Deepseek Peak

OpenCode V2 plugin that reports whether a request falls in an official peak or off-peak pricing period.

The package name is historical: the plugin now works across providers. It uses built-in official schedules where they are known and explicit pricing headers returned by any API. When neither is available it shows no indicator instead of guessing.

The sidebar evidence also names the next schedule change, for example Official schedule · ▼ 07:00 in your local time (10:00 UTC when the timezone setting is UTC): ▲ means peak starts, ▼ means peak ends, and the weekday is added when the change lands on another day. A trailing * means the period is computed from the schedule before any request was observed. The line never wraps and refreshes itself while the session is idle. The /deepseek-peak dialog lists the full peak windows.

Indicators follow each tab's session and selected model independently, including when only the prompt footer is visible. Changing to a model without a known pricing period hides the previous indicator; late previews cannot replace a newer refresh or provider event.

Screenshot

Providers

• DeepSeek — peak 01:00–04:00 UTC and 06:00–10:00 UTC, Monday through Friday, excluding Chinese public holidays.
• Z.ai Coding Plan — peak 14:00–18:00 Singapore time (UTC+8), Monday through Friday. Applies only to the coding endpoints (/api/coding, /api/anthropic, /api/v1); the standard pay-as-you-go API has flat pricing and is not matched.
• Alibaba Model Studio — busy hours 08:00–22:00 Beijing time (UTC+8), every day, for the deepseek-v4.1-flash, deepseek-v4-pro and deepseek-v4-flash model families (dated snapshots such as -1017 and -0813 included) served from dashscope.aliyuncs.com. The international endpoint is not matched.
• Alibaba Token Plan — full Credits 08:00–22:00 Beijing time (UTC+8), every day; the published night discount (50% off) applies 22:00–08:00 for the deepseek-v4.1-flash, deepseek-v4-pro and deepseek-v4-flash families (dated snapshots included), and 60% off in the same window for qwen3.8-max and qwen3.8-flash (dated snapshots included). This follows the current limited-time promotion.

How detection works

Profiles are resolved in this order:

1. Exact provider ID — this keeps working behind local proxies such as http://127.0.0.1:8787/v1.
2. Request base URL against official endpoints, for custom-named providers.
3. Model filter for provider products that cover only some models. The filter matches model families, so a dated snapshot such as deepseek-v4.1-flash-1017 inherits the deepseek-v4.1-flash schedule without a plugin update.

If nothing matches and no pricing header arrives, the plugin shows no indicator and /deepseek-peak explains that no official schedule is known — it never guesses a period.

API response headers always win over schedules. The accepted header names stay the same:

• x-deepseek-pricing-tier: peak|off-peak
• x-deepseek-price-period: peak|off-peak
• x-deepseek-peak: true|false
• x-pricing-tier: peak|off-peak

A header is cached for the configured TTL. When an API header disagrees with the schedule, the indicator marks it (⚠).

Commands

• /deepseek-peak — show the last observed status and evidence.
• /deepseek-peak-sections — toggle toast, footer indicator, sidebar indicator, sidebar evidence, toast frequency, timezone (local or UTC), and language.
• /deepseek-peak-lang — switch the display language.

The command names are kept for compatibility.

Language

The interface follows the environment language (LANG/LC_ALL) and defaults to English. English, Portuguese, Simplified Chinese, Japanese, and Korean are available; change it from /deepseek-peak-lang or the sections menu. Chinese public holidays are localized too (for example 中秋节 in Chinese and Mid-Autumn Festival in English).

Installation

Once published to npm:

npm install -g opencode-deepseek-peak
opencode-deepseek-peak

The installer adds the package to the OpenCode server and CLI configurations as well as tui.json(c) for OpenCode V1, preserving existing settings. Restart OpenCode after installation.

For manual configuration, add the package to both plugin lists:

{
  "plugins": ["opencode-deepseek-peak"]
}

Optional server settings:

{
  "plugins": [{
    "package": "opencode-deepseek-peak",
    "options": {
      "providerIDs": [],
      "apiSignalHeaders": ["x-deepseek-pricing-tier"],
      "apiSignalTtlMs": 300000
    }
  }]
}

providerIDs is an optional allowlist: leave it empty or omit it to observe every provider. apiSignalHeaders and apiSignalTtlMs are unchanged in meaning: which response headers to accept and how long a header stays cached.

OpenCode V1

OpenCode V1 loads TUI plugins from tui.json (plugin list) and has no server-side plugin hooks, so the V1 build computes schedules locally per session from the provider/model, resolving the same profiles (the DeepSeek schedule is the fallback when the model is unknown). Sidebar and prompt-right indicators, commands, toast on period change, settings, and i18n all work. Only the API header source of truth is unavailable there.

// ~/.config/opencode/tui.jsonc
{
  "$schema": "https://opencode.ai/tui.json",
  "plugin": ["opencode-deepseek-peak"]
}

Development

npm install
npm run typecheck
npm test
npm run build