Can it silently corrupt my session by dropping something load-bearing?

The extension never rewrites content and the session file on disk is untouched, so pruning is a per-request view. Calls in the first message and the newest preserveRecentMessages are pinned, and any Jev error, malformed answer, unfittable state or broken call and result pairing sends the unmodified messages instead.

Are the net percentage gains trustworthy?

The ratios are, and the absolute numbers are not, by the author's own statement. The token model is a formula and pi's reported usage runs at about 0.65 to 0.67 times it on these sessions. The percentages all compare one basis, but both sessions were roughly 90% tool output, mostly base64 screenshots, and no text-heavy session was measured.

What happens if I tune it badly?

The table in the README shows how. Raising the trigger to 0.85 so it only fires on a near-full window drops the net gain to 0.5% because the cache penalty eats the saving. Cutting the state to 8,000 tokens drops it to 14.0%. Disabling the cache gate multiplies Jev requests by 11.

Why does it refuse to delete whole tool calls the way upstream does?

Deleting a call outright would add roughly 10 points more net gain, because it removes the record of write, edit and shell commands that launch harnesses. This port treats those as not provably re-runnable and keeps the call, so you get a smaller saving in exchange for not losing launch commands.

How do I install it into pi?

Install it as a pi package with pi install git:github.com/Quentindanblon/pi-fast-jev-compaction, or clone it into ~/.pi/agent/extensions/fast-jev-compaction, then restart pi or run /reload.

Where does it read the API key, and what happens without one?

It checks FAST_JEV_API_KEY, JEV_API_KEY and TYPESAFE_API_KEY, then apiKey in the config, then ~/.pi/agent/fast-jev-key. Without a key the extension is a silent no-op.

What does it actually ask Jev about each call?

Each non-pinned tool call gets two noul questions: whether the call should stay, and whether its result should stay verbatim. Questions are batched so state plus questions fit under maxRequestTokens.

CommunityJev Agents & MCP

Pi Fast Jev Compaction — QuentinDanblon

pi extension that never summarizes: it asks Jev whether each old tool call and tool result is still needed, then drops or truncates only those, passing text, thinking and images through verbatim in a per-request view.

See highlights

Open github.com

Stars: 3
Language: TypeScript
Published by: Community
Author: QuentinDanblon
Last verified: Oct 2026

QuentinDanblon/pi-fast-jev-compaction

User repository on GitHub

View on GitHub

Verbatim context pruning for the pi coding agent, scored by TypeSafe Jev: stale tool calls and results are dropped or truncated, everything kept stays verbatim.

Forks: 1
Last pushed: Updated Sep 2026

Repo stats from the GitHub API, cached Oct 2026.

This is one of several Jev listings for the same project. The canonical entry is Pi Fast Jev Compaction — joelhooks.

About this project

Verbatim context pruning for the pi coding agent, scored by TypeSafe Jev (System One). Port of tamaratran/fast-jev-compaction, MIT.

Quoted from the project's README. Read the full README.

Tagscommunity jevbest catalog-import

Works forJev for Classification Jev for Tool calling

Used inJev for Software & DevTools

Languages: TypeScript

Highlights

Two noul questions per non-pinned tool call, batched so state plus questions stay under the request token cap, with read-only tools treated specially.
If anything is uncertain, including a Jev error, a malformed answer or a state that cannot be fitted, the unmodified messages are sent.
Decisions are persisted as taken and restored on session_start, so a reload or resumed session does not re-pay for scores already bought.
Upstream deletes whole tool calls; this port removes a call only when it is provably re-runnable, giving up about 10 points of net gain.

Watch out

The percentages are net ratios, not absolute token counts: the author states pi's own reported usage is roughly 0.65 to 0.67 times this model on these image-heavy sessions and that image pricing is a formula, so only the ratios are the defensible claim. Both measured sessions were about 90% tool results, mostly base64 screenshots, and no text-heavy session was available, so do not expect the same result on prose. Thinking blocks, user text and assistant text cannot be touched at all. Absolute net gain with the default background scoring was 31.8% and 37.7%; scoring on the request path gave 33.6% and 40.9% and cost 2,984 ms in the hook over a session versus 322 ms in the background. Without a key the extension is a silent no-op, and there were zero failures in the two runs. Copy reviewed against the repository README on 2026-10-04, not from running it.

Questions about this entry

Short answers drawn from the linked project or page.

More Jev resources like this

duet-agent

Agent harness for work that outlives one chat, with resumable turns, routed state machines, and observational memory across sessions and processes.

Jev Agents & MCP#community#jevbest#catalog-import

Community

dzhng/duet-agentTypeScript

skilled-agent-harness_spec-driven-loops

Assistant framework that stores specs, memory, and specialist agents as plain markdown files in your repo for Claude Code, Codex, Opencode, and similar runtimes.

Jev Agents & MCP#community#jevbest#catalog-import

Community

MichelKerkmeester/skilled-agent-harness_spec-driven-loopsTypeScript

pi-jev-skill-picker

Pi extension that strips the full Agent Skills catalog from every request and loads only the skills a Jev ranking call judges relevant.

Jev Agents & MCP#community#jevbest#catalog-import

Community

safzanpirani/pi-jev-skill-pickerTypeScript

jevgrep

102

Behavioral code search that asks Jev to score repository fragments for a plain-language question and returns exact excerpts with paths and lines.

Jev Agents & MCP#community#jevbest#catalog-import

Community

nassim-arifette/jevgrepTypeScript

Multi-Agent

Local engine with TypeScript and Python SDKs that runs Claude Code and Codex sessions as a warm team with a durable mailbox.

Jev Agents & MCP#community#jevbest#catalog-import

Community

masonlee39/Multi-AgentTypeScript

oh-my-harness

Single-command harness that turns a plain-language stack description into synced docs plus blocking hooks for Claude Code, Codex, and Pi, with deterministic presets or one Jev chooser call.

Jev Agents & MCP#community#jevbest#catalog-import

Community

kyu1204/oh-my-harnessTypeScript

Back to all resources

From the Jev community

Posts from builders shipping with Jev right now.

Follow @typesafeai