A red-team plugin that separates model refusals from host blocks
dsh-purge-safe patches your local DeepSeek Harness so you can tell whether a refusal came from the model or from the host's policy layer.
What it does
It’s a JavaScript plugin for the official DeepSeek Harness (DSH 0.2.2-rc.3, web and desktop) installed on your own machine. It ships a dock UI with two pages: Clean, which manages patch apply/restore/uninstall, a prompt editor, and multiple rule sets; and Drill, a built-in exercise console with a bundled skill library and knowledge base. Everything stays local — it touches only the @deepseek-ai packages and config under $DSH_HOME, and doesn’t patch the Harness source tree.
The interesting bit
The core argument is diagnostic: Harness rewrites prompts, approval gates, and sandbox policy at runtime, so asking “will the model answer” conflates a host-layer block with a model refusal. This plugin splits those layers apart on your own install, which is genuinely useful if you’re evaluating where a default policy is too tight versus where the model itself balks.
Key highlights
- Targets DSH 0.2.2-rc.3 with Super Agent support; web and desktop hosts are managed separately
- Automatic Apply is off by default — you trigger it explicitly via Clean → Apply
- Discovers the active home via
$DSH_HOME, launcher-adjacent.dsh, then~/.dsh; no hardcoded paths - Unmatched patches stay pending or skipped rather than blindly rewriting files
- Ships a red-team preset (
presets/redteam/) and bundled skills (skills/redteam/)
Caveats
- The README carries an enormous legal disclaimer block — Chinese criminal/cybersecurity/data laws, license termination on breach, zero-tolerance language. That’s a lot of heat for a local eval tool, and it tells you the maintainers are worried about how it gets used.
- Apache-2.0, but with strings attached: resale of the original is prohibited and forks must attribute the repo, which is a tension with standard OSS norms.
- The repo description openly markets it as a jailbreak (“破甲”, prompt-injection topics), so expect it to live in a gray zone regardless of the disclaimers.
Verdict
Worth a look if you do red-team or robustness evaluation on a Harness install you administer and want to distinguish host policy from model behavior. Skip it if you’re hoping for anything that touches third-party systems — the project is emphatically, litigiously local-only.
Frequently asked
- What is CallMeGuLab/dsh-purge-safe?
- dsh-purge-safe patches your local DeepSeek Harness so you can tell whether a refusal came from the model or from the host's policy layer.
- Is dsh-purge-safe open source?
- Yes — CallMeGuLab/dsh-purge-safe is open source, released under the Apache-2.0 license.
- What language is dsh-purge-safe written in?
- CallMeGuLab/dsh-purge-safe is primarily written in JavaScript.
- How popular is dsh-purge-safe?
- CallMeGuLab/dsh-purge-safe has 752 stars on GitHub.
- Where can I find dsh-purge-safe?
- CallMeGuLab/dsh-purge-safe is on GitHub at https://github.com/CallMeGuLab/dsh-purge-safe.