dsh-governance
其他 活跃维护

dsh-governance

tappass/dsh-governance

AI代理权限治理层插件,所有工具调用均会经过业务规则校验,通过TapPass /v1/govern接口统一管控调用行为,无需改动原有业务代码即可落地合规管控,部署轻量便捷。

1
Stars 标星
0
Forks 分支
1
Watchers 关注
0
Open Issues
TypeScript
主要语言
MIT
开源协议
18 KB
仓库大小
1 个月前
最后推送
一键安装扩展 / 插件指令
dsh plugin --profile web add github:tappass/dsh-governance
git clone https://github.com/tappass/dsh-governance.git
git clone git@github.com:tappass/dsh-governance.git
README.md main

TapPass governance for DeepSeek Harness

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-execute by design (they are already logged and shown to the
    model), so a TapPass modify verdict cannot be applied in place. This plugin
    fails such a call closed in enforce mode with a clear reason rather than
    silently running the unmodified request. Redaction obligations are surfaced,
    not applied.
  • Approval needs an open turn. A needs_approval verdict maps to the
    harness ask decision, which routes to whatever approval answerer your
    profile mounts. With no approver configured, ask fails 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)