dsh-llm-quotes

Readme

# dsh-LLM-quotes
<p align="left">
  <a href="https://github.com/deepseek-ai"><img src="https://img.shields.io/badge/DeepSeek_Harness-plugin-blue" alt="DSH Plugin"></a>
  <a href="https://github.com/deepseek-ai"><img src="https://img.shields.io/badge/DeepSeek_Harness-web-orange" alt="DSH Web"></a>
  <a href="LICENSE"><img src="https://img.shields.io/badge/license-MIT-blue.svg" alt="MIT license"></a>
</p>

 [中文](#中文)
 
> Latest LLM provider API prices, right inside **DeepSeek Harness (dsh)** → **Settings → Models**.

A dsh web plugin that shows current LLM API prices for every provider you have configured. Each provider card gets a compact **Quotes** block with per-model input/output prices and a details popup — no tab-switching, no manual lookups.

Data is sourced from [LLMRates.ai](https://www.llmrates.ai) (free, read-only, no API key, updated daily) and cached locally.

![Settings → Models quote blocks](./assets/screenshot-1.png)

## What

`dsh-LLM-quotes` injects a **Quotes** block under every **Settings → Models** provider card in dsh:

- Per-model **input / output** prices (per 1M tokens by default).
- A **Details** popup showing every valued model field — context window, max output, modalities, capabilities, release/knowledge dates, and all price rows (tiers, unit, region, source).
- A **Compare** dropdown in the Details popup — side-by-side comparison against up to 10 other models (see [Compare across providers](#compare-across-providers)).
- Smart matching from dsh provider routes to LLMRates slugs, with manual association only as a last resort.
- A **watchlist** (star toggles in the quote tables and the details popup): followed models are tracked in `~/.dsh/llm-quotes-watchlist.jsonl` with price history — unfollow pauses the record instead of deleting it, and price snapshots are taken after each dataset refresh, so trends stay computable.

Configured models are implicitly priced; the watchlist is an explicit extra for tracking.

## Watchlist & price signals

Click the star next to a model (quote table or details popup) to **follow** it — the model is tracked in `~/.dsh/llm-quotes-watchlist.jsonl` with its price history. Unfollow **pauses** the record instead of deleting it, so history survives; the star goes off and the model stops signaling.

![Watchlist stars in Settings → Models](./assets/my-watchlist.png)

In the composer's **select model menu** (bottom-right of the input box), every currently watched model carries a price-change signal dot: 🔴 red = price(s) up, 🟢 green = price(s) down, ⚪ gray = watched but unchanged (or not enough history yet). Hover a dot for the from → to details per field.

![Model menu signal dots](./assets/model-menu-signal.png)

![Price up tooltip](./assets/price-up.png) ![Price down tooltip](./assets/price-down.png)

Price history grows automatically: after each dataset refresh the host snapshots the prices of all active watchlist records (deduplicated — history only grows when prices actually change), and the comparison only covers the explicit `PRICE_FIELDS` contract (same-currency snapshots; a currency change is reported as such, never compared numerically).

## Compare across providers

From any model's **Details** popup, open **Compare** to check the same model across providers. A provider tree loads each provider's models lazily on expand (each node is disabled while loading); checking models builds a side-by-side table aligned to the target model's field set — the target is the first column, and a model that lacks a target field shows `—`. Up to 10 models can be added (each column has a remove button); prices are normalized to the target's currency via cached ECB/Frankfurter rates; DeepSeek-style off-peak rows are standardized onto the standard tier; the standout value of each row (max for capacity, min for price) is highlighted; and the selection is persisted in localStorage so reopening the popup restores it.

![Model compare in the Details popup](./assets/compare%20prices.png)

## Why

LLM prices change constantly and live across a dozen vendor pages. This plugin keeps current quotes in the one place you already manage your models, so you can sanity-check costs while configuring a provider — without leaving the settings dialog.

## Install

Requires a running DeepSeek Harness (**dsh**) **web** profile (Node.js ≥ 22.19 or ≥ 24).

Install this plugin into your web profile with the dsh CLI:

```bash
dsh plugin --profile web add "github:v587d/dsh-LLM-quotes"
```

For a reproducible, tamper-resistant install, pin a commit SHA instead of a moving branch:

```bash
dsh plugin --profile web add "github:v587d/dsh-LLM-quotes#<commit-sha>"
```

Then (re)launch the web profile and open **Settings → Models** — each provider card now shows a Quotes block:

```bash
dsh --profile web        # or: npx @deepseek-ai/dsh web
```

> Plugins are third-party code that runs with your agent's privileges and may run install-time build scripts. Review the source and pin a commit before installing — see the [dsh plugin docs](https://github.com/deepseek-ai/deepseek-harness).

### Building from source (development)

To hack on the plugin or build it yourself, clone the repo and use pnpm:

```bash
pnpm install
pnpm run build        # outputs lib/ (host half) + lib/client.js (browser half)
pnpm run typecheck    # optional: type check
pnpm test             # optional: run unit tests
```

The bundle patch in `cordis.patch.yml` registers the host plugin row; the browser half is a `window.__ModuleLoader__` module and the host half a Cordis plugin exposing same-origin `/api/llm-quotes` JSON routes (LLMRates.ai is never called directly from the page).

