给 Claude 系列模型添加价格

CHANGELOG.md

@@ -2,6 +2,28 @@
 
 All notable changes to this project are documented here. The format follows [Keep a Changelog](https://keepachangelog.com/en/1.1.0/) and the project uses [Semantic Versioning](https://semver.org/). 中文版见 [CHANGELOG_ZH.md](./CHANGELOG_ZH.md).
 
+## [Unreleased]
+
+### Added
+
+- **Claude cost estimation** — built-in Anthropic Claude official API standard rates (USD per 1M tokens, verified 2026-11-05 against https://platform.claude.com/docs/en/about-claude/pricing): Fable 5.1 / 5, Mythos 5.1 / 5, Opus 5.5 / 5 / 4.8 / 4.7 / 4.6 / 4.5 / 4.1 / 4, Sonnet 5.5 / 5 / 4.6 / 4.5 / 4, Haiku 4.5. Flat across peak/off-peak; CNY converted at the fixed 6.76 rate. Claude Haiku 5.5 is intentionally not listed (prompt-length tiered pricing cannot be determined from per-round totals).
+- **Separate cache-write pricing** — `BucketPrices` gains an optional `cacheWrite` rate (5-minute TTL write price for Claude); `pricing.json` accepts it too. When absent, cache writes are still billed at the cache-miss rate (DeepSeek behavior unchanged).
+- **Model-id normalization** — `normalizeModelId` lower-cases, strips the `anthropic/` prefix and rewrites dotted Claude versions (`claude-opus-5.5` → `claude-opus-5-5`); used by the host resolver, `pricing.json` keys and the client snapshot lookup.
+- **Unpriced rounds** — `RoundCost` / `ChartCost` / resolver results gain `unpriced`. Session totals that include unpriced rounds are shown as `≥` (lower bound) with a panel note; the current-model price line is hidden instead of showing `$0/1M`; compaction cost is left blank instead of `$0`.
+
+### Changed
+
+- **Unlisted Claude models are no longer billed at the DeepSeek fallback rate** — any model id containing `claude` (including Bedrock ids such as `anthropic.claude-…` / `us.anthropic.claude-…` and display names) that has no price resolves to zero rates and is marked unpriced, instead of the `deepseek-flash` fallback estimate that understated them by tens of times. DeepSeek models and a missing model name keep the flash fallback. This also changes the exported `pricingFor()` / `estimateCost()`: for such models they now return `UNPRICED_PRICING` / `0` (still `estimated: true`).
+- **Longest-key price matching** — price-key lookup (`matchPricingKey`, shared by builtin, `pricing.json` and the client snapshot) now picks the exact key, else the longest matching prefix regardless of insertion order, so `claude-opus-5-5-20261101` resolves to Opus 5.5 rather than Opus 5. Claude keys only match the exact id or the id plus a dated snapshot suffix (`-YYYYMMDD` / `-latest`), so an unlisted newer version (e.g. `claude-opus-5-6`) is unpriced instead of borrowing a sibling's rate. Non-Claude keys keep the lenient prefix semantics.
+
+### Fixed
+
+- **Anomaly detection** skips unpriced rounds, so their zero cost no longer drags the baseline down and causes later priced rounds to be flagged as spikes.
+
+### Compatibility
+
+- Checked against DSH 0.2.2-rc.1 (installed app bundle): the peer-range admission check passes (`includePrerelease`), the missing `dsh-client-ui-slots` package is skipped by `dsh.client.inject`, both slots and their standard props are unchanged, and the fold handles 0.2.2-rc.1 session logs. DeepSeek model ids and list prices are unchanged (`deepseek-flash` = DeepSeek-V4.1-Flash-1017, `deepseek-v4-pro` = DeepSeek-V4-Pro-0813).
+
 ## [1.1.8] - 2026-10-22
 
 ### Fixed

CHANGELOG_ZH.md

@@ -3,6 +3,28 @@
 本文件记录本项目所有重要变更。格式遵循 [Keep a Changelog](https://keepachangelog.com/zh-CN/1.1.0/)，
 版本号遵循 [语义化版本](https://semver.org/lang/zh-CN/)。英文版见 [CHANGELOG.md](./CHANGELOG.md)。
 
