dsh-plugin-jev
DeepSeek Harness Web GUI plugin giving the agent a jev_ask tool, usage tracking and an optional command guard with explicit caveats.
Overview
dsh-plugin-jev makes Jev an agent-facing tool in the DeepSeek Harness Web GUI. The model calls Jev through a jev_ask tool whenever a decision needs calibrated judgment, and the browser adds a per-chat usage chip and a settings tab for the API key, default model and overall stats. There is no human query console. Detailed docs, written for AI agents integrating with the plugin, cover the tool, a ctx.get('jev') service, HTTP routes, accounting and module invariants.
It also includes an optional command guard, and the README is notably frank about it: the guard judges commands with a probabilistic model and is not reliable in either direction. It can block safe commands, can allow dangerous ones, sees only the command text, working directory and sandbox mode (not what variables or paths resolve to), and cannot be complete. A clean log is not evidence that nothing harmful happened.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Any typed question the agent sends via jev_ask(inferred) | — |
| Choice | Is this command safe, or does it match a risky rule (command guard)?(inferred) | — |
How it works
- Agent calls jev_ask when it needs a calibrated decision
- Plugin queries Jev and records usage per chat
- Optional command guard judges shell commands before execution
- Settings tab shows key, model and stats
Strengths
- Agent decides when to use Jev without user involvement
- Usage tracking per chat
- Unusually honest guard documentation
Caveats
- Command guard is explicitly unreliable
- No licence file per the listing
Getting started
Install as a DSH plugin and set the API key under Settings → Plugins → Jev (TypeSafe).At a glance
- Type
- Plugin
- Status
- beta
- Language
- TypeScript
- Runtime
- DeepSeek Harness Web GUI plugin
- Has code
- yes
Use cases and integrations
How curated lists describe it
Gives the agent in DeepSeek Harness's web GUI a `jev_ask` tool, so the model calls Jev itself when a decision needs a calibrated answer, with a per-chat usage chip. No licence file.
MrJev/awesome-jev
Frequently asked questions
Who calls Jev?
The agent itself, through the jev_ask tool.
Is the command guard safe to rely on?
The README says no: it can miss dangerous commands and block safe ones.
Where is usage shown?
In a per-chat chip and a settings tab with overall stats.
Related projects
Browser tool that finds bounded decisions in LLM prompts and proposes Jev Choice, Score and Noul questions with SDK code.
ProjectbetaJavaScriptStdlib Python primitive for asking N items K typed questions in bounded, fail-open Jev requests, refusing out-of-set answers.
LibraryprototypePythonEarly language design where typed Jev judgments are a runtime operator, with an engine that batches judgments over shared state.
ProjectprototypeTypeScriptDeepSeek Harness plugin registering a jev_decide tool so the agent can ask Noul, Choice or Score questions and gate on probabilities.
PluginbetaTypeScriptDeepSeek Harness plugin adding a jev_evaluate tool for mixed Noul, Choice and Score evaluation of text or JSON.
ProjectbetaTypeScriptDeepSeek Harness plugin exposing Jev noul, choice and score as agent tools, with configurable guardrails and Web UI result cards.
PluginbetaTypeScript