dsh-trajectory-governance

In DeepSeek Harness, long-running agent tasks often generate massive event logs, making it hard to track branching structures, identify invalid loops, or detect goal drift with a traditional flat log view. This leaves developers struggling to diagnose issues effectively, wasting computational resources. The dsh-trajectory-governance plugin addresses this by rebuilding sessions into a structured multi-branch trajectory tree and introducing observation-layer snapshots, enabling deep visualization of agent behavior. It asynchronously runs three anomaly diagnosis strategies (loop deadlock, invalid retry, and goal drift), highlighting problem areas in an independent tab to help users quickly pinpoint bottlenecks.

The plugin's core capabilities include automatically inferring sub-agent and fork lineages, providing snapshots with context hashes, and enabling stop-loss through alerts and one-click interruptions. All data is stored in a plugin-private database, ensuring zero kernel modification and a purely observational stance that never interferes with agent execution. Users can browse the trajectory tree, view anomaly reports in the graphical interface, or use CLI tools for deeper inspection.

Who is this for? The plugin is ideal for engineers building complex agent workflows, scientists studying AI behavior patterns, and operations teams monitoring resource consumption to optimize costs. Whether you're debugging autonomous tasks or auditing long sessions, dsh-trajectory-governance offers clear insights, making agent governance effortless and efficient.