jev-dsh-decision
Manifest valid★ 237Jev DSH Decision Engine | Structured Decision Plugin for Agent Harness. Native support for DeepSeek Harness, with OpenCode and Codex Harness support via iPolloWork.
Jev DSH Decision Engine
Structured decision-making for Agent Harnesses.
Connect TypeSafe AI's Jev decision capability to your Agent: help select tools, Skills, and task owners, evaluate output quality, and return judgments and probabilities as structured results. Jev assists with decisions, while the original Agent continues to handle planning and execution.
Two installation entry points are provided: a DeepSeek Harness native plugin and an iPolloWork import plugin. Within iPolloWork it can be used by OpenCode, DeepSeek Harness, and Codex Harness. One shared decision core — no engine source code modifications, no takeover of the primary Agent.
Quick start: Install on DeepSeek Harness · Install on iPolloWork
Support and Integration
| Environment | Integration | Install Entry Point |
|---|---|---|
| DeepSeek Harness (DSH) | Native plugin, runs standalone, no iPolloWork dependency | DSH plugin command |
| iPolloWork | Independent import, authorization, start/stop, and uninstall | Extensions → Import Plugin Package |
| OpenCode | Uses Jev via iPolloWork's extension interface | Install once in iPolloWork |
| Codex Harness | Uses Jev via iPolloWork's extension interface | Install once in iPolloWork |
The current support for OpenCode and Codex Harness is via iPolloWork's integration method; this repository does not yet provide their individual standalone native install packages. Other Harnesses can build adapters based on the shared decision core, but that does not imply automatic compatibility.
Core Capabilities
- Recommends from currently available and authorized tools, Skills, and Agents for a task.
- Scores against explicit criteria; independent questions can be batched.
- Returns choice, probability, confidence, and usage; does not execute actions or auto-switch models or route requests.
- Local config checks and real connection tests. Connection tests incur a small TypeSafe API usage.
Jev is TypeSafe AI's structured decision model, not a chat model. The original Agent continues to handle planning, explanation, parameter construction, tool execution, and native sub-Agent collaboration. Jev does not return text reasoning processes; the Agent's explanations cannot masquerade as Jev's raw output.
Install on DeepSeek Harness
For users who use DSH directly. No need to install or run iPolloWork.
1. Install the Plugin
After installing DSH, clone this repo and install it via a local absolute path:
git clone https://github.com/Devin-AXIS/jev-dsh-decision.git
dsh plugin add /absolute/path/jev-dsh-decision
Replace /absolute/path/jev-dsh-decision with the actual absolute path of the cloned directory. The plugin uses the native dsh.bundle format; no source build required.
If you already have the .tgz install package, you can install it directly:
dsh plugin add /absolute/path/dsh-jev-decision-model-0.1.2.tgz
2. Configure Authorization
Configure your own TYPESAFE_API_KEY in DSH's credential management or startup environment. Do not write the key into plugin source code or chat logs. The plugin reuses DSH's credentials, tools, and skills services; the official default configuration provides these services, and custom configurations must retain them.
3. Verify and Use
Restart the running DSH session to apply the new configuration, then say in the chat:
Use Jev to check config first, then test the connection.
The plugin provides three tools: jev_status (local state), jev_check_connection (real connection test), and jev_evaluate (structured evaluation), along with the jev-decision Skill. Connection tests incur a small amount of TypeSafe API usage; a local "configured" display does not mean the key has been verified.
Uninstall
dsh plugin remove dsh-jev-decision-model
DSH credentials are held by the host. Uninstalling the plugin will not delete a shared TYPESAFE_API_KEY; remove it via DSH credential management or the environment when no longer needed. Restart running DSH sessions to load new plugin config.
Install into iPolloWork
For users of the Work GUI. Install once in Work to serve its supported OpenCode, DeepSeek Harness, and Codex Harness engines — no need to install three packages separately.
1. Import the package
Obtain the signed jev-decision-model-0.1.2.ipollowork-plugin package, select it under Extensions → Import plugin package, and confirm the publisher is iPolloWork. See "Development and Testing" below for how maintainers build the package.
When an older version is already installed, importing a new one upgrades while preserving the original plugin identity. The display name is "Jev DSH Decision Engine"; the internal plugin ID jev-decision-model and the DSH package name dsh-jev-decision-model remain unchanged.
Client requirement: a Work version that includes the ipollowork/ipollowork-2026 official signature public key registration. Older clients without this registration will reject the import; merely meeting the minimum version number in the manifest is not sufficient. Do not bypass signature verification.
2. Configure authorization
Open the authorization entry in the plugin details and enter your own TypeSafe API Key. The key is stored in Work's plugin authorization vault and does not need to be sent to the Agent in the conversation.
3. Verify and use
Enable "Jev DSH Decision Engine" in the conversation, then say:
Use Jev: check config first, then test the connection. Once it succeeds, pick a Skill suitable for the current task.
Work provides three actions: status, check-connection, and evaluate, all invoked through the existing extension interface. The "test connection" here sends a real request with a small amount of usage.
Uninstall
Remove it via Extensions → Jev DSH Decision Engine → Uninstall. Work cleans up the Skill, plugin content and dedicated auth along with the package; no global scripts or daemons are installed. Uninstalling does not delete your TypeSafe cloud account.
Examples
- "Use Jev to determine which installed Skills this task needs, then execute it."
- "From the currently available Agents, pick the one best suited to lead competitor analysis."
