DSH Trading Floor (P2)

A live, read-only trading-floor dashboard for DeepSeek Harness itself.

It turns the Session projections the Web UI already maintains into a desk view: active sessions, steps, token mix, cache hit, context pressure, tool time, per-session book bars, and a human risk gate. The visual language is a trading terminal; the data is DSH's own session telemetry.

No market data. No broker. No orders. No external network calls.

What is real

• Sessions with a prompt today (or running) and their projections
• Per-session steps, turns, LLM time, and tool time
• Cache-read, uncached-input, and output tokens
• Context pressure and context window
• Agent preset and model selection
• A per-session lifetime cost estimate at DeepSeek's idle rates
• Subagent labels and last-prompt activity

The dashboard polls the same same-origin Session RPC the Web UI uses, every 10 seconds. A snapshot renders in roughly 0.4–0.6 seconds on a 361-session corpus.

Scope and honesty

The Session projection exposes lifetime token usage per session, not call timestamps. The dashboard therefore calls its scope the active book:

• A session is in the book if its last prompt is today, or it is currently running.
• Spend is the lifetime cost of those book sessions at idle rates.
• It is not today's bill. For exact daily accounting use dsh-token-audit.

The UI labels this everywhere: BOOK LIFETIME, IDLE EST, EST, BOOK CAP, and read-only. The ¥100 cap is a policy guard, not a daily bill. There are no fake P&L numbers or invented agents.

Screens

• Ticker: active, live, steps, tokens, cache, output share, book spend, book cap
• Six lenses: TAPE (sessions), QUANT (steps), MACRO (presets), RISK (context), FLOW (tool time), PM (book spend)
• Gauges: context, cache hit, output share, tool time
• Orchestration web: the harness at the center, six lenses around it
• Human gate: book cap (default ¥100 lifetime), context, and cache checks, with Refresh / Copy
• Desk log: last-prompt and subagent activity
• FULL DESK: full-viewport portal; Esc or EXIT returns to Settings

Install

Use the Plugin Manager (Plugins → Add plugin) or the plugin_manager install_bundle action with this absolute directory. Once the public repo exists:

dsh plugin --profile web add github:lengrufeng2008/dsh-trading-floor

Open Settings → Trading Floor. FULL DESK is in the top-right.

The live profile install is a local link, so client edits apply after a page refresh. A packed tarball is produced with npm pack and can be attached to a GitHub Release.

Verification

• npm test: 4/4 pure projection-fold tests
• Live Web UI: 1600×1000 and 1920×1080, 0 page errors, 0 console errors
• First data render: ~0.4–0.6 s
• Full-desk portal, Escape isolation, responsive Settings pane
• Read-only controls only: Refresh and Copy
• Fresh temp profile: installed the packed tarball with dsh plugin --profile web add dsh-trading-floor-0.2.0.tgz; --dump-config composed the trading-floor row; the Plugins page listed Trading Floor with the locale description
• Reviewer file tour: REVIEW.md

Screenshots: docs/settings-1600.png, docs/full-desk-1600.png, docs/full-desk-1920.png.

Limits / next

• Exact daily bill needs event timestamps or a Host wire.view; P2 intentionally stays on the already-maintained projection API.
• Approval actions are not wired; the human gate reports live policy, goal phase, and pending questions.
• Market data and paper trading are deliberately out of scope.

Compatibility

DSH 0.2.1-rc.2, Web profile. The dashboard depends on the session/list client RPC and the settings.section slot.

MIT. The Grok Bot six-agent desk is a public use case; no code or assets were copied.