Most evaluations that ask “will the model answer” conflate two layers: the host (harness) and the model itself. This plugin isolates the host layer so you can tell whether a refusal comes from a harness rewrite/approval gate or from the model. That separation is the practical insight for accurate red-team or policy tuning work on locally managed DeepSeek Harness installs.
What Sets It Apart
- Host-vs-model separation: applies reversible local package patches and an encrypted prompt-inject so the host’s runtime rewrites, approval gates, and sandbox policies can be evaluated independently of model behavior — so you know which layer to change.
- Integrated UI for evaluation workflows: a right-hand dock exposes Clean (patch management, prompt editor, rule sets) and Drill (asset/skill/environment inspection after authorization), enabling controlled, repeatable tests without manual file edits.
- Safe-by-scope design: targets only the official @deepseek packages and host config on the user’s machine; does not ship network scanners, public exploit payloads, or third‑party attack tooling — so it’s suitable for authorized red-team exercises and local research.
- Practical constraints surfaced: explicitly tied to DSH 0.2 hosts (official web and official desktop EXE), Node.js >=18, and preserves the ability to restore originals, which matters when testing production-adjacent endpoints.
Who it's for and tradeoffs
Great fit if you need to audit or tune prompt/approval policy behavior on a DeepSeek Harness host you control, run authorized red-team exercises, or benchmark how much of a refusal is due to host-layer protections. Look elsewhere if you need a generic external attack tool, a cloud-scale scanner, or cross-host remote exploit capabilities — this project is scoped to local, authorized testing and non-commercial research. Expect to manage DSH version compatibility (targets DSH 0.2) and follow the project's attribution and non-profit usage terms when reusing prompts or code.
