dsh-peak-guard

DeepSeek Harness (DSH) 用のピーク料金ガード。使用中モデルの料金カタログを読み、ピーク倍率が適用される時間だけリクエストを停止する。手動で解除しない限りブロックされる。

ピーク時間をハードコードせず、opencode-go のモデルカタログが返す cost.multipliers を現在UTCで評価する。モデル追加・料金改定に自動追従する。

[PEAK GUARD] deepseek-v4.1-flash はピーク時間のため停止（単価倍増）
  ピーク: 平日 10:00-13:00 / 15:00-19:00 JST (UTC 01-04, 06-10)
  このピーク残り: 約42分
  1時間だけ解除: npx dsh-peak off

Install

npm (recommended):

dsh plugin --profile web add dsh-peak-guard

GitHub (no registry):

dsh plugin --profile web add github:yamahei21python/dsh-peak-guard

A bundle install adds the guard row to the profile; no manual cordis.patch.yml insert is needed. Restart dsh.

Provider configuration

The bundle only installs the guard. The model provider row stays yours (API key, endpoint, models). For OpenCode Go:

# ~/.dsh/profiles/web/cordis.patch.yml
- id: llm-pi-ai
  config:
    providers:
      opencode-go:
        displayName: OpenCode Go
        apiKeyEnv: OPENCODE_GO_KEY
        api: openai-completions
        baseURL: https://opencode.ai/inference/go/openai/v1
        headers:
          x-opencode-session: dsh-<random-stable-id>
        compat:
          supportsDeveloperRole: false
          maxTokensField: max_tokens
          supportsReasoningEffort: true
        models:
          - id: deepseek-v4.1-flash
            name: DeepSeek V4.1 Flash
            contextWindow: 1000000
            maxTokens: 384000
            input: [text, image]
            reasoningEfforts:
              off:
              low: low
              high: high
              max: max
          - id: deepseek-v4-pro
            name: DeepSeek V4 Pro
            contextWindow: 1000000
            maxTokens: 384000
            input: [text]
            reasoningEfforts:
              off:
              high: high
              max: max

- id: agent-default-model
  config:
    provider: opencode-go
    model: deepseek-v4.1-flash

Store the key with the credentials service (Models page) or export OPENCODE_GO_KEY.

Bypass

npx dsh-peak off          # 60 minutes
npx dsh-peak off 30       # 30 minutes
npx dsh-peak on           # block again now
npx dsh-peak status

State file (shared with the OpenCode guard by default):

{"until":1794017000000,"minutes":60,"set_at":"2026-11-07T02:00:00.000Z"}

Override the path with DSH_PEAK_BYPASS. The guard fails closed: a missing or corrupt file means "block during peak".

Test flags for development:

PEAK_GUARD_FORCE=peak dsh web   # always block
PEAK_GUARD_FORCE=off  dsh web   # never block

How it works

• Hooks the llm/stream waterfall.
• Fetches https://opencode.ai/inference/go/v1/models (10-minute cache; last good catalog kept on failure; 60-second negative cache).
• Evaluates each model's cost.multipliers[].when (utc_days, utc_hours with gte/lt, any/all) against the current UTC time.
• On a match with no active bypass, yields a terminal finish chunk with code: PEAK_GUARD and HTTP-429 semantics; the turn fails cleanly instead of crashing the process.
• If the catalog is unreachable, falls back to the documented weekday windows 01:00-04:00 and 06:00-10:00 UTC.

Configuration

Optional plugin config (cordis.patch.yml row can override):

- id: peak-guard
  config:
    providers: [opencode-go]
    modelsUrl: https://opencode.ai/inference/go/v1/models

Scope and limitations

• Only blocks calls routed through the configured provider ids (default opencode-go).
• Other clients (Claude Code, Codex, etc.) are not covered; use a gateway-level guard.
• DSH is a developer preview; engines.dsh is >=0.2.0-rc.1 <0.3.0.
• Prices and peak windows come from the provider catalog; this package ships none.

License

MIT