Agentic AI 的权限层,以 DeepSeek Harness 插件形式提供
The authority layer for agentic AI, as a DeepSeek Harness plugin
安装
dsh plugin --profile web add github:tappass/dsh-governanceGitHub 源码安装:首次需按提示配置 allowBuilds 构建授权后重试
安装与环境配置指引、插件开发教程见 DSH 中文社区文档 ↗
安装即在你的机器上以你的权限运行第三方代码——它可读写文件、使用凭据、访问网络,DSH 的工具审批不会为插件代码加沙箱。「检测到 manifest」仅代表发现 dsh.bundle / dsh.plugin 清单,不构成兼容性或安全审查;安装前请审阅源码,不熟悉的插件先在不含密钥的环境试用。
README
目录
The authority layer for agentic AI, as a DeepSeek Harness plugin.
Everything in DeepSeek Harness is a plugin. This is the one that decides what your agents are allowed to do.
Guardrails and safety classifiers ask "is this output harmful?" That is a property of the model. TapPass asks a different question: "is this agent allowed to do this, under our rules, right now?" That is a property of your business, and no model level tool can answer it, because the answer lives in your organisation, not in the weights.
This plugin intercepts every tool call at the harness's tools/pre-execute
seam, sends it to the TapPass policy decision point (POST /v1/govern), and
allows, denies, or escalates it for human approval.
Why it is different
- Business rules, not model safety. Write the rule once, in your language: "refunds over 500 need a human", "no customer PII leaves the EU region", "this agent may read the CRM, never write it". It is enforced on every tool call, on every harness, under every model. A prompt is a suggestion. A policy is a fact.
- Authority is earned. The plugin ships in observe mode: from the first minute it watches and records every call, and blocks nothing. You see what your agents do before you enforce a single deny. Then you turn on enforcement for what matters. Autonomy is a track record, not a checkbox.
- Harness and model agnostic. The same rules that govern an agent here govern it in Claude Code, in Codex, behind LiteLLM. The harness is interchangeable. The authority is not.
- EU hosted, EU AI Act ready.
Install
# create a profile if you do not have one, then add the plugin
dsh plugin --profile default add @tappass/dsh-governance
# point it at your TapPass workspace
export TAPPASS_API_KEY="tp_dev_..." # a TapPass developer key
# verify the layer without booting, then run
dsh --profile default --dump-config
dsh --profile default
Get a developer key from your TapPass dashboard (Settings, Developer keys) or
POST /api/agents/{agent}/developer-keys. The key is bound to one agent and one
org; TapPass records the audit trail against it.
Configure
Every tool call is governed once it is installed. Configuration is optional; the defaults are safe.
| Key | Default | Meaning |
|---|---|---|
baseURL |
https://app.tappass.ai |
TapPass API base. The plugin POSTs to ${baseURL}/v1/govern. |
apiKeyEnv |
TAPPASS_API_KEY |
Env var holding your tp_dev_ key. A reference, not the secret. |
mode |
observe |
observe: send and record every call, block nothing. enforce: honor verdicts. |
onError |
deny |
In enforce, when TapPass is unreachable: deny (fail closed) or allow (fail open). |
timeoutMs |
4000 |
Hard timeout per verdict call. A slow PDP never wedges the agent loop. |
agentId |
harness agent id | Override the agent id sent to TapPass. |
orgId |
from the key | Override the org id; normally stamped from the developer key. |
Set them in your profile's patch, for example to enforce:
# $DSH_HOME/profiles/default/cordis.patch.yml
- tappass-governance:
config:
mode: enforce
How a verdict becomes a decision
TapPass outcome |
dsh PreToolDecision |
Effect |
|---|---|---|
allow |
next() |
the tool runs |
block |
{ kind: 'deny', reason } |
the model gets an error result with the reason |
needs_approval |
{ kind: 'ask', reason } |
routed to the harness approval flow for a human |
In observe mode every call returns next(), but a would-be block or approval
is still recorded server side and logged locally, so you can size your policy
against real traffic before enforcing.
The plugin sends its mode with each call (enforcement.mode), so the audit
trail can show an observe-mode block distinctly from an enforced one.
Honest limitations
- No argument rewriting. DeepSeek Harness makes tool arguments read only at
tools/pre-executeby design (they are already logged and shown to the model), so a TapPassmodifyverdict cannot be applied in place. This plugin fails such a call closed inenforcemode with a clear reason rather than silently running the unmodified request. Redaction obligations are surfaced, not applied. - Approval needs an open turn. A
needs_approvalverdict maps to the harnessaskdecision, which routes to whatever approval answerer your profile mounts. With no approver configured,askfails closed to a denial, which is the safe default. - Developer preview. DeepSeek Harness and its plugin API are pre-release and
may change. This plugin is deliberately thin: it is a bridge to
/v1/govern, so if the harness API shifts the fix is a small shim, not a rewrite.
Develop
npm install --legacy-peer-deps # dsh rc packages have skewed peer ranges
npm run build
node --test
The verdict mapping and the /v1/govern client are covered by
test/plugin.test.mjs and run against a local mock server, no TapPass instance
required.
Links
- TapPass: https://tappass.ai
- The authority layer for agentic AI. Authority is earned. Even by AI.
License
MIT © Cogniqor BV (TapPass)
原始 README: https://github.com/tappass/dsh-governance/blob/main/README.md ↗
同类插件
查看全部 →
deepseek-harness
从仓库或系统描述生成经过校验的自包含交互式架构图、流程图、时序图、数据流图和生命周期图。

dsh-plugin
通过 DSH MCP 客户端挂载 Ouroboros 的纯配置包,在 DSH 中提供 36 个涵盖需求访谈、Seed、执行、评估与演化流程的工具。

dsh-tongflow
基于 TongFlow 的“片场”插件,用于图片、配音、音乐与视频制作:agent 为每个资产生成 TongFlow 工作流文件(.tongflow.json)并通过 TongFlow 插件执行,内嵌工作流画布,按镜头/角色/take 组织项目,附漫剧模板;以 @tongflow 开头的会话进入 Studio 界面。

helloagents
AI 编码 CLI 的工作流层:技能、项目知识、交付检查、更安全的配置写入与可恢复执行

dsh-ai-novel-writer
安装专用 AI 小说创作预设与工作台:提供带修订号的本地项目资产、紧凑侧边工作台,以及需要原生审批的逐文件变更。

rea
用 agent 逆向任何东西:从应用行为到原生二进制