Mining Claude Code transcripts: dir layout and noise filtering

Past sessions live at ~/.claude/projects/<cwd-slug>/<session-id>.jsonl, where cwd-slug is the working directory with / replaced by -. A leading dash means shell globs break: ls ./-Users-foo works, ls -Users-foo is parsed as flags. Because the slug is per-cwd, one logical project can be split across several dirs (repo dir plus home dir), so grep every dir, not just the obvious one. Ranking candidate files by keyword hit count plus mtime finds the right session fast. Critically, raw grep counts are dominated by noise: Stop-hook feedback text, <system-reminder> blocks, skill preamble dumps, and <local-command-stdout> all appear as type:user entries, so a real signal of 275 human turns can hide behind thousands of matches. Parse the JSONL, keep type==user, extract only content blocks of type text, then drop lines matching Stop hook / Base directory for this skill / command-name / local-command / task-notification before reading.