joelhooks/pi-fast-jev-compaction
User repository on GitHub
Pi extension: verbatim context compaction with TypeSafe Jev decisions
- Forks
- 3
- Last pushed
- Updated Sep 2026
Repo stats from the GitHub API, cached Oct 2026.
Jev directory
CommunityJev Tools & CLIs
Pi Fast Jev Compaction — joelhooks
A Pi extension that keeps conversation text verbatim while pruning stale tool history with Jev keep-or-drop decisions, falling back to Pi's built-in summary compaction when pruning cannot free enough context.
- Stars
- 12
- Language
- TypeScript
- License
- MIT
- Reading time
- 3 min
- Published by
- Community
- Author
- joelhooks
- Last verified
- Sep 2026
About this project
A stock Pi 0.85.1 extension that keeps conversation text verbatim while pruning stale tool history with TypeSafe Jev. This is deliberately layer 1 only. It does not summarize, rewrite, or create memory. When pruning cannot free enough context, Pi's built-in summary compaction takes over. Sends a compact representation of tool calls and surrounding conversation to TypeSafe's /v1/systemone endpoint. Asks two noul questions per eligible tool call: keep the call and keep its result.
Quoted from the project's README. Read the full README.
Highlights
- Asks two noul questions per eligible tool call - keep the call and keep its result - then applies keep, drop_result, or drop_call.
- Keeps the first message and recent messages pinned and filters only the context sent to the model; it never rewrites the JSONL session file.
- Decisions are stored as append-only fast-jev-decisions entries, and the ledger rebuilds after startup, reload, resume, fork, or tree navigation.
- Defaults include keepThreshold 0.5, preserveRecentMessages 6, compactAtPercent 60, truncateHeadChars 300, and a minReductionRatio of 0.25.
- It fails open: a missing key, malformed response, timeout, abort, network error, or insufficient reduction leaves Pi's summary path available.
Quickstart
bash
pi install git:github.com/joelhooks/pi-fast-jev-compaction
pi -e git:github.com/joelhooks/pi-fast-jev-compaction # one run without installing
export TYPESAFE_API_KEY=YOUR_API_KEYWatch out
MIT-licensed with the decision core ported from tamaratran/fast-jev-compaction; needs Node 24.18 and Pi 0.85.1 plus TYPESAFE_API_KEY, every ledger edit invalidates the provider's prompt cache from the first edited message, and decisions are model judgments, not proofs.
Questions about this entry
Short answers drawn from the linked project or page.
Only the context sent to the model. User text, assistant text, thinking blocks, images outside pruned tool results, custom messages, and relative message order all remain unchanged, and the JSONL session file is never rewritten. Decisions live in append-only fast-jev-decisions custom entries, and the ledger is rebuilt from the current session branch after startup, reload, resume, fork, or tree navigation.
That is the main thing to measure. Every ledger edit invalidates the provider's prompt cache from the first edited message onward, and the proactive turn_end trigger repeats that every cooldownTokens of context growth. Each entry records the cacheRead and cacheWrite counts from the last assistant message before the run, so you can compute deltas across runs; if writes dominate, raise cooldownTokens or compactAtPercent to prune less often.
It fails open and hands control back to Pi. A missing key, malformed response, timeout, abort, network error, impossible request budget, overlapping pass, or reduction below minReductionRatio all leave the built-in summary path available, and decisions merge into the live ledger only after a complete, valid result. It is deliberately layer one only, so it never summarizes, rewrites, or creates memory.
Yes to the attached images, and it says so. drop_result removes images attached to that tool result along with the stale body, keeping a deterministic head of truncateHeadChars characters; drop_call removes the call and its matching result entirely. Pairing depends on Pi's tool-call IDs, and the decisions are model judgments rather than semantic proofs, so raise keepThreshold or pin more recent messages if pruning feels aggressive.
Install with pi install git:github.com/joelhooks/pi-fast-jev-compaction, or run it once with pi -e. It targets Node 24.18 and Pi 0.85.1, and a fastJevCompaction block goes in ~/.pi/agent/settings.json or a trusted project's .pi/settings.json, where project values override global.
Export TYPESAFE_API_KEY before starting Pi, or set apiKeyCommand to a helper that prints the key to stdout. The extension trims and caches the key in memory for the session, never logs it, and never stores it in custom entries; failures produce generic errors without echoing output.
Three monotonic ones: keep preserves the call and its result pair, drop_result keeps the call plus a deterministic head of its result, and drop_call removes the call and its matching result. Decisions merge into the live ledger only after a complete, valid Jev result.
Reactions & coverage
Posts, threads, and videos about this entry from around the web.
X: The compaction plugin, in Chinese
X: Decision outputs in Grep
Hacker News: jevmem keeps Claude Code project memory by asking Jev which turns are worth saving.
avinashjetwani61 pointsJevmem – automatic project memory for Claude Code, built on JevRead the thread on Hacker NewsX: A usage router for Grok Bot
Other listings of this project
Forks and copies of this project in the directory. This entry is the canonical one.
Pi package that uses Jev to decide, call by call, which old tool results can leave model context, with a savings gate that rejects prunes too small to be worth the prompt churn and bounded excerpts from status documents.
Jev Agents & MCP#community#awesomejev#catalog-import
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.
Jev Agents & MCP#community#jevbest#catalog-import
More Jev resources like this
A Jev-guided context compaction library for coding agents: instead of summarizing old context, Jev decides which tool calls and results still matter while user and assistant text is kept verbatim.
Jev Tools & CLIs#community#typescript#compaction
A Pi extension that turns constraints stated in conversation into a scoped, replayable policy and checks every side-effecting tool call against it before it runs, with Jev only classifying how each message changes the policy.
Jev Tools & CLIs#community#typescript#guardrails
Community
Nyarlathoteppppp/pi-heedTypeScript
Nyarlathoteppppp/pi-heedTypeScriptA hybrid coding harness for Pi where any LLM writes the code while Jev returns typed probabilities that deterministic policy code turns into allow, confirm, block, warn, or steer.
Jev Tools & CLIs#community#typescript#agents
Community
AbdelStark/bicameralTypeScript
AbdelStark/bicameralTypeScriptA Codex plugin and npm library that wraps native session compaction: Jev scores every tool call and result, and the verbatim history it keeps is re-injected after Codex summarizes. The README does not recommend using it.
Jev Tools & CLIs#community#typescript#codex
A Claude Code plugin and npm library that replaces transcript summarization with Jev keep-or-drop judgments, deleting stale tool calls and results while kept text stays verbatim.
Jev Tools & CLIs#community#typescript#claude-code
An MCP server and agent skill that verifies coding-agent work by driving the real running app and returning pass, fail, or couldn't-tell verdicts with the file:line to fix; its listing frames the approach as Jev-style machine-native runtime perception.
Jev Tools & CLIs#community#typescript#mcp
Community
reticlehq/reticleTypeScript
reticlehq/reticleTypeScriptFrom the Jev guides
Original write-ups on what this entry is for.
From the Jev AI model community
Posts from builders shipping with Jev right now.