dsh-deepseek-pricing

A DeepSeek Harness plugin that answers the two questions that decide a DeepSeek API bill: is the API on peak (full price) or off-peak (exactly half price) right now, and when does it flip. It also reports today's peak windows and the live list prices for the tier you are in.

Registering one tool:

Tool	What it returns
`deepseek_pricing_tier`	`tier` (`peak`/`off`), `reason`, Beijing wall clock, today's peak windows, the next switch time + countdown, per-model prices for the live tier, and optionally the upcoming switches

Install

dsh plugin --profile web add "github:ybbms777/dsh-deepseek-pricing"

The package declares dsh.bundle.patch, so dsh plugin activates it as a profile layer. Restart dsh --profile web afterwards, then check Settings → Plugins. Peer dependencies: @deepseek-ai/cordis (^4.0.1) and @deepseek-ai/dsh-tools (^0.0.1-rc.1).

Why an agent wants this

Off-peak is exactly 50% of peak and is decided by when a request finishes. Before launching a large batch, retrieval sweep or evaluation run, the agent can ask:

peak hours are Beijing (UTC+8) Mon–Fri 09:00–12:00 and 14:00–18:00, excluding Chinese public holidays; everything else — weekday nights, the 12:00–14:00 lunch break, all weekend, all holidays — is half price.

So a job that can wait costs half as much. The tool is pure arithmetic on DeepSeek's published rate card: no network calls, no API key, no credentials, isConcurrencySafe is true, and it never writes anything.

Use

Once mounted, the model calls deepseek_pricing_tier on its own. Typical args:

{ "at": "2026-10-12T09:30:00+08:00", "format": "summary" }  // evaluate a specific instant
{ "days": 7, "format": "json" }                             // now + upcoming switches

Rendered output for a peak instant:

PEAK — DeepSeek bills full price right now
now (Beijing): 2026-10-12 09:30:00 (Monday) — outside the weekday peak windows (Beijing 09:00-12:00, 14:00-18:00)
today's peak windows: 09:00-12:00, 14:00-18:00 (Beijing)
next switch: 2026-10-12 12:00:00 → off-peak (half price) (in 0d 02:30:00)
deepseek-flash: in(cache miss) 2 / out 8 RMB per 1M — DeepSeek-V4.1-Flash, concurrency 2500
deepseek-v4-pro: in(cache miss) 9 / out 27 RMB per 1M — DeepSeek-V4-Pro（0813 版）, concurrency 500

One implementation, two consumers

The peak/off-peak rule is not re-implemented here. tier.js is generated by tools/sync-core.mjs from the /*CORE-START*/ … /*CORE-END*/ block of deepseek-pricing-panel's index.html, at a pinned commit, verbatim:

node tools/sync-core.mjs          # regenerate tier.js
node tools/sync-core.mjs --check  # CI: fail if tier.js drifted from the pin

The generated file carries the source commit and the CORE block's SHA-256, and tests/core.test.mjs asserts the header — so the model-facing tool and the human-facing HTML panel can never disagree about the rule.

Tests

npm ci
node --test

16 tests, no test framework beyond Node's own runner. tests/plugin.test.mjs boots the real harness packages: a real @deepseek-ai/cordis context with a real @deepseek-ai/dsh-tools ToolRegistry, mounts the plugin with ctx.plugin(), and executes the tool through the registry pipeline — asserting the model-facing schema, the rendered summary, the JSON value, the error path for an unusable at, and the prompt-section wiring. The only stand-in is the systemPrompt service the registry injects (a two-method stub). CI runs the suite on Ubuntu + Windows × Node 20/22 plus the drift check.

Model names

deepseek-flash is DeepSeek-V4.1-Flash, not V4-Flash: the legacy ids deepseek-v4-flash / deepseek-v4-flash-vision-exp still resolve, but those models are retired and requests are served by V4.1-Flash at Flash prices. deepseek-v4-pro is DeepSeek-V4-Pro. Both ids and versions were verified against a live GET api.deepseek.com/models.

Caveats

Prices and the schedule are transcribed from DeepSeek's published pricing page and can change; this plugin visualises the published rule, it does not enforce it.
A government make-up workday (调休上班日) is off-peak under the rule, because DeepSeek distinguishes only Mon–Fri from weekend.
The 2026 China holiday calendar is built in; a new year's notice needs a re-pin (tools/sync-core.mjs after updating HOLIDAYS upstream in the panel).

License

MIT.

简体中文

这是一个 DeepSeek Harness 插件，回答决定 DeepSeek API 账单的两个问题：现在按高峰（全价）还是空闲（正好半价）计费、下一次切换在几点，并给出今天的高峰窗口与当前生效单价。

注册的工具：deepseek_pricing_tier —— 返回 tier（peak/off）、判定原因、北京时间、今天的高峰窗口、下次切换时刻与倒计时、各模型当前单价，可选未来若干天的切换列表。

安装：

dsh plugin --profile web add "github:ybbms777/dsh-deepseek-pricing"

包内声明了 dsh.bundle.patch，dsh plugin 会把它作为 profile 层启用；重启 dsh --profile web 后在「设置 → 插件」中查看。

为什么 Agent 需要它：空闲价正好是高峰的 50%，只取决于请求什么时候跑完。高峰 = 北京时间（UTC+8）周一至周五 09:00–12:00、14:00–18:00（不含中国法定节假日）；其余全部时段 —— 工作日夜里、午休、整个周末、整个节假日 —— 都是半价。所以能等的批处理任务，成本直接减半。纯本地算术：不联网、不需要 API key、不写任何文件。

一份实现，两个消费者：峰谷规则没有在这里重写。tier.js 由 tools/sync-core.mjs 从 deepseek-pricing-panel 的 index.html 里 /*CORE-START*/ … /*CORE-END*/ 段按固定 commit 原样生成，文件头记录来源 commit 与 CORE 段的 SHA-256，node tools/sync-core.mjs --check 在 CI 里守住不漂移 —— 因此面向模型的工具与面向人的 HTML 面板不可能对规则产生分歧。

测试：npm ci && node --test，16 个用例、只用 Node 自带 runner。其中 tests/plugin.test.mjs 会真的启动 harness 包（真实的 @deepseek-ai/cordis 上下文 + 真实的 dsh-tools ToolRegistry），用 ctx.plugin() 挂载插件并走注册中心管线执行工具。

注意：deepseek-flash 是 DeepSeek-V4.1-Flash（不是 V4-Flash，旧模型名已下线但由 V4.1-Flash 承接）；调休上班日按规则算空闲半价；单价与规则以官方定价页为准。

安装

说明文档