Starlight-bananice/dsh-status-bar

dsh-status-bar · Know what your agent is doing — at a glance

✅ Adapted to the DSH desktop app — DSH 0.2.0-rc.2. The 0.2.x line is built for the desktop runtime: install it into the desktop profile (in-app Plugins page, or dsh plugin --profile desktop add …). The retired 0.1.x line targets the old 0.1.x runtime and will NOT load on the 0.2.0 desktop.

English · 中文

Overview

The problem: the native DSH bottom status bar is one long, fixed line — the more it shows, the more it overflows, and on narrow windows parts of it get truncated. You cannot see the current model, how full the context window is, how fast tokens are streaming, or what a session has cost — and there is no way to arrange that information the way you work.

Who it is for: anyone running the DSH desktop app daily — power users and teams who want live session telemetry without leaving the composer, and without running a separate monitor.

What it does:

• Desktop-native, not bolted on — rebuilt against the DSH 0.2.0 desktop runtime (@deepseek-ai/dsh-*@0.2.0-rc.2): session facts come from the desktop's own Chat / session / composer services, background jobs from its job service, and the live throughput figure from its own streaming event window. No shims, no legacy compatibility path.
• Near-native experience, fully yours — 16 toggleable, reorderable segments: status dot, model, title, workspace, turns & steps, model/tool time, TTFT & decode speed, cache-hit rate, tokens, context pressure, live TPS, session time, cost estimate, jobs, queue, errors
• Live throughput (TPS) — the bar folds this session's own live stream in the browser, so the speed updates chunk by chunk while streaming and tracks the real generation rate regardless of how finely the stream is delivered; no polling, no host round-trip, no external live-stats plugin
• Cost estimation with a user-maintained model price book — per-model rates, per-model peak/off-peak schedules, each message/step priced with the model that actually produced it (input, cache-hit, cache-write and output priced separately at that step's own time), and a «Usage & cost» dialog with a stacked cost-trend chart (day / week / month), a paged per-step usage history (with a dedicated cache-hit column), and a total-cost hero
• Keeps up with DeepSeek's shifting peak/off-peak rules — no more one hard-coded list of HH:MM windows: billing keys off working day vs weekend vs public holiday, with the holiday / 调休 calendar fetched and cached locally (and editable by hand). When DeepSeek moves the windows, changes the weekend rule, or adds holidays and make-up workdays, you refresh once or edit one field instead of rebuilding the price book
• Zero-config default — 13 segments ship enabled; everything else is a checkbox away
• Useful options — multi-line wrapping (so nothing gets truncated), per-model cost estimation with peak/off-peak pricing, currency choice (CNY / USD), a quick-toggle gear menu, and a dedicated settings page with one-click reset
• 24-hour clock fields — peak windows are typed as 09:00, not whatever AM/PM format the browser locale would impose on a native time input
• Clean takeover — the plugin's bar shadows the built-in stats cell at lower priority: while loaded it renders, when unloaded the built-in line returns untouched
• Bilingual UI — client locale strings ship for English and Chinese, following the DSH locale system