The philosophy of DeepSeek Harness (DSH) is “everything is a plugin.” deepseek-harness-jev is a native DSH plugin that uses Jev by TypeSafe AI or Luna Decisions as its judgment model. It does not modify host source code; instead, it integrates through the public Cordis / DSH plugin API and assists with agent skill and file selection, task supervision, shared-finding corrections, tool output filtering, and single-operation approvals.

This is an independent community project, not an official release from DeepSeek or Jev. It is an early-stage plugin with limited DSH compatibility; its API and model judgments are not guarantees of correctness.

Core Features

The plugin includes the following 12 features. All features are disabled by default; the installation package does not enable them automatically.

  • Skill selection: Ranks skill names and summaries before catalog publication; the main agent still loads the original skills.
  • File ranking: Ranks raw glob path results without rescanning the filesystem or reading file contents.
  • Drift reminders: Checks progress between model steps and can send a non-blocking reminder.
  • Completion checks: Reviews the visible final answer and recorded evidence, and can request at most one supplemental attempt.
  • Goal supervision: Checks native goal completion and pauses after a configured number of turns without progress.
  • Instruction guidance: Reads the current user instruction and applicable agent rules, and provides a non-blocking reminder when needed.
  • Interjection routing: Routes in-flight user corrections to the next step; queues other messages for a later turn.
  • Shared-finding corrections: Compares shared reports and messages, then sends corrections to affected recipients.
  • Long-log admission: After a command returns, can remove obviously unneeded progress notices or repeated notifications, with references for restoring the original output.
  • Test-log admission: Preserves failed, summarized, named, and slow tests while determining whether ordinary pass details are needed.
  • Workspace approval: In workspace-write mode, it can answer eligible native single-operation escalation requests; non-affirmative answers return to human approval.
  • Stage navigation: On request, clusters the complete recorded model steps into stages, then links consecutive stages to their original trajectory evidence.

Installation and Enablement

Currently supported DSH versions include 0.1.7-rc.1, 0.1.7-rc.2, 0.2.0-rc.1, 0.2.0-rc.2, and 0.2.1-alpha.2. Installation does not require source checkouts or manual packaging.

  1. Open the DSH sidebar’s Plugins → Add plugin.
  2. Paste the GitHub repository URL into Package name or address, then click Install.
  3. After installation completes, click Enable now. No restart of the current profile is required unless DSH prompts you to load it on next startup.
  4. In the installed plugins list, select View @dsh-jev/plugin to open the configuration page. Configure the Jev or Luna Decisions endpoint, model, and API key here, and enable the desired features one by one.

Typical Usage

Based on verified usage examples, the plugin performs the following operations in the DSH execution flow:

  • Skills and file processing: Ranks skill names and summaries before catalog publication; ranks glob path results without triggering extra filesystem scans or content reads.
  • Progress and goal supervision: Checks progress between model steps and sends a non-blocking reminder when there is no progress; checks native goal completion and applies a configured pause threshold.
  • Instruction and correction routing: Reads the current user instruction and applicable agent rules, and provides a non-blocking reminder when needed; routes in-flight user corrections to the next step and queues the remaining messages.
  • Answer and evidence validation: Reviews the consistency between the visible final answer and recorded evidence, and requests a supplemental attempt when needed; compares shared reports and messages, and sends corrections to affected recipients.
  • Log and test management: Cleans up obviously unneeded progress notices or repeated notifications after commands return; preserves failed, summarized, named, and slow tests while retaining only necessary pass details.
  • Workspace and stage navigation: Handles single-operation escalation requests in workspace-write mode; clusters complete recorded model steps and links them to their original trajectory evidence.

Notes

  • Compatibility: The plugin is early-stage and has a limited compatibility scope. Its API and model judgments do not guarantee correctness.
  • Default state: All features are off by default and must be enabled manually on the configuration page.
  • Branch status: The main branch includes tool output filtering and stage navigation. codex/jev-tool-output-admission retains a development snapshot. The native web execution feature is located in the codex/jev-native-web-execution branch and is currently paused.
  • Model connection: Official OpenAI calls succeeded in limited samples; the only live OpenRouter diagnostic returned HTTP 403 and was stopped.
  • Accuracy statement: Labels and confidence scores do not establish tool success or classification accuracy.

Summary

deepseek-harness-jev provides DSH with enhanced capabilities based on external judgment models. Through Jev or Luna Decisions, it can assist with skill selection, task supervision, log cleanup, and stage analysis. Before using it, confirm DSH version compatibility and enable each feature independently as needed.