dsh-security-guard

In the dsh plugin ecosystem, third-party code may carry eval calls, postinstall lifecycle scripts, or prompt-injection phrases that can steal credentials or run destructive commands the moment a package is installed. dsh-security-guard adds a lightweight security guard around that lifecycle: it statically scans plugins before and after installation and intercepts dangerous tool calls and file operations at runtime, stopping risk before execution.

Detection is purely static, using only ts.createSourceFile so scanned source is never imported, evaluated, or executed. Rules are plain, auditable JSON files covering four threat classes: malicious code, context injection, token waste, and sensitive-path access. The runtime watcher listens on pre-step, tools/, fs/, and session events to deny destructive shell pipelines, block writes outside workspace roots, and flag suspicious token consumption. An install hook automatically scans every freshly added plugin and records a report, while the web panel, /scan command, and user-managed allowlist close the detect-decide-remit loop.

It suits ops teams managing the dsh plugin ecosystem, platform engineers who need a supply-chain security policy across their organization, and any DeepSeek Harness user who wants to reduce third-party plugin risk without opaque signatures or un-auditable mechanisms.