dsh-lowtide

dsh-lowtide

Queue tasks when it's busy. They run on their own when it's cheap.

Semi-auto or full-auto, off the model's peak hours and peak prices. A must-have plugin for DeepSeek Harness.

English | 简体中文 | 繁體中文 | العربية | Deutsch | Español | Français | Italiano | 한국어 Three tasks waiting in the queue, the price status glowing in the session header, auto-run when your window opens Introduction

lowtide is a plugin for DeepSeek Harness (dsh). The problem it solves is plain and perfectly natural:

Usually, when you want an agent to do some work, you sit at the computer, send the agent an instruction, wait for the reply, and then review it by hand. But this workflow seems to forget that you have plenty of idle time — and a chance to dodge the peak/off-peak pricing that some models charge.

With lowtide installed, the day goes like this: whenever a job crosses your mind during the day, toss it in the queue, glance at it, release it. The tasks pile up until the time you set (say, after 7 PM — that's when DeepSeek is at valley pricing), then run by themselves. Next morning you open the report: keep what went well, send back what didn't.

That's all it is. But use it for a week and your working rhythm genuinely slows down — and don't forget, "time is money, efficiency is life"...

A few key capabilities:Four execution strategies: single, iterative, sampling, review — from "one pass is enough" to "run five candidates and I'll pick"168 unit tests + 10 end-to-end specs, CI green across ubuntu / windows × node 22 / 24One build artifact serves both desktop and web — install once, works on bothOff-peak execution lands in DeepSeek's valley hours: the same batch costs about half of what peak wouldA normal day with an agent might look like this…

Ten minutes before clocking out. You've finished reviewing code, so you file three tickets for tomorrow: a refactor (iterative, 3 rounds), a weekly report (single), and a design you're unsure about (sampling, 4 candidates). Release them all, shut down, leave. Tomorrow at your desk, the morning report says: the refactor's done, the report's drafted, and four candidate designs sit side by side, each with its cost written out.

Friday night. Queue a week's worth of chores in one go: dependency cleanup, missing tests, data scripts. Weekends are valley price around the clock. You go out; it works from home. Monday you check the report — retry what failed, merge what's good.

A 10 AM brainwave. You're mid-conversation with the agent about an urgent bug when you think "hey, update the docs too." The intercept card pops up: running now costs peak price, tonight it's about half — the difference is spelled out. Click "queue for off-peak"; your draft survives untouched, and you go back to the bug.

An always-on server. You've got a machine running dsh 24/7. Switch to L3 full-auto, then file tasks from anywhere through the API (POST /ds-lowtide/tasks). It runs them on schedule and writes the report. Nobody's watching, but the sandbox, the daily budget, and the file locks are all still there.

Deliverables for a client. Use the review strategy: run once, then automatically open an independent session that tears the result apart through your chosen focus (say, "hunt for data-source errors"). In the morning you don't get a bare result — you get a result plus a critical review.