Introduction¶

DeepSeek charges double fees during peak hours: 01:00–04:00 and 06:00–10:00 UTC on weekdays, excluding Chinese statutory holidays; the rest of the time is off-peak at half price. Harness’s interface gives no indication of this—you just know a long task was run tonight, but don’t know which time slot it falls into. Only when the bill comes out the next day do you realize that round of inference was charged at double rates.

dsh-peak-cost fills this gap: putting the current billing period, the countdown to the period switch, and the cumulative cost of the current session directly into the stats row below the composer. Below, I introduce the positioning, capabilities, installation method, and limitations of this plugin.

What is it¶

dsh-peak-cost (npm package name is also dsh-peak-cost) is a DeepSeek Harness plugin maintained by korotky, version 0.1.0, licensed under MIT. The community directory page categorizes it under “Model Inference.”

The problem it solves is very specific: letting you know whether the current unit price is peak or off-peak, and roughly how much this session costs, while you are still typing in the input box. It is about monitoring as you run, rather than checking the bill after the fact.

Community directory sites (like SkillHub) are independent sites with no official affiliation to DeepSeek or HF (Hanfang). They can be treated as a consolidated entry point and should not be considered an official app store.

Core Features¶

The plugin displays a pill in the composer status bar, formatted as follows:

● Off-peak · 9 h 44 m · $1.20

The three fields are: current period, countdown to switching to another period, and current session cost estimate. Peak periods display Peak instead of Off-peak.

Opening the details popover (details layer) reveals a finer breakdown:

• Four token buckets, each showing the actual unit price in effect;
• Subagent session overhead;
• Web call costs;
• Breakdown of peak/off-peak items;
• Today’s and this week’s totals;
• The version of the pricing table currently in use, as well as diagnostic information.

The plugins page also provides these settings: currency, language, refresh interval, alert interval, ledger scope, sidebar footer row, and peak alerts.

The ledger scope setting determines whether statistics cover the entire account or a single workspace; these are two different readings.

Installation and Activation¶

The installation command comes from the original README:

dsh plugin --profile web add dsh-peak-cost

After installation, you need to reload the Web UI page for the plugin to appear in the stats row below the composer.

Code block explanation: --profile web limits installation to the web profile; the plugin package declares dsh.client.platform: "web" and immediately: true, meaning the client loads only on the web platform and immediately injects into the session interface.

Typical Usage¶

View the current period and countdown. Look at the middle cell in the status row. If it shows Peak and the countdown isn’t long, you can choose to wait a bit before starting a large batch task; the countdown changes “It’s expensive now” to “It will be expensive in 24 minutes,” which is the reason countdown is the core interaction in the README.

Open the popover to see where the money went. Once a long task is running, the numbers move in real-time, rather than appearing on the bill the next day. The breakdown of four token buckets plus subagent and web calls allows you to pinpoint whether the main session or a specific subagent drove up the cost.

Verify estimation accuracy using the Usage page. The README mentions a verification: the plugin calculated 1,697 requests and 479.4M tokens for this week, while the Usage page shows 1,769 requests and 482.2M tokens. The difference comes from a platform data latency of up to 5 minutes. The verification method involves checking the same time period on both sides.

Statistics by account or by workspace. Set the ledger scope in the plugin settings. Select the entire account when you need to reconcile with the platform’s total bill, and select a single workspace when you only want to see the cost of a specific directory.

Pricing Rules and Privacy¶

Peak windows are 01:00–04:00 and 06:00–10:00 UTC on weekdays, excluding Chinese statutory holidays; the rest of the time is off-peak at half price. Each request is priced based on the rate at the moment it is initiated, so sessions crossing boundaries won’t mix rates—the two consecutive turns in the same session might be peak and off-peak respectively.

A few details:

1. Cache writes are charged at the cache-miss rate;
2. deepseek-v4-pro uses its own pricing table and does not apply the default table;
3. When encountering unknown models, it marks them rather than substituting a “seemingly reasonable” number.

Regarding privacy: The plugin does not access API keys, and aside from a holiday calendar refresh address you configure yourself, it makes no other network requests. All estimation is done locally, and the data source is the session logs that Harness already retains.

Applicable Scenarios and Notes¶

Suitable for these types of people:

• Frequently running long tasks or batch inference with significant cost fluctuations;
• Sometimes willing to wait, sometimes not, needing a minute-by-minute decision basis;
• Wanting to see how much subagent and web calls each account for.

You need to know these points before using:

• This is an estimation based on token counts, not a bill. The plugin cannot see how much the platform actually deducted, only calculates based on token counts and the pricing table, so there will be deviations from the actual bill.
• The pricing table and holiday calendar are released with the plugin. Requests before 2026-09-10 will be marked as unestimatable and won’t be calculated using guesses. If your billing history includes earlier periods, this part will show as unevaluated in the plugin.
• Under the link: installation method, file changes are not captured by hot reload. A new composition is required. The README points to ADR 0005 to explain why.
• The plugin runs with the permissions of the current dsh process. It reads local session logs, and the permission boundary is your dsh process. It is recommended to review the source code before installation to confirm the license terms (MIT).
• Runtime requirements: The engines field declares dsh >=0.2.1-rc.1 and node >=22.13. Availability is not guaranteed for versions below this.

Conclusion¶

The value of this plugin lies not in saving money, but in moving the question of “What is the current price of this round of inference and how much did it cost” from the time of the bill to the time of execution. The countdown and session accumulation are two available decision inputs: whether to run now, or whether costs are getting out of control while running.

Directory page: https://www.skillhub.cn/plugins/korotky/dsh-peak-cost
GitHub: https://github.com/korotky/dsh-peak-cost