ghalebdweikat / winnow

winnow

winnow sits between Claude Code's tools and your context. After Read, Bash, or Grep returns a large result, winnow splits it into blocks, asks a System One model (TypeSafe's Jev, or an LLM-backed adapter) one yes/no question per block ("is this block needed for the current task?"), and replaces the low-probability blocks with a stub.

What a stub looks like

[winnow] Lines 41-188 (148 lines) hidden: judged unlikely to matter for the current task (relevance <= 0.22).
[winnow] Summary: Boilerplate license header and the argparse setup for flags that are not part of this task.
[winnow] Full text cached as key a1b2c3d4e5f6. Call winnow_recall(key="a1b2c3d4e5f6", start=41, end=188) if you need it.

How to treat it

• The summary is written by a cheap model and preserves identifiers, paths, numbers and error text. Trust it for orientation, not for exact values.
• If the hidden lines might matter after all, call the winnow_recall tool with the key. Add start and end to fetch only part of the cached text. Do not re-run the original command just to see the hidden part; recall is cheaper and does not re-trigger side effects.
• Outputs that show an error are never pruned. If you see a stub, the judge believed the output was clean.
• Line numbers in a stub refer to the original file or output. Claude Code renumbers the rewritten result from 1, so the numbers you see in the margin of a pruned Read do not match the file; use the stub's range with Read offset/limit or winnow_recall to reach the original lines. Edit is unaffected because it matches on text, not line numbers.

Tools

• winnow_recall(key, start?, end?): return the cached full text for a key.
• winnow_stats(): tokens saved, outputs rewritten, and the regret rate (how often a pruned key was later recalled).