jev-pruner
Claude Code plugin: trim long Bash output with TypeSafe Jev before the model sees it.
Claude Code plugin: trim long Bash output with TypeSafe Jev before the model sees it.
After a Bash command, Jev receives the task context, history, command, and output chunks, with a yes/no question for each chunk about whether any line must remain available. The hook preserves chunks Jev marks necessary and removes the rest before returning output to Claude.
Decide whether chunk c1 contains any line that must remain available for the task. The request includes a category derived from the chunk; compare every line with the task and history. Keep uncertain or unclassified information unless all lines are confidently disposable.
For the first output chunk, state includes the agent context, current task goal, history, command, diagnostics/results, and a chunk object with its ID and text. Chunk IDs are assigned in order, so the first is c1. Build/search categories also include category guidance.
Allow the Bash command to complete and capture its output.
Pass through output below the size gate and protected error, structured, code, or whole-document output.
Split eligible output and assemble task, conversation history, command, and relevant metadata.
Ask Jev whether each chunk contains a needed line, then return retained output with pruning handled by the hook.
It combines relevance filtering with explicit protections for short outputs, errors, structured data, source code, and whole-document commands, aiming to trim context without asking a model to summarize it.