dsh-guardian-mode

In AI agent automation, self-modification or unauthorized actions pose significant security risks, especially in handling sensitive tasks or high-risk environments. dsh-guardian-mode, as the fifth mode of DeepSeek Harness, addresses this by offering a preset guardian framework. It combines PTC code presentation, independent review, and a human-approved Cordis remediation loop to maintain standard agent capabilities while preventing dangerous behaviors. Core features include two isolated reviewer roles (summarizer and auditor) with configurable backends like Codex, Claude Code, or DSH LLM runtime, and a cadence-based audit mechanism that automatically triggers reviews. Warning and critical approval processes allow user intervention for fixes, and remote API plus TUI interfaces facilitate monitoring and management. This plugin is ideal for administrators, developers, and security experts needing enhanced agent security, particularly in admin-security scenarios to protect data integrity and operational compliance.