Fast Jev Compaction — npmjs.com

Run npx jev-setup, which installs all CLIs globally and wires the chosen agent: Claude Code via the plugin marketplace, OpenCode via a session hook, Codex via AGENTS.md, Hermes via a plugin, or a generic AGENTS.md. Only an OPENROUTER_API_KEY is required.

Most context compaction asks a model to summarize old turns, and a summary is lossy, so a file path, exact error or command can disappear. This library never rewrites anything: it only deletes tool calls and results Jev says are no longer needed, while user and assistant text stays verbatim and in order.

Input is a JSON array of objects with role, text, toolUses and toolResults; node cli/jev-compact.mjs transcript.json -o dump.md writes the compacted transcript. A typical run goes from 62 to 6 messages and 41K to 4.6K characters for about $0.00002, and the output begins with a stats line.

One batched call of 3 Noul plus 1 Score over the last commit's diff looks for hardcoded secrets, syntax errors, breaking changes and test-failure risk. It exits 0 for ALLOW and 1 for BLOCK, with a default threshold of 0.85 that can be changed via --threshold; a diff containing an sk- key blocked at p=0.99.

The fork's jev-compact and jev-gate CLIs run Jev through OpenRouter's /api/alpha/decisions with model typesafe/jev-1.13, so only OPENROUTER_API_KEY is needed. The upstream library and Claude Code plugin can be pointed at OpenRouter instead by setting the baseUrl library option or TYPESAFE_BASE_URL plus TYPESAFE_MODEL.

An API failure also blocks, so the guardrail is fail-safe rather than silently allowing a push. The check takes roughly 400 ms and costs about $0.00005 per run.

State is fitted into maxStateTokens, 25k by default, and questions are split so state plus questions stays under maxRequestTokens, 30k by default, below Jev's 32k request limit. If the state still does not fit after staged truncation, compaction throws.