+## [未发布]
+
+### 新增
+
+- **Claude 成本估算**——内置 Anthropic Claude 官方 API 标准价（美元 / 1M tokens，2026-11-05 对照 https://platform.claude.com/docs/en/about-claude/pricing 核验）：Fable 5.1 / 5、Mythos 5.1 / 5、Opus 5.5 / 5 / 4.8 / 4.7 / 4.6 / 4.5 / 4.1 / 4、Sonnet 5.5 / 5 / 4.6 / 4.5 / 4、Haiku 4.5。不分高峰/空闲；人民币按固定汇率 6.76 折算。Claude Haiku 5.5 暂不收录（按请求输入长度分档，按轮汇总无法判定档位）。
+- **写缓存单独计价**——`BucketPrices` 新增可选 `cacheWrite`（Claude 取 5 分钟 TTL 写入价），`pricing.json` 同样支持；缺省时写缓存仍按「缓存未命中」价计（DeepSeek 行为不变）。
+- **模型 ID 归一化**——`normalizeModelId` 转小写、去掉 `anthropic/` 前缀、把 Claude 点号版本改为连字符（`claude-opus-5.5` → `claude-opus-5-5`）；宿主解析、`pricing.json` 键与 client 快照查找共用。
+- **未定价轮次**——`RoundCost` / `ChartCost` / 解析结果新增 `unpriced`。会话合计含未定价轮次时显示为 `≥`（下限）并在面板提示；当前模型未定价时不再显示「$0/1M」刊例价行；压缩成本留空而非显示 `$0`。
+
+### 变更
+
+- **未收录的 Claude 模型不再按 DeepSeek 回退价估算**——型号名含 `claude` 且未定价的模型（含 Bedrock 写法 `anthropic.claude-…` / `us.anthropic.claude-…` 与显示名）解析为全零单价并标记未定价，不再套用会低估数十倍的 `deepseek-flash` 回退价；DeepSeek 系列与模型名缺失仍按 flash 回退估算。导出的 `pricingFor()` / `estimateCost()` 随之变化：此类模型现返回 `UNPRICED_PRICING` / `0`（仍为 `estimated: true`）。
+- **价格键最长匹配**——价格键查找（`matchPricingKey`，内置表、`pricing.json` 与 client 快照共用）改为精确优先、其次最长前缀，与键的插入顺序无关：`claude-opus-5-5-20261101` 命中 Opus 5.5 而非 Opus 5。Claude 键只接受精确 ID 或「ID + 日期快照后缀」（`-YYYYMMDD` / `-latest`），未收录的新版本（如 `claude-opus-5-6`）显示为未定价，不借用相近型号的价格；非 Claude 键保持宽松前缀语义。
+
+### 修复
+
+- **成本异常检测**跳过未定价轮次，其零成本不再拉低基线、导致后续正常计价的轮次被误报为突增。
+
+### 兼容性
+
+- 对照已安装的 DSH 0.2.2-rc.1 应用包检查：peer 范围准入通过（`includePrerelease`）；`dsh.client.inject` 中已不存在的 `dsh-client-ui-slots` 会被跳过；两个槽位及其标准 props 未变；fold 能处理 0.2.2-rc.1 会话日志。DeepSeek 模型 ID 与刊例价未变（`deepseek-flash` = DeepSeek-V4.1-Flash-1017，`deepseek-v4-pro` = DeepSeek-V4-Pro-0813）。
+
 ## [1.1.8] - 2026-10-22
 
 ### 修复

README.md

@@ -29,7 +29,7 @@ Click ▸ to open the dashboard panel:
 
