dsh-sandbox-monitor

Real-time sandbox observability for DeepSeek Harness — latency, loss, errors and resource burn in the native right sidebar.

A live dashboard plus a terminal-style sandbox preview, so you can see what the sandbox is doing right now instead of guessing from the chat transcript.

中文说明

Why

DSH runs tools inside a sandbox (PowerShell, file reads/writes, MCP calls, …). When something feels slow or fails silently, the chat only shows the final result. This plugin samples the tool pipeline and answers:

• How long do sandbox calls actually take? (avg / P95 / peak, 60 s sparkline)
• Are calls being blocked, cancelled or timing out? (loss rate)
• What is failing, and with which error code?
• How much are we spending? (tokens, bytes, CPU, memory)
• Which tools dominate the workload?
• What is running right now?

Features

• Sandbox preview — a live console of the most recent calls: time, tool, command/path summary, duration, status. Plus the active sandbox mode, root directory, uptime and in-flight count.
• 14 selectable metrics — add/remove cards at any time; the selection is persisted server-side.
• Live event stream — every call, with human-readable arguments (command, path, pattern), not just ok.
• Tool distribution and error-code aggregation.
• Theme-native — uses DSH design tokens, so light/dark and glass skins work out of the box.
• Zero schema cost — no LLM tool is registered, nothing is added to the system prompt. Sampling is passive.
• Local only — counters live in the host process; config lives in ~/.dsh/sandbox-monitor/config.json. No network calls, no telemetry.

Install

Requires a DSH build with the native right sidebar (@deepseek-ai/dsh-client-ui-sidebar-right), tested on 0.2.0-rc.2.

# recommended
dsh plugin --profile web add github:wuming295/dsh-sandbox-monitor

# or a release tarball
dsh plugin --profile web add https://github.com/wuming295/dsh-sandbox-monitor/releases/download/v0.1.0/dsh-sandbox-monitor-0.1.0.tgz

Replace web with your profile name (e.g. desktop), then refresh the browser.

An npm release is planned; until then install from GitHub or the release tarball above.

From source / without the plugin manager

git clone https://github.com/wuming295/dsh-sandbox-monitor.git
cd dsh-sandbox-monitor

Point your local DSH loader at the folder (the super-injector path), or add it to a profile's node_modules and enable the bundle. There is no build step — lib/ is the shipped JavaScript.

Where to find it

• First install opens itself once, automatically.
• Afterwards: click the pulse icon at the bottom-left of the sidebar, or
• Right sidebar → + → 沙箱监视 (Sandbox Monitor).

The panel is session-independent: it shows the whole host, not just the current conversation.

Metrics

Metric	What it measures
Latency	Tool dispatch duration: avg / P50 / P95 / peak, plus a 60 s sparkline
Loss rate	Calls denied by policy, cancelled, or aborted/timeout
Errors	Failed tool calls + agent-level errors, with error-rate and error codes
In-flight	Sandbox calls executing right now
Call rate	Calls per minute over the sampling window
Tokens	Input/output tokens folded from assistant/message usage
Throughput	Argument + result bytes per second
Exec time	Cumulative dispatch time and average per call
CPU	Host process CPU (sum across cores)
Memory	Host process RSS / heap
Blocked	Policy denials, cancellations and approval rejections
Tool distribution	Calls, errors and average duration per tool
Error codes	Aggregated failure codes
Events	The live call stream

Options

Open the sliders button in the panel header:

• enable/disable any metric (8 of 14 on by default),
• set the polling interval (1 s / 1.5 s / 3 s / 5 s),
• reset the current counters.

How it works

The host half is a passive sampler. It hooks:

• tools/pre-execute — policy decisions (deny/cancel)
• tools/execute — dispatch timing and in-flight tracking
• tools/result — errors, result size, final status
• agent/error — agent-level failures
• session/event — token usage from assistant messages
• process.cpuUsage() / process.memoryUsage() — host resources

Waterfall hooks always call next() and return the original decision/result unchanged.

The client half registers a tab type with sidebarRightTabs, mounts the panel on sidebar.right.pane.tab, and adds a footer button on sidebar.footer.action. It polls the host through a small JSON endpoint.

HTTP API

GET  /sandbox-monitor/api/snapshot
GET  /sandbox-monitor/api/health
POST /sandbox-monitor/api/config   { "enabled"?: string[], "refreshMs"?: number, "windowSec"?: number }
POST /sandbox-monitor/api/reset

Configuration

~/.dsh/sandbox-monitor/config.json:

{
  "enabled": ["latency", "loss", "errors", "inflight", "tokens", "cpu", "tools", "events"],
  "windowSec": 60,
  "refreshMs": 1500,
  "maxEvents": 40
}

Compatibility

• DSH Web with the native right sidebar (ui-sidebar-right).
• Tested on DSH 0.2.0-rc.2 (Windows desktop build).
• Node.js ≥ 20.
• No bundler required; the client bundle is a plain __ModuleLoader__ module.

Development

git clone https://github.com/wuming295/dsh-sandbox-monitor.git
cd dsh-sandbox-monitor
npm run check      # syntax-check both halves
npm pack --dry-run # inspect the package contents

• lib/index.js — host: sampling, aggregation, HTTP routes, config persistence.
• lib/client.js — client: right-sidebar tab, cards, options page.
• docs/ — screenshots.

PRs are welcome. Please keep both halves dependency-free and theme-token based.

License

BSD-3-Clause © 2026 YAKUMI