Overview & result
A DeepSeek Harness plugin that puts Jev in front of the context window. Per the README it prunes long tool output by judging each segment's relevance to the task, screens fetched web pages for instructions aimed at the AI, and recommends at most one matching skill per turn; the repository description also lists gating completion claims. The main model still does all writing, reasoning and tool calls.
From the original post
Jev judgment, not generation: prune long tool output, screen fetched pages for injected instructions, and gate completion claims inside DeepSeek Harness.
How Jev fits in the loop
- When a read, grep, glob, web fetch or web search result exceeds 2,000 tokens, Jev judges each segment's relevance and the plugin keeps the relevant ones.
- Fetched web content is screened for instructions aimed at the AI; past a threshold a warning is attached.
- With 15 or more skills installed, it picks at most one matching skill per turn.
- The model can call jev_ask for any typed question and jev_gate to check completion claims against evidence.
How to reproduce
- Inspect the original showcase and source material at https://github.com/HorusJiang/dsh-jev-tools
- Verify the bounded prompt and input candidate schema configured for Jev
- Benchmark decision latency and classification accuracy against baseline models
Why this build matters
Pruning ranks instead of thresholding and fails open, while the completion gate does the opposite: any unclear path escalates, because failing to "passed" is the most dangerous error.
Limitations
- The README says it has not measured whether the pruning picks the right segments, separately from how much it keeps.
- Whether it keeps more than DeepSeek Harness's built-in truncation depends on payload size.
- The gate only judges what it is given: it does not run tests or apply patches.