winnow

A calibrated context sieve for Claude Code.

Every large Read, Bash, or Grep result is judged before it enters Claude's context. Blocks the judge is confident you don't need are replaced with a three-line stub: what was hidden, a one-paragraph summary from a cheap model, and a key that restores the full text on demand. Nothing is lost; it just stops costing tokens until you ask for it.

Terms used below. The judge is the model that answers one yes/no question per block ("is this block needed for the current task?") with a probability. By default that is Jev, TypeSafe AI's System One model: a model that returns calibrated probabilities for typed questions instead of generating text, so a hundred questions come back in one call in a few hundred milliseconds. Jev is in early access. The adapter is TypeSafe's system-one-adapter package, which answers the same questions by prompting Claude Haiku 4.5; it is not calibrated, but it lets the whole pipeline run today.

What it does

Read big.py  ──►  Claude Code  ──►  tool.call hook  ──►  winnow
                                                             │
        split into ~25-line blocks ◄─────────────────────────┘
        one call to the judge: "is block N needed for the current task?"  ×N, in parallel
        keep confident-yes and uncertain blocks verbatim
        hide confident-no blocks:  cache full text  ─►  summarize  ─►  stub
                                                             │
Claude sees ◄──  { result }  ◄───────────────────────────────┘

A stub looks like this:

[winnow] Lines 41-188 (148 lines) hidden: judged unlikely to matter for the current task (relevance <= 0.22).
[winnow] Summary: Argparse setup for the --export and --format flags, plus the license header.
[winnow] Full text cached as key a1b2c3d4e5f6. Call winnow_recall(key="a1b2c3d4e5f6", start=41, end=188) if you need it.

Without a summarizer the middle line is a deterministic digest instead, so Claude still knows what kind of thing it lost: for a search result, which files the hidden matches came from and how many each; for anything else, the line count, whether it was mostly comments, imports or repetition, and the first line.

Two safety rules are built in. If the judge thinks the output shows an error, nothing is hidden. If a block's probability is merely uncertain (between WINNOW_DROP and WINNOW_KEEP), it is kept. Both thresholds are tunable; the rules themselves are not optional. The default WINNOW_DROP of 0.1 is the bin that came back clean on hand-labeled replay (see Measured); raise it only with your own evidence.

The hook is a function-hook module (see How it hooks in) that talks to a small resident server (winnow serve) on loopback, started at session start, so a judged call costs about 16 ms plus the judge call rather than a Python startup. winnow never judges its own files or its own commands, so recalls and labeling sheets always come back whole.

A second hook runs at prompt time. It ranks the memory files Claude Code keeps for the project (~/.claude/projects//memory/*.md, everything except the MEMORY.md index, which Claude already loads) plus any directories in WINNOW_CONTEXT_DIRS against your prompt, and injects the relevant ones so Claude reads what it needs without a round of Read calls.