dsh-trading
Trading research workbench for DeepSeek Harness — market-data seam, BYO data providers, and model-facing analysis tools
一句话定位
Trading research workbench for DeepSeek Harness — market-data seam, BYO data providers, and model-facing analysis tools
安装方式
在 DSH 的 web profile 中执行下面的命令;装完刷新页面即可。
$ dsh plugin --profile web add dsh-trading@0.0.1
兼容的 DSH 版本
未声明
作者与热度
maddogfinance · ★ 41 · 更新于 2026-11-06
关于这个插件
dsh-trading
A trading research workbench built as plugins for DeepSeek Harness (dsh). No fork, no patched core — just a bundle you stack on the stock web or headless profile. Status: early scaffold. dsh itself is in developer preview and moves fast; expect breaking changes on both sides. The workbench
Under dsh web with the optional shell frame, the app becomes chart-first: a persistent, live chart column on the left, the conversation on the right.
The column is yours: type a symbol, pick a timeframe, and it fetches over a loopback channel without going through the model at all. It also follows the conversation — when the agent charts something, the column loads that instrument *live* rather than mirroring the agent's frozen snapshot. Touching the chart at all — a symbol, a timeframe — pins it, and a 📌 chip says so; from then on the agent's drawings arrive as a Show offer on a pill instead of taking the chart out from under you. Click the chip to start following again.
Drawings work the same way in both directions. annotate_chart levels, zones and paths land on the live column; the *prose* half of the same analysis — the level table, the bull/bear scenario cards — stays in chat, where reading text belongs. And the loop closes: the panel publishes what it is showing back to the host, so the agent can read your chart (get_chart_view, plus a one-line context injection each turn) instead of asking you to screenshot it. The walkthrough
The Market Analyst does not hand you a report. It talks you through the chart the way an analyst sitting beside you would: the answer first, then one piece of evidence at a time, pointing as it goes. Each section is one walkthrough call. In the chat its lines arrive one at a time, at reading pace; on the column, each line moves the chart: the section's timeframe and window, a pen circle around the bar it names, a highlighter over the zone that decides it, a lime pointer on whatever is being talked about. Click any earlier line and the chart goes back to that moment. Drag the chart and the analyst stops and waits for you. When it ends a turn on a question, the replies it offers send a real message.
Every mark is pinned to a bar ({ time, field }) and the tool reads the value off the data, so the analyst cannot circle a number the bar never printed; a level can cite a checkable source ("the lowest low since 9/11"). A wrong number fails the call with the bar where it really occurs, and the model fixes it before the reader sees it. The payload is §7 of CONTRACTS.md. Demo
Live workbench: — a public instance with the keyless market feed and a capped model budget embedded, so there is nothing to install. The deployment itself is kept private.
The column is pinned to Micron. The agent's six marks are for Bitcoin — a different instrument, so the predicate refuses the merge and nothing lands on the wrong chart. They are offered on a pill instead; one click loads that chart with its marks. The clock in the top-right corner is a live feed off a local OpenD, ticking through the whole clip.
▶ Watch the full 90-second demo with narration — the loop above is one beat of it. The full cut also walks the shell, the symbol box and the timeframe row, the agent drawing on the live column, and the per-turn context line that lets it read your chart without asking for a screenshot.
None of it is a mockup: the footage is 1920x1080 Playwright captures of a real session against a live Futu OpenD, and the chart keeps ticking through every shot.
The earlier v0.2 demo — chart cards in the chat feed
相似插件
thoughtdag
An infinite canvas where LLM conversations grow into an editable thought graph. Read papers into maps…
graph-memory
Knowledge graph memory for DeepSeek Harness and OpenClaw — cross-session recall, PageRank, communities, and vector searc…
dsh-Mimir-Academic-research
Mimir — 一站式科研工作台插件:LaTeX 论文边写边编译、arXiv 文献管理、实验追踪、指标图表、GPU 服务器 SSH 任务编排,管理科研全周期…