`kaijia323/dsh-plugin-jev`：DeepSeek Harness 的 Jev 插件（与本库同环境）
DeepSeek Harness plugin that registers a native jev_decide tool and declares its own policy for when the agent should ask Jev.
Overview
dsh-plugin-jev, documented in Chinese, connects TypeSafe's Jev to DeepSeek Harness (DSH) by registering a native tool called jev_decide. Because Jev does not generate text or speak a chat API, it cannot be DSH's main model; the right integration is a tool. The plugin has no dependencies and registers a raw tool definition through the same path MCP tools use.
Its distinctive feature is a self-declared usage policy, attached in two places so the model knows when to call the tool: in the tool description and as a resident system-prompt section for every session. The policy says to ask Jev by default whenever an answer is a judgment rather than a lookup (meaning, intent, relevance, fit, risk), lists four exemptions such as when a mechanical check or readable file output exists, names seven decision moments such as competing plausible options or an irreversible next step, and sets a budget of one call per state with all its questions. Token limits on both policy texts are enforced by tests. A smoke test mocks the official endpoint, and the author measured 0.5 to 0.6 s per call.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Other | Any typed choice, score or noul question the agent sends through jev_decide | — |
| Yes / no | Is the next step irreversible or expensive enough to check first? (decision moment)(inferred) | — |
How it works
- Install the plugin into DeepSeek Harness with a TypeSafe key
- Policy text tells the agent when to call jev_decide
- Agent sends one state with all its questions
- Plugin posts to /v1/systemone and returns the response unchanged
- Agent branches on typed answers
Strengths
- Zero dependencies
- Explicit, tested usage policy
- Response passed through unchanged
Caveats
- Documentation in Chinese
- Depends on the agent following the policy
- Requires a TypeSafe key
At a glance
- Type
- Plugin
- Status
- beta
- Language
- JavaScript
- Runtime
- DeepSeek Harness plugin
- Has code
- yes
- Has tests
- yes
Use cases and integrations
How curated lists describe it
Connect Jev to DSH, **register the native tool `jev_decide`**.
wh000wh000/awesome-jev-live
Frequently asked questions
Why a tool and not a model?
Jev returns typed answers, not streamed text with tool calls, so it cannot be DSH's main model.
How does the agent know when to use it?
The plugin injects a usage policy into the tool description and a resident system-prompt section.
What does a call cost in time?
The author measured about 0.5 to 0.6 seconds per call.
Related projects
Portable decision layer for agent harnesses: Jev picks from host-supplied capabilities while the host keeps permissions and execution.
ProjectbetaTypeScriptLocal MCP server putting Jev in the coding loop: route next steps, pick prepared tool calls, review patches and verify claims.
MCP serverbetaTypeScriptDecision plugin for agent harnesses where Jev recommends tools, skills or owners and scores output, for DeepSeek Harness and iPolloWork.
Agent skillbetaTypeScriptOmO/senpi agent plugin where Jev or Span-01 rates which skills and next tools fit the task and spots loops or completion.
Agent skillbetaTypeScriptClaude Code function hook that sends the agent's this-or-that questions to Jev and optionally ranks installed skills per prompt.
ProjectbetaTypeScriptHermes skill that calls Jev as a typed tiebreaker, returning Choice, Score or Noul probabilities so the agent can act or escalate.
Agent skillbetaPython