fast-jev-compaction
Claude Code plugin that replaces the compaction summary with Jev decisions: every tool call and result is scored in one fast request, stale ones are dro
Overview
fast-jev-compaction is an npm package and Claude Code plugin that replaces summarization with Jev keep-or-drop decisions on paired tool_use and tool_result messages.
Problem: LLM-written compaction summaries drop exact errors, paths, and constraints from agent transcripts.
Built for: Claude Code users who want compaction that deletes stale tool rows without rewriting user or assistant text.
First indexed on Jev Directory: 2026-09-19
Creator and team
- Name
- Tamara Tran
- Handle
- @tamarajtran
“Shrink a messy thread before your agent drowns in tokens. Jev keeps the plot, drops the fanfic.”
How Jev is used
- Role in the product flow
- Context compaction gate on agent tool history
- Primitives
- Noul
- State in
- Full conversation with tool results replaced by short notes, fitted into maxStateTokens (25k default) using staged truncation described in README.
- Decision out
- Per tool call: keep result verbatim, keep call with truncated result, or drop call and result based on keepThreshold.
- Pair tool_use with tool_result; pin recent and first messages
- For each non-pinned call, ask two noul questions (keep call?, keep result verbatim?)
- Shard questions across concurrent requests under maxRequestTokens (30k)
- Rebuild message list without summarizing user or assistant text
Compaction never paraphrases content. Jev only answers whether a tool call or its result still matters given the entire transcript state. Noul questions are explicitly documented in the README, including parallel requests when state plus questions would exceed Jev request limits. Failures throw so the Claude Code hook can fall back to default behavior.
Sourced performance claims
- Default keep threshold applied to keepResult and keepCall noul answers.Source: github.com/tamaratran/fast-jev-compaction README
Features and stack
Features
- npm library plus Claude Code plugin hooks
- Preserves user and assistant messages verbatim
- Concurrent shard requests under Jev limits
- Configurable preserveRecentMessages and thresholds
Stack
- TypeScript
- Claude Code plugin
- TypeSafe System One
- npm
Pricing: Bring your own TYPESAFE_API_KEY per README install section.
Demo
Showcase clip with the same lightbox player as the homepage. Click to play.
Instant context compaction
Shrink a messy thread before your agent drowns in tokens. Jev keeps the plot, drops the fanfic.
Links
FAQ
- Which primitive does compaction use?
- The README specifies two noul questions per non-pinned tool call: whether to keep the call and whether to keep the result verbatim.
Related learn guides
Original Jev guidance that pairs with this product pattern.
- Jev use cases
The patterns builders actually search for: moderation, routing, triage, RAG verify, and agent gates.
- System One model
The model family behind Jev: parallel typed questions, one forward pass, probabilities you can threshold.
Related products
Hand-picked neighbors with rich profiles or overlapping tags.