## Features

- **Per-provider quote blocks** in Settings → Models (one block per configured provider).
- **Model details popup** — every valued model-level field and price row.
- **Compare across providers** — from the Details popup, side-by-side pricing/fields against up to 10 other models, with currency normalization (cached ECB/Frankfurter FX), off-peak tier standardization, best-value highlighting, and localStorage persistence.
- **Slug-based provider matching** — alias map + provider-level association; per-model association only as a last resort.
- **Watchlist with price history** — star toggles follow/unfollow; unfollow pauses the `~/.dsh/llm-quotes-watchlist.jsonl` record (history kept), and active records get a price snapshot after each dataset refresh. The composer model menu shows per-model price-change dots (red/green/gray) with from → to details on hover.
- **Daily auto-sync** — dataset refreshed at most once per calendar day; serves the last good local snapshot on failure.
- **Local-only persistence** — settings & associations in `~/.dsh/llm-quotes.json`; dataset snapshot in `~/.dsh/llm-quotes-data.json`; watchlist in `~/.dsh/llm-quotes-watchlist.jsonl`. No accounts, no telemetry, no cloud.
- **Plan-aware** — token/coding plans are excluded from provider association (only usage-based API pricing is shown).

## Data Source & Attribution

Pricing data is provided by **LLMRates.ai** and is free, read-only, and updated daily. Per their redistribution terms, the source is attributed:

<a href="https://www.llmrates.ai" target="_blank" rel="noopener">LLM &amp; GenAI pricing data by LLMRates.ai</a>

## Roadmap

- **Price-trend tracking** — chart historical prices for the models you actually use and surface changes over time (price history is already collected).
- **Watchlist alerts** — get notified on price drops or increases for followed models.
- **Standalone prices panel** — migrate the built but unregistered prices panel (its own search/compare view) into a Settings seat; the Details-popup compare is already shipped.
- **More locales** — beyond English / Simplified Chinese.

## Contributing

PRs and issues are welcome! The project is iterating quickly right now, so please **rebase** your branch onto the latest `main` before opening a PR to avoid messy merge conflicts.


## License

MIT.

---

# 中文

> 最新的大模型（LLM）厂商 API 价格，直接显示在 **DeepSeek Harness（dsh）** 的 **设置 → 模型** 页面中。

这是一个 dsh Web 插件，会为你已配置的每一个厂商展示当前的 LLM API 价格。每个厂商卡片下方都会插入一个简洁的 **报价（Quotes）** 区块，包含每个模型的输入/输出价格，以及详情弹窗——无需切换标签页，也无需手动查询。

价格数据来自 [LLMRates.ai](https://www.llmrates.ai)（免费、只读、无需 API Key、每日更新），并在本地缓存。

![设置 → 模型 报价区块](./assets/screenshot-1.png)

## 这是什么

`dsh-llm-quotes` 会在 dsh 的 **设置 → 模型** 页面中，为每个厂商卡片下方注入一个 **报价** 区块：

- 每个模型的 **输入 / 输出** 价格（默认按每百万 token 计价）。
- 一个 **详情** 弹窗，展示模型的所有有效字段——上下文窗口、最大输出、模态、能力、发布/知识截止日期，以及全部价格行（档位、计价单位、区域、来源）。
- 详情弹窗内的 **对比** 下拉——与最多 10 个其它模型并排对比（参见 [跨厂商对比](#跨厂商对比)）。
- 智能匹配：从 dsh 厂商路由映射到 LLMRates 的 slug，仅在最后兜底时才需要手动关联。
- **关注列表（watchlist）**：报价表格与详情弹窗中的星标可关注模型，关注记录连同价格历史存入 `~/.dsh/llm-quotes-watchlist.jsonl`——取消关注只会将记录标记为 `paused`（历史保留），每次数据集刷新后会自动为关注中的模型追加价格快照，便于后续做走势分析。

已配置的模型隐式纳入报价；关注列表是额外的显式追踪功能。

## 关注列表与价格信号

点击模型旁的星标（报价表格或详情弹窗）即可**关注**该模型——记录连同价格历史存入 `~/.dsh/llm-quotes-watchlist.jsonl`。取消关注只会将记录标记为 `paused`（历史保留），星标熄灭，同时该模型不再显示信号点。

![设置 → 模型 中的关注星标](./assets/my-watchlist.png)

在输入框右下角的**模型选择菜单**中，每个当前关注中的模型都会带一个价格变化信号点：🔴 红点 = 涨价，🟢 绿点 = 降价，⚪ 灰点 = 价格不变（或历史尚不足）。鼠标悬停在点上即可看到每个字段的 from → to 明细。

![模型菜单信号点](./assets/model-menu-signal.png)

![涨价提示](./assets/price-up.png) ![降价提示](./assets/price-down.png)

价格历史自动累积：每次数据集刷新后，host 会为所有 active 关注记录追加价格快照（已去重——价格没变就不会增长历史）；对比只覆盖显式的 `PRICE_FIELDS` 契约字段，且仅在币种一致的快照间进行（币种变化会明确提示，绝不跨币种比价）。