 - **Session usage summary** — Input (uncached/cached), output, cache-hit percentage, and context occupancy (derived from official adapter `tokenUsage` / `contextPressure` projections).
 - **Context breakdown & compaction diagnostics (v1.1.0)** — Official `contextBreakdown` projection breakdown (System prompt / Tools schema / Message history token counts and percentage with a 3-segment color bar, annotated as heuristic approximations); Host folds `compaction/*` events (which round was compacted, how many tokens were freed, model used, and summarize call cost); provides proactive suggestions (≥75% / ≥90% occupancy) to start a new session or reduce large file injections.
-- **Cost estimation** — Estimated from official list prices (CNY/USD dual-currency per 1M tokens, peak/off-peak tiers) with verified source date; supports user override via `pricing.json`; unpriced models are explicitly tagged.
+- **Cost estimation** — Estimated from official list prices (CNY/USD dual-currency per 1M tokens, peak/off-peak tiers) with verified source date; supports user override via `pricing.json`; built-in Anthropic Claude official API standard rates (USD); unpriced models are explicitly tagged.
 - **Session-cost aggregation (v1.1.7)** — Adds the host's per-round costs using each round's model and billing tier. While a new round is in progress, the indicator and panel share the fetched history plus a marked estimate for new tokens, then refresh the history after token updates pause. The model label follows the same current-history/live-node priority in both places.
 - **DSH 0.1.2+ and themed docks (v1.1.7)** — Reads conversation nodes from the independent `chat` source on newer DSH versions, with a legacy session-snapshot fallback; keeps the expanded panel anchored when a theme creates a fixed-position containing block.
 - **DSH 0.2.x client-SDK compatibility (v1.1.8)** — Declares client peer ranges spanning both DSH 0.1.x and 0.2.x (tested against 0.2.1-rc.2) so DSH 0.2.x's plugin admission gate accepts the plugin at startup; the retired `@deepseek-ai/dsh-client-runtime` peer is gone, and the assistant-node model identity now also reads the 0.2.x `providerMetadata` field (0.1.x `provenance` still preferred when present, `requestConfig` kept as the last fallback).
@@ -165,11 +165,26 @@ Costs are resolved with priority **user override file > builtin list price > fal
 }
 ```
 
-Unit prices are **dual-currency (CNY + USD) per 1M tokens**: `peak` covers peak hours (Beijing time Monday–Friday 09:00–12:00 and 14:00–18:00, charged at 2×), `offPeak` covers the rest (including weekends). Unpriced models are explicitly marked "Unpriced model" in the UI.
+Unit prices are **dual-currency (CNY + USD) per 1M tokens**: `peak` covers peak hours (Beijing time Monday–Friday 09:00–12:00 and 14:00–18:00, charged at 2×), `offPeak` covers the rest (including weekends). Prices may also carry an optional `cacheWrite` (cache-write rate); when omitted, cache writes are billed at the cache-miss rate.
+
+**Models that are not listed** are explicitly marked "Unpriced model" in the UI:
+
+- DeepSeek models (and a missing model name) are **estimated** at the `deepseek-flash` list price, shown with `≈`;
+- models whose id contains `claude` (including Bedrock ids such as `anthropic.claude-…` and display names such as `Claude Opus 5.5`) are **not estimated**: their rates are zero and their cost is excluded, so they are never billed at DeepSeek rates (which would understate them by tens of times). A session total that includes such rounds is shown as `≥` (a lower bound) with a note in the panel; the cost badge and compaction cost do not show `$0`, and anomaly detection skips these rounds.
