Evidence-preserving log reducer

The idea

For a long log that no parser read, a fast model proposes a short receipt. Code checks its hash, exit status and exact quotes before use, and the last 20 lines stay as they are.

The problem
Long logs cost many tokens, though a few lines show the failure.
Why it matters
The worker reads a short, checked receipt, and any failed check sends the original.
Where it stands
In production as of

A step-by-step replay of this idea is not published yet. The library adds one when a recorded run shows it.

  • ObservationPack and folded results

    A tool result of 10 KiB or more is stored whole. The step keeps the head, the tail and exact diagnostic anchors. Older results fold to one recall line, and the worker can read exact pages back.

  • Shadow first, promote on measured outcomes

    A new decision point starts in shadow. It is scored offline on labelled cases and live on outcomes, then promoted or demoted on that evidence. It acts only on the answer keys measured safe.

Turn the numbers into shipped work.

Crewshal runs these choices for you: a persistent AI worker with memory and rules, on your Claude and Codex subscriptions.