fast-jev-compaction

What it does

When a Claude Code session gets long, Claude Code compacts it by asking the model to summarize earlier turns. Summaries are lossy: an exact error message, a file path, or a “never touch this folder” instruction can disappear. fast-jev-compaction replaces that summary with something more surgical. It keeps every user and assistant message word for word and only removes old tool calls and tool results, based on two yes/no questions Jev answers for each call: is the call itself still worth knowing about, and is its full output still needed?

Results Jev considers stale are trimmed to their first few hundred characters or removed along with their call. Everything kept stays verbatim. The same logic ships as an npm library, so you can use it outside Claude Code.

Setup

It’s a Claude Code plugin built on function hooks, an early-access feature that needs Claude Code 2.1.274 or newer and an opt-in flag:

{ "env": { "CLAUDE_CODE_ENABLE_FUNCTION_HOOKS": "1", "TYPESAFE_API_KEY": "<your key>" } }

Then install it from its own marketplace:

claude plugin marketplace add tamaratran/fast-jev-compaction
claude plugin install fast-jev-compaction@fast-jev-compaction

The install prompts for options such as the keep threshold. The API key can be entered there as a plugin option the manifest marks as sensitive, instead of sitting in plain text in settings.json. The package has no runtime dependencies. In our run, the type check passed, all 29 unit tests passed, and claude plugin validate passed with one cosmetic warning (no author listed). An open issue reports that the hooks aren’t recognized on Claude Code 2.1.272, so check your version first.

Using it

We couldn’t run a logged-in Claude Code session in our test environment, so we exercised the library directly: once with a fake Jev to see exactly what it sends, and once through the project’s live demo with a real key, which completed normally.

• The design is thoughtful. The first message and the newest messages are pinned and never touched. A result is never left without its call. When the conversation is too large for one Jev request, it shrinks what Jev sees in careful stages rather than failing outright, and splits questions across as many requests as needed.
• It’s honest about its limits. The README states plainly that a probability isn’t proof a result is safe to delete, and that token counts are estimates.
• The defaults need attention. Several open issues from users who replayed long sessions argue that the default keep threshold doesn’t fit how the two questions are worded, so the plugin may keep or drop less than you’d expect. Treat keepThreshold as something to tune, and our confidence thresholds guide explains how.

Permissions and data

claude plugin validate lists exactly what the plugin can touch: it hooks session.compact and turn.complete, makes HTTP requests, reads session usage, triggers compaction, and reads your TYPESAFE_API_KEY from its options, settings, or environment. It writes no environment variables and no files.

• Your conversation goes to TypeSafe, including tool inputs. We fed it a session in which Claude wrote a config file containing a fake secret, then ran a command that printed a fake password. The written file content and the command appeared in what the library sends to Jev (tool inputs are truncated to at most 1,000 characters). The command’s output did not: tool results are replaced by a one-line size note. In short: what Claude reads stays local, but what Claude writes and runs is shared. If your sessions handle secrets or proprietary code, decide whether that’s acceptable under your policies.
• It changes when compaction happens. By default it asks Claude Code to compact once context reaches 60%, earlier than you may be used to. If Jev fails (including a missing key) or can’t remove at least 25% of the history, it hands off to Claude Code’s built-in summary. In those cases you get the lossy summary you were trying to avoid, and sooner. Raise compactAtPercent if you’d rather keep Claude Code’s usual timing.

Maintenance

This is one of the most-starred Jev projects, and interest has outpaced upkeep. The last commit to the main branch was on September 17, the day it was created. Since then, most of the dozens of open issues haven’t had a single reply. Many are specific, well-written bug reports: no request deadline (a slow Jev call could stall compaction), a race in the auto-compaction lock, logging failures that can break the fallback path, and questions about the threshold’s scale. A community fork or a burst of maintenance could change that quickly, so check the repository before installing. MIT licensed.

Verdict

fast-jev-compaction tackles a real problem with a genuinely good idea: prune what’s stale, never paraphrase what’s left. The library code is careful and well tested. As a plugin, though, it’s early-access software on an early-access Claude Code feature, with open bugs, a threshold that likely needs tuning, and a data flow you should understand before enabling it. Try it on non-sensitive projects, raise compactAtPercent if early compaction bothers you, and watch the toast messages to see when it falls back to the built-in summary.

For a different approach to the same problem, Winnow filters tool output before it enters the context window instead of pruning it afterwards. To see all Claude Code tools built on Jev, browse coding agents and developer tools.

See how it compares with other tools in Best Jev tools, tested hands-on.

Review updated Sep 19, 2026. Numbers quoted from the project are its author's own; we don't publish our own measurements of Jev.