Introduction

DeepSeek Harness (DSH) is a debugging tool based on a plugin architecture. When debugging agents or troubleshooting large-model invocation issues, developers often need to inspect the raw data at the llm/stream layer inside Harness, including specific chunks of streaming responses, reasoning processes, tool call parameters, and session compression logic. External packet capture tools (such as mitmproxy) usually cannot capture these internal states and require additional certificate configuration. The dsh-sseye plugin is directly embedded in Harness and provides deep observability across the full call chain without proxies or certificates.

Plugin Overview

jhuanxx44/dsh-sseye is a DeepSeek Harness plugin, positioned as an “LLM debugging console inside DeepSeek Harness”. It intercepts the native llm/stream waterline of Harness, captures the complete content of each LLM call (including messages, system prompts, tool definitions, streaming response chunks, usage, and wire protocol), and displays it in the Harness interface in a style similar to browser DevTools. The plugin is maintained by jhuanxx44 and uses the MIT license.

Core Features

Full Capture and Protocol Reconstruction

The plugin captures the complete GenerateOptions for each call (system prompt, messages, tools, sampling parameters) as well as every chunk of the streaming response. During capture, this data retains its original identity, order, backpressure, and cancellation state. It parses the actual wire protocol (such as openai-completions, anthropic-messages, etc.), and supports reconstructing tool results into separate role:"tool" messages and reasoning content into reasoning_content.

DevTools-Style Viewer

The viewer is embedded in the right panel of Harness as a collapsible accordion. The interface groups entries by “turn” and displays the number of calls, token counts, and total duration. Each step row shows a status dot, sequence number, content preview, and TTFT/duration/usage indicators. Clicking a row expands inline details, including provider/model information, cache hit rate bar, system prompt, messages, tool names, and response chunks (text/reasoning/tool call parameters).

Comparison and Export

Supports context diff comparison between adjacent calls, highlighting newly added messages. Provides a real-time streaming view that dynamically merges data while calls are running. Supports filtering by session or “All”, and provides JSON export capability (single call or packaged export).

Capture Policy and Privacy

Supports runtime adjustment of the capture policy. Capture can be enabled or disabled by source (agent/compression/session title, etc.) and by field (system/messages/tools, etc.), with uncaptured fields marked accordingly. Provides a regex redaction list to hide sensitive information such as API keys before data enters the buffer. By default, data exists only in an in-memory ring buffer, is destroyed when the process exits, and does not involve persistent storage.

Installation and Enablement

  1. Run the installation command in the terminal:
dsh plugin --profile web add dsh-sseye
  1. After installation, restart DeepSeek Harness. An SSEye button will appear in the session header.
  2. To uninstall:
dsh plugin --profile web remove dsh-sseye

Typical Usage

Language Switching

The default interface is in English. To switch to Chinese, edit the configuration file ~/.dsh/profiles/<profile>/cordis.patch.yml and add the configuration under the sseye plugin entry:

- id: sseye
  config:
    locale: zh-CN

Restart Harness after making the change.

Filtering and Export

Use the buttons at the top of the viewer to filter by session or “All”. Click the export button to save the current capture record.

Redacting Sensitive Data

Add regex rules to the redaction list (for example, sk-.*). The plugin will replace matching values with *** before the data enters the memory buffer, preventing leakage of sensitive information.

Notes

  • Zero Persistence: By default, captured data is stored only in an in-memory ring buffer and is lost when the process closes. Exports must be triggered manually while the process is running.
  • Privilege Isolation: The plugin runs through the Harness process. API keys are stored at the adapter layer, and the llm/stream layer cannot directly access the keys.
  • No Additional Configuration: It can be used directly without proxy or certificate configuration.

Summary

dsh-sseye addresses the pain point of observing LLM calls inside Harness, providing deep debugging capability without requiring additional network packet capture tools. Through memory buffering and real-time streaming display, developers can clearly locate model calls, protocol details, and agent logic.