Small decisions. Interesting possibilities.Submit contentSubmit
yoshi preview

yoshi

Context-pruning proxy for Claude Code and Codex: Jev judges which history is still needed, measured not claimed. POC here now, heading soon into https://github.com/compozy/compozy.

Added to Jevfast

How it uses Jev

Yoshi evaluates new conversation spans with Jev using two questions: whether a required fact would be lost and whether a binding user constraint would be lost. Only validated omissions are removed before forwarding the request to the model provider.

What Jev decides

Passage relevance judgmentExample answers · not a recorded Jev response · Source ↗
Question 1 · c0
YOUR APP
INSTRUCTION

Classify this passage's contribution to satisfying the request: does it contain needed information, is it already covered, is it irrelevant, or is context insufficient?

STATE

The request, retained assistant notes, and one candidate conversation passage.

JEV · CHOICE
  1. needed
  2. covered
  3. irrelevant
  4. uncertain

App workflow

  1. Receive an agent request

    Accept a Claude Code or Codex request at the local proxy.

  2. Identify eligible spans

    Select newly added tool arguments, results, or teammate reports above the configured size gate; keep protected content.

  3. Judge omission safety

    Ask Jev whether removing each span would lose a required fact or user constraint.

  4. Apply and forward

    Mark only validated omissions, preserve the remaining protocol fields, and forward the request to Anthropic or OpenAI.

Why it is interesting

The proxy preserves protocol structure and byte-identical retained fields while representing removed spans with explicit omission markers, making context pruning an auditable intermediary step.