+
+#### Claude models
+
+Built-in standard rates from the [Anthropic pricing page](https://platform.claude.com/docs/en/about-claude/pricing) (USD per 1M tokens, verified 2026-11-05) cover Fable 5.1 / 5, Mythos 5.1 / 5, Opus 5.5 / 5 / 4.8 / 4.7 / 4.6 / 4.5 / 4.1 / 4, Sonnet 5.5 / 5 / 4.6 / 4.5 / 4 and Haiku 4.5:
+
+- No peak/off-peak tiers; cache writes use the 5-minute TTL rate (usage cannot distinguish 1-hour TTL writes);
+- Model-id matching normalizes the `anthropic/` prefix and dotted versions (`claude-opus-5.5`); ids with a dated snapshot suffix (`-20261101` / `-latest`) are priced as that model; any other variant (an unlisted newer version, `-fast`, `-batch`, Bedrock ids, …) is treated as unpriced and never borrows a similar model's rate;
+- **Claude Haiku 5.5 is not listed yet**: it is priced by per-request prompt length (≤100K / >100K), which a per-round total cannot determine;
+- Not included: batch discounts, regional / `inference_geo` premiums, Fast mode, tool fees;
+- Anthropic only quotes USD; CNY amounts are converted at a **fixed rate of 6.76** (informational; they do not follow `cnyPerUsd` or the live rate).
 
 ### Display currency and live FX rate (v0.3+ / v1.0.1 official dual-currency)
 
-Costs are computed with the **official list price of the selected currency** (CNY quote from the Chinese pricing page, USD quote from the English pricing page — **no FX conversion**, consistent with the official bill). The cost section has a one-click **CNY/USD** toggle (remembered in the browser); the indicator, panel, chart and badge all follow it. `config.cnyPerUsd` (default 6.76) and the "Refresh rate" button (via Host `/dsh-usage-chart/rate` proxy) are used only for the informational "1 USD ≈ X CNY" note:
+Costs are computed with the **official list price of the selected currency** (CNY quote from the Chinese pricing page, USD quote from the English pricing page — **no FX conversion**, consistent with the official bill; the exception is the USD-only Claude models, whose CNY amounts use the fixed 6.76 rate described above). The cost section has a one-click **CNY/USD** toggle (remembered in the browser); the indicator, panel, chart and badge all follow it. `config.cnyPerUsd` (default 6.76) and the "Refresh rate" button (via Host `/dsh-usage-chart/rate` proxy) are used only for the informational "1 USD ≈ X CNY" note:
 
 - **Multi-source fallback**: when the custom source (`config.fxUrl`) is unreachable, a built-in fallback source (frankfurter.dev) is tried;
 - **Offline resilience**: the last successful rate is persisted, so a refresh while offline keeps the last real rate instead of the fixed default;

README_ZH.md

@@ -29,7 +29,7 @@
 
 - **会话用量汇总** — 输入（未命中/命中）、输出、缓存命中率、上下文占用（均来自官方 adapter 上报的 `tokenUsage` / `contextPressure` 投影）
 - **上下文构成与压缩诊断（v1.1.0）** — 官方 `contextBreakdown` 投影拆解（系统提示 / 工具定义 / 历史消息 Token 占比与分段条，标注启发式近似）；宿主折叠 `compaction/*` 事件流（哪一轮被压缩、释放了多少 Token、生成摘要花费了多少成本与对应模型）；结合上下文容量占比（≥75% / ≥90%）给出开新会话或精简大文件注入的优化建议
-- **成本估算** — 按官方刊例价（CNY/USD 双币种 /1M tokens，高峰/空闲双时段）估算并标注来源与核验日期；支持用户覆盖 `pricing.json`；未定价模型显式标记「未定价模型」
+- **成本估算** — 按官方刊例价（CNY/USD 双币种 /1M tokens，高峰/空闲双时段）估算并标注来源与核验日期；支持用户覆盖 `pricing.json`；内置 Anthropic Claude 官方 API 标准价（美元）；未定价模型显式标记「未定价模型」
