dsh-loop-detector：DSH 智能体死循环与重复输出的硬中断检测
dsh-loop-detector 是 viego-qiin 维护的 dsh 0.2.x 插件，监听 agent/assistant-stream 与工具执行，检测正文重复与同主题工具空转，命中后 cancel 硬中断并 steer 引导重试。文章说明与官方 repeat-tool-reminder 的差异、v2 相对 v1 的适配、link/复制安装与 bundles 注册，以及主要配置项与日志示例。
Read Moredsh-brake：DSH 智能体会话死循环检测与刹车
dsh-brake 是 Mlte0907 维护的 DSH 插件（admin-security），监听工具执行与注入上下文，检测重复单调用、固定序列循环与重复文本块，默认在下一轮模型请求前注入插件 notice；可选 deny 模式对达到阈值的完全相同调用做 guard 拒绝。安装需在 profile 目录 pnpm add 本地包并写入 bundles 与 YAML 配置。适合长时间跑 agent 时控循环与灌水，需注意近重复不识别、序列循环不 deny 等 README 限制。
Read MoreDeepSeek Harness Plugin: dsh-fs-deny-policy
The DSH plugin `dsh-fs-deny-policy` is designed to block access to sensitive directories before a model performs reads, searches, or Shell invocations, using a root-directory deny list to fill the read-permission gap in sandboxing mechanisms. It supports full-type interception and can inform the model about the restriction to avoid ineffective interactions. Configuration requires setting absolute paths in `cordis.patch.yml` and supports denying reads only. A process restart is required after installation, and it is recommended to use the plugin together with a sandbox to enhance security.
Read Moredsh-codex-approval: DeepSeek Harness's Codex-style Approval Plugin
dsh-codex-approval is an MIT-licensed plugin for DeepSeek Harness (dsh) designed to address the issues of overly coarse or loose tool approval via a finer-grained decision chain. At the `approval/request` stage, the plugin introduces a three-layer mechanism: first, an explicit rule layer that supports matching by command or reason and setting allow, ask, or deny actions, with deny having the highest priority; second, an AI adjudication layer that evaluates risk and maps it to specific actions when rules do not match, supporting the configuration of the model, risk tolerance, and timeout policies; and finally, a human fallback to handle cases where the AI fails or requires confirmation. The plugin offers three modes: manual, ai, and ai-auto, with the latter automatically rejecting or approving pending requests.
Read Moredsh-governed-workflow: Add task authorization, lifecycle, and acceptance boundaries to DSH
dsh-governed-workflow is a community plugin designed for DeepSeek Harness agents, aimed at addressing task boundary drift and permission abuse in agent development. Its core philosophy is to shift agent workflows from direct modification to a strict sequence of "confirming permissions, controlled execution, and independent verification." The plugin introduces a task authority mechanism based on public GitHub Issues and implements lifecycle state machine management (Observing, Admission, Running, Blocked/Completed, Pending Review). Key features include a "RUNNING-only modification guard," which means that sensitive operations such as bash, write, and edit are only allowed when the task is in the RUNNING state and authority is valid; after the task enters a terminal state, modification permissions are automatically frozen, and will...
Read Moredsh-docker: DSH's Docker container management plugin
dsh-docker is a container management plugin for DSH, designed to enable agents to securely operate local Docker objects. The plugin invokes the Docker CLI via the official subprocess service, uses an argv array to pass parameters to mitigate shell injection risks, and retains the default approval mechanism in `docker_exec`. It provides six core tools: `docker_ps` (list containers), `docker_images` (list images), `docker_logs` (view logs), `docker_inspect` (view container details), `docker_exec` (execute commands inside the container), and `docker_manage` (lifecycle management, such as start, stop, restart, and delete). The plugin requires Node.js.
Read MoreDSH-Approval-Gate: DeepSeek Harness Automatic Approval Gating Plugin
dsh-approval-gate is an MIT-licensed plugin for DeepSeek Harness, designed to balance the efficiency of automated approval with security. It leverages Flash models to perform risk pre-assessment on sandbox escape operations, categorizing them into SAFE and RISKY types. For reversible, low-risk operations, the plugin uses a confirmation-based learning mechanism to automatically approve them, reducing repetitive manual confirmations. For the five categories of hard-risk operations—deletions, credential handling, remote/production environments, system paths, and irreversible batch operations—it always routes to manual confirmation to ensure critical security. The plugin provides a manual review UI to display the approval timeline and supports hot-reloading of configurations. Version v0.5.0 introduces file change comparison and one-click revert functionality, allowing users to view unified diffs and restore files based on session snapshots. Users need to install it via the command line.
Read More