dsh-balance-widget 0.5.2 → 0.5.4

package/README.en.md

@@ -142,7 +142,7 @@ If you mainly use DeepSeek-V4-Pro, point the pricing model at it for a more accu
 
 ## Pricing
 
-Built-in DeepSeek official peak/off-peak pricing (CNY per 1M tokens), effective 2026-09-10. Peak windows are Beijing time 09:00–12:00 and 14:00–18:00; prices are double the off-peak rates:
+Built-in DeepSeek official peak/off-peak pricing (CNY per 1M tokens), effective 2026-09-10. Peak windows are Beijing time **Mon–Fri** 09:00–12:00 and 14:00–18:00 (weekends are off-peak all day); prices are double the off-peak rates:
 
 | Model | Window | Cache hit (input) | Cache miss (input) | Output |
 | --- | --- | --- | --- | --- |
@@ -153,6 +153,12 @@ Built-in DeepSeek official peak/off-peak pricing (CNY per 1M tokens), effective
 
 The official page currently lists only `deepseek-flash` and `deepseek-v4-pro`; older names (`deepseek-v4-flash`, `deepseek-v4-flash-vision-exp`, `deepseek-chat`) still work and are billed at the Flash rate. The plugin re-parses the page at startup and every 12h, and falls back to the table above only when parsing fails. Costs are **estimates**; the provider's bill is authoritative.
 
+### Search calls (estimated)
+
+Every **query** of a `web_search` call is a separate auxiliary model call (`@deepseek-ai/dsh-web-search-deepseek`, `deepseek-v4-flash` by default, up to `maxUses` per request). The session log records only the pre-dispatch `web/deepseek-search-llm-request` event — **no usage field** — so the real token counts exist in no local file.
+
+The plugin therefore counts requests × a calibrated constant: ~6k cache-miss input + ~2k output tokens per search, i.e. **¥0.014** off-peak and **¥0.028** at peak (calibrated 2026-09-22 against the balance delta: 14 searches measured **¥0.0133 each**). Searches are attributed to the session, workspace and day they belong to, but each amount is an estimate (±30% order).
+
 ## Security & permission boundaries
 
 This section is for the DSH Store / plugin audit: dependencies, runtime permissions, external services, and failure bounds.
@@ -183,6 +189,15 @@ This section is for the DSH Store / plugin audit: dependencies, runtime permissi
 
 ## Changelog
 
+### v0.5.4 — search calls are now counted (previously dropped entirely)
+- 🐛 **Fixed**: the auxiliary model calls behind `web_search` contributed nothing to cost. DSH writes only a pre-dispatch `web/deepseek-search-llm-request` event per search (no usage), so any log-derived pricing undercounts systematically — measured 2026-09-22: all 1073 searches of the day were missing. Searches are now priced as requests × a calibrated constant and attributed to their session / workspace / day and turn
+- 📏 **Calibration**: ~6k cache-miss input + ~2k output tokens per search (¥0.014 off-peak, ¥0.028 at peak). Measured 2026-09-22 from the balance delta: 14 searches in one window, balance delta ¥0.28, minus ¥0.094 of log-derivable conversation cost → **¥0.0133 per search**
+- ⚠️ **Scope**: the search portion is an estimate (±30% order); conversation text is still priced exactly from the logs
+
+### v0.5.3 — weekends no longer mispriced as peak
+- 🐛 **Fixed**: the peak/off-peak check looked only at the hour and ignored the weekday, so weekends were priced as peak (2×) during 09:00–12:00 and 14:00–18:00, up to doubling "today's cost". The official rule is **Mon–Fri** 09:00–12:00 and 14:00–18:00 (everything else, weekends included, is off-peak); Saturday and Sunday are now excluded first, using the Beijing-time weekday
+- 📄 **Docs**: the pricing section and the peak/off-peak tooltips now state the Mon–Fri restriction
+
 ### v0.5.2 — DSH 0.1.5: new session format & new pricing page
 - 🐛 **Fixed**: DSH 0.1.5 writes session logs under a generation-tagged name (`session.v3.jsonl.zstd`), but the plugin only matched `session.jsonl.zstd`, so every session created after the upgrade was invisible — viewing one produced red error text in the popover, and today's cost was understated (measured ~36% low). Logs are now found by taking the highest generation per session directory; a migrated session's older file is a subset of the new one, so reading only the newest avoids double-counting the session
 - 🐛 **Fixed**: the official pricing page renamed the Flash column to `deepseek-flash` and cut its rates. The parser's model-id anchor landed on a page footnote, so the sync reported success while silently keeping the stale, higher rates — overstating costs by ~1.7x. The parser now anchors on the table's row labels and applies that column to every Flash-family name

