Save Token Jev Clean
A token-saving context compactor that uses Jev Noul judgments to choose which tool calls and results to retain.
Added to Jevfast
How it uses Jev
Asks two Noul questions per call, batches them to fit request limits, then keeps or omits content based on returned judgments.
What Jev decides
Illustrative tool-call retention questionsExample answers · not a recorded Jev response · Source ↗
Question 1 · call {id}
YOUR APP
INSTRUCTION
Whether knowing this tool call ran with its input still matters for the agent next actions.
STATE
Fitted conversation history and each candidate tool call id, name, and result character count.
JEV · NOUL
YesNo
Question 2 · result {id}
YOUR APP
INSTRUCTION
Whether the full result must remain verbatim because exact contents are still needed and rerunning cannot recover them.
STATE
Fitted conversation history and each candidate tool call id, name, and result character count.
JEV · NOUL
YesNo
App workflow
- Collect history
The host invokes compaction and supplies conversation history and candidate tool calls.
- Judge retention
Jev answers call-retention and result-retention Noul questions in bounded batches.
- Pack context
The compactor applies its retention threshold and fits selected history under the token budget.
Why it is interesting
Treats context reduction as a typed retention decision and explicitly budgets request context.