dsh-cn-ai-labeling: An AI Content Labeling Self-Check Plugin for DeepSeek Harness

What Does It Do?

It is a DSH plugin for DeepSeek Harness that turns China's AI-generated content labeling requirements into a checklist you can verify item by item. It removes the guesswork of judging compliance from memory: every rule comes from official source texts, so missing labels in code and exported files can be caught before design sign-off and release. Maintained by feely0208 under the MIT license, it is the AI content labeling piece of the DeepWhale security guard series.

Core Features

• Five static labeling rules: flags generated output returned to users without an explicit AI label, projects with no "AI-generated" wording at all, missing human-review notices, exported files lacking implicit labeling metadata, and copy that says "auto-generated" without mentioning AI, each graded high, medium, or low.
• when guards against false green: on projects that do not use AI, rules are recorded as "not applicable" rather than "passed", and the report states how many rules did not run.
• Bundled evidence script: scans a codebase for labeling evidence and gaps, and parses the AIGC JSON inside real output files such as exported documents or generated images to validate elements, Label values, and character sets.
• Exit codes for gating: 0 means nothing found, 1 means findings exist, so it can be wired into CI and pre-release checks.
• Anti-false-positive boundaries: a dedicated section lists items that are not mandatory, such as digital watermarks being permitted rather than required and font-height rules not applying to text or audio.
• Read-only and offline: it scans only local code and specified files, never modifies the scanned project, and never sends content under review to third parties.

How to Use This Plugin?

After enabling it in DSH, the skill becomes available whenever the agent needs it, so you can ask questions such as whether your product is an obligated entity, which features require explicit labels, or what format AIGC metadata should take. When you need evidence, run the bundled script against a codebase or a single exported file, then work through the report to confirm each gap. It performs static checks only, so findings are prompts for confirmation rather than verdicts, and invisible watermarks in images or audio cannot be seen; whether something violates the rules remains a human call. It complements the sibling plugin that scans outward-facing copy, one checking product implementation and the other checking release text.

How to Troubleshoot This Plugin?

When a rule shows as "not applicable", the project simply does not use AI capabilities, which is the expected result of the when guard and does not mean the check passed. If scan results differ from expectations, first confirm the scanned directory points at the real codebase and that validated files are genuine outputs, then verify rule references against the official texts. The script only gathers evidence, reporting what was found and what is missing, so the final compliance judgment stays with a human.