package/README.md

@@ -139,7 +139,7 @@ DSH 的插件配置统一放在这个文件里：
 
 ## 定价说明
 
-内置 DeepSeek 官方 2026-09-10 峰谷定价（元 / 百万 tokens），高峰时段为北京时间 09:00–12:00、14:00–18:00，价格为空闲时段两倍：
+内置 DeepSeek 官方 2026-09-10 峰谷定价（元 / 百万 tokens），高峰时段为北京时间**周一至周五** 09:00–12:00、14:00–18:00（周末全天为空闲时段），价格为空闲时段两倍：
 
 | 模型 | 时段 | 缓存命中(输入) | 缓存未命中(输入) | 输出 |
 | --- | --- | --- | --- | --- |
@@ -150,6 +150,12 @@ DSH 的插件配置统一放在这个文件里：
 
 官方页当前只列 `deepseek-flash` 与 `deepseek-v4-pro` 两列；`deepseek-v4-flash`、`deepseek-v4-flash-vision-exp`、`deepseek-chat` 等旧名仍可调用，按 Flash 价计费。插件在启动时与每 12h 从官方页重新解析，只有解析失败才回退上表。成本为**估算值**，实际以官方账单为准。
 
+### 搜索调用（估算）
+
+每次 `web_search` 的**每个查询**都会被 DSH 当作一次独立的辅助模型调用（`@deepseek-ai/dsh-web-search-deepseek`，默认 `deepseek-v4-flash`，每个请求最多 `maxUses` 次）。而会话日志只记录派发前的 `web/deepseek-search-llm-request` 事件——**不含 usage 字段**，本地任何文件都查不到这些调用的真实 token 数。
+
+插件因此按「请求次数 × 标定常数」计入：每次搜索 ≈ 6k 缓存未命中输入 + 2k 输出 token，即谷时 **¥0.014**、峰时 **¥0.028**（2026-09-22 用余额差标定：14 次搜索实测 **¥0.0133/次**）。搜索会归到它所属的会话、工作区与「今日·全部」，但单次金额是估算值（±30% 量级）。
+
 ## 安全与权限边界
 
 本节面向 DSH Store / 插件审计，列出依赖、运行时权限、外部服务与失败边界。
@@ -180,6 +186,15 @@ DSH 的插件配置统一放在这个文件里：
 
 ## 版本历史
 
+### v0.5.4 — 计入搜索调用（此前全部漏算）
+- 🐛 **修复**：`web_search` 的辅助模型调用完全不计入成本。DSH 对每次搜索只写一个派发前的 `web/deepseek-search-llm-request` 事件（无 usage），所以按会话日志计价的实现系统性漏算——实测 2026-09-22 一天 1073 次搜索全部没进账。现按「请求次数 × 标定常数」计价，并归入该搜索所属的会话 / 工作区 / 当日及当轮
+- 📏 **标定**：单次搜索 ≈ 6k 缓存未命中输入 + 2k 输出 token（谷时 ¥0.014、峰时 ¥0.028）。2026-09-22 用余额差实测：同一窗口 14 次搜索、余额差 ¥0.28、扣掉日志可算的对话成本 ¥0.094，得 **¥0.0133/次**
+- ⚠️ **口径**：搜索部分是估算值（±30% 量级），会话正文部分仍是按日志精算
+
+### v0.5.3 — 修复周末被误判为峰时
+- 🐛 **修复**：峰/谷判断只看小时、不看星期，导致周末 09:00–12:00、14:00–18:00 被误判为峰时并按 2× 计价，"今日花费"最多翻倍。官方规则为**周一至周五** 09:00–12:00、14:00–18:00（其余含周末全天为空闲时段），现按北京时间的星期先行排除周六/周日
+- 📄 **文档**：定价说明与峰谷 tooltip 补上"周一至周五"限定
+
 ### v0.5.2 — 适配 DSH 0.1.5：新会话格式与新定价页
 - 🐛 **修复**：DSH 0.1.5 将会话日志改为代数命名 `session.v3.jsonl.zstd`，插件此前只认 `session.jsonl.zstd`，导致升级后新建的会话全部不可见——查看这类会话时弹层出现红色报错，今日成本也漏算（实测少约 36%）。现按「每个会话目录取最高代」发现日志；迁移会话保留的旧文件是新代的子集，因此只读最高代，避免整个会话被重复计算
 - 🐛 **修复**：官方定价页将 Flash 列改名为 `deepseek-flash` 并下调价格，插件按模型名定位价格块的锚点因此落到页面脚注，同步"成功"却静默沿用旧高价，成本被高估约 1.7 倍。现改为按价格表行标签锚定，并把该列价格套用到全部 Flash 系名称

