/insights shows where the current session's tokens and time went, ranked by what a change would save, each with a concrete fix. chanto insights does the same for saved sessions in this directory: the last 10 by default, --last N, or --session <id>.
It flags:
- Context bloat: single tool outputs over 10% of the context, large files read whole when only a small part was edited, files read again without changing in between, and commands with long output.
- Prompt-cache breaks: calls where the start of the prompt changed (system prompt, tool list, cleared or edited history), so a local model had to re-process its whole context. Each break is attributed to its cause.
- Wasted turns: the same failing command rerun with no edits in between, loop-guard hits, verification retries, empty or truncated responses, and compactions.
Model time (including prompt processing), tool time and token counts are recorded per session and saved with it. When one tool output takes over a quarter of the context, Chanto shows a one-time tip. Token savings are estimates.