 100 × 28 终端内滚动仍可到达模型费用。
+
+尚未实现（属于后续评估）：真实的分叉／分支树统计、进度条以外的窗口告警色阶、公开价格目录。
+
 ## v0.3.3（GPT 与 Claude Fast 价格）
 
 验证日期：2026-09-23。根据 [OpenAI API Pricing](https://developers.openai.com/api/docs/pricing) 与 [Claude Platform Pricing](https://platform.claude.com/docs/en/about-claude/pricing)，为当前快照中官方支持 Fast mode 的 GPT-5.5、GPT-5.6 Sol/Terra/Luna、GPT-6 Astra/Sol/Luna，以及 Claude Opus 4.8/5/5.5 分别加入独立 Fast 条目。Fast 费率按官方长上下文档位与缓存倍率记录；GPT-5.5 Fast 未公布长上下文费率，超过 272,000 个传入 token 时显示不可估价。条目分别限制到官方支持 Fast mode 的 OpenAI 与 Anthropic provider；快照由 76 个模型增至 86 个。

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@chenlongapps/opencode-token-usage",
-  "version": "0.3.4",
+  "version": "0.4.0",
   "description": "Session-tree token usage, live performance, context window and estimated cost for OpenCode 2",
   "keywords": [
     "opencode",