- "Assess this output for accuracy, completeness, and clarity; hand the uncertain parts to me to confirm."
Recommendations are not execution authorization; external actions still follow Harness permissions.
开发者参考:调用格式、签名与测试
Call Format
evaluate and jev_evaluate use the same parameters:
{
"state": { "task": "为已收集的竞品资料制作汇报", "available": ["research", "slides"] },
"questions": {
"owner": {
"type": "choice",
"instructions": "哪个已提供的 Agent 最适合负责最终交付?",
"criteria": { "research": "负责资料分析", "slides": "负责演示文稿" }
},
"need_review": { "type": "noul", "instructions": "这个任务是否需要审稿?" },
"clarity": { "type": "score", "instructions": "任务要求是否清楚?", "criteria": ["不明确", "部分明确", "明确"] }
}
}
The question type is choice (single-choice), score (level rating), or noul (probability of yes). Multiple complementary Skills may use multiple independent noul questions. Pass only the minimum necessary context; do not pass secrets or unrelated private material. Network requests are pinned to https://api.typesafe.ai/v1/systemone; the default model is jev-latest.
Plugin caps: 64 questions per call, 256 KiB request, 1 MiB response, 20 s total network budget; retries only on 429/529, at most twice. No unbounded retries, background auto-learning, or global request interception. Errors are sanitized; upstream error bodies are not returned.
Dev & Test
Node >=20.11, pnpm; zero third-party runtime dependencies. JSZip is used only during development to build Work packages.
pnpm install --frozen-lockfile
pnpm test
DSH_RUNTIME_ROOT=/path/to/dsh-runtime pnpm test:dsh
pnpm package:dsh
DSH_RUNTIME_ROOT=/path/to/dsh-runtime DSH_CLI_TEST=1 pnpm test:dsh
Work packages use the host-required Ed25519 signature format. The private key is kept outside the repo; not shipped with source, installers, or releases:
IPOLLOWORK_PLUGIN_SIGNING_KEY=/secure/path/ipollowork-ed25519.pem \
IPOLLOWORK_PLUGIN_KEY_ID=ipollowork-2026 pnpm package:work
IPOLLOWORK_ROOT=/path/to/updated/iPolloWork pnpm test:work
pnpm package:dsh
The signing identity is ipollowork/ipollowork-2026; the public key (SPKI DER, Base64) is:
MCowBQYDK2VwAyEAoHA9xxXNZPQ7y+AKpDBIXruNvwVasq7axjilzIY6RDE=
Work integration tests use the real host API, testing signature, icon, import, authorization rotation, service invocation, uninstallation, and tamper rejection in an isolated temporary directory. TypeSafe responses use explicit test data and do not represent verified real API keys. DSH tests use the real tool and Skill registry to check loading and uninstallation.
Sources and Identity
- TypeSafe AI Official API Documentation
- DSH Official Plugin Development Documentation
- The icon is taken as-is from the black icon on transparent background in TypeSafe's official documentation; it is embedded in the installer for offline display, and the pink version is not used. Icon and brand rights belong to the original owner.
This project is maintained and released by iPolloWork. Jev belongs to TypeSafe AI; this project does not imply official endorsement by TypeSafe AI, DeepSeek, or any other Harness project.
Comments
Loading…
Similar plugins
by nanami-0713
DSH plugin: register TypeSafe Jev (System One decision model) as an agent tool — jev_decide returns calibrated probabilities (noul/choice/score) for routing/triage/guardrail judgments, no text generation. 把 TypeSafe Jev 决策模型注册为 DSH agent 工具
★ 3
↓ 313/wk
MIT
JavaScript
Sep 24, 2026
dsh plugin --profile web add dsh-jev-decideby AskTheWay
⚡ Millisecond System-1 judgement for every tool call in DeepSeek Harness — Jev-powered risk classification & evidence-gated auto-approval. Fail-closed by construction. dsh 生态第一个 System-1 决策插件
★ 21
↓ 178/wk
MIT
TypeScript
Sep 25, 2026
dsh plugin --profile web add dsh-jev-interceptorby BetterZflyee
Use the Jev (System One) decision-model paradigm with any OpenAI-compatible LLM — no TypeSafe key required. Registers a jev_decide tool on DeepSeek Harness (dsh).
★ 3
MIT
JavaScript
Sep 21, 2026
dsh plugin --profile web add dsh-jev-adapterby justhalfbit
DeepSeek Harness (DSH) 推理等级自动选择插件:由 Jev System One 模型判断每条消息值多少思考量,按模型声明的等级自动推导档位,上下文信封让「继续」这类追问继承话题深度,低置信度向上取,任何失败都静默沿用原等级。 | Jev-driven reasoning effort per message: per-model ladders derived from wh
★ 6
MIT
JavaScript
Sep 24, 2026
dsh plugin --profile web add dsh-plugin-jev-effort-selectorby kaijia323
TypeSafe Jev (System One decision model) as a native jev_decide tool plugin for DeepSeek Harness
★ 0
MIT
JavaScript
Sep 20, 2026
dsh plugin --profile web add dsh-plugin-jevby yangbobo2021
Codex integration plugin for DeepSeek Harness, providing native Codex conversations powered by the Codex App Server, with workspace, terminal, approval, and DSH tool support.
★ 14
↓ 721/wk
MIT
JavaScript
Sep 19, 2026
dsh plugin --profile web add relay-dsh-plugin-codex