deepseek-AutoReview

The most annoying part of agent workflows is rarely a wrong answer—it is the constant flood of permission prompts. read-only is too restrictive, danger-full-access is too reckless, and workspace-write, while silent inside the sandbox, still throws every escalation at a human. A harmless pip install and a catastrophic rm -rf / look identical in the approval queue. deepseek-AutoReview introduces a fourth permission mode, "review on my behalf", into DeepSeek Harness: safe operations pass through without a single interruption, obviously malicious ones are intercepted outright, and everything in between is judged by an independent reviewer model that falls back to a human prompt whenever it is unsure. The core mechanism is a three-tier funnel: a zero-token whitelist fast-approves structurally verifiable commands (single-verb invocations of npm, git, pip and friends, with no shell metacharacters anywhere); a zero-token blacklist built from 26 deterministic rules blocks catastrophic shapes such as rm -rf /, curl | sh, and writes to /etc paths; and only the gray zone is escalated to the reviewer model, which inspects the real tool arguments recovered from the session log, the current turn user request, and the asker self-reported reason to emit an allow, refer, or deny verdict. The reviewer model is switchable at any time from Settings (follow the session or pin a lightweight model), auto-approvals are rate-limited to 5 per minute and 30 per hour by default, every verdict is logged with an evidence fingerprint, and the whole design is fail-closed by construction. This plugin is aimed at developers who run Daily agent workloads on DeepSeek Harness and want to minimise approval interruptions without surrendering the safety net. If the mechanical repetition of clicking Allow is wearing you down, but handing over full access feels like too big a leap, this preset is the middle layer that keeps both speed and trust in the same conversation.