## 跨厂商对比

在任意模型的**详情**弹窗中打开 **对比**，即可跨厂商对比同一模型。厂商树在展开时才按需加载各厂商的模型（每个节点加载中禁用重复点击）；勾选模型后生成对齐目标模型字段的并排表格——目标模型为第一列，某模型缺少目标字段时显示 `—`。最多可添加 10 个模型（每列都有移除按钮）；价格通过缓存的 ECB/Frankfurter 汇率统一换算为目标币种；DeepSeek 这类 off-peak 优惠价会被统一到标准价档位；每行的“最值”（容量取最大、价格取最小）会高亮；选择结果存入 localStorage，重新打开弹窗即可恢复。

![详情弹窗中的模型对比](./assets/compare%20prices.png)

## 为什么

LLM 价格变动频繁，且分散在各家厂商的页面中。本插件把最新报价放在你本来就管理模型的地方，让你在配置厂商的同时就能核对成本——不必离开设置对话框。

## 安装

环境要求：已运行的 DeepSeek Harness（**dsh**）**Web** 配置文件（Node.js ≥ 22.19 或 ≥ 24）。

使用 dsh 命令行将该插件安装到你的 Web 配置中：

```bash
dsh plugin --profile web add "github:v587d/dsh-LLM-quotes"
```

若希望安装可复现、防篡改，建议固定到具体的 commit SHA，而不是会移动的分支：

```bash
dsh plugin --profile web add "github:v587d/dsh-LLM-quotes#<commit-sha>"
```

随后（重新）启动 Web 配置并打开 **设置 → 模型**，每个厂商卡片下方即会出现报价区块：

```bash
dsh --profile web        # 或：npx @deepseek-ai/dsh web
```

> 插件属于第三方代码，会以你 agent 的权限运行，并可能在安装时执行构建脚本。安装前请审阅源码并固定到具体 commit——参见 [dsh 插件文档](https://github.com/deepseek-ai/deepseek-harness)。

### 从源码构建（开发）

如需改动插件或自行构建，克隆本仓库并使用 pnpm：

```bash
pnpm install
pnpm run build        # 生成 lib/（主机端）与 lib/client.js（浏览器端）
pnpm run typecheck    # 可选：类型检查
pnpm test             # 可选：运行单元测试
```

仓库中的 `cordis.patch.yml` 会注册主机端插件行；浏览器端是一个 `window.__ModuleLoader__` 模块，主机端是一个 Cordis 插件，对外暴露同源的 `/api/llm-quotes` JSON 接口（页面不会直接调用 LLMRates.ai）。

## 功能

- **按厂商展示报价区块**：在「设置 → 模型」中，每个已配置厂商各显示一块报价。
- **模型详情弹窗**：展示模型所有有效字段与价格行。
- **跨厂商对比**：在详情弹窗中与最多 10 个其它模型并排对比价格/字段，支持币种归一（缓存的 ECB/Frankfurter 汇率）、off-peak 优惠档统一到标准档、最值高亮，以及 localStorage 持久化。
- **基于 slug 的厂商匹配**：内置别名映射 + 厂商级关联；仅在最后兜底时做模型级手动关联。
- **带价格历史的关注列表**：星标关注/取消关注；取消关注只是把 `~/.dsh/llm-quotes-watchlist.jsonl` 中的记录标记为 `paused`（历史保留），数据集刷新后自动为关注中的模型追加价格快照。输入框的模型选择菜单中会为关注中的模型显示价格变化信号点（红/绿/灰），悬停可见 from → to 明细。
- **每日自动同步**：数据每天最多刷新一次；失败时继续提供上一次成功的本地快照。
- **纯本地存储**：设置与关联存于 `~/.dsh/llm-quotes.json`，数据集快照存于 `~/.dsh/llm-quotes-data.json`，关注列表存于 `~/.dsh/llm-quotes-watchlist.jsonl`。无账号、无埋点、无云端。
- **区分订阅计划**：厂商关联时会排除 Token / Coding Plan 等订阅型计划，仅展示按量计费（per-token）的 API 价格。

## 数据来源与署名

价格数据由 **LLMRates.ai** 提供，免费、只读、每日更新。根据其再分发条款，须标注数据来源：

<a href="https://www.llmrates.ai" target="_blank" rel="noopener">LLM &amp; GenAI pricing data by LLMRates.ai</a>

## 路线图

- **价格走势跟踪**：基于你常用的模型，绘制历史价格曲线并提示变化（价格历史已在持续采集）。
- **关注列表提醒（Alerts）**：关注模型的降价/涨价提醒。
- **独立价格面板**：把已构建但未注册的价格面板（其自带的搜索/对比视图）迁移到某个设置入口；详情弹窗内的对比已上线。
- **更多语言**：在英文 / 简体中文之外增加更多语言。

## 贡献

欢迎提交 PR 与 Issue！项目近期迭代频繁，请在提交 PR 前将分支 **rebase** 到最新的 `main`，以避免复杂的合并冲突。


## 许可证

MIT。