v0.4.2 — JEV decision mode / 桌面版 JEV 决策模式

JEV decision mode 0.4.2 / JEV 决策模式

Initial public release, independently maintained by QI-seven33. Unofficial community plugin for DeepSeek Harness Desktop 0.2.0-rc.2.

Install / 安装

Open Harness Desktop → Plugins, install and enable github:QI-seven33/dsh-jev-mode#v0.4.2, then restart completely. Alternatively download dsh-jev-mode-0.4.2.tgz below and install its absolute local path through the same UI. Select JEV decision mode in a new session.

Settings → Models → JEV decision model → Edit lets each user enter their own TypeSafe API key. The collapsed card opens a native modal. Keys use Harness's write-only credential service; no shared credential is included. Keep your existing DeepSeek main-model settings. The source zip is for source inspection, not the plugin installer.

桌面版：「插件」安装并启用 → 完全重启 → 「设置 → 模型 → JEV 决策模型 → 编辑」填写自己的 TypeSafe 密钥 → 新会话选择 JEV 决策模式。绿色状态点仅表示密钥已配置，不表示已测试连通性。

Scope

• Explicit package.json / README.md reads use code and native permissions, with zero JEV selection calls.
• Eligible bounded ambiguous reads may use JEV; invalid/low-confidence/unavailable results fall back to ordinary planning.
• Risky operations retain hard rules, semantic authorization and authoritative downstream permissions.
• Ordinary code/document tasks do not receive automatic advisory JEV calls. Quoted document commands do not count as execution.
• Fixed prompt/tool prefixes and producer-owned appended history preserve the host cache structure. Metadata-only ledger.
• No automatic grounded/sufficient loop, per-decision badges, guaranteed speed, fixed cache ratio or calibrated safety accuracy.

Validation

The reviewed source snapshot passed offline tests and native host regressions. The final tgz was installed through the actual host PluginManager in a fresh isolated profile: enabling, preset composition, global settings anchor and browser-module discovery passed. Component browser checks use actual host React/Modal with a stubbed credentials remote.

The decision implementation is unchanged from 0.4.0. Its earlier 56-execution comparison achieved 28/28 scripted task successes in each arm; median 3158.5 ms → 1807 ms and average main-model requests 3.1786 → 2.1071, but P95 worsened 4974 ms → 6250 ms. These small local samples do not establish universal speed or safety improvement. See VALIDATION.md.

Public transport verification passed: anonymous downloads matched SHA256 checksums; both the pinned GitHub source address and public Release tgz URL installed through the native PluginManager into separate fresh profiles, with enabling, preset/settings composition and client discovery confirmed. Node 20/22/24 CI passed. These checks used the published artifacts, not a link to the author's checkout. CI reports native-host skips on GitHub runners explicitly. Compatibility with other host versions/operating systems and real API responses requires user acceptance.

Files: installable tgz, clean source zip and SHA256SUMS.txt. MIT licensed. 中文安装说明.