pengxuding/dsh-plugin-judge
Plugin value auditor: pre-install review (source scan + LLM judge) and post-install audit of installed bundles, with model-switch re-audit reminders.
dsh-plugin-judge is a plugin value auditor: it judges whether a DeepSeek Harness plugin is worth using relative to the current model. Pre-install review (via /plugin-audit, the judge_plugin tool, or the settings page) fetches the plugin source and runs a static scan + LLM judge for a 'worth installing?' verdict. Post-install audit enumerates installed bundles from the profile, classifies each as capability / constraint / cosmetic / hybrid with a constraint-risk score, and shows a report panel in settings. Model-switch reminders watch the default model and flag plugins whose verdict depended on the old model, popping a reminder to re-audit. The core thesis: constraint plugins (system-prompt rules, personas, pre-execute interception, restrict/guard, wholesale prompt replacement) can quietly cap a capable model, so a plugin's value is a binary relation of plugin × current model, not a property of the plugin alone.
Install
dsh plugin --profile web add dsh-plugin-judgenpm dsh-plugin-judge 0.1.1 verified 2026-09-04 (repository field → github.com/pengxuding/dsh-plugin-judge; README EN primary with 中文文档). Install: dsh plugin --profile web add dsh-plugin-judge. Use via /plugin-audit <github:owner/repo or npm:pkg>, the judge_plugin tool, or the settings-page input.
Compatibility
DeepSeek Harness; judges plugins as capability vs constraint vs cosmetic vs hybrid; LLM judge layer uses the current model's identity; deterministic rule-heuristic layer is free.
Details
- Repo: pengxuding/dsh-plugin-judge
- Category: Tools & Capabilities
- Stars: 0
- Version: npm dsh-plugin-judge 0.1.1
- Last push: 2026-08-15
- First seen: 2026-08-15
Recent updates
0.1.1 is the current npm latest (verified 2026-09-04).
FAQ
- What is a 'constraint plugin'?
- A plugin that imposes rules/guardrails on the model — injected system-prompt sections, personas, agent/pre-step or tools/pre-execute interception, restrict/guard, or wholesale prompt replacement — versus capability plugins that add new tools or skills.
- How does the verdict work?
- Two layers: free deterministic rule heuristics scan the plugin source and injected content into a 0-100 constraint-risk score and a capability/constraint/cosmetic/hybrid class; an on-demand LLM judge then weighs scan results plus the current model's identity.
- When does it re-audit?
- It watches the default model — when the model changes, plugins whose verdict depended on the old model are flagged with an overlay reminder to re-audit.
Alternatives
Xrainsmile/DSH-Plugin-Doctor · iiwish/dsh-testkit