When developing DeepSeek Harness (DSH) applications or plugins, developers typically need to understand request latency distribution, model call costs, tool execution status, and cache hit rates for performance optimization and budget control. Existing monitoring solutions often require collecting conversation content or are difficult to integrate seamlessly into local development environments. As a local-first plugin, dsh-telemetry aims to provide observability into Harness runtime behavior while not collecting sensitive conversation content by default.

What It Is

This is a DeepSeek Harness plugin maintained by duyanta123 under the MIT License. It does not perform business analytics, does not modify request content, and does not upload user conversations to any third-party platform.

Core Capabilities

The plugin records lifecycle metrics for requests, models, tools, and plugins, including latency, token counts, cost estimates, error rates, and cache status. It mainly helps answer the following questions:
* How long does a single request take?
* Where is the time mainly spent: model calls, tool execution, or plugin hooks?
* What are the input and output token counts? How much additional cost do retries introduce?
* Which tool calls are the slowest and have the highest failure rates?
* Which plugins threw exceptions or blocked the workflow?
* Was the cache hit? Did model routing save costs?

Installation and Enabling

It is recommended to install it using DSH’s plugin management command.

dsh plugin --profile <name> add "github:duyanta123/dsh-telemetry#main"

After installation, restart the corresponding profile (for example, dsh --profile <name>), and telemetry collection will start automatically.

Usage and Querying

After installation, you can query the status or start the Web UI from the command line.

# 查看运行状态
node <profile>/node_modules/dsh-local-telemetry/bin/telemetry.mjs --status

# 启动本地只读 Web UI（绑定 127.0.0.1）
node <profile>/node_modules/dsh-local-telemetry/bin/telemetry.mjs --ui --port 47610

Through the Web UI, you can view request overview, P95 latency, time to first token, token/cache hit rate, estimated cost, and error categories. Through profile patch configuration, you can set the data storage method, such as using SQLite.

Compatibility and Privacy

  • Node.js version: Supports JSONL and pure CLI mode (Node.js >= 18); the SQLite backend requires Node.js >= 22.5.
  • Privacy protection: By default, it does not collect prompts, responses, file contents, or keys. It only records names, IDs, counts, statuses, durations, and total token counts.
  • UI safety: The Web UI is local and read-only, and is bound only to 127.0.0.1.

Summary

dsh-telemetry provides a local and privacy-friendly way to quantify Harness runtime behavior. Developers can use it to gain deep insights into system performance and cost structure without worrying about sensitive data leakage.