Audit a coding session from its transcript JSONL to find time sinks

Claude Code stores each session as ~/.claude/projects/<slugified-cwd>/<session-id>.jsonl, one JSON object per line. To critique how a user actually worked (not just the output), parse it: filter type==user and skip entries whose text is a Stop-hook or ide_opened_file injection to get real prompts; diff consecutive prompt timestamps to get per-prompt wall-clock and spot idle gaps that separate real working time from elapsed time. Then compute the explanation-to-action ratio: count assistant messages containing a tool_use block vs pure-text ones, and sum assistant text chars — 115K chars of prose is ~90min of human reading and is usually the real bottleneck, not the coding. Finally, regex the whole transcript for the first appearance of a key finding to timestamp when it was discovered versus when the enabling data was already on screen; that gap is the single most actionable speed lever.