What problem does cursor-clijev-compaction solve?

Native compaction in the Cursor CLI can lose file paths, test failures and constraints to summary drift. This plugin captures tool I/O, scores what matters with Jev, and re-injects kept facts verbatim after compaction finishes.

How does it use Jev exactly?

Every non-pinned tool call generates two atomic noul questions sent to POST https://api.typesafe.ai/v1/systemone with model jev-latest: call_ asks whether knowing the call was made should stay in context, and result_ asks whether the full output should stay verbatim.

What is the keep-or-drop policy?

A keep threshold of 0.5. If keepResult is at least 0.5 the output stays verbatim; if keepCall is at least 0.5 the result is dropped but the call and its head chars stay; if both are below 0.5 the call is dropped completely.

Does it modify my Cursor setup?

No. It is pure opt-in via agent --plugin-dir or cursor-jev agent, with zero modifications to ~/.cursor/hooks.json. Cursor's preCompact hook is observational, so native summarization still runs alongside the recovery block.

How is the state fitted to a budget?

Conversation history is fitted into a 25,000-token budget through six progressive stages, while the initial prompt and the newest messages (preserveRecentMessages, default 6) are pinned and never dropped.

What does it require?

The Cursor CLI (agent binary) only, Node 18 or newer, and pnpm 9.15.4 per the README badges. Captured tool I/O is stored under .data/conv_id/tools.jsonl and the scored sidecar under .data/conv_id/sidecar.json.

It cannot replace or suppress native compaction — the preCompact hook only sees the user message — so it works after the fact. The README excerpt is truncated mid-list, so the full state-fitting stages and quickstart are not covered here.

Jev directory

CommunityJev Integrations & Adapters

Cursor Clijev Compaction

Jev-scored context recovery for the Cursor CLI agent: hooks capture tool call and result I/O, two noul questions per call decide what to keep, and kept facts are re-injected verbatim after native compaction finishes.

See highlights Open github.com

Stars: 1
Language: TypeScript
License: MIT
Reading time: 2 min
Published by: Community
Author: kleosr
Last verified: Oct 2026

kleosr/cursor-clijev-compaction

User repository on GitHub

View on GitHub

TypeSafe Jev-scored context recovery for Cursor CLI (agent). Capture tool I/O, score keep/drop, re-inject after native compact.

Forks: 0
Last pushed: Updated Oct 2026

Repo stats from the GitHub API, cached Oct 2026.

Tagscommunity catalog-import logicrw cursor context-management

Works forJev for Decision gating Jev for Scoring

Used inJev for Software & DevTools

Languages: TypeScript
Platforms: CLI
Capabilities: Open sourceRequires API key

Highlights

Two noul questions per non-pinned tool call — one for the call, one for its result — go to POST https://api.typesafe.ai/v1/systemone with model jev-latest.
Keep threshold 0.5: at or above, the result stays verbatim; if only the call passes, the result drops and head chars stay; below both, it all drops.
Fits conversation history into a 25,000-token budget through six progressive stages, pinning the initial prompt and the newest messages (default 6).
Built for the Cursor CLI (agent) only; opt-in via agent --plugin-dir or cursor-jev agent, with zero modifications to ~/.cursor/hooks.json.
Because the preCompact hook only sees the user message, native summarization still runs; a sidecar feeds a stop hook and a sessionStart hook after compaction.

Watch out

Copy reviewed against the truncated README on 2026-10-15; the state-fitting stage list cuts off mid-word and the quickstart section was not included. Per its badges it needs Node >=18 and pnpm 9.15.4, and it works around Cursor's preCompact hook being unable to suppress native compaction.

Questions about this entry

Short answers drawn from the linked project or page.

From the Jev guides

Original write-ups on what this entry is for.

Decision gates in agent frameworks: what each integration ships

Oct 16, 2026

How Pydantic AI, Mastra, TanStack AI, DSPy, Spring AI, and LiteLLM expose Jev: where the typed decision sits in the loop, what each integration returns, and which big frameworks are only adjacent.

Jev Integrations & Adapters#agents#frameworks#integrations

Read article

From the Jev AI model community

Posts from builders shipping with Jev right now.

Follow @typesafeai