IAmUnbounded/save-token-jev-clean
User repository on GitHub
- Forks
- 13
- Last pushed
- Updated Oct 2026
Repo stats from the GitHub API, cached Oct 2026.
Jev directory
CommunityJev Tools & CLIs
Save Token Jev
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.
- Stars
- 81
- Language
- TypeScript
- License
- MIT
- Reading time
- 4 min
- Published by
- Community
- Author
- IAmUnbounded
- Last verified
- Sep 2026
About this project
Portable, Jev-guided context compaction for coding agents. Instead of asking another LLM to rewrite old context into a lossy summary, save-token-jev asks Jev which tool calls and results still matter. User and assistant text is kept verbatim. A tool call can be kept with its full result, kept with a bounded result, or removed together with its result. The algorithm is implemented behind a normalized transcript model and host adapters so the same compaction policy works across multiple coding-agent runtimes.
Quoted from the project's README. Read the full README.
Highlights
- Each tool call is kept with its full result, kept with a bounded result, or removed together with its result; text blocks are never summarized or deleted.
- Adapters cover Codex CLI, OpenCode V2, Claude Code, the Anthropic API, and OpenAI Responses or Chat Completions transcripts, plus a generic format.
- Codex integration scores before built-in compaction and injects retained context at SessionStart; missing keys or low reduction fall back to native compaction.
- Node 20 or newer is required and the runtime package has no third-party dependencies; on macOS the key can live in the Login Keychain.
- The README notes token counts are estimates and that Jev probabilities are decisions, not proofs, so expensive sessions warrant a higher threshold.
Quickstart
bash
npm install
npm run check
export TYPESAFE_API_KEY="YOUR_API_KEY"
save-token-jev compact --format anthropic --input transcript.json --output compacted.jsonWatch out
MIT-licensed. Needs Node.js 20+ and a TYPESAFE_API_KEY (or a macOS Keychain entry); integrations fail open to the host's native compaction.
Questions about this entry
Short answers drawn from the linked project or page.
No. A tool result is paired by call ID and is never left behind when its call is removed, so you cannot end up with a result pointing at nothing. The first message and the newest six are pinned from deletion, and text and opaque blocks are never summarized or selected for deletion. A call can be kept whole, kept with a bounded result, or removed with its result.
Every host integration fails open. Missing keys, malformed responses, unsupported transcript changes, and reductions below SAVE_TOKEN_JEV_MIN_REDUCTION all fall back to the host's own compaction, and a provider timeout at 30 seconds triggers the same path. On OpenCode the plugin simply leaves the compaction result unset, so OpenCode runs its normal compaction unchanged.
That is the deliberate difference from a summarizing compactor. On Claude Code the function hook directly replaces compaction with the retained message list, so old context comes back as real messages rather than prose. Codex exposes no command-hook response that replaces its compacted transcript, so the plugin scores at PreCompact, lets Codex compact, then injects the retained context at SessionStart with source compact.
Treat it as an estimate. The state and request budgets are explicitly estimated token counts, and the local dashboard reports characters saved and estimated tokens saved rather than billed tokens. The README also frames Jev probabilities as decisions rather than proofs, which is why it suggests a higher keep threshold on expensive sessions.
Codex CLI app through PreCompact plus SessionStart(source=compact), which scores before built-in compaction and restores retained context after; OpenCode V2 through the compaction hook, replacing the summary; and Claude Code through function hooks, directly replacing the message list. There are also transcript adapters for the Anthropic and OpenAI formats and a generic adapter.
It needs Node 20 or newer and has no third-party runtime dependencies. Run npm install and npm run check, set TYPESAFE_API_KEY, then compact a transcript with save-token-jev compact --format anthropic --input transcript.json --output compacted.json; doctor inspects environment readiness, and the CLI prints decisions and statistics as JSON with diagnostics on stderr.
Import compactMessages with a TranscriptMessage array and options like preserveRecentMessages and keepThreshold, or use compact with a JevAsker you supply, so tests, gateways, and non-HTTP runtimes can bring their own transport. To add a host, implement TranscriptAdapter with canDecode and decode and pass it to decodeTranscript.
Reactions & coverage
Posts, threads, and videos about this entry from around the web.
X: The compaction plugin, in Chinese
YouTube: Ray Amjad puts Jev inside an agentic coding loop and looks at what it costs to run.
YouTube: RoboNuggets puts Jev to work inside Claude Code.
X: Decision outputs in Grep
More Jev resources like this
A 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
A Stop hook that keeps coding agents from quitting early: rules written in plain language are scored by Jev in about 0.7 seconds, and a violated rule sends the agent back to work instead of letting it stop.
Jev Tools & CLIs#community#python#claude-code
Community
noplan-inc/limpetPython
noplan-inc/limpetPythonA local-first MCP server that gives Claude Code, Codex, Cursor, and OpenCode structured quality scores from Jev across correctness, complexity, tests, security, and other dimensions while the agent writes code.
Jev Tools & CLIs#community#typescript#mcp
Community
NiazMorshed2007/jev-reviewTypeScript
NiazMorshed2007/jev-reviewTypeScriptA local gateway for coding agents that asks Jev which tool to call each turn and steers the LLM when Jev is confident, while every other request goes to your usual provider untouched.
Jev Tools & CLIs#community#typescript#gateway
Community
vinilana/jev-gatewayTypeScript
vinilana/jev-gatewayTypeScriptA Claude Code plugin from cth9191 that gives compaction to Jev: it keeps tool output that is still needed word for word and moves the rest to a drawer file, taking 0.43 to 0.74 s instead of about 35 s on one 591,489-token session. A fork of fast-jev-compaction.
Jev Tools & CLIs#community#github#claude-code
From 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.