Overview & result
yangyu666 shows dsh-jev-prune, built on Jev: Context compaction for DeepSeek Harness, judged by Jev.
From the original post
dsh-jev-prune by @yangyu666: Context compaction for DeepSeek Harness, judged by Jev.
How Jev fits in the loop
- When a long tool result or old history needs reclaiming, the plugin asks Jev yes/no questions about whether each piece is still useful for the current goal.
- The state carries the most recent user instruction, so usefulness is judged against the task.
- Probabilities are used as a relative ranking within the session, never as a fixed threshold.
- Kept text stays verbatim; compacted regions are replaced by a deterministic receipt written by code, not by a model.
How to reproduce
- Review the source documentation and repository at https://github.com/yangyu666/dsh-jev-prune.
- Configure environment with TypeSafe Jev / System One API keys.
- Run test prompts against the decision endpoints.
Why this build matters
Shows Jev in practical use for browser and autonomous agents.
Limitations
- Measured results come from one 37-step file-reading task, compared across three arms.
- Write-type calls and the recent window are protected by hard rules, not by probabilities.
- The author notes a small judge model's output distribution is narrow, so only relative ordering within a session is reliable.