compact-jev

/compact-jev is a Claude Code plugin that frees context without writing a summary. It asks Jev, through Vercel AI Gateway, which old tool calls and results still matter. It keeps your and Claude's messages verbatim, removes stale calls, and shortens stale results. It runs only when you type /compact-jev; Claude Code's /compact and automatic compaction still work normally.

The repository also contains a TypeScript library for using the same compaction logic outside Claude Code.

Install

You need Claude Code 2.1.274 or later and your own Vercel AI Gateway API key.

1. Enable Claude Code function hooks in ~/.claude/settings.json (merge this with your existing env block):

   { "env": { "CLAUDE_CODE_ENABLE_FUNCTION_HOOKS": "1" } }

2. Install the plugin:

   claude plugin marketplace add edoproch/compact-jev
   claude plugin install compact-jev@compact-jev

   On Claude Code 2.1.275+, you can instead run /plugin install compact-jev --marketplace edoproch/compact-jev inside a session.

3. Set the AI Gateway API key in the plugin's configuration (/plugin). The apiKey field is stored as a sensitive value. Alternatively, set AI_GATEWAY_API_KEY in the environment before starting Claude Code. Restart Claude Code, or run /reload-plugins.

Use

Run /compact-jev in a long conversation. Optionally add the task you want Claude to continue:

/compact-jev
/compact-jev fix the failing test in src/cart.ts

Jev judges tool calls against that task. Without an explicit task, it uses your last three requests. The first message and the six newest messages are always preserved. The command reports how many messages it kept, whether it used the no-training fallback, and per-call decisions in the debug log (claude --debug). If Jev or the Gateway fails, the conversation stays as it is; no summary replaces it.

For long conversations, Jev finishes its analysis before Claude Code starts compaction. The plugin keeps at most four Gateway requests active at once. Claude Code exposes at most 4096 messages through the preflight API; at that limit the command leaves the conversation unchanged.

Measured example

I ran four read-only turns in Claude Code 2.1.282 with --model sonnet (resolved to claude-sonnet-5), then replayed the same 14-tool-call transcript through Jev three times per condition. All three runs in each condition produced the same decisions:

	No goal (last three requests)	Goal: fix the cart rounding bug
Transcript characters	60,171 → 11,023 (−81.7%)	60,171 → 8,384 (−86.1%)
Historical tool calls removed	1 / 14	9 / 14
Tool results shortened	9	2

With the explicit goal, old Read calls about billing and release notes and two Glob calls were removed. Three recent calls (Glob, Grep, Bash) were protected in both replays. I also ran /compact-jev inside the actual Sonnet session: it kept 23 of 43 messages, removed 10 historical calls, and reported 86% character reduction without a summary. The files were synthetic and deliberately long, so this is an example, not an expected reduction rate. We have not measured a comparable before/after count of Sonnet context tokens; these percentages count transcript characters. See the full method, per-call decisions, fixture, and live runner.

Local Laya experiment

I tested Laya as a local replacement on the same compaction core, including English and Italian user prompts. On an Apple Silicon Mac with 16 GB of memory, the multilingual checkpoint (322 million parameters, 644 MB of weights) took 25–66 seconds for full-state cases and removed no historical calls; Jev took 1.3–1.4 seconds and removed nine with the same explicit goal. A different Laya question format removed all calls, including relevant code and test reads. Because quality and runtime are not yet acceptable for this workload, the plugin still offers Jev only. See the measured comparison and its limitations.

Privacy and Vercel plans

Each run sends user and assistant text, tool inputs, and short excerpts of candidate tool outputs to Vercel AI Gateway and TypeSafe AI. The state is repeated across batches. Set outputExcerptChars to 0 in /plugin if you do not want tool output sent; this reduces Jev's accuracy. Your full tool outputs are not sent by default.

Vercel plan attached to the API key	Gateway request policy
Pro or Enterprise	Requests first require zero data retention (ZDR) and no prompt training.
Hobby	ZDR is not available. When the Gateway rejects ZDR, the plugin automatically retries with no prompt training still required.

The plugin cannot infer the plan from the key itself. It tries the stronger policy first and only changes it after an explicit ZDR-related error. A different error does not weaken the policy. If no provider meets the required policy, compaction stops and leaves the conversation unchanged. No prompt training does not mean zero retention: on the Hobby fallback, TypeSafe AI may retain data under its applicable terms. Vercel says per-request ZDR requires Pro or Enterprise, while no prompt training is available to all users.

A Vercel Hobby plan and the AI Gateway free credit tier are different things. Check the Jev model page for its current price.

Configuration

The defaults work without tuning. Set options in /plugin:

Option	Default	Purpose
apiKey	AI_GATEWAY_API_KEY	Your Gateway key.
keepThreshold	0.5	Probability needed to keep a call or result.
preserveRecentMessages	6	Newest messages protected from compaction.
outputExcerptChars	240	Characters of each tool output sent to Jev; 0 sends none.
truncateHeadChars	300	Characters kept from a shortened tool result.
maxStateTokens	8000	Estimated state size per Jev request.
maxRequestTokens	25000	Estimated request size per batch.
maxQuestionsPerRequest	20	Questions per batch; two are asked per tool call.
retries	12	Extra tries for Gateway 429 and 5xx responses.
model	typesafe-ai/jev	Gateway evaluation model.

The key is resolved from the sensitive plugin option first, then the process environment, then env.AI_GATEWAY_API_KEY in Claude Code settings. Prefer the sensitive option: settings environment values are also passed to child processes.

TypeScript library

For Node.js 18 or later, install the npm package with npm install compact-jev. Claude Code users only need the plugin installation above. compactMessages(messages, options) uses AI_GATEWAY_API_KEY by default; pass onZdrFallback in its options to detect a no-training-only retry. To supply your own transport, implement JevAsker and call compact(messages, asker, options). See src/index.ts for exports and examples/demo.ts for an example.

Development

npm install
npm test
npm run typecheck
npm run build
npm run validate:plugin

To load the checkout directly, run CLAUDE_CODE_ENABLE_FUNCTION_HOOKS=1 claude --plugin-dir .. Function hooks are an early-access Claude Code API and may change between releases. The checked-in type declarations target Claude Code 2.1.274.

This project is a fork of tamaratran/fast-jev-compaction. Licensed under MIT.