 - **会话成本按轮汇总（v1.1.7）** — 按每轮实际模型和计费时段累加宿主成本；新轮进行中，指示器和面板共用“已读取历史＋新增 token 的估算”，实时用量暂停更新后刷新宿主历史。两处模型名称采用相同的历史/实时来源顺序。
 - **DSH 0.1.2+ 与皮肤兼容（v1.1.7）** — 新版从独立 `chat` 源读取会话节点，旧版回退到会话快照；皮肤创建 fixed 定位包含块时，展开面板仍贴近指示器。
 - **兼容 DSH 0.2.x client SDK（v1.1.8）** — client peer 范围同时覆盖 DSH 0.1.x 与 0.2.x（已在 0.2.1-rc.2 上验证），DSH 0.2.x 的插件准入校验在启动时即可通过；已移除退役的 `@deepseek-ai/dsh-client-runtime` peer，助手节点模型标识新增读取 0.2.x 的 `providerMetadata` 字段（存在 0.1.x 的 `provenance` 时仍优先，`requestConfig` 保留为最后回退）。
@@ -190,12 +190,34 @@ dsh web --profile web
 `offPeak` 为空闲时段（其余时间及周末全天）。兼容旧格式：平铺
 `{ "cacheMissInput": …, "cacheHitInput": …, "output": … }` 视为高峰/空闲同价、
 人民币报价（美元按默认汇率 6.76 折算）。`verifiedAt`（epoch 毫秒）可选，用于面板
-展示核验日期；未收录的模型会在 UI 中显式标记「未定价模型」，不会静默按 0 计。
+展示核验日期。单价可选 `cacheWrite`（写缓存价），缺省按「缓存未命中」价计。
+
+**未收录的模型**会在 UI 中显式标记「未定价模型」：
+
+- DeepSeek 系列（及模型名缺失）按 `deepseek-flash` 刊例价**回退估算**，金额带 `≈`；
+- 型号名含 `claude` 的模型（含 Bedrock 写法 `anthropic.claude-…`、`Claude Opus 5.5` 等显示名）
+  **不估算**：单价记为 0、成本不计入，避免套用 DeepSeek 价造成数十倍低估。会话合计含此类
+  轮次时显示为 `≥`（下限）并在面板提示；成本徽章与压缩成本不显示 `$0`，成本异常检测跳过这些轮次。
+
+#### Claude 系列
+
+内置 [Anthropic 官方价格页](https://platform.claude.com/docs/en/about-claude/pricing) 的标准价
+（美元 / 1M tokens，2026-11-05 核验），覆盖 Fable 5.1 / 5、Mythos 5.1 / 5、Opus 5.5 / 5 / 4.8 / 4.7 /
+4.6 / 4.5 / 4.1 / 4、Sonnet 5.5 / 5 / 4.6 / 4.5 / 4、Haiku 4.5：
+
+- 不分高峰/空闲；写缓存按 5 分钟 TTL 价计（用量无法区分 1 小时 TTL）；
+- 模型名匹配：`anthropic/` 前缀与点号版本（`claude-opus-5.5`）会归一化；带日期快照后缀
+  （`-20261101` / `-latest`）的 ID 按对应型号计价；其余变体（未收录的新版本、`-fast`、`-batch`、
+  Bedrock 写法等）一律视为未定价，不会被套上相近型号的价格；
+- **Claude Haiku 5.5 暂不收录**：其价格按单次请求的输入长度分档（≤100K / >100K），按轮汇总无法判定档位；
+- 不计入：批处理折扣、地区 / `inference_geo` 溢价、Fast mode、工具调用费；
+- 官方只有美元报价，人民币金额按**固定汇率 6.76** 折算（参考值，不跟随 `cnyPerUsd` 与实时汇率）。
 
 ### 币种与汇率（v0.3+ / v1.0.1 官方双币种）
 
 成本按**所选币种的官方刊例价**直接计算（中文定价页 CNY 报价 / 英文定价页 USD 报价，
-**不做汇率换算**，与官方账单口径一致），面板成本区可一键切换 **CNY/USD**（选择在
+**不做汇率换算**，与官方账单口径一致；例外：只有美元报价的 Claude 系列，人民币金额按固定汇率
+6.76 折算，见上文），面板成本区可一键切换 **CNY/USD**（选择在
 浏览器记住），指示器、面板、图表与成本徽章全部跟随。`config.cnyPerUsd`（默认 6.76）
 与「刷新汇率」按钮（经宿主 `/dsh-usage-chart/rate` 代理）仅用于「1 USD ≈ X CNY」
 参考注记：

docs/ARCHITECTURE.md

@@ -57,7 +57,7 @@ Client 半区（浏览器）：src/client/
 | 模块 | 接口（小） | 实现（深） | 接缝 / 适配器 |
 |---|---|---|---|
 | **RoundFold**（`usage/rounds.ts`） | `foldRounds(events) → { totals, rounds }`，`Round = { turn, buckets, model, startedAt, endedAt, durationMs, ttftMs, outputTps, endReason, userSource, cost }` | 四桶 + 轮次耗时/来源/模型/成本；0.2.x 读取 message/attempt 的 usage 与 stream，TTFT 取首 token，TPS 分母为各次流式生成时长之和；旧日志保留 chunk/usage 回退（详见 2.5） | 纯函数，无 IO；测试喂合成事件流 |
-| **PricingResolver**（`pricing/resolve.ts`） | `resolve(model) → { pricing, source, verifiedAt, estimated, known }` | 优先级解析（覆盖>内置>回退）、峰谷时段、未知模型标记、来源/时效标注 | **接缝：PricingSource**（见 2.3） |
+| **PricingResolver**（`pricing/resolve.ts`） | `resolve(model) → { pricing, source, verifiedAt, estimated, known, unpriced }` | 优先级解析（覆盖>内置>回退）、峰谷时段、未知模型标记、来源/时效标注；键匹配走 `matchPricingKey`（精确 > 最长前缀，Claude 仅接受日期快照后缀）；未收录的 Claude 不套 DeepSeek 回退价（`unpriced`，单价全零） | **接缝：PricingSource**（见 2.3） |
 | **PricingCalc**（`pricing/calc.ts`） | `estimateCost / formatTokens / buckets 数学` | 纯共享计算，两个半区 bundle 同一份 | 无 IO；不解析来源 |
 | **HostRoutes**（`src/index.ts`） | 注册 `/usage` `/balance` `/pricing` `/meta` `/rate` | 编排 RoundFold + PricingResolver + 安全守卫（沿用 `isTrustedRequest`） | 适配器：`webServer` 接缝 |
 | **HistoryStore**（`history/store.ts`，v1.2） | `appendSample(round)` / `queryRange(from, to) → rows` | 聚合/按天落盘、索引、一键清除、导出 | **接缝：历史存储**（2.3） |