vercel-compaction

Why it exists

When a long agent session fills its context, the usual fix is to ask a model to summarize the old turns. A summary is lossy in exactly the wrong places: a file path, an exact error, a constraint or a command can vanish even though it matters an hour later.

This plugin never rewrites anything. It only deletes tool calls and tool results that an evaluation model judges no longer needed. User and assistant text stays verbatim and in order.

How it works

Every tool call is paired with its result. The first message and the newest few are pinned. For every other call the judge gets two yes/no questions: should the call stay, and should its result stay verbatim. The judge is Jev, an evaluation model that answers typed questions with calibrated probabilities and is priced per input token with no output cost, so scoring a whole session is cheap.

The conversation sent for scoring is fitted into a token ceiling in stages, each applied only if the one before was not enough. Questions are split into as many concurrent requests as needed and the answers merged. Per call, the outcome is keep both, keep the call and truncate the result to a short head, or remove both. No result is ever left without its call.

Failing safe

A failed request, a malformed or incomplete answer, a missing key, or a history that cannot be fitted all throw, and the plugin falls back to Claude Code’s built-in summary. If pruning would free less than a quarter of the context it also falls back. A failure never costs the session.

What leaves the machine

Prompts, assistant text and tool inputs go to the gateway at each compaction. Tool results do not; each is replaced by a one-line size note. The README says this before the install steps finish, and recommends project scope, so it is never switched on for a repo whose contents may not leave the machine.

The raw gateway key does not have to live in Claude Code either. A launcher asks Keysrs for a short-lived grant and hands the session a token and a loopback URL. When the token expires, compaction quietly falls back to the summary.

Credit

Adapted from fast-jev-compaction (MIT), which does the same pruning against TypeSafe’s own API. This project changes the transport to the Vercel AI Gateway: one key, no proxy process, no SDK dependency. It is also usable as a library.