dsh-save-money

Save-money plugin for DSH (DeepSeek Harness) — define your own "pause / resume" time windows; at pause time running long tasks are paused (not stopped) automatically, and they resume when the window ends. Built for LLM API peak/off-peak pricing (e.g. DeepSeek peak hours 9:00–12:00, 14:00–18:00 Beijing time, off-peak at half price; since 2026-08-23 weekends are off-peak all day — the plugin's weekday switch skips Sat/Sun by default), and equally useful for time-of-use electricity rates, bandwidth off-peak shifting, or any "I don't want the machine working during this period" scenario.

Status: ✅ Implemented, continuously maintained. 中文版

Interface

The colored status text in the top-right of the session header (Save · ⚪/🟢/🟡/🔴, color follows the state) is the single persistent entry — click it to open the settings popover. When a pause is upcoming or active, a reminder banner appears at the top of the page (with the End this save mode button).

With "Show balance" enabled, your official DeepSeek account balance appears next to the status text; clicking the balance opens the last-8-hours spend bar chart:

Each bar is a 10-minute window: the X axis shows whole-hour ticks, the Y axis key-point amount ticks (≤5, with gridlines), and hovering shows each window's exact time range and amount. Windows where the balance dropped without any local activity are marked in warning color (changes may come from another device).

Features

• Multiple time windows: add / remove pause-resume windows freely; supports midnight-crossing windows (23:00–08:00) and per-weekday filtering;
• Automatic pause on schedule: when the pause time arrives, running tasks are safely "frozen" (their progress is preserved exactly — nothing is interrupted or lost) and resume automatically when the window ends; if nothing is running, nothing is paused;
• No requests during the window (the saving core): inside a pause window the AI sends no new requests to the model service, so no cost is incurred; after the window ends (or on disable / end-this-window) everything resumes, with conversation context and in-flight tasks unaffected. The AI not replying inside a window (including new conversations) is expected — to resume right away, click the End this save mode button (takes effect directly, no AI involved);
• Per-model-tier save mode: decide per model tier which ones pause during windows. The settings panel shows two compact rows — Official API (the DeepSeek official API) and Other API (EVERY non-official provider: opencode go/zen, relays, SiliconFlow, …) — each with flash / pro / vision toggles (vision = deepseek-v4-flash-vision-exp), six toggles total. Checked = paused (saving money); unchecked = exempt (requests flow even during a window). Defaults: the three official tiers checked, all "other" tiers exempt; your changes are persisted and never reset. Any model whose name carries no flash/pro/vision marker (legacy chat/reasoner, unknown models, anything else) is always exempt — it can never block your requests;
• Global weekday switch: one row above the window list — 一 ☑️ 二 ☑️ 三 ☑️ 四 ☑️ 五 ☑️ 六 ⬜ 日 ⬜. Saving applies only on the checked days (default Mon–Fri — under the 2026-08-23 peak/off-peak rule weekends are off-peak all day, so nothing pauses on Sat/Sun). Composed with each window's own per-weekday filter: both must match for a window to apply;

…