dsh-autopilot: DeepSeek Harness Acceptance-Driven Automated Completion Engine
dsh-autopilot is an acceptance-driven automatic completion engine for DeepSeek Harness, designed to address the issue that an agent’s claim of completion does not necessarily mean the task is truly complete. The plugin uses independent evaluators to perform deterministic verification of agent outputs and issues a completion certificate only when all required checks pass. It supports multiple evaluation modes, features repair loops, retries, regression detection, and blocking mechanisms, and provides a CLI tool for standalone evaluation and demonstrations. Installation requires Node.js 20.19+, and it is open-sourced under the MIT license.
Read More