bowenliang123/dsh-context

What it solves

dsh-context is a plugin for the DeepSeek Harness that allows developers to visualize and manage the context window of AI agents. It solves the problem of "black box" context management, where it is difficult to see exactly what tokens are being consumed by system prompts, tool schemas, and conversation history, and when compaction or pruning occurs.

How it works

The plugin integrates directly into the DeepSeek Harness web UI, adding a dedicated Context tab and a /context slash command. It tracks the session's context lifecycle, providing real-time statistics on token usage, composition, and events. It allows users to browse the actual content of any request's assembled context, including the exact system prompts and tool schemas being sent to the model.

Who it’s for

Developers and AI agent creators using the DeepSeek Harness who need to maintain a precise understanding of their agent's memory and token budget to prevent performance degradation.

Highlights

• Context Dashboard: A UI tab providing stats on turns, steps, and context events like injections, compactions, and pruning.
• Composition Analysis: A six-color stacked bar chart showing the breakdown of tokens used by system prompts, tool schemas, messages, injected context, assistant replies, and tool results.
• Interactive History: A timeline of context growth and compaction events (marked by ✂), allowing users to scrub through the session to see how the context evolved.
• Context Browser: A detailed view that lets users expand and inspect the actual text and JSON schemas of every element in a request's context window.
• In-session Insights: A /context modal that provides a quick snapshot of current occupancy and recent trends without leaving the chat.