dsh-trajectory-anchor：DeepSeek Harness 插件
`dsh-trajectory-anchor` 是 DeepSeek Harness 的工作流插件，旨在解决智能体任务偏离与行为退化问题。该插件提供观测审计、首轮锚定、双向行为提醒及状态检查四大核心功能，通过权重评分和干预机制防止偏离。它无需配置即可启用，支持自定义工具列表和审计日志，专注于安全监控而非直接执行，适用于 DSH 版本 0.2.0 至 0.3.0。
Read MoreDeepSeek Harness Awareness Metrics Audit Plugin
The DeepSeek Harness (DSH) plugin `slatinwine/dsh-sentience-audit` audits an agent's functional organization based on trajectory structure. Grounded in the indicators of sentience theory by Butlin et al. (2023), it rates sessions on an L1-L5 scale, with fully reproducible results that do not rely on large models. It outputs specific ratings and evidence through the `sentience_audit` tool, and is suitable for assessing whether an agent exhibits structural features such as recursive loops or a global workspace, but it is limited to live sessions and cannot evaluate subjective experience.
Read Moredsh-harness-audit: Agent Harness Failure Mode Audit
`dsh-harness-audit` is a DeepSeek Harness plugin maintained by Leeaoyin, designed to address failures that commonly occur in agent scheduling loops involving tool invocation, state management, and error recovery. Through machine-enforced evidence verification, the plugin performs automated audits of Harness and rejects any findings that cannot be grounded in code. Its core capabilities include 15 check mechanisms, code-driven summary reports, background job execution, and a strict evidence verification mechanism. The audit process is divided into three stages—reconnaissance, dispatch, and aggregation—and supports running either all checks or specific ones via a command-line interface. Reports are saved by default in the `.harness-audit/` directory, and a “rejection rate” metric is provided to monitor the reliability of subagents.
Read More