dsh-auto-review

Long-running AI agents inevitably accumulate trust and write permissions over hours or days, which naturally heightens risk. Without a continuous security layer, agents may easily forget to remain cautious during extended sessions, potentially leading to accidental errors or security vulnerabilities. dsh-auto-review addresses this by providing a native "Auto Mode" for DeepSeek Harness, turning security into a persistent, always-on safeguard.

At its core, the plugin employs a two-layer review mechanism: pre-execution rule interception and delivery-time deep review by an independent read-only subagent. It intercepts sensitive operations—such as writing secrets or running dangerous commands—before they execute, defaulting to a "deny" policy for fail-closed security. This design ensures that security checks are orthogonal to permissions, meaning even agents with full access remain subject to scrutiny, and the system survives restarts to maintain protection throughout the agent's lifecycle.

It is ideal for developers building long-term, stable, and secure workflows, especially in scenarios requiring high autonomy but strict safety boundaries. By offering a frictionless experience with clear status indicators, the plugin ensures that security is prioritized without disrupting development efficiency, truly embodying the philosophy of "Security First, Trust Second."