package/lib/client.js

@@ -476,8 +476,8 @@ window.__ModuleLoader__.load({
 			"todayCostShort": "今日·全部",
 			"peak.active": "峰时",
 			"peak.offpeak": "谷时",
-			"peak.tooltip.active": "当前为峰时（09:00-12:00、14:00-18:00，价格翻倍）：输入 ¥{input}/M · 输出 ¥{output}/M",
-			"peak.tooltip.offpeak": "当前为谷时（其余时段，半价）：输入 ¥{input}/M · 输出 ¥{output}/M",
+			"peak.tooltip.active": "当前为峰时（周一至周五 09:00-12:00、14:00-18:00，价格翻倍）：输入 ¥{input}/M · 输出 ¥{output}/M",
+			"peak.tooltip.offpeak": "当前为谷时（其余时段含周末全天，半价）：输入 ¥{input}/M · 输出 ¥{output}/M",
 			"sessionNote": "基于最近活跃会话：",
 			"hint": "价格为估算值，实际以官方账单为准。",
 			"topUp": "去充值",
@@ -507,8 +507,8 @@ window.__ModuleLoader__.load({
 			"todayCostShort": "Today · all",
 			"peak.active": "Peak",
 			"peak.offpeak": "Off-peak",
-			"peak.tooltip.active": "Peak hours now (09:00-12:00, 14:00-18:00, 2× price): input ¥{input}/M · output ¥{output}/M",
-			"peak.tooltip.offpeak": "Off-peak hours now (rest of day, half price): input ¥{input}/M · output ¥{output}/M",
+			"peak.tooltip.active": "Peak hours now (Mon-Fri 09:00-12:00, 14:00-18:00, 2× price): input ¥{input}/M · output ¥{output}/M",
+			"peak.tooltip.offpeak": "Off-peak hours now (all other times incl. weekends, half price): input ¥{input}/M · output ¥{output}/M",
 			"sessionNote": "Based on most recent session:",
 			"hint": "Prices are estimates; actual billing from the provider is authoritative.",
 			"topUp": "Top up",

package/lib/index.js

@@ -264,11 +264,14 @@ const PRICING_ALIASES = {
 /**
  * Whether `date` falls inside a Beijing-time (UTC+8) peak window.
  * DeepSeek's peak pricing applies to Beijing time 09:00-12:00 and 14:00-18:00
- * regardless of the host's timezone, so the window is computed in UTC+8
- * explicitly instead of the machine's local time.
+ * on weekdays only (Mon-Fri); the whole weekend is off-peak. The window is
+ * computed in UTC+8 explicitly instead of the machine's local time, which also
+ * supplies the Beijing weekday.
  */
 function isPeak(date) {
 	const beijing = new Date(date.getTime() + 8 * 60 * 60 * 1000);
+	const day = beijing.getUTCDay(); // 0 = Sunday, 6 = Saturday
+	if (day === 0 || day === 6) return false;
 	const hour = beijing.getUTCHours();
 	return (hour >= 9 && hour < 12) || (hour >= 14 && hour < 18);
 }
@@ -295,6 +298,40 @@ function priceUsage(usage, modelId, timeMs) {
 	return (input * table.inputMiss + hit * table.inputHit + output * table.output) / 1e6;
 }
 
