jev-dsh-decision: A DSH plugin for Structured Decisions in DeepSeek Harness

What Does It Do?

Jev DSH Decision Engine is a DSH plugin for DeepSeek Harness that brings TypeSafe AI's Jev structured decision capability into an agent harness. It solves the problem of choosing which tool, Skill, or agent should handle a task, and of judging whether an output is good enough, returning choices, probabilities, and confidence as structured results. Jev only assists with decisions; the original agent keeps planning and executing, and the plugin neither takes over the main agent nor modifies engine source code. It is maintained by Devin-AXIS, written in JavaScript, and was created on 2026-09-20.

Core Features

• Tool and Skill recommendation: suggests a choice among the tools, Skills, and agents that are actually available and permitted right now.
• Structured evaluation: supports choice, score, and noul (probability of yes) question types, and can batch several independent questions in one call.
• Explainable output: returns choices, probabilities, confidence, and usage, but never performs the recommended action or automatically switches models or routes all requests.
• Local check plus real connection test: offers a local status check and a real connection test, the latter consuming a small amount of TypeSafe API usage.
• Two entry points: ships as a native DSH plugin and as an iPolloWork import plugin, so one shared decision core can serve OpenCode and Codex Harness through the iPolloWork extension interface.

How to Use This Plugin?

After enabling it in DSH, configure your own TypeSafe API Key, restart the session so the new configuration takes effect, then say in the conversation: use Jev, check the configuration first, then test the connection. The plugin exposes three tools for local status, real connection testing, and structured evaluation, plus a decision Skill. A local "configured" state does not mean the key has been verified, and the connection test sends a real request. Typical uses include asking Jev which installed Skills a task needs, which available agent should own a competitive analysis, or how a result scores on accuracy, completeness, and clarity. Pass only the minimum necessary context when evaluating, never keys or unrelated private material; the plugin caps each call at 64 questions, 256 KiB request, 1 MiB response, and a 20-second total network timeout, retrying at most twice on 429/529. A recommendation is not execution authorization, so external actions still follow the current harness's permission and confirmation flow.

How to Troubleshoot This Plugin?

If an iPolloWork plugin package is rejected on import, the client version usually lacks the official signing public key registration; meeting the minimum version number in the manifest is not enough, so use a Work version that includes that registration and do not bypass signature verification. If local status shows configured but the connection fails, the key has not actually been verified, so review the credential setup per the official instructions and retry the connection test. If the plugin is unavailable under a custom DSH configuration, check that the credential, tool, and skill services it reuses are still present.