Winnow
A Claude Code hook that filters large tool results before they enter the agent's context.
Added to Jevfast
How it uses Jev
For a large tool result, winnow reconstructs the current task from the session's latest user request and assistant intent, then sends that task, the tool name and input, and result blocks as state. Jev answers Noul questions about whether each block is needed and whether the output contains an error. Code keeps confident-yes and uncertain blocks, and hides only confident-no blocks when no error is detected; the hidden text is cached for recall and replaced by a stub with a summary or digest. A separate prompt hook uses Noul judgments to rank relevant memory files for injection.
What Jev decides
Illustrative Jev decisionExample answers · not a recorded Jev response · Source ↗
Question 1 · b001
YOUR APP
INSTRUCTION
Would the assistant have to look at this result block to accomplish the task correctly?
STATE
Current task, tool name and input, and numbered blocks from a large tool result.
JEV · NOUL
YesNo
App workflow
- Extract and split results
For eligible large Read, Bash, or Grep results, extract text and split it into blocks.
- Build task-aware state
Recover the latest user request and assistant intent; send them with the tool input and judged blocks.
- Ask Noul questions
Ask whether each block is needed and whether the output contains an error.
- Apply the pruning policy
Keep missing or uncertain probabilities and all blocks when an error is detected; prune only below the drop threshold.
- Rewrite and preserve access
Cache full text, summarize or digest hidden groups, and return rewritten output containing recall keys.
Why it is interesting
The task state is reconstructed because the hook does not receive it directly, then used for per-block judgments with explicit uncertainty and error safeguards. Hidden text remains locally recoverable, so context reduction does not discard the source material.