deepseek-peak-valley-router

English | 简体中文

A peak/valley pricing auto model router for DeepSeek Harness: it detects whether the DeepSeek API is currently in its peak or off-peak pricing window and automatically picks the model accordingly — use the cheap deepseek-v4-flash during peak hours to save money, and the stronger deepseek-v4-pro during off-peak hours for better quality. A model you explicitly select is always respected; the plugin never takes away your choice.

> Background: since 2026-08-17 DeepSeek prices its V4 API with peak/valley pricing — peak hours cost 2× the off-peak price. The official peak windows are defined in Beijing time as 09:00-12:00 and 14:00-18:00; everything else is off-peak. This plugin schedules the model for you automatically.

Features

• Automatic peak/valley detection: converts the machine's wall clock to a reference time zone (default Asia/Shanghai, configurable) so detection is correct under any system time zone / DST
• Automatic routing: peak → deepseek-v4-flash; off-peak → deepseek-v4-pro
• Your explicit choice wins: any model you pick in the model selector is routed to directly (same provider — regardless of what the request config currently says), detected via changes to the agent-default-model setting
• Never overreaches: requests on non-DeepSeek providers or non-V4 models are passed through untouched
• Auditable: every switch is a proper return from the agent/request waterfall and is recorded in the session's request/header event — no silent drift
• Chat tools:

- peakvalley_status — current window, reference-time clock, target model, routing mode (auto / respect), and your explicit selection - peakvalley_reset — restore automatic routing (clears the in-memory user-selection record until you pick a model again in the UI)

• Robustness: an invalid timeZone config falls back to the default zone with a one-time warning — requests never fail; the startup fallback is lazily loaded on the first request, independent of host service mount order
• Edge handling: when switching pro → flash, reasoningEffort is dropped (flash may only support off) so the request is never rejected

Official windows

> Prices are subject to DeepSeek's official announcements (see links at the bottom). If the official windows change, edit PEAK_WINDOWS in lib/timezone.js.

Installation

Option A: Dynamic plugin (recommended to try, no restart)

In a Harness conversation, create the plugin with the cordis_define tool, paste the full HOST_CODE from [dynamic/host-code.js](dynamic/host-code.js) into the code.host parameter, then activate it with cordis_run.

Option B: Composition plugin (official bundle install, permanent, requires restart)

This package is a standard DSH bundle (declares dsh.bundle + cordis.patch.yml), so use the official install path:

# local checkout
dsh plugin --profile web add ./deepseek-peak-valley-router

# or from GitHub
dsh plugin --profile web add github:Eligahyu/deepseek-peak-valley-router

# once published to npm (recommended — no build authorization needed):
dsh plugin --profile web add deepseek-peak-valley-router

dsh --profile web   # restart to load

Optional time zone config (default Asia/Shanghai): override in your profile's own patch layer:

# append to ~/.dsh/profiles/web/cordis.patch.yml
- id: deepseek-peak-valley-router
  config:
    timeZone: 'America/New_York'

> Dependencies: runs inside DeepSeek Harness; @deepseek-ai/dsh-tools is provided by the host (see peerDependencies in package.json) and resolves through the launcher-maintained flat module fallback — nothing to install manually.

Behavior rules (priority, highest first)

1. A model you explicitly selected — always wins. Every model you pick in the selector is written to the agent-default-model setting; the plugin listens for that change and then routes directly to your chosen model for every request on that provider — even when the request config carries a frozen value from a session/subagent. If the default model is already pro (not the factory-default flash) on the first request after startup, it is treated as your historical choice. 2. Automatic routing (default). When no explicit selection is detected, switch by reference time zone. 3. Never overreach. Requests on non-DeepSeek providers or non-V4 models pass through unchanged.

How it works

• Time zone conversion: lib/timezone.js uses Intl.DateTimeFormat (IANA time zones, hourCycle: 'h23') to convert the machine's wall clock to the reference zone — correct under any system time zone / DST; an invalid time zone falls back to the default zone, and to the UTC+8 offset (Asia/Shanghai has no DST) when Intl is unavailable. The request path never throws.
• Routing seam: hooks the agent/request waterfall — next() yields the frozen LlmCallConfig; returning a replacement switches the model, and the loop records the switch in request/header.
• Your choice: Harness's own session model-selection mechanism (installModelSelection) already guarantees that an explicit user selection wins; the plugin also actively routes to your selection in its own respect check as a second layer.

> Runtime note: for sessions created before the plugin was activated, the model is governed by the session's own selection-freeze mechanism — picking a model in the UI takes effect immediately and is respected; automatic routing takes effect for such a session after the next page refresh / reconnect (agent re-creation), and immediately for new sessions and all subagents. This is the intended behavior of Harness's model-routing architecture (frozen config + session selection priority).

> Variant difference: the dynamic variant additionally exposes the harness.handle('peak-valley-status') host method (for a future Client UI); the composed variant has no host channel and exposes status through the peakvalley_status chat tool. Everything else behaves identically.

Verification

• Run the tests: npm test (Node ≥ 20, node --test, 21 cases: time-zone logic + routing behavior)
• Watch the logs: entering peak/off-peak prints peak-valley: 进入高峰时段(...),DeepSeek 模型路由 -> deepseek-v4-flash; each switch prints peak-valley: deepseek-v4-flash -> deepseek-v4-pro (...)
• Call peakvalley_status in a conversation to see the current state
• The session's request/header events record the model actually used by every request

Repository layout

deepseek-peak-valley-router/
├── .github/workflows/test.yml # CI: node --test on Node 20/22/24
├── cordis.patch.yml           # bundle patch (official dsh plugin add entry)
├── lib/
│   ├── index.js          # composition plugin (ESM, reads config.timeZone)
│   └── timezone.js       # pure logic: time-zone conversion + peak/valley check (zero deps)
├── dynamic/
│   └── host-code.js      # dynamic plugin variant (cordis_define code.host source)
├── test/
│   ├── timezone.test.js  # conversion, boundaries, cross-midnight
│   └── router.test.js    # routing behavior: auto route / user pick / tz fallback / tools
├── package.json
├── CHANGELOG.md
├── README.md
├── README.en.md
└── LICENSE

Disclaimer

• Peak/valley windows and prices are subject to the official DeepSeek announcements; this plugin implements them as announced and makes no guarantee about future pricing policy.
• Automatic model switching may change answer quality/behavior; for important tasks, check the model actually used (see request/header in the session log).
• This plugin is not affiliated with or endorsed by DeepSeek.

References

• DeepSeek API models & pricing
• DeepSeek-V4 API repricing: first peak/valley billing (TechWeb)
• DeepSeek API enables peak/valley pricing (CWW)

License

[MIT](LICENSE)