dsh-performance-optimizer

DSH bundle that exposes host-wide and per-event-loop performance metrics so the user (via a Settings tab) and the agent (via a native tool and a stdio MCP server) can reason about scroll lag, memory drift, and tool loops.

The bundle is intentionally minimal — it samples Node process.memoryUsage() / process.cpuUsage() once per second, keeps a 60-sample rolling buffer, and exposes it through three surfaces:

surface	consumer	mechanism
Settings → Plugins tab	user	client.js mounts <perf-settings-tab> into the per-plugin detail page; live table + Canvas2D chart with CPU / RSS / Heap / Wall tabs.
Floating HUD	user (opt-in)	client.js mounts <perf-hud> into the document body when hudVisible: true.
Native tool	agent (in-session)	ctx.tools.register('performance_report') returns a fresh JSON report on demand.

Files

file	role
index.js	Host half. Owns the sampler, the performanceOptimizer Cordis service, and the performance_report tool.
client.js	Web GUI half. Two custom elements (<perf-settings-tab>, <perf-hud>); routes by ctx.slot. Calls back into the Host via the runtime bridge.
cordis.patch.yml	Loader patch: one insert — the bundle itself.
locale/en.json	English copy.
locale/zh.json	简体/中文 copy.
icon.svg	Bundle icon.

Install

# from a clone
npx @deepseek-ai/dsh plugin --profile web add /absolute/path/to/dsh-performance-optimizer

# from a tarball
npx @deepseek-ai/dsh plugin --profile web add ./dsh-performance-optimizer-0.1.0.tgz

# from npm (once published)
npx @deepseek-ai/dsh plugin --profile web add @local/dsh-performance-optimizer

Restart the Web profile and open http://127.0.0.1:3080/settings/plugins.

npx @deepseek-ai/dsh web

To remove:

npx @deepseek-ai/dsh plugin --profile web remove @local/dsh-performance-optimizer

What it measures

sample field	source	what it answers
t / tHr	Date.now() + process.hrtime.bigint()	When was this sample taken (clock / monotonic, never trust the wall for ordering).
rss	process.memoryUsage().rss	Has the Host process grown? Did release/manual GC help?
heapUsed	process.memoryUsage().heapUsed	V8 heap pressure (RSS includes native).
heapTotal	process.memoryUsage().heapTotal	Did V8 grow the heap (hint of churn).
external	process.memoryUsage().external	Native buffers / modules.
cpuMs	process.cpuUsage()	Cumulative user/system CPU time (ms).
cpuDeltaMs	delta against previous sample	CPU burn between two ticks; correlates with sampler tick & loop work.
wallDeltaMs	now - lastCpu.takenAt	Sampler timer drift — catches stalled timers caused by event-loop saturation.
uptimeMs	process.uptime() * 1000	Host uptime.

History is the last historySize samples (default 60). Sampling interval is sampleIntervalMs (default 1000 ms). When MCP is enabled, the most recent report is also written to a JSON file under the bundle's plugin directory every tick so the stdio MCP server can read it without holding a reference into the Host process.

How it talks to DSH

The Host half uses only Cordis accessors declared in inject:

• tools — registers performance_report. Every MCP tool call routed through DSH eventually surfaces through tools.register, so this is the canonical agent interface regardless of whether @deepseek-ai/dsh-mcp-client is also mounted.
• pluginContext — composition helper; lets the patch layer attach additional reads.

The Client half is mounted once per dsh.client.inject target. The package.json lists three:

• @deepseek-ai/dsh-client-ui-settings-plugins — registers the per-plugin detail tab. This is the canonical user surface.
• @deepseek-ai/dsh-client-shell — registers the HUD.
• @deepseek-ai/dsh-client-ui-conversation — reserved for a future chip; currently no-op.

It looks up the runtime bridge via window.__DSH_HOST__ / window.__DSH__ / window.deepseek.host (whichever is present) and calls plugin:performance-optimizer.snapshot | .history | .config | .sample | .reset. If the bridge is missing, the tab renders an empty state and the HUD shows — — neither blocks the rest of the UI.

Tool surface (native)

{
  "name": "performance_report",
  "arguments": {},
  "result": {
    "kind": "json",
    "value": {
      "intervalMs": 1000,
      "sampleCount": 42,
      "latest": { "t": "2026-10-20T11:39:42.000Z", "rss": 314572800, "heapUsed": 134217728, "cpuMs": { "user": 1234.5, "system": 67.8 }, "...": "..." },
      "history": [ /* last 10 samples */ ],
      "generatedAt": "2026-10-20T11:39:42.000Z"
    }
  }
}

Agent surface

The bundle exposes one native tool:

performance_report

It returns the same JSON-shaped report shown above. Every MCP tool call routed through DSH ultimately surfaces through ctx.tools.register, so this is the canonical agent interface — there is no separate MCP server in this bundle.

If you want external MCP clients (Claude Code, Codex, an inspector) to also pull snapshots, write a small sibling DSH plugin that mounts an MCP server and reads ctx.performanceOptimizer.buildReport() for each tool call. Keeping MCP server concerns out of this bundle avoids coupling to a specific @modelcontextprotocol/sdk version.

Config

sampleIntervalMs: 1000  # sampling cadence in milliseconds
historySize: 60         # rolling buffer length
hudVisible: false       # boot-time HUD visibility
settingsTab: true       # mount the Settings → Plugins detail tab

All keys are optional; defaults are baked into the bundle.

Failure / uninstall

When the bundle is disabled in the Plugins page, ctx.effect cleanups tear down:

• the sampler timer on the Host,
• the performance_report tool registration,
• the HUD element,
• the <perf-settings-tab> element.

No side effects on the Host process remain.

License

MIT — see LICENSE.