Subscription lane = Claude Code on the Max plan. No dollars on purpose — the plan is flat-rate, so the numbers that matter are the plan windows and how much context each session is carrying. Plan limits come from Claude Code's own /usage endpoint (the 5-hour session window, the 7-day window, per-model weekly caps, extra-usage credits), read through the Mac mini bridge with the Claude Code login — refreshed every 3 minutes. Context health, the 5-hour block, the 24-hour pulse and today's rows are live from the transcripts under ~/.claude/projects on the mini (deduped per message); earlier days and the MacBook come from the hourly spend collector. Context window = 1M tokens on every current model (200K on Haiku 4.5); Claude Code auto-compacts a little past the red mark at 90%. Block math follows ccusage: a block opens at the top of the hour of its first turn and lasts 5 hours.

API lane = pay-per-token keys — Martha, Patricia, Robert, Grace, the proxies. Spend, budgets and by-agent come from the WilHQ meter on claude-proxy.php / claude-stream.php (wilhq_ai_usage). Reconciliation and the org-wide tables read Anthropic's Usage & Cost Admin API (every key in the Console org) — the first live through anthropic-cost-proxy.php, the second stored hourly by the collector (archie_ai_spend / archie_ai_billing). Both need Admin-API access, which Anthropic withholds from individual organizations (403) — this Console is one, so these panels stay empty until that changes.