dsh-trust

Trust layer for DeepSeek Harness: audit, rollback, and policy gates for agent actions

A DeepSeek Harness (dsh) plugin.

Why

当 Agent 拿着用户的文件、凭据与支付权限行动时，"它做了什么、能否审计、能否回滚、能否追责"比"它多聪明"更重要。dsh-trust 目标成为 dsh 生态的信任层总入口。

What works in 0.0.1

This is a working plugin, not a name reservation:

• trust_policy_check 工具：对一段即将执行的操作描述做敏感度分级（高危/中危/低危），内置破坏性命令、凭据访问、网络写入等规则，返回命中规则与建议。模型可在执行前自查。

Install it and ask the agent to call the tool — the loop (tool registration, schema validation, model invocation, structured output) is fully exercised.

Roadmap

• 0.1.x — 会话级审计面板：append-only 事件流的可视化（谁、何时、调用了什么工具、动了哪些文件）
• 0.2.x — 行为回滚：基于 Change Ledger 的文件/会话状态回退
• 0.3.x — 策略引擎：可声明的审批门（敏感命令、网络出口、凭据访问前必须人工确认）
• 0.4.x — 插件供应链审计：依赖锁定校验与 inject 清单核对

Install

dsh plugin --profile web add dsh-trust

DeepSeek Harness is in developer preview and states that compatibility-breaking changes will happen. This build was developed and tested against @deepseek-ai/dsh 0.1.0-rc.x (August 2026); re-verify against your host version before relying on it.

License

MIT