improvement-extractionlisted

# Improvement Extraction Scan the current session for candidate CLAUDE.md/AGENTS.md baselines or candidate skills, and write each as a structured note to a configurable output folder. This is a scratch capture step — candidates land here for later manual review, not automatic promotion into this repo's baselines or skills. ## Step 1 — Resolve the output root - Read `IMPROVEMENTS_ROOT` from the environment. - If unset, ask the user for the path once, then use it for this run only. Tell them how to persist it so this step isn't needed again: - Windows: `setx IMPROVEMENTS_ROOT "<path>"` - macOS/Linux: `export IMPROVEMENTS_ROOT="<path>"` in their shell profile - Never write the resolved path into this skill file or any committed content — the environment variable is the only place it lives. ## Step 1b — Full-session sweep mode (only when explicitly asked) Default scope is "the current context" — what's actually visible in this conversation right now, which after one or more compactions is only a summary of most of the session, not the real thing. That's fine for a normal end-of-task run. It costs roughly 100-120k tokens per chunk (a handful of chunks for a normal session, ~8-10 for a multi-day one), so this mode is never the default — only enter it when triggered: **First, check whether it would buy anything.** This mode exists to recover what *compaction dropped*. If no compaction has happened and the whole session is still visible in context, the default scope already