bowenliang123/dsh-context

What it solves

dsh-context is a plugin for the DeepSeek Harness (dsh) that provides deep visibility into an AI agent's context window. It solves the problem of "black box" context management, allowing users to see exactly what is occupying the model's memory, how costs are accumulating, and how the context evolves over the course of a conversation.

How it works

The plugin integrates directly into the DeepSeek Harness UI, adding several monitoring surfaces:

• Context Dashboard: A high-level overview of all sessions, including token usage, estimated costs, and activity heatmaps.
• Context Tab/Panel: A detailed per-session view showing current context composition (system prompts, tool schemas, messages, etc.), timing stats, and a "Context Browser" to inspect the actual content of any specific request.
• Context Trend Chart: A visual representation of how the context window grows or shrinks (via compactions or prunes) turn-by-turn.
• File Activity Log: A tracker that monitors every file the agent reads, writes, or searches.
• Agent Network Map: A visualization of the relationship between the main agent and any sub-agents used in a session.
• Slash Command: A /context command for quick access to composition and browsing without leaving the chat.

Who it’s for

Developers and power users of the DeepSeek Harness who need to optimize their agent's performance, debug why a conversation is degrading, or monitor the financial cost of their AI interactions.

Highlights

• Granular Token Breakdown: Visualizes context occupancy by category (system, tools, user, assistant, etc.) using the same projections as the official token meter.
• Detailed Request Inspection: The Context Browser allows users to expand and view the actual raw or markdown content of every element in a request.
• Context Evolution Tracking: Tracks compactions, prunes, and model switches as "events" to explain why the context window changed.
• Comprehensive Cost Tracking: Estimates costs based on model list prices, including peak/off-peak awareness for DeepSeek models.
• File System Audit: Aggregates all file operations (reads/writes/searches) and links them directly to the tool results in the context browser.