The author of the Tips AI channel tested the open-source fast-jev-compaction plugin, which hands chat compaction in Claude Code over to TypeSafe AI's probabilistic Jev model: instead of a lossy summary from a generative LLM, outdated tool calls are deleted, and important content remains verbatim. The plugin can be installed today with two commands, and the jev-1.13.0 version costs $0.042 per million input tokens. The evidence base is still modest: a single test by the author and vendor documentation, with no measured compaction quality metrics.

image
image

What Happened

The experience was published by the author of the Tips AI channel: they got access to the Jev model, tested the plugin, and broke down how it changes Claude Code's behavior. The tamaratran/fast-jev-compaction repository was created on September 17, 2026, is distributed under the MIT license, and gained 1,739 stars in a matter of days. The plugin relies on the function hooks feature, available in Claude Code since version 2.1.274, and takes over chat compaction: each tool_use and tool_result is evaluated by the Jev model on two typed noul questions, "is the call itself needed" and "is the result needed verbatim." The questions go to the model in parallel and return probabilities, based on which outdated calls are deleted or truncated, and everything kept remains in the context verbatim, without retelling.

Context

Compaction is needed for agentic systems because long sessions hit the context window limit, and until recently Claude Code solved the problem in only one way: a generative LLM wrote a lossy summary of the conversation, and the loss of details was smeared across the retelling and remained invisible to the agent. Jev is built differently: it is TypeSafe AI's flagship System One model, which does not generate text but takes state as input and answers typed questions in parallel, returning probabilities suitable for direct use in code. In the current jev-1.13.0 version, one request holds 64k tokens, of which 32k is allocated for state plus the longest question. This architecture also implies economics: the model returns only structured answers, not generated text, so the decision about context importance is paid for exclusively by input tokens.

Why This Matters for the Industry

Context compaction is a narrow and expensive part of agentic systems, and in this case, for the first time, a probabilistic classifier model is used instead of a generative LLM summary: determining importance turns from a generation task into a classification task, which is orders of magnitude cheaper and, with parallel question evaluation, runs in tens to hundreds of milliseconds. The structural difference from summaries is visible in error behavior: in a retelling, information loss is blurred across the entire text, whereas here classifier errors are concentrated and silent, because what is deleted from the context cannot be recovered, and the agent will not know that something needed was thrown out. If the approach takes hold, "context management through probability-based deletion" will become a separate pattern for agentic frameworks, and System One models will become a new class of infrastructure alongside generative LLMs; in the coming months, expect independent comparisons with summaries, threshold calibration methods, and plugins offering a choice of compaction strategy. In the longer term, the same mechanism is applicable beyond compaction: to task prioritization, filtering tool results, and routing calls.

Why This Matters for Users

The plugin can be installed today: you need to enable the CLAUDE_CODE_ENABLE_FUNCTION_HOOKS=1 variable in ~/.claude/settings.json and run two commands, "claude plugin marketplace add tamaratran/fast-jev-compaction" and "claude plugin install fast-jev-compaction@fast-jev-compaction." After installation, both the manual /compact command and auto-compaction go through Jev, and if the model call fails, a fallback to the built-in summary kicks in, so the risk of breaking the workflow is small. New TypeSafe AI users get $5 for free, which, according to calculations in the post, is enough for about 125 million input tokens, i.e., dozens of large chats. The main practical caveat is related to the 32k token state limit: long sessions of 150k tokens or more will have to be cut into pieces, which the test author specifically warned about.

What Is Still Unknown / Limitations

The evidence base is still weak: a single test by the channel author plus vendor documentation, and the sources have no measured compaction quality metric, whether it is preserving important context or final savings. Price and speed are vendor claims, not independent measurement results. Honestly named limitations restrict applicability: Jev evaluates calls without their results, probability does not guarantee safe deletion, and tokens are counted by a heuristic, not a tokenizer, so the upper bound of deletion error is unknown in advance. Because of this, the conclusion to "deploy today" as a working default is premature: a phased rollout with testing on your own tasks is more reasonable.

Sources

Author

Look at AI, editorial team