DeepSeek Harness Jev Plugin

No. Every feature is independently configurable from the plugin's Jev settings page and all are disabled by default; enabling the package alone enables none of them. Thresholds, counts and the shared timeout can be adjusted without enabling any feature.

It supports Jev through a System One endpoint such as api.typesafe.ai/v1/systemone with jev-latest, and Luna Decisions through OpenRouter or OpenAI. Each profile stores up to three separate connections, each with its own address and credential reference.

Skill selection ranks skill names and summaries before catalog publication; file ranking orders the original glob results without another scan or content read; drift reminders and instruction guidance deliver nonblocking nudges; completion checks and goal supervision review answers and progress; long-log and test-log admission filter command output while protecting failures; workspace approval can answer eligible escalations in workspace-write; and stage navigation links recorded steps back to trajectory evidence.

It is early stage, pinned to DSH 0.2.1-alpha.2 and Cordis 4.0.5-alpha.1 with other harness versions unsupported. A public testing report separates fixed protocol and installation checks, controlled live samples and real tasks: official OpenAI calls succeeded in the limited samples, the only live OpenRouter diagnostic returned HTTP 403 and was stopped, and the report states it does not establish general judgment accuracy or stable task benefit.

No. The main model continues to plan, generate answers and call native tools; DSH configures it separately. Enabled features use the saved Jev or Luna Decisions connection at DSH extension points and then apply the result per feature, through public Cordis and DSH plugin APIs without modifying the host source.

Five skill summaries are ranked, at most 40 glob matches are eligible, and 12 paths are displayed. The main agent still loads the original skill, so ranking changes what it sees first rather than what it is allowed to load.

MIT-licensed and an independent community project, not an official DeepSeek or Jev release. Its APIs and model judgments are explicitly not a correctness guarantee, and a Chinese feature site documents each integration point, what information is sent to the judgment service, and the observed limits.