The pace line shows how far above or below a steady burn you are — not total spend. A steady pace that would land you exactly at the limit when the window resets plots as a flat line at zero. Under-pace dips down (green); over-pace climbs up (yellow → red).
Spending limit: Anthropic doesn't publish a hard dollar number for Max tiers, so we work it out from /api/oauth/usage — the same feed Claude Code's /usage command reads. If that endpoint is unreachable, we fall back to the observed ceiling (~$5,900/week).
Reset time: pulled live from /api/oauth/usage. The reset anchor is usually stable but Anthropic can shift it (bug fixes, billing changes), so we always trust the live timestamp over a day-of-week assumption.
What it shows: the sessions you ran in the chosen window whose peak context hit at least 500K, ranked by how heavy they got. Anything lighter than 500K is filtered out so the list stays focused on the ones where doing something would actually matter. The action column suggests what to do based on peak context:
/compact soon/compact is still safe here/compact can drop important early context this close to the edge. Better play: /save a summary, start a new session, and pick up from that summary.Starting fresh beats compacting when you're near the 1M wall, when you've already compacted recently (less to strip), or when the chat has accumulated dead-end explorations you don't need to carry forward.
The "last turn" column tells you how recent the session is — just now means live right now; older entries help you spot which finished sessions burned the most so you know where to compact, save, or restructure similar work next time.
Session IDs here are the first 8 characters of the Claude Code session filenames under ~/.claude/projects/.
Loading…
Tiles compare the most recent 7 days to the 7 days before that, sorted by direction of change — worse on the left, better on the right. Tap a tile to plot that metric in the chart.
Loading…
What it shows: the chart is daily or weekly aggregates over your lookback. The strip above is a quick read on current pace: either 7-day averages vs the week before, or a rolling 24-hour snapshot vs the same prior-week daily baseline.
Reading the trend: bars show individual days (or weeks, if you toggle Week); the line is a smoothed running average. The two dashed markers point to capability events (1M context turning on, Opus 4.7 launching) — useful anchors for where habits often shift.
How to spend less:
/compact more often, Grep before Read, hand research to sub-agents so their context doesn't pile onto yours.