dsh-code-coverage

AI hands you a mountain of code, but how much of it actually sits under a safety net of tests? Traditional coverage tools hand you a single number and stop there. dsh-code-coverage goes one step further: it does not reinvent coverage collection, but layers attribution on top of c8 so you can see exactly what AI-authored code and human-authored code each look like in terms of test protection.

It parses DeepSeek Harness session logs to precisely identify which files were created or modified by AI tools (write, edit, str_replace_editor, including subagent sub-sessions), then cross-references the result with V8 coverage data gathered by c8 --all. The output includes an AI-vs-human line count and coverage comparison, a risk list of untested AI files ranked by uncovered lines, and a trust score from 0 to 100 with a letter grade. Installed as a DSH plugin, the Agent can call code_coverage_check, code_coverage_fix, and code_coverage_verify directly in conversation, forming a discover-generate-plan-re-verify loop without any manual CLI runs.

If you use DeepSeek Harness in your daily workflow and want a concrete, low-friction way to quantify how well AI-generated code is actually tested, or to back up team discussions about AI code testing habits with hard numbers, this plugin is a focused starting point. It does not judge test quality; it simply puts the blind spots on your desk.