What problem does it solve? Tuning an LLM context-compaction strategy is hard to evaluate without a controlled testbed. This Skill lets you design synthetic coding-agent sessions with known ground truth, run Jev's keep/trim/drop compaction over them, and measure needle recall versus token reduction so you can find settings that hold recall while cutting the window aggressively. ## Core Features & Use Cases - Synthetic session design: Write task vocabularies in session.json and validate them with check.mjs, controlling distraction, target, trimTo, budget, noise, focus, and taskEvery knobs. - Ground-truth measurement: Every tool result is secretly labeled detail needle, gist needle, or junk, so verdict.json reports true needle recall, junk removed, and reduction against a summarize-instead baseline. - Real transcript analysis: Point source at a Claude Code .jsonl transcript to get an offline compaction plan with keep/trim/drop counts, token reduction, cost, and timing, using the mock judge when no Jev API key exists. - Use Case: Start at distraction 0.1, raise it in steps of 0.2, find where recall crosses your target, then adjust target, budget, or trimTo one at a time and report which fix held recall at the largest reduction. ## Quick Start Ask the agent to design a Jev Compactor session with custom task vocabularies, run check.mjs, and report the distraction level where needle recall drops below target along with the settings that fix it.
