dsh-research-nudge

When using AI agents for complex research or coding tasks, a common pitfall emerges: agents can get stuck in a "local trial-and-error loop." They might spend extensive time reading local code, repeatedly editing the same file, executing commands and hitting the same failures, without ever consulting official documentation, searching for specific error messages, or looking up existing solutions. This wastes compute and time, and often leaves the core problem unresolved. The dsh-research-nudge plugin is designed to address this, acting as a quiet observer that calculates the accumulating "research debt."

The plugin's core capability is a deterministic "research debt" assessment mechanism. It monitors tool call types and results—such as routine tool usage, file mutations, execution failures, and repeated equivalent failures—to increment a virtual debt score. When this score exceeds a threshold (default 20), or when an agent has gone too long without engaging in external research, the plugin automatically inserts a concise, advisory reminder message before the next model interaction. This prompt explicitly tells the agent, "you might want to look something up," but it is entirely non-intrusive. It does not block any tool usage, make external calls, or force a browsing action, allowing self-contained work to proceed normally.

This plugin is particularly useful for developers and researchers building automated workflows with DeepSeek Harness, especially those involved in debugging, data analysis, or long-duration research tasks. If you find your AI agent is inefficient, endlessly spinning on the same issue, dsh-research-nudge helps it "break the cycle." For teams looking to improve task completion quality and reduce wasted computation, it offers a lightweight, non-invasive enhancement that helps agents balance head-down execution with strategic, evidence-based problem-solving.