DeepSeek Harness (DSH) provides powerful agent orchestration capabilities, but it lacks systematic trace data support when debugging complex workflows, sub-agent invocations, and tool executions. The Litefuse plugin aims to solve this problem by converting DSH runtime data into compliant, spec-driven observability traces.

What Is This

This plugin is an official DeepSeek Harness plugin for integrating Litefuse’s observability and evaluation capabilities. It converts each user interaction into an independent trace and strictly follows the Litefuse agent-trace spec v1.2 specification.

The plugin is maintained by litefuse and is released under the MIT License.

Core Features

The plugin has the following core capabilities:

  1. Turn-Level Trace Mapping: Each user turn generates a trace with a single agent root node. Under the root node are generation entries for each model call, tool entries for each tool execution, and nested containers for delegation to sub-agents.
  2. Zero Dependencies: It has no runtime dependencies and does not rely on the Langfuse SDK. Data is sent directly to Litefuse endpoints via OTLP.
  3. Sub-Agent Support: It supports sub-agent and delegation scenarios and can accurately record Delegation Overhead.
  4. Code Mode Support: It supports code mode and directly traces tool calls inside run_code programs, avoiding confusion with ordinary tool calls.
  5. OTLP Integration: It supports sending trace data to Litefuse via the OTLP protocol.

Installation and Enabling

Installing the plugin requires the npx command.

npx @deepseek-ai/dsh plugin --profile web add -w dsh-litefuse-plugin

After installation, configure credentials.

  1. Set the Litefuse public key and secret key in ~/.dsh/.env:
    LITEFUSE_PUBLIC_KEY=pk-lf-…
    LITEFUSE_SECRET_KEY=sk-lf-…

These keys must be created in the Litefuse project's settings.
  1. Start DeepSeek Harness:
    npx @deepseek-ai/dsh web

Typical Usage

After completing the configuration above, when you send a message in DSH:

  • Open a browser and visit https://litefuse.cloud, then enter the corresponding project.
  • View the data on the Tracing page.
  • The trace appears when the first span for that turn completes, not when the turn ends.

When reading configuration, the plugin prefers environment variables, but the configuration can also be overridden via ~/.dsh/profiles/<name>/cordis.patch.yml.

Notes

  • Credential Handling: Credentials in the configuration are references, not actual values. The plugin reads values from the harness credential store (~/.dsh/.credentials.yaml) or process environment variables. No configuration item writes keys to files, and only the first 10 characters of the public key are logged.
  • Trace Visibility Timing: The trace is visible only after the first span completes.
  • Permissions: The plugin runs with the permissions of the current DSH process. It is recommended to review the source code and license before installation.
  • Configuration Location: Configuration defaults to reading from environment variables. To override it, edit ~/.dsh/profiles/<name>/cordis.patch.yml.

Ecosystem Context

DeepSeek Harness follows the philosophy of “everything is a plugin.” Community directories (such as SkillHub) provide plugin indexes but are not officially affiliated with DeepSeek or High-Flyer AI.

By directly subscribing to DSH’s session/event event stream, the plugin enables precise recording of the actual execution process (not a post-hoc reconstruction), including real latency, token usage, context compression events, and so on.