+/**
+ * Estimated token footprint of one auxiliary web-search turn.
+ *
+ * DeepSeek runs every `web_search` query as its own Messages call (see
+ * @deepseek-ai/dsh-web-search-deepseek) and DSH records only the
+ * pre-dispatch `web/deepseek-search-llm-request` event — never the response
+ * usage — so the real token counts are not recoverable from any local file.
+ * These constants were calibrated on 2026-09-22 against the account balance
+ * delta (14 searches over one window: ~¥0.0133 per search at off-peak Flash
+ * rates, i.e. ~6k cache-miss input + ~2k output per search).
+ */
+const SEARCH_TURN_INPUT_TOKENS = 6000;
+const SEARCH_TURN_OUTPUT_TOKENS = 2000;
+
+/** Price one auxiliary web-search turn (estimated; see constants above). */
+function priceSearchTurn(modelId, timeMs) {
+	const table = isPeak(new Date(timeMs ?? Date.now())) ? pricingTable[pricingKey(modelId)].peak : pricingTable[pricingKey(modelId)].offPeak;
+	return (SEARCH_TURN_INPUT_TOKENS * table.inputMiss + SEARCH_TURN_OUTPUT_TOKENS * table.output) / 1e6;
+}
+
+/** Cost of one parsed sample — a real usage sample, or an estimated search turn. */
+function sampleCost(sample) {
+	return sample.search === true ? priceSearchTurn(sample.model, sample.time) : priceUsage(sample.usage, sample.model, sample.time);
+}
+
+/** Token totals of one parsed sample (a search turn contributes its estimate). */
+function sampleTokens(sample) {
+	if (sample.search === true) return { input: SEARCH_TURN_INPUT_TOKENS, output: SEARCH_TURN_OUTPUT_TOKENS };
+	return {
+		input: (sample.usage.inputTokens ?? sample.usage.uncachedInputTokens ?? 0) + (sample.usage.cacheReadTokens ?? 0),
+		output: sample.usage.outputTokens ?? 0
+	};
+}
+
 /**
  * Parse the official pricing page into our pricingTable shape.
  * The page is HTML; we look for per-model peak/off-peak "per 1M tokens" rates.
@@ -540,6 +577,7 @@ async function decompressSession(path) {
 function parseSessionUsages(text) {
 	const samples = [];
 	let currentModel = void 0;
+	let currentTurn = 0;
 	let cwd = void 0;
 	let title = void 0;
 	for (const line of text.split("\n")) {
@@ -555,6 +593,8 @@ function parseSessionUsages(text) {
 			cwd = event.cwd;
 		} else if (type === "session/title") {
 			title = event.data?.title ?? title;
+		} else if (type === "turn/start") {
+			currentTurn = event.data?.turn ?? currentTurn;
 		} else if (type === "request/header") {
 			currentModel = event.data?.header?.config?.model ?? currentModel;
 		} else if (type === "assistant/message") {
@@ -567,6 +607,15 @@ function parseSessionUsages(text) {
 					usage
 				});
 			}
+		} else if (type === "web/deepseek-search-llm-request") {
+			// One auxiliary search turn: no usage is recorded anywhere, so it is
+			// priced from the calibrated estimate below.
+			samples.push({
+				turn: currentTurn,
+				time: event.time ?? Date.now(),
+				model: event.data?.body?.model ?? currentModel,
+				search: true
+			});
 		}
 	}
 	return { samples, cwd, title };
@@ -582,9 +631,10 @@ function lastTurnCost(samples) {
 	let output = 0;
 	for (const s of samples) {
 		if (s.turn !== lastTurn) continue;
-		cost += priceUsage(s.usage, s.model, s.time);
-		input += (s.usage.inputTokens ?? s.usage.uncachedInputTokens ?? 0) + (s.usage.cacheReadTokens ?? 0);
-		output += s.usage.outputTokens ?? 0;
+		cost += sampleCost(s);
+		const t = sampleTokens(s);
+		input += t.input;
+		output += t.output;
 	}
 	return { cost, inputTokens: input, outputTokens: output };
 }
@@ -599,9 +649,10 @@ function todaySessionCost(samples) {
 	let output = 0;
 	for (const s of samples) {
 		if (s.time < dayStart || s.time >= dayEnd) continue;
-		cost += priceUsage(s.usage, s.model, s.time);
-		input += (s.usage.inputTokens ?? s.usage.uncachedInputTokens ?? 0) + (s.usage.cacheReadTokens ?? 0);
-		output += s.usage.outputTokens ?? 0;
+		cost += sampleCost(s);
+		const t = sampleTokens(s);
+		input += t.input;
+		output += t.output;
 	}
 	return { cost, inputTokens: input, outputTokens: output };
 }
@@ -650,9 +701,10 @@ function todayCost(sessionList, cwdFilter) {
 		if (cwdFilter !== void 0 && session.cwd !== cwdFilter) continue;
 		for (const s of session.samples) {
 			if (s.time < dayStart || s.time >= dayEnd) continue;
-			cost += priceUsage(s.usage, s.model, s.time);
-			input += (s.usage.inputTokens ?? s.usage.uncachedInputTokens ?? 0) + (s.usage.cacheReadTokens ?? 0);
-			output += s.usage.outputTokens ?? 0;
+			cost += sampleCost(s);
+			const t = sampleTokens(s);
+			input += t.input;
+			output += t.output;
 		}
 	}
 	return { cost, inputTokens: input, outputTokens: output };

package/package.json

@@ -1,7 +1,7 @@
 {
   "name": "dsh-balance-widget",
   "description": "Balance & cost widget for the dsh web GUI: a sidebar footer card shows the DeepSeek account balance and today's costs; clicking opens a popover with a five-tier cost breakdown (last prompt with session name / today-this-session / today-this-workspace / today-all-workspaces), peak/off-peak status tags, term explanations, and a one-click top-up link. Zero external dependencies — works on Node 24.",
-  "version": "0.5.2",
+  "version": "0.5.4",
   "type": "module",
   "main": "lib/index.js",
   "files": [