Problem

The AGENT USAGE panel showed only an estimated cost (the SDK's total_cost_usd, e.g. "$6.40 today · $6.40 wk") for Claude Code credentials. For an OAuth/subscription credential the meaningful number is the real 5-hour + weekly limit consumption — what Claude Code's /usage shows — not a dollar estimate.

Why it wasn't there already

The app already captures _meta._claude/rateLimit from the ACP stream and renders one bar from it (AccountLimitsSection) — but driving a real agent showed that signal is empty (agent_usage meta_keys=[]). The current Claude Agent SDK / ACP doesn't surface subscription usage. The dedicated GET /api/oauth/usage endpoint needs a user:profile scope the inference token lacks (403).

What does work (verified against the real token): the anthropic-ratelimit-unified-5h-* / -7d-* / -7d_sonnet-* headers ride on every /v1/messages response — the exact source Claude Code's /usage reads.

Change

• FastAPI (usage.py): fetch_subscription_usage() makes the cheapest possible /v1/messages ping (max_tokens: 1) with the credential's OAuth token and parses the unified headers into a multi-window { buckets: { five_hour, seven_day, seven_day_sonnet } } snapshot (utilization 0–1, reset timestamps). Best-effort — returns None on any error rather than disrupting a turn.
• Gateway (session_manager.py): at the turn boundary, when the SDK carries no rate-limit meta, fire a debounced (≤1/min per credential) background fetch that resolves the OAuth token, p ings, and persists the snapshot via the existing recordRateLimit. OAuth-only (api-key creds have no subscription windows and are skipped). No schema change — lastRateLimit is already v.any().
• Frontend (usage-display.tsx): accountUsagesFromRateLimit now parses the multi-window buckets shape into an ordered array (keeping the legacy flat single-window as a fallback), and AccountLimitsSection renders one ProgressBar per window — reusing the existing budget-bar UI verbatim.

Verified live

Drove a real Claude Code agent (Daytona sandbox, real OAuth credential). The ping fired (POST api.anthropic.com/v1/messages 200), the snapshot stored, and the panel rendered:

Claude account limits — Current session 14% (resets 3h 42m) · Current week 46% (resets 1d 16h) · Current week (Sonnet) 7%

Tests

• FastAPI: 7 new cases (header → buckets parsing, 429 still reports windows, network error → None). Full suite green.
• Frontend: multi-window parser tests (0–1 → 0–100 normalization, ordering, rejected-window, flat fallback). Full suite green; tsc + biome clean.

Note on labels

Claude subscriptions don't have a literal "daily" limit — the short window is a 5-hour rolling session. The bars are labeled accordingly ("Current session" + "Current week") with live reset times, matching Claude Code's /usage. The estimated cost line is kept below as supplementary detail.

🤖 Generated with Claude Code